Estimating AI energy use
spectrum.ieee.org
spectrum.ieee.org
- 25 LLM queries: ~8.5 Wh
- driving one mile: ~250-1000 Wh
- one glass bottle: ~1000 Wh [2]
- a new laptop: ~600,000 Wh [3]
- round-trip flight from LA to Tokyo: ~1,000,000 Wh
[1] https://ourworldindata.org/energy-production-consumption
[2] https://www.beveragedaily.com/Article/2008/03/17/study-finds...
[3] https://www.foxway.com/wp-content/uploads/2024/05/handprint-...
Just look at your own life and see how much of each you would use.
We need cheaper and cleaner forms of energy. More efficient uses of energy.
I do not agree that we "all" need to use less energy overall. Energy use tracks wealth pretty closely, and manufacturing/creating things tends to be energy intensive.
The more cheap clean energy we make available, the more novel uses will be found for it.
Yep that's the dream, but it's not what I have coming out of my wall right now.
> Energy use tracks wealth pretty closely,
I'm guessing the majority of users on this site are in the 1% globally so it seems reasonable to consider what's produced/manufactured for us and what services like these that we're using
> The more cheap clean energy we make available, the more novel uses will be found for it.
That will be a brilliant future but it's not the reality today.
It would be better to not use this energy, but it won't move the needle either way.
Scratch that, the latter is probably more reliable.
[2] does not cite energy use, only CO2 emissions
Where I live, the utility company bills you at a higher rate if you use more electricity.
You need strong residential consumer protections to avoid this.
Every recent "datacenters are evil" news segment/article these days has a section quoting a few local NIMBYs talking about how they are opposing more transmission lines in their area for various reasons. Then these same folks (usually literally the same person) is quoted as saying that they are "for" investing into the grid and understands America needs more capacity - just not here.
It's pretty frustrating to watch. There are actually large problems with the way many local communities are approving datacenter deals - but people cannot seem to put two and two together why we are where we are. If everyone vetos new electrical infrastructure in their community, it simply doesn't get built.
wait what? consumers are literally paying for server farms? this isn't a supply-demand gap?
In the case of your quoted article - taking it at face value - this means "everyone" is paying .02/khw more on their bill. A datacenter is going to be paying thousands of times more than your average household as they should.
I don't see a problem with this at all. Cheap electricity is required to have any sort of industrial base in any country. Paying a proportionate amount of what it costs the grid to serve you seems about as fair of a model as I can come up with.
If you need to subsidize some households, then having subsidized rates for usage under the average household consumption level for the area might make sense?
I don't really blame the last watt added to the grid for incremental uptick in costs. It was coming either way due to our severe lack of investment in dispatchable power generation and transmission capacity - datacenters simply brought the timeline forward a few years.
There are plenty of actual problematic things going into these datacenter deals. Them exposing how fragile our grid is due to severe lack of investment for 50 years is about the least interesting one to me. I'd start with local (and state) tax credits/abatements myself.
Data centers get commercial or maybe even industrial rates depending on their grid hookup and utilities love predictable loads. Those are lower than residential rates. If you're dishonest and don't understand the cost of operating a grid, you could say that's users paying for data centers. But then you'd need to apply it to every commercial/industrial user.
If the regular users were paying for data centers usage, why are so many of them going off-grid with turbines or at least partially on-prem generation?
The solution is more and cheaper energy.
Ofc, possible long term contracts and options are involved in some of these markets. But there the option sellers would bear the cost.
In Europe it is constantly >"why does the households of half of Europe pay for German unwillingness to have a good power mix? Why should anyone want more cross country or long range interconnects if it drives up local prices?"
Say Norway with abundant hydropower, they should by all right have cheap power. But reality is not so in half of the country because they're sufficiently interconnected to end up on a common bidders euro market and end up paying blood money for the poor political choices of countries they don't even share a border with.
Addition: this also creates perverse incentives. A good solution for many of the interconnected flat euro countries would love enormous hydropower overcapacity to be built in Norway at the cost of the local nature. This is great for whoever sells the hydropower. This is great for whoever is a politician that can show off their green all-hydro power mix in a country as hilly as a neutron star. But this is not great for whoever gets their backyard hiking trails reduced to a hydro reservoir.
But hey we do it with everything else too, "open pit mines are too destructive to have in our country, so we'll buy it from china and pretend we're making green choice. Globalism in a nutshell: Export your responsibility.
PJM claims this will be a 1.5-5% yoy increase for retail power. https://www.pjm.com/-/media/DotCom/about-pjm/newsroom/2025-r...
Datacenters are simply the final straw/tipping point, and make a convenient scapegoat.
At some point you run out of the prior generation's (no pun intended) energy investments. Efficiency gains only get you so far, eventually you need capital investment into actually building things.
