Meta on October 19, 2026 published a detailed technical account of Prometheus, its gigawatt-scale AI supercluster in New Albany, Ohio, confirming the first phase is complete and was used to train the company's Muse family of foundational models1,2,4. The facility is scaling toward roughly 1 million GPUs and 1 GW of total capacity3.
Two years ago, the New Albany campus was a conventional data center serving web traffic; Meta reworked it for frontier-model training and is building Prometheus in phases. The cluster now comprises hundreds of thousands of GPUs spread across traditional and AI-optimized data center buildings and rapid-deployment structures, all operating as a single machine.
Reliability at scale
Meta disclosed granular reliability data that is rare at this scale. The company reported more than 85% goodput on Muse Wonder pre-training runs and more than 90% on reinforcement learning runs, with a wall-clock mean time between interruptions of approximately 47,000 GPU hours during those runs. Firmware fixes rolled out since training have pushed MTBI past 54,000 GPU hours. Meta's stated target is two interruptions per 8,000 GPUs per day, which works out to roughly 100,000 GPU hours between interruptions.
The Muse family includes the Muse Wonder frontier models, which Meta says deliver state-of-the-art performance on personal and professional agentic tasks. Meta engineers Kevin Lee, Yashar Bayani, Ahmed Sharif, and Yee Jiun Song authored the engineering blog post detailing the project.
Hardware and what comes next
Prometheus deploys two NVIDIA GPU platforms currently in production, and Meta says a third based on NVIDIA's Vera Rubin NVL72 is arriving soon. Two key infrastructure components carry most of the load: a Backend Aggregation network, or BAG, that links GPUs across multiple buildings using high-capacity Ethernet connections, and a closed-loop liquid cooling system.
Meta walked through the engineering behind Prometheus at the OCP Global Summit on October 12, 2026, ahead of the October 19 blog post.
ANALYSIS Publishing detailed goodput and MTBI figures sets a public benchmark for large-cluster reliability that other hyperscalers have generally kept private, giving external researchers a concrete reference point for what sustained training at the hundreds-of-thousands-of-GPUs scale looks like in practice.