Who Pays for AI

What the numbers say

A seller's AI revenue is someone else's spend, and that spend is one of two things. Either a buyer pays because a line on its own income statement moved, a cost that fell or revenue that rose, or a buyer pays because it, or the investors behind it, expect that to happen. The first is absorption. The second is belief. The seller's number cannot tell them apart; the buyer's filings sometimes can. The ledger reads both sides so that the two can be told apart wherever the filings allow, and every section below is a reading of one or the other.

Read that way, the 58 companies say this. Almost all the AI money the filings can trace is capital changing hands among the builders: $103B of sized build in the quarter, $85B of chips paid for by the clouds, $20B of compute paid for mostly by payers the filings do not split, labs funded by investors and by the sellers themselves, and $60B of gains on stakes in those labs, booked by the sellers. The part that has reached companies' own operating statements is far smaller: $8.2B a quarter of end-use revenue paid by enterprises, and $1.1B of AI spend across the 37 buyers other than Meta, 0.08% of their revenue. Meta, the one buyer that also builds, spends 90 times that share.

That is the present reading, not a forecast. The sections take the chain from the chip down to the buyer, say what each layer's number is and what it does and does not show, and end with what would move these readings.

AI revenue narrows from hardware to end use

The further the money travels from the chip toward a company's own work, the less of it the filings show. AI revenue is $85B a quarter at the hardware layer, $20B at compute and $10B at end use. The money is widest where chips are sold and narrowest where people and businesses use AI for their own work, which is the only place absorption can show.

Hardware is chips and systems, and Nvidia is 99% of it; AI hardware is 88% of Nvidia's own revenue in the quarter. Compute is accelerator capacity and model access rented to whoever builds on it, sold here by Microsoft, Amazon, Oracle and CoreWeave. End use is AI inside a company's own work and AI products sold to businesses and people: the Copilots, agents and AI features, from 29 companies.

The fourth layer, facilities, is power, space and buildings sold to whoever runs the data centres. It is on the ledger at 9 companies, from power sellers such as Constellation Energy, NextEra Energy and Duke Energy to Equinix's data-centre space and Caterpillar's generators. It carries contracts and demand and no sized dollar: each company names AI beside cloud or other load without stating AI's part, counts gigawatts or deals rather than dollars, or has contracts whose money has not started. The layer that the build most visibly reaches is the one the filings price least.

The build sits outside all of it

The largest AI number on the ledger is capital: cash four companies spent building capacity in one quarter, kept out of every flow total. The ledger can size $103B of AI capital spending in the quarter ($84B to $115B), nearly all of it at Alphabet, Amazon, Oracle and CoreWeave. That is 10 times the cohort's end-use AI revenue, 5.2 times its compute revenue, and 5% of the 58 companies' combined revenue in the quarter.

Capital spending is traced and kept out of every flow total. It reaches the income statement later, as depreciation, leases, power and interest, and those costs count as spend. Adding the cash as well would count the build twice. The build is the belief side of the money in its purest form: it is paid before any buyer has paid for the capacity it creates.

The figure is a floor. Microsoft describes its capital spending as for cloud and AI together, and Meta describes its own as for its AI efforts and core business together. Neither separates AI's part, so the ledger leaves both unsized; only Meta's data-centre ventures carry a size. Ten companies in all name AI capital spending with no size, among them Apple, NextEra Energy and Duke Energy.

Clouds pay for the chips; the filings do not split who pays for compute

The filings name who pays for chips and stop naming at compute, and compute is where the belief side of the money sits. At the hardware layer, the hyperscale clouds and the largest consumer internet companies pay $45.8B of the $85.2B. Another $37.9B comes from AI clouds, model makers, governments and enterprises together, which Nvidia reports as one group the ledger cannot split by payer. The remaining $1.5B is Nvidia's AI workstations and Alphabet's TPU system sales.

At compute, $12.5B is paid by a mix the filings do not split: Amazon's AI run rate, Oracle's AI infrastructure and most of CoreWeave's customers. AI labs pay $6.5B, all of it on Microsoft's frontier-lab line, and enterprises $721M. At end use the payers are named: enterprises $8.2B, consumers $1.6B.

