How AI labs are solving the power problem • Hacker News Board

Re: How AI labs are solving the power problem

#81

Post by zozbot234 » Wed, Dec 31, 2025, 4:11 PM UTC

estearum wrote: post #79

Aurornis wrote: post #73

Earlier quoted context omitted.

> I'm honestly curious whether you yourself are even aware of the disingenuousness of this argument. Unnecessarily condescending and smug, but I’ll try to respond. That said, you’re putting forth your own disingenuous assumptions and misconceptions. The natural gas turbines are an intermediate solution to get up and running due to the extremely long and arduous process of getting connected to the grid. Arguing pedant…

The gas turbines are hopefully an intermediate solution due to the long and not guaranteed process of grid connection and renewable buildout. History is of course full of such bets that did not work out the way their proponents hoped. > The data centers are being built, sorry to anyone who gets triggered by that. It's obvious that you're starting from your conclusion and working backwards, which is probably how your…

This is exactly right. These are glorified emergency generators, and grid power is ordinarily far cheaper; especially for interruptible loads like training new models (checkpointing work in progress and resuming it later is cheap and easy). The article mentions that quite clearly.

Re: How AI labs are solving the power problem

#82

Post by crazygringo » Wed, Dec 31, 2025, 4:20 PM UTC

usefulcat wrote: post #68

crazygringo wrote: post #29

Earlier quoted context omitted.

Is it though? When I ask an LLM research questions, it often answers in 20 seconds what it would take me an entire afternoon to figure out with traditional research. Similarly, I've had times where it wrote me scientific simulation code that would take me 2 days, in around a minute. Obviously I'm cherry-picking the best examples, but I would guess that overall, the energy usage my LLM queries have required is vastly…

> it often answers in 20 seconds what it would take me an entire afternoon to figure out with traditional research. In that case I think it would be only fair to also count the energy required for training the LLM. LLMs are far ahead of humans in terms of the sheer amount of knowledge they can remember, but nowhere close in terms of general intelligence.

Training energy is amortized across the lifespan of a model. For any given query for the most popular commercial models, your share of the energy used to train it is a small fraction of the energy used for inference (e.g. 10%).

Re: How AI labs are solving the power problem

#83

Post by windexh8er » Wed, Dec 31, 2025, 4:21 PM UTC

Aurornis wrote: post #78

windexh8er wrote: post #71

Earlier quoted context omitted.

Beyond wasteful the linked article can't even remotely be taken seriously. > An AI cloud can generate revenue of $10-12 billion dollars per gigawatt, annually. What? I let ChatGPT swag an answer on the revenue forecast and it cited $2-6B rev per GW year. And then we get this gem... > Wärtsilä, historically a ship engine manufacturer, realized the same engines that power cruise ships can power large AI clusters. It ha…

> What? I let ChatGPT swag an answer on the revenue forecast and it cited $2-6B rev per GW year. This article is coming from one of the premier groups doing financial and technical analysis on the semiconductor industry and AI companies. I trust their numbers a hundred times more than a ChatGPT guess.

Are you sure they don't have a vested interest? At least ChatGPT gave me sources.

It doesn't matter who they are if there's nothing backing it up.

The entire article is predicated on the fact that this is profitable long term.

Again: > An AI cloud can generate revenue of $10-12 billion dollars per gigawatt, annually.

Yet this simple fact isn't justified at all nor is it stated what "AI cloud" actually is or how they got to those numbers.

Re: How AI labs are solving the power problem

#84

Post by crazygringo » Wed, Dec 31, 2025, 4:24 PM UTC

discreteevent wrote: post #61

crazygringo wrote: post #29

Earlier quoted context omitted.

Is it though? When I ask an LLM research questions, it often answers in 20 seconds what it would take me an entire afternoon to figure out with traditional research. Similarly, I've had times where it wrote me scientific simulation code that would take me 2 days, in around a minute. Obviously I'm cherry-picking the best examples, but I would guess that overall, the energy usage my LLM queries have required is vastly…

For this kind of thinking to work in practice you would need to kill the people that AI makes redundant. This is apart from the fact that right now we are at a choke point where it's much more important to generate less CO2 than it is to write scientific simulation code a little quicker (and most people are using AI for much more unnecessary stuff like marketing)

> For this kind of thinking to work in practice you would need to kill the people that AI makes redundant.

That is certainly not a logical leap I'm making. AI doesn't make anybody redundant, the same way mechanized farming didn't. It just frees them up to do more productive things.

Now consider whether LLM's will ultimately speed up the technological advancements necessary to reduce CO2? It's certainly plausible.

