Who Is Really Paying for AI?

On August 26, Nvidia filed its quarterly report. Deep in the notes, under a heading about commitments and guarantees, the company disclosed that it had agreed to stand behind somebody else's rent. The guarantees are "capped at a total of $105 billion." They cover leases for "approximately 4.25 gigawatts of IT load" at a data center campus in Pike County, Ohio, signed "on behalf of a customer, an affiliate of OpenAI Group PBC." And they end, the filing says, upon certain events, "including OpenAI achieving a satisfactory credit rating."

Read that last clause again. Nvidia, which reported $215.9 billion of revenue in its latest fiscal year, is backing the lease payments of a customer until that customer's own credit is judged good enough to stand alone. In return, the filing says, the site "will exclusively host NVIDIA AI infrastructure." The guarantees switch on as each of nine construction phases is finished, the first expected in Nvidia's fiscal year 2029, and they cover "defined portions of lease and power payments," not the whole site.

This deep dive asks the question that has been argued loudly all autumn: who is actually paying for AI? The popular answer goes like this. Venture investors pour money into AI labs. The labs spend it renting computing power from the cloud companies. The cloud companies report it as revenue and buy more chips. The money goes around in a circle, and the circle is funded by investors rather than by customers who make money from the result.

We tested that story against four kinds of record: the cash flow statements of the five companies at the center of the build-out (Microsoft, Amazon, Alphabet, Meta and Oracle), the contracted backlogs that the cloud companies report every quarter, the national accounts, and a Census Bureau survey that asks American businesses whether they use AI at all. The answer splits in two. On the spending, the popular story is wrong: the build-out is paid for mostly with cash that advertising, retail, cloud and software businesses already throw off. On the revenue, the story is closer to right than the optimists admit. And the most telling numbers are neither spending nor revenue. They are the guarantees: who has promised to pay, and who has promised to pay if they do not.

Start with scale, because it settles one argument quickly. The Bureau of Economic Analysis counts private investment in information processing equipment (computers, communications gear, instruments and medical equipment) and in software. In the second quarter of 2026 that investment ran at 5.05% of GDP. The peak of the dot-com build-out, in the fourth quarter of 2000, was 4.46%. At the start of 1990 the share was 3.01%.

So the comparison everyone reaches for has already been settled by the data. Measured against the size of the economy, this build-out is larger than the one that ended in 2000. That is a statement about size, not about what comes next. After the last peak the share fell to 3.4% within five years, and it did not climb back above 4% until 2018. Nobody has to predict a repeat to notice that the line is now higher than it was then.

It also shows up in growth. In 2025 real GDP grew 2.3%, and investment in information processing equipment and software contributed 0.7 of a percentage point of it, about 30%. In 1999 and in 2000, at the height of the last boom, the same categories contributed about 18% of growth. The accounts impose one caution of their own. These contributions are gross. Much of the equipment is imported, imports are subtracted from GDP elsewhere in the same tables, and the net effect on growth is smaller than the headline. The direction is not in doubt; the exact size is.

Here is the received view at its strongest. PitchBook and the National Venture Capital Association counted $222.1 billion of venture deals in American AI and machine-learning companies in 2025, 65.4% of all US venture deal value. That is an enormous sum, and a great deal of it is spent renting computing power from the companies in this story.

Now set it beside the spending. Microsoft, Amazon, Alphabet, Meta and Oracle paid $412 billion in cash for property and equipment in calendar 2025, about 1.9 times all US venture funding of AI that year. In the latest twelve months they paid $601 billion. Not every capital dollar is for AI, and not every venture dollar buys computing. But the arithmetic leaves no room for the circle to be the main story. Had every venture dollar raised by every American AI company in 2025 been paid straight to these five as rent, it would have covered roughly half of what they spent.

Where does the money come from, then? Mostly from businesses these companies already run. Their combined cash from operations was $710 billion over the latest twelve months. Capital spending took 85% of it. In fiscal 2016 the five spent $32 billion, 28% of their operating cash. It was 54% in fiscal 2022 and 41% in fiscal 2023. In calendar 2025 it was 68%.

That is the first myth, and the filings bust it. The build-out is being paid for by the search, advertising, retail, cloud and software profits of five companies, with venture money a smaller stream flowing into the same pipes. If the critics' worry is that the whole edifice rests on venture funds, the cash flow statements say it does not.

