AWS's AI business passed a run rate, from over a quarter earlier, inside AWS revenue of , up . The ledger's quarter is , one compute-layer channel paid by AI labs and enterprises; management states no split and the ledger makes none. The backlog rose to after Anthropic expanded its commitment by more than .
The company kept funding the labs that pay it: and into Anthropic preferred stock, the second under the capacity-linked facility, and into OpenAI with after the quarter. The Anthropic preferred stock was marked up on Anthropic's fundings, and the release puts non-operating other income of primarily on Anthropic; all of it is non-operating, left out of totals, and partly set by rounds the company joined. Funding on the lab money reads mixed: the company's purchases and the labs' outside rounds both stand behind it, and no source states a payment drawn from the facility.
Cash capital spending was , and the 2026 expectation rose to about on memory prices, the majority for AI and AWS by the CEO's words; the AI share, , capped by AWS's share of net additions to property and equipment (), is capital and left out of totals. AWS depreciation was , and the ledger's AI part . Interest expense rose to on new notes, which the 10-Q says were issued for general corporate purposes (claim c55); no source names AI as the reason for the debt, so the interest is a funding note on capital spending rather than a channel. Technology and infrastructure carries an energy contract gain of , taken out of the bases for shares.
Steps from Q1. Channels open: the company's own frontier model, shown and left out of totals as overlapping depreciation; Forward Deployed Engineering, an undated amount of left unsized until a source shows the team at work; and AI in fulfillment robotics, a passing mention. AI ad tools move from described to directional and from exploratory to offensive on Ads Agent's measured advertiser costs; engineering moves from directional to described and from efficiency to exploratory, the Q1 measure not repeated; third-party agent referrals and Health AI go unmentioned. On the buyer side the measures are rates comparing users with non-users ( more per order, higher conversion on prompts), not lifts, and the payroll decline in general and administrative costs is not tied to AI (claim c57).
Sized channels against the income statement, Q2 2026
10 of 17 channels sized
Each blue mark is one channel's dollars for the quarter; a bar is the range of an estimate. Grey marks are the company's reported lines. The distance between them is the point: how large the AI channel is next to the line it sits in.
Total net salesreported line
$200.61bn
Total operating expensesreported line
$173.15bn
Cost of salesreported line
$95.78bn
Sales and marketingreported line
$11.70bn
General and administrativereported line
$2.79bn
Gains on the investments in Anthropic (conversions and upward adjustments)revenue in · reported in the filing
Cash capital expenditures on AI data centers, chips, servers and networkingspend · our inference
AWS AI revenue: the annual revenue run rate management gives for the AI business of AWSrevenue in · our inference
Depreciation of the AI infrastructure build in the AWS segmentspend · our inference
The company's own frontier model development (Nova)spend · our inference
Rufus and Alexa+ shopping (Alexa for Shopping): sales lifted by the AI shopping assistantrevenue in · our inference
AI tools for advertisers (Creative Agent, Ads Agent) and the advertising they bringrevenue in · our inference
The company's own engineering work done with agentic coding toolscost displaced · our inference
Sponsored prompts and ads inside the AI shopping assistantrevenue in · our inference
AI in fulfillment robotics (conversational direction of Proteus)cost displaced · our inference
$1mn$10mn$100mn$1bn$10bn$100bn$1000bn
AI channel, dollars for the quarter low to high of an estimate reported lineLog scale: each gridline is ten times the one before.
New money and old money, Q2 2026
4 new12 expanded1 relabelled
Each channel is tagged once for whether its money existed before language models, from the company's annual report and call at the start of the period. A bar splits one flow's sized dollars by that tag. The incremental total is the part that would not be there without the models: a new channel counts in full, an expanded one only for what AI added, a relabelled one at zero.
Flow, sized total
Split by novelty
Incremental total
Paid for AI$2.37bn to $4.27bn sized
expanded $2.37bn to $4.27bn
Incremental total $2.37bn to $4.27bnpoint $3.32bn
Cost displaced by AI$19mn to $523mn sized
expanded $19mn to $523mn
Incremental total $19mn to $523mnpoint $131mn
Revenue arriving through AI$5.09bn to $11.68bn sized
new not sizedexpanded $5.09bn to $11.68bnrelabelled not sized
Incremental total $95mn to $11.68bnpoint $551mn$6.19bn in 1 channel has no traced baseline
The sized total counts every channel the company credits to AI, including relabelled money that existed before language models and the ledger's own estimates for it. The incremental total counts relabelled channels at zero. A flow is split by layer where its dollars sit at more than one: end use, compute sold to builders, and hardware. The same dollar can be a buyer's spend, a cloud's revenue and a chipmaker's revenue, so the layers are never added together.
Paid for AI4 channels · $2.37bn to $4.27bn sized · $2.37bn to $4.27bn incremental · 1 not sized
other
Cash capital expenditures on AI data centers, chips, servers and networking
12.1% to 19.1% of the quarter’s revenue; capital spending on the build, not added into totals (its depreciation is)
our inferencedisclosure: quantified· motive: offensive· before LLMs: expanded· layer: compute
Cash capital spending was against , primarily for AWS and generative AI by the CFO's words (claim c33). The 2026 expectation rose to about from about , on higher memory costs, and the CEO says the majority supports AI and AWS (claims c35, c34). The trailing year's rise of primarily reflects AI investments and free cash flow turned to an outflow of (claim c36). Funding is mixed: operating cash flow of and long-term debt proceeds of , and the CFO ties this year's debt to AWS growth, not to AI by name (claim c41). Interest expense rose to from , against for all of 2024, on debt including new note issuances and finance leases, and the notes include ones issued for general corporate purposes (claims c42, c55). No source names AI as the reason for the debt, and its proceeds could as well have funded the lab stakes, so the interest is noted here and not read as an AI channel. AWS net additions to property and equipment were of , a share of that caps the AI share. The size, , is the AWS share times an assumed AI share of it, capital, left out of totals.
Evidence: 11 quotes, 14 figures, 2 confounds, 4 from before coverage
Cash capital spending, which the CFO says primarily relates to AWS and generative AI and the release says rose primarily on investments in artificial intelligence; it also funds the fulfillment network. Capitalized: it reaches the income statement as depreciation, read on its own channel, so this channel is traced and left out of totals. The payees include NVIDIA (whose revenue it is on the Nvidia ledger), memory and component suppliers, builders and power providers; Trainium is the company's own chip. Funded from operating cash flow and, in 2026, new long-term notes that the 10-Qs say were issued for general corporate purposes; the interest on them is read in this channel's funding note, not as an AI channel, since no source names AI as the reason for the debt. The AI share is capped by AWS's share of net additions to property and equipment in the segment note. Layer compute: the filings put AWS first; the same build also runs the company's own AI products (the store's assistant, its own frontier model), and the fulfillment network's capital spending is outside the AI share.
Why this motive
Capacity built while demand exceeds supply, for priced AI services (claims c33, c34).
