Measuring the environmental impact of AI inference
arstechnica.com
Research paper: https://services.google.com/fh/files/misc/measuring_the_envi...
Google blog post: https://cloud.google.com/blog/products/infrastructure/measur...
arstechnica.com
Google blog post: https://cloud.google.com/blog/products/infrastructure/measur...
Playing whack a mole with individual behavior while the elephant in the room is energy production and transportation remains asinine as always.
1. Google rolled our AI summaries on all of their search queries, through some very tiny model 2. Given worldwide search volume, that model now represents more than 50% of all queries if you throw it on a big heap with "intentional" LLM usage 3. Google gets to claim "the median is now 33x lower!", as the median is now that tiny model giving summaries nobody asked for
It's very concerning that this marketing puff piece is being eaten up by HN of all places as evidenced by the other thread.
Google is basing this all of "median" because there's orders of magnitudes difference betwen strong models (what most people think of when you talk AI) and tiny models, which Google uses "most" by virtue of running them for every single google search to produce the summaries. So the "median" will be whatever tiny model they use for those models. Never mind that Gemini 2.5 Pro, which is what everyone here would actually be using, may well consume >100x much.
It's absurdly misleading and rather obvious, but it feels like most are very eager to latch on to this so they can tell themselves their usage and work (for the many here in AI or at Google) is all peachy. I've been reading this place for years and have never before seen such uncritical adoption of an obvious PR piece detached from reality.
> This impact results from: A 33x reduction in per-prompt energy consumption driven by software efficiencies—including a 23x reduction from model improvements, and a 1.4x reduction from improved machine utilization.
followed by a list of specific improvements they've made?
[1] https://services.google.com/fh/files/misc/measuring_the_envi...
The burden of proof is on Google here. If they've reduced gemini 2.5 energy use by 33x, they need to state that clearly. Otherwise a we should assume they're fudging the numbers, for example:
A) they've chosen one particular tiny model for this number
or
B) it's a median across all models including the tiny one they use for all search queries
EDIT: I've read over the report and it's B) as far as I can see
Without more info, any other reading of this is a failing on the reader's part, or wishful thinking if they want to feel good about their AI usage.
We should also be ready to change these assumptions if Google or another reputable party does confirm this applies to large models like Gemini 2.5, but should assume the least impressive possible reading until that missing info arrives.
Even more useful info would be how much electricity Google uses per month, and whether that has gone down or continued to grow in the period following this announcement. Because total energy use across their whole AI product range, including training, is the only number that really matters.
https://services.google.com/fh/files/misc/measuring_the_envi...
I think you are assuming we are talking about swapping API usage from one model to another. That is not what happened. A specific product doing a specific thing uses less energy now.
To clarify: the way models become more efficient is usually by training a new one with a new architecture, quantization, etc.
This is analogous to making a computer more efficient by putting a new CPU in it. It would be completely normal to say that you made the computer more efficient, even though you've actually swapped out the hardware.
Again, it's talking about "median Gemini" while being very careful not to name any specific numbers for any specific models.
This is the median model used to serve requests for a specific product surface. It's exactly analogous to upgrading the CPU in a computer over time
Gemini app is a specific thing: the Gemini App that actually exists.
How can Gemini App also include their internal augmented functionality on search which itself is not an application?
"Yes, this search is powered by a customized version of the Gemini model for its generative AI features."
Based on that, I'm not sure how it is reasonable to claim that Gemini App has a legal term that is exclusive of its use in search.
Amusingly, it refuses to answer if i ask "is this search powered by Gemini app?"
The fact that they do not report the mean is concerning. The mean captures the entire distribution and could actually be used to calculate the expected value of energy used.
The median only tells you which point separates the upper half from the lower half, if you don't know anything else about the distribution you cannot use it for any kind of analysis.
Of course, Gemini App (singular) means the mobile app. But it seems that the term Gemini Apps (plural) is being used by Google to refer to any way in which users can access the Gemini models, and also they do clearly state that a version of Gemini isused to generate the search overviews.
