Appreciate the people crunching numbers on this. The outrage over personal LLM energy usage seems overblown. Especially for the number of queries most of us run (<1000 / month?) Picking any other climate soap box is a better use of outrage: driving cars, eating meat, green energy providers, etc. - ThreadSky

maggieappleton.com • 21 days ago

Appreciate the people crunching numbers on this.

The outrage over personal LLM energy usage seems overblown. Especially for the number of queries most of us run (<1000 / month?)

Picking any other climate soap box is a better use of outrage: driving cars, eating meat, green energy providers, etc.

Reposted from Martin Kleppmann

Stop stressing about the energy and carbon footprint of ChatGPT. It’s not a big deal www.sustainabilitybynumbers.com/p/carbon-foo...

Comments

gab-lab.bsky.social•21 days ago

As a contribution to the discussion: images seem to have an impact which is orders of magnitude higher https://www.technologyreview.com/2023/12/01/1084189/making-an-image-with-generative-ai-uses-as-much-energy-as-charging-your-phone/

maelp.bsky.social•6 days ago

Not only that, but even LLMs are being added everywhere. It's not just your "10 daily ChatGPT queries" now, it's everytime you visit a site, there will be a request in the background, etc. The issue is that it all adds up

https://bsky.app/profile/maelp.bsky.social/post/3lpof7m55n22i

tfox.bsky.social•21 days ago

Even without any new datacenters, we should expect energy use to continue to increase significantly as power production devices get smaller, cheaper, and more efficient and the majority of the world that has small or even no power footprint gets better access. LLM unrest seems pointless.

sysadmintools.com•21 days ago

The problem arises when you trace the deep supply chain behind the datacenters. There is much damage to the biosphere that isn't accounted for behind that (transport/oil/water/minerals/human degradation). I do agree about our domestic mammal imbalance (eating meat): https://www.pnas.org/doi/10.1073/pnas.2204892120

erisianrite.com•21 days ago

The common moved goalpost I've seen is "well at least those things are useful!" as if LLMs are not (but also simultaneously will take all jobs ever)

charlivb.bsky.social•20 days ago

I mean, we are actively adding another thing to the big pot of things that are destroying the earth. We do literally need to stop immediately and now. Are LLMs the biggest culprit? No, lol. But there’s isn’t one big one, there’s tons of small ones and they keep adding up.

erisianrite.com•20 days ago

That's not even remotely true. You really should read the linked article if it's something you actually care about.

rsms.me•21 days ago

Cars/trucks is a good metaphor & parallel to computing. The impact of personal cars is nothing really, compared to commercial traffic. In modern personal cars the impact from rubber coming off tires can be worse than exhaust (so electric cars, esp heavy ones, do pollute.)

maelp.bsky.social•6 days ago

The issue is that it all add up, and then there's a rebound effect which makes it consume much more than what we thought it would

https://bsky.app/profile/maelp.bsky.social/post/3lpof7m55n22i

agbocsardi.com•21 days ago

also a great piece by @andymasley.bsky.social, with lots of great visuals!

maelp.bsky.social•21 days ago

That could be said for barely everything of modern life. "How, this AC is bad, but I only turn it up when it's hot, so overall it's not that bad", etc

But the point is that we keep adding and consuming more and more stuff, when we should **reduce**. So of course individually, each...

maelp.bsky.social•21 days ago

...of those things are okay, but then
- you query ChatGPT
- each website you visit does too ("oh, it's not that bad")
- you stream Netflix
- you use a 4K VR headset to work in 3D in addition to your laptop
- etc...

and suddenly the aggregate is big

maelp.bsky.social•21 days ago

I can also take any factor with a big environmental impact, and "split it imaginarily" in 100 "different sub-factors" which each are small enough, and then just say "oh, since each of those individually are small, I can do this one, and the, also this one, etc"

Soon enough I'm back with the issue

agbocsardi.com•21 days ago

I certainly agree that taking steps to reduce environmental impact is necessary, but I think the main point is more that ChatGPT usage isn’t a Big New High-Impact Thing, rather than a drop in an already pretty large bucket, which itself deserves more attention.

maelp.bsky.social•21 days ago

I think the main issue today is that we keep adding more drops in the bucket

While we should instead look at stopping useless stuff and concentrate on the sustainable

phillipcarter.dev•20 days ago

The other thing is there's an observed curve of the unit cost per token going down significantly, which matches the economic incentives to continue to push more efficient compute.

That said, energy use for training is extremely spiky, which is impactful. And water use in those regions is too.

phillipcarter.dev•20 days ago

But the larger point remains - things like the US Govt. incentivizing carbon-expensive food sources, fossil fuel extraction, disincentivizing public transit, etc. (not to mention its military whose emissions exceed most nations) are so outrageously impactful in comparison.

phillipcarter.dev•20 days ago

However, as more and more things shift to compute in datacenters, it certainly behooves us in our corner of the economy to push for sustainability and find ways to build systems that ultimately use less energy per unit of compute over time.

maelp.bsky.social•6 days ago

It's not just your "10 daily ChatGPT queries" now, it's everytime you visit a site, there will be a request in the background, etc. The issue is that it all adds up

https://bsky.app/profile/maelp.bsky.social/post/3lpof7m55n22i

maelp.bsky.social•6 days ago

Case in point: it's not just "your 10 daily queries", it's websites starting to add LLMs everywhere, eg on https://www.redpanda.com/blog/redpanda-ts-node when you select text it does an LLM summarization of it, etc.

