The Real AI Race Is Being Decided in Power Substations, Not Model Cards
Up to 11 gigawatts of AI data center capacity announced for 2026 has not moved into construction — and the window to change that is closing faster than the grid can handle it.
That is the implication of a June decision pending before the Federal Energy Regulatory Commission, which will rule on whether the rules governing how power plants connect to the grid can be rewritten fast enough to let the AI industry's electricity demands off a queue that currently takes four to ten years to clear in some regions Data Center Knowledge. The Department of Energy, which pressed FERC for an April 30 answer, will now have to wait until June Utility Dive. The commission's chairman has called the situation dire. One commissioner suggested utilities might exit the regional grid operator altogether if the crisis is not resolved.
The numbers underneath are not ambiguous. Morgan Stanley projects a 126-gigawatt surge in data center demand through 2028 Tech Insider. Global data center electricity consumption was approximately 415 terawatt-hours in 2024; by 2026, that figure is projected to approach 1,000 TWh — roughly equal to Japan's total electricity consumption, ranking fifth among nations Brookings. A single AI task can consume up to 1,000 times more electricity than a traditional web search Brookings, and AI-optimized GPU chips draw two to four times the watts of conventional processors Brookings. The bottleneck is structural: new high-voltage transmission lines can take more than a decade to permit and build Data Center Knowledge.
The pressure is already reshaping competitive positions. AEP Ohio has paused all new data center interconnections due to insufficient power infrastructure Tech Insider. States including Ohio are advancing legislation requiring data centers to pay for their own grid upgrades rather than shifting costs to residential ratepayers Tech Insider. Utilities are making choices about who gets power and who waits — and those choices, not model benchmark scores, may determine which AI companies survive the next two years.
Some companies are not waiting for the queue. Microsoft is restarting the Three Mile Island nuclear plant to supply its AI facilities. Google acquired Intersect Power for direct generation capacity. Oracle signed a deal in April to purchase up to 2.8 gigawatts of electricity from Bloom Energy, with 1.2 gigawatts already contracted — bypassing the interconnection queue through a direct procurement relationship with a generator rather than a request to the grid CNBC. These are infrastructure plays, not environmental statements.
The DOE announced last week a $4.2 billion expansion of high-voltage transmission lines in southern Ohio — one of the most congested corridors in the eastern grid — to relieve the kind of bottlenecks that caused AEP Ohio to pause new hookups energy.gov. Whether that building effort moves fast enough to matter for AI's timeline is the central question.
If FERC's reforms accelerate interconnection timelines, new entrants get a path to power. If the rules remain tangled in federal-state jurisdictional fights, the companies that already locked up generation relationships will have a structural advantage that no model improvement can close. The AI race has always been measured in parameters, FLOPs, and benchmark scores. The next race will be measured in megawatts.
The binding constraint is no longer a transistor. It is an electron.