Every technology boom eventually gets a villain the public can picture. For the railroads, it was land grabs. For artificial intelligence (AI), it became a data center quietly drinking a town’s water while a chatbot wrote someone’s emails.
That picture is not imaginary. Massive computing warehouses run hot, and the cheapest way to keep thousands of chips from cooking themselves has long been water, pumped through cooling loops and often evaporated straight into the air.
As the AI race accelerated, local fights over wells, aquifers, and permits spread from Arizona to Ireland, and “thirsty AI” became shorthand for everything people distrust about the buildout. Google (GOOGL) and Amazon (AMZN) have spent recent weeks defending their cooling to skeptical communities, and the industry’s answer has mostly been the same promise that efficiency will eventually shrink the footprint.
I have watched that narrative harden for two years, which is why a claim out of London this week caught my attention. Nvidia (NVDA), whose chips sit at the center of nearly every AI data center, now says the water problem is close to fixed.
Why data centers got so thirsty
Cooling is the hidden cost of computing. Data centers use water two ways, directly to carry heat off the chips, and indirectly through the power plants that feed them electricity.
The direct number is already large, and the indirect number is larger.
- U.S. data centers consumed about 17 billion gallons of water for cooling in 2023, a total on pace to double or even quadruple by 2028, according to Lawrence Berkeley National Laboratory.
- The power plants feeding those data centers used roughly 12 times more water than the cooling itself, the same Lawrence Berkeley National Laboratory report found.
- Training the GPT-3 model consumed about 700,000 liters of fresh water on-site, according to researchers at the University of California, Riverside.
Numbers that big are hard to feel, so here is the translation. That 17 billion gallons is roughly the yearly water use of 160,000 American homes, and it is the floor, not the ceiling, of where this is heading.
Related: Nvidia’s $25B bond deal sends investors a clear signal
When I lined those figures up, the takeaway was simple. The water story was never really about your chatbot. It was about thousands of buildings, each one hot, each one local, each one fighting a city council somewhere over a permit.
It also comes down to a brutal trade-off baked into how these buildings stay cool. Evaporative systems spray water to shed heat and lose much of it into the air for good. Closed-loop systems recycle the water but burn far more electricity running chillers to cool it back down.
Operators have spent years choosing which resource to waste. That is the choice Nvidia says it can erase.
What Nvidia actually changed
Nvidia’s pitch is mechanical, not magical. Its newest AI systems are designed for liquid cooling that still works when the liquid is warm, using a recirculated mix of water and propylene glycol, similar to automotive antifreeze, that can run at about 113 degrees Fahrenheit, according to Axios.
That temperature is the whole point. Because the coolant can stay hot, a data center can lean less on the energy-hungry chillers that themselves burn power and evaporate water, or skip them entirely. “The water consumption challenge for data centers is largely solved,” said Josh Parker, Nvidia’s chief sustainability officer, according to Axios.
More Nvidia:
- Nvidia’s Jensen Huang has 2 words for your Al job fears
- Nvidia’s $500 billion Al opportunity gets real
- Bank of America resets Nvidia stock forecast after CFO meeting
Nvidia is not the only one who sees the upside. The claim could remove the need for a mechanical chiller in most climates most of the time, even somewhere as hot as Arizona, said Steve Solomon, Microsoft’s (MSFT) vice president of data center engineering.
“It would be a big deal for everybody if we got all of the chips to do that,” Solomon said, according to Axios.
When I read the spec, the part that mattered to me was that temperature, not the press release. Warm-water cooling has existed for years in high-performance computing. What is new is a dominant chip vendor designing its flagship product around it and telling operators they will save money doing it.
The number that scared everyone is shrinking
Here is the part most coverage skips. The scariest water statistic in AI was already falling before Nvidia said a word.
The viral “bottle of water per query” figure traces to that University of California, Riverside team, which estimated GPT-3 drank roughly a 500-milliliter bottle for every 10 to 50 responses. The lead researcher has since said newer, far more efficient models use a small fraction of that, closer to 15 milliliters per prompt and about five milliliters inside the data center itself.
So the per-question footprint is collapsing as the per-building footprint keeps climbing. My read is that those two facts are the entire debate. Each AI answer gets cheaper to cool, while the sheer number of buildings drags the total higher.
Nvidia’s claim lands right in that gap. The company is offering hardware that cuts the cost per building at the exact moment the public has fixated on a per-question number that no longer holds.
What it means for your portfolio
Strip out the environmental guilt and the investor stakes get clearer. Water fights are permitting risk, and permitting risk is the quiet threat to the AI infrastructure trade.
If warm-water cooling removes the most visceral local objection, it makes data centers easier to approve and faster to build, which feeds the chip demand priced into NVDA today.
Local opposition to AI infrastructure is already growing, which is exactly why Google and Amazon have been on defense, according to Axios. Every delay an angry community wins pushes back the revenue that justifies a chip order today.
The reasons for caution are real, and Nvidia named most of them. New systems take years to spread, older data centers keep running older cooling, the company declined to discuss costs, and the water used to generate electricity does not go away, according to Axios.
There is also a trap inside the good news. Efficiency that makes each AI system cheaper to cool can also accelerate the buildout, pushing the industry’s total water and energy use up even as every unit gets leaner.
That is the tension worth watching as an investor. The chips can run warm. Whether the backlash cools with them is the one part Nvidia cannot engineer.
