I've never been so conflicted about a technology

They would if the patient is alive and fighting them! In our case, the patient is society and it is quite definitely not all unified that we cannot pick and choose. If the doctors want to save their patient, they either need to tie them down and do what is needed, or they need to work within the confines of what the patient allows while trying to convince them to allow more.

According to https://www.theverge.com/24066646/ai-electricity-energy-watt... a reasonable estimate is half a percent of global energy usage by 2027. On the one hand, this is a lot of energy. On the other, that's in the same general neighborhood as what we could save by maintaining proper tire pressure on our cars. (See https://www.factcheck.org/2008/08/the-truth-about-tire-press... for more about that.)

Naive question: why doesn’t the market regulate electricity consumption?

I’ve heard the answers pre-AI, but I wonder how this new general purpose use of electricity changes the calculus?

> massive amounts of electricity and water ... we need to succeed on all fronts

Datacenters' used for training or batch jobs can be placed in places where water and power are plentiful, ambient temperature are relatively low and government/society are stable.

A datacenter is going to be built at the bottom of New Zealand where all these things are true. There must be plenty of places in the world where this holds true.

At least for training, I think it is possible to have our cake and eat it. For real-time inference, probably not.

Okay, but there is some reason why you personally are still spending electrons posting on the internet and not, presumably, entertaining yourself by rolling coal, and it's probably because the amount of energy you're using to browse and post on HN seems trivial to you. Does the energy use differential between Python and C++ seem important enough to always pick C++ for everything?

What I'm saying is that casual LLM use is also essentially trivial, and scolding people for using querying ChatGPT when you wouldn't scold people for, say, making a dozen Google searches or watching a Netflix show or taking a long shower or driving to the mall, is quite silly and a waste of time.

Similarly, there's a reason why doctors will tell you to worry more about your cigarette habit than almost any other bad habits you have, because it's so much more likely to kill you that it practically demands to be prioritized.

I used to live in a place where there are a lot of data centers now. There were only a few just a few years ago. Now there are tens and more under construction.

I returned to that place to visit family and friends last year. It was an eye opening experience. The people who live there have taken a keen interest on curtailing further development of any new data centers. One of the chief complaints is the constant power issues that didn’t exist before 2021-2022.

The locals argue that this is the result of the dramatic infusion of AI into every technology product. They’re likely not wrong. The communities in the area became quite politically active over the issue and have retained all manner of analysts and scientists, journalists and investigators, and so on, to aide them in making a political case against future data center development.

The thing that got me was the complaints about the noise. Over the past few years the locals and those in their employ have been monitoring noise levels near the data centers and they’ve tracked sustained increases in noise pollution with the timeline over which the use of AI has exploded. There it has become sort of a trope to measure the working day by how much noise pollution is produced by any data center nearby. Mostly belonging to a hyperscaler known by an acronym.

The data centers have reportedly not been the boon that was promised, which seems to be seen as insult to injury. The area already has a vibrant tech scene independent of data center operations. So the locals don’t really see value in allowing more data centers to be built, and they’re starting to organize politically around the idea of preventing future data center construction and implementing heavy usage based taxation on utilities used by the existing ones.

Whataboutism is bad, but there's a distinction between whataboutism and making comparisons to get a sense of the true scale of the problem. If you're not allowed to ever compare relative scales of things without being accused of whataboutism then we're paving the way to making every molehill a mountain.

Two average American's beef consumption for training GPT-3 would place the cost in water of training GPT-3 at 1/150000000 of the US beef industry's average water consumption. That means it's a rounding error relative to all the other uses we place on water, and pointing that out is not whataboutism, it's putting the problem in its proper scale.

We have a constraint, say keeping warming below 1.5 or 2C, and we want to achieve it with as little pain as possible.

You use a laptop, you probably play some video games, you see movies, you probably eat meat, you probably drive a car around town instead of a velomobile.

So the question is, how much energy is used by AI and what does it get us? Google's latest TPUs consume ~250W per fp8-petaflop at the data-center level (ie including cooling etc, according to their recent presentations). Even assuming 50% utilization which is a bit poor, that's say 0.5kW per fp8-petaflop, or ~1kW per 10^15 dense-parameter-equivalent tokens per second. So using a huge model, say 10T dense-equivalent parameters, and doing inference at the rate of o3-mini or Gemini 2.5 Pro (around 150 tokens a second) consumes 1.5kW during generation time. But maybe a better way to think about it is just that you'd be using 10J of energy per token, or ~15J of energy per word.

In that context then, generating a book the length of Fellowship of the Ring (~180k words) would consume ~0.75kWh of energy. That's about like playing a video game on a PS5 for ~3 hours. In the US, that's on average ~270g of CO2, like driving ~1.5 miles or so in a Prius. It'd be about like riding ~30 miles on an e-bike.

