AI Absorption Ledger

ExxonMobil enters with Q1 and Q2 2026. All the steps are in Q2: AI disclosure widened from a silent first quarter, and three channels opened. Q1 has no channel, and none of its stored sources mentions AI, including the prepared remarks and slides published with the call. Its call was taken up by the war in the Middle East, refining, LNG in Qatar, Guyana and Venezuela; the CEO reported a milestone in the enterprise-wide process and data platform transformation (claim c2) and a fully autonomous deepwater well section built on rig automation (claim c3), neither with AI named.

Exploration prospects. The Q2 prepared remarks say the company has built exploration models powered by AI and its seismic database; one model, trained on Guyana discovery data, identified 90% of known discoveries in validation testing and has found 4 new Guyana prospects not found by traditional methods, and the slides list the same opportunities. On the call the CEO relays the count from an earlier conference remark and says much more work is needed to confirm the prospects. The only measure of output is a count of unconfirmed prospects, so the channel is read directional and left unsized, with no ballpark: a count sizes nothing and no prospect has been drilled. It is tagged relabelled: on the Q1 2024 call the CEO described the same work, updating reservoir models from each well and each seismic survey to look for new opportunities in the block, and in coverage no reserve, production, revenue or expense line moves because of AI. Exploration expenses were $155mn against $251mn a year earlier, a line the 10-Q does not explain by AI.

Permian and Guyana drilling. The CFO named AI machine learning beside extended reach laterals and surfactants as contributing to very strong Permian performance (claim c5), and a Guyana slide credits the developments' pace and cost advantage to the design one, build many model, AI-enhanced drilling performance and strategic partnerships together (claim c15). Each is registered as a joint-cause channel, as UPS's network planning remark was: described, unsized, with the other causes as confounds and no ballpark, the Guyana one also because the saving would fall on capitalized drilling. The prepared remarks and slides do not change the Permian reading, and for Guyana they give the measured advantage to the project model without naming AI. Both are tagged relabelled against the anchor's drilling and completions technology.

Beside Chevron. Chevron, the other Energy company here, carries two facilities channels for Project Kilby, a power plant contracted to Microsoft's data center, and keeps its CEO's remark that AI will change exploration as context because it was in the future tense. ExxonMobil reads the other way round. Its data-centre power is still a conversation: in Q1 the CEO said the company is not interested in the utility business of providing power and is in continuing discussions with a number of hyperscalers about low-emissions power from natural gas with carbon capture (claims c4 and c5), with AI not named and nothing contracted, and Q2 does not mention it. Its exploration use is in the past tense, with a tool in use and a count, so it opens a channel where Chevron's did not.

What is not a channel. The CEO said the data platform transformation will help accelerate the adoption and value of AI (claim c1), an aim with no tool or measure, and the Q2 10-Q and release added AI-enhanced technologies to their cautionary statements (claims c6 and c7); both are kept as context. The structural cost program, on the roster as a possible AI door, reached $16.30bn of cumulative savings since 2019, which the CEO credits to the transformation and centralized organizations and the 10-Q defines as operational efficiencies, workforce reductions and divestments; no source attributes any of it to AI, so it is a confound and not a saving. ExxonMobil's Q3 2026 report is due around the end of October.

Reading associate-productivity in Q3: was inferred, sized at the former estimate wmt-2026-cq3-f18, which repeated the Q1 decomposition; now described and unsized, because the CEO's statement that the company uses AI to make associates' work easier names no tool, deployment or measure and the CFO credits wage leverage to technology tools and supply chain automation without naming AI. The step comes from the wording, not from a change at the company.

Reading sparky-shopping-agent (all quarters): unchanged, because an order-value gap between users and non-users is a measure for AI-attributed demand, not a count.

Readings inventory-fulfillment-ai in Q1 and Q2: were inferred, sized at the former estimates wmt-2026-cq1-f23 and wmt-2026-cq2-f21; now described and unsized, because the CFO credits inventory efficiency to technology, AI and automation together and the CEO pairs AI with supply chain investment, and nothing separates AI's part. Automation is added as a confound.

Readings inventory-fulfillment-ai in Q1, Q2 and Q3: motive was narrative-defensive, now exploratory, because AI is named as one of several enablers with no measure and no line moving, and no narrative-defensive tell of its own is quoted; Q3 carries it.

Figure wmt-2026-cq2-f23 (Q2 advertising revenue from AI features in ad tools): was the ledger's own phasing of the prior fiscal year's advertising total, now the fixed annual-plan band on the fiscal 2026 amount, $6.40bn, because an amount for a year converts to a quarter only through that band; the amount is a completed year's actual, not a plan. The estimate is now $6.4mn to $96mn.

Reading associate-productivity in Q3: motive was narrative-defensive, now exploratory, because the CEO's statement names AI with no measure and no line moving and the quarter quotes no narrative-defensive tell of its own; Q1 keeps narrative-defensive on its training and heavy-talk tell.

Readings engineering-ai-coding in Q1, Q2 and Q3: motive was unknown, now exploratory, because unknown is only for sources that contradict each other and the reference quarter gives a usage share with no cost movement.

Channels inventory-fulfillment-ai and associate-productivity: domain was cost-of-revenue and other, now operations, because inventory positioning, fulfillment and store, club and supply chain work are now an operations domain.

Channel ai-vendor-bill: counterparty was model-provider, now mixed, because the bill pays model partners, vendors of AI tools for associates and the software and cloud providers whose AI share sits in it.

Not changed: ad-tools-ai stays sized, since the CFO lists AI features among the toolkit's enhancements rather than as a joint cause of a measured movement, and associate productivity keeps its Q1 narrative-defensive motive.

Wells Fargo enters with Q1 and Q2 2026 and these channels: investment in AI inside the technology budget, customer self-service through the Fargo virtual assistant, headcount efficiency credited partly to AI, and financial advisor productivity from GenAI desktop tools. The technology investment and the advisor tools are existing work into which AI was put, read as expanded. Fargo and the headcount program are activities the anchor already shows (the FY2024 annual report already said the company uses artificial intelligence in operations and customer service, and the Q1 2024 call already credited efficiency initiatives with a headcount falling every quarter since 2020), read as relabelled because no movement is attributed to AI.

Q1 2026. The CEO named AI inside a wider increase in technology investment, and said Fargo, the AI-powered virtual assistant, passed 1,000,000,000 customer interactions since its launch. Technology, telecommunications and equipment expense rose 14.2% to $1.40bn, which the 10-Q explains by software, hardware depreciation and internally developed software. Headcount fell −6.7% over the year, the CEO's 23 consecutive quarters of reductions, credited to efficiency initiatives and not to AI.

Q2 2026, the steps. New channels opened, each described and unsized. The CFO said technology and AI help the company run with fewer people in a different way or faster than in the past, while the CEO, the CFO and the release credit the reduction, 7% on the year, to efficiency initiatives; personnel expense of $8.85bn is the ceiling. The CEO said Advisor Gateway, a desktop with GenAI capabilities, launched in the quarter and credited productivity and advisor hiring to investments like it, part of a platform modernization costing over $1.00bn; wealth revenue of $3.89bn is the ceiling, and since advisers in consumer branches moved to the consumer segment in Q1 2026 it may not cover them all. Fargo went unmentioned, a step from directional to not mentioned.

Context, not a channel. Asked by an analyst about exposure to the AI industry, the CEO described the bank financing the pieces of the data-centre build-out it is used to underwriting and named the LLM provider renting the space as the credit risk behind long-dated loans. Management ties no balance, fee or revenue to AI, so the ledger keeps the remarks as context rather than a lending channel, unlike Bank of America and JPMorgan, whose finance chiefs credited part of loan growth to AI.

What remains. No channel is sized and no ballpark is built. Fargo's count is cumulative with no start date and no prior-year level, so it supports direction, not dollars; a stored reference quarter with an earlier count, and a sourced cost per contact, would allow a saving against the prior-year volume. The other channels name AI beside another cause and nothing separates AI's part, so the ledger quotes the line the AI part sits inside. The quiet is itself the reading: the technology and personnel lines the AI remarks would move are explained in the filings without AI.

Reading network-operations-ai in Q1: was described and inferred, sized at the former estimate vz-2026-cq1-f30; now directional and unsized, because the quarter's measures are a share by count, 85% of issues resolved autonomously, and a count of about 20 kits. Q2, which states a rate of the work (fixes in minutes rather than hours), keeps its size.

Reading central-office-edge-inference in Q2: was inferred, sized at the former estimate vz-2026-cq2-f30; now described and unsized, because the money has not started: the retrofits are in the early stages, the new AI infrastructure revenue begins to layer into results starting in 2027, and the only use is a small trial.

Reading ai-infrastructure-capex in Q2: was inferred, sized at the former estimate vz-2026-cq2-f31 on a judgment share of capital expenditures; now described and unsized, because the CFO names AI infrastructure builds among the fiber deployed to capture growth opportunities and nothing separates AI's part. Capital expenditures, $4.01bn, are quoted as the ceiling. It is capital, so no flow total moves.

Channels central-office-edge-inference and ai-infrastructure-capex: layer was end-use (absent), now facilities, because the money is power-ready space in central offices sold to data-centre customers and the cost of providing fiber routes and space to them. Novelty stays expanded on both, since capacity is being retrofitted and contracted for data-centre customers in coverage. Channel ai-connect-fiber is left as it is.

Figure vz-2026-cq1-f31 (Q1 network energy saving): was the stated amount over the ledger's own span of quarters, now the fixed undated band, because an amount with no period converts only through that band; it is now $17mn to $75mn, with the same low end and a lower point and high end.

Reading care-voice-agents in Q1: was inferred, sized at the former estimate vz-2026-cq1-f28; now described and unsized, because the CEO credits lower costs to AI and automation together, beside a push to digital channels, and the voice agents are still being tested.

Reading ai-offer-personalization in Q1: was inferred, sized at the former estimate vz-2026-cq1-f32; now described and unsized, because AI-enabled processes are one of several named causes of the lower cost of acquisition and retention, 35%, with no AI share.

Readings network-operations-ai and network-energy-ai (both quarters): motive was unknown, now exploratory, because the sources do not contradict each other; AI is named with no cost line moving.

Channels network-operations-ai and network-energy-ai: domain was cost-of-revenue, now operations.

Channel care-voice-agents: counterparty was unknown, now mixed: the company's own care staff and outsourced care.

Channel ai-connect-fiber: counterparty was enterprise, now mixed, because the buyers named are hyperscalers (Google and Meta), alternative cloud providers and large enterprises.

Verizon enters with these channels. On the cost side: care work done by AI voice agents and tools, software delivery and vendor support done with an AI coding tool, network operations done by AI models, network energy saved by AI, acquisition and retention spend lowered by AI micro-segmentation, the AI vendor bill, and the internal AI tech stack build. On the revenue side: fiber transport sold for AI infrastructure (AI Connect) and, from the second quarter, central offices retrofitted as edge data centers for inference, with the capital spending on those builds shown and left out of totals. The vendor bill is new money; care, coding, network operations, the stack build and AI Connect are existing lines AI changes; energy and micro-segmentation were already described as AI or machine learning at the anchor and are tagged relabelled, so they count for nothing in the incremental total.

Q1 2026. The new CEO named voice agents from Sierra, ElevenLabs and Google in some customer service operations, still being tested, with customer satisfaction up 12.8%; an AI coding tool across software development with a vendor-support cut of over 70% as a target; 85% of network issues resolved autonomously; over $200mn of energy savings; and a layered AI tech stack built with Google and Anthropic. The CFO credited the quarter's savings to advertising, network operations and a workforce reduction of 13,000, and the 10-Q credits every personnel decrease to workforce reduction initiatives; the operating expense savings target of $5.00bn carries no AI share. The release says nothing about AI.

Q2 2026. Steps: AI Connect moved from described to bounded with the Google agreement and other deals expected by year end; edge inference space and AI infrastructure capital spending opened; network energy moved from bounded to not mentioned; care, the coding tool, micro-segmentation and the vendor bill moved to not mentioned. The call's remaining cost statements are AI models fixing network issues in minutes and a company-level tie between becoming AI-centric and the operating leverage, which is not read as a care saving. The 10-Q records severance charges of $397mn for workforce reduction initiatives, and the program targets at least $9.00bn of operating and capital savings, neither attributed to AI.

The ledger's own sizes in the latest quarter: network operations $1.8mn to $54mn, the AI tech stack build $4.4mn to $94mn, AI Connect revenue $716k to $36mn (revenue booked since 2024-CQ4, called small then), edge inference revenue $0 to $7.2mn, and AI infrastructure capital spending $0 to $160mn out of quarterly capital expenditures of $4.01bn. In the first quarter the ledger also put care savings at $1.5mn to $98mn, energy at $17mn to $75mn, the vendor bill at $3.4mn to $52mn, the coding saving at $0 to $87mn and the micro-segmentation saving at $0 to $41mn. AI Connect is capacity sold to the buildout and is read as the buildout's money moving. The roster expected a care door confounded by workforce programs; the sources bear that out, and management gives no AI share of any care cost.

Reading ai-platform-integrations in Q1: was inferred, sized at the former trace upwk-2026-cq1-f35; now described and unsized, because the Upwork app in ChatGPT launched the month after the quarter and the earlier OpenAI partnership states no money or term.

Readings human-supervised-agents in Q1 and Q2 and ai-data-opportunity in Q1: were inferred, sized at zero with a trace or pilot high end (upwk-2026-cq1-f36, upwk-2026-cq2-f44, upwk-2026-cq1-f37); now described and unsized, because the agents product was in testing with an undated launch and the data opportunity was conversations and pilots, and a product not yet launched is never sized at zero.

Readings uma-and-ai-features in Q1 and Q2: were inferred, sized at the former estimates upwk-2026-cq1-f34 and upwk-2026-cq2-f42; now described and unsized, because the 10-Q names AI capabilities as one of several initiatives that together lifted the take rate, and the filing and the call credit ads, monetization products and Business Plus. Marketplace revenue is quoted as the ceiling.

Reading ai-search-acquisition-toll in Q2: was directional and inferred, sized at the former estimate upwk-2026-cq2-f41; now directional and unsized, because the new-client decline is credited to AI beside the labor market, macroeconomic uncertainty and the company's own strategy, with no split. The GSV decline is quoted as the ceiling.

Checked and not changed: low-end work automated away in Q2 stays sized, since the CEO places the erosion inside the stated pool of exposed GSV.

Channel ai-product-build: counterparty was internal, now mixed, because the build pays the company's own engineers in research and development and the third-party model and hosting providers behind Uma, which no source splits.

Reading ai-build-investment in Q2: was inferred, sized at the former estimate ups-2026-cq2-f17 (a reference-class share of revenue); now described and unsized, because the CEO's only statement about the money names RFID and AI together and nothing separates AI's part, and wave C put capital named together with AI under the joint-cause rule. Capital spending on information technology stays in metrics as the ceiling, not as a base.

Figures ups-2026-cq2-f15 and ups-2026-cq2-f17: kept in the exhibit, unreferenced as sizes and relabelled as former estimates, because the published joining entry cites both.

Reading network-planning-ai in Q2: unchanged, described and unsized under the same joint clause.

Reading network-planning-ai in Q2: was inferred, sized at the former estimate ups-2026-cq2-f15; now described and unsized, because the CEO names RFID and AI together as helping the company gain efficiencies and gives no measure of AI's part.

Readings network-planning-ai and ai-build-investment in Q2: motive was narrative-defensive, now exploratory, because AI is named with no measure and no line moving, and no narrative-defensive tell of its own is quoted: the position cuts are credited to the Driver Choice Program and the network reconfiguration, not to AI.

Channel network-planning-ai: domain was cost-of-revenue, now operations, because network planning, routing and execution is now its own domain.

UPS enters with Q1 and Q2 2026 and with channels opened in Q2 only: operating cost avoided by AI network planning, routing and execution (savings), and the company's investment in AI for the network (spend). The anchor already shows network planning without AI: on the first call of 2024 the CFO credited Total Service Plan and network planning tools with cutting operational hours by 6.6%, and the 2024 10-K credited the company's systems with network efficiency and end-to-end visibility, so the saving is read as relabelled. The investment is an existing technology budget, expanded and sized as a level with no traced baseline. The CEO's joint RFID and AI remarks about winning and keeping customers support no AI revenue channel: every customer measure she gives is RFID's.

Q1 2026, the quarter ended 2026-03-31. No source names AI and no channel is registered. The 10-Q credits program cost savings of $600mn to the Network Reconfiguration and Efficiency Reimagined initiatives; the CFO counts operational positions down by 25,000; the CEO credits productivity to automation, with 67.5% of buildings automated and cost per piece 28% lower in an automated building. None of it is attributed to AI, so it is a confound and not a saving.

Q2 2026, the quarter ended 2026-06-30. Steps: AI disclosure widened, from a call with no AI passage to the CEO's prepared remarks on RFID and AI; both channels opened, each read described. The CFO says operational hours moved down with volume in the first half. The 10-Q names no AI: it explains lower labor cost by headcount reductions, the Ground Saver outsourcing and lower volume, and a rise of $42mn in technology expense by software costs and application fees.

Both sizes are the ledger's own: $2.4mn to $19mn of U.S. Domestic Package compensation ($4.79bn) avoided by AI planning, held small because hours tracked volume; and $2.3mn to $16mn of income-statement AI spending, a share of revenue at or below Progressive's rate. The roster expected a sparse ledger and the sources bear it out: on the supply chain and automation door it shares with Walmart, UPS credits its savings to network reconfiguration and automation, and names AI only as a capability.

Channel corporate-function-automation: withdrawn from the ledger as not an AI channel under the rule for a one-clause aim with no tool, deployment or measure. Its readings in Q1 and Q2 were inferred, sized at the former estimates unh-2026-cq1-f42 and unh-2026-cq2-f41; now not-mentioned and unsized in both quarters. The id is kept because the entry 2026-10-06-unh-joins cites its former Q2 size. Its passages stay in the exhibits as context claims: claim c35 now carries no channel, and claim c30 moves to the enterprise investment reading as context; the other claims keep their remaining channels.

Readings member-service-automation in Q1 and Q2: were described and inferred, sized at the former estimates unh-2026-cq1-f38 and unh-2026-cq2-f37; now directional and unsized, because the only measures are counts and shares by count: over 80% of consumer contacts through digital formats and 73,000,000 digital visits in Q1, and virtually every provider and consumer interaction using AI in Q2, with no cost, rate per unit or headcount.

Figures unh-2026-cq1-f35 and unh-2026-cq1-f36 (Q1 spend on AI products and on AI across the company): was a management-bound estimate with the ledger's own phasing of the year, now a decomposition on the stated level, $1.50bn, with the fixed annual-plan band, because a plan for a year converts to a quarter only through that band. The values do not change, since the old phasing equalled the band.

Readings ai-product-investment and ai-enterprise-investment in Q1: strength was implied, now inferred, because the quarter's share of an annual amount is the ledger's, not management's.

Figures unh-2026-cq2-f34 and unh-2026-cq2-f35 (Q2 spend): was the ledger's own phasing of the year, now the fixed annual-plan band; the values do not change.

Reading pharmacy-contact-center-automation in Q1: was directional, inferred and efficiency, sized at the former estimate unh-2026-cq1-f39; now described, unsized and exploratory, because the call-volume cut of 25% is credited to digital and AI-enabled self-service together and nothing separates AI's part. The Q2 reading carries the exploratory motive.

Reading payment-integrity-ai in Q2: was inferred, sized at the former estimate unh-2026-cq2-f43; now described and unsized, because AI-enhanced fraud work is named as one of several bases of the margin recovery, beside administrative cost efficiency and medical cost affordability.

Reading optum-insight-delivery-automation in Q2: was inferred, sized at the former estimate unh-2026-cq2-f42 and left out of totals; now described and unsized, because the head of technology credits some of the segment's quarter to AI beside client volume moved into the first half.

Not changed: the operating cost reductions given on the Q4 2025 call, nearly $1.00bn for 2026, stay a ceiling on the UnitedHealthcare savings ranges, which sit under it on the annual-plan band as well.

UnitedHealth Group enters with Q1 and Q2 2026, with channels registered from management’s words: spend on AI products and platforms and on AI across the company’s own processes and functions, revenue from Optum Insight’s AI-first products and AI consulting, and savings in UnitedHealthcare member service, Optum Rx contact centers, prior authorization, claims processing, clinical and care-manager time, corporate functions, Optum Insight’s own service delivery and fraud, waste and abuse work. All but the fraud channel are tagged expanded: the FY2024 10-K already describes AI and machine learning in the company’s internal operations and customer-facing products, and Optum Insight already sold administrative, payment integrity and revenue cycle technology to health plans and providers. The fraud channel is tagged relabelled, since the anchor already ties AI and machine learning to fraud detection and nothing is shown to move. The anchor call, held in the quarter of the Change Healthcare cyberattack, carries no AI passage. The Q4 2025 call (reference quarter) is where the investment plan, nearly $1.50bn, was first given, and where UnitedHealthcare expected nearly $1.00bn of 2026 operating cost reductions, many AI-enabled, which the ledger uses as a ceiling on its UnitedHealthcare savings ranges. Subsidiary heads (UnitedHealthcare, Optum, Optum Insight, Optum Health) give most of the detail.

Q1 2026. The CEO repeated fiscal 2026 AI investment at nearly $1.50bn and the head of Optum Insight at about the same, 0.33 of it into software products and platforms and the rest across processes and functions, with a return of 2.00 expected over the next few years. The ledger caps the quarter at the stated level and takes part of the product spend as capitalized. The release and the CFO tie the operating cost ratio, 13.8% against 12.4%, partly to AI investment; the 10-Q names people, process and technology and does not say AI. The only measured movement is Optum Rx contact call center volume down 25% through digital and AI-enabled self-service, with no comparison base stated, while scripts fell −6.1%. Operating costs ran $1.53bn above their prior-year share of revenues, but incentive compensation up $865mn, restructuring adding $415mn, portfolio gains and the AI spend itself leave $167mn, so the line shows no direction for savings. Optum Insight sizes its AI products in transactions and clients, not revenue.

Q2 2026. The steps: the spend channels move from quantified to withdrawn and from implied to the ledger’s inference, because the level, given on the Q4 2025 call and in Q1, is not given in Q2; the call repeats the product share without a base, and the CFO still names AI investment as a reason for the higher-end operating cost ratio. Whether an in-year plan given once should carry forward as management’s level is a methodology question the ledger has not settled; prior authorization moves from exploratory to efficiency, because the tool built from the internal use case is credited with about 500,000 authorizations processed and 69,000 administrative hours saved year to date, though no cost line moves; Optum Rx contact centers go unmentioned; and the claims, Optum Insight delivery and fraud channels open. Management says AI now automates complex claims, the head of technology credits part of Optum Insight’s quarter to AI efficiency gains while the 10-Q credits business services growth, and the head of the commercial business names AI-enhanced fraud, waste and abuse work in a margin recovery now running past 2027. Optum Insight’s costs held flat as its revenues excluding investment income rose, −$101mn against the prior-year rate; consolidated operating costs ran $439mn above theirs, about the size of the AI spend. Rates replace costs: ambient listening available to 70% of employed providers, case summaries 40% faster for nurse care managers, and contact centers far more efficient with savings partly reinvested.

The ballparks are shares of reported lines times assumed savings, each written down. Savings are read on operating costs: claims-processing savings at $0 to $51mn, member service at $0 to $91mn, corporate functions at $0 to $57mn. Optum Insight’s delivery saving, $0 to $81mn, is left out of totals as an overlap, since the segment does part of the group’s own claims and authorization work. Revenue from AI-first products is a share of Optum Insight’s outside revenue, $7.8mn to $62mn, and excludes the group’s own use, which consolidation eliminates. The fraud channel is tagged relabelled and counts nothing in the incremental total. Provider coding intensity, which management names as a commercial cost driver, is not attributed to AI by management and is not registered as a toll. Spend is the largest money here and the only part management sized; whether the savings it is meant to produce reach the operating cost line is what later quarters may show.

United enters with Q1 and Q2 2026 and no steps: both quarters are silent on AI in every source, so there is no channel to open and no disclosure count to thin or widen. The roster's note that United had talked about AI in customer communications and operations is not borne out by the covered sources. The only AI text the ledger read is in the anchor: the FY2024 10-K's risk factors say the company depends on technology and automated systems including AI, among them reservation systems and demand prediction software, and that its use of AI applications had led to immaterial cybersecurity incidents. A risk factor names no function, money or measure, and the anchor is read for novelty, not for channels, so it opens nothing. The Q1 2024 call does not mention AI.

Q1 2026. Management explained the quarter by fuel, fare increases, capacity cuts and storms. The digital topics (self-service tools, record app use and digital check-in, live security wait times and weather maps in customer messages, and nested selling on the website, which the chief commercial officer valued in the hundreds of millions of dollars a year) were not described as AI. Salaries and related costs grew 9.8% on more flying and a 5.9% rise in headcount; non-fuel unit cost rose 5.9%.

Q2 2026. The release lists an automated customer service channel on WhatsApp for customers in three countries, the nearest passage to a support door; the company calls it automated, not AI, and it stays context. The CFO's efficiency remarks and the technology behind the new fare display are not attributed to AI. Salaries and related costs grew 6.2% on more flying, a 5.6% rise in headcount and the flight attendant agreement; other operating expenses, which hold information technology projects, grew 8.6%; non-fuel unit cost rose 6.1%.

Against Delta. Delta's Q1 was silent too; in Q2 Delta named an AI-powered assistant in its app and an AI baggage tool, and the ledger reads the assistant as an unsized saving and the baggage tool as a saving sized by its own estimate, $117k to $18mn. On the same doors (customer service, operations, revenue management, digital) United's Q2 sources describe the app, messaging and automation without the word AI. The difference is in what each company says, not a measured difference in what each does: United may use the same tools and not name them, and the quiet is the reading.

Channel ai-build-investment: was a spend channel with no capital flag, sized as an AI share of growth capital expenditure, so capitalized spend sat in the spend total; now it carries the capital flag and is left out of flow totals, because a channel that sizes capitalized AI spend is capital (methodology, Capital spending). The sizes ttec-2026-cq1-f28 and ttec-2026-cq2-f28 stay; the high-end assumption no longer stands in for the expensed part of the build, which stays unsized.

Readings engage-ai-volume-displacement and ai-pricing-passback in Q1: were inferred, sized at the former estimates ttec-2026-cq1-f30 and ttec-2026-cq1-f31 from zero to a small share of Engage revenue; now bounded and unsized, because the CEO says volumes are not being reduced by AI and that savings are not given away, and the CFO that pricing has not moved on AI, so the money has not started. Pricing pass-back also enters only through an analyst's question. It stays registered because the displaced-vendor design reads price passed back as its own toll.

Reading engage-ai-volume-displacement in Q2: was inferred, sized at the former estimate ttec-2026-cq2-f30; now described and unsized, because no Q2 source names clients' AI as removing volume: the automation named is TTEC's own offer and the decline is credited to rationalized clients, a public sector client, attrition and a completed contract.

Reading engage-ai-productivity in Q2: was inferred, sized at the former estimate ttec-2026-cq2-f31; now directional and unsized, because the quarter's present-tense statements name AI beside automation and best-shore delivery, and the one AI-alone line concerns margin to come. The step comes from the wording, not from a change at the company; Q1, which names the AI tools alone, keeps its size.

Readings digital-ai-services in Q1 and Q2: were inferred, sized at the former estimates ttec-2026-cq1-f29 and ttec-2026-cq2-f29; now described and unsized, because the professional services growth outside the legacy practices, 15.3% in Q1 and 13% in Q2, is credited to the AI, data, observability and security remix together. The other practices are added as a confound.

Reading engage-ai-productivity in Q1: state was quantified, now directional, because interview-to-hire gains and associates and clients on the coaching platform are counts and rates of AI work with no dollar level. The Q2 step from quantified to directional is gone with it.

Claims ttec-anchor-c16 and ttec-anchor-c20: speaker was CEO, now other executive, because the turn's role is President and CEO of TTEC Engage, a segment.

Claims wmt-anchor-c8, wmt-anchor-c16 and wmt-anchor-c17: speaker was CEO, now other executive, because the turn's role is CEO of Walmart U.S., a segment.

The rule is in the validator: a CEO or CFO title followed by "of" and a unit is a segment head.

Reading cortex-code-revenue in Q1: was quantified and inferred, sized at the former estimate snow-2026-cq1-f82 (a customer count dated after the quarter times an assumed revenue per customer); now described and unsized, because the 10-K places general availability after the quarter ended and the only measure is a count, so the money had not started.

Reading cortex-code-revenue in Q3: was quantified and inferred, sized at the former estimate snow-2026-cq3-f95 (the stated account count times an assumed revenue per account); now directional and unsized, because the quarter's only measures are counts of accounts. Q2 stays sized at $9.6mn to $45mn: its estimate rests on the dollar residual of the guidance beat, not on a count. The step from Q2 comes from the wording, not from a change at the company.

Reading snowflake-intelligence-revenue in Q1, Q2 and Q3: was inferred, sized at the former estimates snow-2026-cq1-f83, snow-2026-cq2-f87 and snow-2026-cq3-f96 (an account count, stated or multiplied, times an assumed revenue per account); now directional and unsized in every quarter, because accounts are a count and no revenue is given for the product. Q1 and Q3 were quantified and are now directional.

Reading other-ai-workloads-revenue: was the residual of the ledger's AI revenue estimate after both agent products (former estimates snow-2026-cq1-f84, snow-2026-cq2-f88, snow-2026-cq3-f97); now the whole estimate in Q1 (snow-2026-cq1-f79) and Q3 (snow-2026-cq3-f92), and in Q2 a new residual after Cortex Code alone, snow-2026-cq2-f100, because a metric whose parts have no evidence of their own is not split by judgment shares. The split changes as follows: in Q1 the channel goes from $11mn to $46mn to $20mn to $55mn, taking what the two products held; in Q2 from $0 to $68mn to $8.8mn to $84mn, taking Snowflake Intelligence's part; in Q3 from $1.2mn to $58mn to $60mn to $116mn, taking both products' parts.

Channel other-ai-workloads-revenue: label and description now say it holds whatever part of the AI revenue estimate no agent product is sized for in the quarter. Channels cortex-code-revenue and snowflake-intelligence-revenue: descriptions now say each is sized only in a quarter where its own evidence carries dollars.

Figure snow-2026-cq2-f85 (the Q2 estimate of all AI product revenue) still deducts the ledger's Q1 Cortex Code estimate, snow-2026-cq1-f82, from its prior-quarter base; it is kept, since that figure, with a low end of nothing, is the ledger's reading of preview usage inside the Q1 estimate, and rebuilding it would move the Q2 point by less than the Q1 estimate itself.

Kept as it was, after review: the Q3 core consumption attributed to AI, since a passage that quarter names AI alone as lifting platform consumption.

Figure snow-2026-cq1-f88 (Q1 AI compute for the company's own development): was the fiscal-year increase, $18mn, weighted toward the fourth quarter by the ledger's own share, now the fixed annual-plan band, because an amount stated for a year converts to a quarter only through that band; the amount is a completed year's actual, not a plan. It is now $1.1mn to $5.4mn, lower across its range.

Figure snow-2026-cq1-f91 (Q1 outside spend displaced by internal agents): was a quarter of the replaced system's cost, $5.0mn, read as annual, now the fixed undated band, because the CEO gives the cost with no period. It is now $104k to $2.0mn, lower at the point and the low end.

Reading core-consumption-attributed-to-ai in all three quarters: motive was unknown, now exploratory, because the filing's cause, consumption by existing customers, is generic and does not contradict management's credit to AI, which carries no measure.

Reading ai-native-customer-revenue in Q3: motive was unknown, now exploratory, because the sources are silent on why these customers spend rather than contradicting each other.

Channels ai-cost-to-serve and outside-spend-displaced-by-internal-agents: counterparty was unknown and enterprise, now mixed, because the cost to serve is paid to model providers and cloud providers, and the displaced bills went to software vendors, an agency and cloud providers.

Channel internal-ai-compute: counterparty was unknown, now cloud-provider, because the spend sits in third-party cloud infrastructure expenses inside research and development.

Channel workloads-and-budget-lost-to-ai: counterparty was ai-lab, now mixed, because the money at risk goes to model providers building their own data layers or to customers' own AI-built software and capped usage.

Not changed: the Q3 agency saving, $50k to $2.1mn, stays a quarter of an annual rate of spend eliminated, the run-rate reading, with the retired system kept in its high end only.

Sherwin-Williams enters with Q1 and Q2 2026 and with one channel, opened in Q2: coatings demand from AI data-centre construction (revenue, in the facilities layer). The anchor already shows the activity without AI: the 2024 10-K lists protective and marine products among what the Paint Stores Group sells, within segment net sales of $13.19bn, and on the first call of 2024 the CEO pointed to Protective & Marine demand strength. The channel has no measure of AI’s part, so it is read as relabelled and adds nothing to the incremental total.

Q1 2026, the quarter ended 2026-03-31. No source names AI and no channel is registered. Selling, general and administrative expense moved by $176mn on a year earlier, explained by employee-related and marketing costs, the Suvinil acquisition, the new global headquarters and technology center and currency; the CFO credits Consumer Brands margin to supply chain efficiencies and simplification. None of it is attributed to AI.

