My boss is always telling me to think about demand and supply (cc'ing @santiagoroel). So I did that today, and thought about an interesting trade. I'm probably wrong, but I think it makes sense? THE WATER BOTTLENECK 🧊 TL;DR: The hyperscalers are building $200B+/year of AI infrastructure on top of a physically finite resource that's governed by laws written in 1902. The market cap ratio between water infrastructure and AI chips is 100:1. When voters realize data centers consume more water than their entire town, the political hammer falls. The trade: Long PHO/PHO/ PHO/AWK/$XYL. Catalyst: Colorado River negotiations conclude Dec-26 I. THE SETUP: Everyone's Betting on the Wrong Bottleneck. The consensus AI infrastructure trade has been painfully obvious: semiconductors (already ran 200%+), power grid buildout (already consensus), uranium (up 69% YTD). Everyone's solving for electricity.Nobody's solving for what the market is missing: you cannot train GPT-5 without water. Not metaphorically. Literally. The H100 clusters generating all this "intelligence" produce so much heat that evaporative cooling is the only economically viable solution at scale. And evaporative cooling has a dirty secret—the water evaporates. It's gone. Into the atmosphere. Forever. The market is pricing AI infrastructure as if we have infinite water. We don't. Western US water supply is capped by snowpack in the Rockies and aquifer recharge rates that haven't changed since the Pleistocene. Meanwhile, demand is growing parabolically. This is the trade. II. THE MATH MATHING: Direct + Indirect Consumption Most analysis stops at "data centers use X gallons for cooling." That's not the full picture. The actual formula is: Total Water Impact = Direct Cooling (On-site) + Indirect Generation (Off-site) The Direct (On-site): Modern AI facilities: 1.8-12 liters per kWh for evaporative cooling "Typical" hyperscale cluster: 15 million gallons/day (5.4B gallons/year) Water Equivalence: Same as 50,000 new homes The Indirect (Off-site) — This Is What Everyone Misses: The thermoelectric power plants (coal, gas, nuclear) generating the 24/7 baseload power for these facilities use ~10x more water than the cooling systems themselves for steam generation and condenser cooling. When you run the full stack: 2030 US AI Infrastructure Projection: Total water consumption: 1.1-1.7 TRILLION gallons annually Context: Equal to every household in California combined Let me repeat that. By 2030, AI infrastructure will consume the same amount of water as the entire state of California's residential use. Nvidia can print more chips. Utilities can build more generation capacity. But you cannot VC-fund more rainfall. The Colorado River flow is determined by Rocky Mountain snowpack, and last time I checked, there's no Series B for the hydrological cycle. ...