Prometheus wakes in Ohio / hundreds of thousands of chips today, a million at full size — but not yet
First phase complete — and it already trained Muse
Two years ago, Meta says, the New Albany site was an ordinary web data center, serving the traffic of websites and apps. The company rebuilt it for AI, in phases, and the milestone announced this week is deliberately precise: the first phase is finished, and it did real work. Prometheus's opening act was training the Muse foundational models, including what Meta calls the Muse Wonder frontier models, which it says deliver state-of-the-art performance on personal and professional agentic tasks — AI that does not just answer, but acts. That sequencing matters. Superclusters are usually announced at groundbreaking, rendered in artist's impressions and measured in promises. Prometheus was announced at the point where its first product — a trained frontier model family — already exists. Everything else in Meta's long engineering account, published by Engineering at Meta on October 19, 2026, is the company explaining what it took to get there, and what scaling the rest of the way will require.
The performance figures Meta chose to publish are reliability figures, and that choice is itself informative. The company reports more than 85 percent 'goodput' on Muse Wonder pre-training runs — goodput being the share of compute time that actually advances training rather than being lost to failures and restarts — and more than 90 percent on reinforcement-learning runs. Its wall-clock mean time between interruptions reached about 47,000 GPU hours during those runs, and later fleet-wide changes, including firmware fixes, raised that to more than 54,000 GPU hours. Its stated north star is about 100,000 GPU hours — a measure of how much headroom Meta thinks remains.
Hundreds of thousands of chips, miles of fiber, one machine
Today the site holds hundreds of thousands of GPUs across several buildings — traditional data center halls, a purpose-built AI-optimized building and rapid-deployment structures — linked by miles of fiber so they function as a single machine rather than a campus of separate ones. The engineering detail shows what 'single machine' costs. Each building-level Ethernet fabric can connect up to 20,736 GPUs; above that sits a Backend Aggregation network spanning distances from 2 kilometres to more than 100 kilometres, knitting building fabrics together; and across the whole cluster Meta counts millions of miles of fiber strand, tens of thousands of switches and millions of optical links. At full size, the company says, Prometheus will comprise roughly one million distributed GPUs, all functioning in lockstep. Meta is explicit that the million-chip machine is the destination, not the present state — the present state is hundreds of thousands of chips and a completed first phase.
Keeping a million-chip machine from stopping
The hard problem, as Meta tells it, is not arithmetic but failure. A single faulty host can slow or halt an entire training job, because tens of thousands of chips advance together and wait for each other. Prometheus's hardware generations sharpen the risk: its racks are built around Nvidia's NVL72 design, in which 72 GPUs act as one tightly coupled accelerator — Catalina, based on the GB200, and Clemente, based on the GB300, are deployed today, with a Vera Rubin-based platform, San Miguel, arriving next. A 72-GPU domain multiplies bandwidth and multiplies the blast radius of a fault by nine compared with the previous eight-GPU generation, so Meta runs each domain as a 64-GPU working slice and holds eight in-rack spares ready to step in the moment a GPU fails, restarting from the last checkpoint instead of waiting for a repair. Meta says interruption rates that peaked near 1,000 events per 8,000 GPUs per day in mid-2025 have since fallen by 99 percent, enough to sustain several training jobs of tens of thousands of GPUs at once. Whether those reliability claims survive independent scrutiny is unknowable from outside — no outsider audits Meta's training logs — but the direction of the engineering is unmistakable: at this scale, uptime is the product.
Liquid to the chip, and the power problem
Heat and power are the other two walls. Prometheus cools its newest halls with direct-to-chip liquid cooling in a closed loop — coolant circulating to the chips themselves and back, sealed rather than evaporated away — in air-assisted form in its free-air buildings and rapid-deployment structures, and as full facility liquid cooling in the AI-optimized building. On power, Meta discloses a problem few operators discuss publicly: when tens of thousands of GPUs swing between compute and communication phases in unison, the cluster's load can surge or drop by 30 to 40 percent in moments, shifting tens of megawatts in seconds — a pattern no grid is built to absorb gracefully. Its answer is a power-management framework that dims GPU power gradually near the site's electrical ceiling, backed by rack-level batteries and facility-scale storage, in coordination with its utility. And on the politics of that power, Meta states its position in terms: it partners with local utilities in New Albany to pay the full cost of the energy the site uses, tries not to add unnecessary strain to the local grid, and matches its use with new clean and renewable energy developed with partners.
