Meta is turning a former conventional data center operation into one of its largest AI computing deployments, with plans to scale its Prometheus supercluster toward roughly 1 million GPUs.
Originally a conventional facility routing routine web traffic, the campus now harnesses hundreds of thousands of Nvidia GPUs acting as a unified compute engine. Meta plans to scale the deployment toward roughly 1 million GPUs. The first phase has already completed pre-training runs for Meta’s Muse foundation models, including its Muse Wonder agentic models.
For infrastructure providers and channel partners, the project highlights a growing challenge in AI deployments: delivering enough computing power is only part of the equation. Keeping high-density GPU systems running reliably also requires increasingly sophisticated networking, cooling, power management, and maintenance.
Engineering around the failure domain
Running frontier models at gigawatt scale introduces severe reliability hurdles. Because an issue on a single host can drag down an entire job, Meta emphasized fleet stability. The company achieved over 85% goodput during pre-training runs and more than 90% for reinforcement learning on Muse Wonder. Mean time between interruptions reached more than 54,000 GPU hours following fleet-wide firmware fixes.
Much of this resilience stems from handling Nvidia’s NVL72 rack architecture. Shifting from eight-GPU systems to 72 tightly coupled GPUs expanded the failure blast radius by nine times.
To counter this, Meta runs 64 active GPUs per rack while keeping eight as immediate in-rack spares, alongside roughly 2% whole-rack reserves. When a GPU drops, an onboard spare steps in so the workload can restart from the last checkpoint rather than waiting on physical repairs.
Prometheus currently runs two primary platforms: Catalina, based on the Nvidia GB200 NVL72 across a two-rack configuration, and Clemente, which packs 72 Nvidia Blackwell Ultra GPUs into a single GB300 NVL72 rack drawing about 135 kilowatts. Meta will soon introduce San Miguel, a liquid-cooled rack featuring 72 Vera Rubin GPUs that pushes power beyond 170 kilowatts per rack.
Bridging campuses over Ethernet
Rather than relying on proprietary fabrics between structures, Prometheus runs an Ethernet-based back-end network connecting up to 20,736 GPUs per building fabric. To bridge these facilities across distances ranging from 2 kilometers to over 100 kilometers, Meta deployed a multi-layer Backend Aggregation network built on Arista chassis with Broadcom Jericho3 silicon.
Physical deployment was expedited using rapid-deployment structures—heavy-duty tent enclosures spanning roughly two football fields that cut build schedules down to months. Inside the halls, Meta paired air-assisted liquid cooling with facility closed-loop liquid cooling.
To manage sudden 30% to 40% cluster power swings, Meta relies on its “Dynamo Dimmer” software, local battery backup modules, and capacitor trays to smooth electrical transients.
The blueprint shift for enterprise providers and channel partners
Meta’s Prometheus disclosures highlight an important transition for the enterprise channel: AI infrastructure value is migrating rapidly from bare silicon toward systems integration, thermal facilities, and interconnect stability. As hyperscalers validate hardware architectures, corporate data centers looking to stand up on-premises generative AI cannot treat high-density GPU nodes as simple plug-and-play servers.
For value-added resellers, systems integrators, and managed service providers, this operational shift creates a major advisory opening. Mid-market enterprises lack the co-engineering resources to embed vendor engineers on-site or design custom power-capping frameworks like Dynamo Dimmer.
Most mid-market enterprises will not deploy AI systems approaching Meta’s scale. However, the underlying challenges of maintaining reliable, high-density compute environments can still affect smaller installations.
Channel partners that can assess power and cooling capacity, integrate specialized hardware, and support ongoing infrastructure maintenance may be better positioned to help customers expand AI workloads without disrupting existing IT operations.
Other news: Revenium has launched its first partner program, bringing together companies including Insight, Flexera, and World Wide Technology (WWT) to help businesses track AI spending and measure return on investment.