I think some of the biggest uses of AI will come from work we currently don't bother doing. Too expensive. Not enough time. A potentially useful idea, but nobody available to investigate it. I recognize this from working in finance. You finish a reporting cycle knowing there are questions you would have liked to spend another day on. Then the next deadline arrives. You deal with what needs doing and leave the rest, even though a closer look might have been worthwhile. I wonder how that changes when investigating those questions costs very little. We would still need to check the work and take responsibility for the conclusions. But we could afford to ask more questions, follow up on smaller discrepancies and revisit assumptions that otherwise sit untouched for another year. The Jevons paradox argument is familiar by now. What interests me is what businesses actually choose to do with the lower cost. Reviewing every transaction instead of a sample, for example, or investigating a problem that was previously too small to justify someone's time. My bet is that there is a lot of demand we cannot measure yet because doing the work has never made financial sense. Reading the Anthropic Institute's paper on economic scenarios for AI brought me back to this. Its extreme scenario is extraordinary, but I found the middle one interesting enough: a substantially larger economy while most work still happens without AI. These are scenarios, of course. I wouldn't build a portfolio around a particular GDP number for 2030. What I take from it is that AI doesn't have to replace entire professions to have a very large economic effect. Then there is research, which is probably where my imagination runs furthest ahead of what I can put into a financial model. What could an engineering team attempt if it could explore many more designs before choosing which ones to build? How much could a small research team accomplish with analytical resources it could never previously afford? Physical testing would still take time and money. Plenty of ideas would fail. I just suspect we would attempt more, and eventually some of those attempts would produce things worth having. This is a big part of why I keep coming back to AI infrastructure. Broadcom and Marvell, memory, networking, power and cooling. We research these businesses separately, but they are serving different parts of the same system. If AI takes on more useful work across the economy, that system has to support it. The difficult part is figuring out who earns attractive returns from supplying it. Customers can get tremendous value from AI while an infrastructure shareholder earns very little. Our Flex analysis is a recent example of why I can like the business and still want a better entry price. But I remain extremely bullish on the broader opportunity. I suspect that, a few years from now, businesses will routinely do work they wouldn't even consider starting today. I'd rather judge the infrastructure opportunity against that possibility than assume today's chatbot habits tell us how much capacity we will eventually need.