OpenAI secures up to 750MW of low-latency AI compute from Cerebras through 2028

OpenAI has agreed a multi‑year, multi‑billion dollar purchase with Cerebras Systems to acquire up to 750 megawatts of low‑latency AI inference compute capacity delivered in tranches through 2026–2028. The deal deploys Cerebras’ wafer‑scale, inference‑focused processors — high‑core, large on‑chip memory chips with very high fabric bandwidth — to reduce latency and accelerate real‑time tasks such as code generation, image creation and AI agents. OpenAI says the capacity will be phased into its inference stack beginning this year and complements broader infrastructure moves including a $500m co‑investment with SoftBank’s SB Energy to build multi‑gigawatt AI data campuses (including a 1.2GW site in Milam County). The agreement aims to diversify OpenAI’s supplier mix away from general‑purpose GPUs and hyperscaler dependence, target inference bottlenecks, and lower per‑inference operating costs over time. Analysts cited in coverage expect significant efficiency gains for inference versus general‑purpose GPUs, and view the deal as validating inference‑specific hardware as a new battleground that could pressure GPU suppliers. Community concerns about data‑centre impacts (water use, noise) are noted; OpenAI and partners plan mitigation steps such as lower‑water cooling, workforce investment and grid upgrades. Primary keywords: OpenAI, Cerebras, AI compute, inference hardware, low‑latency. Secondary/semantic keywords included: wafer‑scale chips, inference efficiency, data centre build, SoftBank, SB Energy, Milam County. (Main keyword "AI compute" appears multiple times for SEO.)