The original press release and report are at [1], couldn't find a link to them in the article.
> In total, the median prompt—one that falls in the middle of the range of energy demand—consumes 0.24 watt-hours of electricity
If they're running on, say, two RTX 6000s for a total draw of ~600 watts, that would be a response time of 1.44 seconds. So obviously the median prompt doesn't go to some high-end thinking model users have to pay for.
It's a very low number; for comparison, an electric vehicle might consume 82kWh to travel 363 miles. So that 0.24 watt-hours of energy is equivalent to driving 5.6 feet (1.7 meters) in such an EV.
When I hear reports that AI power demand is overloading electricity infrastructure, it always makes me think: Even before the AI boom, shouldn't we have a bunch of extra capacity under construction, ready for EV driving, induction stoves and heat-pump heating?
> When I hear reports that AI power demand is overloading electricity infrastructure ...
It feels like dog-whistle tactics. "Aren't the technology companies bad for the environment!" "What about the water usage?" "What about the electricity?"
For me the peak of this is complaining about water consumption at the Dalles datacentre [0]. The buildings are next to the Colombia river and a few miles away from the Dalles Dam [1] which generates an average of 700MW. The river water should be used for cooling, taking out some of the water, warming it up by a few degrees and returning it to the river; one might argue that this is simply returning the heat to the river that would have come from the water flowing downhill.
What's the dog whistle? People are concerned about the impact industry has on the environment, and they are stating those concerns plainly. I don't think the non-profit WaterWatch's true goal is to destroy big tech.
I think you're oversimplifying the "just use rivers" idea. Most data centers (80% for Google) require potable water for cooling, and it can't come straight from a river. Plus, using potable water in cooling adds mineral deposits to the water and will require treatment to be consumable again.
Because in most cases it kind of is. It's not that the H2O molecules are forcefully disintegrated, but most data centers use evaporative cooling, meaning that whatever water is fed to the datacenter through the municipal water system ends up as moisture in the atmosphere. This is in effect equivalent to a sizable leak in the water infrastructure.
Yes, but you can neither drink rainwater nor use it to cool a data center. And the rain may end up falling thousands of miles away. Excessive use of water reduces flow in natural bodies of water and can mess up local ecosystems.
Not sure I'd drink most river water either, and I would hope most data centers don't pull water straight from the aquifer (though maybe they do). Fair points though.
Surely all uses of water are part of the closed-loop water cycle? Other than launching it into space for astronauts to drink, and using it in permanent reactions like concrete hydration?
Drinking water, spraying it on crops, using it to clean a car, or using it to flush a toilet all end up with the water evaporating, or making its way to the ocean and evaporating from there.
Ultimately, if a river provides a certain number of acre-feet of fresh water, evaporating it to cool a data centre uses it just as much as to evaporating it to grow alfalfa in a desert, except perhaps more usefully.
Fresh water isn't meaningfully a closed loop. We are draining fresh water aquifers, causing the land above them to sink downwards eliminating the voids where fresh water was stored, and moving the formerly fresh water into the ocean where it is no longer drinkable, usable for growing crops, or for most industrial purposes.
We do get new fresh water at a reasonable pace thanks to rain - but in many parts of the world we are using it faster than that, and not just depleting the stored volume of fresh water but destroying the storage "containers" themselves.
That's not what a dog whistle is. A dog whistle is when someone isn't doing something, but their ideological opponent wants to imply they are doing that thing, so they accuse them of "dog whistling" the thing. Like if Elon Musk says something that categorically isn't racist but his opponents want to call him racist anyway then they just say he's "dog whistling" to racists.
That's not what "dog whistles" are, lol. Dog Whistle means "coded language" basically.
Dog whistles are where someone says something that their audience will understand to mean a specific thing, but will be inaudible or neutral sounding to people who are not in their audience. They are named that because they are like the whistles only dogs can hear, while most people cannot.
"Inner city" is a canonical example of a dog whistle. Where the literal meaning is the districts in a city in the urban center, but is often used to denote poor minority communities. (If the literal meaning is only "city centers", then would you describe Manhattanites as inner city?)
On the left, "tax the rich" might be a dog whistle that carries a similar literal meaning disjoint from the understood meaning within the community.
> Dog whistles are where someone says something that their audience will understand to mean a specific thing, but will be inaudible or neutral sounding to people who are not in their audience. They are named that because they are like the whistles only dogs can hear, while most people cannot.
That's basically what I said, except you're missing that more often than not it's an intentional stretching of a literal phrase in order to cast aspersions on someone who didn't do the thing you're mad about.
For example, here was one of the top results when I googled "trump dog whistle",
> In February 2018, during Trump’s first term as president, the Department of Homeland Security issued a 14-word press release titled “We Must Secure The Border And Build The Wall To Make America Safe Again.” I and other investigators of far-right extremism attributed this phrase’s use to a clear dog whistle of the common white supremacist saying known as “the 14 words” – “we must secure the existence of our people and a future for white children.”
Or this top result from the search "musk dog whistle",
> Omar Suleiman has called on Elon Musk to stop blowing political "dog whistles of Islamophobia"
> Yet, for the past week, you have blown every conceivable dog whistle of Islamophobia, by highlighting a select group of (horrifying) incidents supposedly in the name of Islam
In this case absolutely no examples were given, but that's the great thing about accusing someone of dog whistling - you don't need to provide any evidence! In fact, literally any evidence you can provide would only serve to weaken your accusation because by definition anyone who isn't whichever -ist you're accusing them of will literally be unable to decode the -ism in their phrasing. If it sounds obviously -ist then by definition it can't be a dog whistle.
Just because you can find a bad article with bad examples, and some are for sure coincidences, but that doesn't mean its not true. Musk did heil, Musk does post well known white supremacy signals. Trump might be a racist and like the fascist power but he is not a white supremacist christian like the rest of his cabinet of project2025 people.
I don't know why, but it always irks me when a corporation puts out a document like this cosplaying as peer-reviewed research, and they don't bother to put in even a lip-service "conflict of interest" disclosure. I should expect it at this point.
I'm not sure why they would report on the median prompt, and not the average, which would give a better sense of (well average) consumption in this case
100%, and I say that as someone who often think average is misleading, but in this case it makes no sense to report median (unless you are working at Google and trying to chase tail usage).
> I'm not sure why they would report on the median
The why is an easier question. They probably picked the lower of the two numbers because it lies in their interest to state they are energy efficient.
>Even before the AI boom, shouldn't we have a bunch of extra capacity under construction, ready for EV driving, induction stoves and heat-pump heating?
When it comes to the EV, the answer is simple: the EV takeover "by 2030" was 100% wishful thinking - the capacity is nowhere near there, starting from scaling the battery production, never mind the charge capacity.
No, mostly misunderstanding. ~95% of all cars sold in Norway are EV, yet only ~25% of the cars on the road are EV's. Most cars predate the EV transition. It'll take another ~20 years until the 95% of the cars on the road are EV's.
We'll have the battery capacity and charge capacity to allow 100% of cars sold in 2030 to be EV's. We only need 2 capacity doublings for batteries, and currently doublings happen every ~18 months. Charge capacity is even easier, we just need to increase electricity production by 1-2% per year for a couple decades to support the transition to EV's.
Existence of “2030 deadline” was/ is significant factor by itself. (Current sate would be less electrified without that arbitrary and over optimistic fantasy deadline)
That seems irrelevant to the question. If I took your prompt, my response would be "do you suggest we should run on thin margins of transmission capacity while we watch an explosion of demand coming at us?"
> When I hear reports that AI power demand is overloading electricity infrastructure, it always makes me think: Even before the AI boom, shouldn't we have a bunch of extra capacity under construction, ready for EV driving, induction stoves and heat-pump heating?
One of the problems might be that a data center puts a lot of demand into a small area and it needs that power soon.
Those other things are being phased in over time, so we only need modest annual capacity growth to deal with them, and they are spread out.
China however is continuously providing double the energy they currently require, only to notice that every two years or so it actually did end up getting used.
>~If they're running on, say, two RTX 6000s for a total draw of ~600 watts, that would be a response time of 1.44 seconds. So obviously the median prompt doesn't go to some high-end thinking model users have to pay for.