Before LLMs: expanded
Capital spending on data centers predates the build: total capital spending was in the first quarter of 2024, when the CFO already tied the coming step-up mostly to AWS and generative AI. No AI part of that quarter can be traced, so the baseline is total capital spending, fulfillment and non-AI AWS included: a wider scope than the channel, which makes the size above it a low reading of the increment. The size is the whole of an activity that existed before.
“We do see, though, on the CapEx side, that we will be meaningfully, you know, stepping up our CapEx, and the majority of that will be in our to support AWS infrastructure and specifically generative AI efforts.”
AWS segment net additions to property and equipment · 2026-CQ2
Net additions to property and equipment, all segments · 2026-CQ2
AWS share of net additions to property and equipment · 2026-CQ2
Proceeds from long-term debt · 2026-CQ2
Net cash provided by operating activities · 2026-CQ2
Interest expense · 2026-CQ2
Interest expense, prior year · 2025-CQ2
Interest expense, 2024 · FY2024
What else could explain it
line composition: Capital spending also funds fulfillment capacity and general compute.
other: Part of the rise is memory and other component inflation, which management names as the reason for the higher plan.
Quotes
“Now turning to our cash CapEx, which is $53.1 billion in Q2. This primarily relates to AWS and generative AI as we invest to support strong customer demand.”
“We now believe we will spend approximately $220 billion in cash CapEx in 2026. The higher cost of memory pushing this number up from our prior estimate of about $200 billion.”
“Free cash flow decreased to an outflow of $7.6 billion for the trailing twelve months, driven primarily by a year-over-year increase of $66.1 billion in purchases of property and equipment, net of proceeds from sales and incentives. This increase primarily reflects investments in artificial intelligence.”
“Cash capital expenditures were $31.4 billion and $53.1 billion during Q2 2025 and Q2 2026, and $55.6 billion and $96.3 billion for the six months ended June 30, 2025 and 2026, which primarily reflect investments in technology infrastructure (the majority of which is to support AWS business growth) and in additional capacity to support our fulfillment network, both of which we expect to increase in 2026.”
“We're on pace with the capacity build that we talked about a few quarters ago, where we said we expect to have double the power capacity by the end of 2027 that we had in 2025, and we continue to be on that track.”
“Interest expense was $516 million and $1.3 billion during Q2 2025 and Q2 2026, and $1.1 billion and $2.1 billion for the six months ended June 30, 2025 and 2026, and was primarily related to debt, including new issuances of Notes, and finance leases.”
“As of June 30, 2026, we had $132.1 billion of unsecured senior notes outstanding (the "Notes"), including foreign currency-denominated Notes issued for general corporate purposes, the carrying values of which are subject to foreign exchange rate fluctuations.”
Depreciation of the AI infrastructure build in the AWS segment
1.2% to 2.1% of the quarter’s revenue
Incremental total: counts in full.
our inferencedisclosure: described· motive: offensive· before LLMs: expanded· layer: compute
AWS depreciation was against ; the 10-Q again puts the rise in technology and infrastructure down to infrastructure spending including depreciation, without naming AI (claim c43). The line also carries an unrealized gain of on energy contracts, mostly in AWS (claims c44, c45). The CEO says most AI capacity is contracted for terms about as long as the servers' useful lives (claim c11). The size, , is AWS depreciation above its 2024 quarterly average times the AI share the release's bound gives: the increase in the build primarily reflects investments in AI (claim c36), read through the words table. Applying that bound to depreciation is the ledger's step, so the size is inferred.
Evidence: 6 quotes, 5 figures, 1 confound, 3 from before coverage
Depreciation of the servers, accelerators, networking and data centers purchased for the build, which the 10-Q says drives the rise in technology and infrastructure costs and allocates mostly to AWS. The filing never splits depreciation by purpose; the company's attribution of the capital spending to AI is what ties it to AI, and the AI share is assumed. Read against AWS segment depreciation. Layer compute: the cost of providing AWS capacity, which also runs the company's own AI products.
Why this motive
Carried: the depreciation of capacity built for priced AI services (claim c33).
Before LLMs: expanded
AWS depreciation was in 2024, in an average quarter, and the anchor annual report shortened server lives because of the pace of AI and machine learning development. The size is AWS depreciation above that quarterly level, times an assumed AI share. The size is the change AI made, not the whole line.
“We do see, though, on the CapEx side, that we will be meaningfully, you know, stepping up our CapEx, and the majority of that will be in our to support AWS infrastructure and specifically generative AI efforts.”
“These two changes above are due to an increased pace of technology development, particularly in the area of artificial intelligence and machine learning.”
AWS segment depreciation and amortization · 2026-CQ2
AWS segment depreciation and amortization, prior year · 2025-CQ2
Technology and infrastructure · 2026-CQ2
Technology and infrastructure, prior year · 2025-CQ2
Net unrealized gain on energy contracts recorded within technology and infrastructure · 2026-CQ2
Reported line it is matched to
AWS segment depreciation was against a year earlier.
2026-CQ2: 2025-CQ2:
What else could explain it
line composition: AWS depreciation covers all AWS property.
Quotes
“Now turning to our cash CapEx, which is $53.1 billion in Q2. This primarily relates to AWS and generative AI as we invest to support strong customer demand.”
“The increase in technology and infrastructure costs in Q2 2026 and for the six months ended June 30, 2026, compared to the comparable prior year periods, is primarily due to an increase in spending on infrastructure, including depreciation and amortization.”
“The servers currently have a useful life of at least five to six years, and most of our AI capacity these days is being contracted for at least five-year terms.”
“The impact of these fair value measurements on our consolidated statements of operations was not significant in Q2 2025 and for the six months ended June 30, 2025, and resulted in net unrealized gains of $551 million in Q2 2026 and $599 million for the six months ended June 30, 2026, recorded within "Technology and infrastructure" and primarily impacting our AWS segment.”
“Second, we recorded a separate benefit of approximately $600 million related to the change in fair value measurement of energy contracts subject to derivative accounting. This primarily impacts the AWS segment.”
“Free cash flow decreased to an outflow of $7.6 billion for the trailing twelve months, driven primarily by a year-over-year increase of $66.1 billion in purchases of property and equipment, net of proceeds from sales and incentives. This increase primarily reflects investments in artificial intelligence.”
The company's own frontier model development (Nova)
0.17% to 1% of the quarter’s revenue; overlaps another channel, not added into totals
our inferencedisclosure: described· motive: exploratory· before LLMs: expanded
The channel opens this quarter. The CEO says the company is pursuing its own frontier model, first for control over the cost of its own consumer applications and of customers' inference (claims c46, c47). No cost is given. The size, , is an assumed share of technology and infrastructure before the energy gain and is the ledger's. That line contains the server and data center depreciation read on ai-infrastructure-depreciation, so training compute would be counted twice: the reading lists that channel in overlaps and is shown, left out of totals.
