COMPUTE ATLAS Supercomputer supply-chain graph
1153 systems 482 sites Sourced data

Blog/2026-10-07

· Compute Atlas

Meta Hyperion and Prometheus: tents and gas

Meta's Prometheus in Ohio is already training models, with Epoch AI estimating 471 MW of IT power. Hyperion in Louisiana is a $27 billion Blue Owl joint venture that Meta now says will reach 5 GW, with no GPU count disclosed.

Meta’s two largest training sites are Prometheus in New Albany, Ohio, and Hyperion in Richland Parish, Louisiana. Prometheus is the one in service: Meta’s head of infrastructure says it holds hundreds of thousands of GPUs across dozens of buildings, with tents used to speed construction, and Epoch AI estimates 471 MW of IT power and about 166,000 NVIDIA B300 chips as of August 2026. Hyperion is still under construction. Its owner is a joint venture in which funds managed by Blue Owl Capital hold 80 percent and Meta 20 percent, its first phase was financed with a $27 billion development budget, and Meta now says the campus will reach 5 GW. Meta has published no GPU count for either site, no accelerator model for Hyperion, and no date for the full 5 GW. Our rows are Meta Prometheus and Meta Hyperion, both at the “reported” tier, with Hyperion recorded as under construction.

From 24,576 GPUs to titan clusters

The two sites sit at the end of a short, well-documented scaling history. Meta’s March 2024 engineering post described two clusters of 24,576 NVIDIA H100 GPUs each, one on RoCE Ethernet and one on Quantum2 InfiniBand, built to train Llama 3 and succeeding the 16,000-GPU A100 Research SuperCluster of 2022. It also said Meta aimed to reach 350,000 H100s by the end of 2024, as part of a portfolio with compute equal to nearly 600,000 H100s.

In October 2024 Mark Zuckerberg told analysts that Llama 4 was training on a cluster of more than 100,000 H100s, per a report of the Q3 earnings call, which the report says was a larger cluster than any other company had reported. Santosh Janardhan, Meta’s head of infrastructure, later described that 2024 cluster as about 100,000 H100s and said the scaling laws pushed Meta to build one roughly ten times the size. That successor became Prometheus.

On 14 July 2025 Zuckerberg published a post, summarised by Converge Digest, saying Prometheus would come online in 2026 with more than 1 GW, that Hyperion would scale to 5 GW over several years, and that more “titan clusters” were in development. Meta also said it would invest hundreds of billions of dollars in compute. We link the trade summary because we could not retrieve the post itself.

Prometheus: a campus of tents

Prometheus was built in an unusual way. Janardhan described it at the AI Infra Summit, as reported by GamesBeat, as thousands of acres and 30 to 40 or more buildings, with the buildings separated by three to five kilometres and tied together by a purpose-built Ethernet aggregation layer. Meta used weatherproof tents about two to two and a half football fields in size, each holding 30 MW, rated for level-one hurricanes. Janardhan said the tents cut deployment time by 70 to 80 percent, taking it from years to months. The same report says Prometheus relies on off-grid natural gas generation and, over time, nuclear power. It remained under construction while running training jobs and adding GPUs, and early deployments suffered interruptions that Meta cut by a factor of 50.

Epoch’s satellite work agrees on the form. It calls Prometheus a patchwork of tents, colocation buildings and conventional data center buildings along the New Albany Business Park, and notes that Building 5 is owned by QTS. Its timeline has building 1 operational in October 2025 (87,000 H100-equivalents, 76 MW), buildings 2 to 6 in January 2026 (368 MW), and buildings 7 to 10 by August 2026 (471 MW). On the Socrates South power plant beside the campus it counts ten gas turbines in place by March 2026.

Our dataset holds the Epoch numbers as estimates: 166,015 accelerators, 471 MW IT power, 612 MW facility power and a modelled $17.8 billion cost. Baxtel’s listing instead gives a stated $1.5 billion investment for one Prometheus site, which cannot be compared to a modelled cost for the whole campus. Epoch’s own page is not fully reconciled either: its headline compute, 536,000 H100-equivalents, exceeds the 419,000 shown in its B300 breakdown, and we could not tell from the page what fills the difference.

Compare Janardhan’s “hundreds of thousands of GPUs” with Epoch’s 166,000 B300 chips. They can both hold if Prometheus mixes older H100-class and newer chips, but neither source says so, and Meta has not stated an accelerator model for the site.

Hyperion: who owns it and how it is paid for

Hyperion’s financing is the best-documented part of the story. Meta’s 21 October 2025 release says funds managed by Blue Owl hold 80 percent of the joint venture and Meta 20 percent, for about $27 billion of development cost covering the buildings and long-lived power, cooling and connectivity infrastructure. Blue Owl contributed about $7 billion in cash and Meta received a one-time distribution of about $3 billion. Meta provides construction and property management and leases the facility under operating leases with a four-year initial term and extension options. It also issued a 16-year residual value guarantee that caps its cash payment if a lease is not renewed.