There is a second reason the cash cushion matters, and it arrives later. Capital spending does not hit profits when the check is written. It arrives as depreciation, spread over the years the equipment is expected to last. Meta's cash flow statement shows the wave forming: depreciation and amortization of $19.5 billion in the first nine months of 2026, against $13.2 billion a year earlier, up 48%. Nearly every dollar of this year's spending becomes a charge against future earnings, whatever the revenue turns out to be. A company paying from operating cash can absorb that. A company paying with borrowed money has to absorb it and the interest.

But the cushion is getting thinner, and debt and new stock are filling the gap. In the latest twelve months Amazon, Alphabet, Meta and Oracle raised $272 billion of new debt (for Oracle, its fiscal year to May). Microsoft raised none. Between April and August 2026, Alphabet and Oracle also sold $69.4 billion of new stock, and Alphabet's filing says its preferred stock will fund, among other things, "capital expenditures to scale AI infrastructure and global compute." Two of the five now spend more than their operating cash. Amazon's capital spending was 107% of its operating cash flow over the twelve months to June. Oracle is the extreme case: $76 billion over the twelve months to August, 161% of its operating cash flow, and its annual report says plainly what that exposes it to: "disruptions in credit markets" could raise its borrowing costs or "limit our ability to access debt financing on acceptable terms, or at all."

If the spending is paid for with real cash, what is supposed to pay it back? The cloud companies report the answer every quarter, as remaining performance obligations: revenue that customers have signed for but that has not yet been earned. Call it the backlog. At Microsoft, Oracle, Alphabet and Amazon together it was $608 billion in June 2024. In June 2026 it was $2,338 billion, 3.8 times as much. In two years the backlog grew by $1.7 trillion.

The backlog did not grow the way a broad market grows, a little every quarter. It arrived in steps. Oracle's rose by $317.5 billion in the quarter to August 2025. Microsoft's rose by $233 billion in the quarter to December 2025. Alphabet's rose by $224.8 billion in the quarter to March 2026. Amazon's rose by $120 billion and then by $132 billion in the first two quarters of 2026. Steps that size come from very large contracts. Two companies have said whose, and both answers are AI developers.

The first is Microsoft. On its earnings call on January 28, 2026, the chief financial officer, Amy Hood, said: "Approximately 45% of our commercial RPO balance is from OpenAI." Microsoft's commercial backlog was $625 billion that quarter, up 110% in a year. The rest of it, she said, "grew 28% and reflects ongoing broad customer demand across the portfolio."

Run the counterfactual the call invites. Take OpenAI out, and Microsoft's commercial backlog grew 28%: a healthy software business, growing about as you would expect. Leave OpenAI in, and it grew 110%. On the next call, on April 29, 2026, the pattern held: excluding OpenAI, Hood said, commercial backlog "grew 26% in line with historic seasonality." The difference between a good year and an extraordinary one is a single customer.

The second is Amazon, and its quarterly reports name the contracts behind its two steps. In the first quarter of 2026, its filing says, AWS and OpenAI "announced an expansion of the existing $38.0 billion multi-year commitment" by "$100.0 billion over 8.0 years." In the second quarter, AWS and Anthropic expanded their commitment "by more than $100.0 billion over 10.0 years." Amazon's backlog rose by $120 billion and $132 billion in those two quarters. Two AI developers account for most of Amazon's increase in 2026.

Microsoft and Amazon are unusually candid. Oracle, whose backlog reached $664 billion in August 2026, does not name its customers. It says no single customer accounted for 10% or more of its revenue, which describes revenue already earned rather than the backlog. Its annual report adds the sentence that matters: "In certain OCI offerings, we are more concentrated among a number of large customers, which could increase these risks." Alphabet, whose backlog nearly doubled in the first quarter of 2026, does not say who signed. Part of that quarter's rise, $7.3 billion, came from a change in what Alphabet counts.

Further down the chain the concentration is on the record. CoreWeave, a cloud company that rents out Nvidia's chips, got about 67% of its 2025 revenue from Microsoft, and has disclosed contracts under which OpenAI committed to pay up to about $11.9 billion and, in a later order, up to $6.5 billion. Nvidia sold 22% of its fiscal 2026 revenue to one direct customer and 14% to another. It does not name them. Its quarterly report puts the pattern in one line: "Our revenue is concentrated among a limited number of direct and indirect customers and this trend may continue."

At the top of the chain sits a company that does not file with the SEC. On November 6, 2025, OpenAI's chief executive, Sam Altman, wrote that the company expected to end that year "above $20 billion in annualized revenue run rate" and was "looking at commitments of about $1.4 trillion over the next 8 years." That is an eight-year total of about 70 times the one-year run rate he gave. He also wrote that he expected revenue to grow "to hundreds of billion by 2030." The commitments are contracted; the growth is a projection. And a large part of the build-out is being ordered by a customer whose chip supplier's guarantee ends, among other events, when it achieves "a satisfactory credit rating."