Q2 2026, the quarter ended 2026-06-30. Steps: AI disclosure widened, from a call with no AI passage to two answers by the CEO; the data-centre coatings channel opened, read described. The CEO names AI data centres built by hyperscalers as a tailwind, while the prepared remarks list data centres beside semiconductor infrastructure and onshoring as drivers of Protective & Marine growth and give no share; Paint Stores Group net sales grew by 5.1%, credited in the 10-Q primarily to selling price increases. The CEO also names leveraging AI where it makes sense among the aims for a more productive store platform, naming no tool, function or saving, so no channel is registered on it. The store platform gains are credited to closing 57 stores below the profitability threshold; restructuring actions are expected to save about $17mn a year, and the CFO names digital, ERP and CRM work, none of it attributed to AI.

No channel is sized, by management or by the ledger. The data-centre remark names AI construction as demand for the company’s own products; on the ledger’s rules a data-centre amount would only be a ceiling on the AI part, and none is given. On the door it shares with UPS, Sherwin-Williams credits its savings to pruning, restructuring and simplification and names AI only as an aim; no AI savings or spend channel is registered, because no source names an AI tool, saving or spending.

Reading merchant-ai-tools in Q1: was described and inferred, sized at the former estimate shop-2026-cq1-f32; now directional and unsized, because the only Q1 measures are counts (weekly active shops, custom apps, Flows, theme edits) and the CFO declines to tie sign-ups to AI.

Reading workforce-productivity in Q2: was inferred, sized at the former estimate shop-2026-cq2-f34 on an AI share held from Q1; now described and unsized, because the CFO credits operating leverage to headcount discipline without naming AI, credits AI only with output quality, and the 10-Q names no AI cause. The step from Q1 comes from the wording, not from a change at the company.

Figure shop-2026-cq2-f34: its rationale said the channel is relabelled and counts zero in the incremental total; the channel is tagged expanded and sized as an increment, and the text now says so.

Readings merchant-support-automation in Q1 and Q2: motive was unknown, now exploratory, because the sources do not contradict each other; support efficiencies are named with no measure and no line moving because of AI.

Channel merchant-ai-inference: counterparty was model-provider, now mixed, because the bill pays model providers for model calls and cloud providers for the AI-related usage inside cloud and infrastructure costs.

Channel internal-ai-usage: counterparty was model-provider, now mixed, because the bill pays frontier model providers, coding-agent vendors and the hardware and software behind the internal proxy and distilled models.

Channel ad-production-automation: counterparty was enterprise, now mixed, because the displaced cost is outside production companies and shoots and the company's own in-house creative time.

Channel ai-initiative-investment: counterparty was internal, now mixed, because the channel carries internal technology and business staff and the outside model and tooling bills beneath them.

Procter & Gamble enters with its fiscal third and fourth quarters of fiscal 2026 (calendar Q1 and Q2 2026). Its fiscal year ends in June, so the fourth quarter's lines are the 10-K year less the nine months in the Q3 10-Q. Every channel sits on the end-use layer and is tagged relabelled against the anchor, which already named Supply Chain 3.0 and digital acumen as focus areas and described sharper media targeting; its risk factors also mentioned technologies enabled by machine learning or artificial intelligence in general terms. No covered source shows a line, rate or volume moving because of AI alone, so nothing enters the incremental total.

Q3 FY2026, the quarter ended 2026-03-31. An analyst described Supply Chain 3.0 as the company's way of deploying AI; the CFO answered that he would not call it AI, that some of it is AI and a lot of it is more basic automation (unattended shifts and warehousing, touchless quality), all inside the productivity commitments. That opens manufacturing and supply chain work automated in part with AI, read described and unsized under the rule for AI named beside another cause, with cost of products sold of $10.61bn net of incremental restructuring charges as its ceiling. The 10-Q credits manufacturing productivity savings of 2.1% of gross margin without naming AI, and the CFO's toolboxes for content creation and molecular discovery are not called AI in this quarter.

Q4 FY2026, the quarter ended 2026-06-30. Steps: channels opened for brand-building tools, internal work and molecular discovery, and the supply chain channel moved from described to not mentioned. In prepared remarks the CEO named AI-enabled tools scaled with integrated workflows from creative development to media activation, as part of a brand-building transformation credited as a whole with trial, awareness, loyalty and growth; integrated data platforms, AI capabilities and programmatic shelf tools that move discovery-to-execution time from weeks to hours in many cases; and AI-enabled molecular discovery as one of the new technologies that, beside the company's existing research capabilities, will drive faster innovation. AI is named beside other causes in each, with no measure of its part, so none is sized and none gets a ballpark; the internal work channel reads directional for its weeks-to-hours rate, the others described. The CFO credits fiscal 2026 with $2.80bn of productivity improvement, the restructuring plan cuts up to 7,000 non-manufacturing overhead roles, and the 10-K credits a change in employees of −4% to that program; none of it is attributed to AI.

Beside PepsiCo, the reading is similar in kind and wider in scope. Both companies name AI in the supply chain only beside automation and other causes, and both credit their measured savings to productivity plans and headcount and plant actions. Procter & Gamble also names AI in marketing content, internal work and research, where PepsiCo names it nowhere, and its CFO explicitly declined the analyst's framing of the supply chain program as AI. Shopping agents and AI-based search are named as changes in how consumers shop, with no spend or lost sales, and are context, not a toll. No AI vendor, AI spend or AI-attributed revenue figure appears in any source. Both calls carry prepared remarks and Q&A in one transcript.

PepsiCo enters with Q1 and Q2 2026, its first and second fiscal quarters; North America reports weekly fiscal periods shorter than a calendar quarter, and international operations report calendar months (January and February in Q1; March, April and May in Q2). It enters with a single channel: supply chain, transportation and route decisions made with AI (savings, operations). The anchor already shows the activity under the AI name: the 2024 10-K says the company was optimizing its supply chain through advanced technologies like artificial intelligence, and credits automation with better optimization across transportation and fleet networks; shipping and handling expenses inside selling, general and administrative expenses were $16.00bn for 2024. With no line, rate or volume shown to move because of AI, the channel is relabelled and adds nothing to the incremental total.

Q1 2026, the period ended 2026-03-21. In an answer on productivity, the CEO named AI in the supply chain and in how the company does transportation and optimizes routes, together with global shared services, technology deployed across the company, digital ordering and data, and gave no measure. The 10-Q credits operating profit growth of 24% first to productivity savings from the 2019 productivity plan, which names new technology and automation, not AI; the CFO credits supply chain productivity to headcount reduction, plant closures and fewer SKUs. The channel reads described; under the rule for AI named beside other causes it gets no ballpark, and the ceiling is cost of sales plus selling, general and administrative expenses of $16.10bn net of the plan's charges.

Q2 2026, the period ended 2026-06-13. Steps: AI disclosure thinned, from the AI passage on the Q1 call to none in any Q2 source; the supply chain channel moved from described to not mentioned. The CEO describes the U.S. cost program as expanding automation and digitalization, combined mixing centers and tests of combined delivery and fleet, without naming AI, and the 10-Q credits operating profit growth of 125% to prior-year impairments, productivity savings, pricing and lower restructuring charges.

No channel is sized, by management or by the ledger. On the doors Walmart and Costco open (supply chain, marketing, sales and workforce), PepsiCo names AI only in that Q1 answer and only beside other causes; digital ordering for salesmen and the optimization of advertising, marketing and trade budgets are described without AI and are context, not channels. No AI spend, vendor, AI-platform demand or toll is named. The call transcripts are the question and answer sessions only; the prepared remarks, published separately on the company's investor site, are stored and read, and they do not mention AI.

Readings embedded-ai-in-applications in Q1, Q2 and Q3: were inferred, sized at the former estimates orcl-2026-cq1-f80, orcl-2026-cq2-f97 and orcl-2026-cq3-f88 (an assumed share of cloud applications revenue); now directional and unsized, because the only measures are counts of agents, uses and tokens and the features carry no price, so the products-with-no-price rule applies as well. Cloud applications revenue is quoted as the ceiling.

Readings priced-agentic-capacity in Q2 and Q3: were inferred, sized at the former estimates orcl-2026-cq2-f99 (customers pre-purchasing tokens times an assumed purchase) and orcl-2026-cq3-f90 (that count times an assumed multiple); now unsized, Q2 directional on the customer count and Q3 described, because a customer count with no price or revenue sizes nothing.

Readings ai-halo-cloud-demand in Q1, Q2 and Q3: were inferred, sized at the former estimates orcl-2026-cq1-f79, orcl-2026-cq2-f96 and orcl-2026-cq3-f87 (a judgment halo share of cloud infrastructure revenue outside AI capacity); now described and unsized, because management names the pull together with the applications, sovereignty and partner clouds beside AI, and in Q2 does not name it at all. Cloud infrastructure revenue outside AI capacity is quoted as the ceiling.

Readings embedded-ai-inference-cost in Q1 and Q2: were inferred, sized at the former estimates orcl-2026-cq1-f87 and orcl-2026-cq2-f105 (a token volume borrowed from the Q3 disclosure and scaled back, times a price class); now described and unsized, because no usage is stated for those quarters and only a measured volume with a period times a public price sizes a spend channel. Q3, with tokens consumed during the quarter, keeps its size.

Readings saas-displacement-by-ai in Q1 and Q3: were inferred, sized at the former estimates orcl-2026-cq1-f91 and orcl-2026-cq3-f101; now described and unsized, because management denies the effect in both quarters, so the money has not started. Q2, where management concedes delayed decisions, keeps its size.

Readings ai-equipment-depreciation, ai-data-center-lease-cost, ai-capacity-power-and-operating-cost and ai-buildout-financing-cost: unchanged, because the build's own costs may carry an AI share where the company bounds AI's part of the build, and Oracle names the capital program for AI alone (AI infrastructure is capital-intensive; the program supports AI cloud infrastructure; general-purpose cloud needs some capital but not as much as the AI clusters), with no cloud-and-AI pairing of the kind that unsized Microsoft's.

Reading database-demand-attributed-to-ai in Q1, Q2 and Q3: was inferred, sized at the former estimates orcl-2026-cq1-f78, orcl-2026-cq2-f95 and orcl-2026-cq3-f86; now described and unsized, because management credits database growth to AI adoption and to regions opening in partner clouds, and nothing separates AI's part.

Reading ai-attributed-restructuring-cost in Q1, Q2 and Q3: was inferred, sized at the former estimates orcl-2026-cq1-f88, orcl-2026-cq2-f106 and orcl-2026-cq3-f96; now described and unsized, because the expense is the cost of a plan whose purposes include acquisitions, strategic measures and other operational activities beside AI, and no AI share is given.

Reading engineering-ai-code-generation in Q2 and Q3: was inferred, sized at the former estimates orcl-2026-cq2-f107 and orcl-2026-cq3-f98; now described and unsized, because the fall in research and development comes under that plan, and the filing explains the line by employee-related and computer equipment expenses. Q1 stays sized at $0, because the release names AI code generation alone as the reason for smaller product teams.

Reading ai-adoption-in-other-functions in Q2 and Q3: was inferred, sized at the former estimates orcl-2026-cq2-f109 and orcl-2026-cq3-f100; now described and unsized, because management credits the lower costs to efficiency actions and a simpler go-to-market model and the filing names AI only among the plan's purposes.

Reading embedded-ai-in-applications in Q1, Q2 and Q3: state was quantified, now directional, because agents delivered, uses and tokens are counts of AI work with no dollar level.

Reading embedded-ai-inference-cost in Q3: state was quantified, now directional, because the tokens consumed, 900,000,000,000, are a count of AI work with no cost.

Channel embedded-ai-inference-cost: counterparty was unknown, now mixed: outside model providers and the company's own cloud capacity.

Channel ai-infrastructure-revenue: counterparty was ai-lab, now mixed: AI labs and other advanced AI customers, from start-ups to investment-grade companies, which management does not split.

Channel ai-capacity-power-and-operating-cost: domain was cost-of-revenue, now operations, because power, network and the other running costs of the data center fleet are now an operations domain.

Motive of ai-adoption-in-other-functions, ai-attributed-restructuring-cost, engineering-ai-code-generation (2026-CQ1, 2026-CQ2, 2026-CQ3): was efficiency, now exploratory, because the shrinking line is credited to AI together with other causes and nothing separates AI's part, so the result does not meet the efficiency tell for any one of them.

Channels ai-infrastructure-revenue, customer-funded-ai-contracts, ai-adjacent-cloud-services, ai-data-center-capital-expenditure, ai-equipment-depreciation, ai-data-center-lease-cost, ai-capacity-power-and-operating-cost and ai-buildout-financing-cost: layer compute. Every other channel is end-use.

Lab funding is unchanged: the sources show no Oracle investment in its lab customers, so the AI infrastructure readings keep their funding.

Readings ai-coding-tools-bill in Q1, Q2 and Q3: were inferred, sized at the former estimates nvda-2026-cq1-f53, nvda-2026-cq2-f57 and nvda-2026-cq3-f54; now unsized (Q1 directional on the share of coders using the tools, Q2 and Q3 directional on the CFO's named acceleration in AI tool use), because the size rested on the research and development headcount, 31,000, a count at a date, times judgment shares and prices. A measured usage volume with a period times a public price may size a spend channel; a count of employees may not.

Nvidia enters with Q4 fiscal 2026 (calendar 2026-CQ1) and Q1 and Q2 fiscal 2027 and these channels. Revenue: data center AI infrastructure bought by the hyperscale clouds, and by AI clouds, model makers, sovereigns and enterprises (the company’s ACIE sub-market); inside them, as overlaps that are not added, sovereign AI, compute bought by and for the frontier labs, and physical AI; and AI workstations at the edge. Spend: cloud capacity rented back from cloud providers for research, capacity commitments to AI clouds (from Q2 fiscal 2027), outside AI coding tools, amortization of the Groq inference licence, and the cost of the AI systems sold, paid on to foundries, memory makers and system builders. Savings: engineering productivity from AI tools (from Q1 fiscal 2027). Non-operating: gains on equity stakes. The labs and the coding bill are new money by the ledger’s test; the Data Center channels, research cloud capacity, the Groq licence and the cost of revenue are expanded; physical AI and workstations are relabelled groupings of lines the anchor already sold.

Q4 fiscal 2026. Data Center revenue was $62.31bn of $68.13bn. The CFO commentary put hyperscalers at slightly over 50% of it; the next release restates the quarter at $33.81bn Hyperscale and $28.50bn ACIE, and the ledger reads the AI part from those levels (hyperscale $29.83bn to $32.37bn, the rest $20.54bn to $27.29bn), each less a non-AI part capped from the fiscal 2022 Data Center level of $2.65bn a quarter. The 10-K estimates that one AI research and deployment company contributed a meaningful amount of revenue by purchasing cloud services from the company’s customers; the call announced a $10.00bn investment in Anthropic. Sovereign AI was over $30.00bn for the year and physical AI north of $6.00bn, converted to quarters with the annual band. Cloud service agreements, capacity the company rents back from its own buyers for research, stood at $27.00bn.

Q1 fiscal 2027, the steps. The company began reporting Hyperscale ($37.87bn) and ACIE ($37.38bn) revenue, so both Data Center channels move from bounded to quantified; the ledger already read Q4 fiscal 2026 from the levels this release restates, so the strength stays inferred. The restated Q4 hyperscale share is 54.3%. Sovereign revenue reads directional, with growth of more than 80% and no level. The CFO named an acceleration in AI tool use, to enhance productivity, among the reasons operating expenses grow: the coding-tools bill moves to directional and an engineering productivity channel opens, unsized by management. The 10-Q says the quarter’s $18.60bn of investments in private companies and infrastructure funds include AI model makers that may purchase or use the company’s products in the cloud.

Q2 fiscal 2027, the steps. Hyperscale funding stays mixed, as at the buyers, with extended payment terms for large multi-quarter purchases by certain investment-grade customers as the company’s own part (receivables at 60 days). The frontier lab channel stays mixed under the rule for labs: within the quarter the company’s part is its equity stakes, nearly $50.00bn invested in the labs by the call date; the credit support capped at $105.00bn for leases to an OpenAI affiliate was entered in August 2026, after the quarter end, and takes effect from fiscal 2029; the CFO says the company will provide credit enhancement for another lab, and expects labs backed this way to contribute a share of next year’s business, a forward statement. The CFO also says independent capital still underwrites every deal. A capacity commitment channel opens: AI clouds procure the company’s systems and it commits $36.00bn to purchase their capacity, a take-or-pay floor that lets lenders finance them, with no payment until fiscal 2028. Physical AI reads withdrawn: the level given in each prior covered quarter is absent. A company was moved from ACIE to Hyperscale, recasting prior periods by $5.18bn in Q1 fiscal 2027.

The ballparks and the layers. The widest ranges are the frontier labs ($7.12bn to $22.26bn, an assumed share of Data Center revenue), workstations ($259mn to $1.61bn), the coding bill ($1.3mn to $93mn) and engineering productivity ($8.5mn to $317mn). The cost of the AI systems sold, $17.24bn to $21.61bn, is the next layer of the same money: hyperscalers’ capital spending (traced at MSFT and ORCL, out of their totals) becomes this revenue, and part of it is paid on to suppliers. By layer, the two Data Center channels (the largest), the named payers inside them, physical AI, workstations and the cost of the AI systems sold are hardware; the capacity commitments to AI clouds are compute; research cloud capacity, the coding bill, the Groq licence, engineering productivity and the equity marks are end-use, the company’s own use of AI. The ledger does not net the layers, and cohort totals are kept per layer, since a total across them would count the same dollar more than once. Equity gains of $7.80bn in the quarter are non-operating and out of totals.

Reading ai-control-tower in Q1 and Q2: was inferred, sized at the former estimates now-2026-cq1-f76 and now-2026-cq2-f77 (a customer count times an assumed contract value, the Q1 count borrowed from the Q2 call); now directional and unsized, because the only measures are a deal-size multiple and a count of customers live, and a count sizes nothing. Q2's state was quantified and is now directional. The channel overlaps servicenow-ai-contract-value, so no flow total moves.

Reading employeeworks-conversational-ai in Q1 and Q2: was inferred, sized at the former estimates now-2026-cq1-f77 and now-2026-cq2-f78 (an assumed recurring revenue level, its base from an announcement outside the stored sources); now directional and unsized, because the only measures are counts of large deals and the growth of deal volume. The channel overlaps servicenow-ai-contract-value, so no flow total moves.

Reading internal-service-desk-automation in Q1 and Q2: was inferred, sized at the former estimates now-2026-cq1-f84 and now-2026-cq2-f85 (assumed support shares of cost of subscription revenues times the share of cases resolved, carried into Q2); now directional and unsized in both quarters, because a share of cases by count and, in Q2, a count of AI specialists deployed separate nothing in dollars. Q2's state was described and is now directional on that count. The channel overlaps now-on-now-productivity, so no flow total moves.

Assessments of Q1 and Q2: the sentences that read AI Control Tower among the widest ranges and as quantified are rewritten; the steps from Q1 no longer list the service desk saving moving to described.

Kept as they were, after review: the Q2 inference cost, sized as a cost-to-serve ratio on the AI revenue estimate rather than as a share of the margin movement management names jointly; and the Q2 Now on Now productivity size, which carries the Q1 figure through the undated band.

Figures now-2026-cq1-f83 and now-2026-cq2-f84 (cost avoided through the company's own AI): was the stated figure spread over the ledger's own reading of one year, at a low end of two years, now the fixed undated band, because an amount with no period converts only through that band; the point is now $10mn to $125mn in each quarter, with a lower low end and the same high end.

Reading internal-service-desk-automation in Q1: state was quantified, now directional, because shares of the company's own cases resolved by agents, 90% and 90%, are measures of AI work with no dollar level and are applied by the company to no reported line.

Reading core-workflow-demand-from-ai in Q1 and Q2: motive was unknown, now exploratory, because AI is named as a tailwind for the core products with no price, attach rate or measure, and the sources do not contradict each other.

Channel ai-adoption-services-investment: counterparty was unknown, now mixed: the company's own forward deployed engineers and customer excellence staff, and contracted services partners.

Channel ai-inference-and-cloud-cost: counterparty was unknown, now mixed: model providers, and the hyperscalers and data centers that supply capacity.

Steps. Between Q1 and Q2 the Rewire products for the utility industry went from described to not mentioned. No other channel changed state, strength or motive.

NextEra Energy enters with Q1 and Q2 2026. On the facilities layer, its channels are power and generation sold to hyperscalers and other data-centre loads (Energy Resources contracts, the data-centre hubs, the U.S.-Japan projects, the Duane Arnold recommissioning with Google and large load at FPL), and the capital spending that builds it, capital and out of totals. Both are expanded by the ledger’s test: at the anchor Energy Resources already had 3.5 GW in operation for technology customers, and in coverage new capacity is contracted for hyperscalers, so volume moved; with no traced quarter before AI they are undetermined in the incremental total. On end use: the Rewire AI tools in the company’s own operations, and Rewire products delivered to the utility industry with Google, both read as relabelled, because the anchor call already describes dozens of proprietary artificial intelligence tools running the fleets and serving customers, the sources never say the new tools are LLMs, and no line moves.

Q1 2026. The CEO launched Rewire, a company-wide initiative to reimagine how the company works paired with an enterprise-wide AI transformation, in partnership with Google Cloud, and named Conduit, Generation Entitlement and Grid Composer; he said the tools have the potential to drive significant savings, without a number. Roughly 30% of the quarter’s backlog additions were driven by hyperscalers; FPL had 12 GW of large load in advanced discussions and no signed customer, at roughly $2.00bn of capital per gigawatt; Energy Resources was selected for 9.5 GW of gas-fired generation for large loads under the U.S.-Japan framework. Capital expenditures were $11.06bn.

Q2 2026. The call did not mention Rewire; the CEO named artificial intelligence only in a list of practices the pending merger with Dominion Energy would share. FPL raised its large-load expectation for 2032 from 6 GW to 8 GW, still with no customer signed, and Energy Resources added 3.6 GW to its backlog with no hyperscaler share given. Capital expenditures were $8.33bn in the quarter against $5.68bn a year earlier. Other operations and maintenance rose on growth across the NEER businesses; no operating line is attributed to AI.

Sizes. No channel carries a size. The facilities channels stay unsized because no covered source names AI as a cause of data-centre demand, the anchor names it only beside other causes, no data-centre amount is stated, and much of the business (FPL large load, Duane Arnold, the federal projects) has not started; the hyperscaler share of backlog additions and the gigawatt counts measure the channel’s volume, not AI’s part, and separate nothing in dollars. The Rewire operations channel stays unsized because the savings are expected from Rewire as a whole, which pairs reimagining the work with AI, and only as a potential; the products channel because no price or customer is given.

What the roster expected and what the sources show. The roster note expected data-centre and large-load demand, contracts with hyperscalers and a large capital plan. The sources show the demand and the plan, and a hyperscaler collaboration by name (Google, on Duane Arnold), but no contract with a hyperscaler whose money, term or start is stated, and no covered statement that ties the demand to AI; the 10-Qs never mention data centres. Beside Duke Energy and Constellation, read at the same layer, NextEra Energy differs in where its AI cause for data-centre demand rests, on the anchor alone, where Duke Energy’s rests on a covered call too and Constellation’s CEO named AI in Q2; it is also the only utility here with a named internal AI program, Rewire, and a product line built on it. None of the utilities earns anything measurable from AI in these quarters. Hyperscalers on this ledger would carry the same money as data-centre cost; totals are not netted across companies, and cohort totals keep the facilities layer apart.

Readings ai-infrastructure-depreciation, datacenter-lease-cost and finance-lease-interest in Q1 and Q2: were inferred, sized at the former estimates msft-2026-cq1-f72, msft-2026-cq1-f79, msft-2026-cq1-f80, msft-2026-cq2-f79, msft-2026-cq2-f92 and msft-2026-cq2-f93 (each the line above its fiscal 2024 quarterly average times a judgment AI share); now described and unsized, because the build's own costs carry an AI share only where the company bounds AI's part of the build, and Microsoft's only words are capital for growth in the cloud offerings and AI infrastructure and training, cloud and AI infrastructure, and short-lived assets for both AI and non-AI infrastructure. Depreciation, lease cost and lease interest are quoted in metrics as the ceiling lines. Lease interest moved the spend total; depreciation and lease cost overlap the reported cost-of-revenue increases and did not.

Reading ai-capital-expenditure in Q1 and Q2: was inferred, sized at the former estimates msft-2026-cq1-f66 and msft-2026-cq2-f73 (a judgment AI share of capital expenditures including finance leases); now unsized, with total capital expenditures as the ceiling, because capital named together with AI falls under the joint-cause rule (wave C). It is capital, so no flow total moves.

Readings foundry-model-and-agent-services in Q1 and Q2: were inferred, sized at msft-2026-cq1-f35 (customers at a trillion-token pace times assumed tokens and price) and msft-2026-cq2-f43 (that estimate rolled forward); now directional and unsized, because a count of customers is not a measured usage volume with a period, and Q2's revenue growth in words has no base once the Q1 estimate is not a size.

Readings security-copilot in Q1 and Q2: were inferred, sized at msft-2026-cq1-f46 and msft-2026-cq2-f48 (an assumed share of the line that holds the security suites); now unsized. Q1 is directional, not quantified, because a customer-count multiple and an alert count are counts; Q2 gives no measure for the launched product, only a system in private preview.

Readings of agent governance (the Agent365 channel) in Q1 and Q2: were inferred, sized at msft-2026-cq1-f47 and msft-2026-cq2-f50 (companies times a paying share times a price); now unsized. Q1 is described, because the product appears to have been in preview, and Q2 is directional, because agents registered and companies are counts.

Reading rd-ai-compute-and-talent in Q2: was inferred, sized at the former estimate msft-2026-cq2-f106 (a judgment share of the rise); now described and unsized, because the annual report names compute, AI talent and data together with XBOX impairment. Q1, where the filing names only compute, AI talent and data, stays reported. The step comes from the wording, not from a change at the company.

Reading copilot-advertising in Q2: was directional and inferred, sized at the former estimate msft-2026-cq2-f108; now described and unsized, because the annual report names commercial sales beside Copilot advertising and the advertising expense line covers all advertising. Q1 keeps its size.

Reading forward-deployed-engineers in Q2: was inferred, sized at the former estimate msft-2026-cq2-f115 (a project count over the past year times judgment staff and cost); now described and unsized, because the organization launched after the quarter and a project count is a count.

Readings github-copilot, m365-copilot-seats, copilot-consumption-credits and ai-pull-through-cloud-demand: unchanged. Paid seats of a priced product are units sold, credit consumption is billed money, and each pull-through quarter gives a database revenue growth rate the CEO ties to AI workloads, a rate of the AI part's own under the upstream joint-cause ruling (left for the orchestrator to confirm).

Reading dax-copilot-clinical-documentation in Q2: state was quantified, now directional, because encounters automated, 28,000,000, are a count of AI work with no dollar level.

Reading foundry-model-and-agent-services in Q1: state was quantified, now directional, because customers on track to process a trillion tokens, 300, is a token volume, a count of AI work. Q2 stays quantified on the stated customer level, 100,000, so the channel now moves from directional to quantified in Q2.

Reading openai-revenue-share-paid in Q1 and Q2: motive was unknown, now exploratory, because the sources are silent on the payment's motive rather than contradicting each other.

Readings openai-equity-method in Q1 and Q2 and anthropic-investment-gain in Q2: motive was unknown, now exploratory, because a non-cash mark meets none of the operating tells and nothing contradicts; both stay non-operating and out of totals.

Reading ai-pull-through-cloud-demand in Q1 and Q2: motive was narrative-defensive, now exploratory, because AI is credited for growth in existing products with no measure of its part and no narrative-defensive tell of its own is quoted.

Channel ai-business-run-rate: counterparty was unknown, now mixed: enterprises buying the Copilots and model services, and AI labs renting accelerator capacity.

Not changed: the lease and interest baselines, a quarter of the anchor's fiscal-year lines, are reported lines rather than amounts stated for another period, so no band applies.

Channels frontier-lab-cloud-revenue, foundry-model-and-agent-services, ai-capital-expenditure, ai-infrastructure-depreciation, datacenter-lease-cost, finance-lease-interest, intelligent-cloud-ai-infrastructure-cost and openai-revenue-share-paid: layer compute. Every other channel is end-use.

Reading frontier-lab-cloud-revenue in Q1 and Q2: funding was investor-capital, now mixed. The labs pay from outside funding rounds and from money Microsoft put in: it holds about 27% percent of OpenAI and had funded $11.80bn of its commitments by Q1 (claim c29), and it is an investor in Anthropic (claim c42).

Morgan Stanley enters with Q1 and Q2 2026 and these channels: the model access and AI tool bill, internal investment in AI and agentic infrastructure, operations and surveillance work displaced by AI, financial advisor effectiveness from AI co-piloting, client questions answered by an AI agent on the equities trading platform, cyber defense against AI-enabled attacks, and investment banking and markets activity from AI themes and the AI buildout. The model bill is new money by the ledger’s test. The build investment, the operations and equities-agent channels and advisor co-piloting are existing work into which LLM tools were put, read as expanded. Cyber defense and the capital markets activity are existing activities the anchor already shows, read as relabelled because no movement is attributed to AI. The anchor itself names AI only in risk factors.

Q1 2026. Asked whether AI is a risk to Wealth Management, the CEO called AI the firm’s friend. Asked by an analyst about Anthropic and its Mythos model, he said the firm is permissioned on Claude Mythos Preview and working with the beta version (Anthropic is private and not on the ledger). He said efficiency around classic operational flow and surveilling is under way and that technical client questions can be answered by a client agent in the electronic trading platform; earlier replacements of call-center or operational work he called not a new phenomenon, and the ledger does not credit them to AI. He said Wealth Management leadership is spending a lot of time on co-piloting for advisors, with agents that are going to drive efficiency and effectiveness; the sources do not show those tools live, so the channel is described and unsized. The CFO said, in his Wealth Management remarks, that the firm is investing in its agentic infrastructure, and that AI themes, beside geopolitical uncertainty and market dispersion, contributed to client engagement. A March workforce action cut about 2% of the workforce at a severance cost of $178mn; the 10-Q ties it to operational efficiency and performance, not AI, so it is a confound.

Q2 2026, the steps. AI disclosure thinned on the call, and the model bill, advisor co-piloting, the equities client agent and the cyber toll went unmentioned. The CFO said higher technology-driven spend relates to investments to support infrastructure, AI-enabled efficiencies and business growth; because AI is named beside infrastructure and business growth, build investment moves from a reference-class estimate to described and unsized. The step reflects the joint-cause rule applied to Q2’s wording, not a change at the firm; Q1’s statement named AI alone, and whether the rule follows the quarter’s evidence or the channel awaits a ruling. Its ceiling is the rise in information processing and communications, $114mn. The 10-Q named the adoption of AI beside investor sentiment as support for active capital markets, and the CEO put the 2026 data center spending forecast at about $850.00bn while declining to give a percentage for the firm’s role in raising it; Institutional Securities net revenues rose 44.4%, and nothing separates the AI part.

The ballparks. In Q1 the build investment, $14mn to $81mn, is a share of revenue taken from other buyers’ readings, capped at the Booking reading; the operations saving, $2.7mn to $108mn in Q1 and $2.6mn to $106mn in Q2, and the equities agent, $326k to $24mn, are decompositions whose every multiplier is a judgment. The model bill is unsized because no charge or paid use is shown, and the cyber toll because AI is named beside a broad rise in cyber risk and the step-up is stated as intent. The roster’s note that the firm has deployed AI assistants for advisors with a model provider is not in any covered source and is not used; the only provider named on the calls is Anthropic, by an analyst.

Merck enters with Q1 and Q2 2026 and one channel, opened in Q1: what it pays Google Cloud under the multi-year partnership the CEO announced in the week before the Q1 call (spend, end-use layer). The anchor already shows the activity: the FY2024 10-K describes a growing use of AI systems to automate processes, analyze data and support decision-making, and a dependence on cloud service providers, with no amount, so the channel is tagged expanded and sized as a level with no traced baseline. No IT line is reported; selling, general and administrative expenses were $10.82bn for FY2024.

Q1 2026, the quarter ended 2026-03-31. The CEO announces the Google Cloud partnership, an expanded collaboration with Tempus AI for precision oncology and an agreement with Mayo Clinic for clinical and genomic data, and says the efforts together support improved productivity and a faster pipeline, with no measure. The release lists the Google Cloud partnership and the Mayo Clinic collaboration for AI-enabled drug discovery among recent news releases. Only the Google Cloud partnership is registered, because only it names a vendor bill: a counterparty Merck pays, a scope and a multi-year duration; it is read described and unsized, since its money had not started in the quarter. Research and development expense of $12.59bn is explained by the $9.00bn Cidara charge and clinical spending, and the 2025 Restructuring Program, expected to save $1.70bn a year, eliminates positions in sales, administration and research and development; neither is attributed to AI.

Q2 2026, the quarter ended 2026-06-30. Steps: AI disclosure thinned, from a call and a release that named AI to no AI passage in any source; the Google Cloud channel moved from described to not-mentioned. Selling, general and administrative expenses rose by $255mn, and the 10-Q names investments in IT among the higher administrative costs without saying AI, so nothing is credited to the channel. Research and development expense of $9.74bn is explained primarily by business development charges, including $5.70bn for Terns.

No channel is sized, by management or by the ledger. On the door it shares with UnitedHealth Group, Merck gives no AI investment amount, no AI saving and no AI-attributed cost movement; its savings come from a restructuring and optimization program it does not tie to AI. On the door no other ledger company shares, AI in drug research and development, Merck names collaborations and an aim and no result.

