Discussion
Does that use a lot of energy?
bena: My first question was: "Is this whitewashing LLM energy usage?"And yes, that seems to be the undercurrent here. Complete with linking to themselves to validate the data they used to make their estimates.Either these companies need to build these massive data centers that consume massive amounts of electricity OR these LLMs don't use a lot of electricity.You don't get both. If LLMs don't require a lot of electricity, then why are we building so much more capacity? If all of that capacity is required, then what is the real cost of sending a query to these LLMs?
quotemstr: I like the comparison concept. It's like that "order of magnitudes every programmer should know" list, but applied to anyone who cares about energy.That said, and hot take: people shouldn't worry about energy independent of what they pay for it. The whole point of a price is to fold a complicated manifold of scarcity-allocation into a set of scalars anyone can rank against each other. Appealing to people's sense of justice or duty to get them to use less energy than they'd otherwise be willing to buy is just asking them to lead a less utility-filled life than they can because you think you can allocate scarcity better than the market. I can't, and you can't either. Nobody can.If you claim that people should listen to moralized pleadings and not the market because prices don't internalize certain externalities, duty is on you to get those externalities accounted so they can properly factor into prices, not apply ad-hoc patches on top of markets by manipulating people's emotions.
mltvc: How do you propose to convince people to get those externalities accounted without emotions? How do you convince people of the value of externalities that are far away in place or time (but not less real)?
quotemstr: > You don't get both. If LLMs don't require a lot of electricity, then why are we building so much more capacity?A small number times a large number is often a large number. Have you heard of the concept called "per capita"? In any case, electricity is going towards data centers in proportion to the degree to which these data centers do useful work. AI companies buy the electricity fairly on an open market, sometimes even subsidizing this market by funding new generating capacity.If all these people and companies are making electricity allocation decisions that make sense to them with their own money, who are you to stop in and say that their voluntary transaction is incorrect? Who died and made you the king?
spencerflem: It’s new capacity!They’re not even saying they shouldn’t do it or that they’re not useful or not worth it but you Cannot logically say both “these things do not use a lot of power” and “we need to build more power plants to handle these things”
spencerflem: How am I as an individual supposed to get externalities priced in?And given that right now they are clearly not, what’s your plan until then?
alphazard: Which specific externalities are you concerned about? Do they affect you directly?
AgentME: LLMs don't use a lot of electricity per user. Why should the fact that the energy usage happens in data centers instead of each user's house be an important moral factor?
oulipo2: Indeed, looking at a "single median query" totally disregard the fact that:- first, those queries are mostly useless and we could totally do without them, so it's still a net pollution- they are being integrated everywhere, so soon enough, just browsing the web for a few hours is going to general 100k+ such equivalent "small queries" (in the background, by the processes analyzing what the user is doing, or summarizing the page, etc). At that time, the added pollution is no longer negligible. And most of this will be done just to sell more ads
janalsncm: Wow, putting everything in the same units is really informative. Running my 450 watt gpu for a day is approximately equivalent to driving a car 10 miles.
burkaman: > As for getting externalities internalized: as a society, we call the procedure for updating rules "politics", and it's as open to you as to anyone else.Ok so I do need to worry about energy so that I can identify these unaddressed externalities and work towards updating the rules. You can to care before you can get involved in this stuff. You can't tell me not to worry about it and then also say that it's basically my fault for not getting involved if the price is wrong.> any such rebuttal must involve numbers, not emotional appealsWho are you arguing with? You're commenting about a website that has strictly numbers and nothing else.
jwilliams: It was genuinely a surprise to see how much relative energy petrol cars use (and shame on me - I'm an electrical engineer). I mean I think I knew it intuitively, but this simple chart blew my mind.
MerrimanInd: When one gets in the weeds on EVs for ICE cars two things become shockingly clear: internal combustion is hilariously inefficient YET gasoline is hilariously energy dense. Most people's intuition is wrong on both of these points but then they cancel each other out.
Terr_: [delayed]
dr_dshiv: One hour of Claude code— well, I’d guess it would be comparable to an hour of driving an electric car. How to know?
