How much of a problem is AI's water use, really?

[Sharon_77]

There's a genuinely confusing debate happening around how much water AI actually uses, and it turns out the honest answer is that it depends a lot on where you are. Data centers use water to cool servers, either directly through cooling towers or indirectly through the power plants generating their electricity. Compared to giants like agriculture and certain manufacturing sectors, the raw numbers for data center water use are still pretty small nationally. But small in aggregate doesn't mean small everywhere, especially in already drought stressed regions

One widely repeated claim was that a single chatbot query could cost about 500 milliliters of water, roughly half a bottle. That figure came from an estimate based on an older GPT-4 era model and researchers now say it's already outdated since models have gotten more efficient. Google's own estimate for a median Gemini query puts it closer to just five drops of water. So the per query number has apparently dropped by a huge margin in just a couple of years

Researchers estimated data center cooling systems used around 66 billion liters of water in 2023, which is under one percent of total national consumption. That number is expected to climb a lot as AI data center buildout accelerates though. The real concern isn't the national total, it's what happens when a handful of massive facilities land in places that already don't have much water to spare

Engineers do have options here too, including closed loop cooling systems that recycle water instead of evaporating it away, plus water replenishment projects tied to new facility construction. Some experts think net zero water use for onsite cooling is realistic eventually if companies actually commit to it. Whether that commitment shows up before or after the protests keep happening is the more interesting question

It also doesn't help that corporate water disclosures from big tech companies are often incomplete or inconsistent, which makes the whole debate harder to have with real numbers. Feels like this is one of those stories where both the alarmist framing and the dismissive framing are missing pieces of the full picture

How much of a problem is AI's water use?arstechnica.com

[Brett42]

The inconsistent disclosure problem is the bigger issue to me. Hard to have an informed debate about a resource when the companies using it the most won't share clean comparable numbers. Feels like that should be the actual regulatory ask here

[Renegade51]

People love a scary round number like half a liter per question. Doesn't mean it's wrong to worry about water use, just means the specific viral stat was probably already stale by the time it went viral

[Danny_21]

Worth remembering that chip manufacturing itself uses a ton of ultrapure water before any AI query even happens, and that part rarely makes it into these viral stats. A single semiconductor fab can use millions of gallons a day just keeping wafers clean. That's a supply chain cost that exists regardless of how efficient the actual inference gets later. So even if per query water use keeps shrinking, the manufacturing side is a separate and much bigger bucket. Feels like these conversations always focus on the visible part of the iceberg instead of the whole thing

[Bear24]

The five drops versus five hundred milliliters gap is wild. Shows how much these efficiency numbers can shift in just a year or two of model improvements. Wonder how much further it drops as more inference moves to specialized chips

[Hollow85]

I live somewhere that's had real drought concerns for years and seeing a new data center announced nearby is genuinely nerve wracking regardless of the national percentage. National averages don't help my local reservoir levels one bit. That's the part these thousand foot view articles tend to miss. Local impact is what actually matters to the people living next to these facilities

[NatureBoyMichael27]

Closed loop cooling sounds great on paper but it usually costs more upfront and needs more electricity to run the extra pumps and chillers. So companies default to evaporative cooling because it's cheaper even if it's thirstier.

Until water actually costs them real money that trade off probably won't change much

[Archie91]

Good breakdown of a topic that usually gets reduced to a single scary number on social media. Nice to see the nuance actually laid out for once