The Invisible Trillion · The Water Question, Honestly

Putting It Together

The honest summary of the water question looks like this.

Globally, chip fabs and data centers use a small fraction of human water — somewhere around 0.1–0.2% of the total. This is real but small compared to almost anything else (agriculture, lawns, golf courses, the cooling needs of power plants themselves).

Locally, in specific drought-affected regions — Taiwan, Arizona, parts of Spain, parts of Texas — the water use is a real problem that is worth paying attention to and regulating carefully. The question is less “should data centers exist” and more “what kind of cooling should they use, and in which watersheds.”

The technology is moving in the right direction (closed-loop, immersion, recycled water), but not fast enough that you should ignore the problem. It’s also not the case that every new data center is making things worse; some of them are genuinely well-designed and use essentially no fresh water.

The “bottle of water per query” meme is based on a real study, but that study’s numbers are mostly upstream power plant water, which drops to near-zero when the grid is clean. As grids decarbonize, the water footprint of AI drops with it — the same coupling we saw with energy in Chapter 5.

The companies’ disclosures are getting better but still aren’t good enough. “Water positive by 2030” is a promise, not a result, and promises are worth exactly what past promises from the same companies have been worth, which is a mixed record.

The Checklist, Again

When someone tells you the water cost of a technology, the questions to ask, in order:

1. What’s the unit? Gallons? Liters? Per day? Per year? Per query?
2. Withdrawn, consumed, or potable? These are different by a lot.
3. Where? A data center in Arizona is not a data center in Finland.
4. What kind of cooling? Evaporative, closed-loop, immersion?
5. Is the upstream power plant water included? Because that’s often the biggest piece, and it depends on the grid.
6. Compared to what? 500 gallons a day is almost nothing for a data center and almost everything for a house.

A number without answers to these questions is not an argument. It is a vibe. Vibes are not what you use to decide things.

A Personal Observation

The water chapter is, honestly, the chapter I find hardest to write, because the emotional valence of water is so strong.

Energy is abstract. Kilowatt-hours don’t evoke much. But water — water is what you drink, it’s what you shower in, it’s what keeps your body alive, it’s what goes into your dog’s bowl. When someone says a data center “drank” water, you viscerally feel like it took something from you. The metaphor is doing a lot of the work.

The reality is more mundane. Most data center water is returned to the local watershed, often cleaner than it came in (because it had to be filtered to ultrapure standards before use). The fraction that’s actually consumed — evaporated away — is small. And almost all of it is in specific places that are specific problems, not a diffuse global crisis.

I want you to be able to tell the difference between a visceral claim and a quantitative one. Visceral claims are easy to make, hard to argue with, and frequently wrong. Quantitative claims take more work to state, more work to understand, and more work to criticize — and they are what you need if you are going to make actual decisions about actual policy.

The data center in Arizona that is drawing down the aquifer is a real problem, and you should care about it. The ChatGPT query you sent this morning is not. These are different facts, and keeping them separate is most of what good thinking about this stuff consists of.

A note on sources: the TSMC water figures come from TSMC’s own sustainability reports, which are unusually detailed. The UC Riverside study is Li et al. (2023), “Making AI Less ‘Thirsty.’” Data center industry water figures come from the Lawrence Berkeley National Laboratory’s data center reports and from the hyperscalers’ own annual sustainability reports. The agricultural water comparisons are from the Water Footprint Network, which is the standard reference. When figures conflicted — and they often did — I used the midpoint of credible estimates and flagged the uncertainty.

All 19 chapters of The Invisible Trillion · Library · All courses