Inside Prometheus: Meta says its Ohio supercluster reached more than 85% goodput for pre-training

Meta has published a detailed engineering account of Prometheus, the AI supercluster at its data centre in New Albany, Ohio. The post, on Meta's engineering blog on 19 October 2026 and written by Kevin Lee, Yashar Bayani, Ahmed Sharif and Yee Jiun Song, covers the hardware, network, reliability and power choices behind it. Everything below is Meta's own account; we have not independently verified any of the figures.

Where it stands

Meta describes Prometheus as "our gigawatt-scale supercluster" and says its first phase "is already complete". It says that "Today, Prometheus comprises an array of hundreds of thousands of GPUs across multiple interconnected buildings", and that "At full scale, Prometheus will comprise roughly 1 million distributed GPUs". Meta says the first phase was used to train its Muse models, including Muse Wonder (our earlier report: Meta releases Muse Wonder, which it calls its most capable model).

The site is not one building. Meta says it spans "traditional and AI-optimized data center buildings and rapid deployment structures", and that it also integrates "leased capacity from strategic partners".

The headline reliability numbers

Meta says that by working on stability it was able "to achieve more than 85% goodput for pre-training runs and more than 90% goodput for reinforcement learning (RL) runs of Muse Wonder". Meta explains that large jobs need everything working, because "an issue on a single host can slow down or stop the entire job, lowering its ‘goodput’".

Other figures in the post:

• Mean time between interruptions (MTBI) of "~47k GPU hours" during those runs, which Meta says firmware fixes have since raised "to more than 54K GPU hours".
• In mid-2025, interruptions "reached a peak of nearly 1,000 events per 8K GPUs daily". Meta says a year of work produced "a 99% reduction in interruptions".
• Its target, or "reliability north star", is "two interruptions per 8K GPUs per day, which is ~100k GPU hours MTBI".

A chart in the post plots interruptions per 8,000 GPUs per day from June 2025 to September 2026, on a log scale, for GB200 and GB300 hardware labelled "on-prem" and for cloud capacity. On our reading of the chart, the GB300 line ends just below the two-per-day goal line and the GB200 line just above it by September 2026, while the cloud line, which starts in September, sits higher. Seven of the chart's eight annotations involve firmware or driver changes; the eighth is NVLink link-recovery checks.

The hardware choices

Meta says Prometheus runs two NVIDIA platforms, with a third, based on Vera Rubin, "arriving soon":

• Catalina (GB200 NVL72). Meta chose a two-rack "Ariel" variant that gives each GPU 800G of scale-out bandwidth. It says the GB200 raised compute and memory figures over the H100 while scale-out per GPU would have stayed at 400G, and "We expected that imbalance to cost performance".
• Clemente (GB300 NVL72). One rack of 72 GPUs, at "~135 kW rack power". Meta lists a "5-10% performance improvement over the GB200 NVL72".

The bigger 72-GPU domain cuts both ways, Meta says: it brings bandwidth but increases "the fault blast radius by 9X" compared with eight-GPU systems. Because "Training parallelism maps most efficiently onto power-of-two group sizes", Meta says, each rack runs a 64-GPU slice and keeps the other eight as in-rack spares, plus "roughly 2% additional racks in reserve". Its worked example: a 32,768-GPU training run means "provisioning 523 NVL72 racks, or 37,656 GPUs", about 13% above the active footprint, of which Meta says most "(11%) falls out of the 64-of-72 slice geometry rather than being an absolute reliability tax".

The network

Meta says it runs a separate Ethernet "Back-End (BE) network" for GPU-to-GPU traffic. Each building-level fabric "interconnects up to 20,736 GPUs". A Backend Aggregation network then links buildings "spanning 2 km to over 100 km", and Meta says running pre-training across it "required a dedicated optimization effort", "especially given speed-of-light constraints at these distances". The post puts the whole network at "millions of miles of fiber strand, tens of thousands of switches, and millions of optics".

Power and cooling

When tens of thousands of GPUs switch between compute and communication in step, Meta says they cause "sudden load fluctuations of 30-40% within the cluster", and that starting or pausing a large job "can shift loads of more than tens of megawatts in just seconds". Its software, Dynamo Dimmer, "gradually adjusts GPU power limits" near the electrical ceiling rather than stopping work, backed by rack batteries, capacitor trays and site-level battery storage. Cooling is a mix of air-assisted liquid cooling and facility liquid cooling.

Meta says it pays "the full cost of energy for our site" in New Albany. It says its next supercluster, Hyperion in Louisiana, "is being designed from the ground up with insights from Prometheus in mind".

Why it matters

On our reading, the post says as much about failure as about speed. It sets out the interruption rates, the spare capacity Meta holds back, and the power swings a single training job can cause. These are Meta's own figures for its own systems, and the post cites no outside audit, but they give a rare view of what it takes to keep a frontier training cluster running.