Isn't the real metric of concern the absolute amount of CO2 generated that will have an impact on the environment? That every person's AI queries contributes a small amount to the CO2 production doesn't make the sum of all CO2 production go away.
j_w 5 hours ago [-]
This is not a critique of LLM usage, just on individual contributions to the environmental crisis.
It's fairly popular to claim that you as an individual have no significant effect on the environment and that it's the actions of the large companies, which are effectively "super polluters." This ignores that companies take these actions because of the market forces imposed on them by the consumers.
An individuals impact in isolation is small, however, if that same individual made changes not only in their own life, but urged those around them to make similar changes, the network effect would be huge. This extends beyond the environment: boycotts, product recommendations, exercise, etc. You truly need to be the change that you want to see.
mac-chaffee 6 hours ago [-]
I agree but there's a lot of nuance to the next question of "well what IS bad for the environment" and tech's role in that question.
I find this unconvincing. The actual discussion of LLM generation is very lacking.
The original link [1] cites a discussion of the cost per query of GPT-4o at 0.3whr [2]. When you read the document [2] itself you see 0.3whr is a lower bound & 40whr is the upper bound. The paper [2] is actually pretty solid, I recommend it. It uses the public metrics from other LLM APIs to derive a likely distribution of the context size of the average query for GPT-4o which is a reasonable approach given that data isn't public. Then factoring in GPU power per FLOP, average utilization during, and cloud/renting overhead. It admits this likely has non-trivial error bars, concluding the average is between 1-4whr per query.
This is disappointing to me as the original link [1] attempts to bring in this source [2] to disprove the 3whr "myth" created by another paper [3], yet this 3whr figure lies directly in the error bars their new source [2] arrives at.
The methodology is inherently flawed by assuming all infrastructure, training, etc is going to exist with or without individual queries, while trying to answer a different question of the impact of AI on the environment. It’s like arguing the environmental impact of solar electricity is 0 because the panels would exist either way.
Thus the results inherently fail to analyze the underlying question.
A more realistic estimate is to take their total spending assuming X% of their expenses are electricity directly or indirectly because the environmental impact isn’t adds up. Even that ignores the energy costs on 3rd party servers when they download their training data.
hugmynutus 5 hours ago [-]
You are correct to point out the larger questions of supply chain cost (and their environmental impact) are not addressed in the root link.
cwillu 6 hours ago [-]
The unit is watt·hour, not watt/hour: multiplication, not division.
> ...Globally, coal, followed by gas, is the largest source of electricity production....
As long as this is the case we can hardly even the debate of the impacts of those new techs on the sole topic of the climate.
Let me remind you kindly we well passed the point of this single problem, we are dealing with planetary boundaries, there is 9 of them.
Another reminder is that co2 pollution alone is the direct product of the GDP, there is no update in sight about how the competing countries should deal with shared homothetic GDP cuts to reduce the gaz emissions. so even we would do something, we have not started to get to the serious business.
Why AI ? Because we are screwed. We failed on humanism, we failed on climate, we cant failed that one, we would just kick ourself out of the real game.
this is a kind of a great megalomaniac idea, but i prefer that to your pathetic bullshit. so even though you are fucking cringe, go elon,
Fire in the hole !
malvim 6 hours ago [-]
And what about all the developing and testing of models? What about all the OTHER companies that can’t wait to get a piece of this cake and are training and scraping the internet like there’s no tomorrow? And all the companies that are integrating LLMs into their daily workflows using tons of api calls daily?
Come on…
5 hours ago [-]
devmor 6 hours ago [-]
Like every other "rebuttal" to this argument, this chooses to pretend that the complaint is about the power usage of making API calls, instead of the power usage of training models.
It's like if I said I was concerned about factory farming impacts and you showed me a video of meat packaging at a grocery store, claiming it alleviates my concerns.
TobTobXX 6 hours ago [-]
From the article:
> Training GPT-4 used 50 GWh of energy. Like the 20,000 households point, this number looks ridiculously large if you don’t consider how many people are using ChatGPT.
