Cerebras Secures OpenAI 750 MW for Inference

The key shift is that Cerebras is no longer just selling exotic training hardware, it is locking in long dated demand for real time model serving at industrial scale. Six inference datacenters, thousands of CS-3 systems, and a contracted 750 MW OpenAI deployment through 2028 mean Cerebras is building a dedicated fast lane for workloads like coding agents and enterprise search, where a few hundred milliseconds or a 10x faster decode rate directly changes product behavior and spend.1 2 3 4 5

• This matters because the hardest part of inference is not raw FLOPS, it is delivering tokens fast enough for interactive use. OpenAI is already using Cerebras for GPT-5.3-Codex-Spark, and Cerebras says its cloud can deliver 1,000 plus tokens per second for coding and up to 3,000 tokens per second on very large models, which makes multi step agent loops feel instant instead of sluggish.4 5 6
• Cerebras is also moving up the stack from one time box sales to recurring cloud revenue. Its cloud and other services revenue reached $82.8M in Q1 2026, up 178% year over year, and internal research estimates cloud revenue doubled to $152M in 2025, lifting inference to about 30% of revenue from effectively zero in 2023.7 8
• The closest analog is Groq, another specialist built around low latency inference, but Cerebras now has a very different scale profile. Groq was expanding toward 200 MW of capacity by 2027, while OpenAI alone committed 750 MW to Cerebras, giving it unusual demand visibility and making it easier to win enterprise buyers that want proof the platform will still be there years from now.5 9

The next phase is a race to become the default serving layer for the largest models that cannot tolerate lag. If Cerebras executes the 2026 to 2028 rollout and keeps winning latency first workloads, inference can become its center of gravity, with hardware acting less like the product and more like the engine behind a high margin compute utility.1 3 7 8