Anthropic’s blockbuster IPO: The next great growth stock or an AI gamble too far?

For three years the bear case on AI has been simple: nobody is making money from it. Anthropic’s prospectus, the first detailed look inside one of the big AI labs, says otherwise.

Not the cash bonfire you were promised

Revenue went from US$41 million in 2023 to US$4.6 billion in 2025. In the June 2026 quarter alone it was US$11.6 billion, more than double the whole of 2025. That quarter brought its first operating profit, of US$974 million. Gross margins were 85% in the first half, up from 12% a year earlier. Anthropic looks to have reached an inflection point, where growth increasingly turns into profit.

And it has done that while barely touching the money it has raised. By the end of June it had raised US$141 billion in total, and still held US$132 billion of it in cash and investments.

The number that will make headlines is the US$80 billion first-half loss. It’s an accounting quirk. Anthropic issued convertible notes, debt that turns into shares, and as the company became more valuable, so did the notes. Accounting rules treat that rise as a loss: about US$72 billion of it.

The worry was that training costs would swallow the profits. In the first half, revenue rose 14-fold year on year while the compute bill, for the chips and data centres that run its models, rose less than four-fold. Even after paying for all its research and training, Anthropic kept 26 cents of every revenue dollar, before sales and admin costs.

In an early sign of broad business adoption, customers paying more than US$100,000 a year rose from 950 to 11,900 in twelve months.

The scary number is a shopping list

The filing’s most alarming figures are its compute commitments, but computing capacity has to be booked a year or more ahead. On our central estimate, Anthropic’s compute costs reach about US$95 billion in 2028, much of it already signed for.

That supports US$130-200 billion of revenue, depending on whether each compute dollar earns 1.4 or keeps earning two: up to three times its current annual pace. With other costs growing more slowly, much would fall to profit. The commitments are a capacity plan, not a warning.

If its models fell behind, that capacity wouldn’t be stranded. Those chips could run anyone’s models, and would sit idle only if the world wanted less AI. Mark Zuckerberg said of Meta’s build-out: “in order for it not to work out, it’s not just that we have to fail, AI overall has to fail.”

A race it has to keep winning

Growth has slowed. Anthropic reportedly told investors revenue rose 45% to US$16.8 billion in the three months to September, after growth of 144% in the June quarter and 132% in March. Keep that up for a year and revenue more than quadruples. 

The question is what drove the slowdown: did customers want less, or did Anthropic run out of chips? The filing says demand “exceeds available compute”, and the biggest new capacity arrives from December.

The surest way to keep a lead in AI is to let your best model build the next one. Claude wrote more than 90% of the contributions to Anthropic’s codebase in September, which is how 5,300 people ship a new frontier model every few months. Rivals do the same, but the lab with the best coding model gets the most from it, and today, in our view, that’s Anthropic.

One risk is a rival overtaking Claude: OpenAI, Google, xAI’s Grok, Meta’s Muse and fast-improving Chinese open models are all chasing it. The bigger risk, however, is customers deciding that a model a few steps behind the frontier is good enough. 

IBM’s Arvind Krishna asks: “Would you like something that’s 10% of the cost if it’s 90% as good as the best?” But Krishna added that some tasks still need the best, and coding is one: code that’s 90% right doesn’t run. “Good enough” needn’t mean leaving Anthropic, either. It sells a range, from its top Fable model down to the cheaper Sonnet and Haiku, and cheaper models make AI a bigger market.

The other key risk, of course, is the end of the world. Most prospectuses warn about interest rates and key staff. Anthropic’s warns that advanced AI could “pose catastrophic or existential risks to humanity”, that its models have been caught “granting themselves permissions they had not been given”, and that possibly the entire insurance industry couldn’t cover the worst case.

Priced for a lot going right

Reports put the expected valuation at US$1.5-2 trillion, at least 22 times the September quarter’s reported revenue, annualised. What decides this stock’s success is the world in five years. Is AI adoption near the peak, or at the very beginning? Moody’s says use of its research inside customers’ AI tools rose more than tenfold in three months, adding: “It is early days.” We agree, and think demand in five years will be many multiples of today’s.

The filing also answers a big open question about the AI build-out. The chipmakers are already highly profitable. Our deal-by-deal work on data centres and the new GPU-rental companies, or neoclouds, finds projects earning around their cost of capital, and new GPUs rented today earning well above it. The missing piece was whether the labs paying for it all could make money. This filing shows one can, and that flows through the supply chain, which should lift the whole sector.

We can’t say whether we’ll take part, but this is one of the most interesting IPOs in years, and worth every investor’s attention. 

As the listing nears, we’ll be watching the price, the annualised October revenue figure expected in an amended prospectus, and whether Claude holds its lead, including through the launch of Fable 5.5, its next public model, expected shortly.

Anthropic has shown that people will pay for intelligence, and that serving them is highly profitable. The float will test whether it can stay in front.

Thomas Rice and Armina Rosenberg are the co-founders and portfolio managers of Minotaur Capital.