Can the AI labs pay for the boom?

The most revealing document about Anthropic this year was not one the company wrote. It sits on page 14 of a prospectus filed by SpaceX, which owns Elon Musk's xAI, and it says the maker of Claude "has agreed to pay us $1.25 billion per month through May 2029." Once a cut-price ramp-up in May and June 2026 gave way to the full rate in July, that came to $15 billion a year, paid to a company whose own chatbot, Grok, competes with Claude.

The money buys time on machines in the American South. SpaceX's prospectus describes Colossus as its flagship data center, on Paul R. Lowry Road in Memphis, Tennessee; Colossus II spreads across Memphis and Southaven, Mississippi, part of what the company calls a "coherent gigawatt-scale AI training cluster." Anthropic's share includes "approximately 325,000 NVIDIA GPUs."

Multiply that contract across the industry and you have the American economy's biggest construction project. Can the customers at the end of the chain, the frontier labs selling Claude, ChatGPT and their rivals, earn enough to pay for it? Our answer, for now, is no. The build-out is being carried by the cash flows of four of the world's most profitable companies and, increasingly, by debt. The labs' revenue is the justification for it rather than the means. That can work for a while. It cannot work for ever.

Start with the size of the bet. Alphabet's second-quarter call in July raised its 2026 capital-spending range to $195 billion-205 billion; Anat Ashkenazi, its finance chief, said she expects capex "to increase significantly in 2027." Meta, the same week, guided to $130 billion-145 billion, a figure that includes principal payments on finance leases. Amy Hood of Microsoft put calendar-2026 spending at about $175 billion after an accounting change that moves some data-center leases out of the capex line, and promised more in the fiscal year to June 2027. Amazon's written guidance, from its February results release, is still about $200 billion: Andy Jassy, its boss, said then that "we expect to invest about $200 billion in capital expenditures across Amazon in 2026." Its second-quarter release did not update that full-year figure. Add those four plans with care and they come to roughly $700 billion this year—an illustrative sum, not a single industry total. Meta's range includes finance-lease principal; Microsoft's $175 billion is after the lease reclassification; the figures are not like-for-like.

The justification, in writing, is customer demand already booked. In his 2025 letter to shareholders, Mr. Jassy said: "We're not investing approximately $200 billion in capex in 2026 on a hunch." Of the AWS portion, "we already have customer commitments for a substantial portion of it."

The suppliers' books show where the money lands. Nvidia's data-center revenue was $89 billion in the quarter to July 26th, up 117% on a year earlier, according to its results on August 26th, and it guided to total revenue of $108 billion, give or take 2%, for the quarter now ending. Oracle reported $664 billion of remaining performance obligations at the end of August. Google Cloud's backlog reached $514 billion in the second quarter, Sundar Pichai told analysts.

Now set the labs' revenue beside that. Anthropic said in May that its run rate, which extrapolates recent sales over a year, had crossed $47 billion. On October 9th Sarah Friar, OpenAI's finance chief, wrote on LinkedIn: "We ended September with an annualized revenue run rate of ~$50 billion, up 77% in just three months." SpaceX's AI segment, which includes Grok and X, recognized $2.56 billion of revenue in the second quarter. Annualize xAI's quarter and add the two run rates—mixing run rates with recognized revenue, and figures months apart—and the three together are selling something near $110 billion a year, against a single year's illustrative capex of roughly $700 billion by just four buyers.

The comparison is unfair in one way: data centers last for years and revenue recurs, so a year's revenue need not match a year's capex. Microsoft has just stretched the assumed life of its data-center buildings from 15 years to 25. But chips wear out much faster than buildings, and roughly two-thirds of Microsoft's capex last quarter went on short-lived assets, mainly CPUs and GPUs, Ms. Hood said. Those have to earn their keep within a few years. The revenue that must pay for them is, for now, mostly a run rate.

That is why Anthropic matters beyond its own valuation. Its $965 billion post-money price, set in its Series H on May 28, is the market's most explicit wager that lab revenue will grow into the infrastructure built for it. And Anthropic has gone quiet. Its run rate stood at $14 billion in February and above $47 billion by May; since the Series H release, almost five months ago, it has published no new figure. On June 1 it said it had confidentially submitted a draft S-1 to the Securities and Exchange Commission. No public filing has followed.

Nobody outside can say what Anthropic earns, what it keeps after paying for chips, or what it owes Microsoft and Google in any given year.

What is known is the bills. Beyond the Memphis contract, Anthropic has said it is committing more than $100 billion over ten years to Amazon Web Services; spread evenly, that is at least $10 billion a year. With SpaceX, it carries at least $25 billion a year of compute spending on disclosed terms, roughly half its May run rate. Its $30 billion Azure purchase, agreed in November 2025, came with no term, and its April deal with Google and Broadcom for multiple gigawatts of TPU capacity from 2027 came with no price, though Anthropic called it "our most significant compute commitment to date."

No lab publishes a gross margin, so assume, crudely, that compute eats half of revenue. Anthropic would then need about $50 billion a year to carry its two contracts with known terms, and the May run rate is just short of that. Stretch the Azure deal over five years, a term nobody has disclosed, and the figure rises to about $62 billion. On that arithmetic a company valued near $1 trillion is running at about compute break-even, provided the run rate turns into revenue.

OpenAI looks healthier on paper, and the calendar helps it. Against its $852 billion valuation of March 31st, Ms. Friar's September figure implies a multiple of about 17, below Anthropic's roughly 20 times on its May number. Its AWS and CoreWeave contracts with disclosed terms come to roughly $19 billion a year, which on the same half-of-revenue assumption needs about $38 billion of sales. But the $250 billion of Azure services it agreed to buy in October 2025 has no stated term; if it ran ten years, the break-even would rise to about $88 billion. The two firms do not calculate their run rates alike, and Anthropic's latest number is more than four months older than OpenAI's.

