Cerebras chief executive Andrew Feldman has given his account of why OpenAI walked away, and the account is notable for what it blames. In remarks reported by BigGo Finance, Feldman said the defection stemmed from a shortage of data center capacity, not from any shortfall in Cerebras' chips.

That distinction matters for a company whose entire pitch rests on silicon. Cerebras builds wafer-scale processors, an architecture that puts an entire wafer to work as a single large device rather than dicing it into many small ones. If the product performed adequately and the deal still collapsed, the bottleneck sat elsewhere in the stack.

Data center capacity has become the binding constraint across the AI hardware business. Graphics processors remain scarce, yet power, cooling and physical floor space have proven harder to scale than transistor counts. Feldman's framing places OpenAI's decision in that second category.

The claim carries obvious self-interest. A vendor that loses a marquee customer prefers an explanation that spares the product. Still, the explanation is plausible on its own terms, because the industry's most conspicuous shortages this cycle have appeared in megawatts rather than in wafer starts.

What does the explanation change for Cerebras? If the chips were not the obstacle, the company can argue its technology remains competitive for customers that control their own facilities. Enterprises building private AI infrastructure, or hyperscalers with committed power budgets, face a different calculus than a lab renting capacity on the open market.

The episode also sketches the current hierarchy of AI procurement. A model developer of OpenAI's scale buys compute the way a manufacturer buys electricity: in bulk, on schedule, with little patience for missing capacity. Suppliers that cannot guarantee the surrounding infrastructure, regardless of how novel their silicon, will lose the order to those that can.

For Cerebras, the loss removes its most visible reference customer at a moment when the company has been working toward a public listing. Wafer-scale computing remains a genuine engineering achievement, and the CEO's version of events keeps that achievement intact. It does not, however, restore the order.

The commercial question now is whether other buyers with firmer control of data center capacity will step into the space OpenAI vacated. Feldman has offered the diagnosis; the revenue line will supply the verdict.

By Nathan Brooks