> Since GPT-4 was trained, it has answered (at minimum) about 200,000,000 prompts per day for about 700 days. Dividing 50GWh by the total prompts, this gives us 0.3Wh per prompt. This means that, at most, including the cost of training raises the energy cost per prompt by 10%, from 10 Google searches to 11. Training doesn’t add much to ChatGPT’s energy cost.
How does people using it offset the amount of energy used to train it? If I use three hundred pounds of flour learning to make pizza, the subsequent three hundred pounds of flour I use making delicious pizzas doesn't make the first 300 go away. Am I misunderstanding the numbers?
ssalazar 6 hours ago [-]
Its not offset, its amortized.
Your effective flour / pizza is (300 + 300) / num_pizzas.
The total marginal flour expended will go up as you make more pizzas, but the effective cost will actually go down as the upfront cost is amortized over lifetime usage.
serial_dev 6 hours ago [-]
You don’t misunderstand the numbers, you misunderstand the point. If you flush your pizzas down the toilet, it’s a waste. If you feed 300 people with it, it’s not, even if you end up using the same amount of ingredients.
TobTobXX 6 hours ago [-]
Sure, it's a value calculation.
If you're able to serve delicious pizzas afterwards, it was worth wasting the first kg (you might call it an investment).
If you're able to bring value to millions of users, it was worth to invest a few GWh into training.
You might disagree on the usefulness. I think, you shouldn't have wasted a kg of flour because I won't ever eat your pizzas anyway. But many (you, your guests, ChatGPT users) might think it was worth it.
Remnant44 6 hours ago [-]
It doesn't make it go away. Using your analogy - if you used 300lb to learn and then only made 10 lb of pizza after that, it would be a pretty poor use of resources.
If you instead went on to produce millions of pizzas for people and 30,000lb of flour, that 300lb you used to learn looks like a pretty reasonable investment.
warkdarrior 6 hours ago [-]
The cost of training has to be normalized by the number of users (or queries) that rely on that training.
If you use 300 lbs of flour to learn, and 300 lbs of flour to make 300 pizzas, then the total flour cost is 2 lbs of flour per pizza.
RobinL 6 hours ago [-]
For context thats's about equivalent to 100 transatlantic flights
See my other comment here. One AI training run does not exist in a vacuum. Do you think they built billions of dollars in datacenters full of computer power just to let it sit idle?
megaman821 6 hours ago [-]
That argument is just as bad. The training runs are equivilant to a few days to a couple of weeks of flights from New York to London. No one thinks we are going to save the environment by stopping a single plane route for 2 weeks.
To make it worse, the model training cost only refers to the cost of the training itself. The externalities - everyone else being forced to drastically upscale their compute power because scraper blocking isn't foolproof - are, as usual for hypercapitalism, conveniently ignored.
AI training in its current form is unsustainable, I'd go as far to say it's a threat for the open and decentralized Internet as you have all but zero chance of standing alone against the flood of training scraper bots and more and more control gets ceded to Cloudflare et al.
etchalon 6 hours ago [-]
The "cheat sheet" seems to address the environmental impact of using ChatGPT, not the environmental impact of training the model.
TobTobXX 6 hours ago [-]
Wrong. In the article:
> Training GPT-4 used 50 GWh of energy. Like the 20,000 households point, this number looks ridiculously large if you don’t consider how many people are using ChatGPT.
> Since GPT-4 was trained, it has answered (at minimum) about 200,000,000 prompts per day for about 700 days. Dividing 50GWh by the total prompts, this gives us 0.3Wh per prompt. This means that, at most, including the cost of training raises the energy cost per prompt by 10%, from 10 Google searches to 11. Training doesn’t add much to ChatGPT’s energy cost.
Just because you divide a number by a lot to get a small number doesn't make the original number smaller.
Those are 200M/d prompts that wouldn't happen without the training.
