Cerebras CTO: our plate-sized chip drew more power than an entire 5 kW rack
AccBalanced · x · 2026-11-04
Cerebras co-founder and CTO Sean Lie recounts the company's early infrastructure challenges in a new interview:
- When Cerebras started, data centers were built for 5 kW per rack — a single wafer-scale chip needed more than that.
- Today NVIDIA and AMD AI racks draw 100–200 kW, and the power and liquid cooling infrastructure Cerebras pioneered has become industry standard.
- Topics covered include why a dinner-plate-sized chip needs special data centers, how liquid cooling enables denser racks, the path from delivered racks to production tokens, how they keep serving through constant hardware failures, and what changes when running bigger models.
More from Infra
- Meta raises $15B in bonds for AI buildout, drawing $58B in orders at 3.9x oversubscription — rohanpaul_ai · 2026-11-04
- BIOS ports RFD4-Proteina to Mac: local protein binder design 2.2x faster with MLX — AllThingsApx · 2026-11-04
- fal Co-founders on Generative Media Rise and AI Infra Economics — gorkem · 2026-11-04
- Rust inference engine beats SGLang 13.4x on 250K-context agent TTFT, open-sourcing soon — ahstanin · 2026-11-04
- Data centers turn to discarded jet engines as grid waits stretch for years — TechCrunch AI · 2026-11-04
- Nvidia lost credibility in Washington by exaggerating China's AI chip capabilities, CFR fellow says — mchorowitz · 2026-11-04