HomeFull speed ahead. Keep track of advancements in every frontier

TheTechno Optimist.

Prometheus and HyperionMeta—Prometheus and HyperionMeta—Muse modelsMeta—Meta—Muse modelsMeta—Muse modelsMeta—Muse modelsMeta—
Prometheus and Hyperion
Courtesy of Meta · Press kit, editorial use (opens about.fb.com)

Prometheus and Hyperion

by MetaUS

DeployStage 4 of 5

Prometheus (about 1 GW, Ohio) is due online in 2026. Hyperion (Louisiana) is growing toward several GW.

Updated 13 Jul 2026Checked 25 Sep0 updates this week

Milestones

Next · Prometheus online
  1. Two 24,000-GPU training clusters detailedMar 2024Complete.
  2. Prometheus and Hyperion announced14 Jul 2025Complete.
  3. Hyperion financing joint venture with Blue OwlOct 2025Complete.
  4. Prometheus onlineTarget 2026Current milestone.

Most important updates

  • 13 Jul 2026
  • 21 Oct 2025
  • 20 Aug 2025
  • 14 Jul 2025
  • 4 Dec 2024

Upcoming

  1. 2026Prometheus online (next)

Current obstacles

  • Power supplyHyperion relies on new gas power plants built by the local utility. Grid timelines limit how fast it grows.

Physics limits

  • A gigawatt is a power plant1 GW running all year is about 8.8 TWh, the output of a large nuclear reactor. Sites need dedicated generation or new high-voltage lines, which take years to build.
  • Almost all power becomes heatNearly every watt a chip draws ends up as heat. A 72-GPU rack draws about 120 to 130 kW, over 10x an old server rack, so air can't cool it; liquid loops must.
  • Light is too slow for far-apart GPUsSignals in fibre travel about 200 km per millisecond. Training syncs GPUs many times a second, so one tightly coupled cluster must sit on one site or a few nearby ones.

How it works

4 parts
Fibre network gear in a Meta data centre: links like these tie thousands of GPUs across buildings into one training job
Fibre network gear in a Meta data centre: links like these tie thousands of GPUs across buildings into one training jobCourtesy of Meta · Press kit, editorial use (opens about.fb.com)
Tents

Build first, walls later

Prometheus in Ohio houses GPUs in weatherproof tent structures, so capacity comes online before conventional buildings would be done.

Power

Generation of their own

Prometheus adds on-site gas generation to its grid supply; Hyperion in Louisiana relies on new gas plants built by the utility Entergy.

Network

One job, many buildings

GPUs are joined by high-speed fabrics so one training job spans whole buildings, scaled up from Meta's 24,000-GPU clusters of 2024.

Finance

Shared ownership

Hyperion is owned by a joint venture 80% held by Blue Owl funds; Meta keeps 20%, builds and runs it, and leases the capacity.

Update log

6 updates

Mon 13 Jul

  • Minor: PressHardware

Tue 21 Oct 2025

  • Major: BlogFunding

Wed 20 Aug 2025

  • Minor: PressRegulatory

Mon 14 Jul 2025

  • Major: PressManufacturing

Wed 4 Dec 2024

  • Major: PressHardware

Tue 12 Mar 2024

  • Minor: BlogHardware

About Meta

The team behind Prometheus and Hyperion

Meta

Llama open models, Muse, AI data centers

Meta, owner of Facebook, Instagram and WhatsApp, made open-weight AI mainstream with its Llama models from 2023. In 2025 it formed Meta Superintelligence Labs. In April 2026 it released Muse Spark, its first closed frontier model.

  • Llama was downloaded over a billion times, but Llama 4 (April 2025) was seen as behind the leaders.
  • Paid about $14.3B for 49% of Scale AI and hired its CEO Alexandr Wang to lead its new AI lab (June 2025).
  • Muse Spark (April 2026) is closed, a shift away from open weights.

ai.meta.com (opens ai.meta.com)Company profile →

Founded
200422 yrs
Headquarters
United States
Status
Public
Listed
META (opens google.com)NASDAQ
Valuation
public
Staff
~74,067Dec 2024
Works in
AIOpen-weight modelsComputing
Coverage
3 programs · 19 updateslatest 17d agochecked 25 Sep
Partners
Scale AIBlue Owl CapitalMicrosoft
Suppliers
NVIDIACoreWeave
People
Mark ZuckerbergFounder, Chairman and CEOAlexandr WangChief AI Officer
Primer