Cerebras inks a 750 megawatt OpenAI deal and hands over 10 percent of the company

SemiAnalysis published a detailed look at Cerebras and its OpenAI deal that finally puts the wafer-scale chip company in the same conversation as the GPU giants. The hardware side is familiar. WSE-3 packs 44 gigabytes of on-wafer SRAM, 21 petabytes per second of internal bandwidth, and 15.6 dense FP16 petaflops on TSMC's N5 process. Only 150 gigabytes per second leaves the wafer, which is why large models have to be split across multiple systems through pipeline parallelism. The pitch is raw speed: GPT-5.3-Codex-Spark runs at about 2,000 tokens per second per user on Cerebras, against roughly 40 tokens per second on standard Opus 4.6.

The deal terms are the more interesting part. OpenAI committed to 750 megawatts deployed between 2026 and 2028, with an option for another 1.25 gigawatts, and the contract carries $24.6 billion in remaining performance obligations. Cerebras also extended a $1 billion working capital loan at 6 percent that gets waived if OpenAI takes the capacity, and granted 33.4 million warrants priced at essentially zero against a $82.02 reference. That works out to roughly $2.74 billion in contra-revenue, about 10 percent of the company. SemiAnalysis pegs a CS-3 rack at around $450,000 after the recent memory price spike.

Why it matters

If you build inference-heavy products, this is the first OpenAI compute deal where the non-NVIDIA alternative is being priced at scale, with terms you can compare against. Track p99 latency rather than peak tokens per second; pipeline parallelism on Cerebras pays in throughput but constrains how you shape your serving workloads.