A few people asked over the weekend what the calls to slow the pace of frontier AI mean for the buildout. Some thoughts: Even if models never improved from here, just rolling out what they can already do would take more compute than the world can build for years. The debate about how fast AI should be allowed to improve is a fair one to have. It's about future generations of models. Existing demand is the part I think people are misreading. Anthropic CEO Dario Amodei said in May they'd planned for 10x growth and were running at an 80x pace in the first quarter, and that's why they've had trouble supplying compute. OpenAI president Greg Brockman said in July they'll be in a compute shortage no matter what, and are choosing which products to scale. Google says it's processing 7x the tokens it did a year ago. Hundreds of millions of people use these models today, and most of them use a small part of what the models can already do. And every time more compute comes online, usage steps up again: limits come off for people already using it, customers who were turned away get on, and new use cases show up that weren't in anyone's plan. Now supply. The constraint is HBM, the memory that sits inside every major AI chip. Three companies make it and all three are sold out this year. A new memory plant takes years to build. TrendForce has HBM shipments growing 50-60% next year. NVIDIA, the biggest buyer of it, expects its revenue to grow about 70% next year and calls that outlook 'supply-constrained', noting its customers' forecasts point closer to 100%. On our own bottom-up work, the memory constraint lands in about the same place as NVIDIA's growth number. Then the chips need a building with power connected, which takes longer again. Goldmans reckons only about half the US capacity scheduled over the next two years will actually be built on time. The risk to demand continues to seem heavily weighted to the upside. The risk to supply continues to seem weighted toward less capacity getting built, not more.