🚰 A single golf club uses more water than ChatGPT

Pessimist fog

In social media, a claim is spreading that a single question or instruction (prompt) to ChatGPT consumes a half-liter water bottle. This is a distortion of an older claim that 20 to 50 prompts use up half a liter of water.

These claims are far too bad to be true. Instead, it's a fog of pessimism being spread.

But before we get to that, let's understand how water consumption is connected to AI.

AI data centers consume water primarily for cooling. When the servers are working, they get hot, and the heat has to be removed for the systems to function. In many facilities, water is used in the cooling systems, and most of the water that is actually consumed doesn't become polluted — it evaporates as the heat is carried away.

A smaller portion remains as what's called blowdown water, where salts and sometimes chemicals used against corrosion, bacteria, and deposits have become concentrated. That water may need to be treated before being discharged or sent to treatment facilities, much like other wastewater.

So data centers don't pollute the water, so the key issue is how much water they consume.

One person who has done significant work to both map and visualize this is Andy Masley. He's a former physics teacher, holds dual master's degrees in physics and psychology, and now does research on AI. He's also vegan. That last part is relevant because, given the context, you might otherwise suspect he's some kind of fossil-fuel-loving far-right extremist, or something like that.

He has written extensively about AI and both energy and water consumption.

One problem — which helps claims like the ones above gain traction — is that people in general have no clue how much water goes into anything. That makes it easy to be horrified.

How much water ChatGPT actually consumes

If you factor in electricity production as well as data center cooling, a single ChatGPT prompt consumes roughly 0.31 milliliters. That's 0,06 percent – or 0,6 permille – of a 500-milliliter water bottle.

To put that in perspective, the entire production chain for a beef hamburger consumes around 2,500 liters of water. Switch to a chicken burger and you're looking at roughly 650 liters, and for a plant-based burger about 150 liters. (These figures are based on numbers from the Water Footprint Network, and I haven't checked whether they're exaggerated.)

AI compared to golf courses

Andy Masley has mapped and compared AI data centers with several other activities, including golf.

ChatGPT accounts for roughly 0.018 percent of the water use of all golf courses in the US. America's golf courses use approximately 5,474 times more water than ChatGPT.

Golf courses in Arizona alone use roughly 148 times more water than ChatGPT.

There are individual golf courses that consume more water than ChatGPT does worldwide!

Quail Creek uses roughly 6 times more water than ChatGPT, Omni Tucson National about 4 times more, and Ventana Canyon about 3 times more, according to Pima County's records.

All data centers in the US (not just those used for AI) consumed, about 3.3 percent of the water used by golf courses, according to the most recent figures (2023).

Golf courses are thirsty, but they're nothing compared to corn

ChatGPT accounts for:

• roughly 0.0035 percent of all irrigated corn farming in the US
• roughly 0.0088 percent of irrigated corn farming in Nebraska

All US data centers account for:

• roughly 0.62 percent of all irrigated corn farming in the US
• roughly 1.58 percent of irrigated corn farming in Nebraska

Corn production thus uses approximately 28,900 times more water than ChatGPT, and roughly 161 times more than all US data centers combined.

What about bottled water?

ChatGPT's water consumption amounts to roughly 0.43 percent of the water the US uses for bottled water production. Not even all data centers in the US combined reach the same amount, coming in at approximately 76 percent.

Going forward?

Unfortunately, we then have to rely on projections, and when something is growing as fast as AI, my starting assumption is that they're far too low. But if we look at the ones that exist, here's what we get:

All AI use in the US will reach roughly 8.5 percent of total golf course water consumption by 2030. ChatGPT will account for 0.047 percent. By 2030, ChatGPT won't even reach two percent of Arizona's golf courses. It takes all AI use in the US to surpass Arizona.

All AI use in the US will surpass bottled water production by 2030 and consume roughly twice as much. But corn production in the US will still use over 60 times more water.

Shouldn't we care at all?

Of course we should try to minimize water consumption related to AI and data centers in general. And that's exactly what's being done.

To reduce water consumption in data centers, the main focus has been on reducing the need for cooling where water evaporates. This is achieved through more closed-loop cooling systems, liquid cooling closer to the chips, better management of temperature and airflow, more sensors to detect leaks, and increased use of recycled water instead of drinking water.

The impact has been clear: Microsoft has reduced its water intensity by around 45 percent in just a few years, AWS by 40 percent since 2021, Meta by roughly 40 percent since 2020, and Google says new low-water solutions can reduce water use by up to 50 percent at a facility.

It's good for these companies to face continued pressure to reduce both energy and water consumption, but that pressure should be based on facts. Not a bunch of made-up numbers spread in a fog of pessimism.

Mathias Sundin
Angry Optimist