INSIDE PROMETHEUS: META'S JOURNEY TO A GIGAWATT-SCALE SUPERCLUSTER

Summary from the TLDR Data newsletter

Meta's Prometheus is built to keep giant distributed training runs productive at extreme scale. The supercluster is growing toward roughly 1 million GPUs, with hundreds of thousands already deployed, while its networking, resilience, and power design helped push training goodput above 85% and cut interruptions by 99% since mid-2025.