Is Using AI Bad For The Environment? How Much Energy Does AI Use?
Welcome to Saving Money with Andrew!
After posting Saving Money…With AI?, three readers chided me for my AI usage.
The most detailed comment came from Logan Terheggen:
The use of AI is not a net benefit and is not a harmless tool which can save money. There is a hidden cost to these tools which are built on theft and an outrageous energy and water cost. Data centers are driving the cost of energy up for everyone in states right now. AI is not a tool that should be used cavalierly or regularly. Love the stuff, but just can't accept that we should be advocating for AI usage in our daily lives.
And so I spent a week researching the environmental impact, a topic recently discussed by fellow Substack-er Marcel Salathé in his newsletter Engineering Prompts. I tried to delve further than his post, but owe a lot to his excellent work for inspiration and helping me get started.
First, a disclaimer: this analysis involves a lot of guesswork, and may have basic (or not-so-basic) errors. If you think I’m off by an order of magnitude or more, please let me know!
Summary/TLDR - by my estimates, AI’s energy usage and environmental footprint are modest. There are plenty of reasons to be cautious about AI, but this probably shouldn’t be one of them.
Let’s start with the energy usage of a Google search:
An oft-cited 2009 Google blog post states the average Google search uses ~0.3 watt-hours (Wh) of electricity, about as much as a 10-watt LED light bulb uses in just under two minutes.
Alternatively, this works out to 0.2 grams of CO2 (based on the US’s average carbon intensity). Our Honda CR-V currently gets about 30mpg, or 290 grams of CO2 per mile, so by that measure a Google search has as much carbon impact as driving our car less than four feet.
Estimates from a more recent article suggest that tech improvement has reduced the average Google search query’s power usage by nearly 90% to ~0.04 Wh, a negligible amount.
But what about an AI query?
The same article cites several estimates of ChatGPT energy usage, concluding a single ChatGPT query consumes ~1-10 Wh, or ~25-250x the amount from a Google search. That sounds like a lot!
But if you put it into context, even assuming a high-end estimate of 10 Wh and 10 queries per day (far more than most people perform, including myself), you yield ~36.5 kWh of electricity usage per year. That’s about as much as 10 hours of air conditioning (or keeping an extra LED light bulb on in your house for about five months). I expect these estimates to decrease over time with further improvement, as a more recent post predicts.
But aren’t you using Deep Research?
There’s a catch. Increasingly, I use ChatGPT’s “Deep Research” mode, which takes far longer and does far more work than a single ChatGPT query. A regular query might take 10 seconds, while a Deep Research query easily takes several minutes or more.
ChatGPT limits the number of Deep Research queries you can submit, suggesting it uses far more energy.[1]
I was unable to find any estimates of the usage of a Deep Research query, but given the length of time for a response and the limits ChatGPT places on use, it is likely that Deep Research energy usage exceeds that of a typical query by an order of magnitude or more.[2]
And so, some rough math:
If a Deep Research query uses 10x the energy of a typical ChatGPT query, that would imply each query uses 10-100 Wh, and that a year of max usage by a Plus user (10 queries/month, 120 per year) would use between 1.2 and 12kWh.
If a Deep Research query uses 100x the energy of a typical ChatGPT query, that would imply each query uses 100-1,000 Wh, and that a year of max usage by a Plus user (10 queries/month, 120 per year) would use between 12 and 120kWh.
Is this a lot of electricity? Not really. The average American home uses just under 11,000 kWh per year, so even the extremely high-end estimates of AI energy usage would increase a typical home’s energy usage by about 1%.
In addition, AI queries often substitute for many more Google searches, and sometimes even real-world activities that may have a larger environmental footprint.
My use of Deep Research for car shopping likely replaced tens of searches and reduced our dealership visits (less driving!). In aggregate, my energy footprint may have been greater for using AI, but negligibly so.
Similarly, I’ve used AI to help with home repair questions that would have otherwise necessitated calling a professional, saving a car trip to my house.
What about data centers’ power and water usage?
A 2024 Goldman Sachs post notes that data centers account for 1-2% of power consumption, expected to reach 3-4% by the end of the decade.
By 2028, they expect AI to account for about 19% of data center power demand, with the AI portion reaching about 200 TWh by 2030. This would represent tremendous growth from 2024 estimated levels (about 30 TWh), and even at that level the AI portion would be less than 1% of world energy usage.
