Signal
OpenAI has signed a definitive multi-billion dollar agreement with Cerebras Systems valued at over $10 billion to secure 750 megawatts (MW) of dedicated computing power through 2028.
Backdrop
The deal is the largest single infrastructure commitment in AI history, targeting the severe “compute famine” facing OpenAI’s 900 million weekly active users.
- Performance Leap: Cerebras’s Wafer-Scale Engine (WSE) delivers up to 21x faster inference than Nvidia’s current clusters in complex reasoning benchmarks.
- The Power Factor: The 750MW capacity—comparable to the energy grid of a major metropolitan area—will be deployed to bypass public grid congestion and enable real-time agentic workflows.
Why it matters
This is a strategic decoupling from the Nvidia ecosystem.
- Ending the “Nvidia Tax”: By committing $10B to a non-Nvidia vendor, Sam Altman is proving that GPU-free high-performance inference is commercially superior, effectively capping Nvidia’s pricing power.
- Winning the “Inference War”: As models commoditize, speed-to-token becomes the primary differentiator. Cerebras’s latency advantage allows OpenAI to launch autonomous agents with human-like responsiveness that competitors cannot replicate.
- Financial Kingmaking: This deal secures Cerebras’s position as the clear “Second Pole” in the AI hardware sector, providing the revenue surge needed for its high-multiple 2026 IPO at a $22 billion valuation.
What to watch
- If OpenAI migrates its most compute-intensive products—Sora (video) and o-series (reasoning)—entirely to Cerebras-powered “Inference Factories.”
- Nvidia’s pricing and software response to prevent a broader “Hyperscale Defection” by other major AI labs.
