$Meta Platforms(META.US) just showed what it takes to run a gigawatt AI supercluster, and it's scaling the site toward roughly 1 million $NVIDIA(NVDA.US) GPUs

Meta engineers published a rare look inside Prometheus in New Albany, Ohio on Monday. The first phase is done and trained the Muse models, including Muse Wonder

Two years ago that same campus was a regular data center serving web traffic

It runs $NVIDIA(NVDA.US) GB200 and GB300 NVL72 racks today, with Vera Rubin NVL72 racks coming next

The wildest numbers are on reliability. In mid 2025 the cluster peaked at nearly 1,000 interruptions per 8,000 GPUs per day. A year of work with $NVIDIA(NVDA.US) engineers stationed on site cut that by 99%

Each 72 GPU rack only puts 64 GPUs on the job and keeps 8 as spares, so a failed chip gets swapped without waiting on a repair. A 32,768 GPU training run actually needs about 13% more GPUs than it uses

Power is just as extreme. Tens of thousands of GPUs moving in lockstep swing the load 30% to 40%, and starting or pausing a big job can move tens of megawatts in seconds

Next is Hyperion in Louisiana, which $Meta Platforms(META.US) is designing from scratch

Buying the chips is the easy part of the AI buildout. Keeping hundreds of thousands of them alive is where the real spending goes, and spare GPUs sitting idle in every rack mean demand for $NVIDIA(NVDA.US) hardware runs even higher than the compute numbers suggest

Source: StockMarket.News