Notably, and a bit alarmingly, this is far less than the energy consumed by cryptocurrency-related activities (high-end estimates of up to 240 TWh in 2023, and likely rising) which, unlike AI, are less likely to reduce other energy-intensive activities.
Water is a tougher question, and involved even more guesswork than the numbers above. Figures vary widely, but some credible sources suggest global water usage of ~1-1.25 quadrillion gallons of water annually, with agriculture about 70% of that.[3]
Google’s data centers, for example, used 6.1 billion gallons of water in 2023 (in the company’s words, about as much water as 41 golf courses in the Southwestern United States), and the company noted that their water usage had increased similarly to increases in energy usage.[4]
One post that examined water usage by the largest tech companies noted that an average data center uses 1.8 liters per kWh but that the data centers owned by the larger tech companies are significantly less water-intensive. For example, Amazon’s data centers use 0.19 liters/kWh and Microsoft’s use 0.49 liters/kWh.
Assuming, however, that the average data center uses 1.8 liters per kWh and applying that to Goldman Sachs’ estimate of ~1,000 TWh of data center power usage by 2030 would imply that in 2030 the data center industry may use 400 billion gallons of water, or less than 0.04% of world water consumption, compared to the 70% used by agriculture.
To the industry’s credit, they also appear to be making good-faith efforts to improve water intensity. It appears highly likely that data center water intensity will be well below 1.8 liters/kWh by 2030.
Five Energy/Water-Saving Tips
Researching this post made me significantly less concerned about the environmental footprint of AI. But it’s still an issue, and we should all do more to conserve energy and water where possible. You can even save money by doing so!
Five great tips:
Replace All Non-LED Bulbs In Your Home With LED Bulbs - our unscientific experiment showed that replacing the remaining non-LED bulbs in our house reduced our energy footprint by tens of kWh per month, saving a lot of money.
Consider A Hybrid Car - by switching to a hybrid car, we reduced our gasoline usage by >10 gallons per month, with an attractive payback period for the additional cost of buying the hybrid. More recently, we leased a plug-in hybrid. Its battery gets us only ~30 miles, but by charging almost every night, we’ve gone over a month on a single tank of gas.
Switch To A Time Of Use Plan With Your Electric Company - we have a plan that offers deep discounts for energy usage during off-peak hours. We’ve configured our charger to charge our plug-in hybrid during those windows, saving us hundreds of dollars per year.
Economize On Your Water Usage With Smart Sprinklers - we reduced our water usage during the spring/summer months by 18,000 gallons (saving over $80/year) by switching to a smart sprinkler system.
Consider The Water Intensity Of Your Diet - different foods use dramatically different amounts of water. This brief post from the University of Michigan summarizing a recent study discusses some dietary changes that can significantly reduce your water footprint.
Finally, I’d love to hear your comments. Did I make any major errors? Are there any major points I’m missing? I’d love to hear from you.
And now, Andrew’s pick(s) of the week:
I found myself torn reading It’s Legal to Pay US Workers With Disabilities as Little as 25¢ an Hour, which discusses whether the federal government will eliminate a program that allows employers to pay sub-minimum wage rates to certain individuals with disabilities. Employment rates among the disabled are extremely low, but work provides a great sense of purpose, especially for people who often feel overlooked. I’m not sure what the right answer is here.
I hope this has been helpful. If you liked it, please share it on social media! Also, please send me your feedback, requests, and success stories.
[1] Notably, as of this post, ChatGPT limits one’s usage of Deep Research to 10 queries per month (with a $20 Plus subscription) or 120 queries per month (with a $200 Pro subscription).
[2] This is the weakest link in my analysis. Who’s to say that a Deep Research query doesn’t use 1000x the energy of a single ChatGPT query? But it seems unlikely given the time the query takes (not even close to 1000x as long as a regular query), among other things.
Others might try to estimate energy usage based on media reporting of OpenAI’s cost structure. OpenAI has a lot of costs other than energy, so one could make assumptions as to the % of their costs represent electricity, and then try to back into estimates based on OpenAI’s total number of users and estimates of average number of queries per month.
Alternatively, one might estimate that is unlikely that OpenAI would set the Deep Research quota at a level that the associated energy usage would dwarf the revenue that they collect from a subscriber. This could then be used to back into an extremely high end estimate (unlikely, but a reasonable ceiling) of the amount of energy a query could use.
[3] See Reassessing the Projections of the World Water Development Report (a bit dated) and The United Nations World Water Development Report 2024: water for prosperity and peace; facts, figures and action examples.
[4] See Google 2024 Environmental Report.