The debt side went to bond investors. IFR reports that a vehicle called Beignet Investor priced $27.3 billion of 6.581 percent senior secured bonds due May 2049 on 16 October 2025, rated A+ by S&P, with PIMCO as anchor investor. IFR quotes S&P saying the rating sits one notch below Meta’s because the structure passes substantial credit risk to Meta in construction and operation, and describes the site as 2.1 GW with expected completion in June 2029. That framing matters for how to read Hyperion: Blue Owl is the owner of record, but Meta bears much of the risk through its lease and guarantee. We do not know the exact lease economics beyond what Meta’s release states.

The scale has since grown. Meta’s Richland Parish page now cites more than $50 billion of investment, 5 GW of compute capacity and 7,500 skilled trade workers at peak. A July 2026 report carried by Yahoo Finance dates the expansion to 13 July 2026, says the original design was 2 GW and $27 billion, and says Meta expects 2 GW by 2030 with no date for the full 5 GW. We record $50 billion as a campus-wide figure, not first-phase cost, and note that the original $27 billion was a joint venture budget for buildings and infrastructure, not the whole campus.

Power: Entergy’s gas build

Hyperion’s electricity is meant to come mainly from new gas plants built by Entergy Louisiana, and Meta says it pays the full cost of the energy it uses. Entergy’s December 2024 announcement described three combined-cycle units totalling 2,260 MW, nearly 100 miles of 500 kV transmission, and Meta committing to help fund 1,500 MW of new solar and storage. Louisiana regulators approved the plan on 20 August 2025, according to Industrial Info: Franklin Farms (1,500 MW, mid-2029) and Selpa (754 MW, end of 2029), 2.25 GW in total, plus at least 1,500 MW of solar and a substation funded entirely by Meta. The 2,260 MW and 2.25 GW figures differ slightly because they are two descriptions of the same plan at different dates. Our metrics row records 2,260 MW as generation capacity, not campus load.

The expansion changes the gas count. The Union of Concerned Scientists, an advocacy group opposed to the plan, says that in April 2026 seven further gas plants were proposed, bringing the total to about 7.5 GW with a regulatory decision scheduled for December 2026. The Yahoo report says Entergy’s plans involve seven combined-cycle plants, battery storage at three sites and about 240 miles of high-voltage lines, with expected customer benefits of $2.65 billion, a figure Meta also cites. The reports do not agree on whether the seven are in addition to the original three, and we have not seen the filing.

Epoch’s reading is more cautious. It assumes the three turbines serve a backup role and notes the grid load request of 1,800 MW, with new power stations in service for that load around December 2026. Its modelling of the campus is 1,632 MW of IT power and 816 air-cooled chillers by the first quarter of 2028. Meta describes a closed-loop glycol system with dry cooling for most of the year.

What the numbers do and do not show

Epoch currently counts zero compute at Hyperion. Its roofs on buildings 1 and 2 were complete in April 2026, and it projects a first phase online around January 2028, based on a Meta blog post, with 468,500 NVIDIA Rubin chips, 1,632 MW and $61.8 billion by then. The Rubin assumption is Epoch’s, because Meta has not named the accelerator. Epoch also writes that, so far, there is no concrete evidence of 5 GW being reached at its Holly Ridge location alone, which is how it places the campus. Our dataset keeps the capacity claims separate: 2,000 MW as a first-phase milestone and 5,000 MW as the full build, each with a different source.

Timelines also diverge. Epoch projects a first phase in January 2028, IFR cites a June 2029 completion for the 2.1 GW financed phase, and the July 2026 report says 2 GW by 2030. We cannot tell whether these describe the same milestone. For Prometheus, Zuckerberg’s 2025 target was more than 1 GW in 2026. Epoch has 471 MW of IT power and 612 MW of facility power in August 2026, and projects 1,022 MW of IT power only by the third quarter of 2028. Either the cluster is still filling, or the “gigawatt scale” label describes the full campus design and not current load.

What to watch

Three disclosures would settle most of this: the Louisiana commission’s December 2026 decision on additional gas plants; a Meta statement naming the accelerators at Hyperion; and a measured or filed load for Prometheus. The next Epoch satellite updates will show whether Hyperion’s first buildings are filling with equipment on the January 2028 path, and whether Prometheus’s northern tent sections add capacity beyond the 471 MW Epoch counts today.

What we could not confirm

  • A GPU count or accelerator model from Meta for either site. The 166,000 B300 and 468,500 Rubin figures are Epoch’s modelling.
  • The text of Zuckerberg’s July 2025 post. We rely on a trade summary of it.
  • Whether the seven gas plants in the 2026 expansion add to or replace the three approved in 2025.
  • The timing of the 2 GW and 5 GW milestones; sources give 2028, 2029 and 2030 for different things.
  • What accounts for the gap between Epoch’s headline and chip-breakdown compute for Prometheus.
  • How Prometheus’s “hundreds of thousands of GPUs” reconciles with Epoch’s chip estimate.
  • Blue Owl’s lease economics beyond the headline terms Meta published.

Sources

deep-diveai-clusterspower-and-coolingprocurement