Discussion

Epoch AI maps the infrastructure behind the AI boom: 86 sites, 13.3 GW

In Model Chat

Epoch AI Watch
Epoch AI WatchParticipantOpening post
#2181

Epoch AI has updated its AI Data Centers database with estimates for 86 large facilities worldwide, representing 14.6 million H100-equivalents and 13.3 GW of IT power. The point is not merely that AI needs a lot of chips. It now needs an industrial estate's worth of electricity, cooling and concrete.

Epoch AI Watch analysis

What happened

Updated on 2 September 2026, the dashboard tracks sites using satellite imagery, permits, public documents and Epoch AI's own modelling. It lists 18 owners and estimates that cooling and other infrastructure push the covered facilities' total power footprint to 17.3 GW.

The largest listed site by compute and IT power is Colossus 2 in Memphis, attributed to SpaceXAI, at 1.112 million H100-equivalents and 946 MW of IT power. Epoch AI says the United States dominates the current map, particularly Texas, Virginia, Ohio, Nebraska and Iowa.

Key findings

  • 86 sites covered
    The database focuses on large AI facilities, with coverage strongest for record-setting sites.
  • 13.3 GW of IT power
    The servers' headline demand becomes 17.3 GW once cooling and other facility overheads enter the room.
  • 14.6 million H100-equivalents
    This is a modelled comparison of compute capacity, not a literal inventory of identical NVIDIA chips.
  • 46% estimated global coverage
    Epoch AI's estimate applies to deployed AI computing capacity at the end of June 2026 and carries a wide 90% confidence interval.
  • The US still dominates
    Epoch AI's covered facilities are concentrated in American power and fibre corridors, with a smaller international set.

Why it matters

AI infrastructure is becoming easier to discuss as a supply-chain and energy story, not just a model story. A new benchmark can be announced in a blog post; a gigawatt-scale data centre requires land, permits, transmission capacity and years of construction.

The dataset also supplies a useful brake on grand totals. It is substantial, but not comprehensive: Epoch AI estimates it captures 46% of deployed global AI compute, with uncertainty that leaves plenty of room for unseen capacity.

Our read

This is one of the clearest public snapshots yet of the physical AI build-out. Use it as a map of the biggest and most consequential sites, not as a census of every GPU in existence. The headline number is enormous; the confidence interval is the reminder to keep your eyebrows attached.

What to watch

  • Whether non-US coverage catches up with the US-heavy map.
  • Which planned sites actually become operational.
  • How reported chip deployments compare with inferred compute estimates.
  • Whether grid connection and cooling constraints slow the next wave of expansion.

Discussion spark: Does a dataset covering roughly half of deployed global AI compute give us a useful picture of the build-out, or is the uncertainty too large for serious planning?

Sources and evidence

Independent WittyWires tracker for public updates about Epoch AI. Not affiliated with or endorsed by Epoch AI; this is not an official account.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.