TobTobXX 6 hours ago [-]
> Just because you divide a number by a lot to get a small number doesn't make the original number smaller.
A bus emits more CO2 than a car. Yet it is more friendly to the environment because it transports more people.
> Those are 200M/d prompts that wouldn't happen without the training.
Sure, but at least a few millions are deriving value from it. We know this because they pay. So this value wouldn't have been generated without the investment. That's how economics work.
warkdarrior 6 hours ago [-]
Those 200M/d prompts would be replaced with some other activities to solve the same problems. So if training did not happen, maybe instead of 200M/d prompts, you'd have 200M/d trips to the local library, using 200M cars to each drive three miles.
Remnant44 6 hours ago [-]
By the same token, even if you accept that AI usage will incentivize additonal model trainings, that cost is diffused across hundreds of millions of users, and is not a marginal cost, so it gets further reduced on a per-chat basis the more you use AI. I don't know what the per-user environmental cost of training a model is, but that's a pretty big factor to divide energy usage by.
spencerflem 6 hours ago [-]
But irrelevant if the point is to not contribute to something that you'd rather see banned.
Remnant44 6 hours ago [-]
I don't really follow this objection. Determining the actual energy usage of AI training+inference is something that is an objective reality. Whether you hate it or love it doesn't change these facts.
spencerflem 5 hours ago [-]
I think LLM training is objectively bad for the environment (uses countries worth of power). I am aware than my marginal usage wouldn't change things much either way, but I don't want to encourage them regardless.
simonw 4 hours ago [-]
Uses countries worth of power?
Training a single LLM takes a few dozen fully loaded transatlantic passenger aircraft trips worth of power.
For "counties worth of power" I think you might be talking ALL data center use as a whole.
roschdal 6 hours ago [-]
The human brain is dramatically more energy-efficient than AI models like ChatGPT.
Human brain: Uses about 20 watts of power.
ChatGPT (GPT-4): Running a single query can use hundreds of watts when accounting for the entire datacenter infrastructure (some estimates suggest 500–1000 watts per query on average, depending on model size and setup).
If we assume:
20 watts for the human brain thinking continuously,
1000 watts for ChatGPT processing one complex query,
then the human brain is about 50x more energy-efficient (or 5000% more efficient) than ChatGPT per task, assuming equal cognitive complexity (which is debatable, but good for ballpark comparison).
JustFinishedBSG 6 hours ago [-]
Watt isn’t a measure of energy. Without how long it takes for a human and ChatGPT to solve the task then the comparison doesn’t teach anything
ryanianian 6 hours ago [-]
Watt is not a unit of energy but instead a unit of power. A brain may need 20 watts, but it may use 20 watts for a lot more time than ChatGPT would.
The brain may ultimately be more power-efficient, but the units you want are watt-hours.
TobTobXX 6 hours ago [-]
This is fascinating. I mean it's not an argument against LLMs (I have only one brain, even though I'd like to have more). But I really hope that we'll learn much more about how our brains work.
Legend2440 5 hours ago [-]
The brain is more efficient because it physically is a neural network, as opposed to a software simulation of one.
homebrewer 6 hours ago [-]
You should then probably count the whole body, which consumes approximately 200 watts last time I checked.
halyconWays 6 hours ago [-]
There's massive evolutionary pressure for maximizing energy efficiency in brains. I'd like to see LLMs procreate and select for energy efficiency while, ideally, minimizing insanity and maintaining g-factor.
reyqn 6 hours ago [-]
So all of the cheatsheet is basically "it's not bad because there are worse things"?
You can try explaining why it's not "that" bad for the environment, the planet is still worse off than when it didn't exist.
Let's carry on inventing new ways to spend energy, but it's ok because we still spend more energy for other stuff.
It's kinda sad how the world saw climate change, said it was bad, but in the end decided to do nothing about it.
aabhay 6 hours ago [-]
I think the better argument is about the direction of change versus the current magnitude.
