Local vs National: Where the Real Story Is

This is the single place where the "AI water myths are overblown" thesis needs to do nuance, not bombast. The national picture is unambiguous — AI is a rounding error on US freshwater. The local picture in a handful of regions is genuinely contested and worth taking seriously.

The national picture (overwhelming)

Metric	Value	Source
US daily freshwater consumption (all uses)	~132 billion gal	USGS
Agriculture share	~70–80%	USGS
Thermoelectric power share (consumption, not withdrawal)	~3%	USGS / Macknick
All US data centers	~0.2%	derived from EESI
AI workloads specifically	~0.008%	Masley

At national scale this is over before it starts. Doubling AI water consumption — which would require something like a 25× increase from today — would still leave it below the noise floor of agricultural variability year over year.

The local picture (real, sometimes serious)

Roughly 40% of US data centers sit in regions classified as "high" or "extreme" water stress by WRI Aqueduct, per Lincoln Institute. The top three concentrations:

Phoenix metro (Arizona)

~60 data centers (Stanford & The West)
~177 million gal/day of water consumed (Stanford)
Local agriculture share: ~86% of total water consumption (Stanford)
Headline: data centers are a much smaller user than alfalfa, but they are a visible and new user, and the politics is loud
Several local moratoria proposed; some passed in suburban municipalities

Newton County, Georgia (Atlanta exurb)

Meta's massive build-out triggered well-water failures for nearby residents, primarily during construction (sediment runoff, not operational draw) (NYT via Spokesman)
Operational water draw ~500K gal/day = ~10% of county consumption (PPC.land)
The political damage from the construction phase is lasting

Texas (Austin / DFW corridor + West Texas)

HARC / U Houston projection: ~49 B gal/year by 2025; 29–161 B gal/year by 2030 depending on scenario (HARC)
The high end of the 2030 range is roughly equivalent to one mid-sized Texas city
Texas has both abundant gas-fired generation (water-intensive) and serious aquifer drawdown — the worst combination for scope-2 water concerns

Northern Virginia ("Data Center Alley")

~35% of the world's hyperscale capacity sits here (Lincoln Institute)
Water consumption is moderate (climate is favourable for evaporative cooling); the concentration concern is electrical, not water
Useful counterpoint: large data center footprint, water is not the controversy here — which itself supports the "water issue is regional, not technological" framing

The honest framing

The legitimate water criticism of AI is not "your ChatGPT use is killing the planet." It is:

"Operators have an incentive to site near cheap power, which historically means near coal in arid Western states or gas in Texas, and when they do, they should pay full freight for water and submit to community-level permitting like any other heavy industrial user. The current setup often gives them sweetheart utility deals and minimal local accountability."

That is a real critique. It does not require any of the inflated per-query numbers to land. The editorial should make this concession explicitly — and then point out that the per-query / national-scale alarmism actively gets in the way of the legitimate local-policy conversation, because it crowds out the discussion that actually matters with one that is mathematically incoherent.

What good policy looks like (for the editorial's "what now" beat)

Locational accounting. Data centers should report water use per facility, not company-wide averages. Ren et al.'s central policy ask.
Pay full retail. End discounted municipal water rates for hyperscalers in stressed basins.
Tie permitting to siting class. Treat new builds in WRI "extreme stress" basins under stricter rules than ones in Quincy, WA or Iceland.
Disclose scope-2. Companies report on-site water; force them to report power-plant water proportional to their PPAs.
Decarbonise the grid. The single biggest water-saving lever, and it's not one that data center operators control — but they are some of the largest renewable PPA buyers in the world, which should be acknowledged.

None of these require believing that ChatGPT drinks a bottle of water per query. All of them are improvements over the status quo.

What the editorial should not do

Frame Phoenix's water issues as "agriculture's fault and not data centers'." That deflection has been used by both industries against each other and is unhelpful — both are real users and both are contested.
Pretend construction-phase impacts don't count. The Newton County episode is real and matters, even though it doesn't fit the "operational water per query" narrative.