Meta details Prometheus, its Ohio AI supercluster heading for about 1 million GPUs
Meta’s engineers say the New Albany site trained its Muse models, now runs hundreds of thousands of GPUs, and cut training interruptions by 99% from a 2025 peak.
Illustration · not a photoHow this story was checked
- ResearchedWith AI assistance · 2 sources
- AI source checkEach sentence matched to its source
- Editor approvalMayank Sahu · 20 Oct, 13:37
- Published20 Oct, 13:37
Verify this story
- 2 sources, 1 official · 2 with an archived copy · see sources
- Every fact has a label and a source number · fact ledger
- What we searched, read and could not verify · research note
- Machine-readable: facts as JSON · text as Markdown
How this was made: researched, drafted and source-checked by AI (Claude, made by Anthropic), then published under the editor’s standing approval. A person did not read it line by line before it went live. AI policy
In short
Meta published a detailed look at Prometheus, its gigawatt-scale AI supercluster in New Albany, Ohio, on 19 October 2026. Meta says it runs hundreds of thousands of GPUs today, is being scaled toward roughly 1 million, and trained its Muse family of models.
What Meta shared
Four Meta engineers described Prometheus in a post on Engineering at Meta.1 About two years ago, the New Albany site ran ordinary web traffic; it is now a “gigawatt-scale supercluster”.1 Meta does not give an exact power figure.1
It has “hundreds of thousands of GPUs” and is being scaled toward “roughly 1 million”.12 Its first phase trained the Muse family of models, including Muse Wonder.12
The hardware
- Catalina: Nvidia GB200 NVL72, about 120kW per two-rack pair.12
- Clemente: Nvidia GB300 NVL72, 288GB of HBM3e per GPU, about 135kW per rack.12
- San Miguel: Nvidia Vera Rubin NVL72, coming soon, over 170kW per rack.12
Each building’s network connects up to 20,736 GPUs over Ethernet, with 800G of bandwidth per GPU, and a regional network links buildings from 2km to over 100km apart.1 Cooling mixes air-assisted liquid cooling, facility liquid cooling and closed-loop systems.1
Reliability
Meta reports more than 85% goodput on Muse Wonder pre-training and more than 90% on reinforcement learning runs.12 Interruptions fell 99% from a mid-2025 peak of nearly 1,000 per 8,000 GPUs per day, and mean time between interruptions rose past 54,000 GPU hours, with a target of about 100,000.12 These are Meta’s own measurements.
What comes next
Lessons from Prometheus feed into Hyperion, Meta’s planned supercluster in Richland Parish, Louisiana, including closed-loop liquid cooling and a grid connection with Entergy.1
Why it matters
Few companies publish this level of detail. The reliability figures show how much engineering goes into keeping very large training runs going, beyond buying chips.
Fact ledger
How labels workQuestions readers ask
Where is Meta’s Prometheus supercluster?
In New Albany, Ohio.
How many GPUs does Prometheus have?
Meta says hundreds of thousands today, scaling toward roughly 1 million.
What is Hyperion?
Meta’s next planned supercluster, in Richland Parish, Louisiana, designed using lessons from Prometheus.