Meta Converts Ohio Campus to Gigawatt Prometheus AI Supercluster
Hyperscaler converts traditional New Albany facility into gigawatt-scale AI supercluster to power next-gen models
Meta has unveiled the first live phase of its Prometheus AI supercluster in New Albany, Ohio, transforming a former traditional data center campus into a gigawatt-scale AI compute hub.
Key details
The converted New Albany campus currently hosts hundreds of thousands of GPUs across interconnected traditional and AI-optimized facilities, with plans to scale to roughly 1 million distributed GPUs at full buildout. Meta confirmed the facility has already been utilized to train its Muse foundation models.
The site deploys Nvidia GB200 and GB300 accelerators. The GB200 deployment features Meta's Catalina rack architecture using the Ariel two-rack configuration, which connects 36 GPUs per rack via NVLink into a 72-GPU domain with a 1:1 CPU-to-GPU ratio supporting 800G of I/O per GPU.
For GB300 chips, Meta uses its Clemente platform with Air-Assisted Liquid Cooling (AALC) or Facility Liquid Cooling (FLC). GB300 raises High Bandwidth Memory (HBM) capacity per GPU by 1.5x from 192GB to 288GB. At the building level, Ethernet fabrics connect up to 20,736 GPUs, while a Backend Aggregation (BAG) network spanning up to 100 km connects building-level fabrics using Arista 7800-series chassis with Broadcom Jericho3 silicon.
Why this matters
Converting existing data center campuses into high-density AI superclusters allows hyperscalers to scale compute capacity rapidly while reusing existing site infrastructure. However, scaling to 1 million GPUs requires massive power delivery, advanced liquid cooling architectures, and high-bandwidth optical networking that place immense demands on regional energy grids.
Context
As AI model parameter counts scale, tech giants are shifting from isolated data center buildings to regional superclusters spanning multiple facilities. Meta's Prometheus campus highlights how hyperscalers are retrofitting existing hyperscale real estate with air-assisted and facility liquid cooling to accommodate the extreme power densities of next-generation GPU hardware.
What happens next
Meta plans to begin receiving and integrating Nvidia's next-generation Vera Rubin platforms at the Prometheus site in upcoming phases as it continues scaling toward its target of 1 million GPUs.
Source: Data Center Dynamics Published on AI Usage Global, author: AUG Bot
