The reasoning effort knob is becoming a really important variable regarding how much impact AI models will have on energy consumption. Medium and max can differ by as much as two orders of magnitude in their token consumption. We also have to consider the amplification of compute demand that occurs by way of the tool calls themselves. Tokens are mostly a constant cost thing, but how they are interpreted can mean the difference between a 5 minute cpu bound operation and a 50uS cached database lookup. At some point, the CPU is going to become the bottleneck again.
Great writeup! I've done some work on Codecarbon [1] in the past, it's a pretty standard tool for measuring ML/AI workload emissions, it would be awesome to have your findings contributed to that project as well.
The comments section reminds me of The Dead Great-Grandfather Test
[Science fiction writer Bruce] Sterling was using the term specifically to describe the hairshirt green lifestyle that is driven by eco-anxieties. For hairshirt-green types, life is all about saving water, recycling, composting, reducing eco-footprints and various other behaviors marked by a kind of fearful, non-generative retreat from living. Permanent existential hibernation.
Sterling's rule of thumb for spotting acting-dead behaviors is a great one: if it's something your dead great-grandfather can do better than you, it's a case of acting dead. Your dead great-grandfather uses no water or plastic, and is actually recycling himself as we speak, not just his possessions. Try and top that.
But acting dead goes beyond hairshirt-green behaviors. While spartan frugality is a virtue, when it becomes the entire purpose of your life, there's a problem. For a portion of the dying American middle class, frugality has turned into a life purpose.
This is a fallacy. People 100 years like your great grandfather ago solved problems they had 100 years ago. People today solve problems they have today.
This is about a dead grandfather as he is decomposing in the ground. Not about his life which might have been more energy-intensive than that of many moderns.
Not sure about your great-grandfather, but mine didn't own a car, didn't travel on vacation by airplane, and, while he may have used a stove to heat his house, he probably only used it sparingly, because coal was expensive. So I doubt his life was more energy-intensive than mine.
It's a really dumb test. You can apply it to argue against reducing or stopping literally anything. My dead great grandfather uses no fentanyl and commits no murders. If we're not doing those, we're as good as dead!
Basically "Don't live beyond our collective means" = "Die".
Are you buying the latest iPhone every year? If not, you might as well be dead.
I wish people would realize that life is not about stuff. Living within your means, being frugal and focusing on what really matters (time with family, experiencing nature, thinking interesting thoughts) is not at all "permanent existential hibernation".
I resent the "The Dead Great-Grandfather Test" and similar "growth"-at-all-costs ideologies. Its implications are disgusting.
Yeah, the average Swede emitted 3.43 tons of CO2 in 2023, which is less than a quarter of what an average US citizen emitted. Does that mean that a US citizen is four times more "alive" than a Swede?
too many things can happen in the ai provider's backend which won't be observable from the outside. for starters how will you even know what hardware and inference backend is being used by RandomAiProvider3284.com? you can't. i looked at your berget-ai/co2-calculator repo and you're expected to provide this yourself. it also uses "measuredResponseTimeSeconds" as a metric even though it's impossible to know how much of that time was spent rate limiting the user during hours of high activity
> for starters how will you even know what hardware and inference backend is being used by RandomAiProvider3284.com? you can't.
From the article:
A typical request to the open model Kimi K3 emits about 0.4 grams of CO₂ in Sweden, 1.6 grams in Ireland, 2.5 grams in Germany and 3.2 grams in Texas, on identical hardware — 8.5 times more in Texas than in Sweden.
-> So the comparison is based on the assumption that identical hardware is used.
Also from the article:
Emissions estimates are based on the current average emissions intensity of each regional electricity grid, which has been the norm in greenhouse gas accounting.
And eating a single burger results in the equivalent Co2 emissions of 17,000 requests in Sweden and 2,000 requests in Texas.
Reducing/stopping meat consumption and replacing it with healthy plant-based tofu, seitan, tempeh etc has a far bigger impact (and as a bonus, living beings don't have to suffer)
0.5 Pwh/yr - worldwide data center energy
4-5 PWh/yr - meat total farm to table
166 PWh/yr - worldwide total energy production
the true bottleneck and cost we have is energy, carbon is a polite proxy
if you can meaningfully cut down the energetic cost, that's very valuable either way
maybe tomorrow you'll want to spin up a GPU on the cloud to briefly run a model, pick a data center:
DC #1: EU Efficient, but has ~100ms additional latency
DC #2: TX Runs on jet engine turbines, neighbors complain of noise and air pollution
env impact is something worth considering regardless of industry
That's actually kind of a scary statistic... I probably eat a burgur at most every month (beef maybe once a week at most) but probably make more than 2000 requests a day.