Readings ai-technical-talent in Q1 and Q2: were directional and inferred, sized at the former estimates meta-2026-cq1-f32 and meta-2026-cq2-f43 on a judgment share of the compensation rise; now directional and unsized, because compensation growth is credited to technical hires, particularly AI talent, with no rate or share for AI's part. The rise in compensation is quoted as the ceiling.

Readings buildout-debt-interest in Q1 and Q2: were inferred, sized at the former estimates meta-2026-cq1-f43 and meta-2026-cq2-f53 on judgment shares of interest above the anchor level; now unsized, because Q1 names infrastructure and AI initiatives beside other uses of cash flow and financing and ties no debt to AI, and Q2 ties long-duration capital to long-horizon initiatives, especially AI infrastructure projects, which names AI first and bounds nothing. Interest on the build carries an AI share only where the company bounds AI's part. Interest expense is quoted as the ceiling.

Readings ai-capital-expenditure in Q1 and Q2: were inferred, sized at the former estimates meta-2026-cq1-f15 and meta-2026-cq2-f21 on a judgment share of the level; now described and unsized, because the filing gives the spending as support for AI efforts and the core business together; its statement that infrastructure investment increased in connection with AI initiatives speaks to the increase, not the level. It is capital, so no flow total moves.

Reading genai-ad-creative-tools in Q2: was directional and inferred, sized at the former estimate meta-2026-cq2-f70; now directional and unsized, because the quarter gives only adoption counts, 9,000,000 small businesses using a tool and image generation adoption more than doubling, with no measured lift. Q1, with a tested conversion rate, stays sized.

Meta enters with Q1 and Q2 2026, its channels in groups. The build: capital expenditures (capital, out of totals), depreciation, third-party cloud capacity, AI token costs, compensation for AI talent, interest on the debt that now helps fund it, and off-balance-sheet data center ventures (capital). The ads system: revenue from AI ad ranking, from engagement gained through AI recommendations, and from generative creative tools (the one sized ads channel). Products and internal use: business agents, Meta AI, Meta One, the model API, AI glasses, the advertiser support assistant and engineering productivity. Token bills, the products built on LLMs and the model API are new money by the ledger’s test; the build channels and the creative tools are existing activities into which LLMs were put, read as expanded. Ad ranking and recommendations are relabelled: the 2024 anchor call already credited ranking architectures with measured lifts of the same kind. The glasses were sold as AI glasses at the anchor and are relabelled too. Every channel is read at the end-use layer, the build channels (the largest) included, since the filings put the build first to Meta’s own products and model training; the model API alone is read at the compute layer, as model access sold to builders at a price.

The steps, all in Q2. Capital spending’s funding moved from operating cash flow to mixed: in Q1 capital expenditures of $19.84bn sat well inside operating cash flow of $32.23bn, while in Q2 $31.08bn took nearly all of $31.86bn, free cash flow was $784mn, the company issued $25.00bn of notes and the CFO said outside capital supplements cash flow. Third-party AI token costs went from a CEO remark about prototyping on other companies’ APIs to a named primary cause of the marketing and sales rise of $452mn. Interest went from described to directional as the CFO tied long-duration debt to AI infrastructure. Meta One and the model API opened, both launched at about the time of the call with no revenue. The advertiser support assistant, credited in Q1 with resolving account issues at a 20% higher rate, went unmentioned.

The roster note said the buildout is paid from advertising cash flow. The sources support it for Q1 and not for Q2, and not for the ventures in either quarter: in the Louisiana campus, which the 10-Q ties to AI capacity needs, the parties fund pro rata shares and the company holds a minority interest, with a lease of $12.31bn from 2029 and residual value guarantees with a threshold of about $28.00bn; an El Paso campus with BlackRock was agreed in July, after Q2, with residual value exposure of about $13.00bn expected at closing. Leases not yet commenced reached $278.99bn and commitments, mostly cloud capacity and servers, $349.31bn. Meta names no cloud provider; CoreWeave (CRWV), on this ledger, names Meta as a significant customer in its own Q1 2026 10-Q, and the CEO names NVIDIA systems among the servers. Those dollars are recorded on the suppliers’ own ledgers, not netted here.

The ballparks are the ledger’s. Capital spending on the AI build is $15.54bn to $27.97bn for Q2, an assumed share since the same servers run the ads models, which Meta itself calls AI. Depreciation is $1.31bn to $2.07bn, third-party cloud $492mn to $3.98bn (a level, net of the token estimate), AI talent $1.20bn to $3.00bn, interest $302mn to $574mn, and tokens $158mn to $1.32bn after an assumed part of the severance is taken out, against $35mn to $531mn in Q1. The ads lifts are management’s and are kept as rates: a 15.7% uplift in conversions and 8.3% more clicks on Facebook in Q2, more than 6% on landing page view conversions in Q1. Advantage+, at over $75.00bn a year, is recorded as a metric, not a size. Generative creative tools, $30mn to $623mn, are the one ads ballpark, and they count in the revenue total, since the ranking channel they share a path with is unsized. Engineering productivity is read small, $29mn to $897mn, bounded by the output-per-engineer rate the CFO gave on the reference call and by management’s statement that it builds more with the gains; the May reduction of about 8,000 employees with severance of $1.18bn is not attributed to AI in any source. AI glasses, $129mn to $302mn, are relabelled and count nothing in the incremental total.

Reading atlas-vector-search in Q1, Q2 and Q3: was directional and inferred, sized at the former estimates mdb-2026-cq1-f54, mdb-2026-cq2-f45 and mdb-2026-cq3-f42 (an assumed share of Atlas-related revenue); now unsized, because the only measures are a customer-count multiple and a share of large customers by count, credited to vector and full-text search together. The channel overlaps the cohort channels, so no flow total moves.

Reading voyage-ai-models in Q1, Q2 and Q3: was inferred, sized at the former estimates mdb-2026-cq1-f55, mdb-2026-cq2-f46 and mdb-2026-cq3-f43 (one quarter of an assumed annual level); now unsized, because the measures are customer-count multiples and a share of new customers by count, and the 10-K's not material, in Q1, is a phrase the table does not convert. Q1 stays bounded in state on that phrase. The channel overlaps ai-native-cohort-revenue, so no flow total moves.

Reading coding-agent-referrals-and-integrations in Q1: was described and inferred, sized at the former estimate mdb-2026-cq1-f57 (a multiple of the Voyage estimate, set with hindsight); now described and unsized, because agent APIs and integrations were a roadmap for the coming year, so the money had not started.

Reading coding-agent-referrals-and-integrations in Q2 and Q3: was directional and inferred, sized at the former estimates mdb-2026-cq2-f48 and mdb-2026-cq3-f46 (multiples of the Voyage estimate, itself now unsized); now unsized, because the only measures are growth in MCP usage, a share of referral traffic and a rising count of MCP clusters. The channel overlaps the Voyage and cohort channels, so no flow total moves.

Reading ai-native-cohort-revenue in Q3: was described and inferred, sized at the former estimate mdb-2026-cq3-f39 (an assumed share of Atlas-related revenue); now directional and unsized, because the quarter's measures are a record count of new customers, many of them AI natives, and a share of new Voyage customers by count, and the CFO's bound, small, is never converted. This moves the Q3 revenue total. Q1 and Q2, which give words and no count, keep their ballparks, so the step comes from the quarter's wording, not from a change at the company. The derived share of revenue, mdb-2026-cq3-f55, leaves the reading's metrics.

Reading ea-search-vector-search-addon in Q3: was described and inferred, sized at the former estimate mdb-2026-cq3-f45 (an assumed share of the increase in Enterprise Advanced and other revenue); now unsized, because the CFO credits the line's growth to the product's growing strategic importance beside the launch, and the charge covers full-text search, which is not AI, with vector search. The line's increase, $48mn, is quoted as the ceiling. The reading overlapped ea-demand-from-ai, itself unsized, so no flow total moves.

Assessments of Q1 and Q3: the sentences that sized Vector Search, Voyage, the cohort in Q3 and the add-on are rewritten.

Kept as it was, after review: enterprise AI workloads on Atlas, since each sized quarter has a passage naming AI alone.

Reading frontier-lab-revenue in 2026-CQ3: was left out of totals as overlapping ai-native-cohort-revenue, now counts in totals, because that channel is unsized in the quarter under the rules above and a component sized on its own evidence counts when its umbrella does not (overlap rule, 2026-10-06).

Reading ea-demand-from-ai in Q1, Q2 and Q3: was inferred, sized at the former estimates mdb-2026-cq1-f56, mdb-2026-cq2-f47 and mdb-2026-cq3-f44; now described and unsized, because each quarter names AI beside other causes: one of a variety of reasons to keep data on premises in Q1, traditional and AI applications together in Q2, and AI readiness beside resilience, sovereignty and capacity in Q3. The Q3 state stays directional on management's words.

Readings ea-demand-from-ai, ai-model-serving-and-training-compute, ai-tools-software-bill and internal-ai-productivity in all three quarters: motive was unknown, now exploratory, because the sources are silent or name joint causes rather than contradicting each other, and AI is named with no measure of its own reason or result.

Channel ai-model-serving-and-training-compute: counterparty was unknown, now mixed, because the bill is paid to cloud providers for serving and training compute and to third-party AI providers behind the assistants.

Not changed: the Q3 AI tools bill, $7.0mn, stays reported, since the 10-Q attributes the software cost increase to AI tools on its own.

Reading ai-platform-distribution in Q1: was inferred, sized at the former estimate klar-2026-cq1-f37 (a share of GMV at the take rate); now described and unsized, because the CEO calls the Gemini placement this week's launch on a call held after the quarter ended, and Stripe Link, the one placement live in the quarter, carries no stated agent volume, so the money had not started.

Figure klar-2026-cq2-f36 (Q2 size): the low-end basis no longer cites the Q1 estimate as its starting point, since that estimate is no longer a size; the values do not change. The Q2 reading stays sized because the Shopping Search app is live in ChatGPT; the Gemini surface is still described as coming.

Readings customer-service-automation in Q1 and Q2: were inferred, sized at the former estimates klar-2026-cq1-f35 and klar-2026-cq2-f34; now described and unsized, because the release credits operating leverage to AI-enabled productivity gains and cost discipline together and names no function. Cost discipline is added as a confound.

Readings workforce-productivity in Q1 and Q2: were inferred, sized at the former estimates klar-2026-cq1-f36 and klar-2026-cq2-f35; now described and unsized, for the same joint clause; revenue per employee, $1.4mn in Q1, is not itself attributed to AI in the sentence that gives it.

Readings customer-service-automation and workforce-productivity in Q1: motive was efficiency, now exploratory, because a result credited to several changes together does not meet the efficiency tell for any one of them. In Q2 the motive was unknown, now exploratory, because the sources do not contradict each other; the Q2 step from efficiency to unknown is gone with it.

Readings ai-vendor-bill in Q1 and Q2: motive was unknown, now exploratory, because the sources are silent on the bill rather than contradictory.

Channel ai-vendor-bill: counterparty was ai-lab, now mixed, because the bill pays model providers for models and tokens and compute providers for AI-specific compute.

Channel customer-service-automation: counterparty was unknown, now mixed, because the displaced cost is in-house and outsourced agents.

Reading ai-cash-tool in Q1 and Q2: was inferred, sized at zero with a pilot-scale high end (jpm-2026-cq1-f30, jpm-2026-cq2-f50); now described and unsized, because the CFO says the tool is not live and the CEO calls it a test case, and a product not yet launched is never sized at zero.

Reading operations-servicing-automation in Q2: was directional and inferred, sized at the former estimate jpm-2026-cq2-f46; now directional and unsized, because the only measures are shares of jobs cut in discrete areas, a count with no period or area, and the CEO places the cuts in a long run of automation beside AI. The job-cut shares and the operating-center headcount stay in its metrics.

Readings operations-servicing-automation and employee-productivity-tools in Q1: were described and inferred, sized at the former estimates jpm-2026-cq1-f26 and jpm-2026-cq1-f27; now not mentioned and unsized, because their only support was the CEO's company-wide remark on AI and the efficiency ratio, which names no function and is kept as context (claims c8, c9 and c12 now carry no channel). Q2 stays as read, with its own claims.

Channel ai-buildout-financing: was a registered revenue channel, described and unsized in Q2; now withdrawn from the ledger as not an AI channel and read not mentioned, because a lending door is a channel only where the company itself ties the money to AI, and AI entered through the analyst's question while the CFO declined to tie the loan growth to AI. The channel is cited by the published entry 2026-10-06-jpm-rules, so it is kept with its id; its description says it is withdrawn, and its claims are kept as context.

Reading ai-buildout-financing in Q2: was inferred, sized at the former estimate jpm-2026-cq2-f51; now described and unsized, because the CFO credits the loan growth to capital spending of which the AI buildout is an unseparated part. Other capital spending is added as a confound.

Readings ai-buildout-financing in Q2 and ai-prospecting-personalization in Q1 and Q2: motive was unknown, now exploratory, because AI is named as a cause with no measure and the sources do not contradict each other.

Steps. Q2 carries two: AI disclosure widened, from a Q1 with no AI passage in any source to a Q2 call that names AI once and names the AI product again (two of the three Q2 passages count only through the product name); and the CARTOSOUND SONATA channel opened, read described.

Johnson & Johnson enters with Q1 and Q2 2026 and one channel, opened in Q2: sales of electrophysiology mapping and imaging equipment that the CEO credits with AI-powered capabilities (revenue, end-use layer, hospitals as payers). The anchor already shows the activity: the FY2024 10-K lists electrophysiology products in the Cardiovascular portfolio, and Biosense Webster sold high density mapping catheters and ultrasound catheters; electrophysiology sales were $5.27bn in FY2024. With no measure of AI-attributed demand, the channel is tagged relabelled.

Q1 2026, the quarter ended 2026-03-29. No source names AI. Sales were $24.06bn, up 9.9%; research and development expense grew 9.4% to $3.53bn; selling, marketing and administrative expenses rose by $922mn on advertising phasing and launch investment. The robotic surgery programs, OTTAVA and MONARCH, are described as robotic, not as AI.

Q2 2026, the quarter ended 2026-06-28. The CEO says the debut of CARTOSOUND SONATA brought AI-powered imaging and mapping capabilities to electrophysiology; the CFO expects its continued U.S. adoption, and the release lists the debut as CARTO-powered innovation, neither saying AI. No sales, price or unit count is given, the 10-Q credits electrophysiology growth to new ablation catheters and does not name the product, and the AI capability is one attribute of a product with its own hardware, so nothing separates the AI part and no ballpark is built. Electrophysiology sales were $1.53bn, up $65mn. OTTAVA is described by automated features and the data-driven insights of the Polyphonic digital ecosystem, never as AI, and is not a channel.

No channel is sized, by management or by the ledger. The margin improvement the CFO raised to 0.75%, the Surgery franchise restructuring, a new Innovative Medicine supply chain restructuring that recorded $200mn in Q2, and a multi-year initiative to standardize back-office processes and systems are not attributed to AI, so none is an AI saving.

Beside the pair. Merck, the ledger’s other pharmaceutical company, named AI on its Q1 call through a Google Cloud partnership and drug-research collaborations and then went silent; Johnson & Johnson is silent in Q1 and names AI only in Q2, in a device sold to hospitals. Neither gives an amount, and neither names an AI result in drug discovery or clinical development, so the research and development door the roster expected is still unopened at both. The difference is the door: Merck’s one channel is a vendor bill it pays, while Johnson & Johnson’s is a product feature it sells. UnitedHealth Group, the health care company on the ledger that gives AI amounts, does so on the payer and services side; nothing at Johnson & Johnson is comparable to it.

Reading ai-research-and-development in Q1 and Q2: was directional and inferred, sized at the former estimates ibm-2026-cq1-f87 and ibm-2026-cq2-f78 (an assumed AI share of the whole research and development line); now described and unsized, because the filing names AI beside hybrid cloud and quantum as what the investment drives and nothing separates AI's part. The line is kept in metrics as the ceiling and the other causes are a confound. This moves the spend total in both quarters.

Reading lightwell-open-source-remediation in Q2: was described and inferred, sized at the former estimate ibm-2026-cq2-f83 (nothing at the point, an allowance for early adopters at the high end); now unsized, because general availability was announced on 2026-07-08, after the quarter ended, and client sign-ups are spoken of in the future tense, so the money had not started. Only the revenue total's high end moves.

Reading mainframe-ai-capacity in Q2: was bounded and inferred, sized at the former estimate ibm-2026-cq2-f76 (an assumed AI share of derived IBM Z revenue); now directional and unsized, because the quarter's new measure, nearly half of z17 customers investing in the AI accelerator, is a share of customers by count, and capacity growth is credited to AI, analytics and Linux workloads together. The derived line is quoted as the ceiling. Q1 keeps its size, since the CEO named AI alone as a kind of capacity sold; the step comes from the wording, not from a change at the company. This moves the Q2 revenue total.

Assessment of Q2: the sentence that put Lightwell's revenue at a size is rewritten, and the mainframe and research and development readings are said to be unsized.

Not changed, for review: lightwell-build-investment in Q2 still lists ai-research-and-development in its overlaps and so stays out of the spend total, although that channel now carries no size; dropping the overlap would add the Lightwell build to the total.

Kept as they were, after review: the Q1 mainframe reading, and forward-deployed AI talent, which names a staff role and an investment.

Reading lightwell-build-investment in 2026-CQ2: was left out of totals as overlapping ai-research-and-development, now counts in totals, because that channel is unsized in the quarter under the rules above and a component sized on its own evidence counts when its umbrella does not (overlap rule, 2026-10-06).

Figure ibm-2026-cq1-f82 (Q1 AI software revenue): was a quarter of the trailing-year figure, $1.50bn, scaled by the ledger's own seasonality multiple; now the fixed annual-plan band on that amount, $300mn to $450mn, because an amount for another period converts only through the band, and a completed trailing year is an actual that takes the annual-plan band. The seasonality of transactional software stays a confound. The share of software revenue, 5.3%, and the model bill built on the estimate, $20mn to $247mn, move with it.

Figure ibm-2026-cq2-f73 (Q2 AI software revenue): was a quarter of the Q1 trailing-year figure times the ledger's own roll-forward; now the annual-plan band on that amount times a separate growth assumption, $255mn to $630mn, with the same point and a wider range.

Figure ibm-2026-cq2-f84 (Lightwell spending in Q2): was the commitment, $5.00bn, over a span of years the ledger borrowed from another plan; now the fixed undated band, $62mn to $1.25bn, because the commitment states no period. The reading overlaps research and development spending on AI, so no total moves.

Readings client-zero-ai-productivity in Q1 and Q2: were inferred, sized at the former estimates ibm-2026-cq1-f89 and ibm-2026-cq2-f80; now described and unsized, because the selling, general and administrative benefit is credited to productivity actions in which AI is one lever beside spend cuts, procurement, sales efficiency, supply chain and workforce rebalancing, and no share is given for AI.

Readings consulting-delivery-productivity in Q1 and Q2: were inferred, sized at the former estimates ibm-2026-cq1-f91 and ibm-2026-cq2-f82; now described and unsized, because the filing credits the consulting margin to productivity actions generally and AI is one lever of them.

Readings distributed-infrastructure-ai-demand in Q1 and Q2: were inferred, sized at the former estimates ibm-2026-cq1-f86 and ibm-2026-cq2-f77; now described and unsized, because the CEO names generative AI beside the new flash and Power products in Q1, and AI adoption beside data growth and buying ahead of shortages in Q2, with no split. The Q1 motive was unknown, now exploratory, because the call and the filing agree; the Q2 motive stays unknown, because the CEO and the filing give different causes for the same growth.

Readings model-and-inference-cost in Q1 and Q2: motive was unknown, now exploratory, because the sources are silent on cost rather than contradictory.

Readings platform-and-data-demand-from-ai in Q1 and Q2: motive was narrative-defensive, now exploratory, because AI is named as a cause with no measure, and answering an investor question about software durability is not a narrative-defensive tell.

Channels model-and-inference-cost and lightwell-build-investment: counterparty was unknown, now mixed: frontier model providers and the compute behind the company's own models, and the company's own and Red Hat engineers beside outside frontier capability.

Motive of client-zero-ai-productivity, consulting-delivery-productivity (2026-CQ1, 2026-CQ2): was efficiency, now exploratory, because the shrinking line is credited to AI together with other causes and nothing separates AI's part, so the result does not meet the efficiency tell for any one of them.

Channels mainframe-ai-capacity and distributed-infrastructure-ai-demand: layer hardware. Every other channel is end-use.

Home Depot enters with every covered quarter of its fiscal year ending near the end of January and these channels: sales caused by Magic Apron, AI in search, recommendations, chat and review summaries, Pro sales through the AI material list builder and Blueprint Takeoffs, store labor saved by AI tools for associates, the AI tooling bill, and sales lost to AI shopping tools. The tooling bill and the lost sales are new money; Magic Apron, the Pro list tools and associate tools are existing activities AI changes; the search features are an activity the anchor already shows, now described with AI, and stay out of the incremental total. Order routing by machine learning (ship from best location) is described without AI and is not a channel.

Q4 FY2025, the quarter ended 2026-02-01. The 10-K describes every tool once and adds AI tools and generative and agentic AI to the competitive risks; on the call the CEO says AI tools let the company introduce list builders and an AI takeoff for Pros, and the head of Pro calls the takeoffs in strong early days. Labor productivity on the call is credited to moving store tasking to the merchandising execution team, not AI, and the SG&A rate is explained by payroll and a prior-year legal benefit. Net sales fell −3.8% against a quarter with an extra week.

Q1 FY2026, the quarter ended 2026-05-03. Steps: Magic Apron, the search features, associate tools, the tooling bill and the competitive risk moved from described to not mentioned; only the Pro list tools are named, now in one digital workspace. The merchandising head credits online growth of 10.5% to search, recommendations and delivery speed described as technology.

Q2 FY2026, the quarter ended 2026-08-02. Steps: AI disclosure widened on the call; Magic Apron returned as directional, with a monthly question count and an in-store version for associates and customers; associate tools returned as described; the Pro list tools moved from described to directional, the head of Pro saying Pros purchase more with the AI-powered material list builder, without a number. Online sales grew 11%, credited to traffic, conversion, the app and delivery speed rather than to Magic Apron.

Every size on the page is the ledger's own: shares of online sales for Magic Apron, the search features and the Pro list tools, and a share of the operating expense line for associate labor (counting only the share of saved hours that leaves payroll) and for the tooling bill. Sidekick, which the anchor call described as machine learning, and Computer Vision, which the fiscal 2025 10-K describes as technology for seeing where product is located, are not called AI in any covered source and are confounds, not channels. Beside Walmart the doors match (a shopping assistant, associate tools, a model bill, a competitive toll), but Home Depot gives no usage share or order measure for its assistant, only a question count, so its ranges are wider and rest on judgment alone. Home Depot gives no coding, customer care or workforce statement about AI at all.

Readings core-platform-demand-from-ai in Q1 and Q2: were inferred, sized at the former estimates gtlb-2026-cq1-f31 and gtlb-2026-cq2-f32; now unsized (described in Q1, directional in Q2), because in Q1 AI is credited with code volume and the revenue remark does not name AI, and in Q2 the revenue-tying statements are future or name AI beside security and compliance while the measures are code pushes and pipelines. The subscription increase is quoted as the ceiling. The channel stays relabelled.

Reading ai-native-customers in 2026-CQ2: was left out of totals as overlapping core-platform-demand-from-ai, now counts in totals, because that channel is unsized in the quarter under the rules above and a component sized on its own evidence counts when its umbrella does not (overlap rule, 2026-10-06).

Reading premium-seat-growth-pressure in Q1: was inferred, sized at the former estimate gtlb-2026-cq1-f34; now described and unsized, because management names rising AI code experimentation together with a Premium price increase, flat SaaS budgets and its upmarket shift, with no split.

Reading premium-seat-growth-pressure in Q2: was inferred, sized at the former estimate gtlb-2026-cq2-f36; now described and unsized, because the CFO credits the extra seat contraction to customer layoffs and mergers, and the CEO names spending shifting to AI only among several reasons for shorter agreements.

Reading internal-ai-process-automation in Q3: was inferred, sized at the former estimate gtlb-2026-cq3-f50; now described and unsized, as in Q2, because AI automation of reviews, approvals and handoffs is one of the restructuring's several operational changes and no source gives its share of the affected roles.

Reading core-platform-demand-from-ai in Q1, Q2 and Q3: motive was narrative-defensive, now exploratory, because AI is credited with code volume, activity rates and part of large-deal growth with no revenue measure, and no narrative-defensive tell of its own is quoted; answering investor uncertainty is not one.

Reading ai-native-customers in Q2 and Q3: motive was unknown, now exploratory, because the sources say nothing of why the labs subscribe and do not contradict each other.

Channel model-inference-cost: counterparty was model-provider, now mixed: third-party model vendors and the cloud providers that host the AI features.

Goldman Sachs enters with Q1 and Q2 2026 and these channels: internal investment in AI deployment (cloud, data and engineering), staff work across the firm displaced or avoided by AI, engineering work saved by large language model tools, the model access and AI tool bill, cyber defense against AI-enabled attacks, advisory, underwriting and markets activity from AI themes and the AI buildout, and lending and financing for the AI infrastructure buildout. The model bill is new money by the ledger’s test. Build investment, firm-wide productivity and engineering productivity are existing work into which LLM tools were put, read as expanded; the anchor call already described engineers applying AI with a focus on developer productivity. Cyber defense, capital markets activity and buildout financing are relabelled: the anchor already shows them, the anchor call already expected AI infrastructure financing to be a tailwind, and no movement is measured.

Q1 2026. The CEO said the firm implemented new technologies across the initial OneGS 3.0 work streams, and the CFO said early learnings led the firm to accelerate investment in cloud migration and data, which he called critical to deploying AI and expects to unlock productivity over time. OneGS 3.0 is an operating program of its own, so it is a confound and nothing is credited to AI; of the roughly $750mn rise in non-compensation expenses the CFO put about $650mn on transaction-based expenses. Asked about AI risks to banking infrastructure, the CEO said the firm is accelerating its cyber investment as the models improve, and that it has Anthropic’s new model and works closely with Anthropic. That opens the model bill, described and unsized: no charge, seat count or term is stated, as at Morgan Stanley, which opened its bill on the same model. Anthropic is private and not on the ledger. AI-related capital investment and AI-driven disruption were named among several forces behind client activity and volatility.

Q2 2026, the steps. AI disclosure widened on the call; engineering productivity and buildout financing opened, and the cyber toll went quiet. The CFO said AI use started with productivity in engineering, through several large language model providers and engineers hired from outside, which opens engineering productivity with a ballpark of $122k to $35mn, every multiplier a judgment and the share taken as lower cost set low because the CFO calls the expense base structurally unchanged; the model bill stays unsized, since the providers are named with no charge or term. The CEO and CFO tied the firm’s financing to the AI capital spending cycle, which opens buildout financing as a channel, unsized: no amount, share or count is given, and FICC and Equities financing rose 61.6%. The CFO said technology lets people do more and that teams may not replace all leavers, while calling the expense base structurally unchanged; headcount fell −1.7% over the quarter, credited to the firm’s efforts without separating AI from OneGS 3.0. The cyber toll reads not-mentioned: the call is silent, and the 10-Q sentence on AI-enabled fraud is carried word for word from the anchor’s annual report, which is not a mention in the quarter.

The roster’s notes were tested against the sources. A firm-wide AI assistant is in no covered source and is not used. AI in engineering is in the Q2 call and is a channel. OneGS 3.0, the operating-efficiency program, is named beside AI in both quarters and is kept as a confound; nothing it saves is credited to AI.

Reading cloud-ai-solutions in Q2: was directional and inferred, sized at the former estimate googl-2026-cq2-f19; now directional and unsized, because core GCP, AI solutions and AI infrastructure were all named as important drivers of Cloud growth with no rate or ranking for any. The Q1 rate and ranking are not repeated, so the step from Q1 comes from the wording, not from a change at the company. Q1 stays sized.

Readings cloud-ai-infrastructure in Q1 and Q2: were directional and inferred, sized at the former estimates googl-2026-cq1-f19 and googl-2026-cq2-f20; now directional and unsized, because AI infrastructure is named beside AI solutions and core GCP as a driver of one rise with no rate or ranking of its own. Cloud revenue is quoted as the ceiling in each quarter; the non-AI remainder figures googl-2026-cq1-f64 and googl-2026-cq2-f77 are no longer cited.

Readings ads-customer-support-automation and smb-agentic-acquisition in Q2: were directional and inferred, sized at the former estimates googl-2026-cq2-f66 and googl-2026-cq2-f69; now directional and unsized, because the only measures are a share of support queries by count, 75%, and a count of new customers given in words.

Figures googl-2026-cq1-f44, googl-2026-cq1-f47, googl-2026-cq2-f52 and googl-2026-cq2-f55 (depreciation and data-centre running costs of the AI build): the AI share of the infrastructure added since 2024 was the ledger's judgment; it now rests on the CFO's bound on capital expenditures, the overwhelming majority in Q1 and the vast majority in Q2 for AI technical infrastructure (claims c13 and c20, now cited in the readings), read through the words table. Depreciation becomes $2.12bn to $2.65bn and $2.46bn to $3.11bn, running costs $319mn to $1.33bn and $369mn to $1.56bn.

Alphabet enters with Q1 and Q2 2026 and these channels: Google Cloud AI solutions (products built on its generative models), AI infrastructure rented to AI labs and other firms, TPU systems delivered to customer data centers, consumer AI plans in Google One, AI Overviews and AI Mode in Search, Gemini in the ads systems, capital spending on AI technical infrastructure, depreciation and running costs of the build, shared AI research and development in Alphabet-level activities, the research and development compensation rise for AI talent, marketing for the Gemini app, agentic coding in engineering, and from Q2 ads support handled by Gemini agents, Gemini pitch tools in the sales force, agentic products bringing new small advertisers, and third-party capacity used as a bridge. AI solutions and consumer AI plans are new money by the ledger’s test; rented capacity, TPU systems, the Search features, Gemini in the ads systems and the cost channels are existing activities AI changed, read as expanded. The build’s cost channels take the exploratory motive, the least durable of the tells: compute goes first to frontier models, Search AI is served with no price change, and Cloud capacity is priced.

Q1 2026. The CEO said enterprise AI solutions became Cloud’s primary growth driver for the first time, with revenue from products built on its generative models up nearly 800%; Cloud revenue rose $7.77bn to $20.03bn and the Cloud backlog reached $462.30bn, including multi-gigawatt TPU hardware agreements with capital markets firms and, in certain cases, frontier AI labs. For certain of those agreements the company backstops data center and power obligations of third parties; the buyers pay the price, so funding is read as mixed. The ledger’s estimates, built from the Cloud rise so that non-AI Cloud still grows at their high ends, are $2.62bn to $4.81bn for AI solutions and $1.39bn to $3.41bn for rented AI capacity. Capital expenditures were $35.70bn, the overwhelming majority for AI, which the words table puts at $28.56bn to $35.70bn, and Alphabet-level activities, primarily shared AI research and development, lost −$5.39bn.

Q2 2026, the steps. TPU system revenue began, a small amount the ledger puts at $245mn to $2.02bn, and the channel moves to offensive; engineering coding moves to directional on one Chrome team’s projected speed-up. Channels that open: ads support, where Gemini agents address 75% of queries, the efficiency tell; Gemini pitch tools in the sales force, with up to 20% higher win rates; agentic products that brought hundreds of thousands of new small advertisers; and third-party capacity, already in use at a cost not disclosed and to expand in Q3, so it is inscrutable this quarter. Booking Holdings (BKNG, on the ledger) expanded its Cloud commitment, a ledger company’s spend in Google Cloud revenue. Free cash flow was negative by $5.90bn and the company raised equity of $49.60bn with AI infrastructure among the stated uses, so the build is now funded beyond operating cash flow.

What is not sized, and why. Search reads described: AI Overviews and AI Mode are said to drive query growth and Search & other rose $9.08bn, but management credits the growth to many parts of the business together, so nothing is attributed to AI. Gemini in the ads systems and marketing for the Gemini app are unsized for the same reason. The roster expected equity marks on stakes in AI labs: other income held gains on equity securities of $36.91bn in Q1 and $99.03bn in Q2, which the 10-Q attributes mainly to SpaceX and an unnamed private company, and the company committed $40.00bn to a private company it does not name. No source names Anthropic or any AI lab among the holdings, so no equity-mark channel is registered; these are recorded as context.

Layers. The largest revenue channels, AI solutions and rented AI capacity in Google Cloud, are compute: model access and accelerator capacity sold to builders, with Gemini Enterprise inside AI solutions as an end-use part the filings do not size apart. TPU systems are hardware. The build’s capital spending, depreciation and running costs, and shared AI research and development, are end-use: the 10-Q allocates the infrastructure by usage and Google Services carries the larger share, with the capacity Google Cloud sells as the other use; the third-party bridge capacity is compute. No source names a lab buyer as a company Alphabet has put money into, so lab payments read as investor capital inside a mixed payer base. Part of the capital spending is accelerator purchases that are revenue at Nvidia (NVDA), and where AI labs rent capacity or pay for TPU systems the revenue is the buildout’s money moving; the third-party capacity bought as a bridge may be revenue at another capacity seller. The ledger records Alphabet’s own flows and does not net them. Shared AI research and development includes the technical infrastructure usage costs of model development, which the depreciation and operating cost channels already count, so the ledger takes an assumed infrastructure share out of the shared research size and counts each dollar once; the research and development compensation rise is left out as overlapping.

