A new piece making rounds on Hacker News pushes back hard on the narrative that AI data centers are guzzling water at civilization-threatening rates. Water researcher Jay Lund ran the numbers for California and found that all the state’s data centers combined consume somewhere between 0.055% and 0.7% of annual human water use. That’s not a typo. Less than one percent.
Lund didn’t just guess. He started from physics, calculating heat dissipation across California’s roughly 15 million square feet of data center floor space, then sanity-checked his work against four AI models (ChatGPT, Claude, Gemini, and Copilot). The estimates landed in a wide but consistent zone: 20,000 to 290,000 acre-feet per year against California’s total human water use of about 40 million acre-feet.
His closing line is the one to remember. Lund jokes that his breathing while writing the post may have evaporated more water than the incremental impact of running those four AI queries.
Why the panic doesn’t match the data
The reporting gap is real. AI companies don’t publish detailed water and energy figures, and Lund acknowledges that’s a legitimate transparency problem driven by competitive secrecy. But what’s filling that vacuum, he argues, is speculation dressed up as analysis. Journalists, academics, and advocates have used the data void to spin worst-case scenarios that don’t survive contact with basic thermodynamics.
Here’s what stands out: every one of the four AI models, when asked to estimate California data center evaporation, converged on ranges that overlap with Lund’s physics-based math. ChatGPT said 20 to 400 thousand acre-feet. Claude said 14 to 21. Gemini said 2 to 40. Copilot said 30 to 50. The narrower consensus zone sits around 20,000 acre-feet annually.
For scale, that’s enough water to irrigate 10,000 to 100,000 acres out of California’s 7 million irrigated acres. A rounding error in agricultural terms.
What’s actually changing
The industry context matters. Data center buildout is accelerating, energy grids are straining, and local communities are pushing back on new sites. The water angle has become a rallying point partly because it’s tangible in a way that megawatts and FLOPS aren’t. People understand a swimming pool. They don’t understand a gigawatt.
But the geography is critical. Lund is careful to note that California is not the country. States with heavier data center concentration and weaker water infrastructure (think Virginia, Arizona, parts of Texas) face genuinely different math. The headline finding here is location-specific, not a blanket dismissal.
He also flags an underrated upside: in regions where urban water demand is dropping due to conservation, data centers can become paying customers for excess capacity. That reframes the story from “AI is stealing water” to “AI is buying surplus water from cities that need the revenue.”
What practitioners and operators should take away
A few practical moves:
- If you’re building or siting infrastructure, push hard on transparent water and energy reporting. The competitive secrecy argument is wearing thin, and voluntary disclosure beats regulated disclosure every time.
- If you’re communicating about AI, ditch the speculation and run the physics. Lund showed how trivial the calculation actually is once you start from heat dissipation and cooling efficiency.
- If you’re a journalist or analyst, treat “AI is destroying water supplies” claims with the same skepticism you’d apply to any other industry-scale assertion. Ask for the math.
- If you operate data centers, the efficiency range Lund cites (60 to 90%) is wide. The companies on the high end of that range have a real story to tell. Tell it.
The broader pattern is familiar. New technology arrives, fears outrun facts, advocates seize the attention window, and eventually the numbers catch up. Lund’s piece is one of the first attempts to do that catching up with rigor instead of vibes.
More detail, including the full calculations and source links, sits in the original Hacker News discussion.