Analysis: Artificial intelligence (AI) will soon use 3.7 trillion litres of water a year, the same amount of water as 1.3 billion people living in sub-Saharan Africa. That’s the alarming headline from a recent UN report. The claim made its way around the world in the blink of an eye, with expert commentary from the Science Media Centre repeating this dire warning.
By Thomas Scrimgeour
Concerns about AI water use have been around for some time. AI critic and author of Empire of AI, Karen Hao, points out in her book that a single data centre in Chile was using 1000 times as much water as the local town.
Alternative view: Dire warnings on AI's water use
When such anxiety-inducing numbers come to our attention, it’s worth pausing to ask a few questions. Let’s start with a basic one: how does a data centre even use water?
A data centre is little more than a warehouse full of computers. Cram enough of these into a confined space and very quickly you have a problem with overheating. Air cooling works to a point, but water is often much more effective.
An evaporative system is one cooling method, where water is sent into the atmosphere along with heat energy. Alternatively, there are more modern circulating systems, where the water is cycled through and piped back out as needed.
The first point to make is obvious, but still worth mentioning: water is not destroyed by either process. Consumptive use of water, evaporation, still returns it to earth elsewhere. In non-consumptive, circulating systems, the water is available again immediately.
At first glance, most people will assume that the UN report is warning about water evaporated for cooling. But the volume used by these processes doesn’t even come close to the 1.3-billion-person figure. The UN report is concerned with the “associated water footprint” needed to produce the electricity used by data centres.
In many parts of the world, New Zealand included, power is produced by hydro. It is renewable, affordable, and dispatchable—a wonderful bit of technology.
A hydro dam uses water to spin a wheel, after which it flows away unharmed. The amount of water used is enormous, but it’s non-consumptive: downstream, it is available to use again.
So what’s the problem with the UN report?
A hydroelectric system has a fixed amount of water available for electricity generation. That amount is determined by how much water flows into the catchment. If that water is not used to generate electricity, it passes through the system as spill. Reducing electricity consumption does not save the water for some alternative use. It must either be used to power something else or flow through unused.
This means water use is already baked into the cost of hydroelectricity. It is not a separate line item. Complaining about both the electricity used by a data centre and the water used to make that electricity counts the same cost twice under different names.
If the data centre disappears, you don’t get back both the electricity and the water—you only get one.
Even more importantly, water is not fungible. An extra litre of water in a Swedish hydro reservoir is entirely meaningless to a poor household in sub-Saharan Africa walking to their local well. No amount of Scandinavian water will help them.
Of course, there are parts of the world where water is scarce, groundwater is overdrawn, and humans face risk of drought. Several American localities have mismanaged their natural water resources. But none of this is an AI-specific problem, and it is easily managed by boring, old-fashioned tools like resource management regulation and water rights.
The UN report’s invocation of sub-Saharan households is designed to provoke and outrage, not inform.
But perhaps I protest too much. 1.3 billion is a big number — 16% of the world’s population. So what is this figure as a percentage of total human water use?
America has more data centres than any other country on earth, and they only use 0.2% of their total human water consumption.
Even the inflated “associated water footprint” number from the UN forecast, 3.7 trillion litres globally, is little more than double what California uses to grow almonds.
The fact is that AI won’t be causing us to drink out of our roof gutters.
Humans use very little water in our homes. We use an awful lot to grow food and manufacture goods. Incidentally, Karen Hao overstated the Chilean data centre’s water use by more than 1000 times. She relied on a government report that confused litres with cubic metres, resulting in a wild miscalculation.
For better or worse, AI is here to stay. It will present many challenges that we need to address: energy, e-waste, data sovereignty, and education.
In order to actually address these challenges, we need to keep a sense of proportion. AI water use is moderate and manageable.
There are enough important questions that need our attention. We don’t need the distractions.
Author: Thomas Scrimgeour is a researcher at Maxim Institute, a New Zealand public policy think tank.
SHARE ME