MichaelDickens: OP says one query uses 0.3 Wh. Driving an electric car for 10 miles = 3,000 Wh which is roughly 10,000 Wh per hour.I'm not sure how many queries is equivalent to an hour of Claude code use, but maybe 5 seconds, which means an hour of continuous use = 216 Wh, or ~50x less than an electric car.OP has a longer article about LLM energy usage: https://hannahritchie.substack.com/p/ai-footprint-august-202...
loganp: This is neat. I think I'm actually more interested in avg ChatGPT query than median single query so that I can enter a large query # and be confident in the associated energy cost for that larger number (e.g. what's the energy cost for 1,000 chat gpt queries)
nikcub: I can't find a github or email for Hannah - if you're reading this i'd like to add Australian energy price data via Open Electricity[0] to the data (reach out via my profile)[0] https://explore.openelectricity.org.au/
PowerElectronix: Useful work is debatable here, a lot of people just talk to the thing or use it instead of searching the internet.The owners surely think, or at least want us to think that it is very useful indeed, otherwise we'd see no point in burning through piles of investors cash to buy overpriced ram, storage, gpus, cpus, nics, secure the power to run it and then subsidise the users to use it.I do think that transaction is wrong and it's going to bite them in the ass in the long term, but I don't have the money to outbid them for the power. I do get to see them crash and burn when the investors get impatient.
jfengel: I'm surprised that cooling takes less energy than heating. I imagine that depends a lot on the temperature range; they only need so much to cool a room even on a "hot" day in the UK.Still... AC still feels like magic. I know how it works and understand the over-unity factor. But it feels like it ought to take enormous energy for it to work at all.
lm28469: Yeah without knowing the climate, temperature delta and insulation these values don't really mean much.
creeble: I doubt your 450w gpu runs at that wattage 24 hrs, unless you're mining with it.My (admittedly old) gpu+CPU idles around 50-75w.
janalsncm: During training it’s running at around 410. But yeah idling at 450 would be pretty crazy.
rsolva: It is not only about raw power consumption. Comparing driving an electric car with using AI only in kW hides a major point: Hyperscale datacenters are massively centralised, which brings it's own problems; a lot of energy is used for cooling, and water consumptions is enormous. Charging electric cars at home is distributed and does not suffer from the same problems as the centralised hyperscalers do. Also, running AI models at home is not much different than a gaming session :)
DoneWithAllThat: What’s startling to me is how many comments in this thread just take the provided values as gospel without asking questions that methodology answers either in the abstract or barely describes. Also going giving a cost for “United States” is absolutely nonsense - electricity, gas and gasoline prices vary widely across the country. There is no one cost for each, and the average is worthless for this kind of thing (especially since the average of each - gas vs. electric vs. gasoline cost - are independent variables that have no relation to each other on a region by region basis).Why are people so gullible?
Rebelgecko: I think specifically it's comparing gas heating vs AC. Heat pump heating would probably do better. In other words, it takes less energy to move heat inside/outside than to "create" it(With caveats like heat pumps are much less effective in extreme cold)
tl2do: In Japan, my country, this looks a bit different. A lot of electricity still comes from oil- and gas-fired plants. The mechanics differ (gas turbines vs. car engines), but in both cases we’re still relying on combustion. I suppose some countries have the same issue.
jolmg: > Desktop computer - 1 hour of use - 50 WhThat seems low...
dylan604: Train locomotives have used diesel powered generators that then powers electric motors. Would this be less efficient than battery powered EVs? Or better asked, what would be the most efficient use of gasoline?
ZeroGravitas: Nissan makes a range of these under the e-power branding:https://www.nissan-global.com/EN/INNOVATION/TECHNOLOGY/ARCHI...
sixo: Your dismissal of moral concerns is not convincing.Imagine a world where the only energy you do is use was generated by a stationary bike you had to ride yourself. You would, generally speaking, use that energy differently than energy you would pay for--you would generally reserve your effort for worthwhile things, and would be averse to farming energy yourself just to power frivolity or vice. How you determine what to put your energy into would explicitly be a moral question.Instead in our world we an abstractions conceals the source of the energy. But if the moral concerns from the first world had any weight, they haven't lost it now; if energy is anything short of completely free we should by the same logic be averse to expending energy on worthless work or vice. The human being is not a utility monster, but something very different, and moral questions of this sort are central to how it navigates the world, they should not be dismissed.
stdbrouw: Doesn't this argument hinge on equivocating between two different definitions of aversion, though? I'm averse to bananas, but that doesn't mean I think it's immoral to eat them. The moral dimension kicks in if somebody else had to ride that stationary bike for you, because then you'd be wasting their time on frivolities.
