How Should We Think About AI Waste?

By Sarah Currie-Halpern

AI waste is fast becoming the preeminent environmental problem of our time. Right? 

To begin understanding AI’s environmental impact, we first need to acknowledge that we’re missing the full picture. How much water AI removes from the water cycle is still debated, as are the technological benefits AI might have on ecologically restorative or sustainability efforts. This technology arrived and became central to environmental debates almost overnight. No wonder researchers are still trying to catch up.

Our understanding of AI’s environmental impact begins with the numbers. We know the data centers on which AI depends can consume up to 5 million gallons daily, equivalent to the water use of a town populated by 10,000-50,000 people. This is particularly troubling when we consider that two-thirds of U.S. data centers built since 2022 are located in high water-stress areas.

Data centers’ energy demands are also outsized compared to non-AI servers. Data center energy consumption is estimated at 415 terawatt-hours per year, equivalent to about 1.5% of global electricity consumption in 2024. These facilities also take up vast tracts of land, which inevitably creates questions about land use and habitat destruction.

Land being cleared for development

It isn’t productive to moralize our way out of this or say, “Just stop building.” A Pew Research Center survey found that 62% of American adults interact with AI “at least several times a week,” and one-third of American adults under 30 interact with AI daily. This technology is already established and widely used. We can’t undo it, so we need to figure out new solutions within that paradigm.

Most data centers use evaporative cooling to keep their machinery from overheating. The gallons of water evaporated in this process are lost to the atmosphere, and thus central to the water waste debates. Alternatives to evaporative cooling, however, already exist. Air cooling is the classic option, long used to cool other types of computer servers. Liquid-immersion cooling doesn’t require potable water and reduces overall cooling demand. Newer, closed-loop immersion systems recycle cooling water continuously, eliminating the need to pull from local drinking water supplies. Some methods forgo water altogether, dunking servers in non-corrosive oil.

The less water you use, however, the more energy you need. Google found that its water-cooled data centers use about 10% less energy than air-cooled data centers. The only cooling alternative that conserves both water and energy is free cooling, where cold air from outside is vented into a facility—but this is only possible year-round in a few remote regions around the planet. Suffice it to say that there’s no perfect solution.

The conversation around AI’s environmental impact tends to fixate on how to reduce energy and water use at data centers. More recently, researchers have been investigating these facilities as an energy resource of their own. A new vision of a water-positive and carbon-negative data center has emerged.

The waste data centers create is not without environmental potential. New analysis by the European Union has found that data centers may be useful tools for water purification and carbon capture processes. At present, the waste heat generated by the facilities isn’t being recycled anywhere, and this heat is the key to making a data center sustainable.

There are two areas in particular that show potential: direct air capture (DAC) and thermal water purification. AI’s waste heat can be used to power DAC functions—so effectively, in fact, that researchers estimate it could remove as much as 1,000 megatons of CO₂ from the atmosphere each year. Thermal water purification also requires heat energy to make saltwater or brackish ground water drinkable. In theory, a data center sustaining a big enough water purification plant would be a net water producer.

Researchers emphasize that neither process is perfect or ready to be implemented at scale. If the waste heat produced isn’t hot enough, it won’t be useful in a DAC system. Even if it is, the amount of waste heat currently produced by AI well exceeds the heat needs of existing DAC facilities. In other words, there isn’t a DAC center big enough to use all the waste heat generated by one data center. This is not to say it can’t be built. There might even be an economic incentive, and the analysis suggests that an AI-heated DAC program could generate up to $100 billion in economic value.

Beautiful lake drying up

Pessimism is rarely useful in discussions of waste, AI waste included. Widespread awareness of AI’s water and energy use has created an environment conducive to new technological and ecologically-minded solutions. The point is not to roll back the clock on data centers, but to pursue the kind of ambitious sustainable models that will make AI a net gain for the planet rather than a net drain on our much-needed aquifers.