Is AI's Environmental Cost Worth Its Benefits?

AI has a physical footprint

AI may feel like software that exists somewhere in the cloud, but it depends on physical infrastructure: data centers, electricity, cooling systems, networking equipment, and specialized chips. Training and running models consume energy, and some cooling systems also use water.

As demand grows, communities and policymakers are asking how AI infrastructure affects power grids, local water resources, and emissions.

The impact is not uniform. It depends on the efficiency of hardware and software, how often models are used, the energy mix powering facilities, and the location and design of data centers.

Training versus everyday use

Training a large model can require substantial computing resources, but the total cost of serving that model to millions of users can also be significant. A single prompt does not tell us much about the overall footprint without considering model size, hardware, workload, and scale.

Comparisons such as “one AI question uses exactly this much energy” should therefore be treated cautiously unless their assumptions and measurements are clear.

AI can also support sustainability

AI may help optimize electricity demand, improve building efficiency, identify equipment failures, and accelerate certain kinds of scientific research. Those benefits are promising, but they should be measured rather than assumed.

There is also a rebound effect to consider: when computing becomes cheaper and more efficient, people may use more of it. Efficiency improvements do not always reduce total resource consumption.

My perspective

I do not think the right question is simply whether AI uses energy. Every major technology has costs. The better question is whether its benefits justify those costs, and how we can reduce the burden.

Companies should publish meaningful environmental information, invest in efficient systems, and consider local impacts when expanding infrastructure. Researchers should explore smaller models and more efficient methods when they meet the task’s needs.

Question for readers: Should AI companies disclose the energy and water impacts of their largest systems?