Your comment makes me really contextualize my AI usage and want to reduce it for ecological reasons.
I've seen people who hate AI instantly become defensive as soon as someone brings up the water usage for a kg of beef. It's selective outrage.
I don't like whataboutism, but this isn't whataboutism because the meat industry requires way more water per capita like you pointed out. You have to create products twice, once for feed and once to raise the cows.
The meat industry helps keep humans alive however, whereas the AI industry does not. I think the argument between the two needs to factor in the actual reason for the emissions.
Eating meat is ecologically unsustainable but it has an actual reason for it to exist.
Eating meat is not necessary by any means to 'be alive' in fact many studies show that a plant based diet is better for longevity. So, in the same way that AI is not 'necessary' to do a job, it's not necessary to eat meat. It's a matter of taste and societal conformity/norms.
Because some causes are significantly worse than others. If you want to have a big impact, you start with the big levers.
I'll give you an even better example than meat: golf courses in the US consume 2 to 100 times as much water as US datacenters depending on how you calculate.
How many golf-course-water-saving initiatives have you seen? Have they even tried anything at all?
Have a look at a graph of meat consumption in the US. We are also adding to that. That's my point, we don't need to add to meat consumption, we can make reductions.
The relative resource consumption of meat vs. plants is well established. Chicken is much more resource efficient than beef if you want to eat meat but consume less resources, and mussels are even more efficient.
That part I'm well aware of, I wanna know how the parent reached the figure 17000. What does count as a request even nowadays? When I tell an agent to make the button blue, how many requests is that? And how about when I tell it to throw together a whole website? Or rewrite the codebase in Rust?
“A typical request to the open model Kimi K3 emits about 0.4 grams of CO₂ in Sweden, 1.6 grams in Ireland, 2.5 grams in Germany and 3.2 grams in Texas, on identical hardware — 8.5 times more in Texas than in Sweden.”
0.4 x 17,000 = 6,800g, a fairly typical estimate for the co2 emissions of a burger.
Most people in the West eat too much meat. Red meat in particular is a group 2 carcinogen according to the WHO, so the body isn't meant to consume a lot of it. Yet many people eat meat every day, because "It's not a complete meal without it" or "I somehow need 10 times more protein than people a millenium ago while claiming to eat like them".
Worry not, AI people! Greenhouse gas exemptions will be made for the future of work, maths and art, the salvation of the humankind, the genie that is out of the bottle, etc, if only at the expense of taxing your and your dog's fart.
One of the reasons to build them near populated areas is that it makes it easier to repurpose waste heat to, for example, heat houses in colder climates.
Probably so they have no problems getting people who work there and provide technical services. I mean that seriously without any sarcasm or tongue-in-cheek, it seems the most likely explanation to me.
This is why people in left wing areas need to oppose datacenters bans. Regulate them. Require green energy investments that more than offset their load. Use some of that flowing AI money for the betterment of society.
You don't need to specifically regulate green investments for datacenter. Just introduce CO2 emission tax and the industry will find efficient ways to reduce CO2 emissions.
Datacenter wouldn't stop, and there would be some country or area which allows it. Unless they regulate the entire world, we should focus on lowering CO2 globally, and that means allowing datacenter in your neighour if your energy is better.
We can pretend that international trade and energy price arbitrage doesn't exist to make us feel nice and happy; but the reality is a Watt spent in Sweden increases the marginal cost of energy by the same amount as a Watt spent in Texas. These things are tied together by foreign exchange, imported energy intensive goods, petroleum products and directly imported power. It's time we stop being naive about total energy use.
Sweden is not grid connected to Texas, well, nor is also any other grid in the USA actually. Smelting was placed near cheap hydro resources (and why Boeing built planes in Seattle) for good reasons. Just replace cheap watts with clean watts.
Because Sweden's clean energy capacity is fixed. Which means a "clean" Watt spent on inference is taking away a "clean" Watt that would have been spent on heating or steel. Which means Sweden needs to buy energy or energy intensive goods on the global market, which raises prices, which means Texas is being paid.
This is a weird assumption to make. Why would you assume they wouldn't just build out more capacity as demand increases?
