The only data in the article that helps to understand the scale in context is this:
> It found data centers will account for 8 percent of total electricity use in the United States by 2030, a near tripling of their share today. New solar and wind energy will meet about 40 percent of that new power demand from data centers, the forecast said, while the rest will come from a vast expansion in the burning of natural gas. The new emissions created would be comparable to that of putting 15.7 million additional gas-powered cars on the road.
So all data center power (not just "AI") is currently 3-4% of US electricity consumption, and is expected to grow to 8% in 6 years.
I think we so have upcoming energy crunches between data center buildouts, EVs, heat pumps etc. I am not sure why the current news cycle is about AI. But I guess I don’t fully understand the amount of disdain AI gets as seen in the comments here.
In my opinion forcing the grid to modernize is a good thing.
One reason AI gets extra criticism might be that all the things you mention displace usage of fossil fuels. EVs may be powered partially by natural gas, for example, but that is still better than inefficiently burning gasoline. Heat pumps displace natural gas heaters or fuel oil heaters.
Because GPUs are so power hungry that in a conventional data center, a fully loaded HPE Apollo 6500 (which has 8x H100 cards) uses more power than a rack’s total power budget.
Running an LLM just for fun requires tremendous amount of energy. Training one from scratch is even more energy intensive.
Real AI applications are not much behind either. If you are processing any serious amount of data with AI, you use a lot of energy, again.
A smallish water cooled “super computer” class CPU cluster needs “half a megawatt” to operate.
As I said elsewhere, you can’t build a multi-megawatt data center “just like that”. Making the power you need available there is a multi-year, multi-institution endeavor on many cases.
Similarly, you can’t “just force” the power grid to modernize, because live infrastructure servicing millions moves very slowly for good reasons (like multi-day boot up sequences if it shuts down).
What is an LLM for fun compared to “real AI”. Hard to take your word for anything based on that.
Regardless, you absolutely can force the grid to modernize which is what is happening now with DC colocating near new and existing plants along with large inflows of investment into generators.
I work as a sysadmin at an academic HPC center, and some projects we support use AI for real causes like marine life and water health, and other ecosystem related work. These are production systems which provide real benefits and returns to humanity, and tries to secure our and other species future and on this planet.
When compared to these projects, an LLM is just a shiny toy which can do many tasks with worse quality, with no guarantee of correctness and with much higher energy usage.
Other numbers I gave is fresh from our data center, about the systems I work with directly.
You can dismiss all my comment just because I call a big, energy wasting NN a toy, but nothing I say on this comment is wrong, and I’m reporting them from the trenches directly.
> DC colocating near new and existing plants along with large inflows of investment into generators.
You can get tons of servers in six weeks, install them in two days, and get them running in a day, tops (for us it’s shorter).
If you can get your generators manufactured and delivered to your location in a year, you’re lucky.
This still doesn't help me understand, because we don't know what non-DC electricity usage will be in 2030. That could vary wildly depending on the adoption of electric cars, HVAC systems, water heaters, clothes dryers, et cetera.
As I said in comments on that one: This article claiming that "AI is exhausting the power grid" is surprisingly light on evidence that AI is exhausting the power grid. (In fact, I think it contains no such evidence, though it does contain evidence that tech companies have datacentres that use electricity and seems to think that that is an exciting and damning discovery.)
Energy would be a non-issue if we'd have went full-in on nuclear. It would be so cheap we would all be heating our houses with resistive heaters and running all the AI farms we want in our basements without a second thought. Instead we let stupidity win.
Tell that to the Navajo Nation. They're sitting on and around the 2nd largest source of uranium in the country. Between all the miners getting lung cancer and the high rates of blood cancer in everyone else due to the environmental neglect from the mining companies they eventually declared a moratorium on mining within their land.
I had a brief teaching stint out there on Shiprock. One of the first peculiar things I noticed was a clinic that I first thought offered some kind of radiation therapy, like an oncology clinic or what have you. Upon closer inspection I realized it was a clinic for those affected by radiation-related illnesses.
Meanwhile, given all the oil and gas activity, it certainly seems like that has given them far fewer problems.