Evidence: 3 quotes, 1 figure, 1 confound, 2 from before coverage
Training compute, data and staff for the company's own frontier model, which the CEO says it is pursuing for control over cost, priorities and speed. No cost is given; it sits in technology and infrastructure, whose server and data center depreciation is also the base of ai-infrastructure-depreciation, so this channel lists that one in overlaps and is shown but left out of totals. Layer end-use: the CEO puts the company's own consumer applications first; the model is also offered to customers on Bedrock.
Why this motive
Pursued for control over cost, priorities and speed, with no measure (claim c46).
Before LLMs: expanded
At the anchor the company already built first-party Titan models and its engineering payroll sat in technology and infrastructure, in 2024. A budget redirected to AI work is expanded, sized as a level with no traceable quarter before AI. The size is the whole of an activity that existed before.
“We expect spending in technology and infrastructure will increase over time as we add computer scientists, designers, software and hardware engineers, and merchandising employees.”
“In the last few months, Bedrock's added Anthropic's Claude 3 models, the best-performing models on the planet right now, Meta's Llama 3 models, Mistral's various models, Cohere's newest models, and new first-party Amazon Titan models.”
Technology and infrastructure before the energy contract gain · 2026-CQ2
What else could explain it
line composition: Model work may be booked with the capacity the company also offers to customers, so part may sit in AWS capital spending.
Quotes
“All that said, we are pursuing our own frontier model, and we're doing it for a few reasons. First of which is it just gives us additional control over cost.”
“The impact of these fair value measurements on our consolidated statements of operations was not significant in Q2 2025 and for the six months ended June 30, 2025, and resulted in net unrealized gains of $551 million in Q2 2026 and $599 million for the six months ended June 30, 2026, recorded within "Technology and infrastructure" and primarily impacting our AWS segment.”
AWS Forward Deployed Engineering: AI engineers embedded with customers
Not sized
The amount is undated and its announcement may postdate quarter end; no stored source shows the team operating in the quarter, so the undated band is applied from the first quarter a source does.
inscrutabledisclosure: quantified· motive: exploratory· before LLMs: expanded· layer: compute
The channel opens this quarter. The release announces an investment of to create AWS Forward Deployed Engineering, AI engineers embedded with customers (claim c48). The amount has no period, and the release lists it among highlights since the prior earnings report, a window that runs past quarter end; no source shows the team at work in the quarter. The undated band applies from the first quarter a source does, so the channel is unsized here.
Evidence: 1 quote, 1 figure, 2 confounds, 1 from before coverage
A team of AI engineers embedded with customers to co-develop and deploy agentic AI, announced in the Q2 release, among highlights since the prior earnings report, with an investment of a stated amount and no period. That window runs past the end of the quarter and no source shows the team at work in Q2, so the amount converts to a quarter through the undated band only from the first quarter a source does. Layer compute: a cost of marketing and deploying AWS's AI services; what the team builds is the customers' end-use.
Why this motive
A team created to get customers' agentic AI live, with no result yet (claim c48).
Before LLMs: expanded
AWS sales staff, who work with its customers, were already paid through sales and marketing at the anchor; no headcount or cost is given for them. The team is sized as a level of staff with no traceable quarter before AI. The size is the whole of an activity that existed before.
“Sales and marketing costs include advertising and payroll and related expenses for personnel engaged in marketing and selling activities, including sales commissions related to AWS.”
Investment announced to create AWS Forward Deployed Engineering, no period · as-of 2026-07-30
What else could explain it
relabel: The team may be assembled from solutions and services staff the company already employed.
one time item: The amount may cover more than a year or include customer credits.
Quotes
“Announced an investment of $1 billion to create AWS Forward Deployed Engineering, a team of AI engineers embedded directly with customers to co-develop and deploy agentic AI solutions in days rather than months.”
Cost displaced by AI2 channels · $19mn to $523mn sized · $19mn to $523mn incremental
engineering · cheap to verify
The company's own engineering work done with agentic coding tools
0.01% to 0.2% of the quarter’s revenue
Incremental total: counts in full.
our inferencedisclosure: described· motive: exploratory· before LLMs: expanded
Kiro is described as built first for the company's own engineers (claim c22); no measure is repeated. Employees were against , and the 10-Q explains lower general and administrative costs by lower payroll without naming AI (claim c57). The base is technology and infrastructure before the energy contract gain, . The size, , is the ledger's.
Evidence: 2 quotes, 5 figures, 1 confound, 1 from before coverage
Software engineering inside the company done with agentic coding tools, Kiro among them, which management says began as a tool built for its own engineers. The CEO gives one project's staffing before and after; no cost line is shown to move, and the payroll declines the filings report are not tied to AI.
Why this motive
AI named for internal use with no measure and no line moving this quarter (claim c22): the quarter is not silent on the channel, so the motive is read from its own words, a passing mention, rather than carried from the Q1 measure.
Before LLMs: expanded
Engineering payroll sat in technology and infrastructure at the anchor, in 2024. Putting LLM coding tools into that work is expanded; the size is a saving against the prior rate, an increment. The size is the change AI made, not the whole line.
“We expect spending in technology and infrastructure will increase over time as we add computer scientists, designers, software and hardware engineers, and merchandising employees.”
Technology and infrastructure before the energy contract gain · 2026-CQ2
Net unrealized gain on energy contracts recorded within technology and infrastructure · 2026-CQ2
Employees, full-time and part-time, at quarter end · as-of 2026-06-30
Employees, full-time and part-time, a year earlier · as-of 2025-06-30
What else could explain it
transformation program: Payroll declines the 10-Q reports are not attributed to AI.
Quotes
“Some of this is born out of what customers tell us they wish they had and they want to be using. Some of it is born out of just needing to provide those capabilities to ourselves inside Amazon. Kiro, which is our agentic coding service, is an example of that.”
“The impact of these fair value measurements on our consolidated statements of operations was not significant in Q2 2025 and for the six months ended June 30, 2025, and resulted in net unrealized gains of $551 million in Q2 2026 and $599 million for the six months ended June 30, 2026, recorded within "Technology and infrastructure" and primarily impacting our AWS segment.”
AI in fulfillment robotics (conversational direction of Proteus)
0% to 0.06% of the quarter’s revenue
Incremental total: counts in full.
our inferencedisclosure: described· motive: exploratory· before LLMs: expanded
The channel opens this quarter. The release says employees can now direct the Proteus robot in plain language using AI (claim c49). The CFO credits fulfillment gains to inventory placement and robotics and automation integral for decades, without naming AI (claim c50). Fulfillment expense was against . The size, , is the ledger's, an increment against the prior rate.
Evidence: 2 quotes, 2 figures, 1 confound, 1 from before coverage
AI named in the fulfillment network: employees can now direct the Proteus robot in plain, conversational language. The cost-to-serve gains management describes are credited to inventory placement, consolidation and robotics and automation, not to AI, and are confounds here.
Why this motive
AI named as a cause with no measure and no line moving: a passing mention (claim c49).
Before LLMs: expanded
Robotics and automation in the fulfillment network were already a cost lever at the anchor, with no language interface named; fulfillment expense was in 2024. A conversational AI interface put into fulfillment work during coverage is an LLM tool deployed in a cost line the anchor shows, so the channel is expanded, sized as the saving it adds against the prior rate, though no saving is measured. The size is the change AI made, not the whole line.
