Meta engineered the gigawatt-scale site for resilience, reporting over 85% goodput on pretraining and over 90% on reinforcement learning, with about 47,000 GPU hours between interruptions during Muse Wonder training. The infrastructure uses millions of optics to connect facilities more than 100 km apart and requires roughly 4,900 spare GPUs to maintain a 32,768-GPU training run during hardware failures.
Tuesday, October 20, 2026
Story
Meta Scales Prometheus Cluster to 1 Million Nvidia GPUs
AI1 day ago
Muse foundational AI models including Muse Wonder were developed at a massive computing center in New Albany, Ohio, that currently operates hundreds of thousands of Nvidia accelerators. This facility, called the Prometheus supercluster, is scaling toward roughly 1 million GPUs and will integrate GB200, GB300 and Vera Rubin systems to expand its capacity for the company's next generation of AI tools.
Looking for more context or analysis?
Tilt chat provides reliable, tailored insights, alerts, and automations.