AI’s next bottleneck? It’s the power bill

The real AI bottleneck might be the grid

Everyone loves talking about semiconductors as the brains of the AI boom. But this piece argues the bigger choke point may be something much less glamorous: electricity. The Kobeissi Letter projects global AI power demand could jump more than 1,100% by 2033, reaching roughly 315 gigawatts — and the U.S. is expected to shoulder about 200 gigawatts of that growth.

That’s not a small “incremental tailwind.” That’s a whole new layer of demand slapping an already-strained grid in the face.

Why investors should care

AI data centers are power hogs, especially during training runs when GPUs fire up in waves and electricity use can spike as much as 50% above design capacity. Translation: the companies that generate, move, and manage power may get a lot more attention than the ones just making the chips.

The article puts names like Constellation Energy, Vistra, GE Vernova, and NextEra in the spotlight. In other words, the market is starting to treat power infrastructure like the backstage crew that suddenly has to run the entire concert.

The hyperscalers are promising to pay up

The twist is that the big AI spenders aren’t just asking the grid to keep up and crossing their fingers. Amazon, Alphabet, Meta, Microsoft, and Oracle have signed Trump’s Ratepayer Protection Pledge, promising to build, bring, or buy new electricity and cover grid upgrade costs.

That matters because if the hyperscalers are on the hook for more of the infrastructure bill, the AI story stops being just about software margins and starts looking a lot more like a capital-intensive utility race.

Big picture

You can’t run the future on vibes and venture funding alone — you need megawatts. If AI demand keeps compounding like this, the next big winners may be the companies that keep the lights on, not just the ones training the models.