Why America’s Grid May Become the Biggest Constraint on AI Growth

Photo: Ziko, Wikimedia Commons, CC BY-SA 4.0

01 The constraint that outranks chips

Ask what limits AI growth in 2026 and the candid answer from utilities, hyperscalers and regulators is converging on one word: wires. Not compute, not capital, not cooling. The U.S. interconnection system — the queue through which new generation, storage and large loads must pass — ended 2025 with 2,061 gigawatts of active requests: 1,312 GW of generation and 749 GW of storage. That is more than the roughly 1,280 gigawatts of installed generation capacity in the entire country. The pipe applicants must wait in is now bigger than the pipe.

The demand side is not waiting politely. NERC's January 2026 Long-Term Reliability Assessment revised its ten-year peak demand forecast upward by 24 percent, projecting summer peak to climb from 127 gigawatts in 2026 to 143.7 gigawatts by 2035, and flagged MISO, PJM and ERCOT as the highest-risk regions — precisely where hyperscalers have been buying land and pouring shells.

Analysis — not prediction. N43 and Hermes AI grounds every scenario in the documented record and verified reporting as of September 19, 2026; where evidence is incomplete we say so.

02 The queue: structural scarcity, not a backlog

The comforting reading of these numbers is administrative: a backlog that better software will clear. The record does not support the comforting reading. Median time from queue entry to operation ran 55 months for 2024 completions, and PJM projects that entered service in 2025 had averaged about eight years in queue. ERCOT's large-load queue exploded to 410 gigawatts by April 2026 — 87 percent of it data centers — up from 233 GW at the end of 2025 and 63 GW at the end of 2024, nearly five times ERCOT's own peak demand. Cluster-study reforms cut the total active queue 12 percent in a year, but 700 gigawatts of projects withdrew in the process — much of it capacity the queue had been double-counting all along, and much of it real projects that gave up.

The result is a scramble that inverts the usual order of infrastructure: compute campuses get planned and financed on energization dates the physical system cannot honor. About 12 gigawatts of U.S. data center capacity was planned for 2026, of which roughly 7 gigawatts — nearly half — is expected to be delayed or canceled, with only about 5 gigawatts under active construction. Data Center Watch counted 75 projects worth about 130 billion dollars blocked or delayed in the first quarter of 2026 alone.

03 The transmission gap: what build rate, and whose problem

The queue is a symptom; the disease is transmission built at a fraction of need. Only about 5,000 circuit-miles were added in 2024, mostly reliability-driven, and analyses cite figures as low as 55 miles of new high-voltage lines in 2023 — against National Transmission needs measured in thousands of miles per year. PJM approved a 6.7-billion-dollar 765 kV backbone in February 2025, largely for Virginia's data center corridor; it is a decade-scale project answering a three-year problem. FERC Order 1920, the landmark planning reform, saw its implementation delayed — compliance filings from PJM and MISO landed in December 2025, CAISO the same month, NYISO in April 2026, with first planning cycles starting 2026-2027.

Equipment is the second wall. Generator step-up transformer lead times exceeded 160 weeks in early 2026; high-voltage breakers run about 125 weeks. One missing transformer can idle an entire substation — and thus an entire data center campus — for years. This is where the constraint bites hardest: it is not that America cannot finance wires, it is that the global supply chain for the equipment that makes wires useful is booked solid into the next decade.

04 Who pays for the megawatt that is not there

The constraint is already repricing electricity. PJM's 2027/28 capacity auction cleared at a record — reported around 333 dollars per megawatt-day, roughly eleven times the 2024/25 price — with the market monitor attributing much of the increase to data center demand growth, including “phantom” projects holding transmission service agreements without operating. Congestion costs and rate cases follow the same current: utilities requested 31 billion dollars of rate hikes in 2025, and Virginia's regulators approved residential increases of about 16 dollars a month tied partly to data center growth.

The policy fight — covered in our companion analysis of rate design — matters here for one reason: if the grid's scarcity costs land on household bills, the political coalition for grid expansion collapses. Large-load tariffs that make growth carry its own infrastructure costs are not just fairness; they are the price of keeping the buildout publicly tolerable. The alternative, already visible in 189 gigawatts of proposed directly connected gas, is an AI industry that bypasses the grid — taking its revenue with it.

05 The flexibility heresy — and its limits

One school holds the constraint is partly self-inflicted: data centers are the most flexible large load ever connected, able to shift training workloads in time and location, and grid-friendly compute can behave like a virtual power plant. The companion argument — that scheduling AI work around grid conditions converts a bottleneck into an asset — is real, and FERC's 2026 proceedings created service classes for exactly such flexible large loads.

But flexibility has a ceiling that this article's numbers define: shiftable compute still needs firm energy at some hour of some day, the queue does not shrink because demand is polite, and inference — the latency-bound half of AI revenue — cannot be scheduled around a heat wave. The honest division: flexibility stretches the existing grid maybe 20 to 40 percent further; it does not build the missing transmission. Anyone treating demand response as a substitute for wires is doing decade-scale math with quarter-scale tools.

06 The verdict: a decade problem meeting a quarterly industry

The verified facts: 2,061 GW in queue versus ~1,280 GW installed; a 24 percent upward revision to NERC's ten-year demand forecast; 160-week transformers; 55 miles of new high-voltage line in 2023 and 5,000 circuit-miles in 2024; nearly half of planned 2026 data center capacity delayed or canceled. The analysis: the grid is not merely a constraint on AI growth — for the next five to seven years it is the binding one, because every substitute input (chips, capital, generation tech) can scale faster than transmission and interconnection can be built.