Keynote
summary
Meta reports concrete scale: 1.2 million H100s (~zetaflop class), a 129k‑GPU cluster and repeated 24k training clusters, plus rapid‑deployment Prometheus gigawatt tents already in production. New silicon work ships fast — MTIA: four chips in two years aimed at training and inference — and a MetaRocky transport stack and optics choices support heterogeneous, long‑distance fabrics. Agentic AI runs planning, delivery and fault‑finding across millions of NICs and optics.
AI-written summary · about
chapters (5)
- 0:00Scaling AI infrastructure
- 2:47Evolving AI workloads
- 8:59Network architecture design
- 16:00AI in network lifecycle
- 19:12Future and mission
