The infrastructure buildout powering the AI boom is colliding with a physical constraint that no software update can fix: the electrical grid. New projections from PJM Interconnection, the largest US grid operator covering 13 states and Washington D.C., show that data center load in its footprint is expected to grow by 40 gigawatts between 2025 and 2028 — roughly double what grid planners modeled just 18 months ago. The revised forecast has triggered emergency capacity procurement, accelerated transmission upgrade timelines, and, in several US markets, outright moratoriums on new large-load interconnection requests.
The Scale of the Demand Shock
The numbers are striking. The Lawrence Berkeley National Laboratory’s most recent data center energy report, released in March 2026, estimates that US data centers consumed approximately 176 terawatt-hours in 2025 — 4.4% of total US electricity generation. Under a high-growth AI scenario, the same report projects that figure reaching 12% of US electricity consumption by 2028, or roughly 500 terawatt-hours annually.
To put that in context: 500 TWh per year is approximately equivalent to the entire electricity consumption of France. It represents a tripling of data center energy demand in three years — a pace of growth that no grid infrastructure investment program, however well-funded, can fully match.
The concentration of this demand in specific geographies compounds the problem. Northern Virginia — home to the world’s densest data center market, accounting for roughly 70% of global internet traffic transiting at any given moment — has already hit physical limits in several substations. Dominion Energy, the primary utility serving the region, has disclosed a multi-year queue of over 90 gigawatts in data center interconnection requests, most of which cannot be accommodated under current grid topology.
Hyperscaler Responses
The four major hyperscalers — Microsoft, Alphabet, Amazon, and Meta — have all disclosed aggressive power procurement strategies in recent earnings calls, but the approaches diverge significantly.
Microsoft, which spent $22.6 billion on capex in its most recent quarter (predominantly AI infrastructure), has signed power purchase agreements for approximately 10.5 gigawatts of new renewable energy capacity, according to company disclosures and Bloomberg NEF data. It has also announced plans to restart the Three Mile Island Unit 1 nuclear plant in Pennsylvania — a 835-megawatt facility — under a 20-year power purchase agreement with Constellation Energy. The plant is scheduled to return to service in Q4 2026.
Alphabet, fresh off its Q1 2026 earnings beat, has committed $17.2 billion in quarterly capex and is pursuing both wind and solar PPAs alongside geothermal exploration contracts in Nevada and Utah. The company has internally targeted 24/7 carbon-free energy matching by 2030 — a significantly more stringent commitment than Renewable Energy Certificate purchasing.
Amazon Web Services is the largest corporate buyer of renewable energy globally, with over 43 gigawatts of contracted capacity worldwide, but analysts at Wood Mackenzie note that contracted capacity and delivered electrons are different things: transmission constraints mean much of that contracted power cannot physically reach the data centers it is intended to serve during peak demand hours.
The Grid Infrastructure Gap
The fundamental problem is a mismatch in investment timescales. A large hyperscale data center can be designed, permitted, and constructed in 18 to 36 months in favorable jurisdictions. A new 500-kilovolt transmission line — the kind needed to move bulk power across regions — takes 8 to 12 years from planning to energization in the United States, primarily due to siting and permitting timelines.
The Federal Energy Regulatory Commission’s Order 1920, finalized in 2024, mandated long-term transmission planning on a 20-year horizon, but critics argue the rule lacks enforcement teeth and relies on voluntary regional coordination that has historically moved slowly. The Biden-era Transmission Facilitation Program deployed roughly $2.5 billion in federal loans for transmission projects, but that figure is an order of magnitude below what LBNL and the American Council on Renewable Energy estimate is needed by 2030.
“We are building the AI economy on a grid that was designed for a different era,” said one grid planning official at a major US regional transmission organization, speaking on background. “The demand growth we are seeing right now was not in any baseline forecast we published two years ago.”
Implications for the AI Infrastructure Build
For investors and technology companies, the power constraint is beginning to surface as a material risk factor in ways that financial models have not fully priced. Data center REIT stocks — including Equinix, Digital Realty, and Iron Mountain — have all cited power availability as the primary constraint on expansion in core markets in recent investor communications.
Hyperscalers with the deepest pockets are increasingly pursuing what industry analysts call ‘power-led’ site selection: identifying locations with stranded power capacity — often near retired fossil fuel plants with existing transmission infrastructure — before identifying real estate, rather than the reverse. This dynamic is reshaping data center geography, pulling new construction toward the Midwest, the Gulf Coast, and rural Texas, away from the traditional coastal clusters.
The near-term implication is that AI infrastructure spending will continue to accelerate regardless of grid constraints — but the marginal cost of power procurement, and the risk of stranded assets in undersupplied markets, will increasingly differentiate winners from those who over-built in the wrong places.
Discussion
Sign in to join the discussion.
No comments yet. Be the first to share your thoughts.