Cooling choices, not slogans
Evaporative cooling, closed-loop dry cooling, and direct-to-chip liquid cooling are not the same water story. This page separates them and gives a recommendation.
A large AI data centre turns electricity into heat. Liquid cooling by itself does not tell us how much water the site consumes. What matters is how much water the facility loses every day while rejecting that heat.
There are two separate design choices: how heat is picked up inside the server, and how that heat is rejected outside. Direct-to-chip liquid cooling can be part of a very low-water design if the outside heat rejection is dry. A wet cooling tower can use a lot of water even if the inside of the building is well engineered.
These are simplified buckets for public review. Real engineering drawings may combine them, but the water-loss question still lands in the same place.
This is the traditional cheap-and-efficient answer in a lot of large cooling designs. It works. That is not the issue.
The issue is that it works by deliberately losing water to the air. At gigawatt AI scale, that becomes a public water issue.
Water loss: very high. For a 1 GW full-wet thought experiment, evaporation alone is roughly 35 million litres per day before site-specific factors.
The cooling fluid stays in a sealed loop. Heat is moved outside and rejected to air through dry coolers, basically industrial radiators with fans.
Water still exists in the system, but it is not routinely being consumed to reject heat.
Water loss: very low. Mostly initial fill, maintenance, treatment, repairs, and leaks.
This moves liquid right to the hottest parts of the server: GPUs, CPUs, and high-density AI hardware. It is the best fit for dense AI racks.
It should still be paired with closed-loop dry heat rejection if the goal is strong cooling with low water loss.
Water loss: very low, with better cooling performance for high-density AI loads.
| Cooling approach | Water-loss profile | Efficiency / practicality | Recommendation |
|---|---|---|---|
| Evaporative cooling | Very high. Water is intentionally lost as part of normal operation. | Thermally efficient and often cheaper, which is why it has been common in many U.S. builds. | Reject as the default for large Alberta AI data centres. Do not approve it on vague “efficient cooling” language. |
| Closed-loop + dry coolers | Very low. Losses should be maintenance/leak related, not daily heat rejection. | Solid low-water option, but needs enough dry-cooler capacity and honest hot-weather design. | Acceptable. Publish the loop design, expected losses, and post-commissioning numbers. |
| Direct-to-chip + dry heat rejection | Very low. Same low-water principle, better heat pickup at the server. | Best fit for modern AI density if engineered properly. | Recommended. This is the cleanest answer to “cool the servers without wasting water.” |
Evaporative cooling is not technically stupid. It is efficient, proven, and often cheaper. That is exactly why it has been attractive in many large U.S. data centre builds.
But Alberta should not treat “cheap and efficient” as the whole answer. If a project relies on routine evaporative water loss, the proponent should have to publish the full water balance: source, daily makeup demand, seasonal peak demand, discharge, drought plan, and public reporting.
For gigawatt-scale AI, the better public standard is simpler: do not build the normal cooling plan around throwing water into the sky.