The Invisible Trillion · The Water Question, Honestly
Fresh Water vs. Withdrawn vs. Consumed
One more technical distinction, and then we’ll be done with the unpleasant part.
When people talk about “water use,” they often conflate three different things:
Withdrawn water is water taken from a river, lake, or aquifer. Most of it is returned to the source (sometimes warmer, sometimes dirtier, but returned). This is the biggest number when you see industrial water figures.
Consumed water is water that is actually used up — evaporated, absorbed into products, or otherwise removed from the local water system. This is typically 10–20% of withdrawn water for industrial uses, and it’s the number that actually matters for whether the local water table is being drawn down.
Potable fresh water is water that was suitable for drinking before it was used. A data center that uses drinking water is competing with human and agricultural needs. A data center that uses recycled wastewater or seawater or reclaimed greywater is not.
When a data center reports “5 million gallons of water used per day,” it might mean 5 million gallons withdrawn (most of which goes back to the river), or 5 million gallons consumed (which doesn’t), or 5 million gallons of drinking water (which competes with human use) or 5 million gallons of reclaimed water (which doesn’t). These are all very different, and the headlines almost always fail to distinguish them.
When you read a scary number about data center water use, the question to ask is: withdrawn or consumed? Fresh or reclaimed? In what watershed? If the article doesn’t answer those questions, the article doesn’t actually know what it’s claiming.
The Technologies That Change the Answer
The water story is not static. Several things are changing fast enough that the numbers in this chapter might be outdated by the time you read this book, and mostly in the direction of lower water use.
Closed-loop cooling. Newer data centers increasingly use systems where the cooling water circulates continuously through a closed loop and is cooled by contact with outside air, rather than evaporating into the atmosphere. These systems use a tiny fraction of the water of evaporative systems — sometimes less than 1% — but they use more electricity, so they’re most efficient in cool climates. Microsoft’s data centers in Ireland and Sweden use essentially no water for cooling. The tradeoff is real (more electricity vs. less water), and the right choice depends on the local grid and the local water situation.
Immersion cooling. A newer approach where servers are literally submerged in a liquid that isn’t water — a dielectric fluid that conducts heat but not electricity. Zero water consumption. Higher capital cost, but catching on fast, especially for AI workloads that run very hot.
Reclaimed and non-potable water. Several data center operators now use recycled wastewater or seawater for cooling, which means they’re not competing with drinking water supplies at all. Google’s data center in Douglas County, Georgia, runs on treated sewage. This is weirder to think about than it sounds, and it is also obviously the right answer in a lot of places.
Water recycling in fabs. TSMC now recycles over 85% of the water it uses, up from around 30% a decade ago. Newer fabs are being built with even higher recycling rates. The chip industry’s per-chip water use has dropped significantly over the last 15 years, even as total chip production has gone up.
None of this erases the concerns, but it does mean that the direction of travel is better than it used to be, and the “water use will spiral out of control” scenarios assume a continuation of old practices rather than current best ones.
All 19 chapters of The Invisible Trillion · Library · All courses