The AI Power Problem Doesn't Get Solved, It Moves

In the June heat wave, the Department of Energy gave PJM — the grid operator for 13 states — emergency authority to order data centers onto their backup diesel generators so that homes wouldn't lose power ^[1]. The generators were supposed to be the industry's escape hatch: the way to keep building when the grid couldn't deliver. Instead they became the grid's relief valve, burning diesel through a heat emergency while the people who live next to them watched. A Pennsylvania official put it plainly.

The billionaires that are driving this AI revolution that we hear about all the time in the news, they want communities like Lackawanna County to bear the burden while they reap all the rewards. — Bill Gaughan

The Sierra Club's Tom Schuster was blunter still.

They’re even worse than coal. — Tom Schuster

That scene is the entire buildout in miniature: each fix for one bottleneck becomes the next bottleneck, and the problem never resolves — it migrates. Start with the grid. Developers can't wait the decade-plus it takes to expand transmission, so they stopped waiting. They build their own power behind the meter, bypassing the utility entirely: Meta has deployed more than 800 mobile mini-turbines in El Paso, and OpenAI signed a $1.25 billion deal for jet-engine turbines ^[2]. Developers added only 25 gigawatts to their pipeline in the fourth quarter of 2025 — half the prior quarter — and on-site gas is the workaround for the delay ^[3]. That workaround now has its own opponents. In Richmond County, North Carolina, more than a hundred residents turned out against air permits for 600 diesel backup generators at an Amazon site ^[4]. In Mingo County, West Virginia, residents protested gas plants running more than a hundred methane engines to power data center microgrids ^[5]. And the EPA has turned its attention to the turbines themselves.

To meet this need for speed, on-site generating capacity and behind-the-meter distribution are becoming standard considerations. — Burns & McDonnell

The fix for the grid delay has become a regulatory problem of its own ^[3]. The gas plants, in turn, need the high-voltage transformers that connect a power plant to the grid. Eaton's U.S. data center infrastructure backlog stands at 307 gigawatts — fifteen years of work at last year's build rate ^[6]. Transformer makers have put more than $3 billion into new capacity, but the constraint isn't money; it's specialty steel and the people who wind the coils ^[7]. No amount of capital accelerates a skilled-trades shortage. Which brings the cascade to capital itself. Apollo's chief economist Torsten Slok estimates the buildout may need more than $2 trillion in debt, while public bond markets can absorb no more than $1 trillion through 2030 — a trillion-dollar gap that has to be filled by private credit ^[8]. You can see what that looks like in practice: Oracle's $16.3 billion Michigan financing closed only after PIMCO anchored $10 billion of it, once other banks had withdrawn ^[9]. The money exists, but it's concentrating in fewer hands, on worse terms. And where construction actually succeeds, the result is its own kind of problem. Meta, unable to monetize its models directly — its chatbots hold a sliver of the market despite $130–145 billion a year in capital spending ^[10] — launched Meta Compute to sell its excess GPU capacity to outside customers ^[11]. Meta's stock rose 10%; Nebius and CoreWeave fell 17% on the news, because a hyperscaler dumping spare compute undercuts the claim that AI compute is scarce ^[11]. That's the paradox the whole cascade was building toward. The scarcity thesis justified the spending; the spending produced the glut. Where projects can't get built, there's scarcity — grid, gas, transformers, capital, each one a wall. Where they can get built, there's oversupply, and the price of compute collapses. The industry spent a trillion dollars to escape a shortage and ended up with both: a shortage everywhere it couldn't build, and a glut in the one place it could.