AI Data Center Power Crunch: The Grid Is the New GPU Shortage

In the $655 billion bet I looked at where the AI capex goes. This piece is about what it runs on—because in 2026 the binding constraint on AI is no longer GPU supply. It's electricity.

Nvidia can print more chips. Nobody can print more grid. Interconnection queues run five to seven years, gas turbines are sold out into the 2030s, and the auctions that keep the lights on are clearing at prices nobody modeled. The power market has quietly become the purest expression of the AI trade—and the most mispriced.

Demand: A Doubling, Not a Drift

Start with the demand curve that broke everything.

FIG. 01 // US POWER DEMAND

US data-center power demand is compounding, not growing

Goldman Sachs projects US data-center demand more than doubling between 2025 and 2027. The share of total summer peak load doubles to 8.5% in the same window.

DATA: GOLDMAN SACHS RESEARCH · 2026

Goldman Sachs projects US data-center power demand climbing from 31 GW in 2025 to 66 GW in 2027—more than doubling in two years—with the share of US summer peak load going from 4.1% to 8.5%. Gartner sees global data-center electricity consumption growing 26% in 2026 alone, to 132 GW. The IEA's base case has global consumption roughly doubling to ~945 TWh by 2030.

For context: 66 GW is about sixty-six nuclear reactors' worth of continuous load, being added to a grid whose planning cycles are measured in decades. Utilities that spent twenty years managing flat demand are now staring at compound growth. That mismatch—not chip supply—is the story of 2026.

Price: The 11x Nobody Backtested

Demand shocks meet inelastic supply in exactly one place: the auction.

FIG. 02 // PJM CAPACITY AUCTION

The price of keeping the lights on rose 11x in two years

11×

CAPACITY PRICE, 2024/25 → 2026/27 DELIVERY YEAR

PJM is the largest US grid operator (65 GW peak, 13 states). Its capacity auction cleared at $28.92/MW-day for 2024/25 and $329.17/MW-day for 2026/27 — and still fell 6.8 GW short of its reliability target for the third consecutive auction.

DATA: PJM · IEEFA · CANARY MEDIA · JUL 2026

PJM—the grid operator covering 13 states and 65 million people—runs a yearly auction to secure generating capacity. In two years, the clearing price went from $28.92 to $329.17 per megawatt-day: an 11x repricing of the same product. IEEFA attributes 63% of the 2025/26 increase directly to data centers. Wholesale power in PJM averaged $136.53/MWh in Q1 2026, up 75.5% year-over-year—the largest single-year jump in the market's history.

And here's the detail that should stop you: even at 11x the price, the last auction fell 6.8 GW short of its reliability target—for the third time in a row. Price is doing what price does when quantity can't respond. This is not a spike; it's a regime.

The Bill: A Political Fuse

Someone pays for all this, and it's not only the hyperscalers.

FIG. 03 // THE BILL

Who pays for the grid the AI boom needs

Virginia electricity already consumed by data centers

26%

projected 41-59% by 2030 (EPRI)

Planned capex by 51 US utilities through 2030

$1.4T

30+ cite data centers as a top growth driver

Extra cost passed to PJM ratepayers, 2025/26 auction

$9.3B

≈ $21/month on a DC-area residential bill

Grid capex lands on regulated rate bases, which means consumers co-finance AI infrastructure through their utility bills — a political fuse burning under the entire build-out.

DATA: EPRI · POWERLINES · IEEFA · CANARY MEDIA · 2026

Data centers already consume 26% of Virginia's electricity, heading for 41-59% by 2030 on EPRI's projections. US utilities have $1.4 trillion of capex planned through 2030. And the 2025/26 PJM auction alone pushed $9.3 billion onto ratepayers—about $21/month on a Washington-area residential bill.

That last number is the one to watch. Utility capex lands on regulated rate bases: households are co-financing AI infrastructure whether they use ChatGPT or not. Rising residential bills in an election-heavy decade are exactly the kind of input that produces price caps, data-center moratoria (Ireland and parts of Virginia already have them), and windfall clawbacks. The AI build-out's biggest unhedged exposure isn't technical. It's political.

The Quant Angles

Power is the purest AI trade left. AI equities price perfection; capacity auctions still clear on scarcity math. Independent power producers, grid-equipment makers and uranium have outrun most of the Mag 7 since 2024—and the forward curve says the scarcity persists into the 2030s.
Electricity became an alt-data stream for AI itself. Interconnection-queue filings, PPA announcements and auction results are public, slow-moving and legally binding—a far more honest signal of real AI compute growth than any earnings call. If the capex is fake, the megawatts won't show up.
Basis trades are opening everywhere. Regional spreads (constrained PJM vs. gas-rich ERCOT), capacity vs. energy prices, regulated utilities vs. merchant generators—the demand shock is local, the repricing is uneven, and dispersion within the power complex is wide.
Model the political scenario, not just the demand one. A ratepayer backlash that shifts grid costs onto data centers changes hyperscaler opex materially—and it's the scenario nobody's DCF includes. The $21/month is the fuse; watch state utility commissions, not tech conferences.

The honest assessment: the market spent three years asking whether AI demand is real. The grid already answered—demand is real enough to break a fifty-year-old auction mechanism. The open question is who ends up paying for it, and positioning for that answer is a better use of research budget than another token-per-dollar benchmark.

Trading power markets, or building the models that should? The forward curve tells one story; the interconnection queue tells another.