{"id":47781,"date":"2025-05-16T14:45:44","date_gmt":"2025-05-16T18:45:44","guid":{"rendered":"https:\/\/mjtsai.com\/blog\/?p=47781"},"modified":"2025-07-21T13:10:37","modified_gmt":"2025-07-21T17:10:37","slug":"why-using-chatgpt-is-not-bad-for-the-environment","status":"publish","type":"post","link":"https:\/\/mjtsai.com\/blog\/2025\/05\/16\/why-using-chatgpt-is-not-bad-for-the-environment\/","title":{"rendered":"Why Using ChatGPT Is Not Bad for the Environment"},"content":{"rendered":"

Andy Masley<\/a>:<\/p>\n

It’s not bad for the environment if you or any number of people use ChatGPT<\/a>, Gemini<\/a>, Claude<\/a>, Grok<\/a>, or other large language model<\/a> (LLM) chatbots. You can use ChatGPT as much as you like without worrying that you’re doing any harm to the planet.<\/p>

[…]<\/p>

Throughout this post I’ll assume the average ChatGPT query uses 3 Watt-hours (Wh) of energy, which is 10x as much as a Google search. This statistic is likely wrong. ChatGPT’s energy use is probably lower according to EpochAI<\/a>. Google’s might be lower too<\/a>, or maybe higher now that they’re incorporating AI into every search. We’re a little in the dark on this, but we can set a reasonable range. It’s hard for me to find a statistic that implies ChatGPT uses more than 10x as much energy as Google, so I’ll stick with this as an upper bound to be charitable to ChatGPT’s critics.<\/p>

[…]<\/p>

If you multiply an extremely small value by 10, it can still be so small that it shouldn’t factor into your decisions.<\/p>

[…]<\/p>

They hear about AI data centers rapidly growing, look around, and see that everyone’s using ChatGPT, and assume there must be some connection. […] The mistake they’re making is simple: ChatGPT and other AI chatbots are extremely, extremely small parts of AI’s energy demand.<\/p><\/blockquote>\n\n

Via Adam Engst<\/a>:<\/p>\n

\n

Masley calculates that, on a daily basis, the average American uses enough energy for 10,000 ChatGPT prompts and consumes enough water for 24,000–61,000 prompts.<\/p>\n<\/blockquote>\n\n

Wayne Williams<\/a>:<\/p>\n

\n

The power needed to run generative AI is pushing infrastructure beyond what traditional air cooling can handle.<\/p>\n

To explore the scale of the challenge, I spoke with Daren Shumate, founder of Shumate Engineering, and Stephen Spinazzola, the firm’s Director of Mission Critical Services.<\/p>\n

[…]<\/p>\n

A typical Chat-GPT query uses about 10 times more energy than a Google search – and that’s just for a basic generative AI function. More advanced queries require substantially more power that have to go through an AI Cluster Farm to process large-scale computing between multiple machines.<\/p>\n<\/blockquote>\n\n

Dan Drake<\/a>:<\/p>\n

If you’re measuring energy consumption, you need to do a kind of “lifecycle analysis” -- if the choice is between using a traditional search engine and asking a chatbot, you should compare the entire workflow with each.<\/p>

If I do a regular web search for something, I will frequently click three to four of the results and open them in new tabs, because I’m not sure exactly which one will answer my question; I might do another search. Each of those loads a website, with all the accompanying HTML, JS, and so on.<\/p>

With chatbots, I find it’s more common for the response to have exactly what I want. “One and done”, as they say.<\/p><\/blockquote>\n\n

Also, as AI gets better, people will use it more. They will ask it to do deep research tasks that they would not have even attempted with Google. Or that perhaps they would have paid a person to do.<\/p>\n\n

Update (2025-05-19<\/a>): There are Hacker News<\/a> and Lobsters<\/a> pages for Masley’s post. Simon Willison<\/a> says it’s “by far the most convincing rebuttal of this idea that I’ve seen anywhere.” Michael Lazar<\/a> wrote a rebuttal, which I find to be long on axe grinding and rhetorical criticisms and short on substance (via Dustin Westphal<\/a>). Masley<\/a> has a follow-up post about what he got wrong.<\/p>\n\n