So it still seems reasonably likely, until they confirm otherwise, that this median includes search overview.
Look, there's not enough information to answer this within the paper. I'm not willing to give Google the benefit of the doubt on vague language, and you are. I'm assuming they're a huge basicappy evil corporation whose every publication is gone over and reworded by marketing to make them look good, and you're assuming... whatever.
That's fine by me, we disagree. Let's stop here.
I’m inclined to believe that they are issuing a misleading figure here, myself.
Again this is pretty similar to how CPUs have changed
> To calculate the energy consumption for the median Gemini Apps text prompt on a given day, we first determine the average energy/prompt for each model, and then rank these models by their energy/prompt values. We then construct a cumulative distribution of text prompts along this energy-ranked list to identify the model that serves the 50-th percentile prompt.
They are measuring more than one model. I assume this statement describes how they chose which model to report the LM arena score for, and it's a ridiculous way to do so - the LM arena score calculated this way could change dramatically day-to-day.
What if they are serving more requests?
But, wasn't it always so?
Wasn't it always so in business of all kinds?
Why should we expect anything different? We should have been skeptical all along.
- The report doesn't name any averages (means), only medians. Why oh why would they be doing this, when all other marketing pieces always use the average because outside of HN 99% of Joes on the street have no idea what a median is/how it differs from the mean? The average is much more relevant here when "measuring the environmental impact of AI inference".
- The report doesn't define what any of the terms "Gemini Apps", "the Gemini AI assistant" or "Gemini Apps text prompt" concretely mean
Now, I do agree it would have been nice to demonstrate this, however it could be genuine.
I can see the median being useful for answering what the cost of one more server/agent/whatever would be, but that’s not what this paper is asking.
In reality, we know what Google means by the term "Gemini Apps", because it's a term they've had to define for e.g. their privacy policies[0].
> The Gemini web app available through gemini.google.com and browser sidebars
> The Gemini mobile apps, which include:
> The Gemini app, including as your mobile assistant, on Android. Note that Gemini is hosted by the Google app, even if you download the Gemini app.
> The Gemini app on iOS
> Gemini in the Google Messages app in specific locations
> The Gemini in Chrome feature. Learn more about availability.
That established definition does not include AI summaries (actually AI Overviews) on search like you very claimed. And it's something where Google probably is going to be careful -- the "Gemini Apps" name is awkward, but they need a name that distinguishes these use cases from other AI use cases with different data boundaries / policies / controls.
If the report was talking about "Gemini apps", your objection might make sense.
The rest stands though - no models, no averages. User tovej below put it better than I did:
> The median does not move if the upper tail shifts, it only moves if the median moves.
> The fact that they do not report the mean is concerning. The mean captures the entire distribution and could actually be used to calculate the expected value of energy used.
> The median only tells you which point separates the upper half from the lower half, if you don't know anything else about the distribution you cannot use it for any kind of analysis
49% of queries could be costing 1000x that median. Stats 101 combined with a sliver of critical reading reveals this report isn't worth the bytes it's taking up.
It's very concerning that you claim this without previously fully reading and understanding Google's publication...
> It's very concerning that this marketing puff piece is being eaten up by HN of all places as evidenced by the other thread.
It's very concerning that you can just make shit up on HN and be the top comment as long as it's to bash Google.
> Never mind that Gemini 2.5 Pro, which is what everyone here would actually be using, may well consume >100x much
Yes, exactly, never mind that. The report is to compare against a data point from May 2024, before Gemini 2.5 Pro became a thing.
I don’t think that’s fair. Same would’ve happened if it were Microsoft, or Apple, or Amazon. By now we’re all used to (and tired) of these tech giants lying to us and being generally shitty. Additionally, for decades we haven’t been able to trust reports from big companies which say “everything is fine, really” when they publish it themselves, about themselves, contradicting the general wisdom of something bad they’ve been doing. Put those together and you have the perfect combination; we’re primed to believe they’re trying to deceive us again, because that’s what happens most of the time. It has nothing to do with it being Google, they just happened to be the target this time.
