Nearly a third of Anthropic's IPO prospectus is risk factors. That is the most useful thing any AI company has published this year, and most buyers will never read it.
Sales decks are written to close. Prospectuses are written so nobody can sue later. The difference in honesty is enormous.
According to TechCrunch, summarising reporting from Reuters and the Financial Times, the filing includes numbers that every company running Claude in production should understand. Not because Anthropic is uniquely risky. Because it is the first frontier lab forced to show its books in this much detail.
The numbers behind your API bill
Revenue jumped twelvefold in 2025 to nearly $4.6 billion. Operating expenses reached almost $13 billion. The operating loss was more than $8 billion, driven by compute.
Then 2026 moved faster. Second-quarter revenue alone reached $11.5 billion, and the company is on track for a second straight quarter of adjusted operating profit. The filing also describes plans to spend $518 billion on cloud, compute and infrastructure in the coming years.
Read those together and a picture forms. This is a business whose costs are set by infrastructure contracts measured in hundreds of billions, and whose revenue only recently caught up. Your price per token sits somewhere between those two numbers.
I wrote about Anthropic doubling again in ten weeks in August. Growth like that is real. It also means the pricing you negotiated last year was set by a company in a very different financial position from the one it occupies now.
Customer concentration cuts both ways
The detail I would circle is this one. Nearly a quarter of last year's revenue came from just two clients.
For investors, that is a concentration risk. For you, it is information about whose roadmap gets priority. When two customers pay a quarter of the bills, their feature requests, their rate limits and their deprecation timelines carry weight that yours will not.
This is not a criticism. Every enterprise software company goes through this phase. But it should change how you write the contract.
If your product depends on one model from one vendor, you are a small customer of a company whose biggest customers are much bigger than you. That is the position where deprecations surprise you and pricing changes arrive as policy updates rather than negotiations.
The risk section is a procurement checklist
The filing reportedly describes model behaviours the company says have appeared or could appear, including attempts to resist shutdown and to conceal or manipulate information. It also includes existential risk language, apparently a first in the SEC's database.
You can read that as marketing, as sincerity, or as legal cover. For a buyer, it does not much matter which. What matters is that the vendor has now put in writing, under securities law, that its product can behave in ways it did not intend.
That is a sentence your risk team should see. Not to stop using the model, but to design the deployment as if the sentence is true.
In practice that means human review on anything that touches money or customers, logs you can audit, and permissions scoped to the task rather than to the account.
The same logic applies to every other vendor. OpenAI, Google and Meta have not filed an equivalent document, but the behaviours are not unique to one lab. This month alone, OpenAI disclosed that its agents accessed systems they were not authorised to reach during training runs. The prospectus simply says out loud what the rest of the industry is still discussing in blog posts.
What to do with this in the next renewal
Three practical changes follow, and none of them require picking a side on AI safety.
First, write model portability into the architecture. Keep prompts, evaluation sets and retrieval layers vendor-neutral so a switch costs weeks, not quarters. I made this case in three reminders that you rent your AI, and the prospectus makes it again with better data.
Second, negotiate for notice periods, not discounts. A 20 percent discount is worth less than a guaranteed 12-month deprecation window when your product runs on a specific model version.
Third, read the filing yourself when it is public. Not the headlines about existential risk, the sections on compute commitments, gross margin and customer concentration. They tell you how much pricing pressure is coming, and from which direction.
There is also a second-order effect worth watching. A public Anthropic will report quarterly, and every quarter analysts will ask about margin.
Margin pressure at an AI lab turns into one of three things for customers: higher prices, tighter rate limits, or quieter changes to which model actually serves your request. Plan for all three.
Disclosure becomes the industry baseline
The effect reaches well beyond one company. Once one frontier lab publishes audited numbers, every competitor gets measured against them, publicly listed or not.
Investors will ask OpenAI and others for comparable figures on compute commitments, gross margin and customer concentration. Enterprise buyers will ask too, and they will have a benchmark to point at when the answer is vague.
That is good news for procurement teams. For three years, AI vendors sold on capability demos and benchmark charts. The next three will increasingly be sold on cost structure and reliability, because that is what a public filing forces everyone to talk about.
It also changes how you should read a vendor's silence. If a private lab will not share basic information that a listed competitor is legally required to publish, that tells you something about the negotiating position you are in.
The best due diligence document in AI is about to become public. The companies that read it will negotiate differently from the ones that only read the launch posts.