Inside Prometheus: Meta Details Gigawatt-Scale AI Supercluster Scaling to 1 Million GPUs
Meta details Prometheus, its gigawatt-scale AI supercluster in Ohio scaling to 1 million GPUs across NVIDIA GB200, GB300, and upcoming 2.52 EFLOPS Vera Rubin NVL72 racks.
Meta has unveiled the engineering architecture of Prometheus, its gigawatt-scale AI supercluster in New Albany, Ohio, scaling toward 1 million GPUs. Spanning NVIDIA GB200, GB300, and upcoming 2.52-EFLOP Vera Rubin NVL72 platforms with closed-loop liquid cooling, the facility achieved over 90% goodput on Muse Wonder training runs through innovative in-rack sparing and power management.
Scaling toward 1 Million GPUs Across Three Hardware Generations
On October 19, 2026, Meta detailed Prometheus, its most ambitious AI infrastructure initiative located in New Albany, Ohio. Currently scaling toward approximately 1 million interconnected GPUs operating as a unified supercomputer, Prometheus deploys two generations of NVIDIA silicon—the GB200-based Catalina and GB300-based Clemente platforms—while preparing the upcoming San Miguel platform powered by NVIDIAs next-generation Vera Rubin NVL72. Each San Miguel rack integrates 72 Vera Rubin GPUs with 288 GB of HBM4, delivering 2.52 peak EFLOPS of FP4 training compute and 216 TB/s of NVLink bandwidth at 170 kW per rack with direct liquid cooling.
Advanced Ethernet Fabric and 100km Hub-and-Spoke Aggregation
Interconnecting up to 20,736 GPUs per building fabric, Prometheus combines Disaggregated Scheduled Fabrics (DSF) and Non-Scheduled Fabrics (NSF) running Metas open-source FBOSS network stack on 51.2T Minipack3 switches with ConnectX-8 800G SuperNICs. Across regional spans from 2 km to over 100 km, Meta deploys a Backend Aggregation (BAG) network leveraging Arista 7800 chassis and Broadcom Jericho3 silicon with 800G-ZR coherent optics, ensuring petabit-scale throughput without stalling synchronized model parallel collectives.
72-GPU Fault Domains, 99% Interruption Reduction, and Dynamo Dimmer
To overcome the 9x blast radius expansion of NVL72 racks, Meta maps 64-GPU data-parallel slices per rack and dedicates 8 GPUs as hot in-rack spares, reducing rack failure impact and achieving 47k to 54k GPU hours mean time between interruptions (MTBI). Combined with Parallelism-Aware Fault Tolerance (PAFT) and asynchronous delta checkpointing, Meta achieved a 99% reduction in training halts. To handle violent 30-40% electrical transients during synchronous bursts, Metas Dynamo Dimmer dynamically caps GPU power alongside rack battery backup units (BBUs), capacitor trays, and facility battery energy storage systems (BESS).