The Infrastructure Behind the Race to Superintelligence

Meta said on October 19 that the first phase of Prometheus, its computing cluster in New Albany, Ohio, was complete and had trained its Muse models. The company is expanding the cluster toward future models from Meta Superintelligence Labs. Meta’s latest engineering account.

An October 12 announcement named Muse Wonder and Muse Wonder Full as models trained at Prometheus and described them as a step toward Meta’s goal of personal superintelligence. Meta’s model announcement; official location confirmation.

In Richland Parish, Louisiana, Meta is preparing Hyperion on a larger scale. Its July expansion announcement set out 5 gigawatts of planned “compute capacity” and more than $50 billion of investment in the region. That announcement described a future training cluster; the Ohio disclosure identifies models already trained. Meta’s Louisiana expansion.

Connecting a distributed training cluster

Prometheus is spread across conventional data-center buildings, weatherproof outdoor structures and adjacent leased space. Meta’s October 12 engineering account describes a network layer that joins those locations into one cluster and puts its 2026 scale at roughly 1 GW. The later October 19 account says hundreds of thousands of GPUs are in place, with roughly one million planned at full scale. Completion of the first phase is not confirmation that the whole 1 GW buildout is commissioned; neither account gives measured cluster-wide power demand. October 12 network account; October 19 phase update.

The network is part of the infrastructure challenge. Meta says longer links complicate traffic management and can allow congestion or failures to affect the wider system. Its designs seek to keep training running as traffic moves among buildings and equipment fails.

GamesBeat’s September reporting on a presentation by Meta infrastructure chief Santosh Janardhan described training jobs running while construction continued. He also described construction dust causing problems for fiber connections, a practical hazard when operating and unfinished facilities share a site. GamesBeat’s report.

Socrates South, a gas-fired plant built to supply Meta’s New Albany data-center campus, entered service in late July, Williams reported in August. At that time, its sister North facility was still under construction, with service expected in the fourth quarter. Williams’ filing, page 52; Data Center Frontier’s account of the supply arrangement.

Nuclear power is part of Meta’s longer-term supply plan. Agreements announced in January cover up to 6.6 GW of existing and prospective generation by 2035 across grids serving its operations. The package spans existing Vistra plants, planned increases in their output and advanced-reactor development. The figure describes that wider portfolio rather than power already delivered to Prometheus. Meta’s nuclear agreements.

New Albany’s government clarified that the agreements do not involve a nuclear plant in the city. City statement.

Hyperion requires a larger power buildout

Louisiana regulators approved the initial package of three gas-fired generating units and associated transmission in August 2025. Entergy broke ground that December on the two Franklin Farms units in Richland Parish, targeting late-2028 operation. Commission order; Entergy’s groundbreaking announcement.

The expansion has a separate regulatory track. In March 2026, Entergy proposed seven additional gas-fired units totaling more than 5,200 MW, alongside transmission and other resources. An October settlement involving Entergy, commission staff and Walmart remained subject to commission approval. The October 19 administrative-law-judge report recorded unresolved contested issues and set post-hearing briefs for November. Entergy’s March proposal; October settlement filing; administrative-law-judge report.

WVUE’s May visit documented construction at Entergy’s facilities and the local pressures accompanying the buildout. Entergy said Meta would cover construction costs. The Alliance for Affordable Energy questioned the risk to other customers if Meta’s demand did not endure. WVUE’s site report.

The gigawatt figures also measure different things. Generation capacity describes what power plants can produce. Facility demand includes cooling and power-delivery equipment as well as computers; IT power capacity covers the computing equipment. The Department of Energy makes the facility-versus-IT distinction in its data-center efficiency guidance. DOE guidance.

Meta labels Hyperion’s 5 GW target “compute capacity.” How that planned figure relates to Entergy’s generating capacity and the campus’s eventual electricity demand remains unspecified.

Financing the buildings and infrastructure

In October 2025, Meta and funds managed by Blue Owl Capital formed a venture with an estimated $27 billion development cost covering buildings and long-lived power, cooling and connectivity infrastructure. Blue Owl-managed funds own 80%; Meta retains 20% and provides construction and property management. Joint-venture announcement.

Meta’s second-quarter 2026 filing says leases for the Louisiana venture are scheduled to begin in 2029, with initial four-year terms and renewal options. It also discloses conditional guarantees tied to the properties’ residual value if Meta terminates or declines to renew a lease. The 2029 date applies to those leases; the filing gives no completion date for the full 5 GW expansion. Meta’s quarterly filing, Note 5.

The $27 billion venture and the later announcement of more than $50 billion of regional investment cover different scopes and cannot simply be added together. The July expansion announcement leaves its effect on the venture’s financing unexplained.

Stargate and Colossus provide related examples

OpenAI has also identified a model trained at a named campus. In April, it said GPT-5.5 was trained at Stargate Abilene in Texas on Oracle Cloud Infrastructure using Nvidia GB200 systems. Abilene is one site in Stargate, the wider compute program supporting OpenAI’s artificial general intelligence ambitions. OpenAI’s account.

On its September 10 earnings call, Oracle said six of the campus’s eight buildings had been delivered to its customer. Developer Crusoe has separately announced an adjacent Microsoft campus. That development is separate from the existing OpenAI/Oracle site. Oracle management’s account; Crusoe’s project boundaries.

Colossus, developed by xAI in Memphis, supplies another documented training example. In July 2025, xAI said it used the cluster for Grok 4’s reinforcement-learning training. SpaceX acquired xAI in February 2026. Grok 4 disclosure; acquisition announcement.

In May, Anthropic announced that it had agreed to use all computing capacity at Colossus 1 to expand Claude capacity, with access expected that month. On May 28, Anthropic said its SpaceX agreements covered GPU capacity in both Colossus 1 and Colossus 2. SpaceX’s October 20 results describe the continued buildout of Colossus II. These disclosures do not provide a current site-by-site account of workloads or utilization. Anthropic’s initial announcement; May 28 scope update; SpaceX’s results.

From construction to capability

Prometheus has produced named models, according to Meta. Hyperion’s larger plan depends on construction and electricity resources arriving in phases. Across these projects, buildings, working training clusters and released models mark different stages of progress. Whether the models achieve superintelligence is a separate question about their capabilities, one that the infrastructure announcements cannot answer.