Liquid Cooling in AI Data Center

Our Take

MEP Academy lays out the plumbing side of the AI boom clearly: air cooling tops out somewhere around 15-20 kW a rack, and today's GPU clusters are blowing past 100 kW, so the industry is racing toward direct-to-chip cold plates and immersion tanks just to keep hardware from cooking itself. It's a solid technical walkthrough, credit to the channel for making CDUs and two-phase immersion legible to non-engineers.

But every kilowatt these systems are built to move is a kilowatt someone nearby is paying to generate, and every liquid loop is a water and chemical supply chain that has to come from somewhere. Direct-to-chip and immersion cooling can cut on-site water use compared to old evaporative towers, sure, but they also lock in enormous, concentrated power draws that utilities are now asking ratepayers and taxpayers to help finance through new substations, transmission upgrades, and rate structures nobody voted on directly. The engineering is getting more efficient per rack; the community-level footprint is still growing.

Densification like this is exactly why we track where these facilities land and what deals got them there — check our facility map to see what's proposed or under construction near you, and if a project's headed your way, our take-action page has the questions worth asking your local utility board before the permits get signed.