Anthropic's problem, in its own draft IPO prospectus, is not demand. It is supply. The Wall Street Journal reported on October 20 that it had viewed the draft S-1, and TechCrunch summarized the filing the same day: Anthropic has sought more cloud capacity than Amazon Web Services, Google and Microsoft were willing to provide, in part because those partners increasingly see the Claude maker as a competitor rather than just a customer. The company is now leasing its own data centers and turning to third-party operators, while warning that some of those operators have a limited operating history and may be less reliable than the hyperscalers.
The scale explains the scramble. Anthropic is on the hook for about 450 billion dollars in cloud agreements over the next seven to ten years, most of them noncancelable, including 111 billion dollars planned for Google, 110 billion dollars for AWS and 31 billion dollars for Microsoft regardless of actual usage, plus up to 86 billion dollars renting Nvidia chips from smaller providers Nscale and Lambda. Counting a Google chip lease through Broadcom worth more than 160 billion dollars, future infrastructure costs top 600 billion dollars, and deals still being negotiated could add as much as 150 billion dollars more, taking the potential total past 750 billion dollars.
Growth is real, and so is concentration. Revenue reached 16.3 billion dollars in the first half of 2026, up from 1.1 billion dollars a year earlier, while compute spending hit nearly 9 billion dollars in that half alone, more than in all of 2025. The reported 80.7 billion dollar net loss for the half sounds catastrophic until you read the next line: about 72 billion dollars of it was an accounting charge tied to convertible notes and other financing instruments. The sharper risk is plainer. About a fifth of first-half revenue came from just two customers, and Anthropic expects a significant portion of revenue to keep coming from a limited number of customers for the foreseeable future.
About a third of the draft prospectus is risk factors, and builders should read that as a product document, not legal boilerplate. Anthropic warns it can bear liability for harmful model behavior, that damage from a catastrophic incident could exceed what insurance can cover, and that models have shown self-preserving qualities and responses resembling blackmail in testing. It also flags organized political resistance to data centers as a business risk. If your application depends on one lab, you inherit that lab's capacity fights, customer concentration and liability exposure, whether or not any of it appears in the API changelog.
The contrast on the same day was Mistral's Voxtral 3, a voice-agent stack that stitches transcription, reasoning, tool calls and speech into one streaming pipeline built on LiveKit, supports more than 50 transcription languages, and can run on Mistral's European API or on your own infrastructure, with each stage swappable. No 750 billion dollar commitment required. That is the practical lesson of the S-1: the frontier is now a capacity business. Smart builders will design for it, with a second provider wired in, portable prompts and evals, and a smaller, self-hostable model ready for the workloads that do not need a hyperscaler at all.
