Everyone keeps asking: "Won't better chips solve the power problem?" No. And here's why. Chip efficiency is improving. That's real. NVIDIA's chips improved performance per watt by about 20%. Liquid cooling is reducing energy waste. Better architecture is squeezing more compute out of every kilowatt. But here's what people miss: AI compute demand is growing faster than efficiency improvements. Much faster. I've seen forecasts where AI compute demand doubles every 3 months. And I believe it... Chip efficiency is improving maybe 20-30% per year. Do the math: If demand doubles every 3 months, that's 16x growth in a year. If efficiency improves 30% in a year, you're using 0.7x the power per computation. 16x demand × 0.7x efficiency = 11x more power needed. Efficiency helps. But it doesn't come close to keeping up. This is why Google is betting $40 billion on Texas transmission upgrades. This is why Microsoft is restarting Three Mile Island. This is why Anthropic is spending $50 billion on infrastructure. They've done the math. And the math says: chip efficiency buys you time, but it doesn't solve the problem. The only solution is more power infrastructure. A lot more. So when someone tells you "better chips will reduce power demand," they're technically right. But they're practically wrong. Because the demand curve is steeper than the efficiency curve. And that gap is the entire power infrastructure build-out for the next decade. If you're betting on chip efficiency to slow down power demand, you're going to lose. The hyperscalers aren't betting on it. They're building power infrastructure as fast as they can. That should tell you everything you need to know.
That 11x math really puts it in perspective. Makes you wonder how tech professionals should be weighing power infrastructure access when evaluating companies to work for, since that gap is going to separate who can actually compete in AI versus who's just claiming they will.
Lachlan Trinh
6mo
16x demand versus 0.7x efficiency is honestly brutal math. We're optimizing the car while the highway is on fire.
Ryan H. Vaughn
6mo
The efficiency trap is so seductive because it feels like something you can control and optimize. Meanwhile the actual constraint is sitting there screaming at you, but nobody wants to look at it because the infrastructure build is messy, slow, and unsexy.
Randy Hill
6mo
So basically the companies moving slowest on infrastructure are already out of the race. Who can actually afford to keep up with that demand curve besides the hyperscalers?
This is Jevons' Paradox in real-time, Ryne Ogren. The real insight here: efficiency gains don't reduce absolute power consumption when they make a technology economically accessible to exponentially more use cases. Every efficiency breakthrough unlocks new applications, new workloads, new demand.