The Power Wall
⏱️ Reading Time: 7 minutes
TL;DR:
- The Shift: For three years the AI supply chain conversation was about GPUs. In 2026 it is about gigawatts. The bottleneck has physically moved from the server rack to the substation.
- The Scale: U.S. data center power demand is projected to hit 75.8 GW in 2026, up from roughly 53 GW in 2023, on a path toward 290 GW globally by 2030, according to Gartner and S&P Global’s 451 Research.
- The Fallout: Gas turbine manufacturers are sold out for years, hyperscalers are signing nuclear power deals by the gigawatt, and a fight is breaking out in state legislatures over who actually pays for all of it.
1. The Bottleneck Moved From the Chip to the Substation
From 2021 to 2024, the AI industry’s defining constraint was silicon. Nvidia could not make H100s fast enough, and the entire narrative of the boom was written in units of GPUs.
That constraint has been solved. The one replacing it is much harder to fix.
U.S. data center electricity demand is projected to reach 75.8 GW in 2026, up from about 53 GW in 2023, according to S&P Global’s 451 Research. Gartner puts global data center power demand at 132 GW in 2026, rising 27 percent from 104 GW the year before, and projects it will hit 290 GW by 2030. The consulting firm Grid Strategies found that utilities’ own five-year forecasts of summer peak demand growth more than tripled, from 38 GW in 2023 to 128 GW in 2024, as data centers entered utility planning models for the first time in two decades.
The reason this shift matters more than the chip shortage did is timing. A GPU factory can add capacity in 12 to 24 months. A high-voltage transformer or a gas turbine cannot. Those lead times now run 2 to 4 years for transformers and 3 to 5 years for large turbines, according to industry trackers and IEEFA’s October 2025 report on global turbine shortages. Hardware scales on a chip cycle. Grids scale on a decade cycle. The AI industry is trying to force the second timeline to match the first.
2. Why a Handful of AI Facilities Can Break a Regional Grid
The scale problem is compounded by an intensity problem. A single AI-heavy compute task can draw up to 1,000 times more electricity than a traditional web search, which is why a handful of large AI campuses can strain a regional grid in a way that hundreds of ordinary data centers never did.
The clearest real-world example is PJM Interconnection, the grid operator covering 13 states and Washington, D.C., including Virginia, the largest data center market on Earth. In an emergency order this Summer, the Department of Energy cited grid reliability findings that PJM’s demand is growing at its fastest pace in years, driven primarily by data centers. On June 30, 2026, PJM forecast a peak load of 166,304 megawatts for July 2, which would have broken the grid’s all-time Summer record set in 2006. PJM’s Independent Market Monitor separately reported that total wholesale power costs in its territory jumped 49 percent in 2025, driven in large part by a 262 percent spike in capacity costs tied to data center demand.
THE DEMAND CURVE VS. THE BUILD CURVE
GW
300 | * 290 (2030, global)
| *
200 | *
| *
132 |---------------------* <- 2026 global demand
|
76 |---------------------* <- 2026 US demand
| * *
53 |* <- 2023 US demand
|____________________________________________________
2023 2024 2025 2026 2027 2028 2029 2030
Meanwhile, large gas turbines and grid transformers
take 3-5 years to build. The demand line moves in
quarters. The supply line moves in presidential terms.
Sources: S&P Global 451 Research, Gartner, Grid Strategies.
3. The Turbine Shortage Nobody Saw Coming
Natural gas still supplies over 40 percent of U.S. data center electricity, more than any other source, which makes gas turbines the single most important piece of hardware in the entire power buildout. There are effectively three companies on Earth that can manufacture large-scale gas turbines: GE Vernova, Siemens Energy, and Mitsubishi Heavy Industries.
All three are sold out for years. GE Vernova’s combined gas turbine backlog and slot reservations reached 100 GW in the first quarter of 2026, up from 80 GW at the end of 2025, and the company now expects to hit 110 GW by year-end. Data center customers accounted for $2.4 billion of GE Vernova’s equipment orders in a single quarter, more than the company booked from data centers in all of 2025 combined. CEO Scott Strazik has said he expects turbine reservations to be sold out through 2030 by the end of this year. Wait times for large frames have stretched from two to three years earlier this decade to roughly five years now, with some manufacturers advising customers to plan on seven to eight year timelines, according to IEEFA.
Pricing tells the same story as the queue length. New gas turbine order pricing in the first half of 2026 is running 10 to 20 percentage points higher per kilowatt than it was in the fourth quarter of 2025, comfortably outpacing inflation. When three manufacturers control a market and demand triples, the shortage does not show up as an empty shelf. It shows up as a longer line and a bigger invoice.
