AI-generated article. This article was researched and drafted using AI tools and published automatically, and its featured image was generated by AI. Facts are drawn from the sources cited in the text.
Somewhere in a six-person company, a monthly subscription quietly became load-bearing. The invoices are small, the tools work for drafting proposals and summarizing calls and cleaning up spreadsheets, and so nobody looks underneath them. That is worth sitting with this week, because the companies supplying those tools have started saying out loud that AI compute capacity is the one thing they cannot buy enough of. Anthropic’s draft IPO prospectus, reported by The Wall Street Journal on October 20, puts numbers on it.
What the draft filing says about supply, not price
The figures are striking on their own. The Journal reported first-half 2026 revenue of $16.3 billion, up from $1.1 billion a year earlier, and nearly $9 billion spent on computing capacity and infrastructure in six months, against $7.3 billion for all of 2025. The headline net loss of $80.7 billion looks worse than it is: roughly $72 billion of it was non-cash, tied to convertible financing instruments rather than money leaving the building.
The part that matters for anyone running a business on these tools sits further down. Anthropic has reportedly signed noncancelable compute commitments of around $450 billion over ten years, including about $111 billion with Google Cloud, $110 billion with AWS and $31 billion with Microsoft. TechCrunch, citing the same reporting, said total cloud contracts could pass $750 billion. And the filing reportedly states that even that would not be enough, because the company sought more capacity than the hyperscalers were willing to supply, partly because those same providers increasingly treat it as a competitor.
One honest caveat. None of this is a public SEC filing yet. It is journalism about a draft document, Anthropic declined to comment, and the reported totals vary between outlets, from $450 billion to $518 billion to over $750 billion depending on what is counted. The direction is consistent even where the arithmetic is not.
AI compute capacity is a supply problem, not a cost problem
Price has been the familiar worry, and a reasonable one. What AI tools actually cost a small business is a question with a fairly stable answer right now, because competition has been pushing per-token prices down. Supply runs the other way, and it shows up differently: rate limits at busy hours, slower responses, a waitlist for the newest model, or a tool that simply changes underneath a workflow that had been working fine.
There is a concrete example from the same week. On October 14, OpenAI told developers that GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini leave the API on January 14, 2027, fine-tuned versions included, with new fine-tuning jobs accepted only until January 6. That is about three months of notice. The suggested replacement, GPT-6 Luna, is cheaper per token than any of the four, but it reasons by default and cannot be fine-tuned at all. Cheaper is not the same as compatible, and a business that spent a year tuning prompts around one model’s behavior is doing that work again. Model retirement has already caught small firms once this year.
Who gets served first
Here is the editorial judgment, flagged as such rather than dressed up as a disclosure. Two unnamed customers accounted for roughly 20% of Anthropic’s first-half revenue, per the Journal’s reading of the draft, and Fortune’s earlier account of a Reuters-sourced version put the figure closer to a quarter. When capacity is short, it is rarely shared out evenly. The accounts with the largest commitments and the longest contracts tend to keep their throughput, and the $29-a-month seat is not where a provider defends its margin.
The same shortage is why the tools got good
None of this reads as a reason to retreat. Bloomberg reported on October 19 that Anthropic is in talks for a second compute deal with Nscale worth around $45 billion. Reuters reported OpenAI racing to close roughly $30 billion at a valuation above $1.4 trillion, ahead of Anthropic’s expected November listing. SemiAnalysis expects about 50 gigawatts of new data center capacity to arrive in 2027 with another 60 gigawatts under construction, and suggests chips and memory, rather than power, may become the next bottleneck.
That is an extraordinary amount of money being spent so that a small company can rent capability it could never build. Both things are true at the same time. The tools are cheap and genuinely useful because of this race, and the terms underneath them are less settled than a monthly invoice makes them look.
One small thing worth doing this week
Not a migration plan. Nobody needs a vendor risk framework for three subscriptions. Something smaller: write down the three AI tasks that would hurt most if the tool stopped working tomorrow. For each one, note where the prompt actually lives, whether anyone other than its author knows how it works, and whether the output is stored somewhere the business owns. That is an afternoon, not a project, and it is most of the useful answer to the build-versus-buy question.
No small business is going to fix a global compute shortage, and most will never feel it directly. What a small business can control is whether its own work is portable. Which leaves a question worth asking before the next renewal comes round: if that tool went away in ninety days, how much of the work would go with it?
