OpenAI adds 750 MW of AI power via US firm Cerebras

OpenAI has announced a partnership with Cerebras to add 750 megawatts of AI computing power to its platform.

Cerebras, a US-based company, designs custom AI systems using large chips to improve speed and reduce latency.

OpenAI will integrate Cerebras’ technology into its systems gradually, aiming to boost response times for AI tasks such as code generation, image creation, and real-time interactions.

The new computing capacity is expected to be introduced in phases through 2028.

OpenAI described the move as part of its broader strategy to diversify its infrastructure and support real-time AI for more users.

Cerebras’ CEO, Andrew Feldman, said the collaboration would enable new ways to build and interact with AI models.

🔗 Source: OpenAI

🧠 Food for thought

Implications, context, and why it matters.

The 750 megawatt (MW) claim lacks feasibility details critical to 2028 delivery

OpenAI plans 750 MW via Cerebras. The announcement omits site locations and power sourcing via power purchase agreements (PPAs, long-term contracts to buy electricity). It also skips build timelines and whether deals are firm or goals.
Oklahoma City plans 300+ CS-3 systems (its third-generation wafer-scale AI computers) in June 2025, with Montreal in July 2025, covering only a slice of the 20x 2025 capacity plan 1.
Phasing through 2028 signals limits. Power, cooling, and chip supply constrain how fast new compute can land.
Cerebras leans on one buyer. G42 (an AI and cloud company) made up 87% of 1H 2024 sales, which raises capacity questions for a large OpenAI deal 2.

Hardware diversification creates demand for cross-platform AI tooling

OpenAI adding Cerebras gear makes AI stacks more mixed. That opens space for portability and orchestration tooling (software that schedules and manages where workloads run).
Cerebras builds on OpenXLA (an open-source compiler stack that lets AI frameworks run across different hardware). Developers using OpenXLA-ready tools can pitch infrastructure-agnostic options as enterprises mix GPU and Cerebras setups 3.
Inference startups can ride Cerebras’ 3,000 tokens per second (a measure of text generation speed) with tools that shift workloads between GPUs and Cerebras with little code churn 4 5.
Through 2028, vendors can ship tools for mixed fleets with GPUs and wafer-scale engines (extremely large single chips built from a full silicon wafer). Targets include monitoring, cost control, plus workload schedulers.