John Mac Ghlionn calls the AI data center boom a run of "genuine environmental monstrosities" in a July 11 opinion column for The Hill. The buildings drain local water. They leave behind an automated skeleton crew. His argument runs through labor economics: electricians pulling five-figure monthly paychecks to wire facilities that will run on a handful of people once commissioning ends. The line that matters to this desk is the one he never breaks down. Cooling.
Mac Ghlionn writes that these buildings need billions of gallons of water to keep servers from overheating, often in states already short on it. That claim holds for one architecture only: open-loop evaporative cooling towers, the method hyperscalers keep specifying because it is cheap, proven, and rated for gigawatt campuses. It is also the reason U.S. data centers pulled roughly 17 billion gallons from public and groundwater supplies in a recent accounting, a figure that keeps climbing as training clusters push past 100 megawatts. Direct-to-chip liquid loops and closed-circuit chillers cut consumptive draw by an order of magnitude, at higher upfront cost, and most operators fighting for a permit right now keep the towers running rather than retrofit mid-fight. The backlash Mac Ghlionn describes, drought-state residents watching water disappear into a windowless building, traces back to that procurement choice. Most of the volume never touches a server rack at all: 72% of data center water consumption happens off site, at the power plant supplying the electricity the chillers and cooling towers burn through.
The detail Mac Ghlionn treats as an aside, an automated site staffed by a skeleton crew of fifty, is the harder engineering problem. A facility that goes from a construction site crawling with electricians to a nearly unmanned data hall inside months has to get its cooling distribution units, leak detection, and redundancy paths right before the last trade truck leaves. There is no bench of on-site engineers left to catch a mistake. Cooling is already the single greatest risk in commissioning a 100 MW AI data center, and that risk does not shrink once the ribbon gets cut. It moves onto whoever is left running the building remotely.
That headcount is also why remote monitoring and predictive maintenance on liquid loops have turned into procurement requirements instead of add-ons. A clogged manifold or a slow leak with nobody walking the floor turns into multi-day downtime, the kind a staffed facility catches in hours. Mac Ghlionn counts fifty people running a gigawatt-scale campus as proof of automation eating jobs. Engineers on this beat read the same number as a redundancy spec.
Mac Ghlionn frames the debate as a false choice: hand water and grid capacity to domestic tech monopolies, or cede ground to Beijing's algorithms. That framing skips the decision actually sitting on an engineer's desk right now: which cooling method gets specified before the concrete pours. Evaporative towers pull in the drought headlines. Closed-loop and direct-to-chip liquid systems pull in the permits. The Hill column reads like a labor and monopoly story. Underneath it, and under every drought-state hearing it will feed, sits a spec sheet decision nobody in that piece names.