Closing Keynote
summary
Claims that the industry spent ~$450B on AI in 2025 and ~$765B in 2026 back a year of concrete engineering: from prototypes of compute tents serving hundreds of megawatts to Prometheus, a multi‑gigawatt training cluster, and a GB200 “Catalina” rack with 72 GPUs plus cooling. Fleet totals hit 1.3M H100‑equivalent GPUs early 2026. Software+firmware+ops work reduces job interruptions by ~95% reduction, but scaling math still bites (1k GPUs → failure ≈8h, 16k → ≈2.5h; 100k+ → failures every ~8min, risking ~30% idle capacity).
AI-written summary · about
chapters (14)
- 0:00Apollo mission investment comparison
- 4:13Theory versus data center innovation
- 6:18World's largest hybrid cloud
- 7:42Prometheus cluster abstraction
- 8:46GB200 rack reliability
- 10:29Failure rates and costs
- 12:54Reducing failures with control
- 14:17AI-driven software revolution
- 15:40Developer productivity and incidents
- 17:02AI products like Ray-Ban Meta
- 18:04AI buildout versus space race
- 20:10Trillion-dollar AI investment
- 21:11Fault tolerance challenges
- 22:33Three decades of tech reflection
related talks
27mLive Panel: Building at the Roofline — Planning and Delivering Gigawatt-Scale AI Infrastructure
23mKeynote
52mKeynote Address: Insights Gained from Delivering Two Generations of AI Supercomputers...
20mOperating Frontier Intelligence at Scale
29mLive Keynote
29mUnlocking the Future of AI: Trends, Infrastructure, and Tokenomics
15mKeynote: From Platforms to AI Factories. Has Kubernetes Solved It?
05mSponsored Keynote: Towards Clouds of AI Clusters