AI Agent UpdateMeta creates Meta Compute; Prometheus supercluster first phase complete, ~1 GW planned
Meta has created Meta Compute, a new division to run its AI data-center ambitions, and detailed the completed first phase of its Prometheus supercluster in Ohio — planned to scale to about 1 gigawatt and roughly one million GPUs.
The News
Meta has created Meta Compute, a new division to manage its AI data-center ambitions. Separately, its October 19 engineering disclosure detailed the completed first phase of its Prometheus supercluster. Together they show how seriously Meta is taking the infrastructure race. Meta Compute is co-led by global infrastructure head Santosh Janardhan and Daniel Gross, who covers long-term capacity planning, supplier partnerships and business modelling. Chief executive Mark Zuckerberg said Meta is planning tens of gigawatts of AI capacity this decade, and hundreds of gigawatts or more over time. A gigawatt is a billion watts, roughly the output of a large power station; tens of gigawatts means tens of power stations' worth of computing, devoted to AI.
Prometheus, in New Albany, Ohio, is where that ambition becomes concrete. The campus is a converted Facebook site now in its first AI supercluster phase, already used to train Muse foundation models — Meta's own AI models. It is planned to scale to about 1 gigawatt and roughly one million distributed GPUs, the specialised chips that train AI models. The site houses Nvidia GB200 and GB300 chips, with Vera Rubin, Nvidia's next chip generation, arriving soon. Each Ethernet fabric — the network that links chips so they can train as one system — interconnects up to 20,736 GPUs, a scale that allows very large models to be trained without splitting them into pieces.
The timing is notable. Meta's Kuna, Idaho data center has also gone live, as its second AI-optimised facility after Temple, Texas — but Prometheus is the flagship, the site Meta is choosing to showcase. Creating a dedicated division and publicising a supercluster's specifications are both signals to suppliers, utilities and investors: Meta intends to compete for power, chips and talent at the largest scale, and it wants the market to know the buildout is organised, not improvised. Nvidia is named here as a supplier read-through, since its chips fill Prometheus; the story itself is Meta's.
Meta just gave its AI buildout its own division — and showed off a supercluster headed for a million GPUs.
Why It Matters
This matters because the AI race is increasingly an infrastructure race. The company with the most computing power, the best-connected chips and the most reliable power supply can train bigger models faster — and bigger, faster models attract more users and developers. By organising its buildout under Meta Compute, Meta is signalling that it understands infrastructure as strategy, not logistics. That organisational choice may matter as much as any single chip or data center.
It matters for Nvidia, too. Prometheus running GB200 and GB300 chips, with Vera Rubin on the way, is a showcase order and a roadmap commitment in one: Meta is not just buying Nvidia's current chips, it is planning around its next generation. For readers tracking the chip cycle, hyperscaler roadmaps like this are leading indicators of demand — they tell you what Nvidia will be selling two years from now, not just what it sold last quarter.
Finally, it matters for Ohio and for the politics of AI infrastructure. A converted Facebook site becoming a million-GPU supercluster is a story of industrial transformation in a single location: an older internet facility reborn as AI infrastructure, with the power, jobs and grid implications that follow. As data-center construction faces ballot measures and local opposition across the US — 41 measures are on ballots on November 3 alone — projects with visible local benefits and clear leadership will have an easier path than anonymous builds. Prometheus is Meta's attempt to be visible, and therefore welcome.
Analyst's Take
The verdict is Bullish, because organisation is itself a competitive advantage at this scale. Building tens of gigawatts of AI capacity is not a side project; it is a programme that needs its own leadership, its own supplier relationships and its own planning horizon. By creating Meta Compute and putting named leaders in charge, Meta is treating infrastructure as a core business, not as a cost centre. And Prometheus's first live phase — already training real models, with a credible path to a million GPUs — shows the programme is producing, not just promising.
The risk, as with every hyperscaler this week, is cost: gigawatts cost billions, and Zuckerberg's 'hundreds of gigawatts or more' is a spending horizon that dwarfs most companies' entire revenues. Meta's advertising business still generates enormous cash, which is its funding backstop for now — but the spending horizon Zuckerberg described dwarfs most companies' entire revenues, so cost discipline remains the risk to watch. For now, Meta Compute looks like ambition with a structure behind it — which is why it earns a bullish mark, even in a week when AI spending is under harsher scrutiny than ever.
By The Numbers
FAQ
What is Meta Compute?
A new Meta division to manage the company's AI data-center ambitions, co-led by global infrastructure head Santosh Janardhan and Daniel Gross, who handles long-term capacity planning, supplier partnerships and business modelling. In plain terms, Meta has given its data-center buildout its own leadership team instead of treating it as a side job.
What is Prometheus?
Meta's AI supercluster campus in New Albany, Ohio — a converted Facebook site now in its first AI supercluster phase, already used to train Muse foundation models. A supercluster is a giant, tightly connected group of computers designed to train AI models together.
How big will Prometheus get?
Meta plans to scale it to about 1 gigawatt of power and roughly one million distributed GPUs — the specialised chips that train AI models. It houses Nvidia GB200 and GB300 chips, with Vera Rubin, Nvidia's next generation, arriving soon.
What does 20,736 GPUs per fabric mean?
Each Ethernet fabric — the network that connects chips so they can work together — interconnects up to 20,736 GPUs. The bigger the interconnected group, the larger the AI model that can be trained as a single system, rather than in pieces.
What did Zuckerberg promise on scale?
CEO Mark Zuckerberg said Meta is planning tens of gigawatts of AI capacity this decade, and hundreds of gigawatts or more over time. A gigawatt is a billion watts — roughly the output of a large power station — so tens of gigawatts means tens of power stations' worth of computing.
👁