Anthropic's compute bill due next year is $70 billion, exceeding its full-year revenue

In the S-1 amendment Anthropic filed, the most eye-catching number is $895 billion—that's its total commitment over the coming years to buy compute and build data centers. After listening to this episode of Prof G Markets, though, the number I remembered was the inconspicuous one next to it: $70 billion.

That's the portion due within the next 12 months. Anthropic brought in $17 billion last quarter, which annualizes to $68 billion. The money it has to pay out next year already exceeds what it can take in for a full year.

$895 billion is the total; $70 billion is the deadline

Ed Elson put these two numbers side by side at the start of the show. He put it bluntly: that $70 billion is more than Anthropic's current annualized revenue. John Foley, editor of the FT Lex column, responded in a flat tone: the compute bills over the next few years will be enormous, and being able to attach a $900 billion figure to it does sound huge—and it may get even bigger.

What I care about more is his next line: more than $400 billion in commitments come due over the next four years. This isn't a forward-dated check that can be dragged out to 2032—nearly half of it has to be paid out in the first four years.

By Foley's own math, that's at least $100 billion a year for the next several years. What that means for a company with $68 billion in annual revenue doesn't require an economist to figure out. After paying for compute, what's left isn't enough to feed the other departments.

Revenue is rising, but the bills are rising faster

Anthropic's revenue this quarter was $17 billion, up 12-fold year over year—it sounds like the beginning of a story. Read one line further and the flavor changes: only 47% quarter over quarter, down from 144% in Q2. The growth rate is falling on its own.

Costs are a different curve. Ed read out three quarters of compute costs: more than $3.3 billion in Q1 2026, $5.6 billion in Q2, $9.9 billion in Q3. Foley added a jab: revenue up 47% quarter over quarter, compute costs up nearly 80% quarter over quarter. Costs are outrunning revenue—this is the one line in the prospectus that needs no explanation.

Prices are still being pushed down. From Q2 to Q3 they did a round of price cuts; the new amendment discloses discounts equivalent to roughly 3% of revenue. The show attributed this to open-weight models flooding the market from China. 3% doesn't sound like much, but for a company whose gross margin is being eaten by compute, this is another slice taken off the revenue side.

Foley had another warning: money is finite. Anthropic holds the enterprise business card—enterprise customers are more willing to pay than consumers and stickier—but enterprise budgets also have ceilings. Google and Meta at least have advertising to fall back on; Anthropic and OpenAI don't. In the end, the question becomes whether to lower service levels, because the tokens genuinely can't be paid for.

The loss figures look okay. Operating loss this quarter was $700 million, versus more than $2 billion a year earlier. But they had just posted a positive operating profit in Q2, and this quarter they fell back into loss. There's also something scarier sitting on the books: non-cash paper losses from the valuation of issued securities and derivatives. Foley said forecasting when this company will be profitable is basically asking someone to predict the unpredictable.

The $200 billion in 2028: you have to believe it before you can calculate it

Anthropic's guidance to the market is $200 billion in revenue by 2028. Foley laid out the math: if it really hits that, a $1.4 trillion valuation alongside something on the order of SpaceX isn't outrageous. He even thinks that, at $100 billion a year in compute spending, $200 billion in revenue would still cover it.

The premise is that you believe it.

Foley was quite candid on this point—he felt he was somewhat abandoning his duty as a financial journalist, because this company asks you to predict things that simply can't be predicted. His exact words were direct: extrapolating next year's growth rate is a foolish thing to do, because the base is too weird—a growth rate like 1,100% carries no information in itself. He added that this company probably shouldn't be stuffed into the casing of a traditional public company at all; the variables are too big.

A new item was added to the amendment's risk factors, and it reads like science fiction. They wrote that the company may already be close to—or may have already crossed—the moment of "AI dramatically accelerating progress," that is, when AI begins to improve itself. At that point, for safety, they would deliberately hit the brakes. What hitting the brakes does to revenue, the prospectus says it doesn't know. A "we may deliberately slow down" written into the risk factors is something that can't be priced.

Rating agencies downgraded the grade first

The first place this IPO gets stuck is the rating. According to the FT, Blackstone has already told investors it expects Anthropic to receive a junk rating—that is, high-yield—after listing. Before that, the market's general expectation was investment grade.

Foley's explanation was blunt: being able to raise money cheaply and quickly is itself a competitive advantage at this stage. Everyone is scrambling, and money is being thrown into the furnace. Without an investment-grade rating, raising capital becomes more expensive and harder. That doesn't mean Anthropic will lose, but it should enter into your calculation of how much you're willing to pay for this stock. What worries him is that many people will treat it as a scarce commodity—there's only one Anthropic in the world, so you'd better get on board first and figure out the rest later.

Meta also did a large equity offering last week. In this race, even the wealthiest players are scrambling for money everywhere.

Ed poured cold water on this: the IPO is aiming to raise close to $100 billion, which at this burn rate is nowhere near enough—they'll have to keep raising more later, most likely by issuing debt.

Debt issuance is replacing cash flow

What Deirdre Bosa discussed, I think, matters more than the Anthropic IPO itself. According to BofA data, AI-related bond issuance was close to $400 billion in 2025, is $655 billion this year, and is projected at $1 trillion next year.

In the past, the balance sheets of these hyperscalers were like fortresses, with free cash flow sufficient to build data centers on their own. This year that changed. Microsoft was the only one that publicly said it could self-fund next year—and then last week it issued $25 billion in debt. Free cash flow at Google, Amazon, and Meta has already turned negative, and Microsoft's modest positive figure is also trending down. S&P also downgraded Meta's rating outlook this month.

She also offered a reminder: this isn't the first time the market has asked about ROI. Generative AI had an earlier bout of anxiety in this cycle, which was later pushed back by demand data. What's different this time is that even the healthiest companies' balance sheets are changing.

Borrowed money has a maturity date; investment returns do not. Bosa used the phrase "cone of uncertainty": the further out you look, the less grasp you have on demand, costs, and returns—yet the money must be committed now, and the fixed debt payments hang on that uncertain future.

The market is already reacting. From October 6 to now, the largest AI and chip stocks have shed roughly $5 trillion in market value, and the Nasdaq 100 has fallen more than 10% from its high. Interestingly, the cloud numbers look great: AWS up nearly 40%, Azure nearly 50%, Google Cloud over 100%. The stocks still fell, because the market is focused on the core businesses being dragged down by AI capex.

San Francisco is still in a bull market; the books are on another table

Bosa is in San Francisco, and she says it doesn't feel like doomsday here at all—optimism is as high as ever. Jensen Huang and Gavin Baker are still talking about demand, and demand is indeed still at a very early stage. She points to the risk in the neoclouds—companies that rent out GPUs to others—as potentially the first link to break in this cycle.

Zandi offers a colder perspective. The total market value of U.S. public equities has doubled in the four years since ChatGPT appeared. Three-quarters of Americans hold few stocks, and a third don't even own a home. Real wages are falling, and inflation is outpacing wages. Back in May he said something: after 36 years as a professional economist, he had never seen the stock market so disconnected from the economy. Before the war, mortgage rates were under 6%; now they're above 7.2%. He grades this economy: the AI part gets an A++, the non-AI part gets a D, and together they don't average out to anything.

Finally, back to that prospectus. The $70 billion comes due next year; the $200 billion in revenue won't materialize until 2028. Two years lie in between—and so does a question no one dares to price: if the $200 billion for 2028 arrives a year late, who fills the hole?

More hardcore podcast content: Podwise