Analysis
Meta said on October 19 that the first phase of Prometheus, its supercluster in New Albany, Ohio, is complete and trained its Muse foundational models, including Muse Wonder. The company describes hundreds of thousands of GPUs across interconnected buildings and a planned expansion to roughly one million. It says the site was converted from conventional web serving over two years.
The hardware choices reveal the engineering priorities. Meta says its Catalina platform combines 72 GPUs across two racks, with 13.4 terabytes of memory and about 120 kilowatts of total system power. Its chosen configuration provides 800 gigabits per second of scale-out connectivity per GPU. The company also describes deploying a newer platform based on NVIDIA’s GB300.
Keeping that equipment productive is a separate challenge. Meta reports more than 85 percent goodput for Muse Wonder pretraining and more than 90 percent for reinforcement learning. It reports roughly 47,000 GPU hours between interruptions during those runs, with subsequent fleet changes raising that measure above 54,000 GPU hours. These are company measurements.
The engineering lesson is that installed capacity depends on sustained operation to become useful training capacity. Yet the account does not establish total project cost, facility power consumption, water use or a completion date. Meta’s gigawatt-scale description therefore should not be read as a measurement of current electricity demand.
Source: Inside Prometheus: Meta’s Journey to a Gigawatt-Scale Supercluster · 2026-10-19 · Primary source context · attributed account