Built at Prometheus: Meta's Ohio Data Center and the Compute Behind Wonder

Model announcements rarely spend much time on where the model was actually trained, and when they do, it's usually a single sentence buried past the benchmarks. For Muse Wonder, that sentence is worth more attention than most: TechCrunch's reporting on the October 12 release states that Wonder and Wonder Full were trained at Meta's Prometheus data center in central Ohio, which came online earlier in 2026.

Why the timing lines up

That sequencing — a major new compute facility coming online, followed within months by Meta's first frontier-tier model release — isn't a coincidence worth treating lightly. Wonder is the first release since Alexandr Wang restructured Meta's AI efforts in August 2025, a period during which Meta publicly shipped Muse Spark and the Muse agent but avoided state-of-the-art model releases of the kind OpenAI and Anthropic treat as routine. Prometheus coming online earlier this year removes one of the most plausible bottlenecks for why that gap existed: training a genuinely frontier-scale model requires frontier-scale compute, and Meta's own infrastructure for that appears to have only recently caught up.

Why Ohio, specifically

Meta hasn't published a detailed rationale for the Prometheus site choice in the sources reviewed for this series, but the general pattern for where AI labs site large training facilities is well established across the industry: access to reliable, large-scale power, available land, and a regional grid that can absorb the enormous and sustained electricity draw a frontier-training run requires. Central Ohio has become a recognizable node in that broader buildout for exactly those reasons, alongside other regions competing for the same kind of large-scale data center investment. None of that is Prometheus-specific confirmation — it's context for why a facility like this would plausibly land where it did — but it's worth noting that Meta choosing a region already positioning itself for this kind of infrastructure, rather than building in a genuinely novel location, is itself a small signal that this wasn't an improvised or rushed decision tied specifically to the Wonder release.

What's confirmed and what isn't

What's confirmed, from TechCrunch's reporting, is the facility's name, its general location, and that it came online earlier in 2026, before Wonder's training. What isn't confirmed, in any source reviewed for this series, is Prometheus's specific compute capacity, how much of Meta's total AI infrastructure budget it represents, or whether it's a dedicated training facility or part of a larger mixed-use data center footprint. Neither Meta's own announcement blog post nor the independent coverage reviewed for this series goes into that level of detail — this is a case where the fact itself (a named facility, a rough timeline, a geographic location) is solid, but the scale and significance implied by "Meta's compute behind its first frontier model" is reasonable inference from context, not a disclosed number.

How this fits the broader AI data center buildout

Meta isn't alone in treating purpose-built compute as a prerequisite for frontier-model work — every major lab racing for the top of the benchmark tables has spent 2025 and 2026 announcing new data center capacity, often in the same breath as announcing new models, because the two have become inseparable at this scale. What makes Prometheus notable in Meta's specific case isn't that a tech company built a data center; it's that this is apparently the first facility Meta has pointed to by name as the infrastructure behind a model explicitly marketed as frontier-tier, rather than as general AI infrastructure supporting ads, recommendation systems, or Meta's broader product suite. That specificity — a named facility tied to a named model release — is itself new for how Meta talks about its own infrastructure publicly, and it suggests Meta wants Prometheus understood as dedicated AI-training capacity, not shared infrastructure doing double duty.

Why this matters beyond trivia

Infrastructure investment at this scale is one of the more reliable leading indicators in this industry, precisely because it's harder to fake or spin than a benchmark claim. A company doesn't stand up a dedicated data center on a multi-year construction timeline for a model it doesn't intend to follow with more models. If Prometheus is genuinely a new, dedicated training facility rather than existing capacity relabeled for a press release, it implies Wonder isn't a one-off response to competitive pressure — it's the first output of infrastructure Meta built specifically to sustain a frontier-model cadence going forward, the same kind of cadence OpenAI and Anthropic have maintained for longer.

That reading is consistent with how Meta has positioned Wonder elsewhere — not as a side project, but as, in the company's own words, "the next step on our scaling ladder." A scaling ladder needs rungs to stand on, and a newly online, named data center is the most concrete rung Meta has publicly pointed to.

The part this article can't tell you

It's tempting to use a data center announcement to make bigger claims than the facts support — to treat "Meta built new compute" as proof of a specific capacity number, a specific investment figure, or a specific future release cadence. None of those are available from the sources checked here. What's actually established is narrower and more useful precisely because it's narrower: a named facility, a rough online date earlier in 2026, and a direct connection to Wonder's training. Reading more into it than that — guessing at GPU counts, dollar figures, or Meta's internal roadmap — would be exactly the kind of unsupported extrapolation this series has tried to avoid throughout its run, on both the Muse side and in how it's covered the other two agent launches this cycle.

What to watch next

Whether Prometheus continues to show up in future Meta AI announcements, whether Meta discloses more about its capacity, and whether a second frontier-tier release follows Wonder on anything resembling the cadence OpenAI and Anthropic have set are all open questions this series will keep tracking. In the meantime, the capability Meta is leaning on hardest as proof this investment paid off — Active Memory, and the separate Active Focus claim that's supposed to keep Wonder working through difficulty rather than giving up — gets its own closer look in Active Focus.

Part 36 of 46 in the Meta Muse series. Previously: No Open Weights: What Changed Since Meta's Llama Era. Next: Active Focus: The Feature That's Supposed to Stop Muse From Giving Up. Browse more builder insights or explore AI skills at aiskill.market.