Dear Reader,
On Saturday, the man who runs Anthropic published an essay asking his own industry to slow down.
"We must slow the pace at which we improve the capabilities of AI models," Dario Amodei wrote. He argued that buying even an extra year or two, before these systems reach what he called critical levels of capability, would be worth it.
Then the unusual part. His competitors agreed with him.
Sam Altman at OpenAI. Elon Musk. Demis Hassabis, who runs Google's DeepMind. Altman went a step further and said OpenAI will not go public this year.
Yesterday the market did exactly what you'd expect it to do. It sold the chips.
Intel, Micron and AMD each dropped about 5%. Over in Europe, ASM International fell nearly 9%. In Asia, SK Hynix and Samsung closed down 4% and 6%.
The reasoning is simple enough. Slower AI means a slower buildout, and a slower buildout means fewer chips. I don't have any quarrel with that logic.
But I think everyone is watching the wrong thing.
Nobody Asked the Power Company
Here's what's been on my mind since Saturday.
Whether these companies slow down on purpose is close to irrelevant — because something already slowed them down, and it happened months ago without a single press release.
They ran out of electricity.
Reuters went through utility and grid filings across the Midwest, the Mid-Atlantic and the South and added up what very large power users — overwhelmingly data centers — have formally requested.
More than 700 gigawatts.
That is over ten times what the entire American data center industry consumes today. It is roughly what it takes to keep the lights on in every home in this country.
Then look at Texas, which spent last year calling itself the epicenter of AI development.
On August 3rd the governor ordered a moratorium on new grid connections for data centers and told regulators to audit every project in the queue. That queue holds more than 1,800 projects asking for 474 gigawatts — five times the most electricity Texas has ever used at one time. Roughly nine out of ten of those requests are data centers.
Texas did not run out of enthusiasm for AI. It stopped taking applications.
And in PJM, the grid that serves 13 states and 67 million people, the price of guaranteeing future capacity went from $28.92 per megawatt-day to $329.17.
Eleven times, in two years. It hit the federal price cap.
Meanwhile 1,200 gigawatts of new power generation sits in PJM's own queue, waiting an average of five years just to be allowed to connect.
You Are Already Paying For This
If that all sounds abstract, it isn't. It's on your bill.
PJM's own market monitor attributed 63% of one auction's price increase directly to data center growth — $9.3 billion in extra capacity costs, passed through to households.
In practice that's about $18 a month more for the average home in western Maryland. Sixteen dollars a month in Ohio.
I'd gently suggest that is what actually decides how fast AI gets built in this country. Not an essay on a Saturday. A utility bill in an election year.
The Question Changed
For three years the only question in this business was who had the best chips.
That question is now settled enough to be boring. Here's the new one, and almost nobody has repriced for it yet:
Who has the electrons?
Consider that the same day the AI story broke, Brent crude pushed above $109 a barrel, after Saudi Arabia shut down the pipeline it built specifically to route oil around the Strait of Hormuz. The ten-year Treasury yield touched 5%, the highest since October of 2023.
Read those three facts together, because they're the same fact.
Energy is scarce and getting more expensive. Money is expensive. And the most capital-hungry construction project in the history of the world needs staggering amounts of both.
The bottleneck in artificial intelligence is not the chip. It's the substation.
Which is why what I've been reading about lately has nothing to do with Silicon Valley.
A Valley in Utah
There is a drilling crew in a valley in Utah that just reached a fuel source the oil industry spent fifty years trying and failing to get to.
Not another well in an old field. A different kind of energy — one that doesn't care what OPEC decides on a Thursday, doesn't care who controls the Strait of Hormuz, and cannot be blockaded by anybody, because it never has to cross an ocean.
It also runs around the clock, which is precisely what a data center requires and precisely what the grid is short of.
Google has written a check. So has Berkshire Hathaway. So has the Pentagon.
It's the Department of Energy's most ambitious drilling initiative since the program that set off the shale revolution — and shale, you'll remember, redrew the entire world's energy map in about ten years.
One company sits at the center of it. The whole thing goes public on October 20th, which gives you roughly five weeks.
I put together a briefing on what they found, why half a century of drilling couldn't reach it, and the company doing it.
The Bottom Line
They can slow the models down. They cannot slow down physics.
The scarce thing in artificial intelligence stopped being intelligence a while ago. It became electricity, and hardly anyone has moved their money to match.
The Federal Reserve announces its decision tomorrow at 2:00 PM, and it's worth your attention. But it won't drill a well, and it won't build a substation.
If you want to see where the electrons are going to come from — and the one company doing the drilling — I'd get through this before October 20th.
"The Buck Stops Here,"
Found this helpful? Share it with others.
Written by Dylan Jovine