Meta powers up Prometheus, its giant training supercluster

Meta details Prometheus, its AI supercluster in New Albany, Ohio. The site has already trained the Muse family and aims to scale to about one million GPUs, with challenges in reliability, energy, and networking.

Meta published a detailed technical post on October 19, 2026, about Prometheus, its AI supercluster located in New Albany, Ohio. The group confirms that the first phase was used to train the Muse models, including Muse Wonder, and that it ultimately aims to scale up to around 1 million GPUs. The challenge goes beyond mere engineering prowess: it concerns the training speed of future models, infrastructure availability, and Meta's ability to keep pace with other computing giants.

A site designed to train models, not just to host servers

Prometheus is no longer just a repurposed data center. Meta explains that the New Albany site, which handled web traffic two years ago, has been deeply converted for frontier model training, with several deployment phases and buildings both conventional and optimized for AI. The supercluster already gathers hundreds of thousands of GPUs, all orchestrated as a single machine, giving a clear idea of the scale sought: computing is no longer rented by the day, it is managed like a factory.

The group says it achieved over 85% "goodput" on Muse Wonder pre-training runs, and over 90% on reinforcement learning runs. Recall that goodput measures the portion of computing time actually useful, after interruptions and efficiency losses. Meta adds that an initial reliability milestone showed a mean time between interruptions of about 47,000 GPU hours, later increased to over 54,000 hours after cluster-wide fixes.

Two generations of NVIDIA GPUs, with a third already approaching

Meta describes Prometheus as a very concrete playground for NVIDIA chips. The first platform, Catalina, is based on the GB200 NVL72 in an Ariel two-rack variant: 72 GPUs total, 13.4 TB of HBM3e, 720 PFLOPS in FP8, and about 120 kW for the whole. The second, Clemente, moves to the GB300 NVL72, with 20.7 TB of HBM3e, over 1 exaFLOP in FP4, and about 135 kW per rack. Both architectures use a domain of 72 GPUs as a unified accelerator, but Meta indicates the second offers a 5 to 10% performance gain over the GB200 NVL72.

The next steps are already outlined. Meta announces the arrival of a system based on Vera Rubin NVL72, then a forthcoming San Miguel platform with 72 Vera Rubin GPUs, 288 GB of HBM4 per GPU, and 2.52 exaFLOPS theoretical in FP4 per rack. Computing capacity thus continues to rise, but the real question remains the same: how to maintain such density without turning every failure into a mini-disaster?

The real issue is large-scale reliability

Throughout the post, Meta emphasizes less on raw power and more on failure management. The shift from NVL8 to NVL72 architecture expanded the "blast radius" from eight to 72 GPUs, a much larger potential failure domain. To limit impact, the company says it operates with slices of 64 active GPUs and reserve GPUs, increasing hardware margin but protecting the heaviest training runs.

Interruptions initially peaked at nearly 1,000 events per day for 8K GPUs in mid-2025, before dropping 99% after a year of adjustments. Meta now targets two interruptions per 8K GPUs per day, about 100,000 GPU hours between interruptions. Fixes addressed NVLink signal integrity, memory reliability, switch monitoring, and backend link rerouting. The group also says it reduced the P90 update time of a pod from over seven hours to two hours, which matters when deploying fixes on over 100,000 GPUs within hours.

The network and power become products in their own right

Prometheus does not rely solely on GPUs. Meta developed a backend Ethernet network dedicated to GPU-to-GPU traffic, with fabrics interconnecting up to 20,736 GPUs per building. Links between buildings and regions rely on an aggregation architecture capable of connecting sites from 2 km to over 100 km apart, with throughput of several petabits per second according to the company. The whole mobilizes millions of miles of fiber, tens of thousands of switches, and millions of optics.

Electrical power follows the same logic. Large training runs trigger load variations of 30 to 40% in the cluster, with jumps of several tens of megawatts in seconds. Meta therefore uses battery modules, capacitor banks, and a framework called Dynamo Dimmer to avoid abrupt outages when the site nears its limit. On the cooling front, the company combines direct liquid cooling, air assistance, and AI-optimized buildings, with rapid deployments when capacity must be pushed without waiting for a new campus. The cost of this scaling is obviously less "magical cloud" and more pipes, electricity, and megawatts; the poetry of data centers has its limits.

What Meta is preparing next

Beyond Ohio, Meta presents Prometheus as a testbed for Hyperion, its future supercluster in Louisiana. The Richland Parish site will adopt several technical choices validated at New Albany, notably closed-loop liquid cooling and full connection to the power grid via Entergy. In parallel, Meta says it is working with NVIDIA on a highly advanced co-engineering model, with shared telemetry and embedded teams on site.

The trajectory is clear: more GPUs, more bandwidth, more physical constraints as well. In this context, Prometheus serves less as a showcase and more as an industrial laboratory for AI now built at the scale of megawatts, racks, and fiber optic cables.

By the numbers

  • 1 million GPUs — Prometheus scaling target, Meta, 2026.
  • 85% — minimum goodput announced on Muse Wonder pre-training.
  • 99% — reduction in interruptions after one year of optimizations, according to Meta.
  • 20,736 GPUs — interconnection capacity per building fabric.
  • 2 interruptions — daily target per 8K GPUs for reliability.

To read