- cross-posted to:
- [email protected]
- cross-posted to:
- [email protected]
The University of Rhode Island’s AI lab estimates that GPT-5 averages just over 18 Wh per query, so putting all of ChatGPT’s reported 2.5 billion requests a day through the model could see energy usage as high as 45 GWh.
A daily energy use of 45 GWh is enormous. A typical modern nuclear power plant produces between 1 and 1.6 GW of electricity per reactor per hour, so data centers running OpenAI’s GPT-5 at 18 Wh per query could require the power equivalent of two to three nuclear power reactors, an amount that could be enough to power a small country.
You must log in or register to comment.
There’s such a huge gap between what I read about GPT-5 online, versus the overwhelmingly disappointing results I get from it for both coding and general questions.
I’m beginning to think we’re in the end stages of Dead Internet, where basically nothing you see online has any connection to reality.
The stock market is barely connected to reality and that is required to be updated every 3 months by every single company. Just imagine what the internet’s going to be like.
Well yeah, it’s a for-profit company. They exist solely to make money, that’s their entire goal.
It’s almost all marketing and has been for a while. ChatGPT peaked with 4o (and 4.5 if you used their API), 4.1 was a step backwards despite them calling it an improvement, and 5 was another step backwards.
They are not any closer to AGI, and we’re not going to see AGI from LLMs no matter how much they claim just how close we are to seeing AGI.
People who fawn over generative AI haven’t tried to use it for more than 5 seconds. I wish it could run a ttrpg game for me or even just remember the details of its original prompt but its not even close.
And how many milliwatts does an actual brain use?
1.21 jigawatts
Ahh but you need 8 for a round trip
It’s shockingly tricky to answer precisely, but the commonly held value is that a human brain runs on about 20w, regardless of the computational load placed on it.
So maybe .5 kwh/d. Still vastly more efficient than a code-based one.
Isn’t this the back plot of the game, Rain World? With the slug cats and the depressed robots stuck on a decaying world when the sapient, organic species all left?
I didn’t know there was such a backstory
Spoilers dude.
For reference, this is roughly equivalent to playing a PS5 game for 4 minutes (based on their estimate) to 10 minutes (their upper bound)
calulation
source https://www.ecoenergygeek.com/ps5-power-consumption/
Typical PS5 usage: 200 W
TV: 27 W - 134 W → call it 60 W
URI’s estimate: 18 Wh / 260 W → 4 minutes
URI’s upper bound: 48 Wh / 260 W →10 minutes
It is also the equivalent of letting a LED light bulb run for an entire day (depending on bright it is, some LED bulbs use under 2 watts of power).
I love playing PS5 games!
I was just thinking, in more affordable electric regions of the US that’s about $5 worth of electricity, per thousand requests. You’d tip a concierge $5 for most answers you get from Chat GPT (if they could provide them…) and the concierge is likely going to use that $5 to buy a gallon and a half of gasoline, which generates a whole lot more CO2 than the nuclear / hydro / solar mixed electrical generation, in reasonably priced electric regions of the US…
That doesn’t seem right. By my calculations it should be like 5¢. Can you show your work?
The last 6 to 12 months of open models has pretty clearly shown you can substantially better results with the same model size or the same results with smaller model size. Eg Llama 3. 1 405B being basically equal to Llama 3.3 70B or R1-0528 being substantially better than R1. The little information available about GPT 5 suggests it uses mixture of experts and dynamic routing to different models, both of which can reduce computation cost dramatically. Additionally, simplifying the model catalogue from 9ish(?) to 3, when combined with their enormous traffic, will mean higher utilization of batch runs. Fuller batches run more efficiently on a per query basis.
Basically they can’t know for sure.
I have an extreme dislike for OpenAI, Altman, and people like him, but the reasoning behind this article is just stuff some guy has pulled from his backside. There’s no facts here, it’s just “I believe XYX” with nothing to back it up.
We don’t need to make up nonsense about the LLM bubble. There’s plenty of valid enough criticisms as is.
By circulating a dumb figure like this, all you’re doing is granting OpenAI the power to come out and say “actually, it only uses X amount of power. We’re so great!”, where X is a figure that on its own would seem bad, but compared to this inflated figure sounds great. Don’t hand these shitty companies a marketing win.
