× The Build-out
Eighty-six frontier AI data centres, each one a band in the stack, stacked to
the world total. The chart runs from January 2023 to January 2030 and can be
read three ways: the compute installed, the electricity it draws, and the
capital sunk into the buildings and the silicon inside them.
How to read it One band, one site. Bands are ordered smallest eventual
build at the base, largest at the crest, and coloured on a fixed scale from
nothing to six million H100-equivalents. A site keeps its colour whatever
you filter or switch to, so the colour becomes an identity rather than a
rank that shifts underneath you.
The dashed line is today. Everything left of it has been
verified — satellite passes, utility filings, permits, company disclosures.
Everything right of it is projection, and is drawn dimmed and hatched for
exactly that reason.
Hover any band for that site at that month; click to pin it.
The scrubber and the space bar move time; shift-arrow jumps a year. A pinned
or hovered site shows a white dot at every date it was actually observed.
Contract and pan the years. The two-dot control above the
transport is the window: drag a dot to pull an edge in, drag the bar between
them to slide the window along, NOW to jump to the present, Reset to go back to
the full span. On the field itself, scroll or pinch to zoom around the pointer,
two-finger swipe or drag to pan, double-click or press 0 to reset. The vertical
scale follows the window, so zooming into 2023 shows 2023 rather than a flat
line at the foot of a 2030-sized axis.
Click the legend to change the colour scale — Turbo, then ten
perceptually uniform ramps from matplotlib and Fabio Crameri's Scientific
Colour Maps, then Distinct, which drops the quantity altogether and gives every
site its own hue so neighbouring bands separate. The legend always says which
of the two you are looking at.
Smoothed or Linear. Smoothed draws a monotone curve through
the monthly values — the shape of the thing. Linear draws the straight
segments the data actually describes, and adds a rug of ticks along the
bottom: one for every date any site on screen was looked at. Where the ticks
crowd, the record is dense; where they thin, the line between them is
inference.
Where the numbers come from
Every figure is Epoch AI's. Their researchers identify candidate sites from
company announcements, permits and utility interconnection queues, then
confirm construction from commercial and public satellite imagery, and
estimate capacity from building footprints, cooling equipment, substation
capacity and disclosed chip counts. Compute is expressed in
“H100-equivalents”: the number of Nvidia H100 chips that would deliver
the same 8-bit throughput, which lets a TPU campus and a GB200 campus be put
on the same axis.
What this chart does not know The plateau after 2029 is the edge of the dataset , not a
forecast. These are the sites someone has verified; announcements made after
the last update are still fields.
Capacity between two observations is interpolated. A
satellite pass in June and another in November do not say what happened in
August. We draw a straight ramp between them — and we never interpolate
across today, so the value shown for the present month is the last verified
reading, not an average of the past and the forecast.
China is under-represented. Three of the eighty-six sites
are Chinese. That is a limit of what can be verified from open sources, not
a claim about where compute is being built.
Owners and users carry confidence markers in Epoch's data —
confident, likely, speculative — and the card shows them. Costs are
estimates in 2025 dollars.
On the satellite imagery
Epoch's own site photographs are licensed commercially from Vantor, Airbus and
Vexcel — their map view tags them “© Airbus DS 2026” — and their
open data licence covers their measurements, not those images. We do not
reproduce them. Every overhead view on this page is ours, cut from the European
Union's freely licensed Copernicus Sentinel-2 mission at 10 metres per pixel,
choosing the most recent near-cloudless pass over each site. Coarser than a
commercial frame, and enough to read the halls, the substation and the raw
ground where the next building is going.
Contains modified Copernicus Sentinel data 2026.
Colour
Bands are coloured by a site's compute when complete, on a fixed scale from
nothing to six million H100-equivalents, so a site keeps its colour across
filters, metrics and time. The scales on offer are Turbo (Mikhailov, Google AI),
matplotlib's viridis, magma, plasma and cividis, and six of Fabio Crameri's
Scientific
Colour Maps v8 — batlow, lipari, navia, hawaii, la jolla and acton — which are
perceptually uniform and readable with colour-vision deficiency. Each is sampled
inside the stretch of itself that clears 3:1 against this page's black; the dark
ends of these ramps are the same colour as the ground, and a band painted in one
is not a dark band, it is a hole in the chart.
Source and licence
Epoch AI, Data on AI Data Centers . Published online at epoch.ai.
Retrieved from epoch.ai/data/ai-data-centers ,
licensed under CC BY 4.0 .
GLOBAÏA (2026). The Build-out: frontier AI data centres, 2023–2030
[interactive visualization]. globaia.org/ai-chain/buildout/. Data: Epoch AI,
Frontier Data Centers (CC BY 4.0).