I have sat through a great many AI sales presentations, on both sides of the table. Last week I read something rarer: an AI company describing itself on the record, with its directors answerable for every line.

On 26 October, Anthropic, the company behind Claude, filed its application to list its shares on the New York Stock Exchange. The document, called an S-1, goes to the US Securities and Exchange Commission. Investors who later feel misled can use any sentence in it against the company and the people who signed it. A keynote carries no such risk. A brochure even less.

The growth is real

Revenue was USD41 million in 2023, USD386 million in 2024 and USD4.6 billion in 2025. In the first six months of 2026 alone, it reached USD16.3 billion. More than 11,900 customers paid the company USD100,000 or more over the past twelve months, against roughly 950 a year earlier. Whatever you think of the AI wave, businesses are paying for this and paying more every quarter.

The number a board should frame

The figure investors will love is called net revenue retention. And in Anthropic’s S-1, it’s about 900%. In plain terms, take the customers who were paying a year ago. Today, as a group, they pay roughly nine times as much.

For the vendor, that is the best number in the filing. For the buyer, it is a warning. Hardly any of those customers signed a contract nine times larger. The spend grew through use: more teams, more tasks, more automated runs, each one too small for anyone to approve. The same filing says the latest model cut customers’ cost per task by about 40% compared with the previous one. The price of each unit falls while the number of units multiplies. Cheaper per task does not mean cheaper per year.

With AI you sign once, then keep buying every day, usually without anyone deciding to.

Read the line under the headline

The filing reports a net loss of USD80.7 billion for the first half of 2026. That is a headline waiting to be written. Read one line further: USD72.4 billion of it is a change in the accounting value of convertible notes and similar instruments, which the company says was driven mainly by those securities becoming more valuable. No cash left the building for that part. On its own adjusted measure, the company reports an operating profit for the period. Both headlines, “record loss” and “profitable”, can be written from the same page. Only the reader who goes one line down knows which one matters for his decision. Every vendor case study you receive deserves the same treatment.

One described outcome

The filing names customers. The one with a concrete account of what was done is NASA, whose engineers used Claude to plan and write navigation commands for the «Perseverance» rover on Mars. Insurer AIG and drug maker Bristol Myers Squibb are named without detail, alongside research institutes and software firms that build products on the models. Where the filing describes use in general terms, the list is worth reading slowly here: insurance underwriting, clinical reports for medicines, detection of cyber threats, portfolio analysis for asset managers. Regulated, document-heavy work. That is the daily business of our banks, insurers and management companies, and it is where production use has gone first, not marketing copy or chatbots on websites. The lesson is about evidence, not about this company. The fastest-growing AI vendor in the world, writing to a regulator, gives one described outcome. Any vendor who offers you ten success stories should be asked which of them you can phone.

What is not in the filing

Africa appears once, in a legal notice restricting the offer of shares in South Africa. Mauritius, Kenya or Nigeria are not mentioned, and it’s not surprising. Governments outside the United States are served through AWS’s cloud. I would argue, nothing’s shocking here. A US company is writing for US investors. It is still a roadmap, and our markets are not on it. We are users of these products, not inputs to their plans.

However, the risk is concrete. In June, US export controls led Anthropic to suspend access to its two most capable models for every customer, for almost three weeks. Nobody in Port-Louis or Ébène was consulted. A contract signed here should say where your data sits, how you leave, and what happens to your processes when a model is withdrawn by a decision taken elsewhere.

Three questions before you sign or renew

Which named customer, which measured result, and may I call them? What will this cost once usage grows, and who in our company is allowed to grow it?

If this model disappeared tomorrow, by the vendor’s choice or a foreign government’s, how long before we run on another? None of these questions needs a technical background. They need someone at the table willing to ask them, and a vendor willing to answer in writing.

Anthropic’s lawyers read every line of that filing before it went out. Most of the buyers I meet have read the slides.