You have to compare nuclear to other actually possible power sources, not some idealized hypothetical. How many coal miner get black lung, for instance? How many excess deaths from burning fossil fuels are there? Without that comparison, it’s a bad argument.
> It would be so cheap we would all be heating our houses with resistive heaters
"Too cheap to meter" was a promise in the 50's regarding nuclear power. Since then it turned out to be one of the most expensive ways to produce electricity if not the most expensive.
Lots of anti nuclear movements, managed to reduce investment into the nuclear power, and increase cost.
We have not opened a plant in 25 years.
Research plants have been closed down by anti nuclear activists.
Some of this movement is funded by Germany Governement.
This and all other articles misrepresent current energy consumption, then wildly predict the future from that baseline. The overwhelming trend is towards efficiency as comically inefficient corporate facilities are shut down in favor of cloud computing. Furthermore, half of current "data center power usage" is actually due to cellular radio networks, and that isn't growing due to AI.
I found a DGX H100 for $300k. The internet tells me such a server uses 10kw. I found a 30kw solar system with 90kw of lithium battery storage for $70k. At utility scale I don't think there's any risk of AI causing a problem to the grid when a dedicated supply of power for an AI server can clearly be purchased for less than 20% of the system cost. (Realistically I expect the lifespan of solar panels and the batteries is easily 2x-4x the lifespan of the server, so it's probably closer to 5% or even lower.)
Solar and batteries make power basically a solved problem. Relative to GPUs it's cheap and scalable.
Does that bump it up to 20kW? I'm not sure it really matters that much (nor does what the power generation method you use is.) I used solar and batteries because you can get them off-the-shelf in small quantities, but my main point was just to demonstrate that buying GPUs requires an incredible amount of capital. Powering them and cooling them requires some capital, but it's a fraction of the cost of the GPUs.
So, your battery is good for 9 hours of compute time. What happens if you have a couple of cloudy days in a row? And what's the insolation during the winter (in many places, it's 1/10th of the amount of summer sunlight)
“If we work together, we can unlock AI’s game-changing abilities to help create the net zero, climate resilient and nature positive works that we so urgently need,” Microsoft said in a statement.
Microsoft choosing to have Windows 11 obsolete hundreds of millions of perfectly functional computers tells you all you need to know about their actual commitment to the environment.
let the irony sink in - well done. Meanwhile, an early research paper on using fundamental ML/AI to measure and improve energy use in commercial buildings -- a public research paper from the USA -- has a "Cloudfront Identity Check" before you can actually download it.. or alternatively purchase it from Elsivier, using acceptable payment credentials you see..
Make the usage of fossil fuels a crime against humanity / all sentient life, and use nuclear reactors for base load with renewables for variable load.
Modernize the power grid by including excess energy storage mechanisms (batteries, water pumps, etc), harden it against solar flares, and create a profitable energy sink to keep reactors running (crypto mining) to further stabilize the grid.
Heavily subsidize electric vehicles to the point of gas trade-ins being a net profit for anyone with a car, and mandate conversion to electric vehicles.
Make all fossil fuels shipping barges illegal. Subsidize the usage of nuclear reactors on shipping barges. This can be done in several ways, such as expanding an agency like the NRC to manage reactors in consumer usages.
Those don't require a miracle. They require a sensible population to wake up and make the world better without compromises.
In order to replace the world's current energy mix with nuclear, we'd have to switch on a new nuclear plant every day for the next several decades. All of that new infrastructure would also require massive amounts of fossil fuels to build (which you've magically made illegal) and would only last a few decades before having to be replaced (presumably without requiring any fossil fuels by then).
There isn't even enough copper on earth to electrify everything.
U235 has an energy density around 80,000,000 megajoules per kilogram (MJ/kg).
Burning coal has an energy density around 24 MJ/kg for lignite (a lower grade of coal) to about 32 MJ/kg for anthracite (the highest grade of coal).
This makes utilizing U235 for energy around 2,500,000 more efficient than coal.
If we used thorium-232 instead, the theoretical energy density is around 79,000,000 megajoules per kilogram (MJ/kg) if considering the full lifecycle of the reaction.
The physics speaks for itself, regardless of people too dense to understand numbers.
