Executive Summary

• Our utilities/power team has published many times on our strong view that the U.S. will significantly overbuild datacenter capacity towards the end of the decade and this report walks through our thought process, how we track both the supply/demand of datacenters, provides the tables for each and hopefully sets clients up to track the overbuild on their own. As we highlighted in our Ten Themes for 2026 report, we expect our Overbuild thesis to go from a fairly unique contrarian view now to more mainstream in 2026.

• The impetus for this report is to create a long-shelf life report to send to clients when we get the numerous clients emails asking us to “walk us through the overbuild thesis” and/or “how are you tracking the overbuild” every time we publish on it. It is really quite simple and maybe the reason a lot of folks aren’t doing it themselves is because it spans two, until recently, drastically different sectors: utilities/power on the supply side and technology on the demand side. At the end of the day, our overbuild view is just based on incredibly simple math.

• On the demand side, we use third party estimates from wall street, consultants, Bloomberg and other experts for 2030 and 2035 AI and datacenter (AI/DC) demand for electricity. We let them do the heavy lifting of forecasting chip volumes, end-use inference demand and hyperscaler training requirements. We then break these down by GW of capacity or TWhs of demand and then make them comparable by applying a capacity factor. We then subtract out 2024 AI/DC demand of 35 GW (GS says Dec ’25 was 45 GW so we might move to 2025YE soon if we see more ’25YE estimates). We are tracking TWh forecasts as well but this report is sticking with GW of capacity.