They certainly aren't buying electricity from Texas, this feels like a tortured argument you're just making to retroactively support whatever opinions you have on the pros or cons of clean energy.
There's only so much hydro power to build out, and nuclear takes forever. Solar only works in summer that far north, and without insane batteries wind power just becomes greenwashed natgas.
I'm a bit confused. For one I am not sure why this must hold for something as difficult to transfer as energy, and I am not sure why it matters. Not all energy is the same, hence the title.
The reasoning effort knob is becoming a really important variable regarding how much impact AI models will have on energy consumption. Medium and max can differ by as much as two orders of magnitude in their token consumption. We also have to consider the amplification of compute demand that occurs by way of the tool calls themselves. Tokens are mostly a constant cost thing, but how they are interpreted can mean the difference between a 5 minute cpu bound operation and a 50uS cached database lookup. At some point, the CPU is going to become the bottleneck again.
Great writeup! I've done some work on Codecarbon [1] in the past, it's a pretty standard tool for measuring ML/AI workload emissions, it would be awesome to have your findings contributed to that project as well.
[1] https://github.com/mlco2/codecarbon
This is true of ~everything that uses electricity, not just AI https://app.electricitymaps.com
The comments section reminds me of The Dead Great-Grandfather Test
[Science fiction writer Bruce] Sterling was using the term specifically to describe the hairshirt green lifestyle that is driven by eco-anxieties. For hairshirt-green types, life is all about saving water, recycling, composting, reducing eco-footprints and various other behaviors marked by a kind of fearful, non-generative retreat from living. Permanent existential hibernation.
Sterling's rule of thumb for spotting acting-dead behaviors is a great one: if it's something your dead great-grandfather can do better than you, it's a case of acting dead. Your dead great-grandfather uses no water or plastic, and is actually recycling himself as we speak, not just his possessions. Try and top that.
But acting dead goes beyond hairshirt-green behaviors. While spartan frugality is a virtue, when it becomes the entire purpose of your life, there's a problem. For a portion of the dying American middle class, frugality has turned into a life purpose.
https://ribbonfarm.com/2011/12/08/acting-dead-trading-up-and...
What a shitty way to dismiss anyone with any concerns whatsoever.
You can apply the same to FIRE if you want to trigger half this board.
Good that he didn't or the comment would be long [dead]!
This is a fallacy. People 100 years like your great grandfather ago solved problems they had 100 years ago. People today solve problems they have today.
This is about a dead grandfather as he is decomposing in the ground. Not about his life which might have been more energy-intensive than that of many moderns.
Not sure about your great-grandfather, but mine didn't own a car, didn't travel on vacation by airplane, and, while he may have used a stove to heat his house, he probably only used it sparingly, because coal was expensive. So I doubt his life was more energy-intensive than mine.
Mine drove a gas-guzzling GMC CCKW 353 truck, putting out up to a kilogram of CO2 per 1 km on inefficiently planned (but scenic) routes.
Is this some sort of straw man against trying to stay within your CO2 Budget? I really don’t understand the aim of this comment.
It's a really dumb test. You can apply it to argue against reducing or stopping literally anything. My dead great grandfather uses no fentanyl and commits no murders. If we're not doing those, we're as good as dead!
It's funny because this comment reminds me of something else... cope.
Basically "Don't live beyond our collective means" = "Die".
Are you buying the latest iPhone every year? If not, you might as well be dead.
I wish people would realize that life is not about stuff. Living within your means, being frugal and focusing on what really matters (time with family, experiencing nature, thinking interesting thoughts) is not at all "permanent existential hibernation".
I resent the "The Dead Great-Grandfather Test" and similar "growth"-at-all-costs ideologies. Its implications are disgusting.
Yeah, the average Swede emitted 3.43 tons of CO2 in 2023, which is less than a quarter of what an average US citizen emitted. Does that mean that a US citizen is four times more "alive" than a Swede?
Yes exactly what's the problem with profiting from child slavery if I only do it every 3 or 4 years? Pfff, get good guys.
[dead]
too many things can happen in the ai provider's backend which won't be observable from the outside. for starters how will you even know what hardware and inference backend is being used by RandomAiProvider3284.com? you can't. i looked at your berget-ai/co2-calculator repo and you're expected to provide this yourself. it also uses "measuredResponseTimeSeconds" as a metric even though it's impossible to know how much of that time was spent rate limiting the user during hours of high activity
> for starters how will you even know what hardware and inference backend is being used by RandomAiProvider3284.com? you can't.
