The AI boom has hit two walls at once.
The first is physical: there isn't enough electricity, transformer capacity or transmission to power the data centers already financed and planned. The second is political: a growing backlash from voters, lawmakers and regulators who worry those data centers will raise power bills, strain local resources and leave existing customers with the bill.
Each wall makes the other worse. Shortages raise bills, higher bills fuel opposition, and opposition slows the infrastructure that would ease the shortage. Earlier this year, two pieces captured the tension from opposite directions. Americans for Tax Reform argued that "data center panic" is delaying the power America needs. Rajesh Beri described an industry with committed capital, real demand and available chips, but not enough electricity to switch them on. Both were partly right. Months later, the evidence on both sides has only grown.
The power wall is real
The money is committed. Beri tallied $660 billion to $690 billion in 2026 capital spending from Alphabet, Amazon, Meta, Microsoft and Oracle combined, plus more than $400 billion committed to the Stargate project. Citing Bloomberg data, he reported in April that roughly half of planned U.S. data center builds for 2026 had stalled or been canceled, and Microsoft CEO Satya Nadella has said publicly that GPUs were sitting in inventory for lack of power to run them.
The equipment shortage is getting worse, not better. According to Reuters reporting cited by DataCenterScope, lead times for high-voltage transformers reached up to 160 weeks in 2026, up from about 143 weeks in 2024. The Department of Energy says large substation and generator transformers have stretched from about three years to as long as four, with some prices up four- to nine-fold over five years.
"You cannot build a data center without electricity," Beri wrote, "and you cannot get electricity without transformers, switchgear, transmission lines, and grid interconnection." Each of those now takes longer than a tech company's planning cycle, the same time-to-power problem reshaping where AI gets built. As I argued in Is Compute the New Oil?, the chips can be bought. Powering them is the hard part.
PJM is where both walls meet
Nowhere is the collision clearer than PJM, the grid operator serving about 65 million people across 13 states and Washington, D.C., including the densest data center corridor in the world.
PJM's 2028/29 capacity auction came in about 6.8 GW short of its reliability target, a capacity shortfall driven largely by data center demand. To close it, PJM designed an emergency "reliability backstop" procurement. As Utility Dive reported, federal regulators rejected most of that plan in late September 2026 over cost allocation and collateral rules, and FERC Chairman Laura Swett said the last-minute filing left no time "to rehabilitate the mess we received."
PJM filed a revised plan on October 29. Procurement now starts March 1, 2027, about seven months late. And something revealing happened along the way: utilities including PPL Electric, FirstEnergy and AES Ohio cut their near-term large-load forecasts by about 6.5 GW. According to a Jefferies note cited in Utility Dive's coverage of the revised plan, the backstop target could shrink to as little as 630 MW.
PJM's shrinking backstop target
Sources: PJM, Jefferies via Utility Dive, Nov. 2026.
The backlash has a point, and so do its critics
Households are already feeling it. In its October 2026 outlook, reported by Carrier Management, the EIA projected wholesale power prices up about 11% nationally in 2026 and about 41% in the PJM region. Beri cited projections that AI-driven grid upgrades could add $15 to $25 a month to the average household bill.
Lawmakers have noticed. More than 300 data center bills were filed in over 30 states in 2026, and more than ten states introduced moratorium proposals. Texas's governor paused state permits for data center projects in September pending grid and water audits. The public trust problem is real, and it's bipartisan.
Americans for Tax Reform's counterargument still deserves a hearing. When projects stall under political pressure, the shortage doesn't go away. It gets worse, and higher prices and reliability risks fall on everyone, including households and manufacturers with no connection to AI. Electrification, reshored manufacturing and population growth need new supply whether or not another data center is built.
Both things can be true. Ratepayers need protection from paying for infrastructure built for someone else. Freezing the buildout leaves those same ratepayers with a tighter, more expensive grid.
What each policy lever actually fixes
| Policy lever | What it fixes | What it doesn't fix |
|---|---|---|
| Moratoriums and permit pauses | Gives communities time to set rules | Adds no supply; shortages and prices persist |
| Large-load tariffs (Ohio, North Carolina) | Makes data centers pay for their connections and carry cancellation risk | Usually doesn't cover new power plants |
| Fast-track federal permitting | Speeds approvals for projects and their energy infrastructure | Can't speed up transformer factories |
| Behind-the-meter generation | Gets projects powered without waiting for the grid | Raises emissions and permitting questions |
| Efficiency and flexible load | Gets more value from every existing megawatt | Doesn't remove the need for new generation |
No single lever solves both walls. The states making progress are combining several.
The workaround: building your own power
Developers who can't wait for the grid are going around it. Beri pointed to Microsoft and Chevron's reported $7 billion gas plant in West Texas, designed to deliver about 2,500 MW to a Microsoft campus, and Oracle's Stargate sites using on-site gas engines and turbines. Amazon and Google are pursuing nuclear and small modular reactors on longer timelines.
Behind-the-meter power can shield other ratepayers from some grid costs. It also creates a new tension: companies marketing AI as a sustainability tool are building gigawatts of new fossil generation to run it. Pairing on-site generation with battery storage, flexible workloads and a path to cleaner firm power will decide whether these private power islands become grid assets or stranded emissions.
The third lever: more value per megawatt
Most of this debate assumes the only choices are building more supply or blocking more demand. There's a third: getting more useful work from every megawatt.
Smaller, distilled AI models can cut compute needs by 40% to 60% for many uses without a meaningful drop in quality, according to Beri. Flexible workloads that shift away from peak hours ease the strain further, the same promise Bitcoin miners are now being asked to prove in Texas. That's the case for intelligence per watt: when power is the binding constraint, the most efficient AI is also the most available.
Efficiency won't eliminate the need for new generation. But every megawatt saved is one that doesn't need a 160-week transformer, a contested transmission line or a rate case.
What a workable middle looks like
- Make developers carry their own risk: minimum bills, collateral, exit fees and long contracts, as in Ohio and North Carolina, so canceled projects don't land on households. PJM's shrinking forecasts show why this matters.
- Decide who pays for generation: the open question in North Carolina and the core issue in Who Should Pay for the AI Power Boom.
- Keep approved supply moving: pausing generation and transmission that's already approved deepens the shortfall everyone pays for.
- Treat on-site power as a grid partner: behind-the-meter generation and storage that can support the grid in emergencies is worth more than an isolated island.
- Reward flexibility and efficiency: pay loads that can curtail on command, and prioritize projects that do more per megawatt.
- Share the benefits locally: communities hosting the infrastructure need a visible stake in what it creates.
What this means for developers
Plan as if both walls are permanent. Order long-lead equipment early, assume tariffs with minimum bills and collateral, and price a scenario where you pay for part of the new generation your project requires. Show up with your own answer to the power question, whether that's contracted new supply, on-site storage, flexible operation or all three. Projects that do will move faster through both the interconnection queue and the politics.
The bottom line
The power wall is physics. Transformers, turbines and transmission lines take years, and no policy shortcut changes that. How long the wall lasts, and who pays for it, is a choice.
Blocking every project leaves everyone short of power. Building without guardrails leaves households paying for someone else's bet, a risk PJM's shrinking forecasts make plain. The states that get this right will build supply quickly, put risk on the parties that create it and treat every megawatt as scarce, because for the next few years it will be.
Chris Kalowes
Founder, WattThe?!
Energy intelligence. Simplified.
