Sam Altman Warns AI Compute Buildout Is Showing ‘Unsustainable Silliness’

Sam Altman is warning that the AI industry’s race to add computing capacity is beginning to outrun its commercial foundation. The OpenAI chief said he sees early signs of “unsustainable silliness” among new neocloud providers proposing huge expansions without enough revenue or an identified buyer to support them.

The warning is not that all new AI capacity lacks demand. It is a test of how the capacity is financed and used: Altman’s concern is directed at projects whose announced scale exceeds the revenue or customer commitments behind them. He separated OpenAI from that group, saying he expects the company to use its planned compute profitably while worrying about the industry-wide buildout.

That distinction puts utilization at the center of the debate. A signed customer commitment can give a data-center operator a clearer path to paying for capacity, while an ambitious development pipeline is only a plan until workloads arrive. It also explains why headline capacity figures alone do not resolve Altman’s concern.

Contracts provide a counterweight, not a clean answer

Several listed infrastructure operators have disclosed demand that complicates any blanket reading of Altman’s warning. CoreWeave reported about $104 billion in revenue backlog as of June 30 and more than $25 billion in net new customer commitments early in the third quarter. Nebius disclosed $37.5 billion in remaining performance obligations, though three customers supplied 24%, 21% and 14% of its second-quarter revenue.

Other disclosed demand markers

• IREN said roughly $4 billion of annual recurring revenue was contracted for its 2026 capacity, with $1 billion operating at the time of its earnings call.
• Hut 8 disclosed 949 MW of contracted IT capacity and about $26.6 billion in expected aggregate base-term contract value.
• Bitdeer signed a 16-year, $4.70 billion colocation lease with AI lab Volta for 121 IT MW at its Tydal, Norway facility.

Those commitments do not erase the financing and execution questions. CoreWeave’s second-quarter interest expense was $640 million, up from $267 million a year earlier. Cipher Mining reported about $6 billion of debt against $562 million in stockholders’ equity, while targeting roughly 5.3 GW of total portfolio capacity by 2030 or later.

Efficiency is the risk hiding inside the buildout

Altman identified a second way the economics can change. If OpenAI substantially reduces compute costs and improves efficiency, he said, expensive capacity projects could become bad financial bets. The risk is not merely too little demand: technology that produces more AI work from the same hardware can alter the value of capacity built around older cost assumptions.

He also acknowledged a macroeconomic exposure closer to home: a broad downturn could strain OpenAI’s ability to pay for capacity it has already committed to, although he characterized that risk as manageable. OpenAI is not currently selling compute to third parties, but Altman did not rule out that business in the future.

Altman did not name public companies in the comments. That leaves the market with a narrower, practical question than a verdict on any one provider: which proposed facilities have durable customers and revenue, and which rely on demand, pricing and efficiency remaining favorable long enough to justify their cost.