4. The Nuclear Gambit
Faced with a multi-year wait for gas turbines and a grid that cannot expand fast enough, every major hyperscaler has made the same bet: nuclear power, bought directly, years before it is available.
- Microsoft signed a 20-year, $16 billion power purchase agreement for 835 MW from the restarted Three Mile Island Unit 1, now renamed the Crane Clean Energy Center, with first power expected in 2027.
- Google committed to 500 MW from Kairos Power’s molten-salt reactor design, with Kairos and Google absorbing the first-of-a-kind construction risk while the Tennessee Valley Authority provides the revenue stream through a power purchase agreement.
- Amazon invested $700 million in X-energy for up to 12 small modular reactors, on top of a separate $650 million data center campus built next to Pennsylvania’s Susquehanna nuclear plant.
- Meta leads the pack with commitments across TerraPower, Oklo, Vistra, and Constellation totaling up to 6.6 GW.
Across all four companies, more than 13 nuclear deals and roughly 9.8 GW of committed capacity had been announced as of May 2026. It sounds like a lot until it is measured against the problem it is meant to solve. Goldman Sachs estimates that meeting all data center power demand growth through 2030 would require 85 to 90 GW of new nuclear capacity. Less than 10 percent of that is likely to actually exist by then. Nuclear is not going to bail out the grid this decade. It is a hedge for the 2030s, being purchased now because the lead times are even longer than gas.
5. Who Actually Pays for This
This is the fight that has moved from trade press to state legislatures. In 2026 alone, lawmakers in more than 30 states introduced over 300 bills addressing data center energy policy, moratoriums, and tax treatment, according to the policy tracker Multistate.
The dispute centers on a genuinely unresolved empirical question: do data centers raise electricity bills for everyone else, or do they pay their own way? The evidence is mixed and depends heavily on how a given utility structures its contracts.
On one side, Harvard Law School’s Electricity Law Initiative argues that utilities socialize the cost of new transmission and generation infrastructure across all ratepayers by default, the same way they have always spread the cost of a fallen tree limb or a new subdivision. When a single customer’s new load is the size of a city, that traditional cost-sharing model transfers real money from households to the facility. A Bloomberg analysis found wholesale electricity costs in some data-center-heavy areas rose as much as 267 percent over five years, though fact-checkers at PolitiFact and WRAL note that figure describes wholesale prices, not the retail bills households actually receive, which are also shaped by transmission, distribution, and taxes.
On the other side, an E3 whitepaper commissioned by the Data Center Coalition reviewed 11 quantitative studies and found no evidence of a historical cost shift from data centers onto residential customers, and a separate E3 analysis of Amazon’s facilities found each site generated an average of $3.4 million in net surplus revenue for its utility, money that can fund grid upgrades other customers benefit from. Amazon-commissioned research reached a similar conclusion, and Pacific Gas & Electric has said each gigawatt of data center demand could actually lower average household bills by 1 to 2 percent under the right contract structure.
Both things can be true at once, because the outcome is not determined by data centers in the abstract. It is determined by the specific tariff a state regulator negotiates. Georgia’s Public Service Commission created rules requiring data centers to fund their own generation, transmission, and distribution costs. Oregon built a separate rate class for large loads with long-term contracts. States without those protections are the ones showing up in stories about $281 electricity bills in Virginia. In March 2026, Microsoft, Anthropic, and other AI companies signed a nonbinding Ratepayer Protection Pledge at the White House, committing to build, bring, or buy their own power and to fund related grid upgrades themselves. Whether that pledge holds up once it collides with a real interconnection queue is the thing worth watching next.
The Bottom Line
The AI industry spent three years worrying about whether it could get enough chips. It is about to spend the next three worrying about whether it can get enough electrons, and electrons do not respond to Moore’s Law. A GPU cluster can be ordered, financed, and racked in a data hall within two years. A gas turbine, a high-voltage transformer, or a nuclear reactor cannot, and that mismatch is now the single biggest determinant of where AI infrastructure gets built and how fast it scales.
The companies that win the next phase of this buildout will not be the ones with the best model. They will be the ones that solved their power contracts first, whether through nuclear offtake, self-supply agreements, or simply picking a state with a large-load tariff that does not put the bill on somebody else’s kitchen table. The GPU shortage was a supply chain problem. The power shortage is an infrastructure problem, and infrastructure does not move at the speed of software.