You're not accounting for batches for the optimal gpu utilization, maybe it can takes 30 seconds but it completed 30 requests.
Yes we should, which is yet another example of ineffective vision from the prior administration and outright incompetence and sabotage by the current one, regarding renewables and transmission infrastructure.
I was taken aback recently when a Gen-ish Z person told me AI was 'destroying all the water'. I've done data center work, and while I know it is used for cooling, I don't think I've ever personally destroyed any large bodies of water.
There is a perception out there about GenAI and water that goes surprisingly deep. I was told we are will be living in a drought-stricken hellscape, and AI is to blame.
I'd like to know the equivalent energy consumption of a single TikTok video, but that is probably arguing the wrong thing. My bigger question is ... where do they think that water goes? Steam? The assumption is that it is gone forever, and I can't get over how people could just take that at face value.
Water shortages are highly local problems. Sure, running a data center in Arizona might have some genuine water concerns. But even then, it can be mitigated by changes like using wastewater. The Palo Verde plant does that for its heat exchangers.
says Google's datacenter water consumption in 2023 was 5.2 billion gallons, or ~14 million gallons a day. Microsoft was ~4.7, Facebook was 2.6, AWS didn't seem to disclose, Apple was 2.3. These numbers seem pulled from what the companies published.
The total for these companies was ~30 million gallons a day. Apply your best guesses as to what fraction of datacenter usage they are, what fraction of datacenter usage is AI, and what 2025 usage looks like compared to 2023. My guess is it's unlikely to come out to more than 120 million.
I didn't vet this that carefully so take the numbers with a grain of salt, but the rough comparison does seem to hold that Arizona golf courses are larger users of water.
Agricultural numbers are much higher, the California almond industry uses ~4000 million gallons of water a day.
I was also surprised when someone asked me about AI's water consumption because I had never heard of it being an issue. But a cursory search shows that datacenters use quite a bit more water than I realized, on the order of 1 liter of water per kWh of electricity. I see a lot of talk about how the hyperscalers are doing better than this and are trying to get to net-positive, but everything I saw was about quantifying and optimizing this number rather than debunking it as some sort of myth.
I find "1 liter per kWh" to be a bit hard to visualize, but when they talk about building a gigawatt datacenter, that's 278L/s. A typical showerhead is 0.16L/s. The Californian almond industry apparently uses roughly 200kL/s averaged over the entire year -- 278L/s is enough for about 4 square miles of almond orchards.
So it seems like a real thing but maybe not that drastic, especially since I think the hyperscaler numbers are better than this.
Data centers don't just heat up the water and return it - they evaporate the water into the atmosphere (yes, I know, the H2O still exists, but it's in a far less usable form when it's gaseous atmospheric H2O)
Destroying the water originates as a NIMBY talking point (to stop data centers) that was co-opted by anti-capitalist groups (higher membership rates among GenZ) as a boogeyman to blame for why AI is bad.
Data centers do consume a lot of water, but even more power. AI is causing us to forget our climate change carbon goals.
The global increase in drought is probably more a result of climate change temperatures than direct consumption. (Yes AI is doing it but not in the way most people thought)
Stopping AI development in the name of Climate Change so that we lose to China (who pollutes 3x as much as we do) is idiotic as best, self-destructing at worst.
Since pulling out of the Paris Climate Accords, defunding climate research and doctoring past data, it's no longer possible to take the moral high ground here.
It's probably best, for your mental health, to ask these questions in earnest, and stop dismissing people as illogical. The communities living near data centers have real water problems, making it believable. If you're wondering why NIMBY happens, just watch the first 30 seconds.
And there isn't solid evidence that this was connected to the data center construction:
> Ben Sheidler, a spokesman for the Joint Development Authority, which manages the industrial park that Meta’s facilities occupy, said the cause of the water issues was unknown. The Joint Development Authority did not do a well water study before construction to determine any potential effects, but the timing of the problems could be a coincidence, he said.
> “I wouldn’t want to speculate that even the construction had something to do with it,” he said. “One thousand feet away is a pretty significant distance.”
In 2011, Google claimed that each search query takes about 0.3Wh [1]. Earlier this year, Sam Altman also claimed about 0.3Wh avg use per query for OpenAI.
I'm honestly surprised that they're so similar. I've thought of LLM queries as being far more energy-intense than "just" a Google search, but maybe the takeaway is that ordinary Google searching is also quite energy-intense.
If I as a user just wanted an answer to a dumb question like, say, the meaning of some genZ slang, it seems about an order of magnitude to ask a small LLM running on my phone than to make a google search.
(Check my math: assuming the A16 CPU draws 5 watts peak for 20sec running Gemma or whatever on my iPhone, that’s 0.03Wh to answer a simple query, which is 10x cheaper)
Are training costs (esp. from failed runs) amortized in these estimates?
Around 2008 a core step in search was basically a grep over all documents. The grep was distributed over roughly 1000 machines so that the documents could be held in memory rather than on disk.
Inverted indices were not used as they worked poorly for “an ordered list of words” (as opposed to a bag of words).
And this doesn’t even start to address the ranking part.
It seems highly unlikely that they did not use indices. Scanning all documents would be prohibitively slow. I think it is more likely that the indices were really large, and it would take hundreds to thousands of machines to store the indices in RAM. Having a parallel scan through those indices seems likely.
Wikipedia [1] links to "Jeff Dean's keynote at WSDM 2009" [2] which suggests that indices were most certainly used.
Then again, I am no expert in this field, so if you could share more details, I'd love to hear more about it.
.3 Wh is 1080 joules. A liter of gasoline contains over 30 million joules. So this is like .034 milliliters of gasoline. But with grid power so even less than that since gasoline is very inefficient.
They just be doing something crazy because any time I query my local llm the lights in my office dim and the temperature rises a few degrees. Definitely far more energy than running the microwave for 1 second.
Google was not using deep learning for search in 2011. Deep learning as a concept didn't really take off until AlexNet in 2012 anyway.
Various ML "learn-to-rank" tooling was in use at Google for a while, but incorporating document embedding vectors w/ ANN search into the ranking function probably happened over the course of 2018-2021 [1], I think. Generative AI only started appearing in ordinary search results in 2024.
With google serving AI overviews, now an average search query should cost more? Compute is getting cheaper but also algorithms getting more and more complex, increasing compute?
> the takeaway is that ordinary Google searching is also quite energy-intense.
A related takeaway should be that machine inference is pervasive and has been for years, and that defining "AI" to mean just chatbots is to ignore most of the iceberg.
> I'd still love to see a report that accurately captures training cost. Today's report[1] notably excludes training cost.
From 2022, so possibly out of date: "ML training and inference are only 10%–15% of Google’s total energy use for each of the last three years, each year split ⅗ for inference and ⅖ for training." That's probably close enough to estimate 50/50, or the full energy cost to deliver an AI result is double the inference energy.
My gosh you're right! The paper in question is https://arxiv.org/pdf/2204.05149, "The Carbon Footprint of Machine Learning Training Will Plateau, Then Shrink"
One thing I'm missing in the full report is what a 'median prompt' actually looks like. How many tokens? What's the distribution of prompt sizes like? Is it even the same 'median prompt' between 2024 and 2025?
The numbers are cute but we can't actually do anything with them without those details. At least an average could be multiplied by the # of queries to get the total usage.
In his last blog post, Sam Altman also revealed how much power the average chatgpt query uses, and it's in the same ballpark.
> People are often curious about how much energy a ChatGPT query uses; the average query uses about 0.34 watt-hours, about what an oven would use in a little over one second, or a high-efficiency lightbulb would use in a couple of minutes. It also uses about 0.000085 gallons of water; roughly one fifteenth of a teaspoon.
What matters most is training, post-training, fine-tuning and data scraping, not inference IMO. The "prompting destroys the environment" deal always seemed sensationalistic to me, I'm glad it's starting to be debunked, but at the same time we can see IRL the heavy toll the new data centers being built are taking on the energy grid. If the tech was lean at it's heart in terms of energy consumption, we wouldn't be seeing that or even the private nuclear reactor arms race by big tech.
To set appropriate goal posts, it shouldn't just the total water/energy usage from start to finish: it's that, relative to the local regions they're being extracted from. Are those local regions at risk of being overused?