Funding follows the same pattern. The ledger records hardware revenue as 100% mixed-funded and compute revenue as 87%: the payers' money comes partly from outside funding rounds and partly from the sellers themselves. Nvidia holds stakes in AI model makers and AI clouds. Amazon put money into OpenAI and Anthropic in the quarter. Microsoft holds stakes in both labs. CoreWeave's build is paid for with loans against customer contracts, equipment-maker financing, customer prepayments and equity, with Nvidia among the buyers of that equity. End-use revenue, by contrast, is 85% paid from customers' operating cash flow.

This loop, in which a seller funds a customer that then pays the seller, is traced on the ledger channel by channel. It matters for the reading because a payment that comes partly from the seller's own money is evidence of the seller's belief as much as of the buyer's demand. The ledger records the loop. It makes no forecast about where it goes.

Gains on stakes in AI companies outweigh compute revenue, and stay out of every total

The gains on lab stakes are what belief looks like on a seller's income statement: they rise when later investors pay more for the labs, and in this quarter they outweigh compute revenue. Sellers booked $60B of gains on stakes in AI companies ($58B to $63B). 84% of it is Amazon's gain on Anthropic. The rest is Microsoft's on OpenAI and Anthropic, Nvidia's on its ecosystem stakes and Salesforce's on Anthropic.

Set against each layer on its own, the marks are 3.0 times compute revenue and 5.9 times end-use revenue, and below hardware revenue. A mark is set by what other investors paid in a later round. It is not cash, and it sits below operating income, so the ledger shows it and adds it to nothing. It is also the same money seen again: the rounds that set these marks are the funding behind the labs' compute bills in the section above.

New money: most at end use, undetermined at hardware

Of the revenue at end use, most would not exist without the models; at hardware, no one can say. The ledger tags each channel against the company's 2024 annual report: new if the activity could not exist without large language models, expanded if it existed and AI changed its size, relabelled if only the name changed. The incremental total is what would not be there without the models, and it is the part of AI revenue that is not an older business under a new name.

At end use, $6.0B of $10.1B is incremental. At compute it is $7.0B of $19.8B; the rest is capacity rental, an activity that predates the models, at sellers with no traced quarter before AI.

At hardware the incremental total is $0 at the point and $88.4B at the high end. Nvidia's data-centre revenue is expanded: data-centre chips were sold before the models, and no quarter before AI traces how many. The ledger does not estimate that baseline, so the amount is undetermined, counted as zero at the point and in full at the high end. The truth sits somewhere between, and nothing in the filings says where.

Buyers spend a sliver of their revenue on AI

This is the absorption reading, and it is small. The 38 buyers on the ledger take in $1.45T of revenue a quarter. Their counted AI spend is $5.2B, 0.36% of that revenue, and 80% of it is Meta's. Meta is a buyer that builds: its spend is its own AI build reaching its income statement as depreciation, cloud capacity and tokens. Its AI hiring is named and carries no size.

Without Meta, the other 37 buyers spend 0.08% of their revenue on AI (0.03% to 0.22%), and their counted savings come to $274M. Meta's share is 90 times theirs: the distance between a company that builds AI and the companies that use it is the largest single contrast on the ledger.

16 of the 38 buyers have nothing the ledger can size. Most name AI, often in detail, and give no figure: Apple, Elevance, Wells Fargo and ExxonMobil among them. United Airlines does not mention AI in any covered source. The chart lists them by sector instead of drawing them as zero, since an unsized channel's amount is unknown. The reading is a floor on what buyers spend and a fair measure of what they disclose.

Sellers report billions; the buyers on the ledger show a sliver

The sellers' end-use revenue and the buyers' end-use spend are two sides of the same money, and the ledger sees 8.9 times more of it on the sellers' side. The sellers and displaced vendors on the ledger take in $9.3B of end-use AI revenue a quarter. The 37 buyers other than Meta show $1.1B of end-use AI spend. Part of the buyers' figure is their own staff and build rather than anything bought from a seller.

