Meta's first gigawatt-scale supercluster Prometheus phase one done, targeting ~1M GPUs
BenBajarin · x · 2026-10-20
Meta's engineering blog details Prometheus, its gigawatt-scale supercluster in New Albany, Ohio. Phase one is complete and was used to train the Muse foundation models, including Muse Wonder frontier models; the site is being scaled toward 1 million GPUs.
Key technical details:
- Closed-loop liquid cooling for efficiency
- Racks use the Ariel variant of GB200 NVL72: each Blackwell GPU pairs with a Grace CPU and 2x ConnectX-7 400G NICs at a 1:1 CPU-to-GPU ratio, delivering 800G of I/O per GPU earlier than the industry to fix NVL72's scale-out bandwidth imbalance
- Muse Wonder pre-training achieved >85% goodput and RL runs >90%
Related event: Meta Unveils Gigawatt Prometheus Supercluster, Targets One Million GPUs(4 posts)→
More from Infra
- vLLM Team Details Agent-Assisted GPU Kernel Work: Kimi K3 Hits 1.73x Stock vLLM — lmoroney · 2026-10-20
- Cloudflare unveils Clairic, a fast System One extraction model that undercuts LLMs on cost and accuracy — irvinebroque · 2026-10-20
- Gartner: $3.5T of AI infra spend may face write-downs by 2030, 80% of neoclouds to fold — rohanpaul_ai · 2026-10-20
- Stripe paid $7.5B for a router, but routing is easy — verifying outputs is the real moat — AIGODSEND · 2026-10-20
- Google Cloud upgrades VPC Service Controls with folder-level perimeters for agent traffic, free — rseroter · 2026-10-20
- Amateur researcher fights Stage 4 colon cancer with ML, hits $500-per-run GPU wall — whafa · 2026-10-20