Focusing on water, Google does reports on their total water withdrawal at all their data centers. It was around 10 billion gallons per year in 2024. That's around 0.03% of water usage in the US (that's just for rough context - not all of Google's data centers are in the US). I don't think that's an eye-popping amount.
But when you consider that more than 1 billion gallons came from Council Bluffs, IA, you have to make sure that's being done responsibly, and that the local ecology can support it. Google admits that 28% of their water use comes from sources with "medium to high risk of depletion or scarcity." [0]
Yupp, plus the fact that I unwillingly send prompts out by visiting a website that now has a chatbot nag me;
Among other annoyances in the economy of scale that say "if it is cheap, just use more of it".
This is a good point. If we are marching towards some sort of LLM-driven world, where every company wants to have automated AI agents for customer service, internal development, marketing and prospecting, the energy usage will compound. A simple RAG system can be expected to make 5 or more prompts per query, and multi-agent systems like Roo Code power through tens of prompts for any given task.
> If the tech was lean at it's heart in terms of energy consumption, we wouldn't be seeing that or even the private nuclear reactor arms race by big tech.
Eh, I somewhat disagree with this. The US energy grid has had almost no extra capacity for a long time. A lot of this was due to efficiency (not a bad thing) and little industrial growth (not a great thing) in the country. Data centers themselves, I don't think are the biggest cause of the issues, but the distribution grid. We've had tons of problems around distribution with new energy sources coming online, but issues distributing the power to where we need to.
We see things like private power plants, not because we can't generate power, but because we absolutely suck at distribution of power.
If I'm doing the math right, you can get ~4000 queries per kWh. A quick Google search gave $0.04 per kWh when bought in bulk. So you can get around 100k queries per dollar of electricity.
That's.... a lot cheaper than I would have guessed. Obviously, the data centers cost quite a bit to build. But when you think of $20/mo for a typical subscription. That's not bad?
The fundamental flaw in AI energy/water doomerism has always been that energy costs money, water costs money, real estate costs money, but AI is being given away for free. There is obviously something wrong with the suggestion that AI is using all the energy and water.
It would be interesting to know how many queries they get per "base" model. I honestly have no idea what the scale is. In any case, that's the only way to know how to amortize training cost over inference.
> One major question mark is the total number of queries that Gemini gets each day, which would allow estimates of the AI tool’s total energy demand.
Yeah, I was more interested in knowing the total amount. A "median" prompt without the information on the total number of prompts is kind of meaningless...
The average is more useful if you wish to answer the question of how much energy the entire Google is using, but there are other ways to answer that question (such as Google's annual report, which tells you to 8 significant digits how much energy they used).
The median is more useful to answer the question of how much energy a typical Gemini interaction uses.
That's very interesting, although I'm still curious about the training resource usage -- not "only" inference. I wonder what is the relative importance of training (i.e., what percentage of the resource usage was in training vs. inference)
One random preprint I found (https://arxiv.org/html/2505.09598v2) estimates inference is 90% of the total energy usage. From some googling around, everything seems to agree that inference dominates (or is at least comparable to) training on the large commonly-used models.
I was less surprised that inference dominates training after I read that chatgpt is serving billions of requests per day.
Many papers on large models make claims about the energy used. Google has consistently said that training and inference use about the same fraction of their data center resources.
Is this just meant to be dismissive, or just like a kind of non-answer? That it could be said for anything doesn't make asking it of this specific thing unimportant or uninteresting.
Right, and the scraping, the extra storage required, the manufacture of all the extra GPUs, etc. This is them hoping people don't understand that the query is only one part of it and aren't curious enough to ask further questions.
Maybe, but I would suggest that larger factors arise from utilization and overhead. These are going to be the competitive advantages that AI operators need to win the arms race.
"A Comparison of the Cerebras Wafer-Scale Integration Technology with Nvidia GPU-based Systems for Artificial Intelligence" (2025)
https://arxiv.org/html/2503.11698v1
> We will design and operate our datacenters to support society’s climate goals and become carbon negative, water positive and zero waste before 2030. [...]
> By 2025, we will procure 100% renewable energy on a global scale, both significantly expanding and decarbonizing local electricity grids.
> Our datacenter designs are more water efficient than traditional enterprise datacenters, and our plan by 2030 is to replenish more water than we consume locally.
I asked GPT 5 to extrapolate this to a Max Claude code monthly usage, two sessions a day, just business days. It guessed that would be 21k prompts a month. Google’s proxy numbers give a monthly footprint of ~5.1 kWh, ~5.53 L water, and ~0.64 kg CO₂e for heavy CC use.
That's equivalent to doing less than two miles driving(CO2), one toilet flush (water) and about three dryer loads of laundry.
I think overall this seems good, we knew this is where the gains from TPUs would lie and there is a nice agreement here between the need for profit and more pressing planetary (anthroposcenic?) concerns. With things like this, one might ask how much do we consume binge watching something on a streaming service, or loading a page full of js and ads and canvas elements, or even just scrolling a feed for an hour on an app.
Still, there are a lot unswered questions here, and its up in the air precisely how this stuff will continue to integrate into services we already use, or habits we have yet to form at large. What does that scaling look like?
But by far the most troubling thing is the continued combination of flippancy, defensiveness, or silence we get from the AI peiple about even attempting to talk about this. If you are a True Believer, don't you want this to be something that is tackled head on, rather than tucked away? When this has come up before, I always end up seeing a bunch of guys who essentially leave the vibe of "well I am plenty above sea level, my AC is pumping just fine, and I just simply don't care because my productivity has doubled!"
Like isn't this community supposed to be excited about a future, eager to tackle problems? Or is there maybe some intrinsic solipsism to the impressive chatbots that ultimately renders this kind of attitude to its users? It feels like right when we were culturally about to age out of this particular form of obstinacy, we set ourselves up to create a whole new generation of "global warming is fake news" people. Which is a shame. If you're going to be like this, just go all in on accelerationism in all its pseudo-facist darkness, don't just borrow a script from baby boomers!
[ETA]
Extending on these numbers a bit, a mean human uses 1.25KW of power (Kardashev Level .7 / 8 Gigahumans) and the mean American uses ~8KW of power according to https://en.wikipedia.org/wiki/List_of_countries_by_energy_co.... So if we align AIs to be eco-friendly, they will definitely murder all humans for the sake of the planet /s
> I’ve yet to use an LLM that can consistently provide sources that back up what it’s telling me.
Mmmh, that would have been my take as well up to around end of Q1 2025.
Theses days, the flagship LLM's have reduced hallucination by quite a bit, and are also way better at citing sources (you sometimes have to nudge them).
ChatGPT 5 has been very decent on that particular axis.
In my chats with 2.5 Flash it gives me the direct sources lol. Also not going to lie, I've found 2.5 Flash generally gets straight to the point with sources a lot quicker than Pro. To the point I don't really use Pro.
Edit - just used Pro, gave me a direct source. Who knows...
You're right. It's clear I should have had an LLM write my comment rather than do it myself before a cup of coffee. I've already spent an order of magnitude more energy thinking about this article compared to an LLM.
Also, since I live a first-world life style which consumes multiple KW of power, I've probably consumed multiple orders of magnitude energy more than an LLM on this topic.
There is a non-trivial chance that the LLM would've added a link to _something_, but links/references seem like a very common thing to hallucinate, no?
The way around that is that is for LLM-based tools to run a regular search engine query in the background and feed the results of that in alongside the prompt. (Usually a two-step process of the LLM formulating the query, then another pass on the results)
The used results can then have their link either added to the end result separately, guaranteeing it is correct, or added to the prompt and "telling the LLM to include it", which retains a risk of hallucination, yes.
Common to both of these is the failure mode that the LLM can still hallucinate whilst "summarizing" the results, meaning you still have no guarantee that the claims made actually show up in the results.
> The way around that is that is for LLM-based tools to run a regular search engine query in the background and feed the results of that in alongside the prompt. (Usually a two-step process of the LLM formulating the query, then another pass on the results)
Would the LLM-based tool be able to determine that the top results are just SEO-spam sites and move lower in the list, or just accept the spam results as gospel?