alnwlsn: I attached a generator with some supercaps and an inverter to a stationary bicycle a few years ago, and even though I mostly use it as a way to feel less guilty watching Youtube videos, it does give me a quite literal feel for some of the items on the lower end of the scale.- Anything even even halfway approaching a toaster or something with a heater in it is essentially impossible (yes, I know about that one video).- A vacuum cleaner can be run for about 30 seconds every couple minutes.- LED lights are really good, you can charge up the caps for a minute and then get some minutes of light without pedaling.- Maybe I could keep pace with a fridge, but not for a whole day.- I can do a 3D printer with the heated bed turned off, but you have to keep pedaling for the entire print duration, so you probably wouldn't want to do a 4 hour print. I have a benchy made on 100% human power.- A laptop and a medium sized floor fan is what I typically run most days.- A modern laptop alone, with the battery removed and playing a video is "too easy", as is a few LED bulbs or a CFL. An incandescent isn't difficult but why would you?- A cellphone you could probably run in your sleepAlso gives a good perspective on how much better power plants are at this than me. All I've made in 4 years could be made by my local one in about 10 seconds, and cost a few dollars.
RobinL: One thing missing but important to understand is the energy embodied in buying 'stuff'. At a very rough approximation, the cost of stuff, especially consumer goods manufactured cheaply, is quite a high percentage energy.When you look at people's energy usage, quite a lot of it ends up being the embodied energy in the stuff they buy. For quite a lot of people, it's probably the largest category of energy consumption. I once had a very rough go at calculating this here: https://www.robinlinacre.com/energy_usage/
jfengel: When it comes to the environment the most efficient use is to leave it in the ground.Hybrids work for trains because they are so large and don't need big swings of acceleration or to climb steep grades. They can run the diesel generators at maximum efficiency.Battery power would be better, because you can build even larger power plants running at higher heats and not have to haul them with you, but the costs of sufficient battery is too large, so far. That is changing.
einpoklum: Isn't it better for trains to just to draw from the electric grid?
ssl-3: The presentation is nice, but some of the conversions are questionable.For instance: The cost section, wherein 1kWh in the US is figured as having a cost of 9.7 cents.In reality, it's not that way at all. Unless we're fortunate enough to live in an area where we can walk over to the neighborhood generating station and carry home buckets of freshly-baked electricity to use at home, then we must also pay for delivery.On average, in 2025, electricity was 17.3 cents per kiloWatt-hour -- delivered -- for residential customers in the US.https://www.eia.gov/electricity/monthly/epm_table_grapher.ph...
philipkglass: I looked at the electric car example for the United States. It has 3 kilowatt hours priced at $0.51, 17 cents per kilowatt hour, which seems about right. The "petrol car" example at the top of the chart isn't powered by electricity so its cost number is not directly comparable to the things that consume electricity.
ssl-3: On the energy tab: It says that driving a petrol car 10 miles uses 10,000 Watt-hours, eg 10 kWh.On the costs tab, for the United States: It says that this has a cost of $0.97.97 cents ÷ 10kWh = 9.7 cents per kWh(I didn't look further than that. Perhaps I should have.)
philipkglass: [delayed]
MerrimanInd: That's a complicated question that unfortunately has quite a bit of "well it depends" in the answer. I worked in the auto industry for a long time - both doing engine development and EVs - so my opinions here are well-informed but not world expert.From a pure energy efficiency perspective you can't beat economies of scale. A stationary power plant (even ones that are just big gasoline engines) run at a constant load and RPM so they can be optimized for pure efficiency, they rarely have to start, warm up, and shut down, and they can use larger and more expensive exhaust aftertreatment systems. Most energy conversions grow more efficient with scale and this is no different. The locomotive powertrain works for a handful of reasons but one of them is you can build much more efficient engines that are optimized for a single constant speed and load. But most of the advancements in internal combustion engines over the last 20-30 years don't increase peak efficiency but increase the conditions in which they're efficient. Variable valve timing and lift are probably the most underrated and overpowered technologies that have transformed engines from having one narrow regime of high efficiency to running well over a huge range of the map. But turbocharging, variable intake geometries, 7+ speed transmissions, and mild hybrid systems like belt-starter-generators get honorable mentions here. However we're not talking about anything close to EV-levels of efficiency. I think the cutting edge research engines are running in the mid to high 40s for thermal efficiency (percentage of fuel energy captured as useful work), most passenger car engines probably peak in the mid 30s.So while there is some efficiency to be gained by a more locomotive-style system it's not as much as you would hope. In the industry that's called a series hybrid system, vs a parallel hybrid system where either ICE or EV power can go to the wheels. The benefits of a series system are more emissions and product features. You can get the full torque and power of an EV, you can start and stop the IC engine in a more emissions optimized way, and and you can filter load spikes to use a small engine that meets average not peak load.From a more pragmatic perspective, with the energy density of gasoline and other liquid fuels it's probably best to use it in applications for which you just can't use full electrification. Planes are currently the best example of this. It's also worth noting that passenger cars benefit massively from strong hybridization because of the uneven load cycles so that's a technology where you can deploy a gasoline engine but then claw back a lot of the efficiency losses with hybrids. That's not always true, for example boats don't really have a regen cycle so hybridization just doesn't get much.
