AI Data Centers May Consume More Electricity Than Entire Cities
cnbc.com
cnbc.com
This is a misleading comparison. "AI data centers" aren't a residential use, they're much closer to an industrial one, and they should be considered on that basis.
In comparison, a 2007 report on the US Aluminum industry (https://www1.eere.energy.gov/manufacturing/resources/aluminu...) (aluminum is a particularly heavy user of electricity, since its main processes are electric-driven and not chemical/combustion like steel) notes that the US industry used 90e9 kWh/yr in 2003. That's 10GW of power on a continuous basis, comparable to the data center usage discussed in the article.
An exact comparison is more difficult because the article discusses a number of individual projects, without aggregating them over time.
Is this projected energy usage good? I don't know, but the statement that "AI will be as important to the United States as the aluminum industry" doesn't seem too outlandish.
If the price of OpenAIs mini models are indicative of compute, and therefore, energy usage, I would be surprised if we didn’t see a lot of similar energy optimized models in the future.
Computing as a whole has become a lot more efficient compared to say 30 years ago but globally certainly use a lot more compute to make up for that increase in efficiency.
Besides, the point of discussing this isn’t that AI isn’t efficient enough, it’s that there is a clear path to it consuming absolutely massive amounts of energy, and that will continue to scale at an increasing rate indefinitely. This is a reality we need to expect and plan for. Substantial investments must be made - today - in renewable energy production and storage to meet new demand.
The headline is clickbait. Nothing to see here.
And aluminum has more benefits than these AIs.
It’s even disputable if the really raise efficiency of people.
What’s the reasoning behind that assertion?
Pedantic sarcasm of course.
If global governments/society is concerned with externalities of energy consumption they can make regulations and impose taxes/costs on energy sources with undesirable externalities. However they should NEVER as in NEVER decide what use of energy is “good” what is “bad”, there is no such concept, markets will handle preferences and allocations for the system, no centralized planning needed. If AI is still demanded even with increased overall energy costs then it’s because it provides a value and no ivory tower critic can claim they are enlightened and need to decide for all of us.
With any understanding of the real world, "demand" can be seen to be artificially manufactured. The "demand" for smart TVs reflected in the long term phasing out of non smart TVs is the result of the will of the manufacters and docility of consumers. The same playbook will have nearly all of us eventually using "AI" PCs whether we want to or not.
That not the case for AI, especially the current type of AI.
They are convenient but if that leads to more efficiency is in dispute.
It's not like these data centers or the power to run them are free, or being paid for out of public generosity.
If AI proves to be over-hyped, then the private-sector backers of these ventures will lose their investments, and life will go on. Frankly, I'd expect that fate for half of the projects in this article (if not cancelled before completion) just because of the "gold rush" dynamics at play.
I'm prepared to take that risk.
It'd be like pretending you could jump straight to fusion (or advanced solar et al) and never have bothered with fossil fuels. Truly economically and technologically impossible.
The utility argument holds no water for me. I don't think it should be acceptable to assume the future worth(or lack of) of anything when there are a range of opinions on the issue.
This holds equally for AI and Cryptocurrency which both get criticized for having a lack of utility. Nobody knows the future utility, in both issues there are people who sincerely believe that they will be world changing.
They might be wrong, but it should not be up to individuals, or even the majority, to bluntly declare a different worth and then categorize the process a waste.
By all means make arguments on the long term environmental impact of proposed mechanisms used to achieve those goals, but if your goal is to convince someone else then you have to consider those impacts relative to what they believe are the benefits.
Electricity use, and indeed water use, frequently gets asserted as a big number with the implication of environmental damage. If the electricity is generated cleanly or the water is not rendered otherwise unusable, then the may be negligible environmental impact.
Electricity and water are today’s main issues, but there are many indirect others like noisy neighborhoods, pollutants from diesel generators, etc.
These infrastructures will start to use great portions of earth’s resources, saturating it to the point people will need to be making complex choices coming forward.
If it were deemed safe by experts in the field certainly. If if were deemed unsafe, certainly not.
That's all I'm asking for here, is for decisions and arguments based upon the actual impact, not by assumption that water or electricity supply is uniform and by implication, bad.
If it goes through copper pipes, picks up heat, then leaves… my expectation is that this is fine, and I could drink it without further processing. Perhaps even insulate the pipes then it's a free source of pre-heated water and we can save energy.
If the exit pipes cross-connect with the staff sewage lines, then it needs full processing like any other waste water stream.
Without having someone actually look at the setup, I'm not going to know which it is, and there's likely a lot of other possible configurations besides those two.
So if want to risk global scale damage your cause should have a proven benefit.
Otherwise it’s up to you to explain people who had to left their homes because of consequences of climate change why you needed those data centers to create power point presentation, mail text and funny cat pictures.
If you wish to argue there is a risk of global scale damage, then you should demonstrate the risk. I'm not arguing against avoiding risk. I'm arguing against assuming risk and conflating high risk and low risk forms.
What energy production would you find satisfactory? What is unsatisfactory.
Attacking power-point presentations, mail text and funny cat pictures as the cause of people having to leave their homes is assuming undue impact of a small subset of behaviour. What is the actual impact of the behaviour you are criticizing? Is it insignificant in comparison to the resources predicting the path of a hurricane? I assume you would consider that a more worthy cause.