From the article:
A typical request to the open model Kimi K3 emits about 0.4 grams of CO₂ in Sweden, 1.6 grams in Ireland, 2.5 grams in Germany and 3.2 grams in Texas, on identical hardware — 8.5 times more in Texas than in Sweden.
-> So the comparison is based on the assumption that identical hardware is used.
Also from the article:
Emissions estimates are based on the current average emissions intensity of each regional electricity grid, which has been the norm in greenhouse gas accounting.
And eating a single burger results in the equivalent Co2 emissions of 17,000 requests in Sweden and 2,000 requests in Texas.
Reducing/stopping meat consumption and replacing it with healthy plant-based tofu, seitan, tempeh etc has a far bigger impact (and as a bonus, living beings don't have to suffer)
why not both? some perspective:
the true bottleneck and cost we have is energy, carbon is a polite proxy if you can meaningfully cut down the energetic cost, that's very valuable either waymaybe tomorrow you'll want to spin up a GPU on the cloud to briefly run a model, pick a data center:
env impact is something worth considering regardless of industryThat's actually kind of a scary statistic... I probably eat a burgur at most every month (beef maybe once a week at most) but probably make more than 2000 requests a day.
Your comment makes me really contextualize my AI usage and want to reduce it for ecological reasons.
True, but with autonomous agents hitting 2,000 requests is a lot easier.
I want you to imagine someone like Jenovah's Witnesses trying to moralize people on the Internet. how successful do you estimate such effort might be?
A costs less than B so we should do A is hardly a moral argument.
I think that is a bit out of the blue. But it is fun to imagine a billboard-sized app for a referendum on burgers or AI.
That will put some sweat on some brows :)
Your comparison only makes sense if you also consider how many requests are being made.
The "only the biggest thing matters" is a faulty logic. If you truly care about the environment then we need to improve things across the board.
And moralizing about living beings doesn't help your cause here.
It's not an or/or is it.
I've seen people who hate AI instantly become defensive as soon as someone brings up the water usage for a kg of beef. It's selective outrage.
I don't like whataboutism, but this isn't whataboutism because the meat industry requires way more water per capita like you pointed out. You have to create products twice, once for feed and once to raise the cows.
The meat industry helps keep humans alive however, whereas the AI industry does not. I think the argument between the two needs to factor in the actual reason for the emissions.
Eating meat is ecologically unsustainable but it has an actual reason for it to exist.
Eating meat is not necessary by any means to 'be alive' in fact many studies show that a plant based diet is better for longevity. So, in the same way that AI is not 'necessary' to do a job, it's not necessary to eat meat. It's a matter of taste and societal conformity/norms.
Why not treat both as bad instead of pointing to whataboutism to make a point?
Because some causes are significantly worse than others. If you want to have a big impact, you start with the big levers.
I'll give you an even better example than meat: golf courses in the US consume 2 to 100 times as much water as US datacenters depending on how you calculate.
How many golf-course-water-saving initiatives have you seen? Have they even tried anything at all?
AI data centres are a thing we are choosing to add, we could just not do that.
What benefit is the AI summary on search giving me Vs the actual search for a start.
Have a look at a graph of meat consumption in the US. We are also adding to that. That's my point, we don't need to add to meat consumption, we can make reductions.
what benefit is it giving to the search engine is what you have to ask.
if you dont leave the page they get more ad-money
Eating more meat is a thing we are choosing to add, we could just not do that.
Flying more is a thing we are choosing to do, we could just not do that.
Source?
The relative resource consumption of meat vs. plants is well established. Chicken is much more resource efficient than beef if you want to eat meat but consume less resources, and mussels are even more efficient.
https://en.wikipedia.org/wiki/Environmental_impacts_of_anima...
That part I'm well aware of, I wanna know how the parent reached the figure 17000. What does count as a request even nowadays? When I tell an agent to make the button blue, how many requests is that? And how about when I tell it to throw together a whole website? Or rewrite the codebase in Rust?
Presumably based on the numbers in the article?
“A typical request to the open model Kimi K3 emits about 0.4 grams of CO₂ in Sweden, 1.6 grams in Ireland, 2.5 grams in Germany and 3.2 grams in Texas, on identical hardware — 8.5 times more in Texas than in Sweden.”
0.4 x 17,000 = 6,800g, a fairly typical estimate for the co2 emissions of a burger.
Could as well stop breathing to make way for the Super Intellect, that's what every sane transhumanist must do!
