Wain AI/Tech Blog

AI news and trends worldwide, updated nearly every day

Anthropic's Amended S-1: Compute Costs Rise to 58% of Revenue in Q3 - Short-Term Capacity Among the Causes of an Operating Loss, and a Development-Acceleration Threshold May Be Near

Anthropic's Amended S-1: Compute Costs Rise to 58% of Revenue in Q3 - Short-Term Capacity Among the Causes of an Operating Loss, and a Development-Acceleration Threshold May Be Near

Anthropic's S-1 amendment, filed with the SEC on November 2, covers July to September. Revenue was about $17 billion, compute costs moved from 48% to 58% of revenue, and operating results turned to a loss. It also discloses about $895.4 billion in spending commitments and says it may soon meet an AI-driven development threshold.

On November 2, 2026, Anthropic submitted an amended registration statement (Form S-1/A) to the SEC for its planned IPO1. The first version, filed October 26, stopped its financials at June 30; this one reaches September 3024, so the costs of the company supplying the models and usage allowances are now known through the latest quarter.

For teams that build on Claude, two things stand out: expensive short-term compute weighed on the July to September results, and the company added language saying it may soon reach its own threshold for AI-driven acceleration of development.

Revenue climbed, and so did the cost ratio

The two most recent quarters in the S-1/A compare as follows1.

April-June 2026July-September 2026
Revenue$11,555 million$16,956 million
Compute and infrastructure as a share of revenue48%58%
Operating result$974 million profit$567 million loss

One of the main reasons Anthropic gives for the swing to a loss is $5.0 billion of compute bought on short-term contracts1. In the company’s words, that capacity is currently much pricier than the equivalent secured through multi-year deals.

TechCrunch puts the quarter’s compute bill at about $10 billion2. From a graph in the document, it says revenue per dollar of compute went just above 2x in the prior quarter and then fell to 1.72x, and it explains the extra cost as the price of keeping up with demand.

About $895.4 billion committed, with $213.8 billion conditional

The amount of spending commitments, left blank in the first filing, is now filled in1. Purchase contracts and leases total about $895.4 billion, nearly all tied to compute and infrastructure. Roughly $447.5 billion is expected to be incurred by the end of 2030, and most of the rest from 2031 to 2045. About $69.8 billion is payable in the 12 months after September 30, 2026.

Not all of this is firm debt. About $213.8 billion depends on exercisable termination rights or on conditions before the start date, such as the other party obtaining financing or facilities1. Cash, cash equivalents, and short-term investments stood at about $131.8 billion at the end of September4.

The sourcing model is widening too. Beyond fixed-term agreements with cloud providers, Anthropic now leases chips directly and runs them in dedicated sites, an approach it calls Anthropic Direct1. Long-dated contracts have also surfaced one at a time, such as the seven-year, roughly $11.6 billion Akamai agreement announced in September. Anthropic expects compute spending to grow in dollar terms and to remain its biggest operating cost for the foreseeable future1.

What the filing says about limits and terms of service

The passages that touch users directly sit in the risk factors. Anthropic writes that it has altered usage limits, rate limits, and model or product capabilities, and may do so again1. It also notes that customers who built on older models may resist moving, and that deprecating a model early could cost customers or lead to contract disputes.

On supply, it warns that even with contracts in place it may get less compute than it wants, or get it late, and it discloses having asked providers for more than they agreed to supply1.

These risk statements were already in the first S-1, and they do not announce any change in price or allowances. Still, Claude’s allowances have moved within the past two months, as with the September change to Claude Code’s weekly limit and the late-October rename to Ultra and its revised five-hour limit. The new figures give a sense of how large the procurement bill behind those limits and prices is. For workloads designed around one model or one allowance, they are a reason to have a way to switch when terms change.

New wording on accelerated development

The other change concerns safety. The amended risk factors now say that internal estimates and expectations about future models indicate Anthropic may soon meet its Responsible Scaling Policy definition of “dramatically accelerated progress attributable to AI”13. No current model has been designated as past that threshold, and the data is still being assessed. The filing adds that a major speed-up has gone unrecognized until afterward in the past and could again.

The filing also addresses human involvement. Anthropic says that even once it becomes technically possible, it does not expect to be able to safely hand Claude the full, autonomous design and construction of later frontier generations without meaningful human oversight or understanding, and that it is weighing Responsible Scaling Policy changes so that people stay in charge1. TechCrunch reported on the additions to the risk factors and read them as a warning about nearing recursive self-improvement, while mentioning the skeptical view that an S-1 is partly a sales document3.

In September, the company published an internal measure putting Claude’s “leading” share of AI R&D work at 26% and said nothing it measured was fully autonomous. Now, the approach to the threshold appears in a filing with the SEC. Anthropic says more reports will follow1.

The listing terms are still open. Anthropic has applied to list on Nasdaq’s Global Select tier with the ticker “ANT”, but the per-share price range remains blank in the amendment1.

Sources

  1. Anthropic, PBC Form S-1/A (Amendment No. 1) - SEC EDGAR (filed November 2, 2026)
  2. Anthropic is still grappling with the economics of compute - TechCrunch (November 3, 2026)
  3. Anthropic warns that it’s very close to RSI capability, aka self-building AI - TechCrunch (November 2, 2026)
  4. Anthropic、IPO申請書を修正 7〜9月売上高は約170億ドル、インフラ契約額は約8954億ドル - ITmedia NEWS (November 3, 2026, in Japanese)

We publish the latest AI news nearly every day.

Subscribe via RSS Get new posts the moment they go live.

Search other keywords →