Post

This guy is peddling the same tired “it’s just like the dot com bubble” talking points that you only draw on if you have no real interest in reality. He’s also the guy who wrote the infamous “advice to a poor black kid” blog. Legendary dipshit

The post-dot com bubble story doesn’t work here. Training ai models at scale is only possible with billions of dollars of funding. Ai data centers are not useful for other stuff at any meaningful scale. A data center will be as expensive to finish building or run in ten years as it is today.

but even if it *is* like the dot com bubble, all that dark fiber was still useful after sitting for years. in ten years today's GPUs in those DCs will be competing with GPUs 6-7 times as powerful.

In commercial real estate there is a facility (f) condition(c) assessment, resulting in an index score. Basically tells an investor if capital should be spent or the building should be trashed.

Typically a building needs 10-20% annually spent in capital to keep in relevant beginning at about year 10 through eternity if the building is maintained properly. Not including opex to run the building.

One (of many) things that bothers me about these datacenters is you see "projections" from AI companies like "we'll be profitable in 2030!" (although even that is based on magically discovering new technologies.) - but the GPUs that are core to their buildout will be obsolete by then.

Fair points but specific to GPUs; they’ll still have a lot of utility come 2030. They’ve been a niche element (ex., VDI) in data centers for many years now. So beyond LLMs there are/will be use cases *Note, I’m not discussing the sheer volume of the build outs making sense, just GPUs as a processor

Those use cases are teeny tiny compared to the amount of data centers being built to the point that they may as well not exist

Well, by 2030 they'll be very outdated, and they're not particularly well suited to other tasks, so they'll be an outdated card being used inefficiently. You can dismantle the datacenter and repurpose the cards, but you're selling them at a tiny fraction of what you paid.

AI inference/training boxes are so weird that it's not really economical to use them for anything else; Nvidia boxes run on their own ARM CPUs and a 4RU box will be pulling like 20kW. They're highly specialized for AI and not a whole lot else - they can't even really do graphics anymore.

I dunno, if I had a business model which was "version 1 loses money, but version 2 will be great and make lots of money" I'd want to spend as little as possible on gear that will only be used on version 1...

Ed called it a long time ago that the Guardian is in bed with ai

No it’s not