groundzeros2015: Exactly. The environmental/social burden isn’t just the energy used in the raw physical form, but the cost to acquire and make it useable.The problem with gas is not that burning it doesn’t maximally capture all energy, but that there are externalities to doing so.
medi8r: Glad it has AI. AI has nothing on cars. Save a car trip a week even if electric is way more than 10k queries.
petercooper: Or even a 5 minute shower. About 2500 queries.
20after4: Yes, similarly in the US: I think the largest portion of the energy in the US is produced with gas fired power plants.
nightski: 1 chatgpt query is a little misleading though. Let's see an 8 hour full bore claude code agent session. Or maybe running 3 agents for several hours a day.
fwip: It also doesn't include the amortized cost of training the models, as far as I can tell. I believe I heard that training the models took more energy than total queries against that model, but I could be mistaken.
kimixa: I believe training currently costs significantly more than inference to all the current vendors, so I'd be surprised if it doesn't also use more power.And by the look of it, that'll be the norm pretty much forever - unless something fundamental about how models can be trained/updated, an "older" model loses value as it's knowledge becomes out of date, even if we no longer get improvements from other sources or techniques.But other things likely change based on "lifetimes" and usage patterns too - e.g. a large battery for an electric car may have a higher upfront energy cost in manufacturing than a small ICE + fuel tank, but presumably there's a mileage that the improved per-mile efficiency overcomes that, and then continues to gain with each additional mile.
jborichevskiy: Amazing stuff, have you written up a blog post? I could see a video being a fun format for this as well. Might help people develop the intuition for watts/power consumption in a different way
alnwlsn: Kind of, it's in bits and pieces here:https://hackaday.io/project/191731-practical-power-cyclingand is also a few years out of date
philipkglass: The author Hannah Ritchie works on Our World In Data and also publishes the fantastic Sustainability by Numbers substack. It's in the same vein as the late, great David MacKay's Sustainable Energy Without the Hot Air.This tool has its own recent substack post. See the comments too, especially the one by Chris Preist that contextualizes the energy usage of streaming video (a topic that has also been discussed on HN before).https://hannahritchie.substack.com/p/does-that-use-a-lot-of
0x53: And wrote a great book: Not the End of the World
sien: Yep. It's a very good book and well worth a read.It's interesting to see how upset people are on Goodreads about that book:https://www.goodreads.com/book/show/145624737-not-the-end-of...The top reviews are mostly people angry with Ritchie for not being a catastrophist.
groundzeros2015: I don’t think the problem Is how many joules of energy are used. But the cost/burdnen to produce them. The costs and forms for each of these examples is very different making their energy use incomparabale.
csours: Do you have to run new electric transmission lines? Will you have to maintain those power lines?
mrguyorama: We once dammed basically every river in the nation because it was in vogue at the time.Maybe building overhead power lines for rail infrastructure should be the "hip" thing right now instead of AI. Maybe building oodles of solar power farms and batteries should be "hip"We built electrical infrastructure to the most remote residences just because we could and because it was an investment in our people. We directly funded our massive and formerly world class rail network because we could, and because it would pay off. We built a world class road network half as a make-work project, and it still pays dividends. We purchased Alaska, with no obvious reason. We built a space program to have slightly better nuclear weapons, and it's part of the reason we were so dominant in computer chips for so long.We have spent something like 40 trillion dollars over the past 25 years, and almost none of it on anything of real value. More than a little of that debt is just handouts to already rich people.We can build new electric transmission lines and I'm so tired of things that we absolutely 100% can do if we just demand it be done being somehow treated as a problem. America can afford infrastructure.
Sharlin: For reference it would be good to have per-passenger numbers for "sitting on a diesel bus", "sitting on an electric bus", "sitting on a tram", "sitting on a commuter train" as well.
20after4: The marginal cost of one extra passenger is going to be very nearly zero. The vast majority of the cost is just moving the bus / train / plane, and the overhead / inefficiencies in the system. I've seen somewhere the numbers for one passenger on trains and planes but I can't remember where that was. Just know it is a very very small amount for the added weight of one more passenger.
