For three years the cost of building a frontier AI company has been a matter of educated guessing. This week the guessing got a document. A draft IPO prospectus from Anthropic, reported from figures circulating ahead of a potential listing, put hard numbers where there had only been rumour, and the numbers are not shy. Roughly $4.6 billion in 2025 revenue. A loss of about $42 billion in the same year. Future cloud and compute commitments totalling around $518 billion. And a valuation the company is said to be chasing above $2 trillion.
Line those up and the shape of the business becomes clear, because it is not really the shape of a software company. Traditional software sells the same product many times at almost no extra cost, which is why software margins are famously fat. Anthropic is doing something closer to running an industrial operation: it must secure chips, electricity, data centres, and cloud capacity before it can sell a single token, and those commitments dwarf the revenue they are meant to produce. The $518 billion it has promised to spend is more than a hundred times last year's sales. That is not a rounding error waiting to be optimised away. It is the business model.
The question the prospectus quietly forces is whether public investors will accept that trade. If they do, Anthropic establishes a new template for valuing a frontier lab, one where locked-in access to compute and power counts for as much as gross margin, and the enormous losses read as an infrastructure down payment rather than a warning. If they do not, every large model developer preparing to raise capital, OpenAI and xAI among them, inherits a harder set of questions about whether revenue growth, however spectacular, can ever outrun the spending it depends on. Anthropic devoted a substantial chunk of its filing to the risks of building ever more capable systems, which is candid, but candour does not change the arithmetic.
Widen the lens and the same tension appears at industry scale. Bain & Company, in its annual technology report, calculated that the AI sector will need to generate roughly $6 trillion in annual revenue by 2031 for the current pace of data-centre construction to make economic sense. Existing consumer and enterprise AI services might supply $1.8 trillion of that. The remaining $4.2 trillion would have to come from categories that barely exist today: autonomous machines, robotics, drug discovery, energy systems. Bain's own technology chief said the buildout requires "a wave of innovation that will dwarf what mobile and cloud unlocked." That is another way of saying the demand has been promised but not yet invented.
Meanwhile the cash that has already been made is unmistakable. Nvidia, sitting at the chokepoint of the whole enterprise, authorised an additional $150 billion in share buybacks this week, one of the largest such programmes any American company has announced. Hyperscaler capital spending at Microsoft, Google, Amazon, Meta, and Oracle could approach $780 billion in 2026, close to five times the level of three years ago. The money at the bottom of the stack, the chips and the clouds, is real and flowing. The money at the top, the businesses large enough to justify multi-gigawatt compute factories, is the part still being sketched.
What makes Anthropic's filing valuable is not that it settles the argument but that it finally shows the ledger both sides have been arguing about. Bulls can point to $4.6 billion in revenue built almost from nothing and read the losses as the price of a generational land grab. Sceptics can point to a $42 billion hole and $518 billion in obligations and ask who, exactly, is going to buy enough intelligence to fill it. Both are looking at the same page now. The next twelve months, as Anthropic tests whether public markets will fund an industrial operation dressed as a software company, will start to tell us which reading the money believes.