The word "fake" draws attention but I think the article obscures two real problems:

Training is missing from the analysis entirely (as someone else noted)

Inference water use is indeed minimal per prompt no argument there, but training the old GPT-3 consumed roughly 5.4 million liters of water. LLaMA 3: ~22 million. These are huge events, happening multiple times a year across the industry, folding them into national averages seems like the statistical simplification he article criticizes everyone else for doing…

"Small nationally" ≠ "fine locally"

The Dalles, Oregon is the clearest example. In 2012, Google used 12% of the city's water supply. Today it consumes a third, around 1.19 million gallons per day, and well a sixth data center comes online in 2026, in the same area.

The city is now pursuing a $260 million reservoir expansion into a national forest (!), where 95% of the projected new water demand will be industrial, not residential. Residents are looking at a potential 99% rate increase by 2036 to fund infrastructure that may exists primarily to serve one company. Apparently the city fought a 13-month legal battle just to keep those numbers secret, that’s like a community being reshaped around a single tenant.

Hays County, Texas residents sharing the Edwards Aquifer with incoming data centers voted to block one. Memphis is watching xAI draw 5 million gallons per day. Bloomberg found two-thirds of new U.S. data centers since 2022 are sited in high water-stress zones. Arizona have already passed ordinances capping data center water use.

This to me looks like a problem in the making, AI water use isn't a national crisis for now, but local impacts are already real, training costs are systematically underreported, and the five year trajectory in water stressed regions deserves serious attention indeed

There is so much nuance and context missing that I can't see this as anything other than astroturfing.

I can get behind "AI water use is not a serious concern" if all you are talking about is selling inference, and you're comparing some sort of usage metric (e.g. "water use per request"). Water and power use for inference is on the level of other heavy Internet products like video streaming or cloud compute.

There is a lot I can't ignore, though. Model training is incredibly demanding, so much that OpenAI was trying to get $1 trillion in investment to practically double the number of data centers in the United States by 2030. That is a serious concern when we have to make decisions between, say, consumer water availability and tech investment in water-scarce areas like Arizona and New Mexico.

In Oregon, there are some unique problems with Amazon's water deals in Umatilla, where they are increasing nitrate concentration of the local groundwater through evaporative cooling, and refusing to pay for on-site treatment.

I can go on about other environmental harms, but I think you should take a more nuanced look at the issue. Having ChatGPT summarize a news article is not an unreasonable demand compared to other compute activities, but AI in general is driving compute demand so high that the general public is forced to reckon with a problem that's been there since the beginning: the expansion, operation and use of the Internet has physical environmental consequences.

Okay. There are other criticisms of datacenter buildout that make this kind of product valuable. Moving on.