FedEx enters with Q3 and Q4 of fiscal 2026, the quarters ended 2026-02-28 and 2026-05-31, mapped to 2026-CQ1 and 2026-CQ2. The anchor already shows much of what the later sources call AI: the fiscal 2024 10-K credited AI and machine learning with supporting the DRIVE transformation (claim fdx-anchor-c1), described AI-enabled robotic sortation tests (claim fdx-anchor-c2), described FedEx Virtual Assistant in the words the fiscal 2026 10-K repeats (claim fdx-anchor-c3), and sold FedEx Returns Technology to merchants (claim fdx-anchor-c4). The robots, the workflows and the Virtual Assistant are therefore read as relabelled and add nothing to the incremental total. Tracking+ and Returns+ are read as expanded, new AI products sold beside Returns Technology, and stay unsized with no stated price.

Q3 FY2026. The CEO calls physical AI a critical element of the longer-term strategy and names robotic trailer unloading and loading pilots with Berkshire Grey and Dexterity, to be deployed further later in calendar 2026 (claims c1 and c2); the Chief Customer Officer announces FedEx Returns Plus, an AI-powered tracking and returns offering (claim c3). Both channels open unsized: the robots are running pilots with no unit count or cost, so they are read as inscrutable, and the product has no stated price. A pricing remark credits machine learning tools and never names AI, so it opens no channel.

Q4 FY2026. Steps: the robots are not mentioned, and the workflow, Virtual Assistant, developer assistant, data-centre demand and spending channels open. The 10-K says AI-enabled workflows, together with the digital backbone, lower the cost to serve, and the CEO says AI is embedded in the DRIVE process; with AI named beside other causes and no measure, the channel stays unsized with total operating expenses of $23.46bn as the ceiling. The Chief Customer Officer calls the AI and data center space a growth engine delivering double-digit revenue growth and declines to size it (claims c14 and c17); that channel sits in the facilities layer, directional and unsized. The 10-K adds an AI Council, an AI policy and an AI Literacy Program for all team members, and an AI assistant in the Developer Portal, introduced in May 2026 and left inscrutable. Partnerships with ServiceNow and Dun & Bradstreet that include AI-powered supply chain workflows are context: announced, with no launch, customer, price or term, and no stated direction of money (claim c13).

Both sizes are the ledger’s own: $2.5mn to $96mn of income-statement AI spending, a share of revenue taken from the ledger’s own estimates for other large buyers, and $0 to $17mn of customer service cost the Virtual Assistant may have avoided against the prior-year rate, with a range that allows no saving at all. The doors FedEx shares with UPS in the ledger are network planning and AI build spending. FedEx’s AI-enabled workflows and UPS’s network planning are both read as relabelled and unsized, with AI named beside other causes and every measured saving credited to network programs; AI build spending is expanded, sized as a level with no traced baseline, at both companies from a reference-class share of revenue. FedEx additionally registers robots, a customer assistant, merchant products, a developer assistant and data-centre demand.

Comparisons. FedEx Freight was spun off on 2026-06-01, after both quarters, so both still consolidate it. The 8-K of 2026-07-21 recasts calendar 2024 and 2025 quarters with Freight as discontinued operations and new segments; it is not a results release and names no AI. FedEx moved to a calendar year from 2026-06-01, with a transition period to 2026-12-31, so its next quarters neither consolidate Freight nor follow the fiscal calendar used here.

Reading ai-skills-hiring in Q1: was inferred, sized at the former estimate expe-2026-cq1-f47; now described and unsized, because the CFO says the company is going to be adding back skills and the personnel line fell, so the money had not started.

Reading traveler-self-service in Q1: was directional and inferred, sized at the former estimate expe-2026-cq1-f48 (the first rules pass kept the size); now directional and shape, unsized, because the only AI measures are a count of service interactions and a share of self-service resolutions by count, and the cost movement is credited to payments and customer service efficiencies with no AI named.

Readings traveler-self-service and agent-assist in Q2: were described and inferred, sized at the former estimates expe-2026-cq2-f46 and expe-2026-cq2-f47 on the prior quarter's claims; now not mentioned and unsized, because the call is silent on both and the only remaining passage is the 10-Q's one-clause aim of integrating AI into customer service operations, kept as context (claims c37, c40 and c42 no longer carry these channels).

Channels supply-content-enrichment and partner-advertising-tools: were sized in Q1 at the former estimates expe-2026-cq1-f54 and expe-2026-cq1-f57; now withdrawn from the ledger as not AI channels and read not mentioned in both quarters, because each rested on one clause in the prepared remarks with no tool, deployment or measure. partner-advertising-tools is cited by the published entry 2026-10-06-expe-rules and supply-content-enrichment by the published blog snapshot, so both keep their ids; their descriptions say they are withdrawn, and their claims are kept as context.

Channel organic-search-traffic: was a toll registered in Q2, directional and inferred, sized at the former estimate expe-2026-cq2-f56; now withdrawn from the ledger as not an AI channel and read not mentioned, because no source names AI as the cause of the SEO softness: management speaks of algorithm and search page changes, and AI appears only as the team's own tool. The channel keeps its id, since the published blog snapshot lists it; its claims are kept as context, those also on ai-platform-referrals stay there.

Checked and not changed: marketing creative in Q2 stays sized, since Q1 already had AI improving creative in production.

Reading agent-assist in 2026-CQ1: was left out of totals as overlapping traveler-self-service, now counts in totals, because that channel is unsized in the quarter under the rules above and a component sized on its own evidence counts when its umbrella does not (overlap rule, 2026-10-06).

Reading marketing-creative in 2026-CQ1: was left out of totals as overlapping marketing-spend-efficiency, now counts in totals, because that channel is unsized in the quarter under the rules above and a component sized on its own evidence counts when its umbrella does not (overlap rule, 2026-10-06).

Reading marketing-creative in 2026-CQ2: was left out of totals as overlapping marketing-spend-efficiency, now counts in totals, because that channel is unsized in the quarter under the rules above and a component sized on its own evidence counts when its umbrella does not (overlap rule, 2026-10-06).

Reading partner-support-automation in 2026-CQ2: was left out of totals as overlapping traveler-self-service, now counts in totals, because that channel is unsized in the quarter under the rules above and a component sized on its own evidence counts when its umbrella does not (overlap rule, 2026-10-06).

Reading marketing-spend-efficiency in Q1: was inferred and efficiency, sized at the former estimate expe-2026-cq1-f50; now described, unsized and exploratory, because the CEO credits the leverage to raised return thresholds and removed unprofitable spend, with AI as a helper in testing and reallocation, and a result credited to several changes does not meet the efficiency tell for any one of them.

Reading marketing-spend-efficiency in Q2: was inferred and efficiency, sized at the former estimate expe-2026-cq2-f48; now described, unsized and exploratory, because the CEO lists the team's use of technology beside measurement and targeting, and the CFO credits spend reductions; measurement and targeting are added as a confound.

Reading partner-onboarding in Q1: was directional and inferred, sized at the former estimate expe-2026-cq1-f53; now described and unsized, because faster AI onboarding is named beside property count growth, 10%, which belongs to all supply and carries no AI share.

Readings ai-vendor-bill, marketing-creative, partner-advertising-tools and partner-onboarding (both quarters) and internal-ai-adoption (Q1): motive was unknown, now exploratory, because AI is named with no measure and no line moving and the sources do not contradict each other; silent quarters carry the corrected motive. Internal adoption in Q2 stays unknown, since management and the filing give different causes for the same personnel fall.

Reading traveler-self-service in Q1: state was quantified, now directional, because interactions handled, 250,000,000, and the AI share of self-service resolutions, 30%, are counts and shares of AI work with no dollar level. The size stays the ledger's estimate.

Channel ai-vendor-bill: counterparty was ai-lab, now mixed, because the bill pays model providers for tokens, AI software licensors and cloud providers for AI-specific capacity.

Steps. In Q2 2026 the Distributed AI Hub, launched in Q1, moved from described to not mentioned. Nothing else moved.

Equinix enters with Q1 and Q2 2026. At the facilities layer, the space, power and interconnection sold to whoever runs the compute, are colocation and interconnection sold for AI workloads in the retail data centers; fees from the xScale joint ventures, which build and lease hyperscale data centers; the equity put into those ventures; capital expenditure on retail capacity; and the Distributed AI Hub, an on-ramp to AI model companies and GPU clouds read as relabelled, since the anchor already sells private on-ramps through Fabric and nothing new moved. At the end-use layer are Fabric Intelligence, an AI feature on the interconnection platform in preview with no price, and the company’s own AI tools, read as a saving and a vendor bill. Capital spending in both forms is traced and left out of flow totals. The facilities channels other than the AI Hub are expanded: the anchor already shows AI-linked colocation and xScale demand and the Singapore data center designed for AI, and during coverage new capacity was built and contracted for data-centre customers, so volume moved. The AI tools bill is new, since a bill for third-party AI tools cannot exist without LLMs; the saving is expanded, since the anchor shows employees beginning to use AI tools.

Q1 2026. Recurring revenues were $2.33bn, against $2.09bn a year earlier, and annualized gross bookings $378mn. About 60% of the largest deals were AI-related, AI model providers and neoclouds had placed more than 110 network nodes, and 36 customer deployments used liquid cooling. The Hampton xScale lease slipped to Q2, so services revenue from the ventures was $38mn; the company sold the Hampton campus to the AMER 3 venture for $459mn. Cash purchases of property, plant and equipment were $1.26bn.

Q2 2026. Revenues were $2.63bn, against $2.26bn, lifted by 134 megawatts of xScale leases that contributed about $120mn of one-time fees; services revenue from the ventures was $152mn. Annualized gross bookings were $424mn. Cash purchases of property, plant and equipment were $1.58bn, the plan through 2029 is up to $7.00bn a year, and the 10-Q adds a warning that a correction in AI-related investment could reduce demand. The new CFO expects automating processes with AI, beside functionalization and standardization, to help margins; no saving is measured.

Sizes. The only sized channels are the company’s own AI tools, both the ledger’s inference: a saving of $351k to $11mn and a bill of $61k to $1.3mn in Q2. The AI part of retail revenue is unsized: the company’s measure is a share of the largest deals by count, which separates nothing in dollars, and recurring revenues and annualized gross bookings are quoted as its ceiling. The xScale fees, the equity put into the ventures and the capital expenditure are unsized because the company names AI beside cloud, or calls it an accelerant of enterprise modernization, and nothing separates AI’s part; their lines are quoted as ceilings. In Q1 the CEO also said AI inference needs no different capital from the existing build.

What the roster expected and what the sources show. The roster note said Equinix builds hyperscale sites through xScale joint ventures with outside partners and talks about AI and inference in bookings. The sources support both, with limits. The ventures are real and off the balance sheet, with the company holding 20% of each venture other than AMER 3 and an effective 25% of the AMER 3 venture’s assets, lending to one and guaranteeing part of another’s debt; but the stored sources name only one partner, PGIM Real Estate at the anchor, and none of the partners the roster named from memory. Canada Pension Plan Investment Board appears instead as the company’s partner in the planned purchase of atNorth, a data center company, which is an acquisition confound and not an AI channel. AI and inference are in every call, as a share of the largest deals and as customer examples, never as a share of bookings or revenue dollars. Some payers of this revenue may be on this ledger as MSFT, AMZN, GOOGL, META, ORCL and CRWV, where the same dollars are lease or colocation spend; cohort totals keep the facilities layer apart from compute and end-use, and are not netted.

Elevance Health enters with Q1 and Q2 2026 and these channels: investment in digital and AI-enabled capabilities (spend); member service through the AI-enabled virtual assistant, provider matching through Sydney, prior authorization through Health OS, medical cost work in trend detection and payment integrity, Carelon risk identification, administrative automation and associate productivity tools (savings); and Star Ratings bonus revenue from AI-powered member engagement (revenue, opened in Q2). Provider matching, prior authorization, medical cost work, Carelon and Star Ratings are tagged relabelled: the 2024 10-K already shows Sydney Health, the HealthOS platform, payment integrity, utilization management and Star Ratings bonuses, and no AI-attributed movement is measured in coverage. The rest are expanded. The first call of 2024 has no AI passage.

Q1 2026, the quarter ended 2026-03-31. Nearly every AI passage is in the CEO’s answer to one analyst question. She gives counts, not dollars: 22,000,000 commercial members using the AI-enabled virtual assistant, more than 20% of members connected through Sydney, more than 60,000 associates with access to AI tools, and missing-information prior authorization denials down by more than almost 70%, as spoken, credited to Health OS and AI together. On medical cost she says AI surfaces outliers and strengthens payment integrity, and that the cost effect comes over time. The 10-Q names AI only among the aims of the Operating Model Transformation Program, whose charges of $129mn were for severance; the program is read as a confound, never as an AI saving. Adjusted operating expense was flat at $5.21bn, credited in the release to disciplined expense management.

Q2 2026, the quarter ended 2026-06-30. Steps: AI disclosure thinned on the call; the member assistant, provider matching, prior authorization, Carelon and associate-tool channels went to not mentioned, because Sydney Health and Health OS are described this quarter without AI; administrative automation moved from the ledger’s ballpark to described and unsized; and the Star Ratings channel opened. The CFO ties targeted investment spending of approximately $1 of EPS to AI adoption, workforce enablement and Carelon scaling, and adjusted operating expense ran $518mn above its prior-year share of operating revenue, which the 10-Q credits to investments in the workforce and technology adoption, regulatory matters and premium taxes without naming AI. AI is one of several named causes, so the investment channel stays described and the line’s movement is quoted as a ceiling, not read as a shape. The CEO frames medical cost management, and the expense effect on administrative cost, as analytics and AI-enabled tools whose effect is expected next year.

On the doors it shares with UnitedHealth and CVS Health (member service, prior authorization, claims and payment integrity), Elevance Health gives counts and rates, names no claims-processing tool, and states no AI-only amount. Utilization management is a regulatory subject here as at its peers, and the ledger quotes only the company’s words on denials. Administrative automation in Q1 carries the ledger’s only ballpark, $0 to $52mn, where the CEO names AI alone as reducing administrative expense; Q2 carries no size, and the joint-cause, count-measured and not-yet-started channels stay unsized.

Reading ai-content-production in Q1: was directional and inferred, sized at the former estimate duol-2026-cq1-f28 (the first rules pass kept the size); now directional and shape, unsized, because the only measures are counts of course units published, and the size rested on judgment shares of Research and development.

Figure duol-2026-cq2-f34 (feature inference bill, Q2): new. The reading ai-feature-inference was sized at the former estimate duol-2026-cq2-f30, a judgment share of cost of revenues, whose text said the CFO's tens of millions is not in the words-to-numbers table; the table does carry the spoken tens of millions, so the reading now converts the phrase through the fixed undated band, since the CFO gives no period, and stays inferred. The point falls from $4.1mn to $29mn to $833k to $25mn.

Figure duol-2026-cq2-f31 (Q2 internal AI use): was the CFO amount times the ledger's own phasing, read as an annual to half-yearly amount, now the fixed undated band, because an amount with no period converts only through that band; it is now $833k to $2.5mn. The quoted figure's label no longer reads it as annual.

Reading ai-content-production in Q1: state was quantified, now directional, because course units published, 20,500, are counts of AI work with no dollar level.

Channel ai-content-production: counterparty was internal, now mixed, because the work is done by employees and contractors whose cost sits in Research and development.

Channel ai-feature-inference: counterparty was model-provider, now mixed, because the bill pays third-party model providers and the hosting and compute behind the open source and local models the company is moving to.

Steps. None between Q1 and Q2: no channel changed state, strength or motive, and the Q2 call carried fewer AI passages than Q1 without meeting the disclosure-thinned test.

Duke Energy enters with Q1 and Q2 2026. On the facilities layer: electricity sold to data-centre customers, chiefly under electric service agreements, and the capital spending on generation and the grid that serves that load, capital and out of totals. These are expanded by the ledger’s test: selling power and building generation predate LLMs, and the anchor already shows data-centre usage in commercial sales; new capacity is contracted for data centres in coverage, so volume moved. With no traced quarter before AI they are undetermined in the incremental total. On end use: AI tools that track construction milestones on the generation build, read as relabelled because construction oversight was already part of the work at the anchor, the sources never say the tools are LLMs, and no line moves.

Q1 2026. The CEO said growth in the company’s regions is driven by innovation in AI technologies and advanced manufacturing, and that ESAs with data center customers rose by 2.7 GW in the quarter to about 7.6 GW, against a late-stage pipeline of 15.4 GW. The contracts carry minimum demand provisions, credit support, refundable capital advances and termination charges. Capital expenditures were $4.09bn, up 29.9% on the year, inside a capital plan of $103.00bn. Operation, maintenance and other rose on a legal settlement and Winter Storm Fern.

Q2 2026. Signed ESAs reached 7.8 GW, and the company expects the rest of the pipeline to convert by the first half of 2027. The CFO put the upside to the capital plan at $5.00bn to $10.00bn if further contracts are signed; it has not started. No Q2 source names AI as a cause of data-centre demand. The CEO again cited AI tools that monitor construction deadlines and progress. Operation, maintenance and other fell on lower storm amortization in Florida.

Sizes. No channel carries a size. The revenue channel stays unsized because the contracted load starts after the quarter and existing data-centre usage is never given as an amount; the capital channel because AI is named only beside other causes and no data-centre share of capital is stated, so the ceiling is quoted instead; the construction tools because the saving would fall on capitalized construction work, which a savings channel cannot carry as capital, and no saving is given.

What the roster expected and what the sources show. The roster note expected large-load and data-centre demand, signed electric service agreements and a large capital plan. The sources show each of them, and show that the company talks about data centres, large loads and economic development far more than about AI: across both quarters AI appears in a remark on demand, in the remarks on construction tools, and in a risk-factor line. Constellation Energy and Chevron, read at the same layer, also earn nothing yet from their data-centre power contracts. Hyperscalers on this ledger would carry the same money as data-centre cost once delivered; totals are not netted across companies, and cohort totals keep the facilities layer apart from compute, hardware and end use.

Reading hyperscaler-lab-training-revenue in Q1: was bounded and inferred, sized at the former estimate ddog-2026-cq1-f33; now described and unsized, because the CEO says the customers landed in Q1 do not contribute any revenue yet and the CFO calls training a future contributor, so the money had not started. The channel overlaps ai-native-cohort-revenue, so no flow total moves.

Channel ai-lab-gtm-partnerships: was a revenue channel sized in Q1 at the former estimate ddog-2026-cq1-f39; now withdrawn from the ledger as not an AI channel under the partnership rule, because the Sakana AI release states no money paid and no term. No published entry cites the channel, but it is kept in company.json with its id, reads not-mentioned and unsized in both quarters, and the Sakana passage stays as a context claim with no channel.

Reading engineering-ai-productivity in Q2: was described and inferred, sized at the former estimate ddog-2026-cq2-f36; now not-mentioned and unsized, because the quarter's only AI passage is a risk-factor heading that the company uses AI in its operations, a statement with no tool, deployment or measure, and the call no longer mentions AI coding tools. The heading is kept as context. Q1, which names engineers enabled with AI coding tools, keeps its size; the Q2 motive is carried from Q1.

Reading enterprise-ai-driven-platform-usage in Q1 and Q2: was inferred, sized at the former estimates ddog-2026-cq1-f36 and ddog-2026-cq2-f34; now described and unsized, because AI adoption is named together with cloud migration and greater product adoption as the cause of non-AI customers' acceleration, and the CFO says the company tries to separate the AI effect and gives no figure.

Reading enterprise-ai-driven-platform-usage in Q1 and Q2: motive was unknown, now exploratory, because AI is named as a cause with no measure and the sources do not contradict each other.

Reading ai-coding-tools-bill in Q1: motive was unknown, now exploratory, because a single mention of engineers building with AI coding tools carries no cost or measure; the silent Q2 reading carries it.

Channel ai-native-cohort-revenue: counterparty was ai-native, now mixed, because the cohort holds investor-funded AI-native start-ups and the AI research divisions of hyperscaler parents, which management counts together.

Channel ai-model-compute-and-services: counterparty was unknown, now mixed: cloud providers for training and inference capacity, and third-party AI service vendors.

Checked and not changed: engineering-ai-productivity stays narrative-defensive, and hyperscaler-lab-training-revenue stays ai-lab, since Datadog sells those labs observability, not capacity.

Reading delta-concierge in Q2: was inferred, sized at the former estimate dal-2026-cq2-f16; now described and unsized, because the NPS gain is credited to proactive communication, simplified rebooking, expanded self-service and the assistant together, and no care volume or cost is given.

Channel baggage-ai: domain was other, now operations, because baggage handling is now an operations domain.

Channel baggage-ai: counterparty was unknown, now mixed, because the cost avoided is the company's own baggage staff, contracted delivery of delayed bags and compensation paid to customers.

Not changed: baggage-ai stays sized, because the release attributes the fall in the Atlanta mishandled baggage rate, 25%, to the tool on its own; the COO's mention of system and process work stays a confound.

Delta enters with Q1 and Q2 2026. The steps are three, all in Q2: AI disclosure widened from a silent first quarter, and two savings channels opened. Q1 has no channel at all: the call explained the quarter by demand, the fuel spike, capacity cuts and operational recovery, and its digital topics (in-flight Wi-Fi, Delta Sync, satellite connectivity) were not described as AI. Non-fuel unit cost rose 6% on lower capacity growth and recovery costs.

Delta Concierge. The CEO and the release describe an AI-enabled digital assistant offering self-service and messaging during travel, rolled out to over 50% of SkyMiles members, with the full rollout due after quarter end. No contact volume, deflection rate or cost is given. The channel is tagged expanded and sized as an increment: at the anchor the care work existed, in reservations centers the company owns and in a Fly Delta app being evolved toward a digital travel concierge for notifications, but no assistant, messaging or self-service deflection was shown under any name. Care staff sit inside salaries, which rose 8.2% on pay increases. The ballpark is a care share of salaries times an assumed saving, with a low end of zero. Its motive stays exploratory: the only number given for the assistant is its rollout, and the NPS improvement the COO cites is credited to the whole set of service changes.

Baggage AI. The release credits proprietary Baggage AI in Atlanta with cutting the hub's year-to-date mishandled baggage rate by over 25% against the prior year, and June by 50%; on the call the COO shares the credit with enhancements to the baggage handling system and processes. The anchor does not mention baggage technology, and the release ties a measured rate to the AI, so the channel is expanded and sized as an increment. The company reports no cost per mishandled bag or mishandled rate, so the ballpark prices avoided bags at an industry average cost per bag, with the hub's prior-year rate a judgment, on a wide range. The channel is read as efficiency: the company attributes a measured operating result to an implemented tool, which the motive table now counts even where no cost line moves.

What is not a channel. Asked about long-term margins, the CEO said AI would make the company more efficient rather than grow it, with a large imprint that may take another year or two; the statement names no function and is kept as context, not as evidence about either channel. Predictive maintenance, better retailing through technology and the CFO's efficiency actions were described without AI being named and are not credited to AI. Revenue management, listed on the roster as a likely AI door, was not described as AI in any covered source. Delta's Q3 2026 report is due shortly after this entry.

Chevron enters with Q1 and Q2 2026. All the steps are in Q2: AI disclosure widened from a silent first quarter and the Kilby channels opened. Q1 has no channel. Its call was taken up by the war in the Middle East, refining, Venezuela and Tengizchevroil; the 10-Q reported an exclusivity agreement with Microsoft and Engine No. 1 to negotiate power from the West Texas project (claim c3), and the CEO spoke of new technologies for exploration (claim c6), neither with AI named. Data-centre demand is not an AI label, so neither opened a channel in Q1.

Project Kilby, revenue. The President of New Energies says U.S. electricity demand is shifting as AI accelerates, and the release quotes the CEO on powering American AI dominance. In June the company signed a take-or-pay power purchase agreement with Microsoft for about 2.7 gigawatts of firm behind-the-meter capacity over 20 years, ahead of a final investment decision later in 2026. No power is delivered and no price or contract value is given, so the channel is read described and left unsized until power flows. It sits in the facilities layer with Microsoft (MSFT) as the counterparty: once power flows it is the same money as power cost inside Microsoft's data-centre spending, and a data centre's power is a ceiling on the AI part. It is tagged expanded: the FY2024 10-K already planned gas-fired power for U.S. data centers, and in coverage new capacity is contracted for a data-centre customer.

Project Kilby, capital. Turbines and small block generation are secured, an EPC contractor is engaged and turbine deliveries begin during 2026; the CFO places the project in both the organic and the affiliate capital outlooks. No amount is given and no filing says which segment carries the project, so the channel is read described and unsized, with consolidated capex of $4.54bn and affiliate capex of $272mn quoted as the ceiling; U.S. capex outside upstream and downstream was $96mn against $111mn a year earlier. It is capital, shown and left out of the spend total. Funding is read as the company's own cash: no source names the affiliate or its partners, and the quarter's affiliate capex shows no power spending.

Exploration, as context. Asked about exploration, the CEO says the company is going to use new tools and that AI will change cycle time and what it can see in the subsurface, and he expects it to change outcomes; the discoveries he cites are credited to the portfolio. The remark is in the future tense, and exploration expenses rose $4.0mn on higher geological and geophysical engineering costs, so no channel is registered and no saving is sized.

What is not a channel. The structural cost program, on the roster as a possible AI door, reached $3.00bn of annual run-rate savings six months early, 70% from efficiency gains that the CFO credits to the reorganization, technology achievements and improvement initiatives; predictive maintenance in shale is described the same way. No source attributes any of it to AI, so it is a confound and not a saving. A Permian remark that there are interesting things going on with AI names no function and is kept as context. The 10-Q's only AI sentence is a risk factor on cyberattacks that leverage AI. Chevron's Q3 2026 report is due around the end of October.

Channel ai-vendor-bill: novelty was new, now expanded, sized as a level with no baselineUsd, because the anchor 10-K already describes investment in AI and in cloud capabilities (anchor claims cvs-anchor-c1 and cvs-anchor-c2, now tagged with the channel) and the channel's own description conceded that the conversational AI part predates LLMs; the tie-break takes expanded before new. The sizes do not change; they move from the incremental total to the undetermined amount.

Readings ai-build-investment in Q1 and Q2: were inferred, sized at the former estimates cvs-2026-cq1-f25 and cvs-2026-cq2-f54; now described and unsized, because every statement of the investment names AI beside technology (AI and broader technology, technology and AI, AI and emerging technologies) and nothing separates AI's part. The corporate line and the technology commitment's pace are quoted as the ceiling, with the technology program added as a confound.

Reading pharmacy-call-automation in Q2: was directional and inferred, sized at the former estimate cvs-2026-cq2-f49; now directional and unsized, because calls moved and pharmacist hours refocused, 1,000,000, are cumulative counts over the last couple of years with no period or start date, and a cumulative count sizes nothing.

Reading clinical-records-review-automation in Q2: was directional and inferred, sized at the former estimate cvs-2026-cq2-f50; now directional and unsized, because the only measure is a cumulative count of pages analyzed, 1,000,000,000, with no period.

Assessments in Q1 and Q2: rewritten so that they no longer cite the former estimates as sizes.

Figures cvs-2026-cq2-f43 and cvs-2026-cq2-f44 (the quarter's share of the cumulative savings figure): was a derived figure on the ledger's own reading of at least three years, now an estimate with the fixed cumulative-years band, because an amount built over the last few years converts only through that band. The points do not change; each now carries a range. The non-AI part stays the traced restructuring figure, $500mn.

Readings prior-authorization-automation and care-navigation-cost in Q1: were inferred, sized at the former estimates cvs-2026-cq1-f22 and cvs-2026-cq1-f23; now described and unsized, because the Q1 results are credited to embedded technology, automation and digital tools beside AI, and the navigation products to AI and digital tools together.

Reading adherence-analytics-revenue in Q2: was inferred, sized at the former estimate cvs-2026-cq2-f52; now described and unsized, because the adherence level is credited to technology, automation and AI together.

Readings care-navigation-cost (both quarters) and adherence-analytics-revenue: motive was narrative-defensive, now exploratory, because AI is named as a cause with no measure and no narrative-defensive tell of its own is quoted.

Reading pharmacy-call-automation in Q2: state was quantified, now directional, because calls moved and pharmacist hours refocused, 1,000,000, are counts of AI work with no dollar level.

Reading clinical-records-review-automation in Q2: state was bounded, now directional, because a floor on pages analyzed, 1,000,000,000, is a count of AI work.

Channels member-provider-service-automation and claims-processing-automation: counterparty was internal, now mixed, because the anchor 10-K says call center operations and claims processing are run directly or through vendors (claim cvs-anchor-c6).

CVS Health enters with Q1 and Q2 2026. The ledger registers channels for Health100, the AI-native engagement platform, as revenue; an AI vendor bill and internal AI investment as spend; prior authorization, claims processing, member and provider service, retail pharmacy calls, clinical record review and AI-guided navigation as savings; and adherence programs as revenue. Prior authorization, clinical record review, navigation and adherence are tagged relabelled: the anchor shows the company running real-time preauthorization tools, AI and machine learning on medical records, AI tools in care management and adherence algorithms before LLMs, and no AI-attributed movement is measured in coverage. The vendor bill is new; the rest are expanded.

Q1 2026. The release announced Health100, a subsidiary that will use Google Cloud AI technologies, and the call framed it as the consumer front door, launching later in the year. Asked for the level and pace of AI investment and when it would turn from net investment to net benefit, management answered with strategy and an AI academy at Aetna and gave no number. The prior authorization results the release reports, 83% in real time and more than 1,000,000 provider calls eliminated, are credited in writing to automation and digital tools. Insurance segment operating expenses ran −$63mn against their prior-year share of revenues, with no cause given.

Q2 2026. The claims, service, pharmacy call, clinical record and adherence channels opened. The release says the second-generation Claims Assist Manager, an AI-powered agentic claims advisor, cuts processing time by over 20% on complex claims that need manual review, and that agentic AI is being deployed on the call center platform for Aetna and CVS Caremark. On the call the Aetna president said case preparation for an advocate fell from 90 minutes to 2, the group president said hundreds of millions of pharmacy calls moved to conversational AI over a couple of years, refocusing 1,000,000 pharmacist hours on clinical care, and the CEO said AI has analyzed over 1,000,000,000 pages of clinical records. The targeted launch of Health100 with the HIO assistant began after the quarter ended.

Insurance segment operating expenses rose 3.1% against segment revenues up 3.5%, running −$20mn against their prior-year share, a small move in the direction the claims predict that the 10-Q explains by increased business investments without naming AI. Pharmacy segment operating expenses, without a prior-year litigation charge, rose 7.9% against revenues up 0.7%, against the claims. The CFO's savings figure was generated over the last few years and probably includes the 2024 restructuring plan, expected to save over $500mn in 2025 and not attributed to AI. Read as cumulative, on the span of years the figure label states, its yearly increment is $50mn to $125mn a quarter, and with the restructuring savings taken out $25mn to $63mn is left as a loose ceiling for AI. The ledger's savings estimates are set so that their high ends sum below it: claims $0 to $12mn, member and provider service $0 to $9.8mn, prior authorization $0 to $5.6mn, pharmacy calls $0 to $8.0mn, clinical records $0 to $4.7mn. The restated commitment of more than $20.00bn to technology over a decade may include capital spending, an inference from the anchor capital spending mix; internal AI spend is put at $2.5mn to $75mn of the expensed part, excluding outside vendors, and the vendor bill at $908k to $21mn. Health100 is left unsized until it is in market.

The steps between the quarters are those openings; no state, strength or motive changed on the channels carried from Q1. The call carried more AI passages in Q2 than in Q1, short of the doubling the ledger records as a disclosure step. Compared with Progressive on the claims door and Airbnb on the support door, CVS Health gives operating measures (processing time, case minutes, calls, hours) where Airbnb's filing gave a dollar saving; it gives one combined technology-and-AI savings figure and no AI split.

Steps. In Q2 2026: managed inference moved from described to quantified and from exploratory to offensive, with booked annual recurring revenue of more than $100mn from $1.0mn; storage, CPU, networking and software moved from described to quantified on a run rate of more than $400mn. Nothing else moved.

CoreWeave enters with Q1 and Q2 2026. Revenue is read by payer, because that is the question the company was added for: hyperscalers and Meta, AI labs and AI-native companies, and established enterprises, each an assumed share of reported revenue whose points sum to it. The lab channel is new money by the ledger’s test, since its payers exist because of LLMs; the hyperscaler and enterprise channels are expanded, since renting GPU capacity predates LLMs at the company itself, which ran crypto mining and launched its cloud in 2020, and with no traceable quarter before AI they are undetermined in the incremental total. Managed inference, priced by the token and read as expanded since the anchor already shows customers running inference on rented capacity, and the services sold beside GPUs, including Weights & Biases, sit inside those dollars and are left out of the flow total; CoreWeave Omni is described and unsized until it earns revenue. Spend: capital expenditures, traced and shown and left out of totals, with depreciation, data center leases, power and interest on debt counted in their place.

Q1 2026. Revenue was $2.08bn, against $982mn a year earlier; Customer A took 45% and Customer B 20%, the top two 65%. The 10-Q names OpenAI, a private company, with commitments of up to $11.90bn and $6.50bn, and Microsoft and Meta, which committed up to $21.00bn in March. Anthropic signed. The CFO said non-investment-grade AI natives and foundation labs were less than 30% of backlog, a ceiling of $29.82bn, a share of backlog that does not limit the quarter’s revenue, while the CEO set financial services, which he says are not AI labs, near $10.00bn of backlog. Cash paid for property and equipment was $7.70bn, management capital expenditures $6.80bn; depreciation was $1.15bn, operating lease cost $388mn and interest on debt $427mn. NVIDIA bought $2.00bn of stock and equipment makers financed $4.70bn outstanding.

