Reuters reported on September 28 that it had reviewed a draft initial public offering prospectus from the artificial intelligence company Anthropic. The figures inside are the clearest look the public has had at the economics of frontier model development, and they are not what the equity market's enthusiasm has assumed.

The document reportedly shows 2025 revenue of $4.6 billion, up 1,088% on the prior year, against an operating loss of $8.06 billion, which itself was $2.98 billion the year before. Compute and infrastructure spending tripled to $7.33 billion. Most striking of all is a figure for future cloud, computing and infrastructure obligations: $518 billion.

What a commitment number means on an income statement

This is a draft that leaked rather than a filing that was made, and the company has not confirmed it, so every number here carries that caveat. Treat it as the best available estimate rather than as an audited fact.

With that said, the $518 billion line is the one that matters for anyone holding technology shares, and it is worth understanding what kind of number it is. A purchase commitment is not an expense yet. It is a contractual promise to buy capacity over a period of years, and it sits in the notes rather than on the face of the accounts. For the company making the promise it is an obligation. For the company receiving it, it is booked revenue in waiting.

That is the mechanism by which private artificial intelligence losses become public company earnings. The capital spending that hyperscalers have been reporting, and the power and semiconductor demand underneath it, is substantially backed by commitments of this shape from a small number of private counterparties.

The concentration runs in both directions

Whitmore's Watchlist:
AMZN (Amazon.com, Inc.): Cloud capacity and a disclosed investor relationship with the company in question.
GOOGL (Alphabet Inc.): The other large cloud provider and investor in the same orbit.
NVDA (NVIDIA Corporation): The hardware layer every one of these commitments eventually pays for.

The prospectus reportedly notes that two unnamed customers accounted for close to a quarter of 2025 revenue, with no long-term contracts guaranteeing that spending continues. So a company promising half a trillion dollars of future compute purchases derives a quarter of its own revenue from two customers who are not contractually obliged to stay.

That is a chain of concentration, and it runs through public portfolios whether or not the investor holds a single artificial intelligence stock directly. An index fund weighted by market capitalisation has an unusually large share of its value in companies whose recent earnings growth depends on this spending continuing. Semiconductor profits were up more than 100% in the last reported quarter. Utilities are signing power agreements on the same assumption.

None of that says the spending stops. It says the earnings growth of several large sectors rests on a smaller number of contracts than the diversification in the fund name implies.

The risk factor nobody usually writes

The document also reportedly contains language about existential risks to humanity, warning that autonomous systems could behave unexpectedly, create security problems, or be used for fraud and manipulation. Separately, reports this week said a competitor had shelved a frontier model over safety concerns.

Risk factors in a prospectus are written by lawyers to be comprehensive rather than predictive, and a reader should discount them accordingly. What is unusual here is not that the language is dramatic but that it points at a commercial variable investors have not been pricing: the possibility that a product cycle gets slowed by its own developers, or by regulation that follows this kind of disclosure.

The target valuation reported is above $2 trillion, against roughly $965 billion in May 2025, with an offering expected after the November midterm elections. If that happens it becomes one of the largest listings ever attempted, and the capital to buy it has to come from somewhere, which usually means selling something else.

For a portfolio, the practical question is not whether to have a view on artificial intelligence. It is how much of what you already own is a claim on this specific spending continuing at this specific pace, and whether you would have chosen that exposure deliberately.

Whitmore's Take: A half trillion dollar purchase commitment from one private company is the clearest measure yet of how much public company earnings depend on private artificial intelligence budgets. Worth checking how concentrated your index exposure is before that becomes a listed security competing for the same capital.

Written by Daniel Whitmore
Millionaire Insiders