Off topic. I wanted to say somewhat counterintuitively I often upvote / submit things I disagree with and dont downvote it as long as sub comments offer a good counter argument or explanation.
Sometimes being top just meant that is what most people are thinking, and it being wrong and corrected is precisely why I upvote it and wish it stayed on top so others can learn.
Figure 2 in the paper shows the LMArena score of whatever model is used for "median" Gemini query. That score is consistent with Gemini Flash (probably 2.0, given the numbers are from May), not a "tiny model" used for summaries nobody is asking for.
I'm sure the relatively clean directed computational graph + massively parallel + massively hungry workload of AI is a breath of fresh air to the industry.
Hardware gains were for the longest time doing very little for consumers because the bottlenecks were not in the hardware but instead in extremely poorly written software running in very poorly designed layers of abstraction that nothing could be done about.
Fun fact: Deep Blue was a dedicated chess compute cluster that ran on 30 RS/6000 processors and 480 VLSI chips. If the Stockfish chess program existed in 1997 it would have beaten it with a single 486 CPU: https://www.lesswrong.com/posts/75dnjiD8kv2khe9eQ/measuring-...
I actually see growth in energy demand because of AI or other reasons as a positive thing. It's putting pressure on the world to deliver more energy cheaply. And it seems the most popular and straightforward way is through renewables + batteries. The more clean and cheap capacity like that is added, the more marginalized traditional more expensive solutions get.
The framing on this topic can be a bit political. I prefer to look at this through the lens of economics. The simple economic reality is that coal and gas plant construction has been bottle necked for years on a lot of things to the point where only very little of it gets planned and realized. And what little comes online has pretty poor economics. The cost and growth curves for renewables+battery paint a pretty optimistic picture here with traditional generation plateauing for a while (we'll still build more coal/gas plants, not a lot, and they'll be underutilized) and then dropping rapidly second half of the century as cost and availability of alternatives improves and completely steam roll anything that can't keep up. Fossil fuel based generation could be all but gone by the 2060s.
There are lots of issues with regulations, planning, approval, etc for fossil fuel based generation. There are issues with supply chains for things like turbines. Long term access to cooling water (e.g. rivers) is becoming problematic because of climate change. And there are issues with investors voting with their feet and being reluctant to make long term commitments in what could end up being very poor long term investments. A lot of this also impacts nuclear, which while clean remains expensive and hard to deliver. The net result of all this is that investments in new energy capacity are heavily biased towards battery + renewables. It's the only thing that works on short notice. And it's also the cheapest way to add new capacity. Current growth is already 80-90% renewable. It's not even close at this point. We're talking tens/hundreds of GW added annually.
Of course AI is so hungry for energy that there is a temporary increase in usage for coal/gas. That's existing underutilized plants temporarily getting utilized a bit more mainly because they are there and utilizing them a bit more is relatively easy and quick to realize. It's not actually cheaper and future cost reductions will likely come in the form of replacing that capacity with cheaper power generation as soon as that can be delivered.
We have seen many technologies which have been made so much more efficient (heat pumps, solar panels, etc). Really great achievements. Yet the amount of (fossil) energy we use still grows.
The benefits of technical solutions is that you get the desired effect without any real trade-offs. I don't really care if I use a boiler or a heat pump to heat my house, because the end goal is to heat my house. I don't really care if I use an electric car or dead dinosaurs car, I just want to get places.
Make the efficient, more climate-friendly alternative a better deal and most people will switch. Tell people that they should give up their cars and AC because the planet will be 3C warmer in 100 years and you'll get an eye-roll. If you want the more environmentally-friendly but also more expensive option to win then the only real option is government subsidies, not preaching - enlightened self-interest trumps all.
For most people, replacing your car with an electric one isn't a big deal. Replacing a car with public transportation is either impossible (living in the boonies), incredibly difficult (suburbia) or merely very annoying (city).
