Meta scales Ohio Prometheus campus toward one million GPUs

Meta now runs hundreds of thousands of GPUs across its Prometheus data center campus in New Albany, Ohio, and plans to scale the site to roughly one million distributed GPUs. The company disclosed the figures in a blog post detailing the conversion of the facility from a traditional Facebook data center into the first phase of its AI-focused supercluster.

The deployment is not a lab. Meta said the cluster has already trained its Muse foundation models. The site combines traditional data center buildings, AI-optimized buildings and rapid deployment structures, all interconnected into a single system.

What hardware sits inside Prometheus?

The campus houses Nvidia GB200 and GB300 chips, with the Vera Rubin generation described only as "arriving soon." No date was given. That absence is worth noting, given how much of the roadmap Meta has specified elsewhere.

The GB200 deployment uses Nvidia's Catalina rack architecture. Meta deployed what its engineers call "the Ariel variant, a two-rack configuration in which 36 GPUs per rack are interconnected via NVLink to form a unified 72-GPU domain." NVLink is Nvidia's proprietary GPU interconnect. Each Blackwell GPU pairs with a Grace CPU at a 1:1 ratio, plus two ConnectX-7 400G network cards, supporting 800G of input-output per GPU.

For GB300, Meta uses the Clemente platform. Cooling is flexible: Clemente accepts either Meta's modular Air-Assisted Liquid Cooling (AALC) or Facility Liquid Cooling (FLC), where the building's water loop connects directly to the racks.

How much faster is GB300 in practice?

Meta quoted concrete uplifts rather than marketing multiples. "GB300 raises HBM capacity per GPU by 1.5x, from 192GB to 288GB, and dense FP4 throughput by a similar factor, while roughly doubling attention (softmax) acceleration," the company said. HBM is high-bandwidth memory; FP4 is a four-bit floating-point format used for inference. Softmax is a mathematical operation central to transformer attention.

The claim covers memory and compute benchmarks. It does not cover end-to-end training throughput, which depends on the network fabric as much as the silicon. Meta addressed that constraint separately.

Did Meta fix the GB200's networking limits?

By its own admission, scale-out was a constraint on GB200 NVL72. Meta stated: "At the platform level, scale-out is no longer the constraint it was on the GB200 NVL72: Nvidia ConnectX-8 SuperNICs deliver 800G per GPU natively, so the capability we deployed early on Ariel becomes the baseline."

In other words, the workaround Meta engineered for GB200 arrives as standard on GB300. That framing concedes the earlier generation required special effort to reach the bandwidth that GB300 provides out of the box.

At the building level, each Ethernet-based fabric interconnects up to 20,736 GPUs through a multi-layer topology. Building-level fabrics then connect through the Backend Aggregation (BAG) network, spanning 2km to over 100km. "It is built with Arista 7800-series chassis running deep-buffered Broadcom Jericho3 silicon, delivering multiple Pbps of bandwidth per region pair," Meta said. One petabit equals one thousand terabits per second.

What remains unspecified?

The blog post omits several operational details. Meta did not state total power draw, the number of buildings completed, the timeline for reaching one million GPUs, or when Vera Rubin hardware will enter service. A campus of that scale implies grid commitments that Meta has not detailed here.

The gigawatt target itself deserves scrutiny. Reaching one million GPUs requires sustained supply of racks, power and cooling at a pace the industry has not yet demonstrated publicly. Prometheus is live, and the first phase has trained real models. The rest of the roadmap remains a claim, not a deployment.