Two numbers decide most AI infrastructure plans: what an hour of GPU time costs, and how much power-ready capacity exists to run it. Compute Census counts both, from published sources, with every figure dated and every source listed.
Operational, 271
Under construction, 249
Announced, 114
Arrange by
Each square is one of the 634 campuses in the census (records to 16 Oct 2026); MW totals count only campuses whose source publishes a figure. Point at a square for its name and capacity.
Series 1 · as of 16 Oct 2026
GPU rental price index
Floor, median and ceiling of on-demand prices, plus 12-month reserved rates, for 10 GPU models, with 73 provider price points behind them. An H100 costs $3.49 per GPU-hour at the median (+79% against 1 Oct 2025); a B200, $6.25.
Table AMedian on-demand price, USD per GPU-hour, 2026-10-16
634 AI-oriented data center campuses in 55 countries. Those under construction or announced add up to 79.9 GW of published capacity; 24 single campuses are rated at a gigawatt or more.
Table BPublished capacity by country, top five, MW
Country
Campuses
MW
United States
233
50,568
Australia
27
7,045
India
51
6,435
Indonesia
31
4,717
Malaysia
26
3,495
All statuses. MW summed only where the source publishes a figure (467 of 634 campuses).
As of 16 Oct 2026, renting an NVIDIA H100 80GB costs $3.49 per GPU-hour at the median on-demand rate, with published offers from $1.30 to $12.29.
As of 16 Oct 2026, the median on-demand price is $4.40 per GPU-hour for the H200 141GB, $6.25 for the B200 and $1.75 for the A100 80GB.
A published 12-month reserved H100 rate of $3.26 per GPU-hour is −7% against the on-demand median (16 Oct 2026).
As of 16 Oct 2026, Compute Census counts 634 AI data center campuses in 55 countries, 467 of them with a published capacity.
AI campuses under construction add up to 48.3 GW of published capacity and announced ones to 31.6 GW (records to 16 Oct 2026).
24 single AI campuses are rated at 1 GW or more, together 37.6 GW of published capacity (records to 16 Oct 2026).
§How the census works
Prices come from provider price pages and comparison trackers, normalised to one GPU-hour. Spot and preemptible tiers are recorded but kept out of the floor. History before August 2026 is seeded from published indices and marked as such on every chart.
Campuses come from a catalogue of 6,018 facility records: a site enters when it is flagged AI-ready or is a project of at least 50 MW under construction or announced. Nothing is estimated — an unpublished figure is shown as not published. Details on the methodology page.
Source GPUs or servers
Renting a block of capacity or buying hardware? Both are handled, through different partners. Tell us the model, quantity and location; we come back with pricing and lead times.