Anthropic has filed to go public, and its filing is the first audited look at what it costs to run AI at this scale. In the first half of 2026 it brought in $16.3 billion and spent $8.9 billion on compute and infrastructure. That's roughly $1.80 of revenue for every $1 of compute. It has also promised to buy a lot more compute in the future, and the customers who pay for it can cut back on short notice.

How we sourced this. Revenue and the compute and infrastructure line are from the S-1 on SEC.gov. The version of the filing we could read leaves the aggregate commitment amounts blank, so commitment totals below come from press and analyst summaries and don't agree with each other. Anything not from the filing is attributed. The $16.3B and $8.9B are rounded from $16,286 million and $8,945 million, and the ratio is our arithmetic.

The number inference people should read first

Fortune reports $11.6 billion of Q2 revenue against $5.6 billion of Q2 compute spend, about $2.08 of revenue per dollar. The filing itself says revenue grew 1,368% year over year in the June quarter while compute and infrastructure expense grew 300%. That is the whole bull case in one line: revenue is outrunning the cost of serving it.

The filing also warns that new compute coming online will compress margins, and, per Fortune, that cuts in token prices could change the ratio. So the ratio is both the strength and the thing to watch. If prices per token keep falling, as they have across the market, revenue per dollar of compute has to be held up by volume.

The loss is mostly not the story

The headline net loss of $80.7 billion is mostly a non-cash charge. Fortune reports about $72.4 billion of it comes from revaluing convertible notes. The operating loss for the half was $693 million, and Fortune reports Q2 operating income of $974 million and positive operating cash flow of $3.9 billion. Cash was $112.4 billion at the end of June, per the same coverage. CNBC adds $19.3 billion of short-term investments, for $131.7 billion of liquidity, and says that total is larger than any non-bank US company's except Berkshire Hathaway and Alphabet, based on S&P Capital IQ data. That's a June 30 figure, before any IPO proceeds.

The commitments are the exposure

Anthropic says it expects “periods of excess or underutilized compute and committed cash outflows that do not produce immediate revenue,” and one risk factor says it may be committed to excess compute that remains underutilized. How big the commitment is depends on who is counting. Fortune says at least $475 billion, up to $750 billion including chip leases. Tomasz Tunguz sums about $637 billion from seven named agreements, with Broadcom, SpaceXAI, Google and Amazon the largest. Others have put it near $602 billion or higher. The filing text we read gives a commitment date range running out to 2045 but leaves the dollar figures blank, so we'd treat every total as an estimate until the filed numbers are confirmed. Even on the low end of that range, the $131.7 billion in liquidity covers well under half of it, and the commitments run for years.

On the demand side, Fortune reports that much of Anthropic's revenue lacks long-term commitments, with customers able to cut usage or terminate on little or no notice. About 80% of first-half revenue, $13.1 billion, was usage-based enterprise fees, and remaining contracted obligations were about $10.8 billion. Nearly half of revenue flows through the Amazon, Google and Microsoft marketplaces, and the top five customers were about 25% of revenue. Anthropic says it is moving customers to multiyear deals with fixed minimums in exchange for discounted token rates.

Why this matters beyond Anthropic

This is one company's disclosure, but it describes the structure the whole inference market sits on: long-dated, fixed compute obligations funded by revenue that can be reduced quickly. It is the same shape we described in our look at the web of GPU-backed financings and in the neocloud debt test. In SOI's view, the useful question isn't whether Anthropic is good for its commitments today. It's what happens to everyone downstream, from neoclouds to chip suppliers, if revenue per dollar of compute stops rising.

What we're not saying. We aren't predicting a shortfall. On the figures reported, Anthropic is growing fast, generating operating cash flow and holding over $100 billion in cash. We also haven't reconciled every secondary figure to the filing. Where outlets differ on cost of revenue, margin and commitments, we've said so rather than picked one.

What this means, in plain English

Anthropic is making more than it spends on compute right now, which is a good sign. But it has signed up to buy far more in the future, and what pays for it is customers who can leave quickly. If revenue per dollar of compute holds up, that's fine. If it slips, those commitments get heavy.

What we're watching

The filed commitment totals once the blanks are resolved. Whether revenue per dollar of compute holds as new capacity arrives. How fast usage-based revenue converts to fixed-minimum contracts. And whether other labs and neoclouds disclose their own ratios once they file.

Sources: Anthropic, PBC Form S-1 (SEC), Fortune, Fortune on compute commitments, Tomasz Tunguz, CNBC, KuCoin News.