Atlas
Where the machines are
Every AI answer is produced in a building. This map shows the buildings: the largest AI sites tracked by Epoch AI, the thousands of ordinary data centers volunteers have mapped in OpenStreetMap, and the electricity grid each one plugs into. Click anything.
—Frontier AI sites
—GW of IT power
—Data centers in OSM
—Frontier sites in the US
Layers
Grid carbon intensity
0 g300600900+ gCO₂e / kWh
Country averages, latest year available (mostly 2024 to 2025). A site may buy or generate power that is cleaner or dirtier than its country's average.
Largest sites by IT power
IT power is what the chips draw. Whole-facility power runs about 1.3× higher once cooling and losses are counted.
Jump to:
Activity
Your prompt's trip
Follow your own AI use to a real building, then to the grid that powers it. The numbers are estimates built from public figures; the point is to see what each number depends on.
What that adds up to
—Wh per day
—kWh per year
—g CO₂e per day, this grid
—kg CO₂e per year, this grid
Same prompts, different grid
Energy per prompt uses Epoch AI's 0.3 Wh estimate for a typical ChatGPT text query and 2.9 Wh per image from Luccioni et al. Grid factors are national averages from Ember and the Energy Institute via Our World in Data. None of this counts training the model, building the chips, or the water used for cooling.
Reflect
Export your work
Everything you typed is saved in this browser only. Nothing is sent anywhere.
About
How this atlas was made
A companion to The Digital Footprint. That activity asks how much energy a prompt costs. This one asks where that energy is drawn, by whom, and from what grid.
Data on the map
- Frontier AI sites: Epoch AI, AI data centers (CC BY 4.0), retrieved 19 Oct 2026. Epoch publishes street addresses, not coordinates; sites were geocoded with OpenStreetMap Nominatim. Each site's inspector states whether its pin is an address, a locality, or a place-level guess.
- All mapped data centers: OpenStreetMap contributors (ODbL), features tagged telecom=data_center, Overpass snapshot of 24 Jul 2026. Volunteer coverage is uneven: dense where mappers are active, thin elsewhere.
- Grid carbon intensity: Ember (2026) and Energy Institute (2025) via Our World in Data (CC BY), grams CO₂e per kWh generated, latest year per country.
- Headline figures: IEA, Energy and AI (2025); per-prompt energy from Epoch AI and Luccioni et al.
- Basemap: Natural Earth 1:50m (public domain), cut into vector tiles and packed as a single PMTiles file served from this folder. No tile server, no API key.
- Terrain: Mapzen Terrain Tiles (Terrarium encoding; SRTM, GMTED2010, ETOPO1 and others), zooms 0 to 4 worldwide and 5 to 6 over the regions the lessons visit, packed as PMTiles.
- Labels: Noto Sans glyphs (SIL OFL) served from this folder.
- Software: MapLibre GL JS and PMTiles, both BSD-licensed, vendored locally.
What the map does not know
- Epoch estimates its hub covers about 44% of deployed AI compute (90% interval 30% to 79%). More than half of the world's AI hardware is in buildings not shown here.
- Owner and user tags carry Epoch's confidence labels: confident, likely, speculative. The inspector shows them.
- Power figures are IT power. Facility power, with cooling and losses, is roughly 1.3× higher.
- Grid intensity is a national yearly average. Many sites sign their own power contracts or run on-site generation.
- Country shading uses Natural Earth boundaries, which are a cartographic convenience, not a political statement.
Credits
Created by Marc Watkins. Free to use, copy, and adapt for teaching under CC BY 4.0.
⚙ AI disclosure: this atlas was designed and coded with Claude (Anthropic) working under Marc Watkins's direction. The data pipeline, geocoding, map style, lesson text and calculator were drafted by the AI and reviewed by a person. Figures are rounded for teaching and were current as of October 2026. Verify before citing in formal work.
© 2026 Marc Watkins · The Data Center Atlas · Made for the classroom.