Meta Details Prometheus, the Ohio Supercluster Heading for 1 Million GPUs
Meta scales its New Albany facility to one million GPUs for AI training.
1Prometheus Scale and Architecture
Meta is expanding its New Albany, Ohio, data center into the massive Prometheus supercluster, targeting a capacity of one million Nvidia GPUs. The facility currently utilizes hundreds of thousands of units across multiple buildings. This infrastructure supports the training of the Muse family of models, including the recent Muse Wonder iterations.
2Hardware Platform Composition
The cluster integrates three distinct Nvidia platforms: Catalina, Clemente, and the upcoming San Miguel. These systems utilize GB200 and GB300 NVL72 racks, with San Miguel incorporating Vera Rubin NVL72 technology. Each rack functions as a single accelerator, with individual buildings connecting up to 20,736 GPUs via extensive fiber optic networking.
3Network Aggregation Strategy
Meta connects these buildings using a sophisticated aggregation network spanning distances from two kilometers to over 100 kilometers. The architecture combines traditional data centers with AI-optimized buildings and rapid-deployment structures. To maintain necessary capacity, the company supplements its internal infrastructure by leasing additional resources from various external industry partners.
4Engineering Reliability Improvements
Managing a million GPUs presents extreme failure risks. In mid-2025, interruptions reached 1,000 events per 8,000 GPUs daily. Through intensive engineering, Meta reduced these failures by 99%. The team addressed signal quality issues across 600,000 NVLink connections and resolved memory faults that emerged while scaling the GB300 rack configurations during training.
5Training Efficiency Metrics
The cluster currently maintains over 85% productivity for pre-training and 90% for reinforcement learning tasks. Mean time between interruptions has improved from 47,000 GPU hours to over 54,000 following firmware updates. Meta aims to reach 100,000 GPU hours between interruptions to maximize the utility of its massive, distributed computing environment.
6Redundancy and Hardware Reserves
Reliability requires significant hardware overhead. Each 72-GPU rack operates with 64 active units and eight spares, while Meta maintains 2% of total racks in reserve. Consequently, a training run requiring 32,768 GPUs necessitates 37,656 units, representing a 13% overhead to ensure continuous operation despite potential hardware failures during large-scale model training.
7Dynamic Power Management
Large-scale GPU clusters experience power draw swings of 30% to 40% during operation. Meta employs software called Dynamo Dimmer to cap power consumption gradually, supported by battery systems at the rack, facility, and grid levels. This prevents massive load shifts when starting or pausing jobs that consume tens of megawatts.
8Sustainability and Future Outlook
Meta matches its New Albany energy consumption with clean and renewable sources, though it exited the RE100 pact in July due to gas infrastructure needs. Lessons from Prometheus are currently informing the design of the Hyperion supercluster in Louisiana. Additionally, Meta has pledged $1 billion to support local data center communities.