Meta has completed the first phase of Prometheus, its gigawatt-scale AI supercluster in New Albany, Ohio — the machine that trained its Muse family of models — and says the site will scale to roughly 1 million GPUs. Why it matters: Meta just published unusually detailed numbers on what it actually takes to keep a cluster that size training instead of crashing.
October 19, 2026 — Meta’s engineering team published a detailed technical account of Prometheus on October 19, covering the hardware, networking and reliability work behind the supercluster.
What did Meta actually announce?
Prometheus now links hundreds of thousands of GPUs across traditional data center buildings, AI-optimized buildings and rapid-deployment structures, all functioning as a single machine. Two years ago the New Albany campus was a conventional data center serving web traffic. The completed first phase trained the Muse family of foundation models, including the Muse Wonder frontier models. At full scale, Meta says Prometheus will comprise roughly 1 million distributed GPUs.
What hardware is Prometheus running?
The cluster runs two NVIDIA platforms today, both built on the NVL72 rack design in which 72 GPUs act as a single accelerator. The first, which Meta calls Catalina, uses GB200 NVL72 systems; the second, Clemente, uses GB300 NVL72 systems with 20.7 TB of HBM3e memory per rack and a stated 5–10% performance improvement over the GB200 configuration. A third platform based on NVIDIA’s Vera Rubin NVL72 is arriving soon. A dedicated Ethernet network links GPUs within and across buildings, spanning distances from about 2 km to more than 100 km.
How reliable is a million-GPU machine?
That is the part Meta was most candid about. The company reports more than 85% goodput — productive training time — on Muse Wonder pre-training runs and more than 90% on reinforcement learning runs. Interruptions peaked at nearly 1,000 events per 8,000 GPUs per day in mid-2025; Meta says a year of firmware fixes, hardware screening, spare GPUs and automated fault recovery has cut daily interruptions by 99%. Mean time between interruptions was about 47,000 GPU hours during training and now exceeds 54,000, against a target of roughly 100,000. Synchronized training also swings power demand by 30–40% in seconds, which Meta manages with a system it calls Dynamo Dimmer, backed by rack and facility batteries. Its newest halls use closed-loop liquid cooling.
Why it matters
Frontier AI is now limited as much by power, networking and failure rates as by chips. Every lab is building at this scale, but few publish their interruption rates, spare-GPU overhead or power swings. Prometheus is the infrastructure behind Meta’s Muse models, and the lessons from New Albany are feeding Hyperion, Meta’s next supercluster in Richland Parish, Louisiana. For anyone buying AI capacity rather than building it, these are the operational numbers that decide how fast the next generation of models arrives — everything AI, tested starts with whether the machine stays up.
Frequently asked questions
Where is Meta’s Prometheus supercluster?
At Meta’s data center campus in New Albany, Ohio. The site previously served conventional web traffic before being reworked for frontier-model training.
How many GPUs will Prometheus have?
Hundreds of thousands today, scaling to roughly 1 million GPUs at full build-out, according to Meta.
What models did Prometheus train?
Meta says the completed first phase trained its Muse family of foundation models, including the Muse Wonder frontier models.
What comes after Prometheus?
Hyperion, a supercluster in Richland Parish, Louisiana, designed from the ground up using lessons from Prometheus.
Sources: Meta Engineering, Unite.AI, StorageReview