The practical and readily-observable-from-output answer is "No, they cannot meaningfully identify spam or misinformation, and do indeed just accept the results as gospel"; Google's AI summary works this way and is repeatedly wrong in exactly this way. Google's repeatedly had it be wrong even in the adcopy.
The theoretical mechanism is that the attention mechanism with LLMs would be able to select which parts of the results are fed further into the results. This is how the model is capable of finding parts of the text that are "relevant". The problem is that this just isn't enough to robustly identify spam or incorrect information.
However, we can isolate this "find the relevant bit" functionality away from the rest of the LLM to enhance regular search engines. It's hard to say how useful this is; Google has intentionally damaged their search engine and it may simply not be worth the GPU cycles compared to traditional approaches, but it's an idea being widely explored right now.
>The way around that is that is for LLM-based tools to run a regular search engine query in the background and feed the results of that in alongside the prompt.
Hardly better, as soon those "search engine results" would be AI slop themselves, including actual published papers (phoned-in by using AI, and "peer reviewed" by using AI from indifferent reviewers)
Happens frequently still with GTP 4o (and now 5) and Claude. Makes up "biographies" unrelated to the actual people, gives me bullshit non-existing API options and cli flags, etc.
And then on the rare occasion they do link to a chat, their prompt is something like:
"Tell about a person of history who was important for their work inthe time of their existence and give quotes of what they said that made them important when they were speaking include notes and other stories about them and give details about their life who they married and their kids and who their parents were and other big things they did do in their lives"
It feels so common actually that I would not even bother sharing them. ChatGPT gives me hallucinated URLs all... the... time. Like several times a day. You can fix it by prompting it to Google the subject or research it otherwise it just vomits garbage.
LLMs are not databases, they are text processors that need to be trained on ungodly amounts of data and can sort of kind of double as a database, though a very fallible one. Inject ground truth and you're cooking; don't and it's a crap shoot. (not saying anything new here, but it bears repeating).
The issue is that now we're over-utilizing prompts everywhere. Every search engine query. Every keypress in an AI editor. Every new website makes queries in the back. Analyze emails. etc
So it's not just about "the one query you ask ChatGPT about what you should write your mum to say you're not coming for Thanksgiving"
It's rather that an AI query is 0.24Wh, but that we are now using thousands of those per users per days, and that we globalize it at the scale of the planet, so 7 billion users... and this becomes huge
Press wire to blogspam as a service! There goes 90% of the "content" on news sites that seem to pride themselves on stretching out a sentence of
info into 4 paragraphs.
I love good journalism because it's adhd crack; in-depth spilling the tea but you have to really dig deep to find it nowadays.
Attempts to reduce human experience down to a single number always seem so wrong-headed. A common one is when articles say "billions of dollars of productivity wasted to X". As if you can just replace any human activity for any human activity at any time and get the same results. I know companies would like to imagine that every employee is a fungible cog, but it seems pretty dismal as a world view.
Unless you're arguing that people should kill themselves to save their 100W of power, the future energy use of every alive human is already committed and doesn't matter.
Not to mention that the energy should also include all the extra energy spent on making converting energy into a form that is usable by humans (ie. food). There is probably at least an order of magnitude.
>In total, the median prompt—one that falls in the middle of the range of energy demand—consumes 0.24 watt-hours of electricity, the equivalent of running a standard microwave for about one second
>The report also finds that the total energy used to field a Gemini query has fallen dramatically over time. The median Gemini prompt used 33 times more energy in May 2024 than it did in May 2025, according to Google.
The original press release and report are at [1], couldn't find a link to them in the article.
> In total, the median prompt—one that falls in the middle of the range of energy demand—consumes 0.24 watt-hours of electricity
If they're running on, say, two RTX 6000s for a total draw of ~600 watts, that would be a response time of 1.44 seconds. So obviously the median prompt doesn't go to some high-end thinking model users have to pay for.
It's a very low number; for comparison, an electric vehicle might consume 82kWh to travel 363 miles. So that 0.24 watt-hours of energy is equivalent to driving 5.6 feet (1.7 meters) in such an EV.
When I hear reports that AI power demand is overloading electricity infrastructure, it always makes me think: Even before the AI boom, shouldn't we have a bunch of extra capacity under construction, ready for EV driving, induction stoves and heat-pump heating?
[1] https://cloud.google.com/blog/products/infrastructure/measur...
> When I hear reports that AI power demand is overloading electricity infrastructure ...
It feels like dog-whistle tactics. "Aren't the technology companies bad for the environment!" "What about the water usage?" "What about the electricity?"
For me the peak of this is complaining about water consumption at the Dalles datacentre [0]. The buildings are next to the Colombia river and a few miles away from the Dalles Dam [1] which generates an average of 700MW. The river water should be used for cooling, taking out some of the water, warming it up by a few degrees and returning it to the river; one might argue that this is simply returning the heat to the river that would have come from the water flowing downhill.
[0] https://www.oregonlive.com/silicon-forest/2022/12/googles-wa...
[1] https://en.wikipedia.org/wiki/The_Dalles_Dam
What's the dog whistle? People are concerned about the impact industry has on the environment, and they are stating those concerns plainly. I don't think the non-profit WaterWatch's true goal is to destroy big tech.
I think you're oversimplifying the "just use rivers" idea. Most data centers (80% for Google) require potable water for cooling, and it can't come straight from a river. Plus, using potable water in cooling adds mineral deposits to the water and will require treatment to be consumable again.
The term "water consumption" makes people think the water is destroyed, which is probably a large part of the controversy.
Because in most cases it kind of is. It's not that the H2O molecules are forcefully disintegrated, but most data centers use evaporative cooling, meaning that whatever water is fed to the datacenter through the municipal water system ends up as moisture in the atmosphere. This is in effect equivalent to a sizable leak in the water infrastructure.
Doesn't it come back down as rain?
Yes, but you can neither drink rainwater nor use it to cool a data center. And the rain may end up falling thousands of miles away. Excessive use of water reduces flow in natural bodies of water and can mess up local ecosystems.
Not sure I'd drink most river water either, and I would hope most data centers don't pull water straight from the aquifer (though maybe they do). Fair points though.
Uhhh I have been living on rainwater directly for several years, and like every other person on the planet indirectly for my entire life.
Somewhere, possibly not in the same watershed. It depends.
> This is in effect equivalent to a sizable leak in the water infrastructure.
Where do you think the evaporated water goes?
Surely all uses of water are part of the closed-loop water cycle? Other than launching it into space for astronauts to drink, and using it in permanent reactions like concrete hydration?
Drinking water, spraying it on crops, using it to clean a car, or using it to flush a toilet all end up with the water evaporating, or making its way to the ocean and evaporating from there.
Ultimately, if a river provides a certain number of acre-feet of fresh water, evaporating it to cool a data centre uses it just as much as to evaporating it to grow alfalfa in a desert, except perhaps more usefully.
Fresh water isn't meaningfully a closed loop. We are draining fresh water aquifers, causing the land above them to sink downwards eliminating the voids where fresh water was stored, and moving the formerly fresh water into the ocean where it is no longer drinkable, usable for growing crops, or for most industrial purposes.
We do get new fresh water at a reasonable pace thanks to rain - but in many parts of the world we are using it faster than that, and not just depleting the stored volume of fresh water but destroying the storage "containers" themselves.
That's not what a dog whistle is. A dog whistle is when someone isn't doing something, but their ideological opponent wants to imply they are doing that thing, so they accuse them of "dog whistling" the thing. Like if Elon Musk says something that categorically isn't racist but his opponents want to call him racist anyway then they just say he's "dog whistling" to racists.
That's not what "dog whistles" are, lol. Dog Whistle means "coded language" basically.
Dog whistles are where someone says something that their audience will understand to mean a specific thing, but will be inaudible or neutral sounding to people who are not in their audience. They are named that because they are like the whistles only dogs can hear, while most people cannot.
"Inner city" is a canonical example of a dog whistle. Where the literal meaning is the districts in a city in the urban center, but is often used to denote poor minority communities. (If the literal meaning is only "city centers", then would you describe Manhattanites as inner city?)
On the left, "tax the rich" might be a dog whistle that carries a similar literal meaning disjoint from the understood meaning within the community.
> Dog whistles are where someone says something that their audience will understand to mean a specific thing, but will be inaudible or neutral sounding to people who are not in their audience. They are named that because they are like the whistles only dogs can hear, while most people cannot.
