The AI boom has spent years being framed as a race for chips. UBS thinks the bottleneck has moved somewhere far less glamorous: the power grid.
In new research, the bank argues that infrastructure constraints, specifically power supply, grid capacity and data center space, are now driving the next phase of the AI trade.
The speed limit on a super-exponential
The report carries the title “Signal over Noise #14: The AI ‘super-exponential’ meets its speed limits.” Its central point is simple. AI demand is growing fast, and the physical world is not keeping up.
According to UBS, AI data centers need a reliable baseload power supply on the scale of a mid-sized city. Baseload means steady, around-the-clock electricity, not power that shows up when the wind cooperates.
Building that supply is slow. UBS points to regulatory hurdles, interconnection problems, supply-chain bottlenecks and community opposition as the main drags on new power infrastructure.
The timeline mismatch is stark. UBS notes a data center can go up in 12 to 24 months, while adding power capacity can take 5 to 10 years.
Advertisement
The queue to connect new projects to the grid tells the same story. US interconnection queues currently sit at around 2.3 TW, roughly double the country’s installed capacity as of 2024. The median time from request to commercial operation is now about 6 years.
The numbers behind the crunch
UBS tracks 276 GW of global data center capacity that is either under construction or planned.
The near-term picture is tighter. UBS sees 57 GW of capacity needed in 2027, against just 38 GW of sites with confirmed power.
The bill for closing the gap is steep. UBS estimates $511 billion will be needed for new US power-generation capacity by 2030 to support a 3% compound annual growth rate in electricity demand.
Equipment is another choke point. Large transformers may face lead times of up to 210 weeks.
Policy is adding friction too. New York has enacted a one-year moratorium on new data centers exceeding 20 MW. Other states are requiring developers to pay for grid upgrades themselves, shifting costs from utilities and ratepayers onto the builders.
Where UBS sees opportunity
On September 23, 2026, the bank initiated coverage on several colocation data center stocks with a bullish stance. UBS cited the currently constrained supply of compute against rising demand, along with strong power development pipelines at these firms. Among its picks, it started Core Scientific (CORZ) and Applied Digital (APLD) at Buy.
The bank projects around 80% annual EBITDA growth for these colocation names from 2027 to 2029.
The longer-term view is more reassuring than the near-term one. UBS projects there will be adequate power, approximately 415 GW, to meet AI demand by 2030, provided the initial projects come online as scheduled.
What this means
A pipeline of 276 GW looks impressive until you compare the 57 GW needed in 2027 with the 38 GW actually powered. Planned capacity and usable capacity are not the same thing.
The practical watch list follows directly from the report: regional regulatory moves, interconnection queue trends, equipment lead times and whether power projects hit their in-service dates.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.