Current gen AI is going to result in the excess datacenter equivalent of dark fiber from the 2000's. Lots of early buildout and super investment, followed by lack of customer demand and later cheaper access to physical compute.
The current neural network software architecture is pretty limited. Hundreds of billions of dollars of investor money has gone into scaling backprop networks and we've quickly hit the limits. There will be some advancements, but it's clear we're already at the flat part of the current s-curve.
There's probably some interesting new architectures already in the works either from postdocs or in tiny startups that will become the base of the next curve in the next 18 months. If so, one or more may be able to take advantage of the current overbuild in data centers.
However, compute has an expiration date like old milk. It won't physically expire but the potential economic potential decreases as tech increases. But if the timing is right, there is going to be a huge opportunity for the next early adopters.
So what's next?
Another possibility is that new breakthroughs significantly reduce computational needs, efficiency significantly improves, or some similar improvements that reduce DC demand.
New nuclear is too expensive to make sense. At most there are small investments in flash-in-the-pan startups that are failing to deliver plans for small modular reactors.
The real build out that will happen is solar/wind with tons of batteries, which is so commonplace that it doesn't even make the news. Those can be ordered basically off the shelf, are cheap, and can be deployed within a year. New nuclear is a 10-15 year project, at best, with massive financial risk and construction risk. Nobody wants to take those bets, or can really afford to, honestly.
That actually sounds awesome, is there a downside I’m not seeing?
From where I'm standing, the immediate capital seems to be being deployed at smaller-scale (2-5MW) natural gas turbines co-located on site with the load. I haven't heard a whole lot of battery deployments at the same scale.
Of course turbines are now out at 2029 or something for delivery.
Only marginally at the edge of this space these days though, so what I hear is through the grapevine and not direct any longer.
https://www.eia.gov/todayinenergy/detail.php?id=65964
Of course, remember that nameplate capacities from different technologies should be corrected for capacity factor, which is roughly 60% for gas, 40% for wind, and 25% for solar, but pre-correction EIA expects
solar: 33.3GW
battery: 18.3GW
wind: 7.7GW
gas: 4.7GW
I'm pretty disconnected from the data center folks, but in general the current political environment is highly disfavorable to solar and batteries, and using them too much could have lots of political blowback that is very expensive.
Of course, small gas also has the benefit that the operating costs are spread over the lifetime, rather than being an up-front cost. So even if solar+batteries is cheaper than gas over the lifetime of the system, gas may seem more expedient if you don't want a lot of capital on the books.
It is not clear to me why we will have a breakthrough after virtually no movement on this front for decades. Backpropagation is literally 1960s technology.
There is an ungodly number of engineering decisions that go into making ML work and any number of stupid things are all over the place that cause stuff to fail.
Like something stupid like your normalization was bad or your mix of data was bad or your learning rates were bad or you have some precision issues or your model has a bad init or some architectural problems cause poor training or straight up there are tons of bugs somewhere like your batching was doing something silly or there is some numerically unstable division or sqrt or somewhere etc etc.
At scale with stupid issues like hardware faults I imagine this only gets exponentially worse.
And then on product sides of integrating stuff more bugs sneak in like so many labs were releasing so many open source LLMs with broken and incorrectly configured chat templates that massively tanked performance.
Or they set up some parmeters in sampling wrong and stuff gets stuck in loops or hallucinates tons or something.
In his 2025 hotchips keynote Noam Shazeer (GDM VP) even says that you need hardware determinism because there are just so many bugs in ML experiments that you need to be able to tweak and test things.
Also there are just so many obvious issues with the way everything works conventionally in GPT2 style like with softmax causing attention sinks at punctuation and creating dispersion over longer sequences because of low sharpness and the whole previllaged basis thing making it so common information takes up a lot of model capacity.
It is so easy to have good ideas broken by random bugs everywhere...
Also adding to that tendency, I suspect as the tech matures more and more consumer space models will just run on device (sure, the cutting edge will still run in server farms but most consumer use will not require cutting edge).
It's going to move fast I think and I would not surprised if the inference cost in energy is 1/10 of today in less than 5 years.
Seems ... low? And it will only get more efficient going forward.
I don't get why this is supposed to be a big deal for infrastructure since there's definitely way more than 14,000 EVs out there and we are doing well.
Before GenAI we were on our way to optimizing this, at least to the level where the general public can turn a blind eye. It was to the point where the companies would brag about how much efficient they are. Now all that progress is gone, and we're accelerating backwards. Maybe that was all a lie too. But then who's to say the current numbers are a lie too to make the pill easier to swallow.
What % of EVs on the market is 14,000?
what do you think the odds of this being accurate are?
zero?
GPT-5 came out on 7th August.
Assuming the 0.34 value was accurate in the GPT-4o era, is the number today still in the same ballpark or is it wildly different?
if I was altman then I'd release a few small numbers to try and get influencers talking about "how little energy chatgpt uses"
and he can never be accused of lying, as without any methodology as to how it was calculated it's unverifiable and completely meaningless
win-win!