I think the best criticism is that the narrow<\/a> question Masley is investigating is not what really matters. If you’re against the idea of LLMs or the overall energy consumption of AI (including training and non-chatbot uses), you don’t particularly care about the incremental cost of one more person using ChatGPT. Also, the numbers for ChatGPT may not apply to other systems such as Grok.<\/p>\n\n

Stephen Hackett<\/a>:<\/p>\n

\n

As I wrote about earlier this week<\/a>, xAI has broken ground on a second data center on Tulane Road here in Memphis that will require an unbelievable amount of electricity.<\/p>\n

[…]<\/p>\n

As seen here, the SELC has photographic evidence<\/a> that some 35 turbines have been in operation at xAI’s initial data center, despite Memphis Mayor Paul Young claiming in mid April that only 15 were in use<\/a>. If 15 strikes you as an oddly specific number, it’s because the Shelby County Health Department’s permit to xAI only covers 15 permanent units.<\/p>\n

If the plan outlined in this documents comes to pass, there could be anywhere between 40 to 90 turbines running in south Memphis across the two sites.<\/p>\n<\/blockquote>\n\n

Matt Birchler<\/a>:<\/p>\n

\n

I could keep going, but I have some very real options for not only offsetting my ChatGPT usage, but also radically reducing my tech energy footprint overall. The easiest win for me is scheduling my Synology to power down overnight.<\/p>\n

[…]<\/p>\n

I didn’t write this post to suggest we should all use as much energy as possible, screw the environment, let’s just burn it all down. My intention was to present the same ChatGPT and other LLM energy use numbers you see in alarmist articles in a different way to show that you can tell different stories depending on how you present the same data. Do LLMs use more energy than a lot of other digital actions? Yeah, they seem to, but the base number is so microscopically small that we still aren’t dealing with large numbers in the grand scheme of things.<\/p>\n<\/blockquote>\n\n

Update (2025-05-23<\/a>): James O’Donnell and Casey Crownhart<\/a>:<\/p>\n

\n

Today, new analysis by MIT Technology Review<\/i> provides an unprecedented and comprehensive look at how much energy the AI industry uses—down to a single query—to trace where its carbon footprint stands now, and where it’s headed, as AI barrels towards billions of daily users.<\/p>\n

[…]<\/p>\n

By 2028, the researchers estimate, the power going to AI-specific purposes will rise to between 165 and 326 terawatt-hours per year. That’s more than all electricity currently used by US data centers for all purposes; it’s enough to power 22% of US households each year.<\/p>\n

[…]<\/p>\n

The Lawrence Berkeley researchers offered a blunt critique of where things stand, saying that the information disclosed by tech companies, data center operators, utility companies, and hardware manufacturers is simply not enough to make reasonable projections about the unprecedented energy demands of this future or estimate the emissions it will create. They offered ways that companies could disclose more information without violating trade secrets, such as anonymized data-sharing arrangements, but their report acknowledged that the architects of this massive surge in AI data centers have thus far not been transparent, leaving them without the tools to make a plan.<\/p>\n<\/blockquote>\n\n

Via Nick Heer<\/a>:<\/p>\n

\n

This robust story comes on the heels of a series of other discussions about how much energy is used by A.I. products and services. Last month, for example, Andy Masley published<\/a> a comparison of using ChatGPT against other common activities. The Economist<\/em> ran another<\/a>, and similar articles<\/a> have been published before. As far as I can tell, they all come down to the same general conclusion: training A.I. models is energy-intensive, using A.I. products is not, lots of things we do online and offline have a greater impact on the environment, and the current energy use of A.I. is the lowest it will be from now on.<\/p>\n<\/blockquote>\n\n

Nick Heer<\/a>:<\/p>\n

\n

Thinking about the energy “footprint”<\/a> of artificial intelligence products makes it a good time to re-link to Mark Kaufman’s excellent<\/a> 2020 Mashable<\/em> article in which he explores the idea of a carbon footprint.<\/p>\n<\/blockquote>\n\n