That's basically what I said, except you're missing that more often than not it's an intentional stretching of a literal phrase in order to cast aspersions on someone who didn't do the thing you're mad about.
For example, here was one of the top results when I googled "trump dog whistle",
> In February 2018, during Trump’s first term as president, the Department of Homeland Security issued a 14-word press release titled “We Must Secure The Border And Build The Wall To Make America Safe Again.” I and other investigators of far-right extremism attributed this phrase’s use to a clear dog whistle of the common white supremacist saying known as “the 14 words” – “we must secure the existence of our people and a future for white children.”
https://theconversation.com/musks-inauguration-salute-is-not...
Or this top result from the search "musk dog whistle",
> Omar Suleiman has called on Elon Musk to stop blowing political "dog whistles of Islamophobia"
> Yet, for the past week, you have blown every conceivable dog whistle of Islamophobia, by highlighting a select group of (horrifying) incidents supposedly in the name of Islam
In this case absolutely no examples were given, but that's the great thing about accusing someone of dog whistling - you don't need to provide any evidence! In fact, literally any evidence you can provide would only serve to weaken your accusation because by definition anyone who isn't whichever -ist you're accusing them of will literally be unable to decode the -ism in their phrasing. If it sounds obviously -ist then by definition it can't be a dog whistle.
https://x.com/elonmusk/status/1890842231180816419
Just because you can find a bad article with bad examples, and some are for sure coincidences, but that doesn't mean its not true. Musk did heil, Musk does post well known white supremacy signals. Trump might be a racist and like the fascist power but he is not a white supremacist christian like the rest of his cabinet of project2025 people.
That's just a press release. Here's the actual PDF of the actual tech report: https://services.google.com/fh/files/misc/measuring_the_envi...
I don't know why, but it always irks me when a corporation puts out a document like this cosplaying as peer-reviewed research, and they don't bother to put in even a lip-service "conflict of interest" disclosure. I should expect it at this point.
I'm not sure why they would report on the median prompt, and not the average, which would give a better sense of (well average) consumption in this case
100%, and I say that as someone who often think average is misleading, but in this case it makes no sense to report median (unless you are working at Google and trying to chase tail usage).
> I'm not sure why they would report on the median
The why is an easier question. They probably picked the lower of the two numbers because it lies in their interest to state they are energy efficient.
Median is _an_ average, are you wanting a mean consumption perhaps when you say average?
>Even before the AI boom, shouldn't we have a bunch of extra capacity under construction, ready for EV driving, induction stoves and heat-pump heating?
When it comes to the EV, the answer is simple: the EV takeover "by 2030" was 100% wishful thinking - the capacity is nowhere near there, starting from scaling the battery production, never mind the charge capacity.
No, mostly misunderstanding. ~95% of all cars sold in Norway are EV, yet only ~25% of the cars on the road are EV's. Most cars predate the EV transition. It'll take another ~20 years until the 95% of the cars on the road are EV's.
We'll have the battery capacity and charge capacity to allow 100% of cars sold in 2030 to be EV's. We only need 2 capacity doublings for batteries, and currently doublings happen every ~18 months. Charge capacity is even easier, we just need to increase electricity production by 1-2% per year for a couple decades to support the transition to EV's.
While it was never going happen.
Existence of “2030 deadline” was/ is significant factor by itself. (Current sate would be less electrified without that arbitrary and over optimistic fantasy deadline)
That seems irrelevant to the question. If I took your prompt, my response would be "do you suggest we should run on thin margins of transmission capacity while we watch an explosion of demand coming at us?"
> When I hear reports that AI power demand is overloading electricity infrastructure, it always makes me think: Even before the AI boom, shouldn't we have a bunch of extra capacity under construction, ready for EV driving, induction stoves and heat-pump heating?
One of the problems might be that a data center puts a lot of demand into a small area and it needs that power soon.
Those other things are being phased in over time, so we only need modest annual capacity growth to deal with them, and they are spread out.
Should we? Yes. Do we? Nope.
China however is continuously providing double the energy they currently require, only to notice that every two years or so it actually did end up getting used.
>~If they're running on, say, two RTX 6000s for a total draw of ~600 watts, that would be a response time of 1.44 seconds. So obviously the median prompt doesn't go to some high-end thinking model users have to pay for.
You're not accounting for batches for the optimal gpu utilization, maybe it can takes 30 seconds but it completed 30 requests.
Yes we should, which is yet another example of ineffective vision from the prior administration and outright incompetence and sabotage by the current one, regarding renewables and transmission infrastructure.
I was taken aback recently when a Gen-ish Z person told me AI was 'destroying all the water'. I've done data center work, and while I know it is used for cooling, I don't think I've ever personally destroyed any large bodies of water.
There is a perception out there about GenAI and water that goes surprisingly deep. I was told we are will be living in a drought-stricken hellscape, and AI is to blame.
I'd like to know the equivalent energy consumption of a single TikTok video, but that is probably arguing the wrong thing. My bigger question is ... where do they think that water goes? Steam? The assumption is that it is gone forever, and I can't get over how people could just take that at face value.
Water shortages are highly local problems. Sure, running a data center in Arizona might have some genuine water concerns. But even then, it can be mitigated by changes like using wastewater. The Palo Verde plant does that for its heat exchangers.
The existing golf courses in Arizona alone use more water than the entire global data center industry.
I’d love to believe this but maybe a citation would help.
https://azallianceforgolf.org/wp-content/uploads/2023/01/C-S...
page 21, says Arizona 2015 golf course irrigation was 120 million gallons per day, citing the US Geological Survey.
https://dgtlinfra.com/data-center-water-usage/
says Google's datacenter water consumption in 2023 was 5.2 billion gallons, or ~14 million gallons a day. Microsoft was ~4.7, Facebook was 2.6, AWS didn't seem to disclose, Apple was 2.3. These numbers seem pulled from what the companies published.
The total for these companies was ~30 million gallons a day. Apply your best guesses as to what fraction of datacenter usage they are, what fraction of datacenter usage is AI, and what 2025 usage looks like compared to 2023. My guess is it's unlikely to come out to more than 120 million.
I didn't vet this that carefully so take the numbers with a grain of salt, but the rough comparison does seem to hold that Arizona golf courses are larger users of water.
Agricultural numbers are much higher, the California almond industry uses ~4000 million gallons of water a day.
I was also surprised when someone asked me about AI's water consumption because I had never heard of it being an issue. But a cursory search shows that datacenters use quite a bit more water than I realized, on the order of 1 liter of water per kWh of electricity. I see a lot of talk about how the hyperscalers are doing better than this and are trying to get to net-positive, but everything I saw was about quantifying and optimizing this number rather than debunking it as some sort of myth.
I find "1 liter per kWh" to be a bit hard to visualize, but when they talk about building a gigawatt datacenter, that's 278L/s. A typical showerhead is 0.16L/s. The Californian almond industry apparently uses roughly 200kL/s averaged over the entire year -- 278L/s is enough for about 4 square miles of almond orchards.
So it seems like a real thing but maybe not that drastic, especially since I think the hyperscaler numbers are better than this.
Data centers use evaporative cooling.
Data centers don't just heat up the water and return it - they evaporate the water into the atmosphere (yes, I know, the H2O still exists, but it's in a far less usable form when it's gaseous atmospheric H2O)
The Meta FTW data centers use evaporative chillers, but they re-condense the water so that it's a closed loop.
The implicit claim that data centers don't recondense the water they evaporate is surprising to me.
Do you have a source?
> Do you have a source?
Just what I've been told by people in the industry. I too would love to see more solid data.
Destroying the water originates as a NIMBY talking point (to stop data centers) that was co-opted by anti-capitalist groups (higher membership rates among GenZ) as a boogeyman to blame for why AI is bad.
Data centers do consume a lot of water, but even more power. AI is causing us to forget our climate change carbon goals. The global increase in drought is probably more a result of climate change temperatures than direct consumption. (Yes AI is doing it but not in the way most people thought)
Stopping AI development in the name of Climate Change so that we lose to China (who pollutes 3x as much as we do) is idiotic as best, self-destructing at worst.
