Meta details Ohio Prometheus AI supercluster
Meta has published a detailed account of Prometheus, its gigawatt-scale AI supercluster in New Albany, Ohio. The company says the first phase trained its Muse foundation models and that it is scaling the system toward roughly 1 million GPUs.

Key takeaways · 4
- 01
Meta says Prometheus’s first phase is complete and trained its Muse foundation models, including Muse Wonder.
- 02
Meta is scaling Prometheus toward roughly 1 million GPUs.
- 03
The reported two-rack configuration draws about 120 kW and delivers 720 PFLOPS of peak FP8 training compute.
- 04
Meta says its shift to multi-tiered cached object storage is intended to reduce GPU idle time.
A cluster built in phases
Meta published a detailed account of Prometheus, which it describes as a gigawatt-scale AI supercluster in New Albany, Ohio.[1] The company says the first phase is complete and trained its Muse foundation models, including Muse Wonder.[1] Meta is scaling the system toward roughly 1 million GPUs.[1] The campus had been a conventional data center serving web traffic two years before the report, giving the account a before-and-after view of the site’s role.[1] For teams evaluating AI infrastructure, the announcement links a large-scale buildout to a stated model-training use.
Compute, racks and memory
Meta says Prometheus currently runs two NVIDIA GPU platforms, both based on the NVL72 rack architecture.[1] The reported configuration has two racks with 36 GPUs per rack, a 1:1 CPU-to-GPU ratio, and 800G of scale-out bandwidth for every GPU.[1] Meta reports 720 PFLOPS of peak FP8 training compute and about 120 kW of power draw across the two racks.[1] The account also lists more than one exaflop of peak FP4 compute.[1]
Storage and training efficiency
Meta says Prometheus has shifted from file-based storage to multi-tiered cached object storage to reduce GPU idle time.[2] The company reports more than 85% goodput during Muse Wonder pre-training and more than 90% in reinforcement-learning runs.[1] Those figures describe reported goodput for two specific training stages; the account does not establish that the same rates apply to other models or workloads.[1] Meta also says firmware fixes pushed mean time between interruptions past 54,000 GPU hours.[1] Together, these details give infrastructure teams several operational measures to examine alongside headline compute capacity.
Network, power and expansion
Meta says San Miguel will bring Vera Rubin NVL72 into the same network fabric as Prometheus.[1] The fabrics are tied together over distances ranging from 2 km to more than 100 km.[3] To manage power, Meta says Dynamo Dimmer gradually lowers GPU power limits as Prometheus approaches its electrical ceiling.[3] Meta also says it pays the full cost of energy at New Albany and matches its use with new clean and renewable power.[3]
For infrastructure and AI teams, Meta’s account offers concrete figures for cluster scale, compute, power, storage and reported training performance. Those details can help frame questions about capacity planning and operational trade-offs, but they should not be treated as benchmarks for different workloads without further evidence.
Why it matters
Test yourself on this story — 1 question.
Create a free account to take the quiz, earn XP, and get a daily session built for your industry.
Take the quizHow this developed
20 October 2026
Meta details Ohio Prometheus AI supercluster