A public, interactive 3D map of every known AI compute cluster on earth —
700+ named sites tracked by Epoch AI's GPU Clusters dataset
plus FLOP Map's curated additions, covering ≈65 GW of power and ≈80M H100-equivalent chips. Built to make the
infrastructure behind the AI boom visible.
How it works
Each dot is a cluster (points view), or rolled up by geography (heatmap / hex).
Click a dot for full specs: power, chips, FLOP/s, cost, sources. Filter by status,
operator, country, sector, or year. Pin up to 4 clusters to compare side-by-side.
Press ?
for keyboard shortcuts.
Why ~700, not 11,000?
Recent estimates put the world's total data centre count at
~11,000–11,800. FLOP Map is a much narrower scope:
named AI compute sites with public disclosure of operator,
location, and hardware. Most of the 11k+ are enterprise IT, telco, retail
back-end, edge, or hyperscale colocation that hosts no AI workload.
Of the few thousand sites that do run meaningful AI compute, only a few
hundred have detailed-enough public information to map. That's our set.
Data & caveats
Source: Epoch AI's GPU Clusters dataset
(CC BY) plus 34 community-curated additions maintained by FLOP Map.
253 anonymised Chinese clusters share one centroid (35.5°N, 101.9°E) —
heatmap reads this correctly but the raw dot count is misleading. Many
planned sites lack power or FLOP/s estimates. China is heavily
under-covered; smaller US hyperscaler footprints aren't enumerated
individually. This is the best public dataset we have.
Embed
Drop this into any page to embed a live FLOP Map (chrome-hidden, attribution kept):
Want the bigger picture?
Read The story behind the dots:
US chip export controls, Huawei vs Nvidia, sovereign AI, the energy crunch, and the scaling law.
Press Esc to close
The story behind the dots
Why this map looks the way it does · the geopolitics, chips, and energy beneath the AI buildout