Yeah, I saw some people speculating GPT-5 is having worse performance in some tasks than older models but still being forced to everyone because it’s a consolidation and cost cutting version, but how could it be cutting costs if it consumes 8 times more power?
This figure is already not bad. 40 watt hours = 0.04kWh - you know kWh? That unit on your electric bill that is around $0.18 per kWh (and data centers tend to be in lower cost electric areas, closer to $0.11/kWh.) Still, 40Wh would register on your home electric bill at $0.0072, less than a penny. For comparison, an average suburban 4 ton AC unit draws 4kW - that 40Wh request? 1/100th of an hour of AC for your home, about 36 seconds of air conditioning. I don’t know that this article is making anybody “look bad” in terms of power used.
What exactly do you get for that power though?
The point is that it’s too much power for little gain in return.
Thats actyally a fav rhetorical trick of mine when arhuing with consummatw bullshitters who have followers.
And an LLM that you could run local on a flash drive will do most of what it can do.
Can you give an example?
I mean no not at all, but local LLMs are a less energy reckless way to use AI
Why not… for the ignorant such as myself?
AI models require a LOT of VRAM to run. Failing that they need some serious CPU power but it’ll be dog slow.
A consumer model that is only a small fraction of the capability of the latest ChatGPT model would require at least a $2,000+ graphics card, if not more than one.
Like I run a local LLM with a etc 5070TI and the best model I can run with that thing is good for like ingesting some text to generate tags and such but not a whole lot else.
How slow?
Loading up a website with flash and GIF in the 90s dialup slow… Or worse?
It’s horrendously slow, unusable imo. With the larger DeepSeek distilled models I tried that didn’t fit into VRAM you could easily wait 5 minutes until it was done writing its essay. Compared to just a few seconds when it does. Bit that’s with a RTX 3070 Ti, not something the average ChatGPT user has lying around probably.
Basicly I can run 9b models on my 16gb gpu mostly fine like getting responses of lets say 10 lines in a few seconds.
Bigger models if they don’t outright crash take for the same task then like 5x or 10x longer so long it isn’t even useful anymore
So very worse.
Like make a query and then go make yourself a sandwich while it spits out a word every other second slow.
There are very small models that can run on mid range graphics cards and all, but it’s not something you’d look at and say “Yeah this does most of what chatGPT does”
I have a model running on a gtx 1660 and I use it with Hoarder to parse articles and create a handful a tags for them and it’s not… great at that.
Probably not a flash drive but you can get decent mileage out of 7b models that run on any old laptop for tasks like text generation, shortening or summarizing.
What do you use your usb drive llm for?
Porn. Obviously.
I don’t care how rough the estimate is, LLMs are using insane amounts of power, and the message I’m getting here is that the newest incarnation uses even more.
BTW a lot of it seems to be just inefficient coding as Deepseek has shown.
For training yes, but during operation by this studies measure Deepseek actually has an even higher power draw, according to the article. Even models with more efficient programming use insane amounts of electricity
This was higher than all other tested models, except for OpenAI’s o3 (25.35 Wh) and Deepseek’s R1 (20.90 Wh).
OK I guess I didn’t read far enough but your quote says that Deepseek uses less than Open AI?
Less than Open AI’s o3, but that’s because o3 was estimated to use even more power than GPT 5’s 18 Wh per query.
BTW a lot of it seems to be just inefficient coding as Deepseek has shown.
Kind of? Inefficient coding is definitely a part of it. But a large part is also just the iterative nature of how these algorithms operate. We might be able to improve that via code optimization a little bit. But without radically changing how these engines operates it won’t make a big difference.
The scope of the data being used and trained on is probably a bigger issue. Which is why there’s been a push by some to move from LLMs to SLMs. We don’t need the model to be cluttered with information on geology, ancient history, cooking, software development, sports trivia, etc if it’s only going to be used for looking up stuff on music and musicians.
But either way, there’s a big ‘diminishing returns’ factor to this right now that isn’t being appreciated. Typical human nature: give me that tiny boost in performance regardless of the cost, because I don’t have to deal with. It’s the same short-sighted shit that got us into this looming environmental crisis.
Coordinated SLM governors that can redirect queries to the appropriate SLM seems like a good solution.
