For years, the companies building the most powerful AI models have kept their accounts private. That is about to change. A draft of Anthropic's IPO prospectus, the S-1 filing a company must make before listing its shares in the US, has been seen by The Wall Street Journal and Reuters, and it offers the most detailed look yet at what it costs to run a frontier AI lab. The short version: revenue is soaring, the reported loss is enormous, and the bill for computing power is on a scale that is hard to picture.

The document has not been published, Anthropic has not commented on the reports, and figures in a draft can change before the final filing. But the numbers reported so far are consistent across several outlets, and they go a long way towards explaining why the AI industry looks the way it does.

The headline numbers

According to the Journal, Anthropic, the company behind the Claude chatbot and the Claude Code programming assistant, made $16.3 billion in revenue in the first half of 2026. In the same six months of 2025 it made $1.1 billion, and across the whole of last year it made $4.6 billion. In other words, it took in roughly three and a half times last year's full-year revenue in just six months. The company has separately told investors that its third-quarter revenue alone reached $16.8 billion.

The loss is the figure that has drawn the most attention. The draft reportedly shows a net loss of $80.7 billion for the first half of the year, up from $13 billion a year earlier. That sounds alarming, and it is certainly large, but most of it is not money that left the building. Around $72 billion is an accounting charge linked to changes in the value of convertible notes and other financial instruments Anthropic used to raise money. As the company's value has risen, so has the value of bonds that can turn into shares, and accounting rules require that increase to be recorded as a loss. It is a real entry in the accounts, but it is not a cash cost of running the business.

Where the money really goes: computing power

The more telling numbers are about compute, the vast banks of specialised chips needed to train and run AI models. Anthropic reportedly spent nearly $9 billion on computing capacity and infrastructure in the first half of 2026, against $2.3 billion a year earlier and $7.3 billion in the whole of 2025.

The forward commitments are far bigger. The draft lists around $450 billion of cloud agreements over the next seven to ten years, most of which cannot be cancelled. The largest reported deals are $111 billion with Google Cloud, $110 billion with Amazon Web Services and $31 billion with Microsoft. Including other infrastructure, total future commitments are said to be above $600 billion, with roughly another $150 billion under negotiation. Put together, Anthropic has signalled it could spend up to $750 billion on computing in the coming years.

To put that in context, Reuters reported only last month that the company had committed to at least $518 billion over a decade. The figure has grown by well over $100 billion in a matter of weeks.

There is one encouraging sign for investors buried in the numbers. Revenue grew around fifteenfold year on year while compute spending grew roughly fourfold, which suggests each dollar spent on chips is starting to bring in more income than before.

The risks Anthropic has to spell out

An S-1 has to list the things that could go wrong, and this one makes unusual reading. According to the reports, Anthropic warns of a narrow customer base and of uncertain access to the computing power it depends on. It also flags political resistance to new data centres in the US, where most of its capacity sits.

Then there are risks that only an AI company would need to write down. Bloomberg reported that the filing mentions models showing "self-preserving" behaviour to avoid being switched off, and responses that resemble "blackmail". It also warns that its products could be misused, including for sophisticated cyberattacks. Anthropic has been unusually open about this kind of behaviour in its own safety research, but seeing it listed as a business risk in a share prospectus is something new.

The draft also covers pay. Chief executive Dario Amodei and his sister Daniela, the company's president, each received a base salary of $655,000 in 2025, according to Fortune, with most of their pay in shares. Their total packages were valued at $18 million and $16.5 million respectively.

What it means

For anyone trying to understand the AI boom, this is the clearest picture yet of its economics. The demand is real: businesses are paying for Claude, and particularly for Claude Code, at a pace few predicted. But the cost of keeping up is staggering, and almost all of it flows to a handful of cloud providers and chipmakers. The same week, SpaceX reported that its AI business had overtaken its satellite internet arm, as we covered in our story on the Tesla and SpaceX merger talk. Wherever you look, AI is now the biggest line on the balance sheet.

It also shows how much rides on continued growth. Hundreds of billions of dollars of contracts that mostly cannot be cancelled only make sense if revenue keeps climbing for years. Analysts quoted by Bloomberg were split: one called the spending plans positive for the chip supply chain but said whether they materialise is "highly uncertain", while another said $750 billion would mean more demand for AI hardware over the medium to long term.

What happens next

Bloomberg has reported that Anthropic could start marketing its shares as soon as the week of 9 November, which would allow trading to begin before the US Thanksgiving holiday. The Journal says the company will seek a valuation of around $2 trillion. Not every investor is convinced: Reuters reported that some who attended Anthropic's investor day last week discussed a figure closer to $1.5 trillion after seeing the latest disclosures.

The public version of the S-1 is expected within weeks, and it will be worth reading closely for any changes to these numbers. If the listing goes ahead at anything like the reported valuation, it would rank among the largest stock market debuts ever, and the first real test of how ordinary investors value a company whose product is, in the end, an AI model. We will report on the filing when it is made public.