How Much Energy Does It Take to Power Billions of AI Queries?
How Much Energy Does It Take to Power Billions of AI Queries?
All generative AI queries could hit 329 billion per day by 2030. See the big picture on AI's energy use, and how it's reshaping our world.
If it's more than a picowatt-hour, it's too much.
LLM delenda est.
These figures are too cherry picked for the shock value. You could go the opposite end and say that (these are all true, I've tried my best to research them):
8.5 Wh (average of all daily queries for a user) is also...
850 MWh (whole consumption of all AI queries in the world) is also equivalent to...
So yes - AI bad... But for other reasons. This is a diversion. Datacentres powered by coal are bad. Cruise ships are worse.
The problem isn't that the whole world needs less than a solar farm's worth of energy for AI. The bigger problem is the social damage of AI - including the fact that this "expansion at all costs" is justifying getting that energy from non-renewable sources.
But seriously, one single cruise ship uses more energy than all of the AI in the world. They serve no useful purpose and there are hundreds of those.
Considering only the queries ommits all the energy used in training models, scraping and preparing data and all the indirect energy from putting a greater load in servers all around the world from scraping them all the time. oh, and all the energy in the manufacturing processes of the hardware, and from building the servers. We must consider the consumption of the industry as a whole, or we're being biased as well.
Articles keep showing up explaining how queries aren't so power-hungry, but the corporations keep draining more and more energy, building more and more power plants that never keep up with the demand, and so on.
Datacentres powered by coal are bad. Cruise ships are worse.
Coal Vs Marine fuel. You've picked the kings of pollution.
Coal is worse for CO2. The free carbon it releases binds with oxygen in the atmosphere to produce more CO2 mass than the coal itself. It's crazy how much CO2 it generates.
Heavy fuel oil / bunker oil / marine fuel is cheap ass shit that contains masses of pollutants. So whilst it won't generate as much CO2, it will create a load of other stuff including Sulphur Dioxide. That creates acid rain.
Here's an idea. Let's do neither of them.
Cargo ships are even worse. The amount they pollute is staggering to such a degree that I'm glad trade between the USA and China is strained.
At least they’re doing something. Rampant consumerism aside they’re shipping physical goods in the most efficient way we have available to us. AI though? What the hell is it even doing? It ain’t being helpful that’s for sure, and that’s not even the goal of it, either, so instead of it being “teething problems” is just the usual bullshit of capitalists hurting everyone else over their latest obssession.
You could me cruise ships
Cargo ships are getting way more efficient and less polluting. I work on container ships which have a capacity ranging from 14k to 20k teu.
Imagine transporting these many containers from 1 continent to another via trucks. The amount of pollution they would be emitting on land.
There are strict regulations in place regarding emission. Now with the introduction of dual-fuel ships, they are going to be more environmental friendly.
It's a concept of the past that cargo ships are pollution factories.
Energy usage may not be astronomical right now but AI is going into everything. Like EVERYTHING! It'll be running even when you don't think it's running, when you think it makes no sense to be running. Usage of AI itself will skyrocket and the fastest energy sources that can be acquired are likely to come first.
More on this subject:
"We Went to the Town Elon Musk Is Poisoning"
"We Found the Hidden Cost of Data Centers. It's in Your Electric Bill"
“We Went to the Town Elon Musk Is Poisoning”
Wait, mobile gas turbines?
What a hack. No permits. And lying about it through gritted teeth.
And they are branded "Solar Turbines"? WTF?
AI bubble will pop by 2030. (I fucking hope.)
Sooner.
None of these AI applications are making money and unlike earlier IT companies (Amazon, Google search, social media site, etc ), the marginal cost of each additional user isn't near zero.
They are having to invest hundreds of billions to cope with demand for applications which lose money on each use.
It's a $50 billion dollar industry priced as a trillion dollar industry.
And there‘s still no compelling use-case for the average consumer. Coders and scientists? Can be. But most people don‘t really have a use for it in most situations, even in business contexts. It‘s mostly a solution in search of a problem, and even then it‘s so unreliable that even things trying to sell you it as a solution have to add the disclaimer that you shouldn‘t use it for anything that‘s remotely important.
So even if the costs were markedly less than they are, there‘s still no real path to profitability because there‘s no real call for it.
The only use I‘ve found as a consumer is using something like Perplexity as a search engine. And that‘s not a testament to how good Perplexity is, but instead a testament to how bad other search engines have become. Perplexity just avoids things like SEO and is mostly quite good at finding sources which aren‘t themselves AI-generated.
And…I really see a near future in which AI-SEO becomes a thing and Perplexity et. al. become just as useless as google.
I'm with you. I think the markets are going to be demanding results very soon now. When they do...Nvidia, Meta, Google, X, Microsoft stock prices are all going to go into free-fall.
And the training beforehand, distributed over models lifetime users?
"Could hit 329bn queries a day by 2030" - my guess is it probably won't, because the bubble will burst before it and most AI companies will crash hard