PROMETHEUS / NEW ALBANY, OHIO
Prometheus campus
Three building types · illustrative layout
Drag to orbit
Sources & model notesMeta report · October 19, 2026
What this model shows
The three building types come from Meta’s article. The rapid-deployment structure follows Meta’s campus film; the other two forms and the site layout are illustrative. The model does not show the real number, size or position of the buildings.
No campus capacity is calculated from the buildings or rack slots drawn here.
- Meta campus film · 00:01–00:03
Curved rapid-deployment roof, ridge caps, louvered end wall and twin entry pods. Site arrangement and the other two building forms are illustrative.
Sources
Engineering at Meta · Published October 19, 2026
Inside Prometheus: Meta’s Journey to a Gigawatt-Scale Supercluster Reviewed October 24, 2026Meta / Open Compute Project · Published September 4, 2025
Clemente compute tray specification v1.01 Reviewed October 24, 2026NVIDIA · Reference hardware
DGX GB rack hardware guide Reviewed October 24, 2026
Reported figures
| Claim | Value | Status / scope | Source section |
|---|---|---|---|
| First phase | Complete | Reported complete Prometheus first phase; used to train Meta’s Muse models | Introduction |
| Reported deployment | Hundreds of thousands of GPUs | Reported operating Prometheus across interconnected buildings | The Supercluster: Unlocking New AI Capabilities |
| Full-scale goal | ~1 million GPUs | Planned Full-scale Prometheus, not the first phase | The Supercluster: Unlocking New AI Capabilities |
| GPUs in the NVLink domain | 72 | Documented configuration One Clemente NVLink domain | Clemente: NVIDIA GB300 NVL72 |
| Compute racks per NVLink domain | 1 | Documented configuration One Clemente NVLink domain | Clemente: NVIDIA GB300 NVL72 |
| GPUs on one building network | Up to 20,736 | Documented configuration One building-level Ethernet fabric | Network Fabrics |
| Distance between connected building networks | 2 km to over 100 km | Documented configuration Backend Aggregation (BAG) network across Prometheus | Aggregation |
| GPUs a job uses in each 72-GPU domain | 64, with 8 in-rack spares | Reported practice Each NVL72 domain in Meta’s training runs | The 72-GPU Failure Domain |
| Clemente GPU memory | 288 GB HBM3e per GPU (20.7 TB per rack) | Documented configuration One Clemente GB300 NVL72 rack | Clemente: NVIDIA GB300 NVL72 |
| Clemente NVLink bandwidth | 130 TB/s per domain (1.8 TB/s per GPU) | Documented configuration (peak) One Clemente NVLink domain | Clemente: NVIDIA GB300 NVL72 |
| Clemente rack power | About 135 kW | Documented configuration One Clemente rack | Clemente: NVIDIA GB300 NVL72 |
| Clemente network per GPU | 800G (ConnectX-8) | Documented configuration Backend network bandwidth per Clemente GPU | Clemente: NVIDIA GB300 NVL72 |
| Catalina racks per NVLink domain | 2 (36 GPUs each) | Documented configuration One Catalina GB200 NVL72 domain (Ariel variant) | Catalina: NVIDIA GB200 |
| Catalina GPU memory | 192 GB HBM3e per GPU (13.4 TB per domain) | Documented configuration One Catalina GB200 NVL72 domain | Catalina: NVIDIA GB200 |
| Catalina NVLink bandwidth | 130 TB/s per domain | Documented configuration (peak) One Catalina NVLink domain, across both racks | Catalina: NVIDIA GB200 |
| Catalina network per GPU | 800G (2 × 400G ConnectX-7) | Documented configuration Backend network bandwidth per Catalina GPU | Catalina: NVIDIA GB200 |
| Catalina system power | About 120 kW across the two racks | Documented configuration One Catalina rack pair | Catalina: NVIDIA GB200 |