Another way to think of it might be that if you are texting with a friend, wherein lets say you both aggressively type at ~120wpm, you'd be using ~15W of power to chat with an AI at a similar pace, about the power draw of your friend's Macbook Air he would be typing on. That's around 0.5-0.7 miles on an ebike per hour of chatting.

So even compared to your typical leisure activities, chatting with an AI is comparable in resource use to whatever else you'd be doing in that time most likely. And of course, a human in the US on average generates ~1.5kg of CO2 per hour they're alive in the US. In Bolivia, it's ~150g/hr, and in Sudan it's ~50g/hr or so. So if you're a company and you want to be climate conscious merely hiring someone who is a ~ subsistence farmer to do a job such that their standard of living rises enough to roughly emit the CO2 of the average Bolivian would mean they emit more carbon in a typical day than running an AI to produce as many words as are in the entire Game of Thrones book series to date. And if you were going to help someone immigrate from Bolivia to the US, you could have the AI write ~10 Songs of Ice and Fire a day for the same net CO2 output.

Not to say that we shouldn't do those things because of climate or whatever, but I'm just saying that the objective energy and climate impact of using these models for things is small compared to basically any alternative for completing a task and most entertainment activities people do (movie theaters for instance use ~7kW or so while running, so in the time you spend watching the Fellowship of the Ring in a theater, an AI could write a text as long as FOTR 28 times).

Of course water is not destroyed. It gets contaminated. And to clean it up, you need energy. Basically, we can simplify those equation to pretty much single argument: energy. Whatever you do, it needs energy. And people really really comes very lighty to that problem and waste management. If we would really like to recycle everything property I suspect global energy usage 10x fold at least. For now, its just better (I mean more economical) to just "store" that waste somewhere.

you could probably save energy by scrapping Python in favor of C++ everywhere but people don't seem too worried about that either

Water use can have big, localized problems, but that's mostly solvable by just putting the datacenters somewhere else.

Some places rely on deep aquifers that don't refill on human lifetimes, essentially fossil water. But that's mostly a local problem and we should just stop building water-hungry industry and agriculture in those places because it's stupid.

I wasn't even really thinking of the actual sausage stuffing process so much as how the meat gets to the butcher (which is really a meat grinding factory in most cases) in the first place.

And what happened after they told you they were shocked? Did they all become vegan once and for all?

People get upset because we use every part of an animal for food? I think it's absolutely great. I am a meat-eater who doesn't feel ethically great about it, but I'd feel much worse if I knew a large percentage of a carcass went to waste.

I'm not sure what documentaries you're watching.

Natural sausage casings are specialty items. If you're buying it at the grocery store, it's probably collagen (closely related to gelatin).

And it's not "all the waste". It includes fatty cuts that people wouldn't want to eat whole, but it doesn't include organ meats outside of specialty items.

Perhaps people find meat-grinding distressing, though it's really not all that different from ground beef. The emulsified filling of hot dogs and bologna looks odd, but the ingredients are inoffensive.

I'm less disturbed by sausage-making than by the slaughter and prime butchering of animals. They're no less dead and dismembered if you're eating a steak or pot roast. I'd rather we at least make use of all of the other parts.

Assuming a disdain toward sausage is straightforward and natural.

Hunger is an unforgiving teacher, and the reality that our forebears grew expert at squeezing every last calorie out of everything in sight might afford a lesson.

Because LLMs are not deeply embedded in all aspects of our society. It is very difficult to reduce existing usages of carbon (eg cars) in society, as opposed to stopping the widespread use of LLMs.

> Except with MCP, essentially most consumers will become "programmers"

You are more optimistic than I am. Most people I have seen are using LLMs are at best as an alternative to Grammarly or a document/web summary, or at worst making decision based on outdated LLM advice or as an inaccurate fact-engine.

The average person could code using Excel, but most don't even if they know how to use IF() and VLOOKUP().

Maybe AI will help us actually create or implement something that fixes it, instead of idealistic approaches that are never going to happen?

The cost of energy generated from burning fossil fuels is dominated by the cost of the fuel itself; the power plant costs much less. Non-combustion energy is the opposite: fuel costs are tiny or zero, so the construction cost for the initial plant is much more important.

If you generate a terawatt hour of electricity with natural gas, most of the cost will be from the fuel. A nuclear plant will have a tiny fraction of the cost come from fuel for the same amount of energy. A solar farm will have none of the cost come from fuel.