Q2 2026. Revenue was $2.58bn, against $1.21bn, with about 93% of the increase from existing customers. The largest customer took 36%, against 71% for the largest a year earlier. Backlog was $104.00bn, before more than $25.00bn of commitments added early in Q3, and remaining performance obligations $103.70bn; neither is revenue. Prices rose about 25% across SKUs in July. Cash paid for property and equipment was $6.42bn, management capital expenditures $9.40bn, and guidance for 2026 rose to $35.00bn to $39.00bn; depreciation was $1.39bn, lease cost $500mn, interest on debt $558mn, and debt principal reached $35.55bn. Jane Street, a customer with about $6.00bn committed, invested $1.00bn.

Sizes. Every size is the ledger’s inference. The payer split rests on Customer A being Microsoft, which the 2025 10-K pins by arithmetic: Microsoft took 67% of 2025 revenue and no other customer reached 10%, a floor of $513mn, while Customer A took $707mn and $861mn in the prior-year quarters of the two 10-Q tables. Customer B is OpenAI or Meta and is not resolved; only a letter a filing pins may set a share, so the ranges span both readings and the points weight them equally. The fleet costs are reported lines times a share near all of it: depreciation $1.25bn to $1.39bn, leases $450mn to $500mn, interest $474mn to $558mn. Power is solved from the year-over-year increase the filing gives, $87mn, and an assumed growth ratio for the line that is not tied to active power, since the increase stayed flat while active power rose to 1,500 megawatts: $131mn to $351mn. Run rates are converted to a quarter and marked inferred: managed inference $2.5mn to $17mn, the services beside GPUs $85mn to $150mn.

What the roster expected and what the sources show. The roster note called CoreWeave the clearest filed view of the investor-funded half of the buildout, since OpenAI and Anthropic do not file. The sources support part of that. They name the labs, describe them as private and possibly highly leveraged, bound their backlog once, and show the company issuing its own stock to OpenAI with its first lab contract ($350mn), which the filings book as a contract incentive, consideration paid to the customer, issued under the contract’s terms for no proceeds. That is a price term, not a stake in the lab, so the lab channel’s funding is read as investor capital. They do not give the labs’ share of revenue, and the CEO’s Q2 description of customers is of buyers monetizing their products, not of how they fund. The clearer filed view is the other side: a build paid for with debt secured on customer contracts, vendor financing and equity from a supplier and a customer. Every channel sits at the compute layer, the largest among them the hyperscaler capacity revenue and the fleet’s depreciation, and cohort totals keep that layer apart from end-use and hardware. A capacity seller’s revenue is the buildout’s money moving, and the incremental total here is the lab share alone. The same dollars appear as spend on the MSFT, META and IBM ledgers, and the capital spending reaches revenue on the NVDA ledger; totals across companies are not netted.

Readings support-agentforce-deflection in Q1, Q2 and Q3: were inferred, sized at the former estimates crm-2026-cq1-f60, crm-2026-cq2-f68 and crm-2026-cq3-f65 (inquiries or conversations times an assumed displaced share times a benchmark cost per contact); now unsized, because inquiries handled, conversations and a resolution share by count are counts of AI work, and Q1 had no measure of its own.

Readings internal-ai-efficiency in Q1, Q2 and Q3: were inferred, sized at the former estimates crm-2026-cq1-f62, crm-2026-cq2-f70 and crm-2026-cq3-f67 (annualized hours saved times an hourly cost and a realized share); now unsized, because hours saved are a count. Q2 also gives a company-wide productivity rate for Slackbot, which is quoted in metrics; no new estimate is built on it in this correction.

Readings slackbot-slack-upgrades in Q1, Q2 and Q3: were inferred, sized at the former estimates crm-2026-cq1-f56, crm-2026-cq2-f62 and crm-2026-cq3-f56 (users times an assumed monthly uplift, Q1 and Q2 borrowing the Q3 user count); now unsized, because active users of a feature included in paid editions are a count, not seats sold at a seat price, and the upgrade multiple has no level. Q3 state was quantified, now directional. Q3 overlapped the Agentforce channel, so only Q1 and Q2 moved a flow total.

Readings ai-company-customers in Q1 and Q2: were inferred, sized at the former estimates crm-2026-cq1-f55 and crm-2026-cq2-f61 (a count of AI companies times an assumed spend); now unsized, Q1 directional on the share of a list by count and Q2 described. Q3, which states a spend growth rate, keeps its size.

Readings ai-pull-through-core-subscriptions and own-sales-agents-pipeline in Q1: were inferred, sized at the former estimates crm-2026-cq1-f54 and crm-2026-cq1-f61; now directional and unsized, because a share of large deals by count and a weekly lead count are counts. Q2 and Q3, with spend and order-value multiples and a stated pipeline, keep their sizes.

Readings model-inference-cost in Q1 and Q3: were inferred, sized at the former estimates crm-2026-cq1-f58 (a share of a cumulative token count with no start date) and crm-2026-cq3-f62 (work units converted to tokens at an assumed ratio); now unsized, because a cumulative count with no period sizes nothing and Q3 states no token volume, while its filing names generative AI only beside hosting. Q2 keeps its size: the volume is the difference between two dated cumulative totals, a usage volume with a period, at public list prices.

Reading headless-platform-access in Q3: was inferred, sized at the former estimate crm-2026-cq3-f55 (an assumed share of Agentforce recurring revenue); now directional and unsized, because the call multiple is a count and the CFO says headless monetization is still being worked through. It overlapped the Agentforce channel, so no flow total moves.

Readings seat-attrition-pressure in Q1, Q2 and Q3: were inferred, sized at the former estimates crm-2026-cq1-f63, crm-2026-cq2-f71 and crm-2026-cq3-f68 (a judgment AI share of attrition); now bounded and unsized, because each quarter management names the thesis and denies the effect, so the money has not started.

Readings anthropic-stake-gains in Q1 and Q2: were inferred, sized at the former estimates crm-2026-cq1-f57 and crm-2026-cq2-f64; now described and unsized, with the gains kept as context, because those quarters' filings report a gain on a privately held investment they do not name, and the Alphabet rule makes a gain on an unnamed private company context. Non-operating, so no flow total moves.

Reading model-inference-cost in Q1 and Q2: state was quantified, now directional, because tokens processed, 19,000,000,000,000 and then 28,600,000,000,000, are a count of AI work with no dollar level. Q3 was already directional.

Reading support-agentforce-deflection in Q2 and Q3: state was quantified, now directional, because inquiries handled, 4,000,000, and the share of conversations resolved, 64%, are counts and shares of AI work that the company applies to no reported line. The efficiency motive stays: a share of contacts resolved by an implemented tool is the measured operating result that tell names.

Reading internal-ai-efficiency in Q2 and Q3: state was quantified, now directional, because annualized hours saved, 3,800,000 and 8,100,000, are counts of AI work, and the company-wide productivity rate is applied to no reported line.

Reading own-sales-agents-pipeline in Q1: state was quantified, now directional, because leads qualified in one week, 50,000, are a count of AI work. Q2 stays quantified on the stated pipeline, $42mn.

Readings informatica-cloud-arr-in-ai-metric (all quarters), ai-company-customers in Q1 and support-agentforce-deflection in Q1: motive was unknown, now exploratory, because the sources are silent on why the money moves rather than contradicting each other, and no measure or line movement is attached.

Reading anthropic-stake-gains (all quarters): motive was unknown, now exploratory, because a non-cash mark meets none of the operating tells and nothing in the sources contradicts; it stays non-operating and out of totals.

Channel model-inference-cost: counterparty was ai-lab, now mixed: third-party labs paid by the token and the cloud and data center capacity behind the company's own models.

Channel seat-attrition-pressure: counterparty was ai-lab, now mixed: frontier-model labs, AI-native vendors and customers' own AI builds.

Channel own-sales-agents-pipeline: domain was distribution, now sales-force, because agents working leads its sellers would not have reached is seller productivity.

Not changed: the fiscal-year Anthropic gain, $1.20bn, keeps its allocation to Q1 by the quarter's share of reported strategic gains, because that phasing rests on reported lines rather than an assumption of the ledger's own.

Costco enters with these channels: sales from members arriving through AI search and large language models (new money), pharmacy sales kept by AI tools that improve in-stock positions (replenishment work the anchor already shows, read as renamed and left out of the incremental total) and, from the second quarter, the AI model and tooling bill (new money). The anchor 10-K mentions artificial intelligence only as evolving regulation and as competitors' faster adoption, and the first call of fiscal 2024 does not mention it. No AI capital spending, savings channel or toll is registered: the CEO calls the technology spend capital light, the productivity the company measures is credited to pre-scanning, pay stations and the mobile wallet, and no source names business lost to AI.

Q2 FY2026, the twelve weeks ended 2026-02-15. The CEO said the company is working closely with the leading AI companies so that its prices are visible when members shop with AI tools, and the CFO said new AI tools were added to improve pharmacy in-stock positions; neither came with a measure. Asked about AI, the CFO said the company applies it only where it makes Costco better at its core business, named employee productivity as one such place and called it early days. Recommendation carousels drove over $470mn of e-commerce sales, without AI named.

Q3 FY2026, the twelve weeks ended 2026-05-10. Steps: AI search moved from described to directional, when the CFO said its traffic is still low in volume, grew at a triple-digit rate and converts at the highest rate of any source, and that the company now uses AI to enhance its product pages for large language models; the AI bill opened, read described, when the CEO, referring back to the CFO's e-commerce remarks, said there is a cost to that AI, offset by greater sales (he does not separate the recommendation carousels from the AI, so they stay a confound, uncounted); the pharmacy tools moved to not mentioned. AI disclosure did not thin or widen.

The ledger's own sizes in the latest quarter: sales from AI search $4.8mn to $39mn, a peer reference class applied to an e-commerce share of net sales of $69.15bn, and the AI bill $619k to $15mn, a small share of a technology spend the company does not report inside selling, general and administrative expenses of $6.19bn. In the first quarter the pharmacy tools were put at $1.2mn to $36mn. The roster note that Costco is a slow, conservative adopter is consistent with the sources but not tested by them: the company says little and sizes nothing, and the quiet is the reading.

Reading internal-ai-productivity in Q1: was described and inferred, sized at the former estimate cnxc-2026-cq1-f27, whose point leaned on the following quarter's figure; now not mentioned and unsized, because the only claim names internal efficiencies with no AI term and the cost program beside it is not attributed to AI (claims c7, c11 and c12 no longer carry this channel).

Readings ai-displaced-services-revenue in Q1 and Q2: were inferred, sized at the former estimates cnxc-2026-cq1-f26 and cnxc-2026-cq2-f33; now unsized (bounded in Q1, described in Q2), because in Q1 the CFO names a little automation inside the volume half of the technology vertical's decline beside lighter client volumes, with no split and the word automation rather than AI, and Q2 gives no split and names offshoring and dropped customer segments; the Q2 size carried Q1's split forward. The vertical's decline is quoted as the ceiling.

Reading ai-services-pull-through in Q1: was directional and inferred, sized at the former estimate cnxc-2026-cq1-f25; now directional and unsized, because deal wins are credited to the iX AI products beside third-party technology partners and domain expertise, and wins with technology are not wins with AI alone. Q2 stays sized.

Checked and not changed: internal productivity in Q2 stays sized, since the CEO credits the reduction in non-billable headcount to the company's own AI tools.

Reading internal-ai-productivity in Q1: motive was unknown, now exploratory, because the sources agree that efficiencies are a lever and do not contradict each other; no measure or line moves because of AI. The Q2 efficiency reading, on revenue per non-billable headcount, 14%, is unchanged.

Channel ai-build-investment: counterparty was internal, now mixed, because the spend pays the company's own forward-deployed engineers and experts and the outside model and compute vendors behind iX Hello.

Comcast enters with a single channel: broadband and wireless sales and retention steered by AI models, which the head of Connectivity & Platforms described on the Q1 2026 call as one example among several changes in the broadband pivot. It is tagged relabelled, since models that score customers for acquisition and retention predate large language models and no line is credited to them, and it carries no size: the CFO credits most of the quarter's improvement in broadband losses to the Legendary February offers (claim c3), so nothing separates AI's part. At Verizon the same door is read as acquisition and retention spend lowered by AI micro-segmentation, also unsized; Comcast speaks of outcomes rather than cost, so it is read here as revenue.

Q1 2026. Broadband losses improved by 117,000 to 65,000; domestic broadband revenue was $6.34bn against $6.68bn a year earlier, which the 10-Q explains by lower rates and fewer customers. Connectivity & Platforms customer service cost was $676mn against $682mn, and the call describes better unassisted channels without naming AI. The release and the 10-Q contain no AI passage.

Q2 2026. Steps: AI disclosure widened on the call, and the AI models channel moved from described to not mentioned. The new passages are about the network, not money: the chairman says AI will demand more data and bandwidth and ties upstream traffic growth to AI queries, and the CFO calls an active network an advantage in an AI-driven world. Domestic broadband revenue was $6.28bn against $6.65bn, and no revenue or cost is credited to the traffic, so no channel is registered for it. Comcast Business revenue included a non-recurring fiber lease renewal the CFO does not tie to AI, and no source names AI or data-centre demand at Comcast Business or AI in NBCUniversal's advertising. The call's Q&A was questions collected in advance and read out by investor relations; only the answers are quoted.

No channel carries a size, management's or the ledger's. Beside Verizon, which names voice agents in care, an AI coding tool, network models and fiber sold for AI infrastructure, Comcast is the quieter telecom: customer service appears in its sources as a weakness addressed with investment, and AI is not named as the means.

Reading ai-operating-efficiency in Q1 and Q2: was inferred, sized at the former estimates chgg-2026-cq1-f24 and chgg-2026-cq2-f25 (an assumed AI share of non-GAAP operating expenses at the prior-year rate); now unsized, because the CFO names expense discipline beside AI (in Q1, AI only as an opportunity still to be found), the CEO the restructuring to become AI-first, and the 10-Q restructuring, and nothing separates AI's part. Non-GAAP operating expenses are quoted as the ceiling. Q1 stays described and Q2 directional on the words. This moves the savings total in both quarters.

Reading study-ai-retention in Q1 and Q2: was inferred, sized at the former estimates chgg-2026-cq1-f22 and chgg-2026-cq2-f23 (an assumed share of Academic Services revenue); now described and unsized, because in Q1 the CEO gives the AI capabilities as only some of the reason the decline is slowing, and in Q2 no source names AI as a cause of retention. Academic Services revenue is quoted as the ceiling. This moves the revenue total in both quarters.

Assessments of Q1 and Q2: the sentences that sized the operating saving and the retained revenue are rewritten.

Reading ai-content-production in 2026-CQ1: was left out of totals as overlapping ai-operating-efficiency, now counts in totals, because that channel is unsized in the quarter under the rules above and a component sized on its own evidence counts when its umbrella does not (overlap rule, 2026-10-06).

Reading ai-content-production in 2026-CQ2: was left out of totals as overlapping ai-operating-efficiency, now counts in totals, because that channel is unsized in the quarter under the rules above and a component sized on its own evidence counts when its umbrella does not (overlap rule, 2026-10-06).

Steps. In Q2 2026 the call carried more AI passages than Q1, which evaluate reads as disclosure widening; Q1’s single passage and one of Q2’s report a state governor’s view, and the other Q2 passage is the CEO naming customers’ AI models. No channel changed state, strength or motive.

Constellation enters with Q1 and Q2 2026 and channels all at the facilities layer: power from the restarted Crane plant sold to Microsoft under a long-term PPA, a cloud-provider counterparty; powered land and co-located power sold to CyrusOne, a data-centre developer whose tenants are not named, beside gas plants in Texas; and the capital spending on the restart and on co-location infrastructure, capital and out of totals. Each is expanded by the ledger’s test: selling a plant’s output and building generation predate LLMs, and the anchor shows the Microsoft agreement and co-location of data centres as strategy. With no traced quarter before AI, they are undetermined in the incremental total.

Q1 2026. Calpine, acquired on 2026-01-07 for $21.80bn, brought $2.40bn of segment revenue for most of the quarter, so operating revenues rose 63.8% on the year, while the other reportable segments’ revenues changed 7.5% (most of the rest is unrealized gains and losses on derivatives, which the 10-Q does not allocate to segments); every comparison of lines against 2025 or the anchor carries that. The company signed 380 MW with CyrusOne at Freestone with exclusivity for a further 380 MW, beside the 400 MW Calpine had signed at Thad Hill before the close, which are excluded from any future size of the channel under the acquisitions rule; the Freestone substation is due in Q4 2026. The presentation bounds further powered-land deals at an EPS impact of 0.20 to 0.50 dollars a share per 1,000 MW, kept as a metric, not a size. The CEO put hyperscaler spending for 2026 nearly 75% above 2025 and named compute, not AI, as the demand. Capital expenditures were $1.27bn, against $806mn.

Q2 2026. Crane won its fuel licence amendment and the transfer of interconnection rights, and the company now expects it back in service in 2027, earlier than the anchor’s 2028. The 10-Q names the restart and co-location infrastructure, beside Calpine, among the causes of higher capital expenditures, $1.25bn in the quarter against $767mn, with no amounts. The company signed 920 MW of nuclear PPAs with investment grade customers, starting in 2029 through 2032, one with Walmart for 176 MW; the CEO declined to say whether any is a hyperscaler and the sources do not say data centres, so they are context and not a channel.

Sizes. No channel carries a size. The revenue channels stay unsized because contracts that start after the quarter are not revenue in it. The capital channel stays unsized because the company names AI only beside the general expansion of data centres and nothing separates AI’s part; the ceiling is quoted instead: the anchor’s estimate of the whole restart, $1.60bn, inside $1.70bn of growth capital planned for 2025 and 2026, and the quarter’s capital expenditures.

What the roster expected and what the sources show. The roster note expected power purchase agreements with hyperscalers tied to named plants. The sources show one: Microsoft at Crane, signed in 2024 and earning nothing until the restart. They do not show a Meta agreement or any agreement at Clinton, whose only appearance is in the Illinois zero-emission credit program, and the counterparties of the quarter’s new nuclear agreements are not named. The company talks about data centres, large loads and the data economy far more than about AI. Microsoft is on this ledger as MSFT, where the same money, once delivered, is part of the cost of running its data centres. Totals across companies are not netted, and cohort totals keep the facilities layer apart from compute, hardware and end use.

Steps. None: the one channel is described in both quarters, and its state, strength and motive did not change between them.

Caterpillar enters with Q1 and Q2 2026 and one channel, at the facilities layer: generator sets, turbines and services sold for data-centre prime and backup power, a revenue channel inside the Power Generation application. The payers are data-centre developers and operators and the cloud companies, reached mostly through dealers and none named, so the counterparty reads mixed. The channel is expanded by the ledger's test: the anchor already shows backup gensets and prime power turbines for data centres, credited to cloud computing and generative AI, and during coverage new capacity was built, restarted and converted for that demand. With no traced quarter before AI, it is undetermined in the incremental total.

Q1 2026. Power Generation sales were $2.82bn, up 41.1%, and sales to users grew 48%, credited primarily to data-centre applications. The company raised its large reciprocating engine capacity target to nearly 3.00 times 2024 levels, with most of the added spending in 2027 through 2029. The large engine backlog has grown more than 3.50 times since January 2024, and a prime power agreement of up to 2.1 GW delivers over five years; backlog and agreements are described, not sized. Asked what drove the capacity decision, the CEO said the demand is still a lot of cloud and not just AI.

Q2 2026. Power Generation sales were $3.10bn, up 28.7%, and sales to users grew 72%. Asked about the durability of AI and data-centre demand, the CEO said no customer is slowing down, and that the capacity was planned on data centres as a big driver beside aftermarket and oil and gas. A medium-speed gas engine platform of about 1.5 GW restarts without significant investment, shipping from the fourth quarter. The firm backlog reached $72.10bn across segments.

Sizes. No channel carries a size. The data-centre power channel stays unsized because the company names AI only beside cloud computing and nothing separates AI's part; the ceiling is Power Generation sales, of which data centres are themselves an unstated part.

What is context and not a channel. The capacity build for large engines and turbines is not registered as AI capital: no capital claim names AI, the CEO credits the build to data-centre capital spending and to the other industries the capacity serves, and the ledger registers capital only where it builds AI capacity itself. Skycatch, the spatial-data company Resource Industries acquired in July, pairs its data with AI capabilities, but under the acquisitions rule it is an acquisition confound, since it is not itself an AI company; a later filing that describes its product or revenue as AI may change that.

What the roster expected and what the sources show. The roster expected data-centre demand for power-generation equipment and AI in the company's products and operations, autonomy and digital. The sources bear out the first, as a joint cause with cloud computing. They do not bear out the second: the company never calls its autonomy or digital work AI, it credits higher Resource Industries SG&A and R&D to compensation and technology including autonomy, and the CEO's remark on factory automation and machine autonomy is an illustration of data use with no tool or measure; none of these is a channel. Data-centre construction also lifts Construction Industries demand, named without AI. Totals across companies are not netted, and cohort totals keep the facilities layer apart from compute, hardware and end use.

Citigroup enters with Q1 and Q2 2026 and these channels: internal build investment and process automation in operations, and, from Q2, staff productivity from AI tools, growth from products brought to market faster, and capital raising for the AI buildout. Build investment, process automation and staff productivity are existing work into which AI is put, read as expanded. Faster product launches and buildout financing are existing activities the anchor already shows, read as relabelled because no movement is attributed to AI. Citi names no outside model or tool bill and states no parts of its AI spending, so any tool and model cost is inside build investment and is not split out by judgment.

Q1 2026. The CEO said Citi is deploying AI at scale and set out an approach by area, of which only process simplification was described as work under way, with AI and automation together; falling transformation expenses were said to create capacity for investment in AI. Asked about headcount and AI, the CFO named automation first and AI as a further opportunity. Direct staff fell −2.2% over the year with severance of about $500mn, and the 10-Q credited a fall in technology expense to fewer contractors from productivity savings, none of it attributed to AI. The 10-Q also named AI risk to vulnerable sectors among the risks the qualitative adjustment of its credit loss allowance covers, beside macroeconomic and geopolitical risk, with no amount or reserve movement attributed to it; the ledger keeps that passage as context and registers no channel for it.

Q2 2026, the steps. Staff productivity opened on the CEO’s statement that nearly 90% of staff use the AI tools, read directional because that share of users is its only measure; growth from faster product launches opened on her naming Payments Express and the Citi Wealth Advisor Insights platform; buildout financing opened on her statement that AI dominates client conversations, beside tech, data centre, energy and defence capital spending, with Investment Banking revenues up 44.3%. The CFO counted more than 100 processes mapped for further automation, still credited to technology and AI automation together. Direct staff fell −4.8% over the year; the CFO credits it to past investments and productivity efforts.

The sizes. Only internal build investment is sized, at $17mn to $235mn, a share of revenue taken from other buyers’ readings and centred on Bank of America’s, because Citi gives no AI budget; it is the ledger’s estimate. Staff productivity stays unsized because a share of staff by count separates no dollars. Faster product launches, process automation and buildout financing stay unsized because AI is named beside other causes, with their ceilings quoted. The roster expected AI tools in use by most employees and a transformation program: the tools are confirmed by the Q2 count, and the transformation, which the filings credit with lower expense, is a confound throughout and is never attributed to AI.

Reading ai-platform-marketing-fees in Q1: was inferred, sized at the former estimate bkng-2026-cq1-f43; now described and unsized, because the CEO speaks of the possibility that AI platforms become performance marketing platforms, the 10-Q's list of performance marketing channels has no AI platform, and the OpenAI cost-per-click test is first disclosed in Q2 with no start date. Q2 stays sized.

Reading organic-search-erosion in Q1: was directional and inferred, sized at the former estimate bkng-2026-cq1-f44; now directional and shape, unsized, because no Q1 source names AI as the cause of the decline in unpaid search traffic (the call reports continued SEO declines and the 10-Q cites evolving search engine dynamics). The CEO's remark on travelers going to a large language model first now sits on llm-platform-demand only.

Reading organic-search-erosion in Q2: was directional and inferred, sized at the former estimate bkng-2026-cq2-f45; now directional and shape, unsized, because the CEO names Google's display changes beside the AI overview and says the overview probably did it, the CFO sees the pressure across consumer internet, and the line also moved on paid mix and chosen investment. The step from Q1 comes from the wording, not from a change at the company.

Reading ai-personalization-conversion in Q2: was inferred, sized at the former estimate bkng-2026-cq2-f50; now described and unsized, because asked for evidence the CEO answers only in the future tense and the 10-Q line is a strategy aim.

Checked and not changed: traveler support automation in Q2 stays sized, since the CEO says cost per interaction is falling because of AI; the CFO's bound on room nights from large language models stays, and so do the reference classes that borrow it.

Reading ai-personalization-conversion in Q1: motive was unknown, now exploratory, because the CEO names AI as raising conversion with no measured movement, price or attach rate, and the call and the 10-Q do not contradict each other.

Reading back-office-automation in Q2: motive was unknown, now exploratory, because the CEO lists functions using AI with no cost effect and the filing explains the lines without AI.

Channel ai-vendor-bill: counterparty was ai-lab, now mixed: model and token providers, AI software licensors and cloud providers.

Channel traveler-support-automation: counterparty was enterprise, now mixed: the third-party customer service providers that carry most of the cost, and the company's own service staff.

Channel back-office-automation: counterparty was internal, now mixed: the company's own corporate-function staff and the outside vendors those functions pay.

Readings erica-virtual-assistant in Q1 and Q2: were described and inferred, sized at the former estimates bac-2026-cq1-f38 and bac-2026-cq2-f53; now directional and unsized, because the only measures are counts (active users, interactions this quarter and a year earlier, CashPro chat interactions) and no cost, deflection rate or cost per contact is given.

Reading employee-productivity-tools in Q1: was inferred, sized at the former estimate bac-2026-cq1-f37; now described and unsized, because every Q1 statement tying a cost effect to AI names it beside digitization, process re-engineering or technology, and the only AI-alone measure is a count of employees with access. Compensation and benefits is quoted as the ceiling.

Readings employee-productivity-tools and call-center-agent-assist in Q2: were described and inferred, sized at the former estimates bac-2026-cq2-f50 and bac-2026-cq2-f52; now directional and unsized, because the only measures are counts (active users, prompts a day, templates and training courses; call-center agents using the tool).

Reading ai-sales-effectiveness in Q1: was inferred, sized at the former estimate bac-2026-cq1-f39; now described and unsized, because the CFO names AI beside technology as improving sales effectiveness and the CEO's market share remark is in the future tense. Wealth and Global Banking revenue is quoted as the ceiling, as in Q2.

Readings ai-coding-tools-bill and engineering-productivity in Q1: were not mentioned but sized on the 2025-CQ4 reference call (bac-2026-cq1-f41, bac-2026-cq1-f42); now not mentioned, inscrutable and unsized, because a quarter silent on a channel carries no size. Both are first described in Q2, where they stay sized.

Checked and not changed: the model and tool bill and the coding tools bill in Q2 stay sized, since seats are the billing unit and seats times a public seat price may size a spend channel; operations in Q2 and buildout financing stay as the first rules pass left them.

Figures bac-2026-cq1-f35 and bac-2026-cq2-f48 (internal build investment): was an even quarter of the 2026 technology initiatives budget, now the fixed annual-plan band, because a plan for a year converts only through that band; the point does not change and the range widens.

Reading operations-process-automation in Q1: was directional and inferred, sized at the former estimate bac-2026-cq1-f36; now described and unsized, because the CEO credits the reduction in manual work to digitization, the application of AI and process re-engineering together. Q2 stays sized, since the CFO credits AI-enabled tools alone.

Reading ai-sales-effectiveness in Q2: was inferred, sized at the former estimate bac-2026-cq2-f54; now described and unsized, because the CFO credits the benefit to advisor productivity, digital engagement and AI-enabled tools together, and the slide gives counts of users, not a lift.

Figure bac-2026-cq2-f60 (buildout financing): was a share of the issuance fee rise plus a share of Business Lending revenue, now the issuance share only, $9.7mn to $121mn, because the CFO says commercial loan growth is broader than the AI theme and gives no AI share.

Readings ai-sales-effectiveness (both quarters) and ai-buildout-financing: motive was unknown, now exploratory, because the sources do not contradict each other; AI is named with no measure of its own.

Channel ai-sales-effectiveness: domain was distribution, now sales-force; counterparty was unknown, now mixed: wealth households and corporations.

Checked and not changed: developer productivity reads inferred, not implied, on the slide's productivity floor, 20%, since its dollar effect rests on the ledger's own captured share.

Bank of America enters with Q1 and Q2 2026 and these channels: the bill for model access and general-purpose AI tools, the bill for coding assistants, internal build investment, operations and manual processing, knowledge-work productivity, developer productivity, call-center agent assistance, client self-service through Erica, advisor and banker sales effectiveness, cyber defense against AI-enabled attacks, and, from Q2, capital raising and lending for the AI buildout. The vendor bills are new money by the ledger’s test. Build investment, the cost channels and sales effectiveness are existing work into which LLM tools were put, read as expanded. Erica, cyber defense and buildout financing are existing activities the anchor already shows, read as relabelled because no movement is attributed to AI.

Q1 2026. The CEO named the application of AI beside digitization and process re-engineering as reducing manual work and unit cost, said all 200,000 employees have access to AI with 90 installations working, and put the headcount reduction since year-end at about 1,070, through attrition. The CFO said the company is expanding its use of technology and AI for operational efficiency and sales effectiveness, and the CEO said AI creates cybersecurity issues the company has invested heavily against. Headcount fell −598 over the year while compensation and benefits rose 4.1% on revenue-related incentives. An analyst raised AI agents moving deposits away; the CEO said there is nothing new about the ability to move deposits, so it is not a channel.

Q2 2026, the steps. Call-center agent assistance opened, on about 5,000 agents receiving automated recommendations, and so did buildout financing, on the CFO’s statement that the company leads in capital raising and financing for the AI investment and that the AI theme has helped commercial loan growth, which is broader than AI. The coding assistants bill and developer productivity are not new: the January 2026 call, stored as a reference quarter, already said AI took 30% out of the coding part of the change process and saved about 2,000 of 18,000 people writing code in 2025, so Q1 carries them unmentioned and Q2 describes them, with the slide’s 19,000 developers and productivity gain of more than 20%. Cyber risk went unmentioned. Headcount fell −2,084 over the year, and the motives on the cost channels stay exploratory: the CFO calls AI something for the future, said in January that AI let the company move people from operational support to client-facing roles with headcount flat, and the CEO says coding tools mean the same money gets more code.

The ballparks. Internal build investment, $40mn to $300mn, is a quarter of the 2026 technology initiatives budget management gave on the January call (more than $4.00bn, all new code) times an AI share whose point follows management’s several hundred million for AI projects. The developer saving, $4.7mn to $160mn, rests on management’s own developer-equivalents with a low captured share; operations is $4.4mn to $99mn and knowledge-work productivity $3.8mn to $123mn. The tool bill, $9.0mn to $24mn, prices the 200,000 Copilot seats management counted in January, and the coding bill is $1.1mn to $11mn. Buildout financing, $9.7mn to $121mn, is a share of the rise in debt and equity issuance fees ($483mn) plus a small share of Business Lending revenue, and is the buildout’s money moving through a bank. Erica was flagged on the roster as an old assistant: it is read as relabelled, since the January call already renamed it the company’s AI agent and explained its interactions by proactive alerts, though its interactions rose to 200,000,000 from 174,000,000 in Q2 after a fall in Q1. The prior-year revenue is the 10-Q’s presentation, $27.44bn, restated for a 2025 change in accounting for tax-related equity investments.

American Express enters with Q1 and Q2 2026 and these channels: internal build investment, the bill for AI usage, programmer productivity, card member servicing and travel servicing assisted by AI tools, marketing campaign production, credit, risk and fraud decisions, agentic commerce, and a ChatGPT statement credit for business card members. The usage bill and agentic commerce are new money by the ledger’s test. Build investment, the cost channels and the ChatGPT credit are existing work or budgets into which AI was put, read as expanded. Credit, risk and fraud models are read as relabelled, since the CEO says AI has been used there for many years.

Q1 2026. The CEO said programmers get about 30% benefit from AI in coding and testing and that it let the company get to more work; that representatives per unit of volume have fallen because fewer customers want to call and because representatives have AI tools, for card servicing and travel; and that the ACE developer kit and Agent Purchase Protection, introduced in April, open an agentic commerce market not yet in its first inning and too early to quantify. He also listed an OpenAI ChatGPT benefit among the commercial card capabilities announced, and said the effect of the new launches would be benign. The 10-Q says the company furthered the development of its AI capabilities, and its list of forward-looking factors names spending on AI initiatives. Salaries and employee benefits rose 17.1% on compensation and incentives, with no AI named.

Q2 2026, the steps. These channels opened: the bill for AI usage, which the 10-Q’s list of forward-looking factors now names among increased technology costs, without saying the cost rose or how large it is; marketing production, which the CEO says AI streamlines; and credit, risk and fraud decisions, where the CEO says an LLM layer is being added to models in use for many years. The ChatGPT credit went live at $300 a year per eligible U.S. Business Platinum and Gold card member and gained a ballpark, so its strength moved from described to inferred. The credit is paid to those business card members against their charge for the subscription; they pay OpenAI, a private lab not on this ledger, and American Express earns discount revenue on the charge. The CEO said servicing and travel representatives have AI-powered tools, their number has not grown as the business grows, and a new servicing portal with AI embedded is meant to cut handle time; he placed agentic commerce in the preseason.