“Within our fulfillment network, we are focused on investing in our inbound network, streamlining and standardizing process paths, and adding robotics and automation.”
transformation program: Robotics, automation and network redesign that management does not attribute to AI.
Quotes
“Introduced the next-generation of Proteus, an autonomous robot that assists Amazon fulfillment center employees by moving goods up to 1,300 pounds, reducing heavy lifting and further increasing safety. Using AI, employees can now direct Proteus with plain, conversational language.”
“We're expanding our deployment of robotics and automation, which have been integral to our operations for decades. We're retrofitting our facilities with our latest generation technology, and we expect to more than double our fleet of robotic arms, like Cardinal and Sparrow, in 2026.”
Revenue arriving through AI11 channels · $5.09bn to $11.68bn sized · $95mn to $11.68bn incremental · 6 not sized
product revenue
AWS AI revenue: the annual revenue run rate management gives for the AI business of AWS
2.5% to 4.7% of the quarter’s revenue
Incremental total: counts at zero at the point and in full at the high end, with no traced baseline.
our inferencedisclosure: quantified· motive: offensive· before LLMs: expanded· layer: compute
The release puts AWS's AI business above a run rate, from over a quarter earlier, and the CEO says well over (claims c2, c1, c3). AWS revenue was , up and on the quarter; the 10-Q again explains the growth by customer usage without naming AI (claim c6). The size, , is a quarter of the run rate scaled for what was in force, the ledger's; AWS revenue is its ceiling. The chips run rate of over includes non-AI chips and sits inside AWS revenue. The CEO says AI margins track the core business at the same stage (claim c5). Management states no parts of the metric, so it is one channel and counts in totals. The counterparty is mixed: AI labs, of which Anthropic and OpenAI are named (claims c8, c56), and enterprises, start-ups and governments. Funding is mixed. The company has put money into the lab payers: into Anthropic preferred stock in the quarter ( and , the second under the facility, which leaves available), and an OpenAI preferred stock investment of on the balance sheet at quarter end, of it in the quarter, with more after it (claims c12, c14, c58, c15, c16). The labs' outside rounds appear in the marks: the Anthropic preferred stock was marked up on Anthropic's fundings, and the second purchase was an option to participate in Anthropic's equity financings (claims c52, c14); the sources do not size OpenAI's outside rounds. The facility opens as AWS delivers compute (claim c13), the vendor-financed case, but no source states that any payment to AWS comes from it, so the funding stays mixed. Enterprises and governments pay largely from operating cash flow, start-ups largely from investor capital.
Evidence: 16 quotes, 14 figures, 2 confounds, 4 from before coverage
Management's headline AI revenue metric, an annual revenue run rate for AWS's AI business, given on the Q1 call and in the Q2 release. The sources do not define it: it appears to span AI accelerator capacity rented to AI labs, Bedrock inference and agent services, SageMaker, and the turnkey agent applications. A run rate is not the quarter's revenue; the size is a quarter of the stated level scaled for what was in force. AWS segment revenue is a ceiling on it, never a level. Management states no parts of the metric, so it is one channel (the methodology rule on umbrella metrics and components), counted in totals, with a mixed counterparty: AI labs (Anthropic and OpenAI are named) and enterprises, start-ups and governments. The lab, Bedrock and agent application channels are registered inside it as described, unsized readings of what the sources say about each group of payers; nothing in them is added to totals. Funding is mixed: the labs pay from outside rounds and from money the company has put into them (preferred stock in both, and a facility for Anthropic that opens as compute is delivered), enterprises largely from operating cash flow, start-ups largely from investor capital. Layer compute: capacity and hosted model access provided to builders, with the agent applications, end-use products, inside the same unsplit figure.
Why this motive
Separately priced AI capacity and services, the run rate climbing significantly on the quarter (claims c1, c5).
Before LLMs: expanded
At the anchor AWS already reported an AI revenue run rate, described without a figure, against an AWS run rate of and quarterly AWS revenue of , and rented NVIDIA instances, Trainium and SageMaker for machine learning. No level was given, and accelerator capacity and machine learning services predate LLMs, so the tie-break gives expanded with no traceable quarter before AI. The lab payers alone would read new (the payer exists only because of LLMs) and the machine learning services expanded; with no stated parts the less durable tag is taken for the whole. The size is the whole of an activity that existed before.
“We see considerable momentum on the AI front, where we've accumulated a multi-billion dollar revenue run rate already.”
“At the bottom layer, which is for developers and companies building models themselves, we see excitement about our offerings. We have the broadest selection of NVIDIA compute instances around, but demand for our custom silicon, Trainium and Inferentia, is quite high, given its favorable price-performance benefits relative to available alternatives.”
“During the first quarter, we saw growth in both generative AI and non-generative AI workloads across a diverse group of customers and across industries, as companies are shifting their focus towards driving innovation and bringing new workloads to the cloud.”
AWS's AI business annual revenue run rate, floor · as-of 2026-06-30
AWS segment net sales · 2026-CQ2
AWS segment net sales, prior year · 2025-CQ2
AWS segment net sales growth · 2026-CQ2
AWS segment net sales, change over the prior quarter · 2026-CQ2
AWS annualized revenue run rate · as-of 2026-06-30
Annual revenue run rate of the chips business, floor · as-of 2026-06-30
AWS's AI revenue run rate, floor · as-of 2026-03-31
Investments in Anthropic preferred stock in the quarter (Series G plus Series H) · 2026-CQ2
Anthropic financing facility remaining after the Series H investment · as-of 2026-06-30
Investment in OpenAI Series C preferred stock on the balance sheet · as-of 2026-06-30
Investment in OpenAI Series C preferred stock in the quarter · 2026-CQ2
OpenAI commitment funded after the quarter · as-of 2026-07-31
Upward adjustment to the Anthropic nonvoting preferred stock on Anthropic's fundings · 2026-CQ2
What else could explain it
relabel: The AI business is not defined, so a widening of its scope between quarters cannot be ruled out.
other: Part of the revenue is paid by AI labs funded in part by the company itself; the same lab money appears as revenue at Microsoft and the accelerators serving it as revenue at Nvidia, and nothing is netted.
Quotes
“Our AI revenue run rate climbed significantly quarter-over-quarter, and is now also over $25 billion, growing triple-digit percentages year-over-year.”
“As I mentioned in my opening comments, we see the AI business following very much the same type of margin trajectory that we saw in the core business before, and it's a little bit ahead of that pace that we saw.”
“AWS sales increased 37% in Q2 2026, and 33% for the six months ended June 30, 2026 compared to the comparable prior year periods. The sales growth primarily reflects increased customer usage, partially offset by pricing changes primarily driven by long-term customer contracts.”
“In Q2 2026, AWS and Anthropic announced an expansion of the strategic collaboration and existing multi-year commitment by more than $100.0 billion over 10.0 years, which includes contractual obligations related to the performance of AWS chips.”