The money also travels in circles. Amazon put $8 billion into Anthropic convertible notes between 2023 and 2025 and bought $5 billion each of Series G and Series H preferred stock in the second quarter of 2026. Microsoft and Nvidia pledged up to $15 billion between them alongside Anthropic's Azure deal. Microsoft owned about 27% of OpenAI when its recapitalization closed in October 2025. Much of that equity returns to the investors as payment for chips and cloud capacity. The cloud firms are financing their own order books.

They are also booking gains on them. Amazon's quarterly report for the period ended June 30th carries its Anthropic preferred stock at about $92.5 billion and its convertible notes at about $97.9 billion, after $50.5 billion of upward adjustments in the second quarter alone; those are estimates derived from private rounds. Microsoft booked $6.5 billion of net gains on OpenAI in the fiscal year to June. Profits that rest on private marks of customers' shares are a thin cushion if the customers stumble.

Increasingly the cushion is debt. Meta, which does not sell cloud capacity, raised $24.9 billion of long-term debt in the second quarter, according to its results. In October 2025 it set up a joint venture with funds managed by Blue Owl Capital to build and own its Hyperion campus in Richland Parish, Louisiana: Blue Owl holds 80%, the partners committed to fund about $27 billion of development costs, and Meta will lease the buildings on four-year initial terms. Meta also gave the venture a residual-value guarantee for the first 16 years, a capped payment if it walks away. Structures like that move the borrowing to outside lenders, partly bonds sold to PIMCO and others, while leaving Meta on the hook if it decides it needs less space.

The build-out is now large enough to show up in the national accounts. In its third estimate of second-quarter GDP, published on September 30th, the Bureau of Economic Analysis said the biggest contributors to the upward revision in private fixed investment were nonresidential structures, "led by commercial and health care, mainly data centers." BEA does not publish an "AI" line in the accounts. The closest related category, private fixed investment in information-processing equipment and software, ran at an annual rate of $1.65 trillion in the second quarter; the contribution of information-processing equipment alone to GDP growth slowed to 0.19 percentage points from 0.79 in the first.

It is also showing up in prices. Lisa Cook, a Federal Reserve governor, said on September 28th that companies "have only spent a small fraction of the $2 trillion in announced plans," and that electricity and water costs "are each up around 5 percent over the past year, potentially attributable in part to AI." She voted with her colleagues to raise interest rates by a quarter-point in September, with inflation at an estimated 3.8% in the year to August. "In the short term," she said, "AI appears to be adding inflationary pressures to the economy."

The grid is being planned around the same expectation. PJM, which runs the grid across much of the mid-Atlantic and Midwest, forecast in January that its summer peak would reach 222,106 megawatts in 2036, an increase of 65,733 megawatts in ten years, with data-center load growth written into the outlook for 14 of its zones. ERCOT, the Texas grid operator, said in an April CEO update that transmission-service providers' preliminary RFI submissions for non-crypto data-center loads rise from 7,401 megawatts in 2026 to 228,420 megawatts by 2032—requested large loads, not firm interconnected demand. Power plants and transmission lines are built on 30-year assumptions. The revenue they are being built for is measured in months.

Further down the market, the price of intelligence is falling. Mistral, a French lab, raised €3 billion in September at a post-money valuation of more than €21 billion, selling open-weight models customers can run on their own hardware. On the companies' own price lists, read on October 15th, Anthropic's flagship, Claude Fable 5.1, costs $10 per million input tokens and $50 per million output tokens; DeepSeek-V4-Pro, whose weights are open, costs $1.32 and $3.96 at peak hours. If cheap models handle most corporate work, revenue per chip will fall just as the chips arrive.

What happens if the money stops? It need not stop to hurt. A listing that prices Anthropic below its Series H valuation would turn up promptly in Amazon's other income, because those marks would then follow a market price. A pause in lab spending would hit Nvidia's order book, then the utilities and lenders that financed the buildings, then the investment line that has been propping up GDP. Ms. Cook's worry runs the other way, too: a slowdown would ease the price pressure she described, but it would also expose how much of recent growth rested on one sector.

A sudden stop looks unlikely in the next year. The hyperscalers' guidance is set, their cloud backlogs are long, and Amazon says it already has customer commitments for a substantial share of this year's AWS spending. The risk is slower and duller: capex keeps rising on the strength of contracts and marks, while the revenue meant to justify it arrives as run rates, from companies that publish when they choose.

For the build-out to pay, several things must be true. Lab run rates have to become audited revenue at something like their stated pace. Compute must take a falling share of that revenue—cheaper chips, or prices that hold up against open-weight rivals. Enterprises, not just labs and hyperscalers trading capacity with one another, must become the paying customers. And the debt now financing the buildings must find tenants who stay.

Over the next 12 to 24 months, watch five things. First, the clouds' third-quarter results at the end of October, and their first formal capex guidance for 2027. Second, Anthropic's public S-1, which should show recognized revenue, losses, customer concentration and the missing Azure and Google terms, and Amazon's next filing, which will show whether it has marked its stake up again. Third, Nvidia's next quarter against its $108 billion guide. Fourth, the BEA's advance estimate of third-quarter GDP, and whether data centers keep carrying investment. Fifth, the Fed: if AI spending keeps pushing up electricity and construction costs, the build-out will be paying higher interest rates of its own making.

The labs have persuaded the world's largest companies to back them against revenue not yet recognized, and to take payment in their own future sales. The economy is leaning on that arrangement too. The first audited numbers will tell investors whether $965 billion bought a company or a promise.