Meta Opened Its Million-GPU Supercluster in Ohio

Meta says the first phase of its Million-GPU Supercluster is complete, and the machine has already trained the company's Muse family of foundational AI models. The announcement came Monday in a detailed engineering post from the team that built the project, which Meta calls Prometheus — a gigawatt-scale training campus in New Albany, Ohio that the company is now scaling toward roughly one million graphics processors running as a single system.

Two years ago, the Ohio site was a conventional data center serving ordinary web traffic. Meta reworked the campus for frontier-model training, building the supercluster in phases across traditional halls, a new AI-optimized building, and rapid-deployment structures joined by miles of fiber. Today the first phase spans hundreds of thousands of chips operating in lockstep, and the next stage is being built with Meta's future frontier models in mind, according to the engineers.

What the training numbers show

Meta published unusually specific reliability figures from the first training runs. During pre-training of its newest frontier models, the cluster sustained more than eighty-five percent goodput — the share of compute time that actually advanced training, rather than time lost to crashes, restarts, or idle hardware. On reinforcement-learning runs, goodput topped ninety percent. The engineers also reported a wall-clock mean time between interruptions of about forty-seven thousand GPU hours during those runs, later raised past fifty-four thousand GPU hours after fleet-wide firmware fixes landed.

Keeping hundreds of thousands of chips from tripping each other up takes serious plumbing. The campus uses closed-loop liquid cooling to pull heat away from the hardware, and a Backend Aggregation network links buildings spread from two to more than one hundred kilometers apart — a distance where Meta had to tune its software around the speed of light itself. The company counts millions of miles of fiber strand, tens of thousands of switches, and millions of optical parts across the cluster. The hardware mixes two generations of Nvidia rack systems, including one with more than an exaFLOP of peak FP4 compute, with a third platform based on Nvidia's next-generation Vera Rubin architecture expected soon.

Fault tolerance turns out to be one of the project's defining costs. Because moving to larger racks multiplied the blast radius of any single failure by roughly nine, Meta over-provisions by about thirteen percent: it dedicates tens of thousands of spare chips across the campus so a single training run of tens of thousands of active processors can survive hardware failures without stopping, according to analysis of the technical disclosures.

Why a million-chip goal matters to you

The models trained on this campus are the engines behind the AI features landing in Meta's apps — the assistants answering questions in your chats, the editing tools in your feeds, the agents the company wants to sell to businesses. A training cluster this large lets Meta build more capable models faster than smaller labs can afford to, and that speed eventually shows up as better free features for the billions of people who use the company's products.

That concentration is also the point. Most commercial AI clusters are measured in tens of thousands of chips, not hundreds of thousands. A facility operating at gigawatt scale — roughly the output of a large power plant — turns frontier-model training into an infrastructure game only a handful of companies can play. When one company can keep hundreds of thousands of processors training a single model for weeks at a time, startups and university labs are effectively priced out of the top tier.

There is a counterpoint, and it showed up in the news the same weekend Meta published its post. The Guardian reported on October 18 that Microsoft had withdrawn from a major AI datacentre project in Britain because the local grid could not supply enough electricity until the 2030s. Gigawatt-scale ambition and gigawatt-scale power have to arrive together — and right now, the power is the harder part to get. Meta's Million-GPU Supercluster proves the engineering works; the open question is whether the world's grids can keep up with it. For more on the infrastructure race, see how Broadcom is hunting over fifty billion dollars for custom AI chips, why SpaceX is borrowing billions to buy Nvidia chips, and how Tencent is leasing advanced chips through Oracle.