Doomsday climate and AI news has been clouding my brain as of late, so I thought I would share something a little more hopeful.
This information is from a study published November 10, 2025, in Nature Sustainability, by researchers at Cornell. The team combined information from sustainability reports (think energy, water, and climate companies), industrial data, and even AI to fill gaps. The researchers noted that "industrial data is hard, because not every company is reporting everything."
The main finding
By 2030, AI could emit 24-44 million metric tons of CO₂ annually and drain 731-1,125 million cubic meters of water. That's equivalent to adding 5-10 million cars to the roads, or consuming the same amount of water as 6-10 million households.
If approached correctly, these numbers could be reduced by up to 73% for carbon and 86% for water.
The points I found most interesting
Location is huge
Most data centers are being constructed in water-scarce areas, including Nevada and Arizona. This has been encouraged by loose regulations, tax incentives, cheap land, and predictable solar energy.
The study states that locating facilities in lower-stress areas with improved cooling efficiency could reduce water demand by 52%. Paired with other "best practices" (I'm sure this is a quickly evolving term in the world of data centers), that could reach an 86% water reduction.
The states proposed to have the best combined carbon and water profile are Texas, Montana, Nebraska, and South Dakota, a.k.a. the Midwest and wind belt states.
Renewables alone don't get us there
Even in a high-renewables scenario, emissions only fall about 15% by 2030 compared to the baseline. Around 11 million tons of emissions would still remain. This is because AI data centers require 24/7 power, and wind and solar aren't 24/7, so fossil-fuel backups are necessary.
To fully reach net-zero, we'd still need roughly:
- 28 GW of wind, or
- 43 GW of solar capacity
And if decarbonization doesn't catch up with computing demand, emissions could rise roughly 20%.
Cooling tech helps too
Some of the proposed efficiency technologies are improved server utilization and liquid cooling, which the study says could "potentially remove another 7% of carbon dioxide and lower water use by 29%," for a total water reduction of 32% when combined.
Liquid cooling refers to submerging servers in non-conductive fluids or attaching liquid-chilled plates to their chips. With this method, local tap water isn't evaporated off.
So how do we design accordingly?
Before reading this article I had come to the conclusion that AI usage will never stop, and I don't think it should. AI has some insane implications for healthcare and several other sectors that will lead to great things. The thought I've been left with is: since this technology has been and will continue to be used, how do we design accordingly?
I am super fascinated by these proposals because they prove that if these structures are initially set in the right place, we can mitigate massive amounts of carbon emissions and water usage.
As this study illustrates, location is huge. Taking advantage of places with renewable energy built in and cool climates for natural cooling, then combining that with liquid cooling, is a strong start.
Beyond physical settings, another concept we can use is carbon arbitrage: routing tasks according to what's available. For example, a tech company might train a model during peak wind hours in South Dakota instead of peak heat in Nevada.
I've seen countless other ideas, like using the heat from data centers to power a desalination plant, that keep me excited about our evolving technology with intentionality.