“In Q1 2026, we and OpenAI entered into (i) a commercial arrangement primarily for the provision of AWS cloud services, which includes the use and performance of AWS chips, and (ii) a joint collaboration agreement pursuant to which certain services using OpenAI models will be made available to the Company and on AWS.”
“Under this financing arrangement, in Q2 2026, we exercised our option to participate in subsequent Anthropic equity financings by investing $5.0 billion in Anthropic Series H nonvoting preferred stock, which reduced the amount available under the facility to $15.0 billion.”
“At inception, there is no amount available to be drawn against and as we reach certain delivery milestones of compute capacity under the amended commercial arrangement, amounts under this facility are made available for Anthropic to draw upon at its discretion.”
“We recorded upward adjustments of approximately $50.5 billion in Q2 2026 and $62.8 billion for the six months ended June 30, 2026 to our nonvoting preferred stock in "Other income (expense), net" to reflect observable changes in price related to Anthropic's fundings.”
AWS capacity and Trainium provided to frontier AI labs (Anthropic, OpenAI)
Not sized
Management gives the labs' commitments, not the quarter's revenue, and states no lab share of the AI run rate. The lab money is sized inside aws-ai-revenue; a metric with no stated parts is not split by judgment (the methodology rule on umbrella metrics and components).
described, no sizedisclosure: quantified· motive: offensive· before LLMs: new· layer: compute
Anthropic expanded its AWS commitment by more than , and the backlog rose to from (claims c8, c7). The CEO calls the labs one end of a barbell consuming gobs of compute (claim c10). No revenue is given; the lab money is sized inside the AWS AI run rate, and this reading is unsized. In the quarter the company invested and then in Anthropic preferred stock, the second under the facility that opens as AWS delivers compute, which leaves available, and in OpenAI, with more after the quarter (claims c12, c14, c15, c16). The labs also pay from outside rounds: the Anthropic mark of follows Anthropic's fundings (claim c52); the sources do not size OpenAI's. Funding reads mixed: the facility drawn against delivery is the vendor-financed case, but no source states that any payment to AWS comes from it. The same labs rent from Microsoft and the chips are partly NVIDIA's revenue; nothing is netted.
Evidence: 12 quotes, 10 figures, 2 confounds, 3 from before coverage
AWS compute, Trainium and other cloud services paid for by the AI labs, of which Anthropic and OpenAI are named, under multi-year commitments the 10-Q describes as including contractual obligations on the performance of AWS chips. No revenue figure is given; the commitments sit in the remaining performance obligation. The lab money sits inside the AWS AI run rate (aws-ai-revenue), which carries the size; this channel is a described, unsized reading of what the sources say about the lab payers. The company has put money into both labs: it owns preferred stock in both, committed to purchase more OpenAI stock in the quarter OpenAI expanded its AWS commitment, and offers Anthropic a financing facility that opens as compute capacity is delivered. Under the lab rule (the methodology rule on lab funding) funding reads mixed; a facility drawn against delivery is the vendor-financed case, but no source states that any payment to AWS comes from it. The same lab money appears as revenue at Microsoft (msft frontier-lab-cloud-revenue) and the accelerators purchased to serve it as revenue at Nvidia; nothing here is netted against those ledgers.
Why this motive
Priced capacity provided under multi-year commitments, most AI capacity contracted for multi-year terms (claims c8, c11).
Before LLMs: new
At the anchor Anthropic was already training its models on Trainium and the company owned convertible notes in it, including a second note of ; no revenue from the lab was given. The payer is an AI lab, so the money is new whatever the product.
“At the bottom layer, which is for developers and companies building models themselves, we see excitement about our offerings. We have the broadest selection of NVIDIA compute instances around, but demand for our custom silicon, Trainium and Inferentia, is quite high, given its favorable price-performance benefits relative to available alternatives.”
“In Q3 2023, we invested in a $1.25 billion note from Anthropic, PBC, which is convertible to equity. In Q1 2024, we invested $2.75 billion in a second convertible note.”
Investment in Anthropic Series G nonvoting preferred stock · 2026-CQ2
Investment in Anthropic Series H nonvoting preferred stock under the financing arrangement · 2026-CQ2
Anthropic financing facility remaining after the Series H investment · as-of 2026-06-30
Investment in OpenAI Series C preferred stock in the quarter · 2026-CQ2
OpenAI commitment funded after the quarter · as-of 2026-07-31
Investment in OpenAI Series C preferred stock on the balance sheet · as-of 2026-06-30
Upward adjustment to the Anthropic nonvoting preferred stock on Anthropic's fundings · 2026-CQ2
What else could explain it
other: The company invests in both named labs while they commit to purchase its capacity; part of the revenue returns that money.
line composition: The commitments include core compute and services and run for up to a decade.
Quotes
“In Q2 2026, AWS and Anthropic announced an expansion of the strategic collaboration and existing multi-year commitment by more than $100.0 billion over 10.0 years, which includes contractual obligations related to the performance of AWS chips.”
“Continued gaining momentum with Trainium, with the two leading AI labs in the world, Anthropic and OpenAI, making multi-year, multi-gigawatt commitments”
“There is, on one end of the barbell, the AI labs are consuming gobs and gobs of compute, and there are a few runaway successful generative AI applications like Claude Code and ChatGPT.”
“The servers currently have a useful life of at least five to six years, and most of our AI capacity these days is being contracted for at least five-year terms.”
“At inception, there is no amount available to be drawn against and as we reach certain delivery milestones of compute capacity under the amended commercial arrangement, amounts under this facility are made available for Anthropic to draw upon at its discretion.”
“Under this financing arrangement, in Q2 2026, we exercised our option to participate in subsequent Anthropic equity financings by investing $5.0 billion in Anthropic Series H nonvoting preferred stock, which reduced the amount available under the facility to $15.0 billion.”
“We recorded upward adjustments of approximately $50.5 billion in Q2 2026 and $62.8 billion for the six months ended June 30, 2026 to our nonvoting preferred stock in "Other income (expense), net" to reflect observable changes in price related to Anthropic's fundings.”
Q1 2026quantified · described, no size · offensive
Q2 2026quantified · described, no size · offensive
product revenue
Bedrock, AgentCore, SageMaker and other AI services used by enterprises and start-ups
Not sized
Management gives growth, customer counts and token volumes, never a level or a share of the AI run rate. The money is sized inside aws-ai-revenue; a metric with no stated parts is not split by judgment (the methodology rule on umbrella metrics and components).
described, no sizedisclosure: direction only· motive: offensive· before LLMs: expanded· layer: compute
Bedrock customers spent more in Q2 than in all prior quarters combined, and its customer count rose (claims c18, c17). Enterprises are described as getting value in cost avoidance and productivity (claim c19). Bedrock offers the models of Anthropic, OpenAI and others, and under a joint collaboration services using OpenAI models are made available to the company and on AWS (claim c56); no source states what the company pays the model providers, so no spend channel to them is registered. No level is given; the money is sized inside the AWS AI run rate, and this reading is unsized. The counterparty and the funding are mixed: enterprises and governments paying from operating cash flow, start-ups largely from investor capital.
