Legitimate Concerns (The Steelman)

The editorial has to make this argument because the discourse has consistently failed to. "AI water use is overblown" gets immediately straw-manned as "AI water use is zero," and the conversation collapses. Here is the strongest version of the critic's case — the version a serious environmental policy researcher would actually advance.

1. Construction-phase impacts are real and chronically under-reported

The most concrete water-related harm directly attributable to a hyperscale data center is the Meta build in Newton County, Georgia. Sediment runoff during site preparation contaminated nearby private wells and rendered them unusable for an extended period (NYT via Spokesman). Operational water draw was a side issue; the construction-phase damage was the actual injury.

This is not unique to data centers — any large industrial site has the same risk profile — but it doesn't fit the "per-query" frame at all. The editorial should not pretend it doesn't exist.

2. Local stress is genuinely contested in arid siting

Phoenix data centers consume ~177 M gal/day (Stanford & The West). That's a small fraction of regional totals — but the regional totals are themselves unsustainable. Aquifer drawdown in the Lower Colorado Basin is a multi-decade crisis; adding any new heavy industrial water user to that basin makes the politics harder, regardless of how their per-unit consumption compares to almonds or alfalfa.

The fair version of the critic's argument here is not "this 0.2% will break the basin" but "we are already over-allocated; everyone needs to pull back, and adding new users in either direction is the wrong sign."

3. Scope-2 disclosure is genuinely poor

Most large operators publish on-site water use with high granularity. Almost none publish water consumed at the power plants generating their electricity. As covered in scope_2_water.md, scope-2 is the dominant share. So the publicly disclosed numbers systematically under-state by 4–10×.

Ren et al.'s central policy ask is for better disclosure, not for AI to use less water. (OECD.AI makes the same point.) That is a legitimate, modest, science-supported request. The editorial should endorse it explicitly.

4. Sweetheart utility deals are a real policy failure

Data centers in many jurisdictions get municipal water at heavily discounted rates — sometimes 30–60% below industrial standard tariffs — as part of the package economic-development authorities offer to attract them. In stressed basins, that pricing distortion is genuinely indefensible. Operators should pay full freight; the cost is small relative to capex and operationally trivial relative to GPU spend.

This is one place where the critics and the steelman defenders converge: nobody serious thinks Microsoft should be paying $0.40/kgal for water in the Phoenix metro.

5. Total water use is growing fast

The 0.008% national share is today (Masley). If hyperscale buildout continues at projected rates (Texas: ~49 B gal/year today scaling to a 29–161 B gal/year range by 2030 — i.e. ~0.6× to ~3× in five years depending on scenario; HARC / U Houston projection), the national share rises from 0.008% to ~0.08% by 2030 (Masley) and could plausibly reach 0.5% by 2035 in aggressive scenarios.

0.5% is still small. But it is no longer trivially small. And growth rates of 50–80% per year compound; saying "it's tiny today" without acknowledging the trajectory is intellectually weak.

The honest framing: AI water use is small and growing fast. Both are true. The editorial's case is that even at 5–10× growth from here, it remains far below the comparison footprints — but the trajectory is worth flagging.

6. Compute scaling vs efficiency scaling is a race

Per-query water consumption has dropped roughly 100× over four years (GPT-3 / 2022 → GPT-4o / 2024 → modern stack / 2026). But total query volume has grown by something like 1,000–10,000×. Net total has gone up, not down.

This is the AI-environmental version of Jevons' paradox: efficiency gains get plowed back into more compute. Critics correctly note that "per-query" numbers improving doesn't tell you what you want to know about aggregate impact.

The defender's response is that aggregate impact is still small in absolute terms, and that it is being driven by demand for a useful service — not by waste — and that this is true of every other industrial efficiency story.

What the strongest critic believes

"AI water use isn't a planetary catastrophe. It's a regional permitting problem with poor disclosure, growing fast, in a country that has not figured out how to govern utility-scale industrial siting in stressed basins. Per-query numbers are the wrong frame. Aggregate growth in specific basins under specific regulatory regimes is the right frame. The viral 500 mL claim is wrong, but using its wrongness to dismiss the entire concern is also wrong."

The editorial should be able to nod at every sentence of that.

What would change the editorial's mind

If credible numbers showed that:

Per-query consumption had stopped falling, and
Query growth was sustained at 5–10× per year, and
New buildout was concentrated in WRI "extreme stress" basins despite policy efforts, and
Scope-2 was not falling fast as the grid decarbonised

…then "AI water use is overblown" would no longer be defensible by the early 2030s. None of those conditions hold today, and the trends on at least three of the four are favourable. But the editorial should commit to revisiting the claim if they shift.