Meta details Prometheus, the Ohio supercluster that trained its Muse models

Meta's engineering team published a detailed look on October 19, 2026 at Prometheus, its gigawatt-scale AI supercluster in New Albany, Ohio. Two years ago the site served ordinary web traffic. The first phase of the rebuilt cluster is complete and was used to train Meta's Muse models, including Muse Wonder, and the company is working toward roughly 1 million GPUs at full scale.

Today Prometheus has hundreds of thousands of GPUs spread across several connected buildings, including fast-to-build temporary structures, linked by miles of fiber so they act as one machine. It runs two NVIDIA platforms, GB200 NVL72 racks called Catalina and GB300 NVL72 racks called Clemente, with Vera Rubin NVL72 systems arriving soon. A separate Ethernet back-end network carries GPU-to-GPU traffic, and each building fabric links up to 20,736 GPUs.

Meta also shared reliability numbers. Training runs for Muse Wonder reached more than 85% goodput for pre-training and more than 90% for reinforcement learning, meaning that share of time went to useful work. The average time between interruptions was about 47,000 GPU hours during those runs, and Meta says firmware fixes have since pushed it above 54,000. The cluster uses closed-loop liquid cooling.

Why it matters

Frontier models are now limited by how reliably giant clusters stay up, and Meta's numbers give a rare public benchmark for what running a million-GPU-class system takes.

Source: Engineering at Meta