Meta and NVIDIA on Prometheus: gigawatt training cluster cut failures by 99%
BenBajarin · x · 2026-10-24
Ben Bajarin interviews NVIDIA's Ian Buck and Meta infrastructure head Santosh Janardhan on Prometheus, Meta's gigawatt-class training cluster in New Albany, Ohio. Key points: Meta scaled across buildings over 100km apart with software presenting them as one cluster; at its mid-2025 peak something broke every few seconds, and after months of joint debugging Meta cut interruptions by 99%; power constrains the energized fleet to roughly 72% of demand, making best-in-class compute a multi-dimensional moat. Full report is paid on DiligenceStack.
More from Companies & People
- Nomic CEO: as AI speeds up engineering reviews, senior engineers' bandwidth is the new bottleneck — andriy_mulyar · 2026-10-24
- Artefact AI Film Festival opens entries: every production phase must use a generative AI tool — lmoroney · 2026-10-24
- Jason Naz Joins Runway to Lead Developer Experience — tlakomy · 2026-10-24
- Too Many AI Startups Sell Picks and Shovels, Not Enough Mine the Gold — andreisavu · 2026-10-24
- Abundant Labs Partners with Baseten to Build Custom Environments for Customers — amankhan · 2026-10-24
- Amazon CEO Andy Jassy: AI will get you to an 'advanced starting spot' faster, but human judgment is your edge — a16z · 2026-10-24