...but then Sora came out.
a) training is where the bulk of an AI system's energy usage goes (based on a report released by Mistral)
b) video generation is very likely a few orders of magnitude more expensive than text generation.
That said, I still believe that data centres in general - including AI ones - don't consume a significant amount of energy compared with everything else we do, especially heating and cooling and transport.
Pre-LLM data centres consume about 1% of the world's electricity. AI data centres may bump that up to 2%
I don't think it shows that training uses more energy than inference over the lifetime of the model - they don't appear to share that ratio.
The amount of energy is insane.
At the end of the day green energy is perfect for AI and AI workloads.
Ok, but heating and cooling are largely not negotiable. We need those technologies to make places liveable
LLMs are not remotely as crucial to our lives
I also wouldn't be surprised if Aramco and Rosneft gave similar estimates on global warming and oil's role in it.
The prompt energy question is something that companies can both actively measure and need to actively measure in order to plan their budgets. It's their job to know the answer.
Those other questions, sadly, fall more into the category of it's their job not to know the answer.
Then when you take into account the amount of water used to cool the data centers as well as part o extraction and production process? Things get insane then https://www.forbes.com/sites/cindygordon/2024/02/25/ai-is-ac...
"How much energy does Google’s AI use? We did the math": https://cloud.google.com/blog/products/infrastructure/measur...
Edit: I asked ChatGPT-5:
https://chatgpt.com/share/68e36c19-a9a8-800b-884e-48fafbe0ec...
it says:
>the manufacturing of GPUs and datacenters themselves consumes a large amount of energy, not just their operation. The operational energy use (for AI training, inference, cooling, etc.) gets most of the attention, but the embodied energy — the energy used to extract raw materials, manufacture chips and components, and construct facilities — is substantial.
and summarizes it with:
4. Bottom Line
• Manufacturing GPUs and datacenters is highly energy-intensive, but operational energy dominates over time.
• For a single GPU, embodied energy ≈ 0.5–1 MWh.
• For a datacenter, embodied energy ≈ 6–12 months of its operational energy.
• Given AI-scale deployments (millions of GPUs), the embodied manufacturing energy already reaches terawatt-hours globally — roughly comparable to the annual electricity use of a small country.OpenAI’s Hunger for Computing Power Has Sam Altman Dashing Around the Globe
Thanks a lot, Chris!
Projections estimate anywhere between 10GW to 30GW of US datacenter buildup over the next few years
1GW of continuous power can support uniform draw from ~2.6M Tesla Model 3s assuming 12,000 miles per year, 250Wh/mile.
So 26M on the lower end, 80M Model 3s on the upper end.
That's 10x-30x the cumulative number of Model 3s sold so far
And remember all datacenter draw is concentrated. It will disproportionately going to impact regions where they're being built.
We need new, clean power sources yesterday
+332TW is like... +1% of US power consumption, or +8% of US electricity. If AI bubble burst ~2030... that's functionally what US will be left with (assuming new power infra actually built) mid/long term since compute depreciates 1-5 years. For reference dotcom burst left US was a fuckload of fiber layouts that lasts 30/40/50+ years. Still using capex from railroad bubble 100 years ago. I feel like people are failing to grasp how big of a F US will eat if AI bursts relative to past bubbles. I mean it's better than tulip mania, but obsolete AI chips also closer to tulips than fiber or rail in terms of stranded depreciated assets.
Bulk of AI capex build out is going to be in specialized hardware and data centers with bespoke power / cooling / networking profile. If current LLM approaches turn out to be deadend the entire data centre is potentially stranded asset unless future applications can specifically take advantage. But there's a good chance _if_ LLM crashes, then it might be due to something inherently wrong with current approach, i.e. compute/cost doesn't make commercial sense, and resuing stranded data centers might not make economic sense.
I honestly have no clue how much trust to place in data from a blog post written by a guy trying to make people give him lots of money. My gut is to question every word that comes out of his mouth but I'm maybe pessimistic in that regard.
But besides that, the cost of this stuff isn't just the energy consumption of the computation itself; the equipment needs to be manufactured, raw materials need to be extracted and processed, supplies and manpower need to be shuffled around. Construction of associated infrastructure has it's own costs as well. what are we, as a society (as opposed to shareholders and executives) going to get in return and is it going to be enough to justify the costs, not just in terms of cash but also resources. To say nothing of the potential environmental impact of all this.
In what world are these sensible units of power? Why can't we just use Watts FFS?
Btw:
1 MWh per day ≈ 42 kW
1 TWh per year ≈ 114 MW
Or:
All Chat GPT users: 850 MWh / day = 310 GWh / year ≈ 35.4 MW
All AI users: 15 TWh / year ≈ 1.7 GW