Meta has finished the first phase of Prometheus, its gigawatt-scale AI supercluster in New Albany, Ohio, and used it to train its new Muse Wonder frontier models, the company said in an engineering post on Monday, Oct. 19, 2026. Meta is now scaling the site toward roughly 1 million GPUs working as a single machine.

Today Prometheus spans "hundreds of thousands of GPUs across multiple interconnected buildings," including older data centers, an AI-optimized building and rapid-deployment structures, Meta engineers Kevin Lee, Yashar Bayani, Ahmed Sharif and Yee Jiun Song wrote. Two years ago the campus was a conventional data center running web traffic.

The hardware inside

  • Catalina: Nvidia GB200 NVL72 in a two-rack "Ariel" layout, 72 Blackwell GPUs with 192 GB of HBM3e each and about 120 kW per rack pair, according to Meta.
  • Clemente: Nvidia GB300 NVL72 in a single rack, 288 GB of memory per GPU, more than an exaFLOP of peak FP4 compute and about 135 kW per rack. Meta puts the gain over GB200 at 5% to 10%.
  • Next: a third platform, San Miguel, built on Nvidia's Vera Rubin NVL72, is "arriving soon," StorageReview reported from the post.

Keeping a giant training job alive

Meta reported more than 85% goodput on Muse Wonder pre-training runs and more than 90% on reinforcement learning runs, with a mean time between interruptions of about 47,000 GPU hours. Firmware fixes have since pushed that past 54,000 GPU hours. In mid-2025, interruptions peaked at nearly 1,000 events per 8,000 GPUs per day; a year of work with Nvidia cut that by 99%, StorageReview reported. Each 72-GPU rack runs 64 GPUs and keeps eight as in-rack spares.

Meta's next site, Hyperion in Richland Parish, Louisiana, is being designed from scratch using lessons from New Albany. Meta also says it has put more than $7.8 million into community grants and sponsorships in Franklin and Licking counties since breaking ground in 2017.

Why it matters for Tesla watchers

Training compute is now a race Tesla is in, too. In its 2026 proxy letter filed Monday, Tesla said it is investing in "additional data center capacity to train our SI systems" and in chip manufacturing through Terafab (see our proxy story). Meta's open account of goodput and failure rates gives a rare public yardstick for clusters like SpaceX's Colossus, which also sells capacity to Anthropic (our Anthropic-Nscale story). Earlier: Meta's Muse Spark Open model.