The ballparks. Build investment, $14mn to $293mn, and the usage bill, $982k to $15mn, are shares of revenue taken from the ledger’s readings of the same channels at other buyers, as JPMorgan’s are; data processing and equipment, which rose $97mn on the prior-year quarter on technology costs the 10-Q explains without AI, is kept as a ceiling on the bill. The ChatGPT credit, $578k to $26mn, rests on Commercial Services cards-in-force and judgment shares for eligibility and use, with the high end narrowed for the CEO’s benign remark; whether OpenAI funds part of the credit is not said. Marketing production is $0 to $26mn and programmer time $0 to $21mn; both ranges start at none, because the CEO describes speed in marketing and says the coding gain is not a saving. Card and travel servicing stay unsized because Q1 credited the same fall in representatives per unit of volume to fewer calls as well as AI tools, and nothing separates them; credit, risk and fraud and agentic commerce stay unsized because their money has not started. The roster note named customer care, fraud and agentic commerce, and each is in the sources, but the fraud models predate LLMs and agentic commerce has no volume yet.

Readings ai-pull-through-core-services in Q1 and Q2: were inferred, sized at the former estimates amzn-2026-cq1-f68 and amzn-2026-cq2-f69 on a judgment share of non-AI AWS revenue; now described and unsized, because AI is named as one of several drivers of core growth (a correlation, beside migrations from on-premises and memory shortages) with no rate or ranking of its own. AWS revenue is quoted as the ceiling.

Readings alexa-for-shopping-assistant, sponsored-prompts-ai-ads and health-ai-care-agent in Q1: were directional and inferred, sized at the former estimates amzn-2026-cq1-f69, amzn-2026-cq1-f70 and amzn-2026-cq1-f73; now directional and unsized, because the quarter's only measures are counts and shares by count: Rufus monthly active users and engagement, the share of shoppers who continue a conversation after a brand prompt, and virtual care visits with the majority via Health AI.

Figures amzn-2026-cq1-f75 and amzn-2026-cq2-f74 (AI depreciation): the AI share was the ledger's judgment, its point above the words-table point; it now rests on the company's bound, the trailing-year increase in purchases of property and equipment primarily reflects investments in artificial intelligence (claims c41 and c36, now cited in the readings), read through the words table for primarily. The low and high ends do not change; the points become $1.97bn to $3.55bn and $2.37bn to $4.27bn. The capital spending reading is left as it is.

Steps, all in Q2 2026. Opened: the company's own frontier model, AWS Forward Deployed Engineering (quantified on an undated amount of $1.00bn, left unsized until a source shows the team at work), and AI in fulfillment robotics, a passing mention read as expanded. AI ad tools: described to directional, exploratory to offensive, on Ads Agent's measured advertiser costs. Engineering with agentic coding: directional to described and efficiency to exploratory, the Q1 example not repeated. Third-party AI agent referrals and Health AI: to not-mentioned.

Amazon enters with two quarters and a registry that reads it as seller first and buyer second. Revenue: the AWS AI run rate as one channel with a mixed counterparty, AI labs (Anthropic and OpenAI) and enterprises, since management states no parts of it; the frontier labs, Bedrock and enterprise AI services and the agent applications are registered inside it as described, unsized readings, and the run rate is not split among them by assumed shares; core demand management attributes to customers' AI, read as relabelled; and the buyer-side revenue from Rufus and Alexa for Shopping, sponsored prompts, AI ad tools, Health AI and referrals from third-party AI agents. Spend: capital spending (capital, out of totals), AWS depreciation, the own frontier model (out of totals as overlapping the depreciation) and Forward Deployed Engineering. Interest on the new notes is noted with capital spending and not read as a channel, since the notes were issued for general corporate purposes and no source names AI as their reason. Savings: engineering with agentic coding and AI in fulfillment robotics. Non-operating: the Anthropic marks. The run rate is expanded by the tie-break, its lab payers new and its machine learning services expanded with no stated split; the lab, application and agent-referral money is new by the ledger's test; Bedrock, the shopping and ad features, the depreciation, the spend channels and fulfillment robotics are expanded; core pull-through is relabelled. Layers: the run rate, core pull-through, capital spending and the AWS depreciation are compute and carry most of the sized dollars; the shopping assistant, sponsored prompts, ad tools, Health AI, the agent applications, engineering and fulfillment savings are end-use; the Anthropic marks are compute, on an investee the sources describe as a lab that commits to AWS capacity.

Q1 2026. The CEO gave AWS's AI revenue run rate as over $15.00bn, growing year over year, inside AWS revenue of $37.59bn, up 28.4%; the 10-Q explains AWS growth by customer usage without naming AI. OpenAI expanded its AWS commitment of $38.00bn by $100.00bn and the company invested $15.00bn in OpenAI and committed to $35.00bn more; after the quarter Anthropic announced a deal of over $100.00bn and was offered a facility of up to $20.00bn that opens as AWS delivers compute. Cash capital spending was $43.20bn, primarily for AWS and generative AI by the CFO's words, funded by operating cash flow of $26.03bn and long-term debt of $53.44bn, the notes issued for general corporate purposes. Pre-tax gains of $16.80bn on Anthropic sat in non-operating income. On the buyer side the CEO gave usage counts for Rufus and one engineering example, a rebuild by 5 people with agentic coding tools; nothing in the stored sources ties a workforce change to AI.

Q2 2026. The AI run rate passed $25.00bn and AWS revenue was $42.23bn, up 36.8%; the backlog reached $496.00bn after Anthropic expanded its commitment by more than $100.00bn. Cash capital spending was $53.10bn and the 2026 expectation rose to about $220.00bn, on memory prices. Interest expense rose to $1.30bn on new notes, noted with capital spending and not read as an AI channel. The company invested $5.00bn and $5.00bn in Anthropic, the second under the capacity-linked facility, and $13.70bn in OpenAI, with $21.30bn more after the quarter; the Anthropic preferred stock was marked up $50.50bn. Buyer-side measures compare users with non-users: Alexa for Shopping users spend over 40% more per order, shoppers who click a sponsored prompt convert 48% more often, advertisers on Ads Agent pay 6% less per acquisition. None is a lift, and the ballparks keep the Q1 shares.

Sizes. Every operating size is the ledger's inference. The Q2 AI revenue, $5.00bn to $9.38bn, is a quarter of the stated run rate scaled for what was in force, and it counts in totals as one compute-layer channel. Core pull-through is $721mn to $3.60bn. Capital spending for AI is $24.24bn to $38.38bn, capped by AWS's share of net additions to property and equipment, capital and left out of totals; AWS depreciation above its 2024 level, times an AI share, is $2.37bn to $4.27bn; the own frontier model $337mn to $2.02bn, left out of totals as overlapping the depreciation. The buyer-side ballparks are small against the store: the shopping assistant $35mn to $1.41bn, sponsored prompts $20mn to $297mn, ad tools $40mn to $594mn, engineering $13mn to $405mn.

What the two quarters show. Amazon shows the buildout's money moving in a circle: the labs that commit to its capacity are funded in part by its own preferred stock purchases and a facility that opens as capacity is delivered, so the run rate's funding reads mixed (no source states a payment drawn from the facility, which would be the vendor-financed case), the marks on those stakes are larger than any operating size here, and the capacity is paid for from operating cash flow and new debt. The same lab money appears at Microsoft and the accelerators at Nvidia; nothing is netted across ledgers. As a buyer Amazon names AI in shopping, advertising, health and engineering, gives rates and counts rather than dollars, and credits its fulfillment and payroll movements to other causes. Amazon's next report is expected around 2026-10-29.

Reading data-projects-from-ai in Q1 and Q2: was bounded and inferred, sized at the former estimates acn-2026-cq1-f73 and acn-2026-cq2-f75 (the attach floor applied to the advanced AI estimate with an assumed relative size and share added); now directional and unsized, because at least one in two advanced AI projects leading to a data project is a floor on a share by count, with no number of projects and no dollars. This moves the revenue total in both quarters.

Reading ai-company-acquisitions in Q1: was described and inferred, sized at the former estimate acn-2026-cq1-f81 (an allowance for deal costs on a press-reported price); now unsized, because Faculty closed the week before the call, after the quarter ended, and the filing separates no deal costs, so the money had not started as far as the sources show. Q2, with the deal closed, keeps its size.

Reading ai-in-corporate-functions in Q2: was not-mentioned and inferred, sized at the former estimate acn-2026-cq2-f87 (the Q1 reference-class share carried onto Q2's corporate lines); now not-mentioned and inscrutable, unsized, because a reading of a quarter whose sources are silent on the channel carries no new size. This moves the Q2 savings total.

Reading ai-in-client-delivery in Q2: was described and inferred, sized at the former estimate acn-2026-cq2-f79 (the Q1 construction carried forward); now unsized, because no Q2 source makes a statement about AI in the company's own delivery: the one passage is a client example whose savings go to the client, and a client's saving is not the company's. This moves the Q2 savings total. Q1, where the CEO says AI is applied in delivery, keeps its size.

Assessments of Q1 and Q2: the sentences that sized data projects and Faculty's deal costs, and the step that carried the corporate-function size, are rewritten.

Kept as they were, after review: training, priced from stated training hours, which are the quantity of the spend itself and not a count of users; and the AI-enabler businesses, which provide services, not power, space or buildings, and stay on the end-use layer.

Reading ai-enabler-demand in Q2: was inferred, sized at the former estimate acn-2026-cq2-f78; now described and unsized, because the CEO names AI beside geopolitical risk as the driver of demand for operational technology security and the stated cybersecurity figure, $10.00bn, is the whole business. Q1 stays sized at $0 to $318mn, where AI is named alone as a catalyst.

Figure acn-2026-cq2-f78: was a quarter of the fiscal 2025 cybersecurity amount taken as an even split of the year, now the fixed annual-plan band on the quoted amount, because a completed year's actual converts through that band; the figure is kept for the record and no longer sizes the reading.

Reading digital-core-ai-readiness in Q1 and Q2: was inferred, sized at the former estimates acn-2026-cq1-f74 and acn-2026-cq2-f76; now described and unsized, because the filing gives becoming AI-ready as one part of transformations that also cover cloud, platforms, security and data. The reading was already out of totals as an overlap with data projects.

Readings ai-enabler-demand and digital-core-ai-readiness in Q1 and Q2: motive was unknown, now exploratory, because AI is named as the cause with no measure and the sources do not contradict each other.

Readings ai-tools-and-model-bill in Q1 and Q2: motive was unknown, now narrative-defensive, because the company evaluates staff on their use of AI tools, a usage mandate, and states no cost; Q2 carries it.

Channel ai-in-client-delivery: counterparty was internal, now mixed: the company's own delivery staff and the subcontractors in cost of services.

Channel ai-tools-and-model-bill: counterparty was unknown, now mixed: model developers, software vendors and cloud providers.

Not changed: platform and asset build stays an even quarter of fiscal 2024 research and development, acn-anchor-f11, because a reported annual line phased evenly over the year is not an amount stated for another period.

Reading host-ai-tools in Q1: was inferred, sized at the former estimate abnb-2026-cq1-f32; now described and unsized, because every Q1 claim is in the future or in development (features held for the May release, engineering capacity to accelerate the host API work).

Reading engineering-productivity in Q1: was directional and inferred, sized at the former estimate abnb-2026-cq1-f31; now directional and shape, unsized, because the only measure is a share of code written by AI, a count of work output and not a rate of a work measure. Q2 gives work rates and stays sized.

Readings ai-coding-tools-bill and engineering-productivity in Q1: state was quantified, now directional, because the share of code written by AI, 60%, is a count of AI work with no dollar level and no rate applied to a reported line.

Reading engineering-productivity in Q2: state was quantified, now directional, because concept-to-launch time, 60%, and features shipped, 80%, measure output, not dollars.

Reading llm-platform-demand in Q1: motive was unknown, now exploratory, because AI-platform traffic is named as converting well with no volume, price or line, and the sources do not contradict each other. The silent Q2 reading carries it.

Channel community-support-automation: counterparty was enterprise, now mixed: the company's own support agents and the outsourced customer support partners.

Channel host-ai-tools: counterparty was enterprise, now mixed: individual hosts and professional property managers, which the company does not split.

Channel ai-vendor-bill: counterparty was ai-lab, now mixed: model and inference providers, AI software licensors and the cloud providers behind data hosting.

Not changed: the reported support saving, $17mn, stays the Q2 size, and the narrative-defensive motives on Q1 coding and engineering stay, since they cite their own tell.

Apple enters with Q2 and Q3 of fiscal 2026 (calendar Q1 and Q2) and these channels: device revenue from Apple Intelligence and Siri AI, Macs bought to run AI models and agents locally, AI investment in research and development, compute for Apple Intelligence in its own data centers and third-party cloud, and capital spending on AI servers. The device and Mac channels are read as relabelled: the anchor already lists Apple Intelligence and the Mac line, and no source measures demand that moves because of AI. The spend channels are expanded: the anchor shows research and development and data centers, and AI work changed their size, with no traced baseline. The boundary between them: development and training sit in the research and development channel, serving user requests in the compute channel, and cash spent on the servers in the capital channel, which is traced and left out of flow totals.

Q2 FY2026. The CEO listed Apple Intelligence among the reasons customers choose the current iPhone family, beside design, performance, durability and the camera, and named AI and agentic tools as a reason Mac mini and Mac Studio demand ran above plan, beside MacBook Neo; the 10-Q explains iPhone growth by Pro models and Mac growth by laptops. The CFO called AI an important investment area funded incrementally on top of the normal roadmap, and research and development rose 33.6% against net sales growth of 16.6%. Asked about foundation models, the CEO said the collaboration with Google is going well; no money or term is given, so it is context. Google is Alphabet on this ledger, and Alphabet's exhibits name no Apple channel.

Q3 FY2026, the step. The compute channel opened: the CEO described a hybrid of third-party cloud and Apple's own data centers running Private Cloud Compute, said AI expenditures sit in cost of sales as well as operating expenses, and said it is early to know what Siri AI costs to serve; the 10-Q added a risk factor on compute from its own data center infrastructure and third-party cloud service providers as industry AI demand raises costs. The 10-Q named investments in artificial intelligence inside the infrastructure-related costs that raised research and development by $2.86bn, beside headcount. Siri AI was introduced and in beta, with no price; the CEO said heavy users may later move up an iCloud+ plan. A capital channel opened too: the CEO said the Houston facility currently assembles advanced AI servers, with no amount, so it reads described and unsized. Payments for property, plant and equipment over the first nine months fell to $6.80bn from $9.47bn; no source ties any part of them to AI.

No ballpark is built. The revenue channels and the research and development channel name AI beside other causes, with nothing separating its part, the compute channel's own data centres are the build's own cost, which carries an AI share only where the company bounds AI's part of the build, and Apple does not, while its rented cloud has no provider, volume or price; the capital channel has no amount. Memory costs and App Store softness are not read as tolls, because Apple names no AI cause for either. The roster's notes check out where the sources speak (Apple Intelligence, Private Cloud Compute, a named model partner), with limits: no covered source says what silicon the Private Cloud Compute servers use, and the partner is named only as a collaboration, not as a bill.

Walmart enters with every covered quarter of its January fiscal year and these channels: sales caused by Sparky, orders through outside AI platforms, store traffic lost to AI shopping platforms, associate labor saved by AI tools, inventory and fulfillment decisions made with AI, the AI model and partnership bill, AI inside capital spending (shown and left out of totals as capital), engineering work done with AI coding tools (registered from the quarter before coverage, where the CEO said more than 40% of new code was AI-generated or AI-assisted, and silent in every covered quarter), and, from the second quarter, AI features in advertising tools. Orders through AI platforms, the lost traffic and the model bill are new money; Sparky, associate productivity and engineering work are existing lines that AI changes; capital spending is an existing budget with no traced AI baseline; inventory decisions and the advertising tools are machine-learning activities the anchor already shows, now called AI, and stay out of the incremental total.

Q4 FY2026, the quarter ended 2026-01-31. Customers who use Sparky had an average order value 35% above non-users, and about 50% of app users had used it; the head of eCommerce called it still early. The CFO said investments in AI are incorporated in the capital spending assumptions and that the company develops AI through partnerships, naming OpenAI and Alphabet, with no amounts. The 10-K adds AI-enabled shopping platforms as a risk to store traffic and says AI-powered tools support associate productivity, while explaining the expense rate by liability claims, a PhonePe charge and depreciation. The associate count stayed at 2,100,000, the anchor's level.

Q1 FY2027, the quarter ended 2026-04-30. Steps: AI in advertising tools opened, read described, when the CFO named AI features that adjust content mix after crediting growth to marketplace sellers; orders through outside AI platforms, the traffic risk and associate productivity moved from described to not mentioned; AI in capital spending stayed described on the CEO's mention of investments in AI, ballparked at $38mn to $382mn on the quarter's capital line. Sparky usage grew from an unstated base: weekly active users up more than 100% in a quarter and units bought through it more than 4.00 times the prior quarter. The 10-Q recorded $200mn of reorganization charges to align global platforms, not attributed to AI.

Q2 FY2027, the quarter ended 2026-07-31. Steps: AI disclosure thinned on the call; associate productivity returned to described when the CEO said AI makes associates' work easier and the CFO credited wage leverage to associates' use of technology tools and to supply chain automation, without naming AI; inventory decisions, the model bill, AI in capital spending and the advertising tools moved to not mentioned. Customers using Sparky were up 70% on a year earlier.

Every size on the page is the ledger's own: a share of eCommerce sales for Sparky and for AI-platform orders, a share of the operating expense line for the model bill and for associate labor, a share of revenue lost to markdowns and spoilage for inventory decisions, and a share of the technology capital line for AI capital spending. The fourth quarter of fiscal 2026 is traced as the year less the first nine months from the prior 10-Q, stored as a reference quarter: Walmart U.S. eCommerce net sales of $29.60bn and technology and supply chain capital spending of $5.18bn. No filing, release or call gives an AI dollar, a bound or an AI share of any line. The roster expected supply-chain and support doors; the sources show the supply chain door credited to automation rather than AI and no support door at all.

Claim wmt-anchor-c15 (the head of Walmart U.S. on the call before coverage): speaker was CEO, now other executive, because the turn's role is President and CEO of Walmart U.S., a segment.

Claims c10 and c11 (the head of Walmart U.S. on the first-quarter fiscal 2027 call): speaker was CEO, now other executive, for the same reason.

Sam's Club is now listed as a Walmart subsidiary, so its chief executive's turns count as management in the passage counts rather than as an analyst's. None of those turns is quoted, and the coverage counts in the published quarters do not change.

The rule is in the validator and in the skill: the CEO and CFO speaker codes are for the company's own officers.

Steps. In Q2 2026: AI platform integrations, described to directional; the AI data opportunity, described to not-mentioned; and a new channel, new clients lost as AI changes search, opened directional. Low-value work automated away stays bounded: the CEO places the accelerating erosion inside the at-risk pool identified in Q1.

Upwork enters with Q1 and Q2 2026. Revenue: fees on AI-related work bought on the marketplace, Uma and other AI features, clients arriving through AI platforms, human-supervised agents, and an AI data opportunity. Toll: low-value freelance work automated away by clients' AI, and, from Q2, new clients lost as AI changes search. Spend: building Upwork's AI products. AI-related work, the displacement toll, the search toll, Uma and the build are tagged expanded against the anchor, where the marketplace already sold an AI Services category, already named generative AI replacing talent tasks as a risk, and already ran Uma on large language models, and already named search algorithm changes as a risk to traffic; the AI-platform, agent and data channels are new.

Q1 2026. Revenue was $195mn on flat GSV of $987mn, held up by an acquisition. Management described AI acting both ways: simple tasks on the smallest contracts replaced by clients' AI tools, and AI-related work growing more than 40%, 8% of marketplace GSV. A new LLM classifier put about 10% of GSV in at-risk tasks, down from 11%. The 10-Q attributed the fall in active clients to slower acquisition and lower retention without naming AI, and management named macro pressure on very small businesses first. The ledger converts both sides through the take rate of 19.4%: AI-related work $10mn to $15mn, fees lost to automation $1.9mn to $6.2mn, set between the net-headwind remark and the classifier's measured fall in the at-risk pool, 9.1% in a year.

Q2 2026. GSV fell −3.6% and revenue −1.7%. Management said AI automation of low-complexity work accelerated, as a pull-forward of erosion inside the at-risk pool already identified, the kind of 10% of the business, and cut guidance on the assumption of a heightened pace; the 10-Q now names the impact of AI on freelance work and on client acquisition among the causes of the decline. Another toll opened: the CEO said AI is changing search, Google's page changes send far fewer searches to any business, and paid search is now the largest acquisition channel, with more marketing planned for the second half. AI-related work grew 22% to an approximate $330mn run rate, and management answered the slowdown with an unmeasured claim that the metric undercounts. The Claude connector joined the ChatGPT app and integrations began to drive referrals, still nascent; the MCP server launched with the results.

What the sources do not support. The roster expected displacement to show as volume lost and as price passed back; no source describes a price conceded because of AI, and the take rate rose to 19.6% as Upwork raised fees where talent outnumbers client demand, so there is no price channel. The restructuring announced with Q1, a 24% workforce reduction, is tied to efficiency and to a nature of work that, in the CEO's release quote, shifts as AI advances; that is a demand-side reason, and no source attributes a saving to AI doing the work of the staff cut, so it is a confound on the build channel and not an AI saving. No model or token bill is given anywhere.

Sizes, latest quarter. Every size is the ledger's inference. Revenue on AI-related work $11mn to $17mn, a level whose baseline is traced to 2025-CQ2: the Q2 2026 run rate and growth rate imply fees of $13mn a year earlier, so only the part above that counts in the incremental total, and the 2025 reference quarters give growth only (25%, 30%); Uma and AI features $0 to $2.5mn; AI platforms $33k to $834k; agents $0 to $83k. Tolls: automation $2.6mn to $6.5mn, search $167k to $3.3mn, their high ends held together inside the reported GSV decline plus the AI-work gain, $10mn. The AI build $5.6mn to $19mn. The data opportunity has no size in Q2.

Steps. In Q1 FY2027, the quarter ended 2026-04-30: the company's own AI compute, state described to bounded and motive exploratory to efficiency; core consumption attributed to AI, described to directional; support and platform operations automation and engineering productivity, described to directional; Snowflake Intelligence, quantified to directional, as an account count gave way to a growth rate; services delivery productivity, directional to described; the cost of acquired AI companies, bounded to quantified; the outside bill displaced by internal agents, quantified to not mentioned. In Q2 FY2027, the quarter ended 2026-07-31: AI revenue outside the agent products, quantified to withdrawn, as the count of accounts using AI was no longer given; model provider fees, described to quantified on the 10-Q's commitment; fees and the cost of serving AI products, motive conduit to product-defensive, and that cost, described to bounded; Snowflake Intelligence, now called CoWork, directional to quantified; the displaced outside bill, not mentioned to quantified; support automation and services delivery, to not mentioned. Revenue from AI-native companies, bounded, and a frontier engineering program, described, opened. Cortex Code revenue, hiring avoided and the toll did not step.

Snowflake enters as a seller with a January fiscal year, so the quarters covered ended in January, April and July 2026. The registry reads what management calls AI revenue as additive channels: Cortex Code, the coding agent for data work; Snowflake Intelligence, the agent product for business users renamed CoWork; and the remainder, on the definition behind the run-rate AI functions inside queries, search, analyst and document processing, tagged new because each calls a language model and is billed as consumption when used rather than folded into existing plans. The CFO's wider list in the quarter ended 2026-07-31, with machine learning and notebooks, is read in that quarter only. Beside them sit core consumption that management attributes to AI, tagged relabelled because no measure of it is given, and revenue from AI-native companies, which overlaps the product channels in part and is kept out of the revenue total. Spend: fees to model providers, which sit inside the cost of serving AI products and are left out of totals; the model and GPU cost of serving AI products; the increase in AI compute for the company's own development; the cost of acquired AI companies, TensorStax and Natoma; and a frontier engineering program paid on outcomes, sized as a level. Savings: hiring avoided across the company, with support automation, services delivery and engineering output read as overlapping parts of it, and outside bills displaced by the company's own agents. Toll: workloads and budget lost to model providers and to software customers build with AI. Observe, bought for about $600mn, is read as an acquisition confound with nothing credited to AI.

Q4 FY2026. Revenue was $1.28bn, up 30.1%, with product revenue of $1.23bn up 30%. Management quantified adoption, not revenue: more than 9,100 accounts using AI features, 2,500 on Snowflake Intelligence and 4,400 customers on Cortex Code weeks after its launch. Asked for an update to a run-rate given before coverage, the CFO answered on cash flow; the ledger's AI revenue estimate of $20mn to $55mn rests on that run-rate, $100mn for the quarter ended 2025-10-31, stored as a reference quarter and quoted from the CEO in the anchor. The 10-K placed GPU and AI inference costs inside third-party cloud infrastructure expenses, up $248mn in the fiscal year, and attributed the increase to customer consumption; the only dollar figure for model providers was the announced size of an OpenAI partnership, $200mn. The CFO attributed a reduction in force of about 200 people and net additions of 37 to AI, while the 10-K showed headcount up 15.6% on the year; the same quarter a year earlier added 11, so the ledger sizes only the reduction in force. One outside bill was stated, a dashboarding system of about $5.0mn replaced by the company's own agent.

Q1 FY2027. Revenue was $1.39bn, up 33.5%; product revenue of $1.33bn grew 33.9%, an acceleration of 4% on the prior quarter, and came in $70mn above the midpoint of guidance. The CFO attributed the difference to Cortex Code, generally available from the first week of the quarter and in use at 7,100 accounts, and raised fiscal 2027 guidance by $180mn. He said the AI products carry a lower gross margin than the core and that a cloud contract of $6.00bn offsets it, holding the margin guide at 75%. The 10-Q reported $86mn more cloud infrastructure expense in cost of product revenue and $19mn more in research and development, both including AI inference and GPUs. The company added 190 employees against about 400 a year earlier, 17 of them outside the Observe acquisition, and support case throughput per engineer rose 25%. Natoma was agreed after the quarter closed for about $110mn.

Q2 FY2027. Revenue was $1.55bn, up 35.1%, with product revenue of $1.49bn up 36.8%, growth faster by 7% than two quarters earlier. The CEO put the AI products at approximately 50% of that acceleration, which on prior-year product revenue is $38mn of added growth. The count of accounts using AI was withdrawn, Cortex Code reached 9,100 accounts and CoWork 5,800, and the CFO now listed machine learning and notebooks inside AI revenue. The 10-Q disclosed the commitment to an AI service provider, made to support the AI products, raised by $190mn during or after the quarter to $390mn with $270mn remaining, and said provider terms compress the margin of the AI offerings; margin guidance fell by 1% to 74% on the AI mix. Cloud infrastructure expense rose $114mn in cost of product revenue and $43mn in research and development, and a note put it at approximately 75% of cost of product revenue against 70% a year earlier. The company added 334 employees in the half against 935 a year earlier, and the CFO said growing cloud costs are offset by slowing headcount expense. Natoma closed for $128mn. The CEO said AI-native companies remain a small part of revenue, and described a frontier engineering team paid only when it delivers customers' outcomes.

Sizes, latest quarter. Every size is the ledger's inference. Revenue: Cortex Code $14mn to $91mn and CoWork $2.9mn to $46mn, each an account count times an assumed revenue per account; the remainder of AI revenue $1.2mn to $58mn, what is left of the whole estimate after the two, on this quarter's wider definition; core consumption attributed to AI $0 to $76mn, an assumed share of the acceleration management does not credit to AI products; AI-native companies $4.5mn to $60mn, an assumed small share of product revenue, which cuts the same revenue by payer, overlaps the product channels in part and is left out of the revenue total as a conservative choice. Spend: the cost of serving AI products $48mn to $70mn, what the 10-Q's cloud-cost share leaves after core consumption at the prior-year rate; model provider fees $48mn to $70mn, a share of the commitment drawdown with an allowance for other providers, capped at that cost, which holds them; the increase in the company's own AI compute $17mn to $43mn, a lower bound on that spend; acquired AI companies $1.1mn to $6.9mn, from Natoma's purchase accounting prorated from the closing date; the frontier engineering program $557k to $5.6mn, sized as a level and counted as undetermined. Savings: hiring avoided $0 to $104mn, from nothing upward since the filings attribute none of it to AI, with engineering output $0 to $7.7mn inside it; outside bills displaced $50k to $2.1mn. Support automation is read as a shape, its headcount line still falling in a quarter that says nothing of it, and services delivery is unsized. Toll: $0 to $15mn, a small allowance against management's account of no loss.

What the quarters show. The roster expected a net payer into the cycle, a company whose model provider bill is larger than its AI revenue. The sources do not establish that. Capped at what the filed cloud-cost lines leave for AI, the fee estimate is below the AI revenue estimate at the point and the ranges overlap; a share of the commitment drawdown, uncapped, would come to more than those lines hold, which suggests the drawdown may hold prepayment or usage outside them. What the sources do show is the shape of disclosure: in the same quarter the cost side gained a traced commitment and a quantified margin effect, while the revenue side gained only a share of growth acceleration. Snowflake's next report is expected around 2026-12-02.

Steps. In Q2 2026 headcount held flat with AI moved from bounded to described, as the share of code written by AI was not repeated and the tie between headcount and AI was dropped; merchant AI tools moved from described to directional on the first outcome management attributes to Sidekick, a 8% increase in new merchants reaching an early order milestone; orders through AI platforms moved from directional to bounded when management called agentic volume small relative to quarterly GMV of $116.00bn; the agentic plan for brands without a Shopify store, described in Q1, went unmentioned. No motive or strength stepped.

Shopify enters with Q1 and Q2 2026 and these channels: a model and inference bill for merchant-facing AI in cost of subscription solutions, AI usage inside the company, internal investment in AI capability, headcount held flat with AI, merchant support handled with AI, AI tools for merchants, orders arriving through AI platforms, and the agentic plan. The inference bill, internal usage, AI-platform orders and the agentic plan are new money; the build, the workforce saving, support and merchant tools are existing activities AI changed. Headcount had already been flat for 3 quarters at the anchor call, credited to structure and automation with internal AI named as an element, so the workforce saving is sized as the AI share of headcount growth avoided rather than as the whole of it.

Q1 2026. The 10-Q named AI related usage inside a $22mn increase in cloud and infrastructure costs within cost of subscription solutions and a $25mn increase in research and development computer hardware and software, and the CFO said rising LLM cost from merchant use of Sidekick offset scale and support efficiencies. The President said the company forced the adoption of AI, that flat headcount is only possible because something changed fundamentally, and that AI writes well over 50% of code; the 10-Q explained every cost line without an AI saving. AI-driven traffic grew 8.00 times and orders from AI searches 13.00 times in a year, at the same economics as a storefront order.

Q2 2026. The 10-Q repeated both increases, $37mn and $34mn, and added that cloud and infrastructure is the significant majority of cost of subscription solutions; the CFO placed most of Sidekick's AI cost in subscription gross profit, held level, and the majority of internal AI spend in research and development. The tie between headcount and AI was not repeated: the CFO credited leverage to headcount discipline and AI to better output, while the 10-Q showed sales and marketing employee-related costs down $16mn with no AI named. The ledger's ballparks for the quarter: the build $31mn to $125mn, undetermined in the incremental total; the workforce saving $1.5mn to $47mn, counted in full as an increment; merchant tools $717k to $43mn.

What the sources did not support. Support efficiencies, tied to AI at the anchor call, are cited in both quarters without that attribution, so the support channel reads not mentioned. No toll appears: management says agents write into Shopify rather than bypass it and that search traffic is still growing. The equity investments marked through each quarter are payments and marketing companies, not AI companies, and carry no channel.

Progressive enters with Q1 and Q2 2026. The ledger registers channels for claims-handling cost displaced by AI, non-acquisition operating expense displaced by AI, investment in the AI initiatives themselves, and advertising production made with generative AI; management names the cost lines and the initiatives, not the channels. Each is tagged expanded: the anchor shows the company using machine learning, predictive models and a document chatbot for years, and its CEO already describing technology as the way to push loss adjustment and non-acquisition expense down. The 10-Qs, the monthly releases and the CEO's own quarterly letters carry no AI passage, and the company names no AI product; everything about AI in the covered quarters comes from the CEO's answers to analysts on the calls, which come weeks after the monthly release that sets the report date. The letters credit the improvement in the non-acquisition expense ratio to expense discipline and operational efficiency.

The reference quarter, Q4 2025, is where the AI framing the CEO points back to was given. The FY2025 shareholder letter says a customer-facing generative AI assistant was put into the claims communication platform during 2025, for automated tasks, information retrieval and follow-up, and names the first AI-generated TV ad, made to learn new creative tools; on the call the CEO said the ad took much less time and money than a regular commercial, that an AI Strategy Council had been formed, and that efficiencies already show in the company's data. None of it carries a figure. The advertising channel rests on those statements alone, is unmentioned in both covered quarters, and is read as exploratory.