Evidence: 4 quotes, 1 confound, 5 from before coverage
Model access and inference on Bedrock, agent infrastructure (AgentCore, Bedrock Managed Agents), model building on SageMaker and AI accelerator instances rented by companies other than the frontier labs: enterprises, start-ups and governments. Management reports growth in customer spend, customer counts and token volumes, never a level. Amazon Connect's AI features are read here as the AI part of an existing service. It sits inside the AWS AI run rate (aws-ai-revenue), which carries the size; this channel is a described, unsized reading. Funding is mixed: enterprises and governments pay from operating cash flow, start-ups largely from investor capital. Bedrock offers models from Anthropic, OpenAI and other providers, and a joint collaboration with OpenAI makes services using OpenAI models available to the company and on AWS; no stored source states what the company pays the model providers, so no spend channel to them is registered.
Why this motive
Separately priced model and agent services, with spend in the quarter above all prior quarters combined (claim c18).
Before LLMs: expanded
At the anchor Bedrock already had customers, Adidas, Pfizer and Toyota among them, and SageMaker, the machine learning service, was offered alongside it; AWS revenue was in that quarter and no AI level was given. Token sales could not exist without LLMs, but the services also carry machine learning work that predates them, so the tie-break gives expanded, sized as a level with no traceable quarter before AI. The size is the whole of an activity that existed before.
“We see considerable momentum on the AI front, where we've accumulated a multi-billion dollar revenue run rate already.”
“Our managed end-to-end service has been a game changer for developers who are preparing their data for AI, managing experiments, training models faster, lowering inference latency, and improving developer productivity.”
“We serve developers and enterprises of all sizes, including start-ups, government agencies, and academic institutions, through AWS, which offers a broad set of on-demand technology services, including compute, storage, database, analytics, and machine learning, and other services.”
“In the last few months, Bedrock's added Anthropic's Claude 3 models, the best-performing models on the planet right now, Meta's Llama 3 models, Mistral's various models, Cohere's newest models, and new first-party Amazon Titan models.”
relabel: Machine learning services and accelerator instances that predate LLMs sit inside the same revenue.
Quotes
“Bedrock not only provides the best selection of leading models at superior performance and with the governance and security controls that companies need, it's also continuing to grow incredibly quickly.”
“hundreds of thousands of customers now use Bedrock, more customers were added in the last six months than in the first two years after launch, and customers spent more in Q2 than all prior quarters combined.”
“On the other end of the barbell are enterprises who are getting real value from AI in cost avoidance and productivity. These are things like automating customer service or business process automation or fraud or things like that.”
“In Q1 2026, we and OpenAI entered into (i) a commercial arrangement primarily for the provision of AWS cloud services, which includes the use and performance of AWS chips, and (ii) a joint collaboration agreement pursuant to which certain services using OpenAI models will be made available to the Company and on AWS.”
Q1 2026direction only · described, no size · offensive
Q2 2026direction only · described, no size · offensive
product revenue · cheap to verify
Turnkey AI agent applications: Kiro, Amazon Quick, AWS Transform, Continuum
Not sized
Management gives usage and customer names, no revenue and no share of the AI run rate. The money is sized inside aws-ai-revenue; a metric with no stated parts is not split by judgment (the methodology rule on umbrella metrics and components).
described, no sizedisclosure: direction only· motive: offensive· before LLMs: new
Kiro's usage rose on the quarter and it is said to be up to more cost-effective than alternatives; Continuum, a vulnerability agent, was released (claims c20, c21). Kiro and the work companion began as tools for the company's own staff (claims c22, c23); that answer is headed as the CFO's in the transcript although it reads as the CEO's. No revenue is given; the money is sized inside the AWS AI run rate, and this reading is unsized.
Evidence: 4 quotes, 1 figure, 1 confound, 1 from before coverage
Separately priced agent applications: the coding agent Kiro, the work companion Amazon Quick (rendered as Amazon Q in the call transcripts), the migration agent AWS Transform and the vulnerability agent Continuum. Management reports usage multiples, hours saved and customer names, no revenue. The revenue sits inside the AWS AI run rate (aws-ai-revenue), which carries the size; this channel is a described, unsized reading. Layer end-use: AI products offered to businesses. Code with tests and migrations are cheap to check.
Why this motive
Separately priced agent products with stated usage growth (claim c20).
Before LLMs: new
Amazon Q, a generative AI assistant for software development and internal data, reached general availability on the anchor call; no revenue was given. LLM agents could not exist before LLMs, so the money is new.
“The top of the stack are the Gen AI applications being built, and today we announced the general availability of Amazon Q, the most capable generative AI-powered assistant for software development and leveraging companies' internal data.”
Kiro cost-effectiveness against alternatives, up to · 2026-CQ2
What else could explain it
bundling: Some agents may be priced inside other AWS services or credits.
Quotes
“Coding agents are a good example, and there are several successful ones, including Claude Code, Codex, and our own spec-driven Kiro, which is up to 50% more cost-effective than others and tripled in usage quarter-over-quarter.”
“Some of this is born out of what customers tell us they wish they had and they want to be using. Some of it is born out of just needing to provide those capabilities to ourselves inside Amazon. Kiro, which is our agentic coding service, is an example of that.”
“We had so many people inside the company using it that they said, "Can't you actually find a way to make it much more productive and easier for us to manage our email, to manage our Slack communications, to manage our calendar, and to use all those things together?"”
Q1 2026direction only · described, no size · offensive
Q2 2026direction only · described, no size · offensive
customer cohort
Core AWS demand (CPU, storage, databases) attributed to customers' AI workloads
Not sized
AI is named as a driver of core growth beside enterprise migrations from on-premises, as a linkage with no rate or ranking of its own (claims c24, c25, c6); nothing separates AI's part (the methodology rule for AI named beside another cause, which holds at the upstream companies). AWS revenue, , is quoted as the ceiling.
described, no sizedisclosure: described· motive: exploratory· before LLMs: relabelled· layer: compute
The CEO says growth in AI drives core growth because post-training and agent tool use run on CPUs, and the CFO sees a strong linkage between AI spend and core consumption (claims c24, c25). Migrations are named too and no figure separates the part, so the channel is left unsized; AWS revenue, , is the ceiling it sits inside.
Evidence: 3 quotes, 2 figures, 1 confound, 2 from before coverage
The part of demand for non-AI AWS services (Graviton compute, storage, vector databases) that management attributes to customers' AI work: post-training, reinforcement learning and agent tool use running on CPUs. Meta's commitment to Graviton cores for agentic workloads is the one named instance (that money is Meta's spend on its own ledger). No figure separates the AI-attributed part. Layer compute: capacity provided to customers running AI work.
Why this motive
AI credited for growth in existing services with no measure of its part (claims c24, c25).
Before LLMs: relabelled
At the anchor AWS already offered compute, storage, databases and analytics, and growth was already credited to both generative AI and other workloads. Management gives a correlation, not a measure of the AI-attributed part, so the movement reads as relabelled.