That basically just sounds like Mixture of Experts
Basically, but with MCP and SLMs interacting rather than a singular model, with the coordinator model only doing the work to figure out who to field the question to, and then continuously provide context to other SLMs in the case of more complex queries
Powered by GNU Hurd
And water usage which will also increase as fires increase and people have trouble getting access to clean water
https://techhq.com/news/ai-water-footprint-suggests-that-large-language-models-are-thirsty/
It would only take one regulation to fix that:
Datacenters that use liquid cooling must use closed loop systems.
The reason they dont, and why they setup in the desert, is because water is incredibly cheap and energy to cool a closed loop system is expensive. So they use evaporative open loop systems.
Closed loop systems require a large heat sync, like a cold water lake, limiting them to locations that are not as tax advantageous as dry red states.
Aw, that’s unfortunate for the big mega tech corps. Anyway.
Unfortunately I wonder if it’s more expensive to set up a closed loop system that’s really expensive or to buy lawmakers that will vote against bills saying you should do so and it’s a tale old as time
Politicians are cheap
Yeah sorry forgot my /s there
That increases your energy use though, because evaporative cooling is very energy efficient.
We can make energy from renewable sources.
Fresh drinking water is finite, especially in the desert.
Also don’t forget how people like wasting resources by asking questions like “what’s the weather today”.
My guess would be that using a desktop computer to make the queries and read the results consumes more power than the LLM, at least in the case of quickly answering models.
The expensive part is training a model but usage is most likely not sold at a loss, so it can’t use an unreasonable amount of energy.
Instead of this ridiculous energy argument, we should focus on the fact that AI (and other products that money is thrown at) aren’t actually that useful but companies control the narrative. AI is particularly successful here with every CEO wanting in on it and people afraid it is so good it will end the world.
Fucking Doc Brown could power a goddamn time machine with this many jiggawatts, fuck I hate being stuck in this timeline.
That are 25 request per kWh. At 10 to 25cents per kWh that’s 1cent per request. That doesn’t seem to be too expensive.
Hmmm. Sure. But I find people don’t understand how much one kWh really is. A 500W drill can twist your arm. Imagine yourself twisting someones arm with all you got for a whole hour. Or idk. Either way it’s a lot of energy.
And then you think about how much more energy a car uses then a human does. And then you find out about hot water…
Which is why they’re giving everybody free access, for now.
I think AI power usage has an upside. No amount of hype can pay the light bill.
AI is either going to be the most valuable tech in history, or it’s going to be a giant pile of ash that used to be VC capital.
It will not go away at this point. Too many daily users already, who uses it for study, work, chatting, looking things up.
If not OpenAI, it will be another service.
Those users are not paying a sustainable price, they’re using chatbots because they’re kept artificially cheap to increase use rates.
Force them to pay enough to make these bots profitable and I guarantee they’ll stop.
Or it will gate keep them from poor people. It will mean alot if the capabilities keep on improving.
That being said, open source models will be a thing always, and I think with that in mind, it will not go away, unless it’s replaced with something better.
I don’t think they can survive if they gatekeep and make it unaffordable to most people. There’s just not enough demand or revenue that can be generated from rich people asking for chatGPT to do their homework or pretend to be their friend. They need mass adoption to survive, which is why they’re trying to keep it artificially cheap in the first place.
Why do you think they haven’t raised prices yet? They’re trying to make everyone use it and become reliant on it.
And it’s not happening. The technology won’t “go away” per se, but these these expensive AI companies will fail.
Well, if they succeed, it’s because of efficiency and lowering costs. Second is how much the data and control is really worth.
The big companies is not just developing LLM’s, so they might justify it with other kinds of AI that actually makes them alot of money, either trough the market or government contracts.
But who knows. This is a very new technology. If they actually make a functioning personal assitant so good, that it’s inconvinient not to have it, it might work.
I can see government contracts making a lot of money regardless of how functional their technology actually is.
It’s more about who you know than what you can actually do when it comes to getting money from the government.
Those same things were said about hundreds of other technologies that no longer exist in any meaningful sense. Current usage of a technology, which in this specific case I would argue is largely frivolous anyway, is not an accurate indicator of future usage.
Can you give some examples of those technologies? I’d be interested in how many weren’t replaced with something more efficient or convenient.
https://en.wikipedia.org/wiki/Dot-com_bubble
There were certainly companies that survived, because yes, the idea of websites being interactive rather than informational was huge, but everyone jumped on that bandwagon to build useless shit.
