Meta Details Prometheus, Its Gigawatt-Scale AI Supercluster in Ohio
An engineering post describes hundreds of thousands of Nvidia GPUs in New Albany, a target of roughly 1 million, and the reliability, network and power work behind training the Muse models.
Photo: LPS.1 / CC0 via Wikimedia Commons
Meta on October 19 published an engineering post describing Prometheus, its gigawatt-scale AI supercluster in New Albany, Ohio, saying the first phase is complete and has been used to train its Muse foundation models, including Muse Wonder.
Scale and reliability
Meta says the site today holds hundreds of thousands of GPUs spread across multiple interconnected buildings, including traditional data centers, AI-optimized buildings and rapid-deployment structures. At full scale it targets roughly 1 million GPUs.
The company reported these reliability figures:
- Goodput above 85% for pre-training and above 90% for reinforcement learning runs on Muse Wonder.
- Mean time between interruptions of about 47,000 GPU hours, later exceeding 54,000 after firmware fixes.
- A 99% reduction in interruptions since mid-2025, when they peaked at nearly 1,000 per 8,000 GPUs per day.
- A target of two interruptions per 8,000 GPUs per day.
Hardware and networking
Meta lists three Nvidia rack platforms: GB200 NVL72 (called Catalina), GB300 NVL72 (Clemente) and the upcoming Vera Rubin NVL72 (San Miguel). Each rack holds 72 GPUs; Meta cites about 135 kW for a GB300 rack and over 170 kW for Vera Rubin.
GPU-to-GPU traffic runs on a dedicated Ethernet back-end network. Each building fabric connects up to 20,736 GPUs using 51.2T switches, and a backend aggregation layer built on Arista 7800-series chassis with Broadcom Jericho3 silicon links sites from 2 km to more than 100 km apart.
To tolerate faults, each rack keeps 8 of its 72 GPUs as spares and about 2% of racks are held in reserve.
Power and cooling
Synchronous training can swing cluster load by 30% to 40%, Meta says. It manages this with gradual power capping, rack battery backup units, capacitor trays and facility-level battery storage. Cooling combines air-assisted liquid cooling in free-air buildings with facility liquid cooling in the purpose-built AI building; the newest sites use closed-loop, direct-to-chip liquid cooling.
What comes next
Meta said lessons from Prometheus will carry into Hyperion, a planned supercluster in Richland Parish, Louisiana, built with utility partner Entergy.