“We serve developers and enterprises of all sizes, including start-ups, government agencies, and academic institutions, through AWS, which offers a broad set of on-demand technology services, including compute, storage, database, analytics, and machine learning, and other services.”
“During the first quarter, we saw growth in both generative AI and non-generative AI workloads across a diverse group of customers and across industries, as companies are shifting their focus towards driving innovation and bringing new workloads to the cloud.”
mix shift: Enterprise migrations from on-premises, which the CEO names, also drive core growth.
Quotes
“We're seeing strong growth across both AI and non-AI, what we call core, and growth in one is driving growth in the other. Growth in AI drives core because post-training reinforcement learning and agent tool use is mostly done on CPUs versus AI accelerators.”
“AWS sales increased 37% in Q2 2026, and 33% for the six months ended June 30, 2026 compared to the comparable prior year periods. The sales growth primarily reflects increased customer usage, partially offset by pricing changes primarily driven by long-term customer contracts.”
Q1 2026described · described, no size · exploratory
Q2 2026described · described, no size · exploratory
search discovery · cheap to verify
Rufus and Alexa+ shopping (Alexa for Shopping): sales lifted by the AI shopping assistant
0.02% to 0.7% of the quarter’s revenue
Incremental total: counts in full.
our inferencedisclosure: direction only· motive: product-defensive· before LLMs: expanded
Rufus and Alexa+ shopping merged into Alexa for Shopping; over customers used it in a year and active users rose (claims c26, c27). Users spend over more per order than non-users, and customers who tried Alexa+ sign up for Prime at nearly higher rates (claims c28, c29). Those gaps compare different shoppers and are not a lift. The size, , applies an assumed reach and lift to online stores revenue of and is the ledger's.
Evidence: 4 quotes, 4 figures, 2 confounds, 1 from before coverage
The free AI shopping assistant in the store, Rufus, merged with Alexa+ shopping into Alexa for Shopping in Q2. Management reports active users, engagement and the spend per order of users against non-users, never incremental sales. An unpriced feature lifting an existing revenue line, sized as an increment.
Why this motive
A free assistant; the spend-per-order gap is a measure but compares users with non-users, and the free-feature tell is the less durable reading (claims c26, c28).
Before LLMs: expanded
At the anchor customers already reached the stores through Alexa and the websites; online stores revenue was in 2024 and no assistant was named. The bookings existed; the feature changes them, so the size is an increment. The size is the change AI made, not the whole line.
“Customers access our offerings through our websites, mobile apps, Alexa, devices, streaming, and physically visiting our stores.”
Customers who used Alexa for Shopping in the last twelve months, floor · TTM as-of 2026-06-30
Spend per order of Alexa for Shopping users against non-users, floor · 2026-CQ2
Prime sign-up rate of customers who tried Alexa+ against others · 2026-CQ2
Online stores revenue · 2026-CQ2
What else could explain it
mix shift: Heavier shoppers are likelier to use the assistant, so the spend-per-order gap is largely selection.
other: Prime Day moved into Q2 in most large countries this year, which lifts the quarter's store sales.
Quotes
“Brought together Rufus and Alexa+ into Alexa for Shopping, an agentic AI shopping assistant that offers personalized recommendations, product comparisons, price history, and the ability to automate shopping through features like Price Alerts and Auto-Buy.”
“Over 350 million customers have used it in the last 12 months, and engagement accelerated in Q2, with active users nearly doubling and interactions up over 5x year-over-year.”
Q1 2026direction only · described, no size · product-defensive
Q2 2026direction only · our inference · product-defensive · $35mn to $1.41bn
marketing
Sponsored prompts and ads inside the AI shopping assistant
0.01% to 0.15% of the quarter’s revenue
Incremental total: counts in full.
our inferencedisclosure: direction only· motive: offensive· before LLMs: expanded
Shoppers who click a sponsored prompt convert more often and spend more than those who do not, as more shoppers discover products in the conversational experiences (claim c30). Advertising services revenue was against ; no revenue is given for the prompts. The size, , is an assumed share and is the ledger's.
Evidence: 1 quote, 4 figures, 1 confound, 2 from before coverage
Ad units offered to brands inside Rufus and Alexa for Shopping conversations: sponsored product and brand prompts. Management reports engagement and conversion rates, no revenue. The revenue sits in advertising services.
Why this motive
A priced ad unit with a measured conversion difference that management attributes to the prompts (claim c30).
Before LLMs: expanded
Sponsored ads were already the advertising business at the anchor, in 2024. The prompts are a new placement for the same sale, so the tie-break gives expanded, sized as the increment they add. The size is the change AI made, not the whole line.
“Advertising services - We provide advertising services to sellers, vendors, publishers, authors, and others, through programs such as sponsored ads, display, and video advertising.”
“The strength in advertising was primarily driven by Sponsored Products , supported by continued improvements in relevancy and measurement capabilities for advertisers.”
Conversion of shoppers who click a sponsored prompt against those who do not · 2026-CQ2
Spend of shoppers who click a sponsored prompt against those who do not · 2026-CQ2
Advertising services revenue · 2026-CQ2
Advertising services revenue, prior year · 2025-CQ2
What else could explain it
mix shift: Shoppers who click a prompt are already closer to buying; the gap is partly selection.
Quotes
“Additionally, increasingly more shoppers are discovering products in our agentic and conversational experiences, including in Alexa+ and Alexa for Shopping. Shoppers who click a sponsored prompt convert to a sale 48% more often and spend 21% more on average than those who don't.”
Q1 2026direction only · described, no size · offensive
Q2 2026direction only · our inference · offensive · $20mn to $297mn
marketing · cheap to verify
AI tools for advertisers (Creative Agent, Ads Agent) and the advertising they bring
0.02% to 0.3% of the quarter’s revenue
Incremental total: counts in full.
our inferencedisclosure: direction only· motive: offensive· before LLMs: expanded
Advertisers using Ads Agent see lower cost per impression and lower cost per acquisition, and the tool has expanded to more countries (claims c31, c32). Those are advertisers' costs, not the company's revenue, and lower prices per result may lower revenue per conversion. Advertising services revenue rose to from . The size, , is the ledger's. The measure is new this quarter and bears out the channel's expanded tag.
Evidence: 2 quotes, 4 figures, 1 confound, 2 from before coverage
Agentic tools that plan, create and target campaigns for advertisers, offered with the ad products. Management says they bring more advertisers and, in Q2, lower advertisers' cost per impression and per acquisition; it gives no measure of the advertising revenue they add.
Why this motive
An implemented tool with a measured result for advertisers, lower cost per acquisition (claim c31).
Before LLMs: expanded
At the anchor advertising growth was already credited to Sponsored Products with continued improvements in relevancy and measurement for advertisers, machine-learning targeting with no generative tools named; advertising services revenue was in 2024. Creative Agent and Ads Agent are unpriced LLM tools built into the existing ad products, which plan, create and target campaigns: a feature lifting an existing revenue line, so the channel is expanded, sized as the increment the tools add. The size is the change AI made, not the whole line.
