Meta Reveals Prometheus AI Supercluster

Meta gave an inside look at Prometheus, a gigawatt-scale AI supercluster built in New Albany, Ohio, in a company technology blog post on the 19th, local time. Meta said the facility supports its recently launched Muse model series. It currently runs hundreds of thousands of GPUs, with plans to scale to about 1 million.

Prometheus links GPUs across multiple buildings so they operate like one computer. Two years ago, Meta began converting a data center that handled general web services into a large-scale AI training site, and it has completed the first phase. Muse Wonder, which Meta recently launched, was also trained there. On the 12th, Meta announced Muse Wonder and Muse Wonder Full, saying they achieved state-of-the-art performance on key evaluations for personal and workplace agents, memory, and coding.

The facility uses Nvidia GB200 and GB300 systems; each GB300 rack holds 72 GPUs. Connecting systems inside and between buildings requires tens of thousands of network switches and millions of optical communications components. Meta said keeping so many GPUs running together without stopping was a challenge. Working with Nvidia, it reduced training interruptions by 99% over about 1 year, and added backup GPUs for failures and liquid cooling.

Meta plans to expand Prometheus with Nvidia’s next-generation Vera Rubin systems. It said it will apply lessons from Prometheus to another supercluster, Hyperion, under construction in Louisiana.

Meta gave an inside look at Prometheus, a gigawatt-scale AI supercluster built in New Albany, Ohio, in a company technology blog post on the 19th, local time. Meta said the facility supports its recently launched Muse model series. It currently runs hundreds of thousands of GPUs, with plans to scale to about 1 million.

Prometheus links GPUs across multiple buildings so they operate like one computer. Two years ago, Meta began converting a data center that handled general web services into a large-scale AI training site, and it has completed the first phase. Muse Wonder, which Meta recently launched, was also trained there. On the 12th, Meta announced Muse Wonder and Muse Wonder Full, saying they achieved state-of-the-art performance on key evaluations for personal and workplace agents, memory, and coding.

The facility uses Nvidia GB200 and GB300 systems; each GB300 rack holds 72 GPUs. Connecting systems inside and between buildings requires tens of thousands of network switches and millions of optical communications components. Meta said keeping so many GPUs running together without stopping was a challenge. Working with Nvidia, it reduced training interruptions by 99% over about 1 year, and added backup GPUs for failures and liquid cooling.

Meta plans to expand Prometheus with Nvidia’s next-generation Vera Rubin systems. It said it will apply lessons from Prometheus to another supercluster, Hyperion, under construction in Louisiana.