Q1 2026. The CEO said several generative AI solutions are in production across personalized experiences for consumers, agents and business owners, a continuation of past technology investment in claims and customer relationship management, that the company is at the tip of it, and in the same answer that its strategic plan expects the non-acquisition expense ratio to keep falling through technology. Loss adjustment expenses rose 12.2% against premiums up 8%, $70mn above the prior-year rate, and the filing explains the loss and LAE ratio by severity and catastrophes. The non-distribution part of other underwriting expenses came in −$45mn against its prior-year rate and the personal vehicle non-acquisition expense ratio fell 0.4%, on a trend the anchor shows running for years before AI was named.

Q2 2026. Asked how AI would show in the expense and loss ratios, the CEO said the company has turned to generative and agentic AI in many areas, counted "dozens" of advanced initiatives producing results, and expected the first effect to be cost reduction on the LAE and expense ratio side, possibly reaching loss costs later. The CEO said the numbers exist but would not be given until they are more complete. Loss adjustment expenses rose 7.8% against premiums up 6.2%, and the non-distribution expense line rose 10.1%, $61mn above its prior-year rate, reversing the first quarter.

No step was recorded between the quarters: the cost channels are described in both, sized by the ledger's inference, and read as exploratory, since no displaced line shrinks. The savings ballparks are a share of each line times an assumed saving, with low ends at zero because both lines ran above their prior-year rate in Q2, headcount rose and the letters credit expense discipline: $0 to $61mn for claims handling and $0 to $43mn for non-acquisition expense. Spend is an AI share, referenced to the ledger's Booking bound, of a technology share that is judgment: $1.8mn to $91mn. Each range spans more than an order of magnitude. Management's own number, when it comes, would be the first step.

Steps. In fiscal Q4 2026 (2026-CQ2): priced agentic capacity opened as quantified, with 33 customers pre-purchasing tokens; staff cost avoided outside engineering opened as directional; the engineering saving moved from narrative-defensive to efficiency as research and development fell on the year; adjacent services moved from bounded to described. In fiscal Q1 2027 (2026-CQ3): inference behind embedded AI moved from described to quantified; priced agentic capacity from quantified to described; forward-deployed engineers opened as described; adjacent services and the coding-tools bill went unmentioned.

Oracle enters with three quarters and a registry that reads it mainly as a seller of capacity and as the builder of that capacity. Revenue: AI infrastructure, with adjacent services and the customer-funded contracts treated as inside it, the contracts recorded as contract value and not sized; database and wider cloud demand that management attributes to AI; AI embedded in the applications at no additional cost; the AI Data Platform; and, from fiscal Q4, priced agentic capacity. Spend: capital expenditures on the build, traced and shown and left out of totals, with their depreciation counted in their place; data center lease cost; power and other running cost of the fleet; interest on the debt raised; the coding-tools bill; inference behind the free features; severance the filing ties in part to AI adoption; and forward-deployed engineers. Savings: engineering cost avoided through code generation and staff cost avoided in other functions. Toll: application revenue lost to software customers build with AI, which management denies for its own suites.

Fiscal Q3 2026. Revenue was $17.19bn, up 21.7%. Management said AI infrastructure revenue grew 243% at a gross margin of 32% on capacity delivered, with more than 400 megawatts delivered, and gave no level; from reported cloud infrastructure revenue of $4.89bn the ledger solved $1.62bn to $2.77bn. Remaining performance obligations were $552.60bn, 12% of it due within twelve months, and $29.00bn of new contract value was signed in which the customer prepays for or supplies the accelerators. Capital expenditures were $39.17bn over nine months; depreciation rose to $2.15bn from $1.00bn and interest expense to $1.18bn from $892mn. The 10-Q did not use the term AI anywhere. The release said AI code generation let the company restructure product development into smaller teams, while research and development rose 7.3%.

Fiscal Q4 2026. Revenue was $19.18bn, up 20.6%, with cloud infrastructure revenue of $5.79bn. The growth rate for AI infrastructure was not repeated; management reported $67.00bn of AI contracts signed, utilization of 97.5% and a backlog of $638.00bn. The annual report showed $4.59bn of customer prepayments, said certain cloud infrastructure offerings are concentrated among a number of large customers, some of whom may be highly leveraged, and rewrote the purpose of the restructuring plan to include the adoption of AI technologies. Capital expenditures were $16.49bn in the quarter and $55.66bn for the year, against operating cash flow of $31.98bn; lease commitments not yet commenced stood at $260.00bn. Research and development fell −1.5% and sales and marketing −10.3% on the year.

Fiscal Q1 2027. Revenue was $19.34bn, up 29.6%, with cloud infrastructure revenue of $7.39bn, up 120.7%, on 850 megawatts delivered. Capital expenditures were $28.50bn, funded in part by $11.36bn of customer prepayments and $19.90bn of new common stock; depreciation reached $3.16bn and lease commitments not yet commenced $288.00bn. One site took up to 43.7% of the quarter’s GPUs and is said to have trained a model released by OpenAI. Embedded AI was given usage figures for the first time, more than 150,000,000 uses and 900,000,000,000 tokens, and the restructuring plan was enlarged by $700mn.

Sizes. Every size is the ledger’s inference. AI infrastructure revenue is solved from the one stated growth rate and rolled forward on reported cloud infrastructure revenue: $1.62bn to $2.77bn, $2.16bn to $3.62bn and $3.28bn to $5.22bn. The backlog is never a size, and neither is the contract value of the customer-funded contracts. Capital spending on the build is sized as a level, the quarter’s cash capital expenditures times an assumed AI share: $13.04bn to $17.70bn, $11.55bn to $15.67bn and $19.95bn to $27.07bn, against a baseline of $1.30bn, the one quarter of fiscal 2024 the anchor traces. It is cash paid and capitalized, and leaves out equipment bought on supplier financing and spending not yet paid; the part that has reached the income statement is read as depreciation, $1.41bn to $2.35bn in the latest quarter, which is what the spend total counts; the capital spending itself is not added in. Lease cost of the build is $481mn to $801mn, power and other running cost $210mn to $519mn, the part of higher infrastructure expenses that depreciation and leases do not explain, and interest $275mn to $550mn. The engineering saving, $9.0mn to $54mn, is capped at the share of the fall in research and development that the ledger attributes to AI, and the saving in other functions, $13mn to $126mn, is an assumed share of the fall in sales and marketing; inference behind the free features, $450k to $18mn, is the only spend figure built on a volume management stated.

New money, and why so little of it. Read as incremental, Oracle’s AI revenue comes to almost nothing at the point, and that is the reading the method intends, not a gap in it. Renting accelerator capacity predates LLMs, so AI infrastructure revenue is tagged expanded and, with no traced quarter of the same activity before AI, counts as undetermined: nothing at the point and in full at the high end. The larger reason sits outside the tag. The buyers management describes are model vendors and companies that raised money, and the build is paid for in part by customer prepayments and new equity. Money that investors put into AI labs, that the labs pay for capacity, and that the capacity seller spends on accelerators is one flow seen at several stops. Only revenue paid by companies absorbing AI into their own operations comes from outside that flow, and the filings do not separate it. A capacity seller shows the buildout’s money moving, not new demand arriving.

What the three quarters show. In the latest quarter the ledger reads $19.95bn to $27.07bn of capital spending on the build, cash that is capitalized rather than expensed, beside $3.28bn to $5.22bn of AI infrastructure revenue, and both are disclosed through quantities other than their own dollars: megawatts, GPUs, contract value and backlog for revenue; capital expenditures, lease commitments and borrowings for cost. The filings attribute those lines to data center expansion and infrastructure expenses; from the annual report on, the risk factors name investments in AI initiatives, infrastructure and headcount among them, and the restructuring plan names the adoption of AI. On the question the roster asked, whether this is revenue paid by AI labs out of investor capital, the stored sources support the description and not the measurement: management speaks of model vendors and of buyers that have raised money, names none, and by the latest quarter calls the demand broad based. Oracle’s next report is expected on 2026-12-14.

Steps, all in Q4 of fiscal 2026 (2026-CQ2). Opened: ambient clinical documentation (unsized: no price can be traced), web grounding sold to AI assistants, forward-deployed engineering teams, and the gain on the Anthropic investment. AI business: quantified to described, as the run rate went unrepeated. Security Copilot: quantified to described, as no count is given. Revenue from frontier model companies: described to quantified, on the annual report’s OpenAI figure. Agent governance: bounded to quantified on an agent count, and motive exploratory to offensive, now sold inside E7. Research and development for AI: reported to inferred, as the annual report mixes XBOX impairment into the explanation. Revenue share paid to OpenAI: described to not-mentioned.

Microsoft enters as seller, builder and investor at once, with a registry of channels in revenue, spend and non-operating groups. Revenue: the AI business as management’s umbrella, with M365 Copilot seats, GitHub Copilot, Foundry, Copilot credits, LinkedIn’s agentic hiring products, Security Copilot, agent governance, ambient clinical documentation and web grounding read inside it; cloud revenue from frontier model companies, also read as inside it. The products count in totals and the umbrella is shown and left out of them; and demand for databases and other services that management attributes to AI, read as relabelled. Spend: the cost-of-revenue and research increases the filings attribute to AI, advertising for Copilot, finance lease interest, depreciation and lease cost (read as components inside the attributed increases), the revenue share once paid to OpenAI, and forward-deployed engineers; capital expenditures are traced as capital. Non-operating: the OpenAI equity-method result and the Anthropic gain. No savings channel is registered: headcount fell in both quarters and management gives pace and agility, not AI, as the reason.

Q3 of fiscal 2026 (2026-CQ1). Revenue was $82.89bn. The AI business run rate passed $37.00bn, up 123%; the ledger sizes the quarter at $7.86bn to $9.71bn. Copilot paid seats passed 20,000,000, LinkedIn’s agentic hiring products passed $450mn of run rate, and Cosmos DB revenue grew 50%, which the CEO attributes to AI workloads. The 10-Q attributes the $4.80bn rise in Intelligent Cloud cost of revenue to AI infrastructure and GitHub Copilot usage, the $680mn rise in Productivity and Business Processes cost of revenue to AI infrastructure for Copilot, the $717mn rise in research and development to compute, AI talent and data, and most of the $602mn rise in sales and marketing to Copilot advertising. Capital expenditures including finance leases were $31.90bn, against $11.20bn in the anchor quarter. The CFO said the new OpenAI agreement eliminates the revenue share the company paid the lab, never sized in the sources.

Q4 of fiscal 2026 (2026-CQ2). Revenue was $90.01bn. Copilot paid seats passed 30,000,000, GitHub Copilot revenue grew more than 60% on the quarter after usage billing, and Foundry reached 100,000 customers. The annual report discloses revenue of $24.10bn from OpenAI, a related party, a ratio of 11.3% to the year’s cloud revenue, though part of it is revenue share rather than cloud revenue, and the CFO says nearly 90% of cloud revenue came from outside frontier model companies; the ledger puts the quarter’s frontier revenue at $5.34bn to $8.30bn. Depreciation was $10.30bn for the quarter against $6.30bn a year earlier, leases signed and not yet commenced reached $329.10bn, and the OpenAI stake produced a gain of $632mn beside a $3.20bn gain on Anthropic, both non-operating.

How to read the totals. The product channels count in the revenue total, each by its own novelty, and the AI business umbrella, $8.65bn to $13.11bn in Q4, is shown and left out of totals as overlapping them. Channels tagged new, Copilot seats, GitHub Copilot, frontier model companies, Security Copilot and agent governance, count in full in the incremental total; Foundry, the credit products and LinkedIn’s hiring products are sized as levels of activities that predate LLMs with no traceable quarter before AI, so they count in the incremental total only at its high end; the relabelled pull-through demand counts in the flow total and not in the incremental one. The spend total is mostly the filings’ own attributed increases, which are changes against the prior year rather than levels. Revenue from frontier model companies is the buildout’s money moving, investor capital paid to labs and spent on capacity the company builds; it counts in the revenue total and is read as inside the umbrella.

What management declines to size is part of the reading. No product revenue level is given for Copilot, GitHub Copilot, Foundry, Security Copilot, agent governance or the credit products; the AI business is not defined and its run rate was given once; the filings name no AI share of capital spending; and the revenue share paid to OpenAI was never quantified. Every size on these channels is the ledger’s, ambient clinical documentation stays unsized for want of a traceable price, with ranges that span an order of magnitude or more where only counts are given.

Figures msft-2026-cq1-f88 to msft-2026-cq1-f106 (the nine-month figures from the third-quarter 10-Q) were recreated in the 2026-CQ2 exhibit as msft-2026-cq2-f116 to msft-2026-cq2-f134 and removed from the 2026-CQ1 exhibit, where no reading, claim or note cited them. Every fourth-quarter figure backed out of the annual report now takes its nine-month input from the same exhibit.

Existing figure ids in both exhibits are unchanged; the 2026-CQ1 exhibit keeps a gap in its numbering where the moved figures were.

The rule is in the methodology (on sources) and the validator: a fiscal fourth-quarter exhibit whose own periodic report is a 10-K may also list the prior quarter's 10-Q.

Steps. In Q1 FY2027, the quarter ended 2026-04-30: the AI-native cohort and enterprise AI workloads, state bounded to directional, as smaller but accelerating drivers gives a direction and no level; Voyage models, state bounded to directional; customers arriving through coding agents, state described to directional and motive exploratory to channel-defensive. Vector Search, the Voyage models and the Voyage acquisition cost stay exploratory: the sentence that gives their movement calls these drivers smaller and early, and the forfeiture of 48.7% of the unvested Voyage stock came in the same quarter. In Q2 FY2027, the quarter ended 2026-07-31: the AI tools bill, state described to quantified and strength inferred to reported; the AI-native cohort and enterprise AI workloads, state directional to described on the word small; frontier lab revenue and self-managed demand attributed to AI, state described to directional; the extra charge for search and vector search on the self-managed product opened. Engineering and marketing spend, model compute, internal savings and the toll did not step.

MongoDB enters as a seller with a January fiscal year, so the quarters covered ended in January, April and July 2026. The registry reads revenue through an AI-native cohort, with frontier labs, Vector Search and the Voyage models as overlapping parts of it; enterprise AI workloads on Atlas and AI-attributed demand for the self-managed product, both tagged relabelled because the products are unchanged and no measure of the AI part is given; the charged add-on on the self-managed product; and customers arriving through coding agents and agent frameworks. Spend: the income-statement cost of the Voyage AI acquisition, engineering on AI capabilities, marketing aimed at AI natives, compute for the company’s own models, and the bill for AI tools its staff use. Savings: what those tools displace. Toll: workloads lost when a coding agent or a prompt-driven platform chooses another database. Clarity, a federal services firm bought in May 2026, is read as an acquisition confound with nothing credited to AI. Gains on private company stakes in Q1 FY2027, $6.5mn unrealized and $9.9mn realized, are not registered as a channel because the filing does not name the investees.

Q4 FY2026. Revenue was $695mn, up 26.7%, with Atlas-related revenue of $503mn. The CEO said AI was not yet a material driver of results, the CFO that AI-native customers were not yet meaningful drivers of revenue, and the 10-K that revenue attributable to Voyage AI since its acquisition for $161mn was not material. The only rates given were customer counts: Vector Search customers nearly doubled in a year and Voyage customers doubled since the acquisition. The 10-K gave the Voyage purchase accounting, from which the ledger computes $3.0mn of amortization and $6.7mn of retention awards a quarter, and it named the risk that AI development tools change how developers choose a database.

Q1 FY2027. Revenue was $688mn, up 25.2%. The CFO listed early enterprise AI deployments, frontier labs and AI-native companies as smaller but accelerating drivers of Atlas growth; the CEO confirmed more than one frontier lab as a customer and said results were driven primarily by core workloads. Voyage customers more than doubled on the quarter and the share of large Atlas customers on two or more features reached 45% from 37%, credited largely to vector and text search. The agent door moved from roadmap to shipped integrations. The 10-Q listed a $4.2mn increase in research and development software costs with no cause, and its restricted-stock table showed 54,584 Voyage restricted shares, 48.7% of those unvested, forfeited and reacquired in April 2026, which takes the acquisition’s award expense down from the straight-line amount of the prior quarter.

Q2 FY2027. Revenue was $772mn, up 30.5%, with Atlas-related revenue up 28.9% and Enterprise Advanced and other up 35.9%. The CFO said the company had started to see some benefit from AI and that it was small. The 10-Q attributed a $7.0mn increase in research and development software costs to increased use of AI tools, 21.8% of that line’s increase, and the six-month increase to the same cause, which dates the unexplained increase of the prior quarter. Search and Vector Search reached the self-managed product on 2026-06-30 as a charged extra, credited in part for raising that line’s full-year outlook to about 11%. Voyage customers roughly doubled again, a majority of the new ones AI natives, and most referral traffic for Voyage came from coding agents; one lab moved more inference-side workloads to Atlas; and the CEO described AI natives that start on other databases through prompt-driven platforms and migrate later.

Sizes, latest quarter. Revenue: AI-native customers $8.5mn to $45mn, an assumed share of Atlas-related revenue set well under the share this ledger reads at Datadog; inside it, frontier labs $500k to $18mn, Vector Search $2.8mn to $20mn, Voyage $1.0mn to $7.0mn and arrivals through coding agents $825k to $5.5mn; outside it, enterprise AI workloads $2.5mn to $25mn of a $127mn increase in Atlas-related revenue, self-managed demand attributed to AI $957k to $14mn and the add-on $478k to $7.2mn. Spend: AI tools $7.0mn, reported, on an increase-over-prior-year basis, with the whole bill put at $7.0mn to $10mn as a comparison; the Voyage acquisition $5.2mn to $8.3mn, of which $3.0mn is amortization traced from the filing’s balances and the rest retention awards on the restricted stock still unvested after the forfeiture; engineering on AI capabilities $10mn to $51mn, a share of research and development after the parts those other channels size are taken out; marketing aimed at AI natives $2.5mn to $15mn; model compute $737k to $8.4mn. Savings: $909k to $15mn, from near nothing upward, since none is claimed. Toll: $0 to $2.8mn, an allowance from nothing.

What the quarters show. The roster expected a seller with no AI revenue metric and nothing to anchor a ballpark on. That held for revenue: management gives adjectives and customer-count multiples, never a dollar, a share or a growth contribution, and the filings attribute growth to large existing customers every quarter. It did not hold for cost: the periodic report sized the AI tools bill, the first reported strength on this company, and the Voyage purchase accounting let the acquisition cost be computed rather than guessed. The asymmetry is the reading: the company can say what its own AI use costs and has not said what AI customers pay. MongoDB’s next report is expected around 2026-12-01.

JPMorgan Chase enters with Q1 and Q2 2026 and these channels: a model and token bill, internal build investment, operations and call-center work displaced by AI, knowledge-work productivity, fraud losses reduced by AI models, AI-improved prospecting and offers, an AI cash-management tool for consumers, cyber defense against AI-enabled attacks, and, from Q2, lending for the AI buildout. Only the token bill is new money by the ledger’s test. Build investment, operations, knowledge work and the cash tool are existing lines whose cost or reach AI changes; fraud models, prospecting, cyber defense and buildout lending are existing activities the anchor already described, read as relabelled because no movement is measured.

Q1 2026. The call spoke of AI as risk and competition: a cash tool with AI in it that the CFO called an experiment and not live, cyber risk the CEO said AI has made worse while the firm tests a frontier model, and the CEO’s view that every bank will deploy AI and the efficiency benefit will be passed on to the marketplace. The CEO also said the firm uses AI against fraud and scams and to prospect better. Nothing carried a number. The 10-Q explained every expense line without AI: technology, communications and equipment expense up 17.2% on investment in technology, compensation up 8.8% on revenue-related pay and front-office hiring. Operating losses, mostly fraud, fell to $286mn from $386mn, the direction the fraud claim predicts, with no cause given.

Q2 2026, the steps. The model and token bill moves from described to bounded: the CFO said token expense was a trivial number for the first half, with a meaningful acceleration forecast for the second, and that the full-year amount is still trivial. The operations channel moves from described to directional on the CEO’s statement that jobs were cut by 30% or 40% in discrete areas; no period or area was given, most of those staff were offered other jobs, and firm employees rose 1.1%, so its motive stays exploratory. The count of 150,000 people in operating and call centers came from a separate answer, on succession. Cyber risk went unmentioned. A new channel opened: the CFO conceded that part of the capital spending and associated loan growth is AI related, without saying how much.

The ballparks. Internal build investment, $36mn to $771mn, and the token bill, $2.6mn to $38mn, are read from the ledger’s own readings at other buyers as shares of revenue excluding the Visa and equity-investment gains ($51.75bn). The operations saving is $9.1mn to $260mn, with a high end that is a labelled judgment; knowledge-work productivity is $12mn to $340mn, with the gain rate taken from the ledger’s readings at the travel buyers and a published study, and the captured share kept low because the CFO lists capacity ahead of efficiency. Buildout lending, $14mn to $208mn, is a conceded part of the increase in Commercial & Investment Bank Lending revenue plus a small share of the existing book, and is the buildout’s money moving through a lender rather than absorption by the borrowers. The cash tool is sized at nothing while it is in test. The roster expected a stated dollar value of AI; none of the stored sources carries one.

Steps, all in Q2 2026. AI software: state quantified to directional. IBM Z capacity bought to run AI: directional to bounded. Storage and Power demand attributed to AI: described to directional. Cost avoided through the company's own AI-enabled transformation: quantified to directional. Software development cost avoided through Bob: bounded to described. Consulting delivery cost avoided through AI: described to directional. Application software revenue exposed to agents: bounded to not mentioned. Opened: Lightwell subscription revenue, the Lightwell commitment and forward deployed engineers. Generative AI consulting, platform and data demand, research and development spending on AI, and model and inference cost did not step.

IBM enters with two quarters and a registry that reads it as seller and internal user at once. Revenue: AI software; consulting engagements tagged generative AI; Red Hat, data and automation demand that management attributes to customers' AI; mainframe capacity and accelerators bought to run AI; storage and Power demand attributed to AI; and, from Q2, the Lightwell subscription. Spend: research and development on AI, model and inference cost, and, from Q2, the Lightwell commitment and forward deployed engineers. Savings: the company-wide productivity program, with developer productivity and consulting delivery inside it. Toll: application software exposed to agents, which management bounds. Confluent, bought for $11.59bn and closed in March, is a data-streaming company and is read as an acquisition confound on the Data line and the cost lines, with nothing of its revenue or cost credited to AI.

The roster's question for IBM was how the ledger handles a cumulative basis. The book of business is cumulative since mid-2023 and adds software transactional revenue to new contract value and consulting signings; the release's second exhibit still defines it in both quarters and says it does not represent revenue. It is recorded as a figure dated to the end of 2025 and sizes nothing. What the sources show is that management itself left the basis in Q1 2026, saying it would speak in revenue terms, and that the replacements are also not a quarter's revenue: a trailing year, an annualized run-rate, then shares of bookings. The period field can date such figures and cannot say that one of them is cumulative; the label carries that.

Q1 2026. Revenue was $15.92bn, up 9.5%. The CFO gave AI software at north of $1.50bn over the trailing year, growing north of 40%, and generative AI consulting at a run-rate above $4.00bn, about 40% of signings and 30% of backlog. Consulting revenue grew 0.9% adjusted for currency. On cost he stated $4.50bn of productivity savings since 2023 with $1.00bn more expected in 2026, from a program in which the CEO named tax, procurement, payables and quote-to-cash as the processes where the company has been capturing knowledge into agents. The 10-Q's expense bullet credited productivity actions with 2% points of selling, general and administrative expense, $98mn, without naming AI, while its MD&A said the company is 'driving efficiency and cost savings with our Client Zero approach, leveraging technology and embedding AI in our own workflows', with no share. Developers were said to be 45% more productive with Bob while research and development expense rose 11.4%. The CEO bounded the application software exposed to agents at about 4% of the portfolio.

Q2 2026. Revenue was $17.16bn, up 1.1%, below the company's expectations. Management attributed most of the shortfall to large deals deferred as clients moved capital budgets to servers, storage and memory ahead of shortages and price increases; Transaction Processing revenue fell $178mn and derived IBM Z revenue changed by −42%, while Distributed Infrastructure grew 37.3%. About 33.3% of the slipped deals had closed by the time of the call. No AI software figure was given and the 10-Q stopped naming generative AI products as a reason for Data growth. Generative AI was about 50% of consulting signings and over 30% of backlog, with no revenue figure. The 10-Q credited productivity actions with 7% points of selling, general and administrative expense, $352mn, and consulting gross margin rose to 28.9% from 27.5%. Lightwell was launched at $1.0mn per client per year against a stated commitment of $5.00bn, and became generally available after the quarter ended.

Sizes for Q2. Every size is the ledger's inference. Revenue: AI software $255mn to $630mn, carrying the Q1 trailing-year figure; generative AI consulting $1.07bn to $1.44bn, with the Q1 floor of over a fifth of consulting revenue as its low end; platform and data demand $4.8mn to $73mn; mainframe AI capacity $10.0mn to $150mn, an assumed share of a derived line, sized as a level; storage and Power demand $21mn to $213mn; Lightwell $0 to $750k. Spend: research and development on AI $231mn to $1.04bn, an assumed share of the whole line; model and inference cost $22mn to $269mn; the Lightwell commitment $62mn to $1.25bn, inside the research and development figure; forward deployed engineers $750k to $90mn. Savings: $18mn to $141mn for the AI share of the company-wide program's selling, general and administrative benefit, with AI read as one of several levers, plus developer productivity at $8.3mn to $146mn and consulting delivery at $7.4mn to $56mn, which are sized on research and development and consulting cost and added.

What the pair of quarters shows. The largest revenue numbers are the least incremental on this ledger's reading: generative AI consulting counts as relabelled because the line it sits in did not grow, and its motive is read as narrative-defensive for the same reason; AI software and mainframe AI capacity are levels whose earlier size cannot be traced. Disclosure moved away from dollars in the quarter the company missed: the trailing-year software figure, the consulting run-rate, the productivity totals and the developer productivity rate were each given once and not repeated, while the filing's own measures of the cost lines moved further than in Q1. Open questions the sources do not answer: whether AI software is mostly transactional, as the Q1 arithmetic implies, or recurring, as the CFO said in Q2; what the Lightwell commitment covers and over what period; and how much of the deferred revenue returns. IBM's next report is expected around 2026-10-21.

Steps. Fiscal Q1 2027 (2026-CQ2): the agent platform went from described to quantified when management gave its run rate; demand attributed to AI moved from described to directional with activity growth rates; new channels opened with the Act Two letter (AI labs as customers, an internal AI tools bill, and internal process automation). Fiscal Q2 2027 (2026-CQ3): the run rate metric widened to include Flex and stays quantified, read on its new definition; model and compute cost moved from directional to bounded when management said the SaaS mix moved the margin more than AI; internal process automation took a size once the restructuring was in effect; the toll from customers' AI code experimentation went from described to not mentioned, and is left unsized.

Registry. Revenue: the agent platform (Duo Agent Platform credits), the older Duo Pro and Duo Enterprise seat add-ons, core seat and platform demand that management attributes to AI-driven code volume (tagged relabelled, since the anchor call of 2023 made the same argument with no line attributed, and sized net of the AI products), and AI labs and AI start-ups as customers (kept out of totals as an overlap). Spend: model fees and AI compute in cost of revenue, engineering and adoption spending on AI, and an internal AI tools bill. Savings: roles removed by internal AI agents, among the operational changes the letter lists. Toll: Premium growth lost to customers' AI code experimentation, which management named in Q4 fiscal 2026 among several causes.

Fiscal Q4 2026 (2026-CQ1). Revenue was $260mn, up 23.2%. The agent platform became generally available in mid-January with usage-based credits; management called its contribution minimal and gave no level. The ledger puts it at $25k to $750k, the seat add-ons at $1.8mn to $28mn, model and compute cost at $849k to $7.1mn and AI engineering at $10mn to $34mn. The toll from AI code experimentation reads at $0 to $5.0mn, from zero.

Fiscal Q1 2027 (2026-CQ2). Revenue was $264mn, up 23.1%. The agent platform's paid run rate was nearly $20mn, more net new ARR than the seat add-ons ever added in a quarter, and the CFO asked that it not be modeled. The ledger reads its quarter at $1.9mn to $3.3mn. After the quarter ended, the Act Two letter announced a restructuring affecting 350 team members, with internal AI agents among the operational changes and an explicit statement that it is not an AI cost-cutting exercise; no saving fell in the quarter. The CFO attributes the quarter's extra seat contraction to layoffs at customers and mergers, not to AI.

Fiscal Q2 2027 (2026-CQ3). Revenue was $286mn, up 21.3%. The calls give inconsistent levels for the end of fiscal Q1: nearly $20mn for the agent platform on the earlier call, $15mn platform-wide on the later; the ledger's agent platform estimate of $3.7mn to $6.4mn spans both. The 10-Q puts Flex commitments not yet provisioned to any product at $40mn. Model and compute cost reads at $1.1mn to $6.3mn, bounded below the SaaS mix's part of the $10mn by which cost of subscription revenue, net of restructuring charges, exceeded its prior-year share. AI engineering reads at $13mn to $47mn net of restructuring charges and AI-attributed role savings at $0 to $6.2mn from zero. The call does not name AI as a pressure on the price-sensitive cohort, whose segments stabilized, and the toll is left unsized.

What the quarters show. GitLab talks about AI more than it sizes it: the agent platform is the only AI product with a disclosed number, and that number changed definition after a quarter. The filings name third-party hosting as the cause of the margin decline and never attribute a cost to AI. The restructuring names AI as a reason for role reductions while saying it is not about AI. A reference quarter is not needed: no AI metric was given before coverage. GitLab's next report is expected around 2026-12-01.

Duolingo enters with Q1 and Q2 2026 and these channels: the model and inference bill in cost of revenues, AI used inside the business, course content produced with AI, experiments per head in engineering and product, the Max tier premium, and Video Call added to Super Duolingo. The spend channels are new money. Content and engineering are existing lines AI changed: content was already generated with large language models at the anchor, and what moves in coverage is the volume published per quarter. Video Call in Super is an AI feature added to an existing tier.

Max is tagged relabelled. It launched in 2023, before the anchor, and on the Q1 2024 call the CEO called the AI features a good excuse to start a third tier the company had wanted for a while, said the top package need not hold only AI features, and described non-AI packaging being tested in it; the CFO called Max subscribers relatively small. No Max price, subscriber count or revenue appears in any stored source, and in Q1 2026 the CEO said metrics showed no big difference with Max hidden from new users. Its ballpark is a flow size that counts nothing in the incremental total, and its motive is product-defensive in both quarters, because Max's defining AI feature is moving into Super with no stated price change.

Q1 2026. The filing and the letter attribute gross margin of 73% primarily to continued reductions in per-unit AI costs, and the CFO guided margin down to 69% by Q4 because the company means to put more AI into the product. The quantified claim was output: 20,500 course units in the quarter, 2.89 times the 2025 quarterly average, credited to AI tools that automate content creation. The 10-Q put the increase in software and AI costs, combined, at $2.2mn of the rise in Research and development, which otherwise grew on headcount. The ledger sizes content production at $0 to $5.9mn and the engineering saving at $0 to $3.4mn, reading most of the gain as more output rather than lower cost; the low end of each is zero.

Q2 2026. The inference bill moved from directional to bounded on the CFO's tens of millions, and internal AI from directional to quantified on the CFO's figure of $10mn, which the ledger reads as closer to an annual level because the 10-Q's increases in software and third-party AI costs leave little room for a quarterly one. Content production and AI in engineering went unmentioned: the course-unit figures were not repeated, and the only per-head remark answered a question about the user-growth bonus plan without naming AI. Gross margin came in at 72.6% and the year-end expectation rose to 70%. The ledger's ballparks are $0 to $34mn for the Max premium and $0 to $4.6mn for Super.

The roster hypothesis was an AI-first memo that replaced contractors, then a partial walk-back. Neither appears in the calls, letters or 10-Qs for the covered quarters: contractors appear only in the boilerplate description of cost of revenues, no AI-attributed headcount change or contractor count is stated, and the covered 10-Qs point to a fiscal 2025 10-K for risk factors that is not among the stored sources. The anchor 10-K records a fall in net contractor costs within Research and development with no cause named. There is no contractor channel, and the content channel rests on output figures alone.

No toll channel is registered: no covered source names another company's AI product as a cost or as lost revenue. Every size on the page is the ledger's inference; management gave levels for the spend only as phrases or without a period.

Chegg enters with two quarters and six channels: Academic Services revenue lost to AI answers (toll), workforce skilling revenue from AI-focused programs (revenue), AI features in Chegg Study holding subscribers (revenue), course content built with AI (savings), AI productivity across the company (savings) and the AI product build (spend). The toll is tagged expanded, sized as an increment: the fiscal 2024 10-K shows search-referred traffic and search rankings lost before LLMs, and AI Overviews and generative AI changed the size of that loss. Skilling programs, content production and the product build are also expanded. Study retention and operating efficiency are tagged relabelled: coverage gives no measure of AI-attributed retention, and the cost decrease credited partly to AI is the one the filings explain by restructuring. Both still carry a size, which counts zero in the incremental total.

Q1 2026. Academic Services revenue fell −56.6% to $46mn. The 10-Q attributes $58mn of subscription and $1.9mn of advertising revenue to reduced traffic and names AI Overviews and generative AI as the cause, and Chegg is suing Google over AI Overviews. After removing a paid marketing cut of $11mn and a recurrence of the fiscal 2023 subscription decline of −4.7%, both at their high end, 72.9% of the decrease is left to AI; the toll is $43mn to $60mn, against the prior-year quarter. The 10-Q attributes $900k of skilling growth to workforce programs, primarily the AI-focused ones, the one AI dollar it reports. The CEO said about 40% of costs had been removed while the business was retooled to be AI first. The 10-Q credits restructuring, so the AI saving is the ledger's small share of the cost at the prior-year rate, $210k to $2.1mn, and the motive is narrative-defensive.

