OpenAI Is Spending $20 Billion on Chips That Are Not Made by Nvidia

OpenAI has agreed to pay Cerebras more than $20 billion over the next three years for computing capacity powered by the chipmaker's wafer-scale processors, more than doubling a deal the two companies struck in January.

The agreement, first reported by The Information, includes warrants that could give OpenAI a stake of up to 10% in Cerebras as its spending grows. OpenAI is also committing $1 billion toward building the data centers that will run Cerebras hardware. Total spending over the full contract period could reach $30 billion.

The deal matters because Cerebras is not Nvidia. Its flagship chip is a wafer-scale processor, literally the size of a dinner plate, designed to deliver AI inference at far higher speeds than conventional GPU clusters. OpenAI has been vocal about wanting to reduce its dependence on Nvidia, and this contract, alongside its own internal chip development and Stargate infrastructure investments, is part of that strategy.

What it means for Cerebras

For Cerebras, the OpenAI relationship is central to its public market ambitions. The company filed for a new IPO last week, targeting a $35 billion valuation and a listing in the second quarter, up from its $23.1 billion private valuation. CNBC reports that having a customer committed to $20 billion in spending over three years is an unusually powerful anchor story for an IPO roadshow.

None of this means the Nvidia relationship ends. OpenAI still runs enormous compute on H100 and H200 clusters. But the scale of the Cerebras commitment, and the equity component, suggests the two companies are thinking about this as something more than a vendor arrangement. Whether Cerebras hardware performs reliably at this kind of production scale will be tested very publicly over the next 36 months.