Since pulling out of the Paris Climate Accords, defunding climate research and doctoring past data, it's no longer possible to take the moral high ground here.
[dead]
It's probably best (for your mental health) to not ask these questions in earnest.
It's probably best, for your mental health, to ask these questions in earnest, and stop dismissing people as illogical. The communities living near data centers have real water problems, making it believable. If you're wondering why NIMBY happens, just watch the first 30 seconds.
https://www.youtube.com/watch?v=DGjj7wDYaiI
The NYT article says that the couple suspects that sediment buildup is blocking pipes, not that there is a water shortage causing taps to run dry: https://www.nytimes.com/2025/07/14/technology/meta-data-cent...
And there isn't solid evidence that this was connected to the data center construction:
> Ben Sheidler, a spokesman for the Joint Development Authority, which manages the industrial park that Meta’s facilities occupy, said the cause of the water issues was unknown. The Joint Development Authority did not do a well water study before construction to determine any potential effects, but the timing of the problems could be a coincidence, he said.
> “I wouldn’t want to speculate that even the construction had something to do with it,” he said. “One thousand feet away is a pretty significant distance.”
There isn't solid evidence that it isn't the data center construction. I do know that correlation + obvious dirt upheaval = likely chance.
So now we're asking data center builders to prove a negative?
In 2011, Google claimed that each search query takes about 0.3Wh [1]. Earlier this year, Sam Altman also claimed about 0.3Wh avg use per query for OpenAI.
I'm honestly surprised that they're so similar. I've thought of LLM queries as being far more energy-intense than "just" a Google search, but maybe the takeaway is that ordinary Google searching is also quite energy-intense.
If I as a user just wanted an answer to a dumb question like, say, the meaning of some genZ slang, it seems about an order of magnitude to ask a small LLM running on my phone than to make a google search.
(Check my math: assuming the A16 CPU draws 5 watts peak for 20sec running Gemma or whatever on my iPhone, that’s 0.03Wh to answer a simple query, which is 10x cheaper)
Are training costs (esp. from failed runs) amortized in these estimates?
1: https://googleblog.blogspot.com/2009/01/powering-google-sear...
14 years of progress on energy efficiency might also have an impact here...
10-ish 18 month doublings would be around 1000x so it explains a lot.
Around 2008 a core step in search was basically a grep over all documents. The grep was distributed over roughly 1000 machines so that the documents could be held in memory rather than on disk.
Inverted indices were not used as they worked poorly for “an ordered list of words” (as opposed to a bag of words).
And this doesn’t even start to address the ranking part.
It seems highly unlikely that they did not use indices. Scanning all documents would be prohibitively slow. I think it is more likely that the indices were really large, and it would take hundreds to thousands of machines to store the indices in RAM. Having a parallel scan through those indices seems likely.
Wikipedia [1] links to "Jeff Dean's keynote at WSDM 2009" [2] which suggests that indices were most certainly used.
Then again, I am no expert in this field, so if you could share more details, I'd love to hear more about it.
[1] https://en.wikipedia.org/wiki/Google_data_centers
[2] https://static.googleusercontent.com/media/research.google.c...
.3 Wh is 1080 joules. A liter of gasoline contains over 30 million joules. So this is like .034 milliliters of gasoline. But with grid power so even less than that since gasoline is very inefficient.
They just be doing something crazy because any time I query my local llm the lights in my office dim and the temperature rises a few degrees. Definitely far more energy than running the microwave for 1 second.
What hardware runs your local LLM?
A RTX 5080, not the lowest power card.
Not the lowest power, but surely it uses less power than a microwave. 360W TDP, 850W required system power, while my microwave is 1000W.
The difference is most prompts take far longer than a single second to return.
Yeah, but couple hundred watts certainly shouldn't dim your lights.
At that point was Google already using deep learning for search? I'd guess the number fluctuated a bit during the rollout of this kind of feature
Google was not using deep learning for search in 2011. Deep learning as a concept didn't really take off until AlexNet in 2012 anyway.
Various ML "learn-to-rank" tooling was in use at Google for a while, but incorporating document embedding vectors w/ ANN search into the ranking function probably happened over the course of 2018-2021 [1], I think. Generative AI only started appearing in ordinary search results in 2024.
1: https://cloud.google.com/blog/topics/developers-practitioner...
With google serving AI overviews, now an average search query should cost more? Compute is getting cheaper but also algorithms getting more and more complex, increasing compute?
> the takeaway is that ordinary Google searching is also quite energy-intense.
A related takeaway should be that machine inference is pervasive and has been for years, and that defining "AI" to mean just chatbots is to ignore most of the iceberg.
I'd still love to see a report that accurately captures training cost. Today's report[1] notably excludes training cost.
Not just "one training run," but the cost of a thousand AI engineers starting failing runs to get to that one deployed model.
1: Link to Google's tech report: https://services.google.com/fh/files/misc/measuring_the_envi... "We leave the measurement of AI model training to future work."
> I'd still love to see a report that accurately captures training cost. Today's report[1] notably excludes training cost.
From 2022, so possibly out of date: "ML training and inference are only 10%–15% of Google’s total energy use for each of the last three years, each year split ⅗ for inference and ⅖ for training." That's probably close enough to estimate 50/50, or the full energy cost to deliver an AI result is double the inference energy.
https://research.google/blog/good-news-about-the-carbon-foot...
It still kills me, every time, that the title embedded in the metadata of that original PDF is "Revamped Happy CO2e Paper".
My gosh you're right! The paper in question is https://arxiv.org/pdf/2204.05149, "The Carbon Footprint of Machine Learning Training Will Plateau, Then Shrink"
One thing I'm missing in the full report is what a 'median prompt' actually looks like. How many tokens? What's the distribution of prompt sizes like? Is it even the same 'median prompt' between 2024 and 2025?
The numbers are cute but we can't actually do anything with them without those details. At least an average could be multiplied by the # of queries to get the total usage.
In his last blog post, Sam Altman also revealed how much power the average chatgpt query uses, and it's in the same ballpark.
> People are often curious about how much energy a ChatGPT query uses; the average query uses about 0.34 watt-hours, about what an oven would use in a little over one second, or a high-efficiency lightbulb would use in a couple of minutes. It also uses about 0.000085 gallons of water; roughly one fifteenth of a teaspoon.
https://blog.samaltman.com/the-gentle-singularity
What matters most is training, post-training, fine-tuning and data scraping, not inference IMO. The "prompting destroys the environment" deal always seemed sensationalistic to me, I'm glad it's starting to be debunked, but at the same time we can see IRL the heavy toll the new data centers being built are taking on the energy grid. If the tech was lean at it's heart in terms of energy consumption, we wouldn't be seeing that or even the private nuclear reactor arms race by big tech.
To set appropriate goal posts, it shouldn't just the total water/energy usage from start to finish: it's that, relative to the local regions they're being extracted from. Are those local regions at risk of being overused?
Focusing on water, Google does reports on their total water withdrawal at all their data centers. It was around 10 billion gallons per year in 2024. That's around 0.03% of water usage in the US (that's just for rough context - not all of Google's data centers are in the US). I don't think that's an eye-popping amount.
But when you consider that more than 1 billion gallons came from Council Bluffs, IA, you have to make sure that's being done responsibly, and that the local ecology can support it. Google admits that 28% of their water use comes from sources with "medium to high risk of depletion or scarcity." [0]
[0]: https://www.gstatic.com/gumdrop/sustainability/google-2025-e...
Yupp, plus the fact that I unwillingly send prompts out by visiting a website that now has a chatbot nag me; Among other annoyances in the economy of scale that say "if it is cheap, just use more of it".
This is a good point. If we are marching towards some sort of LLM-driven world, where every company wants to have automated AI agents for customer service, internal development, marketing and prospecting, the energy usage will compound. A simple RAG system can be expected to make 5 or more prompts per query, and multi-agent systems like Roo Code power through tens of prompts for any given task.
> If the tech was lean at it's heart in terms of energy consumption, we wouldn't be seeing that or even the private nuclear reactor arms race by big tech.
Eh, I somewhat disagree with this. The US energy grid has had almost no extra capacity for a long time. A lot of this was due to efficiency (not a bad thing) and little industrial growth (not a great thing) in the country. Data centers themselves, I don't think are the biggest cause of the issues, but the distribution grid. We've had tons of problems around distribution with new energy sources coming online, but issues distributing the power to where we need to.
