Hacker News Reader

Comment by mountainofdeath | Hacker News Reader

The same thing was said about the crazy build-out of fiber and telecommunications infrastructure. That infrastructure did prove useful but it took about 10-20 years before that was the case. It took 4G becoming broadly available and the ensuing increase of mobile devices to use at least of the overbuilt network capacity.

View on HN

While I get this stream of thought, the major difference is that the internal hardware will age much faster than the core infrastructure. The real question is, how much of these build out costs is the center itself and the hardware within.

That fiber build out will last for decades, a Blackwell GPU, not so much.

I could be wrong but it seems like in the case of a crash no one will be buying new GPUs and thus the existing ones could hold their value longer. Of course that value will no longer be massively inflated by bubble FOMO.

It isn’t about holding value, the cards are going to burn up. If they don’t, in 5 years one could run a rack of 4 cards at home at an affordable rate. Either the cards become affordable again and the datacenter is useless, or they don’t, and nobody can fucking afford to rent them.

GPUs definitely have higher failure rates than CPUs but I'm not sure what the absolute rates will turn out to be. If 10% of GPUs die within 5 years that's very high but also probably economically fine. If 50% die that's a disaster.

Sorry, I meant at some point the current cards in he data centers will be obsolete, financially. They’ll be sold on the secondary market. Buying 8 h200s at $150 a pop will either be a real thing, or they all burn up and capex explodes again, which would be a death knell.

Either way, the moat is about 7 inches wide.

>in the case of a crash no one will be buying new GPUs and thus the existing ones could hold their value longer.

No, because no one has any use for those monstrous GPUs outside of ML and some research projects. They can't even be dropped onto the consumer market because a SOHO is not equipped to house devices like that. The best case scenario is that the boards get dismantled and the VRAM gets salvaged for refurbishing. They've built these machines so specialized that they're essentially disposable.

There's a baseline paid demand for AI inference that can fully occupy today's GPUs (even after a crash) so there's no need to sell or scrap them.

What are you basing that on? Some of the demand that currently exists, exists because of all the money sloshing around the AI ecosystem (i.e. people using AI to sell AI solutions to other people), so how are you so sure demand can fully utilize all existing compute even after a crash?

Chalk an cheese, Wavelength Division Multiplexing took out Globalcrossing Worldcom etc…

But the secondary market that grew out of it was because once it is in the ground it has a long lifespan and low upkeep costs, this is not the same thing as ultra high power density data centers.

Cooling needs to be balanced with demand, they may not work for even cloud scale type loads without serious issues etc…

Not that it matters, my hometown has an announced DC and it is looking more and more like it is a shill, as do several of the others in the area.

The difference is that fibre is infrastructure, LLMs are an application. Who knows, maybe they will pass and leave behind that server infra, and that's where our digital consciousnesses will live once our bodies die.

> The difference is that fibre is infrastructure, LLMs are an application

When I zoom out, I see “token generation” as an infrastructural layer, with applications built upon it.

Reply on news.ycombinator.com