Giving up AI would be much easier and less disruptive.
Most people in the West eat too much meat. Red meat in particular is a group 2 carcinogen according to the WHO, so the body isn't meant to consume a lot of it. Yet many people eat meat every day, because "It's not a complete meal without it" or "I somehow need 10 times more protein than people a millenium ago while claiming to eat like them".
Giving up private jets and cruises would be less disruptive than giving up AI.
same for Germany which has roughly the same Co2 per kwh.
Site is HTTP 503. It does make sense that an nginx 503 would cost much less to serve.
Is this site getting a death hug? I got a 502, then 503, then incomplete page load.
Worry not, AI people! Greenhouse gas exemptions will be made for the future of work, maths and art, the salvation of the humankind, the genie that is out of the bottle, etc, if only at the expense of taxing your and your dog's fart.
it's remarkable how it only took one week of "AI bad for math" headlines for snarky redditors to incorporate it into their collective conscience.
This assumes people did not already have these thoughts/ideas ?
These snarky redditors, their go-to way of making an argument by picking at a single word in the whole sentence checks out every time!
Based on a newly released model for calculating this that is open sourced.
Yeah I don't understand why these new data centres are all in populated areas. AI isn't latency sensitive. Put them in some remote location surely?
One of the reasons to build them near populated areas is that it makes it easier to repurpose waste heat to, for example, heat houses in colder climates.
> don't understand why these new data centres are all in populated areas
Power connections.
Probably so they have no problems getting people who work there and provide technical services. I mean that seriously without any sarcasm or tongue-in-cheek, it seems the most likely explanation to me.
They need drinking water apparently.
[dead]
This is why people in left wing areas need to oppose datacenters bans. Regulate them. Require green energy investments that more than offset their load. Use some of that flowing AI money for the betterment of society.
You don't need to specifically regulate green investments for datacenter. Just introduce CO2 emission tax and the industry will find efficient ways to reduce CO2 emissions.
Or efficient ways around those regulations.
Less waste of resources is still waste of resources.
How much of the AI generated content is actually useful, how much useless, how much harmful?
Except you don't actually ban it, you just move it to texas.
That content is coming either way.
And if you require them to over offset their load to the grid, it's literally fossil fuels burned over all.
Datacenter wouldn't stop, and there would be some country or area which allows it. Unless they regulate the entire world, we should focus on lowering CO2 globally, and that means allowing datacenter in your neighour if your energy is better.
We can pretend that international trade and energy price arbitrage doesn't exist to make us feel nice and happy; but the reality is a Watt spent in Sweden increases the marginal cost of energy by the same amount as a Watt spent in Texas. These things are tied together by foreign exchange, imported energy intensive goods, petroleum products and directly imported power. It's time we stop being naive about total energy use.
Sweden is not grid connected to Texas, well, nor is also any other grid in the USA actually. Smelting was placed near cheap hydro resources (and why Boeing built planes in Seattle) for good reasons. Just replace cheap watts with clean watts.
Because Sweden's clean energy capacity is fixed. Which means a "clean" Watt spent on inference is taking away a "clean" Watt that would have been spent on heating or steel. Which means Sweden needs to buy energy or energy intensive goods on the global market, which raises prices, which means Texas is being paid.
> "Sweden's clean energy capacity is fixed"
This is a weird assumption to make. Why would you assume they wouldn't just build out more capacity as demand increases?
They certainly aren't buying electricity from Texas, this feels like a tortured argument you're just making to retroactively support whatever opinions you have on the pros or cons of clean energy.
There's only so much hydro power to build out, and nuclear takes forever. Solar only works in summer that far north, and without insane batteries wind power just becomes greenwashed natgas.
> this feels like a tortured argument you're just making to retroactively support whatever opinions you have on the pros or cons of clean energy.
My opinion that reality should take precedent over feel good lies?
> Because Sweden's clean energy capacity is fixed
By what?
You're treating electricity as a globally-tradable good. It's not.
Perhaps they are mistaking Sweden with Norway (whose nearly 100% hydropower system is constrained)
Most of electricity in Sweden is produced by:
1. Hydropower 40%
2. Nuclear 27%
3. Wind 22%
https://ourworldindata.org/profile/energy/sweden#what-source...
It is.
I'm a bit confused. For one I am not sure why this must hold for something as difficult to transfer as energy, and I am not sure why it matters. Not all energy is the same, hence the title.