As an example, this is today’s ProductHunt
And yesterday’s was AI, and the day before that it was AI, but most of them are demonstrating little value with high valuations.
LLMs will survive, likely improve into coordinator models that request data from SLMs and connect through MCP, but the investment bubble can’t sustain
Technologies come and go, but often when a worldwide popular one vanishes, it’s because it got replaced with something else.
So lets say we need LLM’s to go away. What should that be? Impossible to answer, I know, but that’s what it would take.
We cant even get rid of Facebook and Twitter.
BUT that being said. LLMs will be 100x more efficient at some point - like any other new technology. We are just not there yet.
@themurphy @rigatti There is one difference … LLM’s can’t be more efficient there is an inherent limitation to the technology.
https://blog.dshr.org/2021/03/internet-archive-storage.html
In 2021 they used 200PB and they for sure didn’t make a copy of the complete internet. Now ask yourself if all this information without loosing informations can fit into a 1TB Model ?? ( Sidenote deepseek r1 is 404GB so not even 1TB ) … local llm’s usually < 16GB …
This technology has been and will be never able to 100% replicate the original informations.
It has a certain use ( Machine Learning has been used much longer already ) but not what people want it to be (imho).
And most importantly the Pandora box has been opened for deep perfect scams and illegal usage. Nobody will put it in the box again, because even if everyone agreed to make it illegal everywhere it’s already too late.
That capital was ash earlier this year. The latest $40 Billion-with-a-B financing round is just a temporary holdover until they can raise more fuel. And they already burned through Microsoft, who apparently got what they wanted and are all “see ya”.
Bit of a clickbait. We can’t really say it without more info.
But it’s important to point out that the lab’s test methodology is far from ideal.
The team measured GPT-5’s power consumption by combining two key factors: how long the model took to respond to a given request, and the estimated average power draw of the hardware running it.
What we do know is that the price went down. So this could be a strong indication the model is, in fact, more energy efficient. At least a stronger indicator than response time.
That’s a terrible metric. By this providers that maximize hardware (and energy) use by having a queue of requests would be seen as having more energy use.
Isnt it just worse than 4 tho? If they didnt make it cheaper, nobody would pay…
They could have kept it at the same price, though.
Help me out here. What designates the “response” type? Someone asking it to make a picture? Write a 20 page paper? Code a small app?
Response Type is decided by ChatGPTs new routing function based on your input. So yeah. Asking it to “think long and hard”, which I have seen people advocating for to get better results recently, will trigger the thinking model and waste more resources.
So instead of just saying “thank you” I now have to say “think long and hard about how much this means to me”?
Once it generates the response, there is a button you can click to make it use the reasoning model.
Why they did it that way instead of giving users the option to just set the model that they want to use ahead of time boggles the mind. Surely it would be more efficient for them to chose a model if they want ahead of time, rather than generating something that’s going to be regenerated with the desired model instead.
FFS, I have been using Claude to code, not only do you have to tell Claude to fix compilation errors, you have to point out when Claude says “it’s fixed” - “no, it’s not, the function you said you added is STILL missing.”
You’re absolutely right!
If you want it to really use a lot of energy on receiving your gratitude, sure I guess^^
How the hell are they going to sustain the expense to power that? Setting aside the environmental catastrophe that this kind of “AI” entails, they’re just not very profitable.
Look at all the layoffs they’ve been able to implement with the mere threat that AI has taken their jobs. It’s very profitable, just not in a sustainable way. But sustainability isn’t the goal. Feudal state mindset in the populace is.
Not just”not profitable”, they don’t make any money at all. Loss only.
that’s a lot. remember to add “-noai” to your google searches.
This is my weekly time to tell lemmings about Kagi, the search engine that does not shove LLM in your face (but still lets you use it when you explicitly want it) and that you pay for with your money, not your data.
I’m just going to ignore the AI recommendations, let them burn money.
i don’t judge you for that. honestly it matters fuck all at this point
Or just use any other better search like Bing or duckduckgo. googol sucks and was never any good. Quit pushing ignorant garbage.
googol sucks and was never any good.
Ha! Kids these days.
duckduckgo yes, but … bing?
ddg is bing
Bing is for porn.