Both can be true. Most of the sellers' customers are outside the cohort: 38 buyers are a small part of Microsoft's or Salesforce's customer base. And much compute is bought by the labs and the clouds themselves, not by companies using AI in their own work.

What the ledger can close is the chain it covers: who sells at each layer, whom the filings name as paying, and how those payers are funded. What it cannot close is the buyer side of the whole economy. Its buyers were chosen sector by sector from the S&P 500 by operating-expense weight, and most of them do not publish what they spend. The ratio above, 8.9 to one, measures that gap in disclosure as much as any gap in demand. What would narrow it is disclosure: a buyer stating an AI dollar, as Airbnb did for one support cost line, moves a channel from the ledger's estimate to the company's own figure, and enough of those would settle which side is nearer the truth.

Outside estimates count different things again, and none of them is a ledger total:

How much weight the numbers bear

Three facts set how far to trust the readings above: nearly every size is the ledger's own estimate, the cohort covers under half the index by weight, and the sector readings stopped moving when the last two waves of companies were added.

Companies name AI often and size it almost never

The 58 companies' latest quarters hold 472 channels. 239 of them have no size. Of the 202 sized channels that count in totals, 194 are the ledger's own estimates, built from a decomposition of reported lines or a reference class with every assumption written down. Three are implied by a bound management gave, and five are reported by the company. Another 31 are sized and kept out of totals as overlaps, capital or marks. The dollar totals are therefore floors with wide ranges, and the ranges are the honest reading.

A channel stays unsized when the company names AI beside another cause and nothing separates AI's part, when the only measure is a count (users, interactions, deals, gigawatts), when the money has not started, or when a size would need a share of the line picked by judgment with no company figure under it. Of the unsized channels, 94 are described with no measure, 66 give a direction or a count, 69 were silent in the quarter, and ten state a bound or a level that is not in dollars.

Those rules got stricter during the build-out. After the fourth wave a correction sweep re-read every earlier company under them. For the 34 companies then on the ledger, compute revenue went from $30.0B to $19.8B a quarter, down 34%, and end-use revenue from $10.9B to $10.1B. The readings got smaller and firmer, which is the direction a stricter rule moves them.

How far it reaches, and why it stops here

The 47 index companies on the ledger account for 45% of S&P 500 operating expenses, the weight the sampling frame uses, and 40% of its revenue. The other 11 companies are outside the index. Every sector above 5% of the weight has at least three companies on the ledger. Operating expenses are the weight because absorption lands in labour and operating costs, which is where a buyer's AI spend and savings would show.

The cohort stops growing here because the readings stopped moving. For each sector, the stability test divides the buyers' counted end-use AI money by their revenue, then adds the next wave of companies and reads it again. In the last two waves, each of the four sectors with a reading that gained a buyer kept its new reading inside the range it already had. Utilities and Energy gained buyers and still read zero. Adding companies was no longer changing the answer, so the sliver in the buyers section is a property of the sectors, not of which companies were picked.

One gap is known. Communication Services' buyer reading, 3.7% of revenue, rests mostly on Meta, which holds 99% of the sector's counted end-use money. Further companies are added only for a named gap of this kind.

What would move these readings

The ledger is re-read every quarter by unattended runs as each company reports, under the same rules. Three readings carry the question of absorption against belief, and each would move in a way the ledger would show.

The buyers' spend share. It reads 0.08% of revenue without Meta. A move toward one percent across sectors, or buyers sizing their own AI spend in dollars, would be absorption arriving in the filings. A reading that stayed here while the sellers' end-use revenue grew would widen the gap above and say the sellers' customers are still elsewhere.

The funding of compute. 87% of compute revenue is mixed-funded. If the labs' payments to the clouds came to be described as paid from their own operating cash flow, belief would be converting into absorption at the top of the chain. If the sellers' share of the labs' funding grew instead, the loop would be tightening.

The marks. $60B of gains are set by later funding rounds, and a mark set by a round can be reset by one. A reversal would show here first, below operating income, before any flow total moved.

This post is frozen at commit 1e96658c; the live numbers are at /ai-absorption, where every figure opens to the quote, filing line and arithmetic behind it.