OpenAI’s Cerebras Deal and the Limits of a 100× Speed Claim

Matthew Berman reports that OpenAI and Cerebras have agreed to deploy Cerebras systems for faster AI inference. In the January 2026 announcement, the companies described a multi-year agreement for 750 megawatts of capacity, to be deployed in stages starting in 2026. The figure describes capacity in power terms, not a measurement of inference speed.

Berman cites reporting that the deal was worth more than $10 billion and describes 750 MW over three years. A later [Cerebras securities filing](https://www.sec.gov/Archives/edgar/data/2021728/000162828026025762/cerebras-sx1april2026.htm) says OpenAI committed to buy 750 MW of inference capacity and related services, with tranches expected in 2026–2028; each tranche has a three- or four-year term, extendable to five years. The filing describes the deal as worth more than $20 billion and includes an option for additional capacity. These details distinguish the rollout period from the term of each tranche.

Berman says he can imagine OpenAI acquiring Cerebras and notes that he recently interviewed Cerebras CEO Andrew Feldman. The acquisition idea is his speculation; the announced agreement is a computing partnership, not an announced acquisition.

The video shows a provider output-speed chart for OpenAI’s open-weight gpt-oss-120B model, with a 1,000-token input. Cerebras leads the visible entries at roughly 3,137 output tokens per second, versus about 863 for the next visible provider—around 3.6 times in that displayed comparison. The chart neither compares ChatGPT as a whole nor demonstrates the headline’s 100-fold claim; speeds depend on model, workload, provider and measurement conditions.

Berman attributes Cerebras’s speed in part to putting memory on its wafer-scale chip and contrasts it with memory supplies affecting other hardware. On-chip SRAM is an important feature, but it is not the whole memory system: [Cerebras specifies 44 GB of on-chip SRAM and external-memory options](https://www.cerebras.ai/press-release/cerebras-announces-third-generation-wafer-scale-engine) for WSE-3 systems. The discussion does not establish that Cerebras is unaffected by all memory shortages or that all workloads keep all data on the wafer.

In a brief clip from Berman’s earlier interview, Feldman says Cerebras does not use the particular memory type under discussion and sees that as a benefit. Berman expects faster inference ultimately to benefit users. The clip does not quantify a universal speedup or the effect of supply constraints on the company’s complete systems.