Q2 2026. One step: AI productivity moved from described to directional. The CFO named enhanced use of AI beside expense discipline for non-GAAP operating expenses of $32mn, against $64mn a year earlier, and the CEO called the workforce restructuring becoming AI-first. Against the prior-year share of revenue those costs moved $904k, above the rate: they fell with revenue, not faster. The saving's range, $0 to $3.1mn, therefore starts at zero. The toll's attribution hardened, the 10-Q now saying AI tools and products have reduced traffic and subscribers. With a smaller marketing cut the floor rises to 85.7%, and the toll is $41mn to $48mn. A content licensing decrease of $5.6mn is explained apart and left out. Skilling growth slowed to $300k. The CEO no longer credited AI features with slowing the decline, and the ledger lowered that range to $0 to $1.4mn. The AI build is now an employability platform in beta and an agentic coach, with no revenue or cost given.

What Chegg adds to the ledger. Concentrix and TTEC, the support outsourcers, say AI is not taking their volume and attribute their declines to other causes. Chegg says the opposite in its filings, and the toll is most of what moves on its income statement: at the point the Q2 toll, $41mn to $48mn, is most of the −$54mn change in Academic Services revenue, against total revenue of $52mn. The toll is measured only against the prior-year quarter, so the revenue already lost before 2025 is not in it. No source in coverage mentions what Chegg pays for models.

Steps, all in the quarter ended 2026-05-31. Delivery cost avoided through AI in client work: state directional to described. Corporate function cost avoided through the company's own AI: described to not-mentioned, with the ledger's size carried through the silence. Training and reskilling for AI: quantified to withdrawn, because the training hours given on the last call before coverage and on the prior call are no longer given, read by the same rule as advanced AI revenue. Client budget diverted to AI infrastructure and tokens: opened, bounded by management as not material. Advanced AI services, read withdrawn in both quarters, data projects, partner bookings, the relabelled demand channels, the tools and token bill, the cost of the Faculty acquisition, platform build and the tolls for work automated away and for price conceded did not step. Management's AI passages on the call and in the release fell in number, by less than the halving that marks a disclosure step.

Accenture enters with a registry that reads it as seller, exposed vendor, user and spender at once. As a seller: advanced AI services, the data projects that follow them, work sold with AI and data partners (inside the first), and relabelled streams: modernization sold as getting ready for AI and the businesses management calls AI enablers. As a company whose billable labor is exposed: tolls for work shortened or replaced by AI, for price conceded on AI productivity, and from the second quarter for client budget diverted to infrastructure and tokens. As a user: savings in client delivery and in corporate functions. As a spender: training, tools and tokens, platform build, and the income-statement cost of the one acquisition read as an AI company, Faculty. The data center, cybersecurity, network data and capital projects firms it bought are acquisition confounds with nothing credited to AI, whatever the call names them.

The roster expected this company to show the withdrawn state at its clearest, and the stored sources bear it out. The anchor call gives generative AI sales of $300mn for all of fiscal 2023 and over $450mn for one quarter, a figure the CEO called pure. The last quarter before coverage, stored beside the anchor as a reference quarter and not extracted, carries the renamed metric: advanced AI bookings of $2.20bn and revenue of about $1.10bn in the quarter, $4.80bn of revenue since the metric began; on that call the CEO said it was the last quarter in which the figures would be shared, because AI is now embedded in larger solutions and isolating it has become less meaningful. Neither covered call, release or 10-Q gives any AI dollar, so the state is withdrawn in both quarters. The withdrawal was announced before coverage, so it is a state in the first covered quarter and not a step between covered quarters. On the first covered call an analyst asked what quantitative evidence shows the company benefiting from AI; the CEO pointed to market share, said AI was no longer isolated, and said the metrics would change.

Q2 FY2026 (ended 2026-02-28). Revenues were $18.04bn, up 4% in local currency. Management counted about 100 more clients starting advanced AI projects, said at least 50% of such projects lead to a data project, and reported 13,000,000 training hours and over 85,000 AI and data professionals, with use of AI tools made part of performance evaluation. The CEO said AI is applied in delivery and that internal efficiencies show in operating income, with no figure; the 10-Q showed cost of services −$71mn relative to its prior-year share of revenues and utilization of 93%, and attributed the margin gain to lower subcontractor costs, after a severance program of $628mn. Asked about AI compressing project timelines and rate cards, the CEO did not deny the compression and called the effect a net benefit; the CFO said pricing improved in some areas.

Q3 FY2026 (ended 2026-05-31). Revenues were $18.72bn, up 3% in local currency, with consulting up 1%. The client count and the attach rate were repeated, the emerging AI and data partners were named, and the CEO added that AI projects are still small with average size rising. The one new number was cybersecurity services of $10.00bn in fiscal 2025, given to frame the purchase of a group of operational technology security companies that lifts the year's acquisition plan to about $9.00bn; it is not an AI figure. Training counts, given on the last call before coverage and on the prior call, were not given, so training reads withdrawn; the statements about AI in delivery and internal operations were not repeated. Payroll costs fell as a share of revenues while non-payroll costs, subcontractors among them, rose by about as much, gross margin slipped to 32.8% from 32.9%, and the workforce grew by 13,000. Asked about clients' token and infrastructure spending, the CEO said it was not material to services spending and that budgets were not increasing even with AI.

Sizes, all the ledger's own. Advanced AI services: $704mn to $1.58bn in the latest quarter and $880mn to $1.32bn in the one before, the last quarterly level the company reported, $1.10bn, carried forward under assumed quarterly growth rates. Data projects added by AI: $23mn to $1.27bn, the stated attach floor on that estimate at an assumed relative size, counting only the assumed share of the data work clients would not have bought without the AI project, since data modernization was already sold at the anchor. Relabelled demand: $0 to $319mn for AI readiness and $0 to $408mn for the enabler businesses, both outside the incremental total; AI readiness overlaps the data projects and is also left out of the revenue total. Savings: $7.6mn to $483mn in delivery and $3.0mn to $59mn in corporate functions, the prior quarter's estimate carried through a quarter in which management says nothing about them. Spend: training $41mn to $585mn, carrying the prior quarter's stated hours; platform build $30mn to $234mn; tools and tokens $2.4mn to $168mn; Faculty $521k to $24mn, on a price bounded by the 10-Q's consideration, net of cash acquired, for the quarter's acquisitions and placed with a dated press report of the deal value. Tolls: $0 to $280mn, $0 to $187mn and $0 to $94mn, each a range that starts at nothing.

What the pair of quarters shows. A company that once reported its AI sales in dollars now describes AI as part of everything it does and offers counts, a rate and a multiple in place of a level; motive on the headline channel and on both savings channels reads narrative-defensive under the rule that takes the less durable tell. On services pricing, the stated fact is that over 60% of work is fixed price and rising, with margins the CFO calls little different from other contracts; the move to revenue not tied to headcount is being made by acquisition and in new categories, and no outcome-based pricing measure is given. The fiscal fourth-quarter release was filed on 2026-10-01; that quarter is extracted once the 10-K and the call transcript are stored.

The test. Each channel is read against the company's anchor: the annual report for the fiscal year it labels 2024 and that year's first call, stored and hashed like every other source (Klarna, not public then, is anchored on its prospectus and its oldest full call). The test runs in order. An activity that could not exist without a language model, or whose payer exists only because of one, is new. Otherwise the anchor must show the activity under some name, or the tag is unknown. If it shows it, the channel is expanded when a line, a price, a volume or a measured rate moves because of AI, and relabelled when only the name changed. When the evidence supports two tags the less durable one is taken. An expanded or relabelled tag must rest on a quote from the anchor or a baseline figure, and the validator now enforces that. No channel ended as unknown.

The incremental total. An expanded channel also records whether its size is an increment (a saving against the prior-year rate, a lift, a measured share of growth) or the level of an activity that existed before (a product line, a budget redirected to AI work). An increment counts in full. A level counts only above a traced baseline, and no such baseline could be traced for any level channel in this cohort, so each of them counts at zero at the point and in full at the high end, and its dollars are shown beside the total as undetermined. That group holds the redirected build budgets at Booking ($8.1mn to $90mn), Airbnb ($20mn to $134mn), Salesforce ($169mn to $844mn), Concentrix ($6.8mn to $23mn) and TTEC, Salesforce's data platform, Concentrix's iX Suite ($19mn to $30mn), TTEC's Digital practices ($16mn to $42mn) and Datadog's in-platform assistant ($1.1mn to $22mn).

Changes to existing tags, each as was, now, because. ServiceNow, core workflow demand from AI: was expanded, now relabelled, because the anchor sells the same products with the same stated driver and the growth rate is no higher. ServiceNow, security demand from AI: was expanded, now relabelled, because the security products are in the anchor and no line, price or rate is shown to move because of AI. ServiceNow, AI contract value and tier price uplift: were new, now expanded, because the tier is the next rung above a Pro tier that already carried a chatbot and predictive features; both are sized as increments and still count in full. ServiceNow, acquisition costs: was new, now expanded, because acquiring companies is an activity the anchor shows. Salesforce, implementation services displaced: was expanded, now relabelled, because the anchor already gives the cause the covered filings give, less demand for large transformation engagements, and no measure ties the decline to AI. Concentrix, iX Suite revenue: was new, now expanded, because the anchor call describes the company's own chatbots already automating client contacts inside services revenue; the suite prices that work separately. Concentrix, build investment: was new, now expanded, because technology development was an existing cost that the company raised for the suite. TTEC, pricing pass-back: was expanded, now relabelled, because the anchor shows price concessions under broader names and no price is shown to move because of AI. Airbnb, AI search and discovery: was new, now expanded, because guest search was already a conversion lever in the anchor and AI search is a feature added to it.

First tags. Booking, Expedia and Datadog had none. The pattern at the two travel companies matches Airbnb's: the vendor bill, demand arriving from AI platforms and fees paid to appear there are new; support, marketing and engineering savings are expanded; ranking, personalization and partner tools the anchor already described as machine learning are relabelled, including the largest revenue ballparks at both ($0 to $147mn at Booking, $0 to $86mn at Expedia). The cost of buying back search traffic is expanded, because the dependence on search engines is an anchor risk factor and the AI-answer part is what changed. At Datadog the cohort, the lab training deals and the products that observe AI workloads are new; growth in ordinary platform usage that management attributes to AI, with no measure of it, is relabelled ($10mn to $82mn).

What rests on judgment. Reviewers worked one company each and read three points differently; the readings were settled across the cohort and written into the methodology. Putting language-model tools into existing work is itself a change, so an unmeasured saving is expanded and not relabelled. An unpriced feature that lifts an existing revenue line is expanded, while a separately priced product is new unless the anchor shows the same work already delivered under another name. Money from payers that exist only because of language models is new whatever they purchase. The least certain tags are Concentrix's iX Suite, where management calls the revenue all incremental and the anchor shows a predecessor, and Salesforce's data platform, where relabelled is arguable and gives the same incremental point.

Nothing in any exhibit changed: no figure, quote, reading or step. The tags, baselines and anchor evidence sit in each company's registry and anchor file, and the pages now show them on every channel card.

Steps, all in Q2 2026. AI contract value: state directional to quantified. Tier price uplift: described to quantified. AI Control Tower: directional to quantified. Security demand attributed to AI: motive exploratory to offensive. Inference and cloud cost: described to directional. Company-wide productivity from the company's own AI: quantified to directional. Internal service desk automation: quantified to described. The Moveworks acquisition cost did not step, and seat and budget loss to customers' AI stayed bounded at nothing by management in both quarters.

ServiceNow enters with two quarters and a registry that reads it as seller, buyer and acquirer at once. Revenue: the AI contract value metric, with tier uplift, assist consumption, AI Control Tower and EmployeeWorks treated as inside it; core workflow demand and security demand that management attributes to customers' AI. Spend: model and cloud cost of serving AI, the income-statement cost of the Moveworks acquisition, and forward-deployed engineering to get customers' AI live. Savings: the company's own use of its AI, company-wide and in the service desk. Toll: seats and budget lost to customers' AI, which management says is nothing. Veza and Armis, security companies bought in 2026, are read as acquisition confounds on the cost lines, with nothing of their cost credited to AI.

Q1 2026. Revenue was $3.77bn, up 22.1%. Management gave no level for AI contract value; it raised the year-end target to $1.50bn from $1.00bn, said large Now Assist customers grew 130%, and restated the definition as the incremental contribution of AI now that every tier contains it. The ledger sized the quarter's AI revenue at $150mn to $218mn, on a contract value it places at $690mn to $870mn by quarter end. On cost, the CFO attributed an operating margin 0.5% above guidance to AI efficiencies and the CEO stated $500mn of productivity, 90% of the company's own cases resolved by agents and flat headcount; the 10-Q attributed every expense increase to increased headcount and the leverage to sales productivity. Subscription gross margin was 78% against 81% on third-party cloud expense and acquired-intangible amortization, with no mention of AI.

Q2 2026. Revenue was $3.99bn, up 24%. AI contract value crossed $1.00bn with net new contract value up 40% on the quarter, under a metric renamed from Now Assist to ServiceNow AI. Price uplift was given as rates, above 30% for ProPlus and from 20% for the AI-native tiers, and AI Control Tower as more than 500 live customers. Subscription gross margin guidance moved to 81% from 81.5%, attributed jointly to hyperscaler usage and customer AI adoption; applied to the quarter the reduction is $19mn. The reported margin fell to 73% from 80%, mostly on amortization after Armis closed for $7.60bn, which is not an AI cost on this ledger. The operating margin beat was attributed to revenue and timing of spend, operating expenses rose to 66.6% of revenue from 66.3%, and the Q1 productivity figure was not repeated.

Sizes. Every size is the ledger's inference. The Q2 AI revenue estimate, $193mn to $250mn, is one quarter of the stated contract value scaled by an assumed recognition share, from the opening level alone to the full stated level. Inside it, tier uplift is put at $143mn to $202mn and assist consumption at $18mn to $77mn by an assumed split; AI Control Tower at $6.3mn to $50mn from the stated customer count and an assumed contract value; EmployeeWorks at $24mn to $50mn on an acquired base taken from a public announcement outside the stored sources. Outside it, core demand attributed to AI is $15mn to $181mn and security demand $8.7mn to $72mn, both assumed shares of reported or stated bases. Spend: inference and cloud $11mn to $110mn; the Moveworks acquisition $41mn to $52mn, of which $41mn is straight-line amortization computed from the filing's purchase accounting; adoption services $8.1mn to $32mn. Savings: company-wide $10mn to $125mn, carrying the Q1 figure, with the service desk at $12mn to $97mn inside it. Toll: $0 to $19mn, from management's zero to a small allowance.

What the pair of quarters shows. On revenue the company discloses more each quarter: a stated contract value level, uplift rates and a customer count by the second quarter, though none of them is recognized revenue. On its own costs it moved the other way, from a dollar figure and case shares to a headcount commitment, while the filing showed operating expenses rising as a share of revenue with acquisitions and severance in them. The contract value figure carries open questions the sources do not answer: whether the acquired Moveworks product is inside it, how the incremental AI contribution is separated now that AI is in every tier, and what the level was at the end of Q1. ServiceNow's next report is expected around 2026-10-21.

Salesforce enters with the quarters ended January, April and July 2026; its fiscal year ends in January, so the oldest is read from the 10-K. Channels are registered in all four flows. Revenue: Agentforce itself; the premium editions and Flex Credits contained in it; the data platform and Informatica Cloud that management counts in the same metric; pull-through to ordinary seats; AI labs and AI-native companies as customers; Slack upgrades driven by Slackbot; headless access for outside agents; pipeline from the company's own sales agents; and gains on its Anthropic stake. Spend: model tokens, coding tools and the build. Savings: its own support, engineering and overhead. Tolls: subscription revenue lost to customers' AI, and professional services displaced by faster implementation.

Q4 fiscal 2026 set the levels. Management stated Agentforce at $800mn, up 169%, inside a combined metric of $2.90bn of which $1.10bn was Informatica Cloud, acquired during the quarter. Credits were 50% of Agentforce bookings. Usage was given in tokens (19,000,000,000,000 to date) and in a new unit of agent work, with the CFO expecting gross margin about neutral. Internal efficiency was asserted without a measure while general and administrative expense rose and $286mn of restructuring was booked, which is why that channel opened as narrative-defensive. Attrition of about 8% and growing seats bound the toll management itself names.

Q1 fiscal 2027 steps. Pull-through moved from directional to quantified on a spend multiple of 1.50 for the heaviest agent users. Support moved from described to quantified and from unknown to efficiency on 4,000,000 inquiries handled by the agent. Internal efficiency moved from described to quantified and from narrative-defensive to efficiency on 3,800,000 annualized hours credited to Slackbot. Slackbot upgrades moved from described to directional and from product-defensive to offensive. AI-company customers moved from unknown to offensive, and their funding from investor capital to mixed once a cash-generating parent was named. Flex Credits fell from quantified to directional when the share of bookings was not repeated. Channels opened for headless access, the coding-tools bill, engineering productivity and displaced implementation services. The filing regrouped revenue under a category named Agentforce Apps, $6.91bn, which is the existing applications and not the size of the AI product.

Q2 fiscal 2027 steps. The Anthropic gain moved from described to quantified and from inferred to reported: the 10-Q names the investment and attributes $2.70bn of unrealized gains to it, 23.8% of the quarter's revenue. Premium editions moved from directional to quantified on 5% penetration at a premium of 60% to 80%. The Agentforce figure was recharacterized: from this quarter it counts Slackbot and Headless, which were outside it before, and the release adds that the metric may be updated for new products or acquired technologies. Its rise from $1.20bn to above $1.50bn is therefore part growth and part reclassification, which management does not split. Informatica Cloud recurring revenue, stated in both prior quarters, was withdrawn: the state moved from quantified to withdrawn, and the data platform from quantified to bounded, because the release still gives the combined metric and Agentforce and no longer the Informatica component. AI-company customers and Slackbot upgrades moved to quantified. Headless moved from described to directional and took its first size as it was folded into Agentforce recurring value. Model cost fell from quantified to directional as the token count was dropped, while the filing for the first time attributed part of the rise in cost of revenues and in research and development to generative AI spend; the coding-tools bill moved from described to directional on the same sentence. The company's own sales agents fell from quantified to described and engineering productivity from directional to described.

Sizes. Management sizes run-rates, counts and shares; none of it is quarterly revenue or cost, so every operating channel is the ledger's inference. Agentforce is read as one quarter of the stated run-rate scaled for recognition: $140mn to $200mn, $201mn to $300mn, then $300mn to $375mn, the last on a wider definition. The data platform and Informatica Cloud add $248mn to $325mn and $275mn to $325mn in the latest quarter, the latter being acquired revenue under an AI and data label. The widest ranges are the build ($169mn to $844mn, an assumed share of research and development), internal efficiency ($0 to $304mn, stated hours at an assumed cost and realized share), pull-through ($0 to $207mn) and model cost ($2.9mn to $154mn, work units times an assumed token count and price). The toll on seats, $0 to $87mn, starts at zero because management's bound does.

What the seller control shows. The roster placed Salesforce as the enterprise-funded seller, and the product channels read that way: operating cash flow throughout. AI labs do appear, in more than one role. They are customers, a ballpark of $3.5mn to $83mn in the latest quarter against the Agentforce figure above. They are suppliers of the models behind the product. And one is an investment whose mark, $2.70bn, is several times the quarter's estimated Agentforce revenue. The gain is non-cash and below operating income; it is recorded as a non-operating gain and left out of flow totals.

TTEC enters with two quarters and five channels: AI design, build and operate services and software in TTEC Digital (revenue), Engage volume displaced by clients' AI (toll), AI productivity passed back in price (toll), AI tools in its own delivery (savings) and the proprietary AI software build (spend). Every channel is tagged expanded: the activities existed before LLMs and AI changes their size, or their name. Digital segment revenue is reported; no AI practice line, automation effect or tool cost is.

Q1 2026. Revenue fell −7.1% to $496mn, Engage by −$32mn with a seasonal public sector client 40% or more of the decline, Digital by −5.7% as recurring revenue fell 7.3% on legacy CCaaS while professional services outside the two legacy practices grew 15.3%. The call was dense with AI and carried no AI dollar: interview-to-hire rates up as much as 25% from AI-aided hiring, over 25,000 associates on the AI-enabled coaching platform, an AI Gateway launched and two platforms in beta. The CEO said the strategy had not hit the numbers column yet, so Digital took an exploratory motive; the delivery productivity channel took narrative-defensive because its tells conflict, a hiring gain against heavy AI talk, no dollar, and a cost of services line that rose as a share of revenue. The one management bound the ledger could apply was the growth share of capital expenditure, $3.5mn to $4.2mn, which brackets the AI software build at $960k to $3.8mn.

Q2 2026. Five steps, four on channels and one on the company. Volume displacement moved from bounded to described: the Q1 statement is not repeated and management now describes deals weighing the mix of technology and human interaction, new programs starting at smaller volumes to validate outcomes, and automation offered to fix underperforming programs; the 10-Q attributes the decline to client attrition and a completed contract, utilization fell to 70% on reduced client forecasts, and offshore delivery rose to 43% of Engage revenue. Pricing pass-back, bounded at nothing in Q1, went unmentioned. Delivery productivity lost its metrics and became directional, the CFO naming shop-floor AI as the main driver of second-half margin improvement while Engage cost of services rose again as a share of revenue. Digital's motive stepped from exploratory to narrative-defensive: a second quarter without an AI dollar, growth outside the legacy practices slowing to 13%, smaller deals, and a sale process framed by AI valuations. The disclosure step fired on the counts: management's AI passages on the call fell by more than half.

Sizes. Digital's AI-led revenue is an assumed share of the segment, $16mn to $42mn. Volume displacement runs from zero to a small share of Engage revenue, $0 to $11mn, with the filing's non-AI explanations keeping the low end at nothing. Delivery productivity is a small share of segment cost of services, $1.4mn to $8.7mn, with the line itself moving the other way. The AI build, $2.0mn to $7.9mn, rests on a capital expenditure base inflated this quarter by equipment pulled forward. Pricing pass-back has no size in Q2.

What the pair now shows. Concentrix and TTEC agree with each other and disagree with the buyers: neither vendor reports AI taking its volume or its price, both attribute their revenue pressure to offshoring, client decisions and portfolio exits, and both put their AI on the revenue side of their own statements. The difference between them is that Concentrix sizes the sale of AI and TTEC does not; TTEC's Q2 quiet, with its AI segment up for review, is the first time a cohort company's AI disclosure has thinned alongside a deteriorating quarter.

Klarna was added as the buyer whose support door swung back: AI agents announced, then human agents rehired on quality. The 2026 sources do not tell that story in words. Across two calls, two press releases, two earnings releases and two sets of interim statements, management's AI statements amount to a clause attributing operating leverage to AI-enabled productivity gains and continued cost discipline together, one revenue-per-employee figure in Q1, one agentic-commerce answer on the Q1 call, and a Q2 placement paragraph for Google Pay inside Gemini and a Shopping Search app in ChatGPT. The customer service assistant, its share of contacts, headcount and the AI cost line are not mentioned in either quarter.

The lines carry what the words do not. Customer service and operations was −$55mn in Q1 against −$51mn a year earlier while GMV grew 33%: −$13mn against the prior-year rate, with a weaker dollar inflating the reported figure. In Q2 the line was −$58mn against −$51mn with GMV up 17.5%: −$1.9mn against the prior-year rate, and growth faster than active consumers. The filing attributes none of it to AI or to hiring. The ledger's decomposition puts the AI saving at $1.6mn to $12mn in Q1 and $0 to $1.7mn in Q2.

Elsewhere the money is inference on inference. AI productivity in the technology, marketing and corporate lines is ballparked at $3.5mn to $21mn and $4.0mn to $24mn from a personnel share and a range for time freed, against non-transaction operating expenses that rose 3.3% in Q1 and 16.1% in Q2 while revenue grew 44.4% and 26.6%. Revenue arriving through AI platforms is ballparked at $506k to $20mn and $1.0mn to $31mn from a share of GMV at the quarter's take rate; Klarna gives direction and no level. The AI vendor bill is registered and unsized in both quarters, because nothing in the sources permits sizing it.

Two notes on method. Klarna is the ledger's first foreign private issuer, and its interim statements arrive as an exhibit to a foreign private issuer's periodic report with no XBRL: every reported line here is a table row or a verbatim quote from the statement, and the Q2 statement was stored as flattened text, so its lines are quotes rather than rows. And the state withdrawn, which this company was meant to test, could not be applied: the metrics Klarna stopped giving were given in releases from 2025, before the ledger's coverage, and the ledger does not store them. Within the record, revenue per employee was said once and not repeated, which the methodology treats as a quarter taking whatever state its own words support.

What would move the reading: a share of support contacts handled by the assistant, headcount, or the support line's next quarter against volume. The company that said the most about AI absorption now says the least, and the one line that was supposed to show it has stopped moving.

Concentrix enters with two quarters and five channels: iX Suite product revenue (iX Hello priced per automated contact, iX Hero per seat), services revenue pulled through by AI, services revenue displaced by AI automation (a toll, the first on a vendor), non-billable headcount displaced by its own AI tools, and the spend on building the products. The product revenue and the build are new money; the pull-through, the displacement and the internal productivity are existing lines whose size AI changes. The methodology now records that a displaced vendor's lost revenue is a toll, sized as what it would have billed at the prior-year rate, so that flow totals stay positive.

Q1 fiscal 2026. Revenue grew 1.9% in constant currency to $2.50bn. Management gave levels for the product: recurring revenue of $60mn at the end of fiscal 2025, a target of at least $100mn by the end of fiscal 2026, 60 enterprise deals closed, contract value for AI-including solutions more than doubled quarter on quarter, wins with technology up 61%. On displacement, the CFO put the technology vertical's decline of about 6% in constant currency half on underlying volumes with a little bit of impact of automation and half on offshore mix; the CEO said AI deals initially compress existing revenue and margin before scaling. The cost side named internal efficiencies and a cost program of approximately $40mn of annualized savings without AI, and the 10-Q attributed every cost movement to currency, severance, volumes and wages. Internal productivity took an unknown motive.

Q2 fiscal 2026. Revenue grew 0.6% in constant currency; guidance was cut on faster offshoring (a headwind of about 3%) and clients ceasing to support customer segments (about 1%), which the CEO said is not volume being automated. Three channels stepped. Pull-through moved from directional to quantified on the influenced-revenue share, 11%. Internal productivity moved from described to directional and from unknown to efficiency on the revenue-per-head tell, 14%, which the CEO attributed to deploying the company's own AI tools and reducing non-billable headcount; the saving is not visible in SG&A, which rose above its prior-year share because $17mn of restructuring sits in it, and the quarter's severance of $42mn affected roughly 20,000 employees, attributed in the filing to cost reduction and offshoring. The displacement toll moved from bounded to described: no split this quarter, volumes consistent, automation at expected levels, and an iX Hello deployment lowering revenue at first.

Sizes. The product revenue is read as one quarter of a run-rate between management's recurring revenue levels, $15mn to $25mn then $19mn to $30mn. The pull-through applies an assumed excess growth rate to the influenced revenue, $7.9mn to $35mn. The displacement range, $962k to $29mn, rests on the Q1 split and assumed automation shares; it is small against revenue because management's own bound is small. Internal productivity, $9.6mn to $29mn, applies management's gain to an assumed support personnel base and an assumed AI share of the gain. The build, $6.8mn to $23mn, is bracketed below the product revenue because management says the AI investments were profitable by the end of fiscal 2025.

What the pair now shows. The buyers report falling support cost per unit and the ledger infers their displaced vendor spend; the vendor reports that what it loses to AI is small and short-lived against what AI-led deals bring in, and that the revenue leaving its income statement is going offshore and to clients dropping customer segments. Both accounts are management's words with no dollar attached. Concentrix's quarter ended 2026-08-31 has a release on EDGAR and no 10-Q yet; the scheduler takes it when the 10-Q is filed.

This entry opens the ledger. Each company-quarter was read in full (the call, the release, the 10-Q), every channel through which AI money arrives or leaves was registered, and sized in two passes: first only where management gave a number or a bound that applies to a reported line, then a ballpark for every remaining channel management had at least described, built from reported lines and stated reference classes with every assumption written down. The two kinds of number are kept apart on the page: solid for management's, outlined for the ledger's.

Booking, Q2 2026. The CFO put AI costs at a low single-digit share of technology spend, which against information technology expense of $263mn gives $2.6mn to $7.9mn for the quarter, and said they are not what drives that line's growth. He put traffic from large language models at significantly below 1% of room nights. Customer service cost per booking is said to be falling at a double-digit rate with no base given; the containing line came in $30mn below its prior-year share of revenue, which the 10-Q attributes to customer service efficiencies without mentioning AI, and the ledger's decomposition puts the AI-attributable saving at $3.1mn to $74mn. Marketing came in $39mn above its prior-year share of gross bookings on declining unpaid search traffic, which the CEO said an AI search feature probably caused; the ledger's range for the replacement cost is $3.4mn to $163mn. The cost program, raised to $650mn of expected run-rate savings, is a confound on every savings channel and is not credited to AI.

Expedia, Q2 2026. Management sized nothing in either quarter and disclosed less in the second: a dollar magnitude for AI in marketing and the service interaction counts were not repeated, and it said in its own words that its conversational features are not driving conversion right now. Technology and content personnel cost changed by −6.8% while licence, maintenance and cloud costs in the same line rose 17.8%; the 10-Q attributes the personnel change to a cost program, recorded as a confound. The ledger's ballparks put the marketing-allocation saving at $3.7mn to $22mn and bookings arriving from AI platforms at $2.2mn to $22mn; several channels stay inscrutable, including competition from AI planning products, where demand that never arrived has no base.

Datadog, Q2 2026, the first seller. The 10-Q bounds the AI-native cohort's contribution to growth, and with the year-ago share from an earlier call that gives cohort revenue of $149mn to $165mn, with the cohort's funding now mixed because hyperscaler labs were folded into it this quarter. The quarter's most material fact is forward-looking and has no slot yet: the largest customer renewed and is reducing usage from the third quarter. The ledger's ballparks for products that monitor customers' AI workloads, in-platform agents and ordinary customers' AI-driven usage are $5.6mn to $56mn, $1.1mn to $22mn and $10mn to $82mn; the labs sub-channel overlaps the cohort and is not additive.

Read together: where management sizes a channel it is small at the buyers and large at the seller, and the ledger's inferences do not change that order. The savings the buyers describe, where the lines support them, appear to stay inside their own margins. Three companies is a test of the procedure more than a finding; the next quarters, expected in early November, are where steps begin.

Airbnb enters with two quarters, Q1 and Q2 2026, and ten channels: a vendor bill, a coding-tools bill inside it, internal build investment, community support displaced by the AI assistant, engineering output per head, AI tools for hosts, guest-facing AI search and generated content, AI ranking and personalization, demand arriving through AI platforms, and competition from AI planning products. Each channel carries a novelty tag: the vendor bill, AI search and AI-platform demand are new money; support and engineering are existing lines whose cost AI changes; host tools and ranking are machine-learning activities now called AI.

Q1 2026. Management gave operating numbers and no dollars: 60% of code written by AI, 40% of issues raised through the AI assistant resolved without a human agent, and customer support cost per booking down 10% year over year, which the CEO tied to improving AI customer support. The 10-Q had no AI passage and attributed every expense line's movement to headcount, compensation, marketing and customer relations. Operations and support came in −$31mn against its prior-year share of revenue, the shape the support claim predicts, while the segment table's third-party services line rose $22mn. The ledger's decomposition put the AI support saving at $3.4mn to $27mn.

Q2 2026. The 10-Q attributes a figure to AI: a $17mn decrease in third-party service provider costs from lower agent contact volume due to AI in community support, against increases in payroll, customer relations and insurance in the same line. The support channel's strength steps from inferred to reported; the ledger's own decomposition for the quarter, $6.4mn to $52mn, brackets the filing's number. Under the ledger's convention the filing's figure is a floor, since it is a year-over-year fall in one component rather than the saving against the prior-year cost per booking at this quarter's higher volume. Management's operating numbers moved the same way: 45% of issues starting with the assistant resolved without a human and cost per booking down 16%.

Two motives stepped in Q2. Engineering output per head and the coding-tools bill move from narrative-defensive to efficiency on the CFO's words that headcount need not grow at past levels and that product development was among the cost efficiencies behind margin expansion; the 10-Q still attributes the line's whole increase to payroll on higher headcount, so the displacement is growth avoided rather than cost removed, and the size stays the ledger's estimate, $941k to $60mn. Demand from AI platforms, described in Q1 when the CEO said ChatGPT traffic converts better than Google traffic and that the platform had closed its app program, went unmentioned in Q2.

The spend side is described in both quarters and never sized: an expense that will ramp in Q1, a material increase assumed in guidance in Q2, inference costs the CEO calls de minimis relative to returns. The ledger's ballparks are a share of the segment line that holds data hosting for the vendor bill and a share of Product development for the build. The revenue channels (host tools, AI search, personalization) remain described with template estimates and no measured lift; competition from AI planning products is named as a risk and stays inscrutable.