Anthropic’s biggest exposure is not its $80.7 billion net loss, according to Tilo Mitra’s chart-by-chart reading of the company’s S-1 on his blog. It is a stack of compute contracts that must be paid whatever happens to demand.
Mitra, writing on tilomitra.com, puts the contracts with seven suppliers at roughly $637 billion. Broadcom accounts for $161 billion, SpaceXAI for $121 billion, Google for $115 billion, and Amazon for $110 billion. Not all of that is firm, since some amounts are caps and some deals carry cancellation rights. The portion Anthropic owes regardless was $265.6 billion as of June, most of it running to 2033 or later. That is about twice the $131.7 billion of cash and investments on the balance sheet.
The mismatch is in the shape of the money. Obligations are locked in for years, while about 80 percent of first-half revenue was usage-based, meaning businesses pay per use through the API and can scale back in any month. Mitra’s point: a demand dip would not shrink what Anthropic owes its suppliers.
His case against alarm is also in the filing. Revenue per dollar of compute spend was $0.40 in the first quarter of 2025 and $2.08 in the second quarter of 2026. Revenue grew 14.7 times over the past year against 4.0 times for compute, and Q2 produced $974 million of operating profit, a first for Anthropic. Mitra adds that training runs sit inside the compute line too, which means margins on serving customers are better than the chart implies.
The headline loss, he argues, is mostly bookkeeping. Of the $80.7 billion, $72.4 billion is a non-cash remeasurement tied to early investors’ convertible notes and SAFEs, instruments that convert into stock down the road. As Anthropic’s valuation climbed, accounting rules forced the company to book the increase as a loss. Cash from operations came in around $3.9 billion above zero.
Customer concentration is the second weak spot he flags. Forty-eight percent of first-half revenue came through the cloud marketplaces of Amazon, Google, and Microsoft, all of which also build competing models. Another 20 percent came from two unnamed customers, a share the filing says is shrinking. On the other side of the ledger, accounts spending $100,000 or more a year rose from 600 in June 2025 to 11,900 by the end of September, and about 250 spend $10 million or more.
Mitra lays out both readings side by side. The optimist sees roughly 12,000 customers where perhaps a million companies could spend that much, and costs per task falling about tenfold a year. The skeptic sees open-source models pushing prices down and pilots that need proven returns before they become company-wide budgets. He also notes that the filing’s risk factors, which run past 90 pages, dwell more on chips, power, and data centers than on customers.
The unknowns are near. The filing leaves third-quarter revenue blank. Mitra cites Reuters for $16.8 billion, still growth but a sharp slowdown from the $11.6 billion quarter before, possibly linked to Anthropic’s crackdown on others copying its models. A revised filing will fix the IPO price range, and he notes expectations of an $80 billion to $120 billion raise.
The figures are the company’s as presented by Mitra, and the Q2 numbers are unaudited. Still, the arithmetic is plain for anyone pricing this offering: the first test is whether third-quarter growth outruns contract payments that start climbing in 2026 and 2027.
Analysis by Tilo Mitra on tilomitra.com, published 31 October 2026, drawing on Anthropic’s S-1 filing.