So on revenue, the critics are closer to right. A large share of the build-out's future revenue rests on contracts with a small number of AI developers, and the one whose chief executive has spoken publicly about its numbers has commitments far larger than its revenue. Where the circular story goes wrong is on who carries the risk if those promises slip. It is not mainly the venture investors.

A guarantee is a way to spend credit instead of cash. It does not show up as capital spending. It does not show up as revenue. On the day it is signed it costs no cash. It shows up in the notes to the accounts, and the notes of the past year read like a map of where the risk of the build-out has been moved.

It helps to ask, as you would of any trade, who is on the other side. When a cloud company signs a contract worth tens of billions with an AI developer, the cloud company takes on the cost of building, and the developer takes on a promise to pay. When the developer's promise is not enough for a landlord or a lender, someone with a stronger balance sheet steps in and lends it their credit. The filings now show several ways that has been done.

Nvidia's is the largest single example: $105 billion for the Ohio campus, plus $3.5 billion of guarantees for other AI cloud companies' leases, $108.5 billion in all. The same filing shows the circle the critics describe, in Nvidia's own words. It has committed $29 billion to cloud service agreements, under which "AI clouds procure our data center infrastructure products and we commit to cloud service agreements." It has committed $25 billion to equity investments in "AI model makers, infrastructure financiers, and other private companies." It sells the chips, rents some of the capacity back, invests in the buyers, and now guarantees a customer's rent. Each of those is disclosed. None of them is hidden. They are simply in the part of the filing that the earnings headlines skip.

Meta's version keeps the buildings off its balance sheet. It holds 20% of two ventures building data center campuses in Louisiana and Texas, with total estimated development costs of about $27 billion and $14 billion. Meta will lease the campuses, from 2029 and 2028, on first terms of four years with renewals up to 20, and has given residual value guarantees to both: if Meta ends or does not renew a lease and the campus is worth less than an agreed threshold, Meta pays the shortfall. Because Meta does not control the ventures, it does not consolidate them. Its maximum exposure to loss, by its own count, was $68.4 billion at the end of September 2026, up from $46.0 billion nine months earlier, mostly because the Texas venture was added in July.

Microsoft's and Amazon's version is ownership and financing. Microsoft holds about 27% of OpenAI, and OpenAI is about 45% of its commercial backlog; the two have, in Microsoft's words, "reciprocal revenue-sharing arrangements." Amazon had $28.7 billion invested in OpenAI preferred stock at the end of June, holds preferred stock in Anthropic, and has agreed to make up to $20.0 billion available to Anthropic under a facility that opens "as we reach certain delivery milestones of compute capacity." The seller of the computing is lending its customer the means to keep buying it.

The capital markets are the last piece, and they work differently: $272 billion of new debt in a year from four of the five, plus new stock, raised on the companies' own credit to build for customers whose credit is weaker. Put together, the risk of the build-out is not held mainly by the venture funds that the circular story worries about. It sits in the footnotes of the companies that sell the computing, in their bonds and new shares, and in the guarantees, stakes and financing they extend to their own customers.

Suppose, as a thought experiment, that the revenue of the largest AI developers grows more slowly than their commitments. Nobody has to predict that to ask where it would show first, and the filings answer the question in order.

It would show first in the backlogs, not in the spending. A backlog is a promise, and promises can be revisited: Microsoft's own quarterly report says that estimating its backlog can involve "potential renegotiation of commitments." The promises are also long. Oracle expects to recognize about 13% of its $664 billion backlog as revenue in the next twelve months. Microsoft, as of December 2025, expected about 25% of its own. Alphabet expects just over 50% within 24 months. Amazon's long-term contracts have a weighted-average remaining life of 6.4 years. The spending is happening now; most of the revenue is scheduled for later.

It would show next at the companies with the most concentrated customers: CoreWeave, whose top customer was about two thirds of its 2025 revenue, and Oracle and Amazon, whose capital spending already exceeds their operating cash flow. It would reach Nvidia's guarantees only later, and only if a tenant defaults. The guarantees switch on phase by phase from fiscal 2029, and Nvidia says its payment obligations "are triggered upon certain tenant defaults."

And it would show last in the spending of the five, because most of that spending is paid from operating cash. That is what the circular story misses. If AI revenue disappoints, the cost does not fall on venture capital alone. It falls on the cash flows of the five companies' existing businesses, on the bondholders who lent against them, on the shareholders of a landlord's guarantor, and on the suppliers who built to order.

