Market2026-07-22
OpenAI said it will add 750 megawatts of computing capacity from Cerebras, whose single very large chip design is built to return long model outputs quickly. The stated aim is faster responses for real-time uses such as coding, image generation and agent workloads. The capacity will be folded into OpenAI's serving systems in stages, with tranches arriving through 2028, so there is no immediate change for users.
What changed
OpenAI ran inference on a mix of conventional accelerator hardware without a dedicated low-latency tier.
What it unlocks
Faster model responses for real-time uses such as coding, image generation and agents, once the capacity is phased in.
- 750MW of low-latency AI compute
- capacity arriving in tranches through 2028
What you need to act on it
- capacity arrives in stages through 2028
Sources