Power grids are being asked to carry data centers before the permits do
The future promised by AI is waiting behind the bottleneck of a grid built for another century.
A year ago, Google told the Louisiana Public Service Commission its planned data center would need “up to 5 gigawatts” of power—the equivalent of four million homes—but lacked a clear path to connect that load before 2034, Network World reported on August 4, 2026. A decade-long build was supposed to chase demand. Instead, the grid is under load now.
Markets are assigning contracts and workloads to facilities whose physical grid connections could take years longer than their construction. Every new permit is shadowed by a private negotiation with local utilities for electricity, yet the grid queues those projects for interconnection on a timeline that neither the data center nor the AI booms can wait for.
The real gap is that promised capacity is not delivered power.
Who carries the exposure? In July, New York imposed the first state-wide moratorium on new hyperscale data center construction, halting permitting and requiring every new build to secure its own energy or pay sharply higher prices, according to Network World on August 4, 2026. Data Center Watch reported at least 75 U.S. data center projects, worth $130 billion, were blocked or delayed in the first quarter of 2026 alone, a figure that matches the total for all of 2025, and the moratorium trend is spreading with similar measures appearing beyond New York.
Attrition is now a baseline risk. For every project that manages to survive five years in line, four drop out—most after being assigned an interconnection upgrade cost that annihilates its economics, PV Tech found in May 2026. Wood Mackenzie reports for failed projects in this queue, upgrade costs represented 30–37% of total budget, compared to 6–8% for those that connect (Spring 2026).
For every project that manages to survive five years in line, four drop out—usually after being assigned an upgrade cost that annihilates its economics.
How grid bottlenecks became structural
Electricity consumption in the United States was flat for thirty years. Data centers mostly built in places where the local utility had enough slack and the grid could be upgraded a few megawatts at a time. AI changed the math. The market can allocate the capital to construct these shells faster than any grid planner can clear the permits to deliver the power.
Flexential, a colocation provider with 42 data centers in 18 U.S. markets, says its capacity out to 2027 is already sold, often years before construction crews finish their work, Network World wrote on August 4, 2026.
The only way to secure a buildable slot is to already own the rights to the power. That queue is longer than the next commodity cycle.
Systemic costs and who carries them
The cost of this standoff is not theoretical. In the PJM market, the largest failure to connect cheaper generation cost consumers $7 billion in a single capacity auction when new resources with lower marginal costs could not clear the interconnection queue, Advanced Energy United reported in May 2026. Net benefits lost by users from delayed transmission now show up directly in utility bills and deferred digital investment.
Operators are adapting. FERC has forced cluster-based “first-ready, first-served” approaches instead of old serial queues. Some build behind-the-meter generation, adding turbines or batteries on-site to bypass approvals. Others create flexible interconnection agreements allowing partial or staged loading, trading certainty and scale for a modicum of speed. Federal entreaties now sit alongside utility rate filings and balance sheets, all to shore up a nationwide power bottleneck for zeros and ones.
The hidden tail here is not only tech. Every municipality, ratepayer, and manufacturer with grid reliance is now carrying a risk for the digital buildout that the grid as a system did not plan or price. Official risk models priced demand growth as incremental, but the reality was lumpy and winner-take-all. The grid’s failure to move as fast as its customers now shows up in wholesale prices, delays, and lost digital capacity. No one is paid to carry it.
A public decrease in queue wait times or attrition rates, if two out of three hyperscale projects announced this fall break ground and power up as scheduled, would confirm this read. The queue would have broken into a real market. A surge in on-time interconnection for hyperscalers, or a structural change that makes queue navigation routine, would break it.