- cross-posted to:
- [email protected]
- cross-posted to:
- [email protected]
As you can compute for yourself, AI datacenter water use is not a substantial environmental problem. This long read spells out the argument numerically.
If you’d like a science educator trying to make the headline claim digestible, see here
You must log in or # to comment.
I’m seeing some glaring problems from the author’s analysis.
The biggest is that the author is mixing together freshwater consumption in industry, agriculture, and power generation vs. potable drinking water. From my quick searches, industry, agriculture, and thermoelectric power generation, generally consume massive amounts of non-potable water. This can certainly put strain on reservoirors, aquifers that are also the source of water which is purified for drinking water in localities with low freshwater overall. . However, this ignores the problem entirely about consuming capacity for purification, which data centers do consume that agriculture and industry, agriculture, and thermoelectric power generation do not. Mixes these two together for percentages, but it conceals the additional strain on local water systems specifically for datacenters that isn’t present with the other industries being called out for high water consumption.
My local water co-op is currently servicing two large datacenters, and two more are going in right now. One is a “good citizen” that is using a closed-loop system that consumes a small fraction of water, similar to another non datacenter business, the other datacenter is using an open-loop system which is the dramatically larger consuming method. Why does a datacenter use one of the other? Open-loop is significantly cheaper to build and power for the owner.
Our local water system is getting strained by the increase in data center growth and we’re having to expand water purification and delivery infrastructure in the county. Data centers aren’t generally employers of large number of workers, so this means we’re losing capacity in our locality to host other businesses that would drive local growth.
Lastly, data the author cites is from 2023 and 2024. The 2023 data would nearly be considered ancient history in the AI/LLM world. Keep in mind, ChatGPT was only released in 2022.
The potable vs freshwater distinction is a new one to me, but makes sense and is interesting. I have tried and failed to find literature on potable water use, and can only find the “1% of freshwater is fit for human consumption” stat parroted (which I am moderately certain refers to all freshwater in the whole water cycle). I agree expanding purification and delivery capacity should be ‘costs’ attached to building a datacenter. Are these costs that deplete a scarce resource or ‘merely’ development in infrastructure that would be sustainable once built?
The article treats the question of outdated data by taking projections and comparing those as well. AI investment and infrastructure has grown, but I don’t think it has grown by the 10-20x required to make the authors comparisons wrong: golf, alfalfa, and steel are much larger freshwater sinks than AI could be.
Well-researched read. I came into it pretty antagonistic but the writer backs up their claims well. Would suggest at least a skim before knee-jerk downvoting
