Epoch AI’s data center map now covers 44% of global AI compute

Epoch AI has expanded its AI Data Centers explorer from 13 facilities to 86, raising its estimated coverage from 15% to 44% of global AI compute. The research group also published its methodology, allowing researchers to inspect the assumptions and reproduce the coverage calculation.

AI chip sales reveal who bought accelerators, but they rarely show where the hardware was installed, when it became operational, or how much electricity it can consume. Facility-level data helps developers, infrastructure planners, and researchers connect chip shipments with physical capacity, regional availability, and grid demand.

From 13 facilities to 86

When the AI Data Centers explorer launched in November, it covered an estimated 15% of global AI compute. Its 86 current records include satellite imagery, construction status, chip types, estimated capital cost, IT power, and compute capacity expressed in H100-equivalents, a normalized measure that compares different accelerators in units of Nvidia H100 capacity.

Coverage is higher for recent deployment cohorts. Epoch estimates that the explorer captures 53% of compute deployed in 2026, 39% from 2025, and 26% from 2024.

Inside each facility record

The landing page provides a timeline slider and filters for owner, primary user, and country. Individual directory records combine hardware estimates with construction evidence and site-specific infrastructure details.

The record for Colossus 2, xAI’s Memphis facility, lists Nvidia B200 and B300 accelerators, cooling equipment, a buildout timeline, and satellite timelapses. Land clearing began in February 2025. About 18 months later, the site had infrastructure capable of supporting roughly 950 MW of IT power, the electricity used by servers, networking equipment, and storage.

The expanded catalog also includes regional facilities such as the 41 MW Southgate Melbourne site in Australia and the 72 MW Oracle Batam site in Indonesia. These records provide visibility into inference capacity deployed near regional users as well as large training campuses.

How Epoch calculates 44%

Epoch derives the coverage estimate by adding the H100-equivalent capacity of operational facilities in the explorer and dividing that figure by an independent estimate of AI compute delivered to major customers. The calculation assumes a three-month delay between chip delivery and deployment.

The denominator comes from Epoch’s AI Chip Owners dataset. Estimated delivered capacity reached 20.3 million H100-equivalents at the end of 2025 and was extrapolated to 31.6 million by September 2026. Epoch has published the full calculation notebook for inspection and reuse.

The gaps vary by operator

Owner-level estimates show where facility attribution is strongest and where the explorer remains incomplete:

Public-cloud infrastructure is difficult to assign because one facility can serve several customers, leaving Microsoft, Amazon, and Oracle with lower attributable coverage. Meta operates more captive infrastructure, which makes its deployments easier to map. China remains the largest geographic gap and a stated priority for further research.

What developers can do with it

The facility records and downloadable methodology support several forms of technical and market analysis:

1. Cross-check disclosures: Compare reported capital expenditure and deployed megawatts with facility records and satellite imagery.
2. Estimate regional capacity: Aggregate compute by country, operator, or grid region to study supply concentration and energy demand.
3. Measure buildout speed: Track the interval between land clearing, construction, infrastructure installation, and estimated operation.
4. Compare hardware choices: Examine site-level deployments of Nvidia H100, B200, B300, and custom accelerators.
5. Model inference availability: Use facility locations as one input when estimating regional capacity and latency, alongside network routing and customer access.

The dataset is available under a Creative Commons Attribution license. A changelog records edits to facility entries, providing an audit trail for citations in research and internal analysis.

The explorer remains an estimated map rather than a complete inventory. Shared cloud infrastructure, uncertain deployment dates, limited Chinese data, and the three-month deployment assumption can all affect its totals. H100-equivalent capacity also normalizes hardware performance without capturing every operational factor, including utilization, networking, memory constraints, and workload mix.