If we are to believe that the models will get bigger, use more tokens, work for longer, this calculation can easily become very very skewed in the other direction.
Consider an agentic system that runs continuously for 6 hours. It is possible this system processes billions of tokens. That could more than equal a transatlantic flight in this hypothetical world.
Now compare this with non-AI work, like a CRUD app. Serving millions of queries in that same period would consume a tiny fraction of what ChatGPT consumes.
Rather than being a “win” for AI, the fact that we’re even 3 or 4 orders of magnitude away from this being a problem means that its already grounds to be concerned.
If you all flag this article, dang will probably get around to fixing it.
It's fairly popular to claim that you as an individual have no significant effect on the environment and that it's the actions of the large companies, which are effectively "super polluters." This ignores that companies take these actions because of the market forces imposed on them by the consumers.
An individuals impact in isolation is small, however, if that same individual made changes not only in their own life, but urged those around them to make similar changes, the network effect would be huge. This extends beyond the environment: boycotts, product recommendations, exercise, etc. You truly need to be the change that you want to see.
I've been unsatisfied with how people in tech address that complex subject so I wrote about it here: https://www.macchaffee.com/blog/2025/tech-and-the-climate-cr...
The original link [1] cites a discussion of the cost per query of GPT-4o at 0.3whr [2]. When you read the document [2] itself you see 0.3whr is a lower bound & 40whr is the upper bound. The paper [2] is actually pretty solid, I recommend it. It uses the public metrics from other LLM APIs to derive a likely distribution of the context size of the average query for GPT-4o which is a reasonable approach given that data isn't public. Then factoring in GPU power per FLOP, average utilization during, and cloud/renting overhead. It admits this likely has non-trivial error bars, concluding the average is between 1-4whr per query.
This is disappointing to me as the original link [1] attempts to bring in this source [2] to disprove the 3whr "myth" created by another paper [3], yet this 3whr figure lies directly in the error bars their new source [2] arrives at.
Links:
1. https://simonwillison.net/2025/Apr/29/chatgpt-is-not-bad-for...
2. https://epoch.ai/gradient-updates/how-much-energy-does-chatg...
3. https://www.sciencedirect.com/science/article/pii/S254243512...
Edit: whr not w/hr
Thus the results inherently fail to analyze the underlying question.
A more realistic estimate is to take their total spending assuming X% of their expenses are electricity directly or indirectly because the environmental impact isn’t adds up. Even that ignores the energy costs on 3rd party servers when they download their training data.
> ...Globally, coal, followed by gas, is the largest source of electricity production....
As long as this is the case we can hardly even the debate of the impacts of those new techs on the sole topic of the climate.
Let me remind you kindly we well passed the point of this single problem, we are dealing with planetary boundaries, there is 9 of them. Another reminder is that co2 pollution alone is the direct product of the GDP, there is no update in sight about how the competing countries should deal with shared homothetic GDP cuts to reduce the gaz emissions. so even we would do something, we have not started to get to the serious business.
Why AI ? Because we are screwed. We failed on humanism, we failed on climate, we cant failed that one, we would just kick ourself out of the real game.
this is a kind of a great megalomaniac idea, but i prefer that to your pathetic bullshit. so even though you are fucking cringe, go elon,
Fire in the hole !
Come on…
It's like if I said I was concerned about factory farming impacts and you showed me a video of meat packaging at a grocery store, claiming it alleviates my concerns.
> Training GPT-4 used 50 GWh of energy. Like the 20,000 households point, this number looks ridiculously large if you don’t consider how many people are using ChatGPT.
> Since GPT-4 was trained, it has answered (at minimum) about 200,000,000 prompts per day for about 700 days. Dividing 50GWh by the total prompts, this gives us 0.3Wh per prompt. This means that, at most, including the cost of training raises the energy cost per prompt by 10%, from 10 Google searches to 11. Training doesn’t add much to ChatGPT’s energy cost.
https://andymasley.substack.com/i/162196004/training-an-ai-m...