Just because fossil fuels own our governments, industry, and most societal mechanisms, doesn't make the world they've fabricated real.
Usage of fossil fuels implies utilizing oil as a fuel. In other words, other usages of oil are still on the table -- plastics, chemistry, etc.
Infrastructure would not be so difficult to electrify. We already all know how to do it, and many countries are much further along than the US.
Water heating can be electrified. Heat pumps are already superior to other methods of heating / cooling. Trains can be electrified. Semis can be electrified.
I believe it would not be much of a challenge to make nuclear powered aircrafts with today's technology. It seems crazy because nuclear anything has been made to seem crazy, because it is literally free energy. The cost is overcoming our fears, and putting in some work to make it infallible on safety.
We have all the tools and all the capabilities. There are no longer excuses. The technology is here and has been here for a long time, to make a society that flourishes in every fashion.
What you are suggesting would require coordination on a scale that's almost unimaginable. The effort and coordination required would be similar to wartime mobilization.
The nuclear fuel cycle is exponentially less ecologically expensive than the fossil fuel cycle -- and it hasn't even been remotely optimized to the extent that it can be.
Nuclear fission is millions of times more efficient than any of the nearest energy sources when considering purely energy density. That efficiency compared to the cost and ecological danger is barely even tapped at this point, because it blows everything else out of the water by orders of magnitude.
The nuclear fuel cycle is economically expensive, not ecologically nor physically. And it is economically expensive because we are all slaves to the fossil fuel industry, who own nearly everything.
Why would slavers / masters living in luxury sacrifice everything so that their slaves would flourish instead? Answer: they wouldn't. Instead, everything would be rigged in their favor. Their most powerful opposition would be made to seem weaker than them.
>Make the usage of fossil fuels a crime against humanity
Delusional.
Fossil fuels have been unbelievably beneficial to humanity and are absolutely vital for modern society.
If a theoretical global ruler immediately outlawed fossil fuels they would be remembered as the greatest villain in world history, killing hundreds of millions and plunging us into chaos.
>They require a sensible population to wake up and make the world better without compromises.
No compromises? Sounds like utopian thinking. Tradeoffs exist everywhere.
This is another reason I am learning CLIPS. It implements a form of the Rete algorithm. A Rete network provides the same benefits of a neural network without the bells-and-whistles, and thus would be a great candidate for anyone who thinks they could benefit from the reasoning power provided by LLMs.
We have one group of people that finds it useful, and another group that doesn't find it useful at all. What are we to make of this? At the very least, it is not that it "definitely" is useful.
We have found there are different personality profiles of communicators that are more or less effective at using the tools.
Those that communicate like supervisors with a wide range of language at their disposal for approaching a problem from many directions can be more effective.
Note that none of these tools are useful for those working at the bleeding edge of human knowledge.
If only there was a power source that could be carbon free and constant (fission) that could be married to a variable load (AI and Bitcoin) to maintain peak efficiency
We do not have limitless power today, and much of the power today is being generated by carbon producing means which exacerbates the climate crisis. So yes, there is plenty of reason to not squander electricity today.
Typical techbro solution: Invent fusion. Thing is, you can take advantage of fusion power today with PV panels you can buy at Home Depot.
I realize it's not quite that simple. You need land, you need sunny skies, you need storage, etc. But all these problems are solvable with today's technology, while fusion reactors are most assuredly not.
It’s really strange, considering “net zero” data centers are all the rage right now. Microsoft, in particular, has solar, and sometimes wind as well, on-site generation, plus new builds require renewables in the power-supply agreements. The money’s there, so where’s the supply?
TFA literally says "The tech giant and its partners say they expect to harness fusion by 2028, an audacious claim that bolsters their promises to transition to green energy but distracts from current reality."
> It found data centers will account for 8 percent of total electricity use in the United States by 2030, a near tripling of their share today. New solar and wind energy will meet about 40 percent of that new power demand from data centers, the forecast said, while the rest will come from a vast expansion in the burning of natural gas. The new emissions created would be comparable to that of putting 15.7 million additional gas-powered cars on the road.
So all data center power (not just "AI") is currently 3-4% of US electricity consumption, and is expected to grow to 8% in 6 years.