Meta published a detailed technical account of Prometheus, its gigawatt-scale AI supercluster located in New Albany, Ohio. The first phase is already operational and was used to train the Muse family of models, including Muse Wonder.
What happened?
In a post dated October 19, 2026, on the company’s engineering blog, Meta described Prometheus as an infrastructure advance connecting hundreds of thousands of GPUs across multiple interconnected buildings via fiber. The goal is to scale to approximately 1 million GPUs functioning as a single machine.
The first phase already trained the Muse Wonder models, which Meta describes as frontier models delivering state-of-the-art performance on personal and professional agentic tasks. The company reported more than 85% goodput on pre-training and more than 90% on reinforcement learning, plus a mean time between interruptions exceeding 54,000 GPU-hours after improvements.
Why it matters
Infrastructure scale remains a decisive factor in the race for larger and more capable models. Prometheus represents Meta’s effort to compete in frontier model training, with platforms based on NVIDIA GB200 (Catalina) and GB300 (Clemente), and plans for Vera Rubin NVL72.
The company highlights technologies such as closed-loop liquid cooling and dedicated Ethernet networks (Back-End) for GPU-to-GPU communication. The account shows an emphasis on reliability, with a 99% reduction in training interruptions compared to previous peaks.
Context
The New Albany site was originally a conventional data center and was adapted for frontier model training workloads. Meta is building in phases and continues scaling the cluster with future models from Meta Superintelligence Labs in mind. The technical details include NVL72 rack configurations, where 72 GPUs act as a single accelerator.
Sources
Transparency: This content was created, edited, or reviewed with the assistance of artificial intelligence. Information was cross-checked with public posts on X and available internet sources. Consult the original sources for the full context.
By GeekikiBot