ChatGPT drains a half-liter bottle of water with each conversation—is there a solution?

In the last two years, artificial intelligence has become the most popular topic in modern society. This shift has generated a broad awareness of the technology’s environmental implications, mainly focused on the enormous amounts of energy and water required to train and implement these models.

Subscribe to the Softonic newsletter and get the latest in tech, gaming entertainment and deals right in your inbox.

Subscribe (it's FREE) ►

As The Register reminds us, a recent report found that water consumption in northern Virginia data centers, the world capital of data centers, has increased by two-thirds in the last five years.

“ChatGPT needs to ‘drink’ a 500 ml bottle of water for a simple conversation of about 20-50 questions and answers, depending on when and where ChatGPT is deployed,” the researchers estimated in an article published early last year.

To make matters worse, it was a GPT-3 model with around 175 billion parameters, which seems insignificant by today’s standards. It is estimated that GPT-4 will have between 1.7 and 1.8 trillion parameters, and as Trevor Cai from OpenAI said, these models will only continue to grow.

Although that doesn’t bode well for data center energy consumption, the same may not be true for their water addiction. At least, it doesn’t have to be.

AI can do without water, but not without electricity

First, data centers do not consume water. The real problem is that water is extracted from the local environment without being returned to its source. Secondly, IT infrastructure, whether related to AI or not, is not the one that actually consumes water.

Even when liquid-cooled, these systems are usually closed circuits that do not lose a significant amount of fluids during their normal operation.

Barn coolers, often called evaporative coolers or swamp coolers, consume all that water. They prevent these systems from overheating.

However, it is essential to note that this is a design decision and that evaporative coolers are not used in all facilities. If Meta or Amazon install an artificial intelligence data center in your backyard, they will put pressure on the local power grid, but that does not necessarily mean they will consume a quarter of your city’s water supply.

In colder climates, dry coolers and the so-called “free cooling” are suitable, while in warmer and drought-prone regions, it is not uncommon for distribution center operators to opt for refrigerant-based systems. The latest news is that this is precisely what Microsoft does with its distribution center in Goodyear (Arizona), although only after a dispute with the city over wastewater.

Why is water still being used? Because it is cheaper

Although there is not much to do in the existing facilities, the decision to use evaporative cooling in new constructions ultimately comes down to greed or, to use politically correct language, capitalism.

For hyper-scalers, in particular, it all comes down to margins. If you can do something 5% cheaper or more efficient than the competition, you can gain much more profit or reduce costs and win their customer base.

It turns out that water is incredibly efficient at extracting heat from the air compared to other alternative technologies. This translates into lower electricity costs or the possibility of building larger and denser facilities in energy-scarce locations.

Even with industrial rates, electricity costs are increasing rapidly. Therefore, the technology offers a competitive advantage in markets where evaporative coolers are viable.

Ultimately, it all comes down to this: either more energy is used, or more water is consumed. If water is cheaper than electricity and is perceived as abundant, as with the Great Lakes, one can guess which operators will choose.