I very much doubt the average person is willing to give up his car for some nebulous greater good of some strangers half a world away, especially when he hears of Jeff Bozos of this world shutting down half of Venice for a wedding so 50 private jets can ferry fellow fat cats to have a good time. But you, Joe Schmo, ought to use paper straws, sit in 30C room in the summer and sit at home instead of traveling for vacations. To save the planet.
The situation isn't much different in non-Western countries. Over the last few years China did more for electrification from renewable sources than the rest of the world combined, and yet they're also building a lot of coal power plants because that's what they have so that's what they'll use, damn everybody else. India isn't going to willingly stay poor so that ivory tower elites can feel good about themselves. Countries with oil reserves, majorly non-western, certainly aren't going to not extract it for the good of the planet.
Also don't have to be sober to go home from the bar. I'm convinced ubiquitous public transport (especially on Friday and Saturday all night) informed German drinking culture.
Similarly you can go from point A to B to C to D to A without having to go back to B to get your big metal box and drag it to D. Exploring the city is way easier. If you've never experienced the freedom of walking around a city designed to be walked around... you should, that's a pretty basic life experience and it's weird how the US government has blocked it from you.
However, you might be too pessimistic here. Fossil fuel usage is actually widely expected to peak in the next few years and then enter a steady decline.
Michael Liebreich of Bloomberg NEF did a pretty interesting editorial on this decline a few weeks ago: https://about.bnef.com/insights/clean-energy/liebreich-the-p...
He uses a simple model with some very basic assumptions (conservative ones) where he shows how short term fossil fuel usage still increases. Mostly this is just market inertia. But then it will start decreasing and then some decades later, it declines all the way to zero with some long tail of hard to shift use cases.
He uses some very basic assumptions about economic growth continuing to grow by an average of 3%, a base assumption of renewables outgrowing energy demand increases by 3%, etc. You get to a modest fossil fuel decline by 2040, majority renewables powered economy by the 2050s. And virtually no fossil fuel left in the economy by 2065. The years change but the outcome stays the same as long as renewables outgrow demand increase.
There are lots of buts and ifs here but he's explicitly addressing the kind of pessimism you are voicing here.
About the “individual choice”: it indeed is, unless tech companies make bad choices. Like GitHub recently showed a button “what are my PRs?” When pressed it asked copilot to give you the list of PRs (incomplete btw). But there already exists a page for that! This is just wasteful and we should blame a company for that.
Thanks. This mindset is not always made this explicit.
That it is an individual-choice is just as true as the claim that it is a choice made by governments, corporations, non-profits, executives, etc. But this atomized fiction is the only one that is given focus. Why?
You said it yourself: the perspective is not even conducive to making any change! (“not a realistic...”) We can’t expect 8 billion to make atomized decisions for the betterment of the planet.
But that’s not what people with this mindset want. They want a scapegoat that (conveniently) cannot change. Or they want an excuse to keep doing what they are already doing. Because hey the entities “that are doing it” cannot change in the aggregate.
Then you have no solution at all.
And yes, keeping people informed is difficult but a crucial effort for a working democracy.
The key is to take advantage of economies of scale: The cost of renewable energy generation is mainly in the initial investment and equipment.
As long as you mass-produce enough equipment, the cost of each device will decrease due to economies of scale. However, thermal power generation is different. The cost of thermal power generation is mainly fuel, and the lower limit of the cost is much higher than photovoltaic power generation.
I don’t understand why so many people are obsessed with using fossil fuels for power generation, as if it is really more efficient... thermal power no longer has a price advantage a few years ago.
If it’s like Marvel sequels every year then there is a significant added training cost as the expectations get higher and higher to churn out better models every year like clockwork.
> LLM training & data storage: This study specifically considers the inference and serving energy consumption of an Al prompt. We leave the measurement of Al model training to future work.
This is disappointing, and no analysis is complete without attempting to account for training, including training runs that were never deployed. I’m worried these numbers would be significantly worse and that’s why we don’t have them.
This is not true of Gemini.