This is EXACTLY the rebound effect we're talking about

maelp.bsky.social•7 days ago

https://www.technologyreview.com/2025/05/20/1116327/ai-energy-usage-climate-footprint-big-tech/

maelp.bsky.social•21 days ago

The problem is the rebound effect, you consume not that much per request, but suddenly it's integrated everywhere, and each time you do an action on a website, the owner runs some LLM behind to try to guess what you want "just because", so in the end you have an explosion of consumption for nothing

j11y.io•20 days ago

Yes!! You can make something a 100k big param LLM inferences before hitting the carbon cost of just ONE bitcoin transaction, so, yeah, kinda weird people are so fixated.

stl8k.bsky.social•21 days ago

https://www.linkedin.com/posts/sashaluccioniphd_the-3-wh-per-chatgpt-query-estimate-is-activity-7325515154719186945-RWeg

hibaiunzueta.eu•21 days ago

The worrying part of LLMs is that they are part of an accelerationist ethos that is not about lean or lagom (just right) but about going exponentially faster at what we do. Loads of greenhouse gases, trash, microplastic, and overconsumption to follow.

hdoro.dev•12 days ago

That!

And the fact that very shady actors are using LLMs to further greed, war and destruction: think finance day trading, AI for targeted genocide, finding more petroleum to burn

As usual, the neoliberal narrative derail us by focusing on the individual's responsibility

maelp.bsky.social•6 days ago

Exactly. We need to go SLOWER, not FASTER

leiningl.bsky.social•21 days ago

This article is narrow in scope focusing only on energy usage. I live in central Texas and my primary concern is the quantity of water that data centers consume. The focus shouldn’t be about an individual’s impact but instead what we allow these major corporations to get away with.

maggieappleton.com•21 days ago

Agree it would be more helpful to focus on macro, corporation-scale energy usage and policies.

But I see a lot of moral high ground, personal-level "you shouldn't use ChatGPT for your daily life" takes. As if refusing to get an LLM to adjust a brownie recipe for you is an act of climate justice.

mhkmhk.bsky.social•20 days ago

You shouldn't use it to adjust your brownie recipe because it doesn't know how to cook

yetanotheruseless.com•20 days ago

but it does indeed know based on a million different brownie recipes across the internet, that it was trained on.

It doesn't need to know "how to cook" to help you adjust your brownie recipe.

mhkmhk.bsky.social•19 days ago

Look Jake, idk how much you know about baking, but it's a science. You can't omit baking soda from one recipe and Frankenstein it to parts of other recipes and I have it work. There are multiple public sources for this kinda info. Putting millions of recipes together is not an improvement

mhkmhk.bsky.social•19 days ago

If anything people that actually bake as a hobby have fav sources (like Sally's Baking Addiction) because one author tested and endorses their recipes. I've never had a bad Sally's recipe. Great for US desserts. She and I have similar tastes so I trust those more then an average across the internet

yetanotheruseless.com•19 days ago

I am not an expert cook or baker, but I get by, for fun and family. And I often have simple suggestions from ChatGPT work out fine, both for myself and my skeptical wife agrees. Data point of one, but that is just my experience.

maelp.bsky.social•7 days ago

https://www.technologyreview.com/2025/05/20/1116327/ai-energy-usage-climate-footprint-big-tech/

maelp.bsky.social•21 days ago

It's more: failing to see that we keep adding up useless charges to the climate, because as westerners we don't feel the direct impact

But if you start adding "some fast fasion here", "some chatgpt request there", "some airplane travel over there", and "a few netfix streaming here", it builds up

maelp.bsky.social•21 days ago

Our children will have to live in the world we're building... the problem is that you can always use your "individual lens" to see every little thing individually and say "oh it's not so bad"

but what's bad is that we're building a culture of senseless accumulation (ie capitalism) which...

maelp.bsky.social•21 days ago

...is, little by little, taking our power, and our means of life

and we're propagating it as just one big "oh it's not so bad", "everybody does the same", "how could you do differently?"

we **can** do differently

it's just that marketers (or AI salesmen) don't want you to know

maelp.bsky.social•21 days ago

Exactly! Also the criticism like "oh it's not that bad let us do it" often comes from white privileged westerners who always have an escape hatch if something goes wrong

They totally miss the impact on environment, culture, rights (privacy, etc), creatives (stolen copyrights) etc

charlivb.bsky.social•20 days ago

This is such an America moment tbh. Data centers powered on petroleum, cooled with transported water. I wish fast-moving, energy-producing, naturally occurring cold water was thing! 🥲
Also, why Texas? It’s not like they pay taxes on the net losses they earn.

leiningl.bsky.social•20 days ago

I think Texas because the state government purposefully blocks municipalities from being able to pass local laws regulating companies causing environmental damage. https://www.texastribune.org/2015/05/18/abbott-signs-denton-fracking-bill/

charlivb.bsky.social•20 days ago

I mean I use LLMs, but I believe people mainly protest the process behind training and maintaining these. This mostly applies to OpenAI and Grok, since instead of waiting for generators they just power with petroleum…. (their use exceeds the grid by a lot, I don’t know why they centralize)

billbuchman.bsky.social•20 days ago

That's like worrying about one lightbulb when the entire grid is coal powered. The energy and resource demands of hyper-scaler data centers is the issue, and it's exacerbated by explosive AI growth

Comments

Posting Rules

Reply