Prometheus by Meta: status and timeline

Prometheus

Bring one of the first gigawatt-scale AI training clusters online fast, in Ohio, to train Meta's next models. Speed of building is a competitive edge when compute decides who leads.

ScalingStage 5 of 5

Operational in 2026 with 500+ MW of AI compute; scaling to 1 GW by Q4 2026 in New Albany, Ohio.

Updated 14 Jul 2025Checked 11 Oct0 updates this week

Milestones

Next · Reaches 1 GW capacity

Prometheus announced2025Complete.
First 500+ MW operational2026Complete.
Reaches 1 GW capacityTarget Q4 2026Current milestone.

Most important updates

Zuckerberg announces Prometheus and Hyperion gigawatt-scale AI clusters14 Jul 2025

Upcoming

Q4 2026Reaches 1 GW capacity (next)
Q4 20261 GW operational milestone

Current obstacles

Power supply and grid connectionGrid connections for a 1 GW facility take years of planning and negotiation; Meta is investing in alternative power sources.
Liquid cooling at scaleCooling thousands of high-power GPU racks with custom liquid systems is an engineering challenge that requires continuous refinement.
Supply chain for componentsSourcing chips, memory, power supplies and networking equipment for a gigawatt cluster requires long-lead agreements with multiple vendors.

Physics limits

Energy density pushes physics limitsTent-style buildings without massive industrial infrastructure are an experiment; extreme power densities (tens of MW per building) stretch cooling and power distribution.
Utilization riskA 1 GW cluster is expensive to build and operate; if utilization drops or training demand slows, return on investment suffers.

How it works

4 parts

Location

Central Ohio for power and talent

New Albany, Ohio offers reliable power grids, nearby universities and access to skilled talent in the Great Lakes region.

Cooling

Tent-style architecture

Prometheus uses unconventional tent-like structures to manage extreme power density and dissipate the heat of thousands of GPUs running simultaneously.

Power

Multi-source energy supply

Meta has signed agreements with nuclear providers (TerraPower, Oklo) and Vistra for natural gas and renewables to reach 6.6 GW of capacity by 2035.

Purpose

Superintelligence training

Prometheus is built to run the longest, most compute-intensive training runs Meta plans to undertake, advancing frontier models and reasoning.

Spec sheet

Spec	Prometheus
Target capacity	1 GW^R (reported)
Operational capacity (as of 2026)	500+ MW^R (reported)
Data center buildings	9^R (reported)
Total building area	1,000,000 sq ft^R (reported)
H100-equivalent chips (projected)	1,031,000^R (reported)
Estimated capital cost	$32.4B^E (estimated)

^R reported by the company · ^E our estimate

Update log

1 update

Mon 14 Jul 2025

Major: Zuckerberg announces Prometheus and Hyperion gigawatt-scale AI clustersPressManufacturing

About Meta

The team behind Prometheus