“Advertising services - We provide advertising services to sellers, vendors, publishers, authors, and others, through programs such as sponsored ads, display, and video advertising.”
“The strength in advertising was primarily driven by Sponsored Products , supported by continued improvements in relevancy and measurement capabilities for advertisers.”
Cost per impression of advertisers using Ads Agent, lower by · 2026-CQ2
Cost per acquisition of advertisers using Ads Agent, lower by · 2026-CQ2
Advertising services revenue · 2026-CQ2
Advertising services revenue, prior year · 2025-CQ2
What else could explain it
other: Advertising growth is also credited to Sponsored Products, Prime Video ads and live sports.
Quotes
“Finally, we make it easy to create, launch, and optimize full-funnel campaigns using AI-powered tools, including Ads Agent, which turns hours of setup and targeting into minutes. Advertisers using Ads Agent targeting see 8% lower cost per impression and 6% lower cost per acquisition, and we've expanded it to 11 new countries this year.”
Q2 2026direction only · our inference · offensive · $40mn to $594mn
distribution
Orders referred by third-party AI shopping agents
Not sized
No source in the quarter mentions the referrals, and the prior quarter's bound is not repeated.
inscrutabledisclosure: not mentioned· motive: exploratory· before LLMs: new
The quarter's sources are silent on referrals from third-party AI agents. The CEO describes the company's own assistant and sponsored prompts instead.
Evidence: 0 quotes, 1 from before coverage
Shoppers sent to the store by third-party horizontal AI agents, which the CEO compares with search engine referrals and calls a small fraction of them. The company is in talks with the agents' makers; no figure is given.
Why this motive
Carried from the prior quarter: the CEO said the third-party agent experience had not gotten great yet.
Before LLMs: new
The anchor lists third-party customer referrals and sponsored search among the marketing channels and names no AI agent among them. Traffic from AI agents could not exist before LLMs.
“We direct customers to our stores primarily through a number of marketing channels, such as our sponsored search, third-party customer referrals, social and online advertising, television advertising, and other initiatives.”
Health AI: the AI health agent that books and routes One Medical virtual care
Not sized
No source in the quarter mentions the agent or its visits.
inscrutabledisclosure: not mentioned· motive: product-defensive· before LLMs: expanded
The quarter's sources are silent on Health AI; the release describes pharmacy growth without naming AI.
Evidence: 0 quotes, 1 from before coverage
A personal health agent in the store's app, backed by One Medical clinicians, that gives guidance and books visits and prescriptions. Management says virtual care visits rose year over year, a majority of them now through the agent; Prime members get some visits free. The revenue sits in Other net sales with shipping services, video licensing and other items. A clinical error is costly, so the work is expensive to check.
Why this motive
Carried from the prior quarter: a free agent routing visits.
Before LLMs: expanded
At the anchor the health services business, pharmacy and One Medical, was already growing, with no AI named. The agent routes existing virtual care, so the size is the increment it adds. The size is the change AI made, not the whole line.
“Our health services business is growing robustly as customers are loving our pharmacy customer experience, and we've launched same-day delivery of prescription medications to customers in eight cities, including Los Angeles and New York City, with plans to expand to more than a dozen cities by the end of the year, with customers now getting first-fill medications 75% faster year-over-year nationwide.”
Gains on the investments in Anthropic (conversions and upward adjustments)
25.2% of the quarter’s revenue; a non-operating gain, not added into totals
reported in the filingdisclosure: quantified· motive: exploratory· before LLMs: new· layer: compute
The 10-Q records an upward adjustment of to the Anthropic preferred stock on Anthropic's fundings, for the half, and the release puts non-operating other income of primarily on the Anthropic investments (claims c52, c51). The preferred stock is carried at (claim c54); six-month tax expense includes of discrete tax on the markups (claim c53). The quarter's markups followed the company's own purchases of and of Anthropic stock (claims c12, c14). Non-operating and left out of totals.
Evidence: 6 quotes, 6 figures, 2 confounds, 1 from before coverage
Gains on the convertible notes and nonvoting preferred stock in Anthropic: reclassified gains when notes convert, and upward adjustments to the preferred stock when Anthropic raises money at a higher price, recognized in other income. Non-cash and not revenue: a non-operating channel, traced and left out of flow totals. Microsoft carries a gain on its own Anthropic stake on its ledger. Layer compute: the sources describe the investee as an AI lab that trains on Trainium, commits to AWS capacity and offers its models on Bedrock, a model builder rather than a business using AI in its own work.
Why this motive
A non-cash gain set by the investee's funding; no operating tell applies and nothing contradicts, so it reads exploratory.
Before LLMs: new
At the anchor the company owned Anthropic convertible notes with an estimated fair value of at the end of 2024. A stake in an AI lab is new money in the ledger's sense.
“In Q3 2023, we invested in a $1.25 billion note from Anthropic, PBC, which is convertible to equity. In Q1 2024, we invested $2.75 billion in a second convertible note.”
Non-operating pre-tax other income, primarily from the Anthropic investments · 2026-CQ2
Total other income (expense), net · 2026-CQ2
Upward adjustments relating to equity investments in private companies · 2026-CQ2
Upward adjustments to the Anthropic nonvoting preferred stock, six months · 2026-H1
Discrete tax expense attributable to the Anthropic upward adjustments, six months · 2026-H1
Anthropic nonvoting preferred stock on the balance sheet · as-of 2026-06-30
What else could explain it
one time item: A mark set by the lab's funding rounds; it can reverse.
other: The company took part in the rounds that set the mark.
Quotes
“We recorded upward adjustments of approximately $50.5 billion in Q2 2026 and $62.8 billion for the six months ended June 30, 2026 to our nonvoting preferred stock in "Other income (expense), net" to reflect observable changes in price related to Anthropic's fundings.”
“Our income tax provision for the six months ended June 30, 2026 was $27.8 billion, which included $15.9 billion of net discrete tax expense primarily attributable to the upward adjustments to our investments in Anthropic.”
“As of December 31, 2025 and June 30, 2026, the amounts recorded on our consolidated balance sheets for nonvoting preferred stock were approximately $14.8 billion and $92.5 billion.”
“Under this financing arrangement, in Q2 2026, we exercised our option to participate in subsequent Anthropic equity financings by investing $5.0 billion in Anthropic Series H nonvoting preferred stock, which reduced the amount available under the facility to $15.0 billion.”
Q1 2026quantified · reported in the filing · exploratory · $16.80bn
Q2 2026quantified · reported in the filing · exploratory · $50.50bn
Reported lines, year-over-year growth
Revenue +19.6%
Q2 2026. Growing slower than revenue: cost of sales (+18.5%), sales and marketing (+2.5%), general and administrative (−6.0%), total operating expenses (+16.6%). A displaced cost shows up as a line that stays under the dashed revenue line. These are the audited lines, as first reported; nothing here is attributed to AI by the filing.
Cost of salesSales and marketingGeneral and administrativeTotal operating expensesRevenue