The last place to look is the one the backlog ultimately depends on: whether ordinary businesses are adopting AI fast enough for the developers' revenue to grow into their commitments. Every two weeks the Census Bureau asks a sample of American businesses whether they used AI in any business function. In the latest wave, covering the two weeks to October 4, 2026, 24.6% said yes. When the current version of the question began, in November 2025, the figure was 17.3%. (The Census's latest files call it "Super Intelligence or SI (formerly referred to as Artificial Intelligence or AI)." The examples it gives are the familiar ones: data analysis, generative tools for marketing and coding, customer-service agents.)

That is fast growth from a low base, and it is very uneven. Over the latest six waves, about 47% of businesses in the information sector said yes, as did 42% in professional, scientific and technical services and 39% in finance and insurance. In accommodation and food services the figure was 9%; in transportation and warehousing, 11%. Three in four American businesses did not say yes in the latest fortnight. The pattern fits a technology that does most of its work on documents, code and customer messages: the industries that run on those lead, and the ones that run on kitchens, trucks and mines trail.

This is the honest counterweight to both sides of the argument. The adoption is real, broad in office-based industries and rising through 2026. It is also a long way from the scale the backlog assumes, which is why the backlog leans so heavily on a few developers whose own customers are, in the end, the businesses in this chart. The data describe use, not spending: a business that tried a chatbot once counts the same as one that rebuilt its operations, so the survey is a ceiling on breadth, not a measure of depth.

Every figure in this deep dive comes from a filing, a federal dataset or a statement on the record, and each is listed in the claims ledger that sits behind the page. Capital spending, operating cash flow and debt are taken from the cash flow statements the five companies file with the SEC, read through the SEC's structured data. Capital spending means cash paid for property and equipment; it covers more than AI, and it leaves out equipment acquired under finance leases, which some of the five use heavily. Microsoft's fiscal year ends in June and Oracle's in May, so their "fiscal 2025" is not calendar 2025; where we compare with calendar-year venture data, we rebuilt their calendar-year totals from quarterly filings (for Oracle, the twelve months to November 2025).

The backlog is each company's remaining performance obligations, and the definitions are not identical: Microsoft reports a total and a commercial portion, Amazon counts only contracts longer than a year, Alphabet began including shorter contracts in 2026, and Oracle's quarters end a month earlier than the others. The national figures are the Bureau of Economic Analysis series for investment in information processing equipment and software, divided by GDP. The adoption figures are the Census Bureau's Business Trends and Outlook Survey, averaged over six waves by sector because single waves are noisy. The venture figure is PitchBook and the NVCA's count of US deal value in AI and machine learning. What we cannot see, we have not estimated: OpenAI does not file with the SEC, and its numbers here are only the ones its chief executive has stated publicly.

For a board approving AI spending, the useful question is not whether there is a bubble. Nobody in the room can settle that. The question that can be settled is whose promises your plan depends on. If your vendor's growth rests on a handful of customers, your capacity, prices and priority rest on them too. Ask your suppliers what share of their backlog comes from their largest customer. Microsoft has shown that the number can be disclosed.

For a supplier selling into the build-out, in power, construction, equipment or real estate, the question is whose credit stands behind the order. A lease backed by Meta's residual value guarantee or by Nvidia's credit support is a different instrument from one backed by the tenant alone, and the filings now tell you which one you are holding. Price it accordingly.

For anyone making a large commitment of their own, the build-out is a case study in what our All-In Bet pack calls the blind spend: how much is committed before outside evidence arrives. Here the commitments run to trillions of dollars, and the outside evidence, in the Census data, is a quarter of American businesses using the product in a given fortnight. That is not proof of failure. It is a sign that most of the bet is still ahead of its evidence.

Spending tells you who has cash. Backlog tells you who has made promises. Guarantees tell you whose promise needed backing. When a boom's numbers look broad, find the footnote that says whose credit stands behind the backlog. When the seller backs the buyer, with a guarantee, a stake or a loan, the seller is carrying part of the buyer's risk, and that is the number to watch.

The next time a company reports a record backlog, ask three questions before you share the headline. How much of it arrived in one quarter? Has the company said who signed it? And is anyone other than the customer standing behind the payments, through a guarantee, an equity stake or financing? A backlog that grew steadily, from many named sources, with no guarantee attached, is demand. One that arrived in a step, from an unnamed customer, with a supplier's credit or capital behind it, is a bet.