Meta says the first phase of Prometheus, its AI supercluster in New Albany, Ohio, is complete and was used to train its Muse models, including Muse Wonder. In an announcement published on October 19, 2026, the company also reported hundreds of thousands of GPUs at the site, with expansion continuing toward a goal of roughly one million. Meta’s engineering post describes a milestone in a much larger buildout.
The milestone covers one phase of a larger cluster
The reported completion applies to Prometheus’s first phase. Meta says the cluster spans interconnected buildings and rapid-deployment structures, and that it is still growing toward its roughly one-million-GPU target. The company describes the project as gigawatt-scale; that refers to the intended scale, not a current operating-power figure.
Meta says the first phase supported training for its Muse foundational models, including Muse Wonder. That makes the milestone concrete: the infrastructure has been used for model training while the wider cluster continues to expand.
Two deployed GPU platforms connect across buildings
Meta identifies Catalina, based on NVIDIA GB200 NVL72, and Clemente, based on NVIDIA GB300 NVL72, as deployed platforms. Each system has 72 GPUs. Meta reports 13.4 TB of HBM3e memory for Catalina and 20.7 TB for Clemente.
The network is designed to link far more than a single rack. Meta says each building-level Ethernet fabric can connect up to 20,736 GPUs. Its Backend Aggregation network, or BAG, links building fabrics over distances from 2 km to more than 100 km. For more background on Meta’s description of Prometheus’s network and construction, see NeoTeo’s earlier report.
Meta also describes two liquid-cooling approaches: air-assisted liquid cooling in free-air buildings and rapid-deployment structures, and facility liquid cooling in a purpose-built AI-optimized building. The company says a Vera Rubin-based platform called San Miguel is planned for Prometheus.
Meta’s reported training and reliability figures
Goodput is the share of training work that makes useful progress. Meta reports more than 85% goodput during Muse Wonder pre-training and more than 90% during reinforcement-learning runs. It also reports about 47,000 GPU-hours of wall-clock mean time between interruptions during those runs, later rising above 54,000 GPU-hours after changes to the fleet.
For a separate reliability measure, Meta says interruptions fell 99% from a peak in the initial phase in mid-2025 of nearly 1,000 events per 8,000 GPUs per day. These are company-reported figures, tied to the training runs and period Meta describes.
Facility listings describe different records
Baxtel lists a Prometheus 2 facility as operational, gives 2026 as its year built and lists a capacity of 100 MW. Interconnection.fyi lists Meta Prometheus as operational at more than 250 MW, with 2024 as an actual operational date and 2026 as a proposed date. Those are the entries each service publishes: Baxtel’s Prometheus 2 listing and Interconnection.fyi’s Meta Prometheus listing.