There are a lot of anecdotal reports of quality differences following some Gemini 2.5 Pro releases earlier in the year.
It’s kind of funny, because they keep talking about how close we are to AGI, and in reality they keep making the models dumber (uh, I mean more efficient).
But this is in the vanishing minority of frontpage AI threads where it's a really interesting concersation about quantifiable things: what quantization, what engagement metrics, what NDGC on downstream IR. People are complaining they gamed the number: that's an improvement! Normally they just lie. This is amenable to analysis and frankly an interesting one.
If it were up to me they'd flat regex ban "llm" and "ai" on HN, thats about the right ROC. But if we're going to have it? I'll take this over "How AI Saved My Vibecode Startup From Vibe Coding".
Is it, though?
There's a post in this discussion claiming that Google rolled out AI summaries on all of their search queries. This means they greatly increased the number of queries by triggering queries at each Google search. These are unsolicited queries that users do not send by themselves or want.
Then the post claims each of these unsolicited queries are executed using small models that are cheaper to run.
The post asserts these unsolicited queries represent half of the queries.
Google's claims are that now the median cost of their queries is lower. The post asserts around half of Google's AI queries are not requested by users and instead forced upon them with searches.
To me, what this spells is the exact opposite of a improvement. It's waste that is not requested by anyone and adds no value. It's just waste.
Consequently, if Google pulled the plug on these queries then the would reduce their total query count by around 50%. How much energy and carbon emissions would that save? Well, if you pick up that value and flip it over to show how much is being wasted, that's your "improvement".
Suppose you were running a computation that requires doing 33,000 multiplies. Later you find a way to do the same computation using only 1,000 multiples
That's basically what happened here
"Reduce 33x" and "make 33x smaller" are ambiguous, unclear, and inaccurate. Is something that's "33x smaller" or "reduced 33x" 1/33 of the original total or is it 1/34? The question can't be answered in the absence of more information.
These are common expressions, sure. They're also awful, belonging to the same category of error as:
* The price is expensive
* It's a good-quality piece
* All but one of my friends speaks like this.
* Here's an author whom we know cares about language.
* As well, this is how some people write.
That is, they're the errors of a normal native speaker.
Is "1x smaller" equivalent to "1x larger?" If it is, then 'it's 1/33' and "2x larger/more" means the same thing as "double the size/amount." But if you have two times more than I have, then you have what I have, plus 2x that amount. So you don't have two times as much as I have. You have three times as much. "2x larger," to my ear, clearly does not mean the same thing as "2x as much." "2x larger" should mean "3x as much." That's why "33x smaller" can be read as "1 part of 34."
When we're even stricter with sense, the expression "33x smaller" becomes completely incoherent, because 1x should represent the original quantity. A 33x reduction should give us a result of -32x.
Obviously that's not what the article means. It's what the words mean, though, when you read them literally mean, rather than reading past their literal meaning to the intentions of the speaker/writer.
Most people don't care whether someone means one thing or the other, because, as you wrote, it's close enough to give the general idea.
The problem that fussy people like me and the commenter above me have is that we want people to say what they mean. And I'd wager that most of us fussy people have to do more mental work in order to get to the result that other people reach intuitively. Having to ignore literal sense in order to read someone's intended meaning is harder for us/me than it is for most people. That's our/my problem. As a matter of sociolinguistics and pragmatics, we're wrong, because literal meaning takes a back seat to idiomatic usage. (It probably does even in this comment that I'm writing.)
That's why I said these are the errors of a normal, native speaker.
If you believe "2x more" and "2x larger" represent two different arithmetic expressions, such that (I assume) one of them means "3x" and the other means "2x," then you must have a much more advanced understanding or the language and of arithmetic than I have. Go with God, and good luck.
It's completely unambiguous and universally understood by people who work on efficiency.
It's an extremely common phrase
Because people can't handle fractions, they say something was made "33x smaller" instead of "1/33 as large".
It makes no sense by the individual meaning of the words, it is just a common (and for many people annoying) idiom.