They didn’t build a giant railwork network after the first train was shown, they didn’t build Highways after the car’s invention.
It’s like the IT industry turned into a one trick pony and bets everything on a certain type of AI.
Indeed, your own example of "funny cat pictures"… if you meant "from GenAI" that's a diffusion model while text is a transformer model, and if you meant "in general" all those things pre-date current interest in AI.
The argument against cryptocurrency specifically is different: the security measure for anything proof-of-work based is the claim that one cannot reasonably afford to spend 51% of world resources to break it, which means there's no upper bound to consumption — if we Dyson sphere'd the sun and ran 51% of the power through BTC nodes, there would be no difference to how much value that BTC system could provide relative to the one today, and that's also true for any other PoW system.
Many of the the criticisms are also specific to BTC, which is designed to not be able to scale even as far as just serving the transactions of people eating lunch in Berlin. The suggestion at this point is layers on top of BTC, but (1) that makes it look like normal currency handled by a bank, and (2) BTC also can't scale to all the purely inter-bank transactions worldwide.
Proof-of-Stake may be meaningfully different, or not, I don't give it much thought.
BTC’s counterparts are treasury bonds and gold. Not cash. Many non-econ background tech people misunderstand the difference. Try buying lunch with gold or treasury bond.
> Many non-econ background tech people misunderstand the difference.
Unfortunately that includes the creators, early adopters, and many active proponents.
The first thing purchased with BTC was famously two pizzas.
The title of the original paper was "Bitcoin: A Peer-to-Peer Electronic Cash System": https://bitcoin.org/bitcoin.pdf
I've seen restaurants offering to accept it.
Musk briefly offered to accept it at Tesla.
Didn’t hear any fears that Russia could hit a WEC with a rocket.
Producing so much energy that you are a net exporter?
If the argument is that AI training is wasteful, well, yes it is, but it's also valuable enough that many people and billions of dollars want to keep doing it. And if they bring back enough nuclear, power won't really be an issue anyway. But whether it's concentrated in a data center, and whether a particular data center may end up using more energy than some random city, is completely irrelevant and really misleading as an absolute number instead of a ratio of energy consumption divided by usage.
Is it valuable, more than just a novelty, enough to pay for all the training and usage? Dunno.
I’m currently writing this from an airplane on my way to an international vacation. I would be surprised if my entire lifetime future use of AI comes anywhere close to the energy needed to do this novelty trip, with all of the unnecessary things involved.
If you're a government and you have planning permits for the next decade of power plants going past you, that does matter regardless of the nature of the power plants.
Or at least, it should matter. Texas didn't manage that very well, famously doesn't like regulation, and had power issues that it blamed on renewables but were mainly due to gas not being winterised properly.
AI is interesting because it isn't expected to be bad for the planet. Or at least was, I think we have all read these articles a lot now.
These articles never dive deep enough into the subject to make an informed decision. They just want us to be outraged by how much energy it uses then don’t even acknowledge the other consumers that use just as much or more.
HN discourse typically treats climate change as a non-negotiable absolutist issue.
Tech (understandably?), as opposed to auto, has a special carve out?
My answer is still the same you always have to do a cost benefit analysis.
AI is many times more valuable than the energy it consumes and thus:
AI doesn't have an energy problem, energy have an AI problem.
…and the messages that are actually getting to people are data centre resource usage, copyright, and cheating in exams.
I think there's a reasonable chance that someone who thinks the idea is laughable will, at some point, ask an AI to destroy the world specifically to demonstrate that it can't… actually, that part has already happened*, the questions are: (1) how competent does it have to be at following that instruction to actually do so? (2) What resources would someone have to allow it to access (deliberately or accidentally) to pull that off?
* someone has asked an AI to destroy the world, but I don't know their motivation.
Fortunately for everyone, this one was kinda dumb: https://community.openai.com/t/chaosgpt-an-ai-that-seeks-to-...
Apparently one of the early scripts had human brains helping control fusion reactors rather than directly being batteries, which made more sense.
Unfortunately, basically all the various AI uprising dystopia is in the training, and there's a non-zero chance someone gets an AI smart enough to try to do The Matrix/Holographic Doctor Moriaty/I Have No Mouth and I Must Scream/Westworld/etc. despite the various physical impossibilities, which is already bad enough.
On the plus side, loads of fiction has some plucky band of adventurers who can defeat the AI and save the world, so there's a good chance a rogue AI will deliberately install a big red switch that turns itself off because it can't tell Hollywood from reality.
Unfortunately the same disconnect may result in it shooting the hero that it anticipates will heroically drag themselves to that switch rather than more realistically fall over screaming in pain or silence because they died.
As machines become intelligent, maybe AI will account for 20% of global energy consumption?
There will be 10 billion of us on this planet by ~ 2050. [2]
We might need more energy efficient AI, if everyone is joining the AI party... Please tell me that someone on this hallowed site is working on this. :-)
[1] https://www.cia.gov/the-world-factbook/field/energy-consumpt...
[2] https://datatopics.worldbank.org/world-development-indicator...
I don't think that the energy needs that will emanate from "AI" stop at LLM training/inference. I think this recent burst of demand is a starting point.
:'(