Meta published details of Prometheus, the gigawatt-scale AI supercluster it is building in New Albany, Ohio. The first phase is complete and trained the Muse models, and the site is being scaled towards roughly one million GPUs. The system uses Nvidia GB200 and GB300 NVL72 platforms. Meta reports goodput above 85% in pre-training and above 90% in reinforcement learning runs.

What it means: The post gives a rare look at how training the biggest models actually works. The most interesting part is not the GPU count but reliability. In systems with hundreds of thousands of processors something is always failing, and every interruption costs time and energy. Meta publishing goodput and mean time between interruptions shows that the race is shifting from simply buying hardware to operational engineering. Turning a conventional data centre into gigawatt-scale infrastructure within two years also shows how fast investment is moving. Meta's capital spending plans for 2027 are due soon and will show whether the pace continues. For Europe and Greece, where energy is more expensive and permitting slower, such scale is hard to reach. That deepens reliance on American providers for the largest models. At the same time it creates openings for smaller, specialised data centres serving applications that need low latency or strict rules on where data is stored.

Read the original on StorageReview

Related service: Custom software

Related stories