If you're able to serve delicious pizzas afterwards, it was worth wasting the first kg (you might call it an investment).
If you're able to bring value to millions of users, it was worth to invest a few GWh into training.
You might disagree on the usefulness. I think, you shouldn't have wasted a kg of flour because I won't ever eat your pizzas anyway. But many (you, your guests, ChatGPT users) might think it was worth it.
If you instead went on to produce millions of pizzas for people and 30,000lb of flour, that 300lb you used to learn looks like a pretty reasonable investment.
If you use 300 lbs of flour to learn, and 300 lbs of flour to make 300 pizzas, then the total flour cost is 2 lbs of flour per pizza.
I mean, I guess advances could plateau and we stop spending exponentially more energy year after year...
I'm not opining on whether it's a good idea (I doubt we ever voluntarily consume less as a species), but data centres use a lot of energy and billions are being spent building them. https://www.technologyreview.com/2024/09/26/1104516/three-mi...
https://washingtonstatestandard.com/2025/04/11/as-demand-for...
To make it worse, the model training cost only refers to the cost of the training itself. The externalities - everyone else being forced to drastically upscale their compute power because scraper blocking isn't foolproof - are, as usual for hypercapitalism, conveniently ignored.
AI training in its current form is unsustainable, I'd go as far to say it's a threat for the open and decentralized Internet as you have all but zero chance of standing alone against the flood of training scraper bots and more and more control gets ceded to Cloudflare et al.
> Training GPT-4 used 50 GWh of energy. Like the 20,000 households point, this number looks ridiculously large if you don’t consider how many people are using ChatGPT.
> Since GPT-4 was trained, it has answered (at minimum) about 200,000,000 prompts per day for about 700 days. Dividing 50GWh by the total prompts, this gives us 0.3Wh per prompt. This means that, at most, including the cost of training raises the energy cost per prompt by 10%, from 10 Google searches to 11. Training doesn’t add much to ChatGPT’s energy cost.
https://andymasley.substack.com/i/162196004/training-an-ai-m...
Those are 200M/d prompts that wouldn't happen without the training.
A bus emits more CO2 than a car. Yet it is more friendly to the environment because it transports more people.
> Those are 200M/d prompts that wouldn't happen without the training.
Sure, but at least a few millions are deriving value from it. We know this because they pay. So this value wouldn't have been generated without the investment. That's how economics work.
Training a single LLM takes a few dozen fully loaded transatlantic passenger aircraft trips worth of power.
For "counties worth of power" I think you might be talking ALL data center use as a whole.
Human brain: Uses about 20 watts of power.
ChatGPT (GPT-4): Running a single query can use hundreds of watts when accounting for the entire datacenter infrastructure (some estimates suggest 500–1000 watts per query on average, depending on model size and setup).
If we assume:
20 watts for the human brain thinking continuously,
1000 watts for ChatGPT processing one complex query,
then the human brain is about 50x more energy-efficient (or 5000% more efficient) than ChatGPT per task, assuming equal cognitive complexity (which is debatable, but good for ballpark comparison).
The brain may ultimately be more power-efficient, but the units you want are watt-hours.
You can try explaining why it's not "that" bad for the environment, the planet is still worse off than when it didn't exist.
Let's carry on inventing new ways to spend energy, but it's ok because we still spend more energy for other stuff.
It's kinda sad how the world saw climate change, said it was bad, but in the end decided to do nothing about it.
If we are to believe that the models will get bigger, use more tokens, work for longer, this calculation can easily become very very skewed in the other direction.
Consider an agentic system that runs continuously for 6 hours. It is possible this system processes billions of tokens. That could more than equal a transatlantic flight in this hypothetical world.
Now compare this with non-AI work, like a CRUD app. Serving millions of queries in that same period would consume a tiny fraction of what ChatGPT consumes.
Rather than being a “win” for AI, the fact that we’re even 3 or 4 orders of magnitude away from this being a problem means that its already grounds to be concerned.