The numbers come first. Data centers are projected to consume nearly 12% of all U.S. electricity by 2030, according to the Lawrence Berkeley National Laboratory, a federally funded research center — nearly six times their pre-AI-boom share in 2018.
The demand curve is steepening faster than forecasters expected. NERC, the grid reliability watchdog, projects summer peak electricity demand in North America will grow by more than 224 GW over the next decade, a figure 69% above what the group projected just a year earlier, driven largely by new AI data centers. In the Western U.S. grid region, planned data centers already average 10% of demand forecasts and run as high as 40% in some areas.
The mismatch is structural, not incidental. 'There is a real disconnect there,' said Rob Gramlich, president of consulting firm Grid Strategies, noting that tech companies move fast while utilities 'notoriously move very slowly' because interconnecting thousands of grid components requires years of 'deliberate study and planning.'
Part of the problem predates AI entirely. Gramlich said the U.S. electricity industry 'got out of practice on building new infrastructure' after a 25-year stretch, from 2000 to 2023, of comparatively flat demand growth following the buildouts of the 1980s and 1990s. Now electric vehicles, electric heating and new manufacturing are colliding with the data-center boom simultaneously.
Kathryn Burke, who leads U.S. specialty energy and power growth at insurance firm Marsh, told Fortune that access to power is 'probably the number one, if not top five bottlenecks for data center development in the U.S. right now.' Companies are requesting delivery in roughly two years; Burke said a single gigawatt of capacity — enough to power 750,000 U.S. homes — takes far longer than that to generate. Separate Berkeley Lab research found U.S. power projects that came online in 2025 spent a median of five years from requesting a grid connection to operating commercially.
The Department of Energy said in July there is a 'pressing need' for more transmission and generation infrastructure 'due to load growth from data centers,' among other factors. Burke predicts 50% to 60% of data center projects will be delayed and will not be operational within the one-to-two-year window companies are hoping for. Utilities, wary of being left with stranded assets, are now asking developers for more money upfront as a hedge against projects that fall through.
Gramlich said the likely outcome is not widespread blackouts, since utilities generally decline to connect customers they cannot reliably serve. Instead, data centers may wait years for full service or accept provisional connections that can be interrupted when the grid is strained. 'There is scarcity on the grid,' he said. 'Not everybody's going to get the full level of service that they want.'
The capital is not the constraint here — OpenAI, Google and Meta have already proven they can finance and build faster than almost anyone. What cannot keep pace is a permitting and utility-planning apparatus built for a 25-year era of flat demand. Communities and lawmakers across party lines have raised concerns about higher electric bills tied to this buildout, but the deeper story is a regulated monopoly system that was never designed to move at the speed of private enterprise.
Power leaves a paper trail, and in this case the trail runs through years of study, review and utility rate cases rather than through Silicon Valley balance sheets. Until permitting timelines shrink closer to construction timelines, the AI economy will be waiting on Washington and state utility commissions, not on Wall Street.