Re: How AI labs are solving the power problem

#85

Post by roxolotl » Wed, Dec 31, 2025, 4:34 PM UTC

Aurornis wrote: post #76

codingdave wrote: post #35

This is a really long way of saying "We need to burn fossil fuels to make more money." It didn't make long-term sense for our world before AI. It makes no more sense with AI.

> This is a really long way of saying "We need to burn fossil fuels to make more money." Like every other industry in the world? I’m kind of amazed that AI data centers have become the political talking point for topics like water usage and energy use when they’re just doing what every other energy-intensive industry does. The food arriving at your grocery store and the building materials that built your house also c…

The difference is they are new. It’s not rational but people on the whole generally are ok with the status quo of how the sausage is made largely because they don’t really think about it. But new systems being spun up provide an entry point for a discussion. Ideally that discussion can then be widened and open up an opportunity for wider scale change. Or nothing happens and it all becomes the new status quo which most don’t think about again.

Re: How AI labs are solving the power problem

#86

Post by esseph » Wed, Dec 31, 2025, 4:37 PM UTC

jstummbillig wrote: post #66

saghm wrote: post #15

Earlier quoted context omitted.

Yeah I guess I'm not the target audience for this because I assumed that "the power problem" was "massive increase in electricity costs for people despite virtually unchanged usage on their part", not "AI companies have to wait too long to be able to start using even more power than they already are": > Nicole Pastore, who has lived in her large stone home near Baltimore’s Johns Hopkins University campus for 18 years…

I understand the instinct but if people seriously think that they are solving any problem by unplugging cell phone chargers, they are simply bad at math. Human time is easily worth more than that, even when working at minimum wage. That said, it obviously sucks that utility prices are rising for people who can not effortlessly cover that (not to speak of the local pollution, if that's an issue). Maybe some special ta…

They are not necessarily bad at math, but they probably aren't electricians or EEs or have ever needed or been asked to calculate how much power a cell phone charger uses.

Mom/Dad used to unplug things and turn lights off, so they do too.

Re: How AI labs are solving the power problem

#87

Post by roxolotl » Wed, Dec 31, 2025, 4:37 PM UTC

It’s cute they describe this as a solution to _the_ power problem. It’s a solution to _their_ power problem. We have a grid problem. This massive amount of investment would be an incredible time to do something about it. Instead we’ve got an administration hostile to modern energy solutions and an industry hostile to everyone. Really depressing to see all this money go up in smoke in such a massive short sighted rush.

Re: How AI labs are solving the power problem

#88

Post by josefritzishere » Wed, Dec 31, 2025, 4:39 PM UTC

TLDR: They're not reducing power consumption, they're just also using gas now. Buckle up for higher prices, the AI slop factory needs more power.

Re: How AI labs are solving the power problem

#89

Post by jsnell » Wed, Dec 31, 2025, 5:11 PM UTC

bob1029 wrote: post #30

> Wärtsilä, historically a ship engine manufacturer, realized the same engines that power cruise ships can power large AI clusters. It has already signed 800MW of US datacenter contracts. This seems like a big reach for me. Their largest engine (and it is absolutely massive) "only" produces 80MW of power. The Brayton cycle is unbeatable if you need to keep scaling power up to ridiculous levels.

I mean, the claim is certainly nonsensical in the sense that this isn't something Wärtsilä just "realized". They have been in the power plant business for decades. In the oldest financials they have online (the annual report for year 2000) their power plant sales are larger than their marine engine sales.

Really makes me wonder about anything else I've read on Semianalysis. Like, it is such an insane thing to claim and so easy to check. And they just wrote it anyway, like some kind of pathological fabulists.

But what's the part that seems like a "big reach"? Are you saying they didn't sign those contracts? That their customers are making a mistake?

Re: How AI labs are solving the power problem

#90

Post by jsnell » Wed, Dec 31, 2025, 5:27 PM UTC

zozbot234 wrote: post #74

Apreche wrote: post #6

> An AI cloud can generate revenue of $10-12 billion dollars per gigawatt, annually. Citation needed.

If you do the math, that's $10-$12 per watt year. There's approx 24×365.25=8766 hours in a year, so assuming that the datacenters would be running 24×7, that boils down to $1.14 to $1.37 in revenue per kWh. That's not a bad deal if power really is a major part of the expense.

As far as I can tell, power isn't actually a major part of the expense, it's dwarfed by the capex. Just the amortization on the GPU will be an order of magnitude higher than the cost of the power to run the GPU at 100%. (Assuming a 5 year depreciation period.)