If AI lowers construction costs, it will improve the relative economics of non-fossil energy compared to fossil energy. A natural gas plant constructed at half the cost will have its final energy cost decrease just a little whereas a half-as-expensive solar farm will have its final energy cost decrease nearly by half. Making clean energy cheaper than fossils means that it will out-compete dirty energy even in locations where there are no explicit policies to reduce CO2 emissions.

You can see the effects on pricing advantage with this interactive simulation of electricity supply in the United States. If you cut the overnight construction cost in half for all generating technologies, solar and wind dominate the country:

https://calculators.energy.utexas.edu/lcoe_map/#/county/tech

Some example modeling of gas/solar electricity economics in the United Kingdom here:

https://electrotechrevolution.substack.com/cp/160279905

Companies have already started using robotics and AI to construct solar farms faster and at lower cost:

https://www.aes.com/press-release/AES-Launches-First-AI-Enab...

https://cleantechnica.com/2025/02/27/leaptings-ai-powered-ro...

https://www.renewableenergyworld.com/solar/cool-solar-tech-w...

It’s the possibility that the AI doesn’t just tell us what we already know but instead comes up with something we don’t. Maybe it comes up with an abundant stable renewable energy source. Maybe a really well working carbon capture.

You mean speeding up the destruction of the internet and media?

The problem is it was falling apart before generative AI. This AI has just hastened its demise.

rage against a tsunami

I know it's cynical. But I see more and more the destructive aspects of the web seeming to now eclipse the positives.

It may be age related too. I had lived some 30-plus years before the Web became a thing ... and I kind of liked the 70's and 80's.

Uehreka, I'm sorry, but this is such utter tosh. This is the sort of rabid utopianism I remember hearing about the web in 1996. But then of course it was predictive - at least a bit excusable. Take a hard a look at the actual history of the web in the last 30 years. "Democracy"...what? It is a story of corporate seizure of commons leading to the frankly nightmarish scenario we are in now, heading full steam into a total surveillance techno-fascist neofeudalism. Here's an innovation in the last millennium that did actually democratize access to information: Libraries. Simple public libraries. Look at the message that is engraved over the door of one of my favorite libraries, the central library in Edinburgh, Scotland. These places were palaces for public learning. Contrast them to the grotty, advert-saturated, casino of the web. Also, the calm, reflective, intelligent state of mind they cultivated. How many people can even read a book anymore, cover to cover? Or even a single page without electronic interruption?

You're probably right. I assume they're moving these servers to places like along the Columbia River where electricity is cheap (bitcoin miners have been there for a decade).

The marginal impact of AI on global climate probably rounds down to zero, or at least is not the thing driving us over the cliff:

> Global energy consumption in 2023 was around 500,000 GWh per day. That means that ChatGPT’s global 3GWh per day is using 0.0006% of Earth’s energy demand.

https://andymasley.substack.com/p/a-cheat-sheet-for-conversa...

The reason we have any environmental regulations at all was from mass death and disease caused by technology. Far more people died due to the agricultural revolution then will ever die due to climate change in the next few decades in the worst predictions.

The Black Death alone was 25-50 million people in 7 years

If we had stayed farmers, we wouldn't have these issues. Most people would be still happy. Intelligence is both destruction and savior.

We don't really need LLMs like we don't really need the internet. We don't really need that blog, nor do we really need Netflix, porn or Facebook.

Individuals value things differently, so attempting to do society-wide prioritization is always going to be a reductive exercise.

For example: Your local cafe doesn't need to exist at all. You could still drink coffee, you'd just have to make it yourself. That cafe is taking up space, running expensive commercial equipment, keeping things warm even when there aren't customers ordering, keeping food items cool that aren't going to be eaten, using harsh commercial chemicals for regular sanitization, possibly inefficient cooling or heating due to heavy traffic going in and out the door, so on and so forth.

Imagine the environmental impact of turning all cafes into housing and nobody driving to go get a coffee.

It's not a large environmental cost, though.

Their use does not come with a large environmental cost. The average American lifestyle has a “water footprint” of 1200 “bottles of water” per day. 10-50 ChatGPT queries == 1 bottle of water. If you decide to use ChatGPT but shorten your daily shower by a second or two you will more than offset your total water usage increase.

Thus LLMs don’t have to be that useful to be worth it. And if used in certain ways they can be very useful.

Source (with links to further sources): https://andymasley.substack.com/p/a-cheat-sheet-for-conversa...

Crossing your fingers and hoping for a sci-fi technology solution to be invented for climate change seems far more realistic at this point than expecting Taylor Swift to not dump 10 tons of CO2 into the atmosphere because she wanted to have lunch in France, so I'm putting all my eggs into that basket.

You do realise most cryptocurrencies aren’t PoW, so your point is living in the past, man!