We see things like private power plants, not because we can't generate power, but because we absolutely suck at distribution of power.
- one second of running a toaster, or
- 1/80th of a phone charge,
- lifting 100 pounds to a height of 6 feet,
- muzzle energy of a 9mm bullet,
- driving 6 feet with a Tesla.
That's surprisingly energy efficient, for a technology that seemed like magic back in 2022.
If I'm doing the math right, you can get ~4000 queries per kWh. A quick Google search gave $0.04 per kWh when bought in bulk. So you can get around 100k queries per dollar of electricity.
That's.... a lot cheaper than I would have guessed. Obviously, the data centers cost quite a bit to build. But when you think of $20/mo for a typical subscription. That's not bad?
Did I do that right?
Yes, you did.
The fundamental flaw in AI energy/water doomerism has always been that energy costs money, water costs money, real estate costs money, but AI is being given away for free. There is obviously something wrong with the suggestion that AI is using all the energy and water.
But what about the training cost?
It would be interesting to know how many queries they get per "base" model. I honestly have no idea what the scale is. In any case, that's the only way to know how to amortize training cost over inference.
If you want to skip the third party article:
Blog post: https://cloud.google.com/blog/products/infrastructure/measur...
Paper: https://services.google.com/fh/files/misc/measuring_the_envi...
> One major question mark is the total number of queries that Gemini gets each day, which would allow estimates of the AI tool’s total energy demand.
Yeah, I was more interested in knowing the total amount. A "median" prompt without the information on the total number of prompts is kind of meaningless...
also an average would make much more sense here than a median, but we can assume it's much higher and that might be why they don't communicate it?
Contentless clickbait headline:
In a first, Google has released data on how much energy an AI prompt uses
One can rewrite clickbait or info-less headlines, while avoiding editorial. A good one is usually either in the sub-head or first couple sentences:
"Gemini apps ... median prompt ... consumes 0.24 watt-hours of electricity, the equivalent of running a standard microwave for about one second."
Possible contentful headline that fits in 80 chars:
Median Gemini text prompt consumes 0.24Wh energy, same as microwave for a second
Learned something even without a click, can decide if you want to!
---
Even better, skip the reblog and link the research, with a headline from the abstract:
• Median Gemini Apps prompt uses less energy than watching nine seconds of TV — https://services.google.com/fh/files/misc/measuring_the_envi...
Isn't it counterintuitive to use the median for this?
In this thread alone there are many comments multiplying the median to get some sort of totalt, but that's just not how medians work.
If I multiplied my median food spent per day with the number of days per month, I'd get a vastly lower number than what my banking app says.
The average is more useful if you wish to answer the question of how much energy the entire Google is using, but there are other ways to answer that question (such as Google's annual report, which tells you to 8 significant digits how much energy they used).
The median is more useful to answer the question of how much energy a typical Gemini interaction uses.
That's very interesting, although I'm still curious about the training resource usage -- not "only" inference. I wonder what is the relative importance of training (i.e., what percentage of the resource usage was in training vs. inference)
One random preprint I found (https://arxiv.org/html/2505.09598v2) estimates inference is 90% of the total energy usage. From some googling around, everything seems to agree that inference dominates (or is at least comparable to) training on the large commonly-used models.
I was less surprised that inference dominates training after I read that chatgpt is serving billions of requests per day.
Many papers on large models make claims about the energy used. Google has consistently said that training and inference use about the same fraction of their data center resources.
Same could be said for anything: human operator, google search algos pre-AI (where indexing = training), etc.
Is this just meant to be dismissive, or just like a kind of non-answer? That it could be said for anything doesn't make asking it of this specific thing unimportant or uninteresting.
Right, and the scraping, the extra storage required, the manufacture of all the extra GPUs, etc. This is them hoping people don't understand that the query is only one part of it and aren't curious enough to ask further questions.
It's rare to see this type of calculation anywhere, though I wish it weren't.
A miles per gallon number for a car doesn't count the diesel that went into the equipment to mine the ore to make the steel for the chassis, etc.
Isn't it less for Google using their TPU compared to everyone else using nvidia?
Maybe, but I would suggest that larger factors arise from utilization and overhead. These are going to be the competitive advantages that AI operators need to win the arms race.
TPUs are more efficient at LLMs because they do more TOPS/KWh.
"OpenTPU: Open-Source Reimplementation of Google Tensor Processing Unit (TPU)" (2025) https://news.ycombinator.com/item?id=44111452
"A Comparison of the Cerebras Wafer-Scale Integration Technology with Nvidia GPU-based Systems for Artificial Intelligence" (2025) https://arxiv.org/html/2503.11698v1
From https://news.ycombinator.com/item?id=44648575 :
> "Next-generation datacenters consume zero water for cooling" (2024) https://news.ycombinator.com/item?id=42376406
>> this design will avoid the need for more than 125 million liters of water per year per datacenter
"Microsoft’s Datacenter Community Pledge: To build and operate digital infrastructure that addresses societal challenges and creates benefits for communities" (2024-06) https://blogs.microsoft.com/blog/2024/06/02/microsofts-datac... :
> We will design and operate our datacenters to support society’s climate goals and become carbon negative, water positive and zero waste before 2030. [...]
> By 2025, we will procure 100% renewable energy on a global scale, both significantly expanding and decarbonizing local electricity grids.
> Our datacenter designs are more water efficient than traditional enterprise datacenters, and our plan by 2030 is to replenish more water than we consume locally.
Here's this about CNT cooling:
"Cyberpower begins selling desktop PCs with carbon nanotube CPU cooling" (2025) https://news.ycombinator.com/item?id=44899495
"A carbon-nanotube-based tensor processing unit" (2024) https://news.ycombinator.com/item?id=41322070
Graphene semiconductors should be at least 10X more energy efficient; but how much less water would graphene-based chips waste?
I asked GPT 5 to extrapolate this to a Max Claude code monthly usage, two sessions a day, just business days. It guessed that would be 21k prompts a month. Google’s proxy numbers give a monthly footprint of ~5.1 kWh, ~5.53 L water, and ~0.64 kg CO₂e for heavy CC use.
That's equivalent to doing less than two miles driving(CO2), one toilet flush (water) and about three dryer loads of laundry.
Using AI to calculate a trivial multiplication...
Yep, I'm weaker than you.
Sure, but now multiply that "new use-case" that we weren't consuming before by 7 billion humans on the planet... that's the issue
We can always keep adding new stuff and say each time "oh but it's small"... sure, but if we keep adding more, altogether it becomes huge
Authors of other studies who aren't affiliated with AI providers have come up with much higher figures: https://www.theverge.com/report/763080/google-ai-gemini-wate...
Any primary sources that aren't subscriber-only content?
The Verge has a soft paywall. Try in incognito/private browsing or archive.is.
What's the point of releasing this number?
Is it a metric for marketing to beat competitors with, like GPU speeds, etc.
"We're more efficient than those Global Warming bastards over at Amazon."
I assume they wouldn't publish them if it cast them in a bad light.
Me using AI to summarize the article;
Ҩ.¬_¬.Ҩ
Me sipping coffee with 1000 times more embodied energy while reading the article.
I think overall this seems good, we knew this is where the gains from TPUs would lie and there is a nice agreement here between the need for profit and more pressing planetary (anthroposcenic?) concerns. With things like this, one might ask how much do we consume binge watching something on a streaming service, or loading a page full of js and ads and canvas elements, or even just scrolling a feed for an hour on an app.
Still, there are a lot unswered questions here, and its up in the air precisely how this stuff will continue to integrate into services we already use, or habits we have yet to form at large. What does that scaling look like?
But by far the most troubling thing is the continued combination of flippancy, defensiveness, or silence we get from the AI peiple about even attempting to talk about this. If you are a True Believer, don't you want this to be something that is tackled head on, rather than tucked away? When this has come up before, I always end up seeing a bunch of guys who essentially leave the vibe of "well I am plenty above sea level, my AC is pumping just fine, and I just simply don't care because my productivity has doubled!"
