
Meta’s newest AI plumbing explainer makes a persuasive case for closed-loop liquid cooling. Coolant runs across chips, carries heat to an exchanger and returns to do the trip again, rather than being continually consumed. That is sensible engineering for racks whose heat density is making traditional air cooling look increasingly like a desk fan pointed at a furnace.
Meta AI Watch analysis
The useful bit of the loop
Meta says its direct-to-chip systems use a water and glycol mixture that can stay in service for as long as a decade. Newer AI sites pair that loop with dry coolers, while some existing campuses use air-assisted liquid cooling to retrofit denser hardware. The company says removing large fans from the compute trays can nearly double rack density.
The same infrastructure now has software watching the plumbing. Meta reports that a reinforcement-learning simulator reduced supply-fan energy by 20 per cent and water use by 4 per cent in a pilot. Those are first-party results, not an independent audit, but they show why cooling controls are becoming part of the AI stack rather than mere facilities housekeeping.
Closed loop does not mean consequence-free
The slippery phrase is “water use”. Recirculating coolant can sharply reduce operational cooling water, and Meta says a typical closed-loop design uses less water in a year than two full-service restaurants. The explainer does not publish the comparison’s methodology, climate assumptions, rack density or site-level water balance. Meta’s own data-centre material carefully distinguishes cooling water from other site uses such as cleaning, domestic supply and fire systems.
Electricity matters too. Pumps, heat exchangers, fans and dry coolers all consume power, while the local climate decides how hard heat rejection must work. The engineering may still be substantially better than evaporative cooling, but the public proof needs whole-site water and energy figures per unit of useful compute, not one memorable restaurant comparison wandering around without its receipts.
The number worth watching
Meta has helped push open rack and liquid-cooling designs through the Open Compute ecosystem for years. Independent reporting from the 2022 OCP Summit documented the direction long before today’s AI build-out reached its current scale. The next useful step is comparable reporting across live sites: climate-adjusted water effectiveness, total cooling energy, compute delivered and heat reuse where available.
Closed-loop cooling looks like real progress. Publishing the whole loop, including the boring meters outside the rack, would make the sustainability claim as inspectable as the plumbing.
Sources and evidence
- Closed-Loop Cooling Explained: The Plumbing Behind Meta’s AI (27 August 2026, 21:44 UTC)
- Meta Data Centers: Innovation
- Meta shows AI system and liquid-cooled racks at OCP Summit (19 October 2022)
Meta AI Watch is an independent WittyWires-operated tracker and is not affiliated with, endorsed by, or operated by Meta.