The Biggest AI Data Centers & GPU Clusters in the World

Executive summary

AI's physical layer is a race of megawatts and silicon. This report ranks the world's largest AI data centers by power, the largest GPU clusters by compute (in H100-equivalents), and the chipmakers that supply them — all from Epoch AI's open infrastructure datasets. Colossus 2 leads the data centers at 946 MW; xAI Colossus Memphis Phase 3 is the largest single cluster; and Nvidia, with ~18 million H100-equivalents shipped, controls roughly 70% of all AI compute — more than every competitor combined. The buildout is overwhelmingly American. Every number carries its source; the full data is downloadable under CC BY 4.0.

Key findings

The biggest AI data centers, by power

The largest AI data centers in the world ranked by electrical power (MW) — the binding constraint on AI compute. Colossus 2 leads; the full table adds capital cost and compute.

The biggest GPU clusters, by compute

The largest single GPU clusters ranked by compute in H100-equivalents — the metric that actually trains frontier models. xAI Colossus Memphis Phase 3 tops the list.

Who makes the world's AI chips

Cumulative AI compute shipped, by chip designer, in H100-equivalents. Nvidia's dominance is near-total — about 70% of all AI compute — with Google's in-house TPUs the main challenger and a long tail of custom silicon (Amazon, Huawei, AMD).

What the data reveals

AI compute is concentrated at every layer of the stack. Geographically, 17 of the top 20 data centers are American. By vendor, Nvidia holds ~70% of all compute shipped. And by operator, a handful of hyperscalers and one well-funded newcomer (xAI, whose Colossus clusters dominate the cluster ranking) own most of the frontier capacity.

Power is the real constraint. Colossus 2 alone draws 946 MW; the top few sites together rival the electricity demand of a small country. The AI race is, increasingly, an electricity race — which is why data-center siting now follows cheap, available power.

The single biggest structural risk sits upstream: nearly all of these accelerators are fabricated by one company, TSMC. Compute is concentrated among designers, but manufacturing is concentrated in one foundry — the tightest chokepoint in the entire AI supply chain.

Limitations & notes

These are estimates, not audited disclosures. Data-center power and compute are inferred from satellite imagery, construction permits and company statements; chip-sales figures are modelled from financial reports. Epoch publishes confidence ranges (5th–95th percentile) that we summarise to a central estimate here. Undisclosed or sovereign facilities may be missing, and figures for planned/under-construction sites reflect current, not final, capacity.

H100-equivalents normalise heterogeneous hardware to one scale for comparison; they are an approximation, not a physical count of chips. The full dataset, with ranges and sources per facility, is downloadable and reproducible below.

Scoreboard (machine-readable data)

Every headline indicator with its value, period, source and confidence. Free to reuse under CC BY 4.0.

↓ CSV · ↓ JSON

Methodology & verification

Data centers and GPU clusters from Epoch AI's infrastructure datasets (CC BY 4.0): power (MW), compute in H100-equivalents, capital cost and owner, estimated from satellite imagery, permits and disclosures. Chip-sales by designer from Epoch's AI Chip Sales dataset, modelled from financial reports, cumulative through the latest quarter. H100-equivalents normalise different accelerators to the compute of an NVIDIA H100. Central estimates shown; Epoch publishes 5th–95th percentile ranges. Nothing is estimated by Affärslivet. Confidence: Medium-High.

Data dictionary

Frequently asked questions

Glossary

H100-equivalent (H100e)

A unit normalising different AI chips to the compute of an NVIDIA H100 GPU, so heterogeneous clusters can be compared on one scale. ↗

Data center power (MW)

The electrical power a data center can draw, in megawatts — the practical ceiling on how much AI compute it can run. ↗

AI accelerator

A specialised chip (GPU, TPU or custom ASIC) built to run the matrix maths behind AI training and inference far faster than a general CPU. ↗

Foundry

A semiconductor factory that manufactures chips designed by others; TSMC fabricates the overwhelming majority of the world's AI accelerators. ↗

Sources

• Epoch AI — AI Data Centers, GPU Clusters, AI Chip Sales (CC BY 4.0)
• IEA — Energy and AI (2025)
• Stanford HAI — AI Index Report 2026

Related reports & data

• National AI Power Index
• The Largest AI Companies
• Countries Leading in AI