Like isn't this community supposed to be excited about a future, eager to tackle problems? Or is there maybe some intrinsic solipsism to the impressive chatbots that ultimately renders this kind of attitude to its users? It feels like right when we were culturally about to age out of this particular form of obstinacy, we set ourselves up to create a whole new generation of "global warming is fake news" people. Which is a shame. If you're going to be like this, just go all in on accelerationism in all its pseudo-facist darkness, don't just borrow a script from baby boomers!
Since a human uses ~100W of power, the .24Watt-hours of energy for an AI prompt is about 40human-seconds [Edit: 9human-seconds] of energy.
And unlike the human who spent multiple hours writing that article, an LLM would have linked to the original study: https://services.google.com/fh/files/misc/measuring_the_envi...
[ETA] Extending on these numbers a bit, a mean human uses 1.25KW of power (Kardashev Level .7 / 8 Gigahumans) and the mean American uses ~8KW of power according to https://en.wikipedia.org/wiki/List_of_countries_by_energy_co.... So if we align AIs to be eco-friendly, they will definitely murder all humans for the sake of the planet /s
>And unlike the human who spent multiple hours writing that article, an LLM would have linked to the original study
Or made up a fake citation, complete with fake or unrelated author names, on the spot
Seconded, I’ve yet to use an LLM that can consistently provide sources that back up what it’s telling me.
> I’ve yet to use an LLM that can consistently provide sources that back up what it’s telling me.
Mmmh, that would have been my take as well up to around end of Q1 2025.
Theses days, the flagship LLM's have reduced hallucination by quite a bit, and are also way better at citing sources (you sometimes have to nudge them).
ChatGPT 5 has been very decent on that particular axis.
The Gemini app is pretty good at giving sources.
Experienced that today with 2.5 Pro. Though I was amused that all the links were links to Google searches for the source, rather than direct links.
In my chats with 2.5 Flash it gives me the direct sources lol. Also not going to lie, I've found 2.5 Flash generally gets straight to the point with sources a lot quicker than Pro. To the point I don't really use Pro.
Edit - just used Pro, gave me a direct source. Who knows...
> Since a human uses ~100W of power, the .24Watt-hours of energy > for an AI prompt is about 40human-seconds of energy.
.24 Watt-hours is 864 Watts for one second, so a 100W human takes ~9 seconds for that output.
You're right. It's clear I should have had an LLM write my comment rather than do it myself before a cup of coffee. I've already spent an order of magnitude more energy thinking about this article compared to an LLM.
Also, since I live a first-world life style which consumes multiple KW of power, I've probably consumed multiple orders of magnitude energy more than an LLM on this topic.
Perhaps sxp used an LLM to do his calculation.
> an LLM would have linked to the original study
There is a non-trivial chance that the LLM would've added a link to _something_, but links/references seem like a very common thing to hallucinate, no?
The way around that is that is for LLM-based tools to run a regular search engine query in the background and feed the results of that in alongside the prompt. (Usually a two-step process of the LLM formulating the query, then another pass on the results)
The used results can then have their link either added to the end result separately, guaranteeing it is correct, or added to the prompt and "telling the LLM to include it", which retains a risk of hallucination, yes.
Common to both of these is the failure mode that the LLM can still hallucinate whilst "summarizing" the results, meaning you still have no guarantee that the claims made actually show up in the results.
> The way around that is that is for LLM-based tools to run a regular search engine query in the background and feed the results of that in alongside the prompt. (Usually a two-step process of the LLM formulating the query, then another pass on the results)
Would the LLM-based tool be able to determine that the top results are just SEO-spam sites and move lower in the list, or just accept the spam results as gospel?
This is an extremely tricky question.
The practical and readily-observable-from-output answer is "No, they cannot meaningfully identify spam or misinformation, and do indeed just accept the results as gospel"; Google's AI summary works this way and is repeatedly wrong in exactly this way. Google's repeatedly had it be wrong even in the adcopy.
The theoretical mechanism is that the attention mechanism with LLMs would be able to select which parts of the results are fed further into the results. This is how the model is capable of finding parts of the text that are "relevant". The problem is that this just isn't enough to robustly identify spam or incorrect information.
However, we can isolate this "find the relevant bit" functionality away from the rest of the LLM to enhance regular search engines. It's hard to say how useful this is; Google has intentionally damaged their search engine and it may simply not be worth the GPU cycles compared to traditional approaches, but it's an idea being widely explored right now.
The only thing that can solve the misinformation from a bad LLM is the misinformation from a good LLM... with a gun.
>The way around that is that is for LLM-based tools to run a regular search engine query in the background and feed the results of that in alongside the prompt.
Hardly better, as soon those "search engine results" would be AI slop themselves, including actual published papers (phoned-in by using AI, and "peer reviewed" by using AI from indifferent reviewers)
This used to be a problem but it's been unheard of for a while
Happens frequently still with GTP 4o (and now 5) and Claude. Makes up "biographies" unrelated to the actual people, gives me bullshit non-existing API options and cli flags, etc.
People say this, but then never link to chats.
And then on the rare occasion they do link to a chat, their prompt is something like:
"Tell about a person of history who was important for their work inthe time of their existence and give quotes of what they said that made them important when they were speaking include notes and other stories about them and give details about their life who they married and their kids and who their parents were and other big things they did do in their lives"
Instead of downvotes, please prove me wrong.
It feels so common actually that I would not even bother sharing them. ChatGPT gives me hallucinated URLs all... the... time. Like several times a day. You can fix it by prompting it to Google the subject or research it otherwise it just vomits garbage.
LLMs are not databases, they are text processors that need to be trained on ungodly amounts of data and can sort of kind of double as a database, though a very fallible one. Inject ground truth and you're cooking; don't and it's a crap shoot. (not saying anything new here, but it bears repeating).
Instead of writing this rebuttal you could have just generated a hallucination and posted it.
If you haven't noticed several by now, then posting some wont do anything, it would be like trying to prove someone we went to the moon
Definitely not unheard of. Claude gives broken links to documentation at least once a week.
Did you use an LLM to do that arithmetic?
The issue is that now we're over-utilizing prompts everywhere. Every search engine query. Every keypress in an AI editor. Every new website makes queries in the back. Analyze emails. etc
So it's not just about "the one query you ask ChatGPT about what you should write your mum to say you're not coming for Thanksgiving"
It's rather that an AI query is 0.24Wh, but that we are now using thousands of those per users per days, and that we globalize it at the scale of the planet, so 7 billion users... and this becomes huge
>And unlike the human who spent multiple hours writing that article, an LLM would have linked to the original study:
This is why journalists are nearly universally hostile towards AI.
Press wire to blogspam as a service! There goes 90% of the "content" on news sites that seem to pride themselves on stretching out a sentence of info into 4 paragraphs.
I love good journalism because it's adhd crack; in-depth spilling the tea but you have to really dig deep to find it nowadays.
Attempts to reduce human experience down to a single number always seem so wrong-headed. A common one is when articles say "billions of dollars of productivity wasted to X". As if you can just replace any human activity for any human activity at any time and get the same results. I know companies would like to imagine that every employee is a fungible cog, but it seems pretty dismal as a world view.
Thanks, I was scouring the article looking for the original study, could not believe that they didn't have it linked.
Great, now consider the human's marginal power consumption for a task instead of their baseline consumption.
Unless your point is that we can kill a bunch of humans to save energy...?
Prompt: what will the best phone be in 2025? Wasted 0.24w/h and 5 drops of water.
Unless you're arguing that people should kill themselves to save their 100W of power, the future energy use of every alive human is already committed and doesn't matter.
Not to mention that the energy should also include all the extra energy spent on making converting energy into a form that is usable by humans (ie. food). There is probably at least an order of magnitude.
Consider this your second upvote from me, for the second paragraph.
I bet the human had LLM help. I bet it didn't take hours to put together.
TLDR; 100 prompts, which is roughly my daily usage, use about 24 Wh total, which is like running a 10 W LED for 2.4 hours.
>In total, the median prompt—one that falls in the middle of the range of energy demand—consumes 0.24 watt-hours of electricity, the equivalent of running a standard microwave for about one second
>The report also finds that the total energy used to field a Gemini query has fallen dramatically over time. The median Gemini prompt used 33 times more energy in May 2024 than it did in May 2025, according to Google.