How many data centers will there be? How big will the biggest AI data centers be? Where should they be located? And do recent announcements connote upside or downside risks? This 15-page report forecasts the build-out of AI data centers, using bottom-up data, statistical distributions and economic modeling.
This note continues our focus on the biggest debate in global energy markets: the size and scale of load growth associated with the build-out of AI data centers.
Why this matters, including controversies over global electricity demand, and US electricity demand, are re-capped on pages 2-3.
After triangulating between dozens of different sources, we have settled upon forecasts for the number of data centers, and the capacity of those data-centers, underpinning the total energy consumption of the internet, as explained on pages 4-5.
What really motivated this note was seeing announcements for some absolute monster-sized installations. Stargate is commissioning a 1.2GW monster at Abilene, and then targeting a 5GW monster in the UAE. There are several others in this range. So does this connote upside to our forecasts?
Interestingly, when we model the distribution of data center sizes, we would expect the distribution to be lognormal (page 6), and thus we would already expect about 20 x 1-5 GW-sized AI data centers to be operational by 2030 (pages 7-8), and if anything, we wonder if the major projects in progress today (pages 9-11) are insufficient, or too speculative, to de-risk our prior 2030 forecasts. This is fascinating statistical stuff.
Where should future data center announcements come from? We have assessed the relative economics of data centers in 20 countries and regions, finding most opportunity in the US and Middle East, but less opportunity in Europe and Asia, per pages 12-14.
Our conclusions into AI data center capacity growth — how many data centers? how big? where? what surprises?– are summarized on page 15.