AI compute capacity is expected to grow substantially by 2027. Taking OpenAI and Anthropic as examples, this article examines the revenue needed to finance the additional capacity.

On 16 September, Ursula von der Leyen placed considerable emphasis on AI in her State of the European Union address: its risks, its use in key sectors, and the need to expand Europe’s compute capacity. Against this backdrop, the expansion of data centres expected in 2027 warrants closer examination. The scale of the planned expansion raises serious questions if capacity and payment obligations grow faster than the revenue needed to finance them.

This article does not question AI’s transformative potential. Demand for compute capacity will rise substantially over the longer term. There are, however, doubts about the current pace of expansion. More specifically, can the two major AI labs, OpenAI and Anthropic, generate enough revenue at current prices to finance all the additional capacity scheduled to come online in 2027?

Most of the estimates below come from Dylan Patel of SemiAnalysis, who discussed them on Dwarkesh Patel’s podcast in August. SemiAnalysis is one of the leading analysis and consulting firms in the AI sector and is regarded as a reliable source of industry data. Nevertheless, these projections rest on assumptions, and actual developments could differ substantially.

How much compute capacity could grow by 2027

OpenAI and Anthropic currently rent most of their compute infrastructure. By the end of this year, each company is expected to have access to approximately 5 gigawatts (GW). Patel considers it possible that 50 GW of additional capacity could come online globally in 2027, with OpenAI and Anthropic accounting for 40–50% of that increase. Using the lower end of this range, 40%, the two companies would have access to approximately 30 GW of combined compute capacity by the end of 2027.

According to SemiAnalysis, the major AI labs allocate roughly 50% of their compute capacity to research and training and 50% to inference. This means that approximately 15 GW of inference capacity would need to generate the revenue to finance both inference and research and training.

The revenue needed to finance the expansion

Renting compute capacity typically costs around $15 million per MW per year, equivalent to $15 billion per GW per year. In some cases, prices are even higher. Particularly specific figures are available for Anthropic. As announced in May, Anthropic agreed with SpaceX to pay $1.25 billion per month until May 2029 for exclusive access to SpaceX’s Colossus 1 data centre, which has a capacity of 300 MW. Dividing the annual payments of $15 billion by 300 MW gives rental costs of approximately $50 million per MW per year.

At an assumed price of $15 million per MW per year, using 30 GW for a full year would cost approximately $450 billion. This would cover compute capacity alone. Both companies also need to cover other expenses and would ideally make a profit. The following calculation assumes that other expenses account for 20% of revenue and that the AI labs target an operating margin before interest and taxes of 20%. This leaves 60% of revenue to pay for compute capacity.

Under these assumptions, the two AI labs would need to generate approximately $750 billion in annual revenue in 2027 to cover compute costs of $450 billion.

For comparison, figures released in August put Anthropic’s annualised revenue at approximately $65 billion and OpenAI’s at approximately $40 billion, giving a combined total of around $105 billion. This annualised revenue figure extrapolates July’s monthly revenue over twelve months. If revenue remained unchanged, the two companies would therefore generate approximately $105 billion over a year. Under the assumptions above, annual revenue would need to rise from approximately $105 billion to around $750 billion by the end of 2027—more than a sevenfold increase.

To put this into perspective, Amazon, the world’s largest company by revenue, generated approximately $717 billion in revenue in 2025.

There are also more optimistic estimates. Some AI analysts, including those at SemiAnalysis, assume that revenue per MW could rise to $100 million or more as hardware becomes more efficient and agentic workflows substantially increase the utilisation of data centre clusters. Under this assumption, 15 GW of inference capacity could generate approximately $1.5 trillion in annual revenue.

Falling prices could put pressure on neoclouds

These are enormous figures. Of course, AI labs could generate more revenue by changing their current subscription plans. Many subscribers consume considerably more tokens than their subscription fees would buy at API prices. This alone, however, would not be enough. Another quite plausible scenario is a decline in the prices of compute capacity purchased by AI labs from neoclouds and other providers. These providers’ revenue might then become insufficient to cover their operating costs, interest payments and principal repayments. This could lead to defaults and a substantial market correction, at least in the short term.

What the financial risks mean for Europe’s investments

The EU should therefore exercise caution when subsidising additional capacity. Public funding should require credible evidence of demand and remain justified even at lower compute prices. Where research access or strategic autonomy justify investments beyond commercial demand, these objectives and their costs should be stated explicitly.

Pension funds should also be cautious about investing in data centres whose financial viability depends on high compute prices or which have yet to secure reliable customers. A potential oversupply in 2027 could push down prices and cause revenue to fall short of expectations.

AI’s extraordinary transformative potential remains undiminished. The current expansion could nevertheless create substantial financial risks in the short term that need to be taken seriously.

A few caveats: these estimates involve considerable uncertainty. Suggestions for more robust figures are therefore welcome. The calculation also rests on several assumptions and simplifications. In practice, compute costs could be lower. Finally, it remains uncertain whether and when the approximately 30 GW of combined capacity projected for OpenAI and Anthropic will actually come online. Even under more conservative assumptions, however, the required increase in revenue remains enormous.

This calculation builds on the work of Jon from the excellent YouTube channel Asianometry. He was the first to calculate the revenue implied by these capacity estimates.

This text is licensed under a  Creative Commons Attribution 4.0 International License