Why this matters: a gigawatt is a power plant, and the race is now measured in them
Why this matters begins with the unit. A gigawatt is roughly the output of one large power station — enough electricity, on conventional measures, for several hundred thousand homes. A typical conventional data center draws tens of megawatts; a gigawatt-scale AI campus therefore represents something like ten to fifty ordinary data centers' worth of load concentrated on one site, arriving in a handful of years rather than decades. That is the context in which to place Prometheus, and the context in which the whole industry now operates: as we reported in our coverage of the projected 32-gigawatt AI power shortfall, analysts now tally the AI build-out in gigawatts the way they once tallied it in chips, because power — not silicon — has become the binding constraint. Prometheus is Meta's proof that it can clear that constraint at the largest scale attempted, and its design is already a template: Meta says its next supercluster, Hyperion in Richland Parish, Louisiana, is being designed from Prometheus's blueprint.
The local argument deserves both of its sides. Meta's case is that a customer paying its full energy costs, coordinating load swings with the utility, matching consumption with new renewable generation and cooling with a closed loop that uses water efficiently is close to the best-behaved gigawatt neighbour the current technology allows — and that the alternative to building in Ohio is not no build-out, but the same build-out somewhere with weaker terms. The critics' case, heard wherever these campuses land, is that 'paying full cost' is contested ground — grid upgrades, water systems and land are shared inheritances, a single customer arriving at power-station scale changes a region's energy future whether or not its own bill is covered, and the jobs inside a finished data center are few relative to its footprint. Both cases can be argued honestly from the same facts, and Prometheus will be a test of which one a community concludes from, because its scale makes the effects impossible to hide in either direction. For Meta itself, the strategic read is simpler. The Muse models are meant to flow into products used by billions of people — the same distribution machine Meta showed off in our coverage of Meta Connect 2026 and its Ray-Ban Meta glasses — and a company whose rivals rent or scramble for training capacity now owns, outright, the factory floor its frontier models are built on. In an arms race where everyone announces gigawatts, Meta has done the rarer thing: it has switched the first one on and published the interruption statistics.
What happens next
The build continues in phases from hundreds of thousands of GPUs toward roughly one million, with the Vera Rubin generation — more compute per watt, fully liquid-cooled at over 170 kilowatts per rack — folding into the same network fabric as it arrives. Three things will tell outsiders how the story is really going. First, cadence: whether the next Muse model generations arrive on a schedule that suggests the cluster's goodput keeps climbing toward Meta's 100,000-GPU-hour target. Second, replication: whether Hyperion in Louisiana rises faster than Prometheus did, which would confirm that the Ohio design is genuinely a blueprint rather than a one-off feat. Third, the local ledger: utility filings and community reporting in New Albany will show, in public, whether a gigawatt of AI load and a town's grid can in fact coexist on the terms Meta describes. The million-chip headline is still ahead. The era it announces — in which the world's largest computers are measured against power stations — has already started.
Sources: Engineering at Meta, 'Inside Prometheus: Meta's Journey to a Gigawatt-Scale Supercluster,' October 19, 2026 (first phase complete; Muse training; hundreds of thousands of GPUs toward roughly one million; goodput and MTBI figures; GB200/GB300 platforms and Vera Rubin to come; network, cooling, power and utility arrangements; Hyperion). Secondary summary: AppVolt, 'Meta is linking about 1 million AI chips in Ohio to work as one giant computer,' October 20, 2026. Related on Signal Post News: The 32-gigawatt AI power shortfall · Meta Connect 2026.