Deep Stackby Burak
All posts
6 min read

The $895 Billion Promise Behind the $17 Billion Quarter

Anthropic made nearly $17 billion in three months and still lost money on operations. The interesting thing is not the quarter. It is the promise underneath it: $895 billion of compute, signed, leased, and still to be priced.

The $895 Billion Promise Behind the $17 Billion Quarter

Anthropic filed Amendment No. 1 to its IPO prospectus on November 2, 2026, filling in the third quarter that the original filing, made public on October 26, had left blank. The numbers, as reported from the filing, are large in both directions at once. Revenue for the three months to September was $16.96 billion, up 47 percent from $11.56 billion in the second quarter and up 1,127 percent from $1.38 billion a year earlier. Over the first nine months of 2026, revenue reached $33.24 billion, against $2.55 billion a year earlier. That is growth at a scale most companies never touch.

And yet the quarter ended in an operating loss of $567 million, after an operating profit of $974 million in the second quarter. The swing is the story. It is not a story about demand falling. It is a story about what the demand costs to serve.

The quarter that got more expensive to win

Compute and infrastructure costs rose 77 percent in three months, to $9.86 billion in the quarter, according to the filing as reported. That is 58 cents of every revenue dollar, up from 48 cents in the second quarter. Anthropic attributes part of the jump to $5 billion of compute bought under short-term deals, which it says currently costs significantly more than capacity under its multi-year contracts, plus $2 billion of non-cash charges tied to its staff donation matching programme.

Strip out stock-based pay and those non-cash charitable gifts and the picture softens but does not change direction: on Anthropic’s own adjusted measure, operating income rose to $2.5 billion, while the adjusted margin fell from 19 percent to 15 percent. Revenue per dollar of compute fell from the previous quarter. The machine is getting bigger faster than it is getting cheaper to run.

The headline net loss needs the same care. The quarter’s net loss was $23.29 billion, and $104.04 billion for the nine months, but most of that is an accounting charge rather than cash leaving the door: $21.94 billion in the quarter came from the rising value of convertible notes, SAFEs, and similar instruments that will turn into shares. Day-to-day operations brought in $6.6 billion of cash over the nine months, and Anthropic ended September with $95.6 billion in cash and $36.3 billion in short-term investments. The company is not running out of money. It is running into a cost structure that grows with its revenue, and sometimes faster.

One reassuring detail: about 79 percent of third-quarter revenue came from customers outside the top five, up from about 64 percent in the first quarter, and no single customer exceeded 10 percent of nine-month revenue. The base is broadening. The bill underneath it is not. It is concentrating into a small number of very large, very long promises.

The bill, in pieces

The biggest new number in the amendment is the size of Anthropic’s commitments. As of and since September 30, it has signed purchase commitments and leases worth about $895.4 billion, substantially all of it for compute and infrastructure. The October filing left that figure blank. A confidential draft reported in September had put the commitments at $518 billion. The new total is not a rounding revision. It is a different scale of bet.

It breaks down, as reported from the filing, into $320 billion of contracts already on the books at the end of September, mostly hosting deals with cloud providers; $451.2 billion of leases not yet started, mostly due to begin between 2026 and 2028 and to run about ten years on average; $100.2 billion of deals signed since September 30, mainly data centre leases and cloud hosting; and about $24 billion of short-term deals. About $447.5 billion is due to be spent before the end of 2030, and almost all the rest between 2031 and 2045, with $69.8 billion due in the twelve months to September 2027.

That last figure deserves a pause. $69.8 billion due within a year, set against $131.8 billion of cash and short-term investments as reported, is payable, not a crisis. It is a treadmill: the next twelve months alone require more compute spending than most companies will ever spend, and it runs, on paper, into the 2040s.

Two caveats keep the $895 billion honest. About $213.8 billion can be cancelled or depends on conditions such as a partner securing financing, sites, or approval. And the total excludes deals with no fixed price, estimated at another $150 billion to $175 billion as capacity arrives. The headline is both overstated and understated at once — overstated as a firm obligation, understated as intent. Either way, Anthropic expects capital spending to increase significantly.

The price range is still empty, and run-rate revenue as of October 31 is still blank. The Nasdaq listing, under the ticker ANT, has not been priced yet. Investors are being asked to back commitments worth about thirteen times the current annual sales rate before they know the price. That is the bet at the heart of the AI bubble debate, and the filing does not hide it. It itemises it.

Why this is a builder’s problem, not just an investor’s

That sounds like a capital-markets story. It isn’t, or not only. Every token price a builder pays is downstream of a commitment like this one.

When a frontier lab signs for hundreds of billions of dollars of capacity, it has to fill it. Unused capacity is the most expensive kind, because the lease is due whether the tokens flow or not. That creates permanent pressure to grow usage, bundle, and discount to fill the pipes — and, when short-term capacity must be bought at a premium, to pass that premium on. The $5 billion of short-term compute this quarter is a small preview: when long-term contracts cannot cover the load, the spot price shows up in the margin.

For builders, the takeaway is practical: do not build a cost model that assumes the current token price is a law of nature. It is the output of a financing decision made by someone else, at a scale you cannot match and on a timetable you do not control. The quarter shows both sides of that in one document — revenue growing twelvefold in a year, and the cost of serving it growing fast enough to erase an operating profit anyway. A price cut announced in that environment is not generosity. It is utilisation strategy. A price rise is not greed. It is the same strategy, on a different quarter.

That is also why the local-model argument keeps coming back, and why it is not nostalgia. Owned weights do not solve the frontier-capability problem, and nobody serious claims they do. What they solve is the exposure problem: a workload that runs on hardware you already own does not get repriced when a lab’s short-term compute bill spikes, and it does not disappear when a lab’s roadmap changes. The $895 billion is, among other things, a measure of how much of the industry’s future compute is rented rather than owned. Builders who want predictable costs should keep at least part of their stack on the owned side of that line, even if the rented side is where the newest capability lives.

The Hacker News discussion, at 150 points and 138 comments when this run fetched the front page, focused on exactly this tension: whether growth this fast can ever outrun a commitment this large, and whether the accounting losses distract from the operating question. The commenters are right that the $104 billion nine-month net loss is mostly a valuation artefact. They are also right to be less comforted by that than the headline suggests, because the operating line — the one that excludes the artefact — still went negative in the quarter that was supposed to prove the model.

Anthropic is not failing. A company with $17 billion in quarterly revenue, $6.6 billion of nine-month operating cash flow, and a broadening customer base is not failing. But it has now told the market, in its own filing, that its future is a $895 billion promise to buy compute before it has sold the tokens to fill it. The interesting thing about that promise is not its size. It is what it commits everyone downstream to: a decade and more of filling someone else’s data centres, at prices set by someone else’s leases.

That is the bill behind the quarter. It will be collected, one token at a time.

Source: Anthropic IPO filing: $17bn quarter, $895bn compute bill

Latest from Deep Stack

5 min read

Open Weights, Closed Map

Baidu's ERNIE 5.5 Flash Base is downloadable, cheap per token, and 280 gigabytes of fine print.