The Energy Bill Always Comes Due

Sometime in the past few months, 49,000 people living near Lake Tahoe learned that their power grid was being quietly redirected toward data centers. Not cut off entirely, not yet, but rerouted in ways that made their service less reliable, their outages more frequent, their complaints less urgent to the utility than the cooling needs of server racks. Fortune covered it in May, and the quote that stuck with me was from a resident: "It's like we don't exist." Forty-nine thousand people, and the grid math just... didn't include them anymore.

That's not a policy failure in the abstract. That's a specific choice someone made in a spreadsheet.

Meanwhile, in Memphis, Elon Musk's xAI facility has been running nearly 50 gas turbines without proper air permits, according to TechCrunch. Unchecked. Just running. The kind of thing that would shut down a manufacturing plant or a refinery inside of a week, but apparently the rules have some flexibility when the turbines are powering a large language model. And in Utah, the state just approved a data center project twice the size of Manhattan, drawing immediate backlash from residents and environmental groups who used the word "irresponsible" in press statements, which is the polite version of what they were actually thinking.

Kevin O'Leary, whose $1 billion Stratos project is at the center of the Utah fight, has a response to all of this: everyone else is doing it wrong, and he's the only data center developer who actually studied the environmental impact. He called the backlash "kind of hypocritical," because people use AI and also don't want the infrastructure near them. That's not wrong, exactly. It's also not an argument. It's a shrug dressed up as a point.

A Gallup poll from this spring found that seven in ten Americans oppose constructing data centers in their local area, with nearly half strongly opposed. Barely a quarter favor them. Seven percent are strongly in favor. That seven percent is building everything right now.

In 1954, Lewis Strauss, the chairman of the Atomic Energy Commission, gave a speech predicting that nuclear power would make electricity "too cheap to meter." It became one of the most quoted promises in the history of technology, and it became that because it was so completely, so perfectly wrong. Not wrong about the physics, not entirely wrong about the potential, but catastrophically wrong about the gap between what a technology can do in theory and what it costs to operate in a society with actual people in it, people who live near the plants, people whose water gets warm, people who have to be told something when things go wrong. The promise was real. The bill was also real. And the bill included things Strauss never put in the spreadsheet.

The AI industry is running the same calculation right now, and making the same omission. The promise is genuine: compute at scale, intelligence on demand, productivity gains that compound. The bill includes 50 gas turbines running without permits in Mississippi, 49,000 residents whose grid reliability is now subordinate to server cooling, and a Utah approval process that moved faster than the communities it affects could organize a response. None of those costs show up in the model until they become a lawsuit or a blackout.

What's actually shifting here is who holds the constraint. For the past decade, the binding limit on AI development was hardware: chips, fab capacity, the TSMC queue. That's loosening, slowly. The new constraint is physical power, and unlike chips, power is local. You can't ship a megawatt. You can't route around a grid that's already at capacity. This is why Fervo Energy, a geothermal startup, popped 33% on its IPO debut last week. Investors aren't buying a clean energy story; they're buying a compute security story. They're betting that whoever controls reliable, scalable, politically defensible power will control the ceiling on AI growth. That's a different kind of infrastructure bet than anything the tech industry has made before, because it's a bet on geology and grid physics, not on software iteration.

The diagnostic question isn't whether AI needs more power. It obviously does. The question is whether the people building the infrastructure have a model that includes the communities sitting on top of the grid. Three things tell you they don't: permits filed after the turbines are already running, utility redirections that treat residential customers as the flexible load, and approval processes that move at the speed of capital rather than the speed of civic input. When all three are present in the same news cycle, you're not looking at isolated bad actors. You're looking at an industry that hasn't yet been forced to internalize the costs it's externalizing.

The geothermal IPO is interesting precisely because it suggests the market is starting to price this in, at least partially. Fervo's 33% pop isn't irrational; it reflects a real scarcity. But it also reflects the same pattern Strauss ran into: the assumption that the infrastructure problem is purely technical, that if you find the right energy source, the siting fights and the community opposition and the permit violations just resolve themselves. They don't. The Lake Tahoe residents don't become less real because the power source is geothermal.

The energy bill for AI is going to be paid. The only live question is whether it gets paid by the companies building the infrastructure or by the 49,000 people who already noticed their power getting worse and felt, correctly, like they didn't exist in someone's model.

The turbines in Mississippi are still running.