The AI infrastructure boom has a bottleneck — and it isn’t silicon. Of the approximately 140 large-scale data center projects representing 12 gigawatts of power planned to go live in the United States in 2026, only one-third are currently under construction. The other two-thirds are delayed or cancelled outright. Not because of a shortage of capital or demand, but because the electrical grid cannot support them.
$1 Trillion in Committed Spend, Nowhere to Plug In
The numbers heading into 2026 were staggering. Alphabet, Amazon, Meta, and Microsoft collectively committed to more than $650 billion in AI infrastructure investment this year alone, part of what analysts at Morgan Stanley and Bismarck Analysis estimate will reach $1 trillion in total tech sector capital expenditure for 2025–2026 combined. The IEA now projects that data centers will consume 1,000 TWh globally in 2026 — equivalent to Japan’s entire annual electricity output.
The problem is that demand has outrun the grid’s ability to respond. Securing grid interconnection for a new hyperscale campus takes three to seven years under current US utility processes. Transmission infrastructure buildout runs on similar timelines. Companies that committed to new campuses 18 to 24 months ago assumed grid capacity would be available. In many markets, it isn’t.
Nuclear, Gas, and On-Site Generation as Emergency Fixes
The hyperscalers aren’t waiting for grid modernization. Facing hard power constraints, they are increasingly building around the problem.
Cleanview’s February 2026 infrastructure report projects that 30% of anticipated US data center energy capacity will come from on-site generation sources by end of 2026, up from effectively zero just 12 months ago. Forecasts suggest that figure could reach 50% as hyperscalers lock in direct generation partnerships — bypassing the grid entirely rather than waiting for it to catch up.
Nuclear is emerging as the preferred long-run solution. GE Vernova and Japan’s Hitachi announced a combined $40 billion investment in small modular reactors (SMRs) earmarked for Tennessee and Alabama — sites chosen in part because of proximity to existing data center corridors. Microsoft, Google, and Amazon have all signed direct power purchase agreements with nuclear operators or SMR developers in the past 12 months.
Natural gas, meanwhile, is serving as the bridge fuel of record. Utilities in Virginia, Texas, and Georgia — the three largest US data center markets — have all filed requests with state regulators to accelerate gas generation buildout specifically to serve data center load.
The Strategic Consequence
The power constraint is doing something that the capital markets have not: it is forcing geographic differentiation in the AI buildout. Regions with surplus generation capacity, favorable interconnection timelines, or proximity to water cooling are commanding significant infrastructure premiums.
Northern Virginia, long the undisputed center of US hyperscale infrastructure, is now losing deals to markets in the Midwest, the Pacific Northwest, and internationally to the Nordic countries and the Middle East — where energy is cheaper, more available, or both.
For the broader AI race, the constraint is an accelerant for two related trends: hardware efficiency (reducing watts-per-FLOP at the chip level) and distributed inference (moving workloads to where the power is, rather than where the talent is). Nvidia’s latest GPU roadmap and the emerging category of power-optimized inference chips from startups like Cerebras, Groq, and the newly-funded Nuvacore all reflect this pressure.
The build-or-wait calculation for the next generation of AI infrastructure increasingly runs through a utility approval process. That is a slow, unglamorous constraint on an industry that has moved fast. It is also one that money alone cannot solve quickly.
Data sourced from Morgan Stanley’s 2026 AI Infrastructure Outlook, Cleanview’s February 2026 infrastructure report, IEA Global Energy Review 2026, and Bismarck Analysis.
