The Power Crunch Got Answered in Rules, Not Steel
The power crunch outran every physical fix, so the year's real adjustments came as rules and metrics — rationing schedules, delayed plant closures, and a new yardstick called intelligence per watt.
PJM Interconnection, the company that runs the electric grid for 13 states, has written a new rule for its newest customers. Any new load of 50 megawatts or more that arrives without its own power plant — which today means an AI data center — goes to the front of the line when the grid runs short: first to be cut off, and paid only half the normal compensation for being shut down. Those same customers have until March 1, 2027, to bring new generation online. [1] The company's stated reason is direct.
The unprecedented addition of large loads, most notably data centers, has given rise to resource adequacy shortfalls — and associated real-time operational issues that Interim Resource Adequacy Service is intended to address — PJM Interconnection
A market operator exists to match buyers and sellers of electricity. PJM is now running a blackout queue, with a 50 percent discount for the inconvenience of being first. It wrote the rule because the faster instruments had already failed. A capacity auction is how PJM pays power plants to stand ready years in advance. In the most recent one, the clearing price jumped from $28.92 to $269.92 per megawatt-day in a single cycle. [2] It hit its maximum allowed price and still failed to buy enough power. [3][4] And every physical fix is dated years out. Large power transformers — the machines that move electricity between voltage levels — now take more than three years to deliver, and the specialty electrical steel inside them comes from exactly one American mill. [5] Connecting a new data center takes one to seven years, and the substation that used to cost $50 million can now run $1 billion. [6] The flagship nuclear deal, Google's $4.3 billion plan to upgrade 11 existing reactors, delivers its first new power in 2028 and finishes in 2032 — most of its 3,590 megawatts is capacity that already exists. [7][8] Transmission grew 0.6 percent a year when demand needs 3 percent, and one large project can require up to 64 permits across 15 federal agencies. [9] Construction moves at the speed of steel and permits. Rules move at the speed of a filing. So the adjustments that actually landed this year were written, not built. PJM's own arc shows the discovery being made. In August 2025, its independent watchdog named the problem, and formally recommended the fix. [10]
The current supply of capacity in PJM is not adequate to meet the demand from large data center loads and will not be adequate in the foreseeable future. — Monitoring Analytics, LLC
By then, 64 percent of the ballooning capacity payments in PJM were tied to data center demand. [11] That November, stakeholders rejected all 12 ratepayer-protection proposals on the table. [12] The board chair, David Mills, was explicit about what the rejection meant.
Just because none of these passed does not in any way mean that the board will not act. — David Mills
He was right. In June, the Federal Energy Regulatory Commission ordered all six US grid operators to rewrite their rules for connecting large new loads. [1] Two months later, PJM's board made the turn concrete with the blackout-first proposal. The industry answered in its own way: tech companies are now adopting a bring-your-own-power model, buying on-site generation to skip the queue entirely and quiet the backlash over strained public grids. [13] The numbers suggest the workaround sticks. On-site generation costs $21.50 to $43.20 per megawatt-hour against a projected $88.60 grid tariff by 2027, so developers may keep the gas plants as primary power and treat the grid as backup. [14] The second adjustment was a change of ruler. In August 2025 the industry was already drifting from scale-first to what it called the "inference economy" — routing work to leaner models that cut compute loads by 75 to 87 percent — and the motive was plainly cost. [15] By this November the motive had hardened into physics. Moving data between chips now burns more power than the math itself, so the headline metric is "intelligence per watt": Samsung is stacking memory on top of processors to cut input-output power by roughly 70 percent, and racks are moving to 800 volts because 54-volt systems can no longer feed them. [16][17] There is another reason to count watts instead of tokens, and it has nothing to do with the grid. Anthropic's realized token price fell from $1.70 to $0.83 per million in a month, and corporate AI spending stalled even as volumes rose. [18] The power crunch and the price collapse push the same direction, which is convenient but not proof that either caused the other. The third adjustment carries the emissions invoice. Two Illinois state reports, published this week, conclude that AI data center demand makes the state's planned fossil retirements impossible on schedule, and recommend 11 oil-and-gas peaker plants — generators switched on only at peak demand — stay open until 2033 instead of closing in 2030 and 2031. [19] Illinois has moved its closure dates the wrong way, even as activists pushing the state's POWER Act campaign for the opposite. [20] The state insists it is not recommending new gas plants and still targets 2050 for closing every fossil plant. [19] Meanwhile, clean power sits switched off for lack of wire. Curtailment — wind and solar told to stop generating because the transmission can't carry it — is accelerating nationwide, even as solar and storage made up 70 percent of new US generating capacity in the first half of this year. [4][3] None of this requires believing every gigawatt in the queue. Bernstein Research figures only about 35 percent of announced US data center capacity is likely to be built; the rest, it says, is "Bragawatts" — requests from inexperienced operators and duplicate ghost entries. [21] Texas's interconnection queue holds 474 gigawatts, five times the state's record peak. [22] Texas power futures are already pricing in a less extreme buildout than feared. [23] And $3 billion of new transformer plants still can't buy the trained labor and single steel mill fast enough. [24] But the collisions that already happened don't depend on the hype. The auction hit its cap and came up short. Fluke's survey of 1,300 data center professionals found only 33 percent say their infrastructure is ready for AI loads, and 78 percent had a power disruption in the past year. [25] The calendar tells you where this lands. Those 11 peakers are now scheduled to keep burning toward 2033, while already-built solar sits curtailed for lack of wire, in a state that still says every fossil plant will close by 2050. [19] The power crunch's most reliable product is delay, and the delay is paid in the years of emissions between now and then.
- 1. PJM Proposes Power Cuts for Unsupported Data Centers
- 2. Data Center Power Demand Drives US Grid Cost Debates
- 3. U.S. Electricity Demand Outpaces Supply Ahead of Midterms
- 4. U.S. Grid Congestion Drives Renewable Energy Curtailment
- 5. US Power Grid Faces Delays Amid Transformer Shortages
- 6. AI Infrastructure Faces Severe Power and Talent Bottlenecks
- 7. Google and Constellation Energy Sign $4.3 Billion Nuclear Deal
- 8. Tech Hyperscalers Commit to 19 GW of Nuclear Energy
- 9. Barclays Warns U.S. Power Grid Expansion Lags Demand
- 10. Grid Watchdog Urges Data Centers to Provide Own Power
- 11. AI Data Centers Drive Electricity Costs for 67 Million Americans
- 12. PJM Stakeholders Reject All 12 Data Center Energy Proposals
- 13. Tech Companies Adopt Self-Powering Models for AI Data Centers
- 14. Data Center Developers Adopt On-Site Gas Power to Bypass Grid
- 15. AI Industry Shifts Toward Efficiency-Driven Inference Economy
- 16. AI Industry Shifts Focus Toward Energy Efficiency and Power Scaling
- 17. AI Industry Shifts Focus to Intelligence Per Watt
- 18. AI Token Spending Stalls Amid Steep Price Declines
- 19. AI Data Centers Delay Illinois Fossil Fuel Plant Closures
- 20. Activists Push Data Center Bans and Regulations in Ohio and Illinois
- 21. Electricity Shortages Delay US AI Data Center Boom
- 22. AI Data Centers Strain Power Grids in US and Europe
- 23. Texas Power Markets Signal Slowdown in AI Data Center Growth
- 24. AI Boom Drives $3 Billion Investment in U.S. Transformers
- 25. Fluke Study Finds Data Centers Unprepared for AI Power