Update (2025-06-11<\/a>): Jay Peters<\/a>:<\/p>\n

OpenAI CEO Sam Altman, in a blog post published Tuesday<\/a>, says an average ChatGPT query uses about 0.000085 gallons of water, or “roughly one fifteenth of a teaspoon.” He made the claim as part of a broader post on his predictions about how AI will change the world. <\/p><\/blockquote>\n\n

Update (2025-07-14<\/a>): David Mytton<\/a>:<\/p>\n

\n

This is several orders of magnitude less than most estimates we’ve seen in the media (mainstream and academic).<\/p>\n<\/blockquote>\n\n

Update (2025-07-21<\/a>): Eli Tan<\/a>:<\/p>\n

\n

The Morrises’ experience is one of a growing number of water-related issues around Newton County, which is a one-and-a-half-hour drive east of Atlanta and has a population of about 120,000 people. As tech giants like Meta build data centers in the area, local wells have been damaged, the cost of municipal water has soared and the county’s water commission may face a shortage of the vital resource.<\/p>\n

The situation has become so dire that Newton County is on track to be in a water deficit by 2030, according to a report last year. If the local water authority cannot upgrade its facilities, residents could be forced to ration water. In the next two years, water rates are set to increase 33 percent, more than the typical 2 percent annual increases, said Blair Northen, the mayor of Mansfield, a town in Newton County.<\/p>\n

[…]<\/p>\n

A Meta spokeswoman said the company had recently commissioned a well study on the Morrises’ property and said it was “unlikely” that its data center affected the supply of groundwater in the area.<\/p>\n<\/blockquote>","protected":false},"excerpt":{"rendered":"

Andy Masley: It’s not bad for the environment if you or any number of people use ChatGPT, Gemini, Claude, Grok, or other large language model (LLM) chatbots. You can use ChatGPT as much as you like without worrying that you’re doing any harm to the planet.[…]Throughout this post I’ll assume the average ChatGPT query uses […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"apple_news_api_created_at":"2025-05-16T18:45:46Z","apple_news_api_id":"9fbfced0-8962-42aa-a119-15c6d0fc41e6","apple_news_api_modified_at":"2025-07-21T17:10:40Z","apple_news_api_revision":"AAAAAAAAAAAAAAAAAAAABQ==","apple_news_api_share_url":"https:\/\/apple.news\/An7_O0IliQqqhGRXG0PxB5g","apple_news_coverimage":0,"apple_news_coverimage_caption":"","apple_news_is_hidden":false,"apple_news_is_paid":false,"apple_news_is_preview":false,"apple_news_is_sponsored":false,"apple_news_maturity_rating":"","apple_news_metadata":"\"\"","apple_news_pullquote":"","apple_news_pullquote_position":"","apple_news_slug":"","apple_news_sections":"\"\"","apple_news_suppress_video_url":false,"apple_news_use_image_component":false,"footnotes":""},"categories":[2],"tags":[1351,2317,2682,1358,2347,2631,2361,96,2773],"class_list":["post-47781","post","type-post","status-publish","format-standard","hentry","category-technology","tag-artificial-intelligence","tag-chatgpt","tag-claude","tag-environment","tag-bard","tag-grok","tag-openai","tag-web","tag-xai"],"apple_news_notices":[],"_links":{"self":[{"href":"https:\/\/mjtsai.com\/blog\/wp-json\/wp\/v2\/posts\/47781","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/mjtsai.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/mjtsai.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/mjtsai.com\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/mjtsai.com\/blog\/wp-json\/wp\/v2\/comments?post=47781"}],"version-history":[{"count":7,"href":"https:\/\/mjtsai.com\/blog\/wp-json\/wp\/v2\/posts\/47781\/revisions"}],"predecessor-version":[{"id":48545,"href":"https:\/\/mjtsai.com\/blog\/wp-json\/wp\/v2\/posts\/47781\/revisions\/48545"}],"wp:attachment":[{"href":"https:\/\/mjtsai.com\/blog\/wp-json\/wp\/v2\/media?parent=47781"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/mjtsai.com\/blog\/wp-json\/wp\/v2\/categories?post=47781"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/mjtsai.com\/blog\/wp-json\/wp\/v2\/tags?post=47781"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}