The AI Shortage Money Can't Fix

Even at the highest allowed price, the market is not attracting the level of new supply the system needs. — Calvin Butler

That is Calvin Butler, the chief executive of Exelon, and he is not complaining that power is too expensive. He is reporting that price has stopped working. Exelon is one of the nation's largest utilities, and Butler is trying to buy electricity inside PJM, the wholesale market that spans much of the eastern United States and is the largest of its kind. Whatever the market is allowed to offer, no one will build what the system needs. What kind of shortage ignores price? One that has left the chip behind. That is now the condition of the physical layer of the AI buildout. General Electric's gas turbines, the machines that burn natural gas to spin a generator and the standard order when a utility needs firm power quickly, are sold out through 2029 ^[1]. The electrical gear is queued even deeper: Eaton alone carries a 307-gigawatt U.S. data-center backlog, roughly fifteen years of work at 2025 build rates, and delivery dates for the largest power transformers now run into the 2030s ^[2]. A transformer steps electricity up or down in voltage so it can travel across the grid; the biggest ones, the kind a hyperscale campus needs, have become a multi-year commitment before ground is broken. Even permission to connect now runs past a decade: Google names electrical transmission its number-one grid challenge, and after one utility quoted twelve years just to complete the interconnection study that decides whether a new plant may plug in at all, the company began building campuses directly beside power plants to skip the queue ^[3]. And the fuel has its own waiting list. The Department of Energy has invoked the Defense Production Act, a 1950s wartime authority now mobilized for server-farm fuel, to convene a consortium spanning uranium mining, milling, conversion, and enrichment, warning that AI data-center demand could push annual grid outage hours from single digits past eight hundred ^[4]. The insider's confirmation comes from Nokia's chief executive, Justin Hotard.

I don't think you can say in any manner we're overbuilding today because reality is that if we could build 2x faster, our customers could build 2x faster, they probably would. — Justin Hotard

Nor has the constraint simply left silicon behind. Musk says the current limit at xAI is memory, and prices for memory chips have nearly doubled on supply shortages, so the scarcity has spread across the whole industrial stack: energy, turbines, and memory alike ^[2]. But there is a complication on the far side of every queue: much of the demand may not be real. Wood Mackenzie estimates only about 28% of the 1,066 gigawatts of power requested for U.S. data centers will actually materialize, the rest speculative filings ^[5]. Texas regulators ordered ERCOT, the state's grid operator, to revise its forecast after it projected peak demand quadrupling by 2032 ^[6]. Price cannot summon supply, and a large share of the demand is phantom. So the physical system has stopped taking orders and started screening them. The screening starts at Exelon itself. The company whose chief executive described the shortage pruned its own pipeline: high-probability data-center load fell nearly 40% once it began demanding transmission security agreements, cash posted up front, and the weeding killed a 1.8-gigawatt, $20 billion facility in Joliet, shrinking its interconnection queue from 43 gigawatts to 25 ^[7]. In Texas, Governor Greg Abbott ordered a mandatory audit of every data center seeking grid access ^[5]. The same screen now runs across the map. Ireland's regulator has proposed gas-price discounts for new data centers that accept planned interruptions during cold snaps, meaning they power down, postpone work, or switch to diesel rather than leave homes unheated ^[8]. New Zealand's opposition proposes requiring AI data centers to bring their own new, firm generation so the buildout stops raising everyone else's power bill ^[9]. More than seventy European projects were rejected or restricted in the first four months of this year ^[10]. The lenders, for their part, are now watching the hardware rather than the models: central banks flagged $450 billion of AI-related debt issuance by September, and analysts say the first cracks will show in the loans' physical collateral, chip resale values and utilization rates ^[11][12]. The credit-desk frame fits the audits and the security agreements exactly; the rationing schemes sit alongside it rather than inside it. And for the sellers, the crunch is an earnings story: utilities carrying long queues and signed partnerships are committing capital, not retreating ^[13]. If the shortage cannot be bought, an industry that sold itself as software has begun answering by owning the physical world instead of renting it. Nvidia, whose chips were the last binding constraint, is now lobbying for nuclear regulatory reform and supplying reactor developers with computer models of their designs ^[14]. Meta's Louisiana data center is being built alongside ten company-funded natural gas plants totaling 7.5 gigawatts ^[1]. Musk has put a billion dollars into a turbine-leasing company and begun making turbine blades in-house ^[15]. OpenAI has signed up 8 gigawatts in Ohio, and Amazon has proposed a 7.7-gigawatt gas plant in West Texas ^[16]. The state is already financing the layer: Southern Company closed a $26.5 billion federal loan to build or upgrade 16 gigawatts of new gas, nuclear, and transmission ^[17]. The industry's rhetoric has migrated into energy units along with its budgets. Sam Altman now concedes that new OpenAI facilities draw five times the electricity of their predecessors.

People said we were using an unacceptable amount of power, and with these upgraded facilities that require so much more of it, we’ve shown that we took those criticisms to heart. — Sam Altman

And its promises are now sized in gigawatts.

Maybe with 10 gigawatts of compute, AI can figure out how to cure cancer. — Sam Altman

That is the load regulators are drafting schemes to interrupt, so home heating keeps running through winter ^[8]. On one desk, the Department of Energy invokes a wartime statute to convene a fuel consortium ^[4]. On the other, a software company asks the government to serve as its insurer of last resort for the financing ^[18]. Fuel on one desk, debt on the other.