Comment by mpalczewski
7 hours ago
there's a big difference between now and then. we built up capacity. dark fiber. and couldn't use until we got broadband. In general. ai capacity is being used right now today.
7 hours ago
there's a big difference between now and then. we built up capacity. dark fiber. and couldn't use until we got broadband. In general. ai capacity is being used right now today.
> In general. ai capacity is being used right now today.
Not according to this article[1], which says:
"I believe we are now in an inevitable overbuild situation, one with no neat, tidy Dot-Com Bubble-style exit story. Demand for NVIDIA GPUs — and those from Broadcom, AMD and other semiconductor companies — is driven by speculative capital believing that the AI industry will become magnitudes larger than it is today, largely driven by the fact that everybody believes there’s far more demand for compute capacity than actually exists. ... NVIDIA has created a remarkable illusion perpetuated by the media — that GPU sales are a direct measurement of the actual demand for AI compute, rather than a measurement of how a few companies are willing to invest in an idea two years in advance, using circular financing as a means of creating the sense that you must buy these GPUs now, or you’ll miss out on the future. ... At the very least, hyperscalers are going to be burdened with brutal depreciation charges or onerous write-offs for years to come, whether their capacity turns into revenue or not. ... I don’t see how 90%+ of NVIDIA’s sales ever end up generating a single dollar of revenue, and considering the amount of project financing-backed data center debt deals, there’s very little that exists to protect investors if AI compute demand never arrives."
[1] - https://www.wheresyoured.at/wherere-all-the-ai-chips/
That's Ed Zitron, and his calls on AI matters routinely get outperformed by that of a broken clock.
The man saw the market demand for saying "AI bubble is going to pop and AI tech is going to wither away and die", and went to meet it on the supply side - truth be damned.
So far, AI companies still keep getting bottlenecked on compute, AI utilization increases - driven by, among other things, increased price-performance of AI making it viable in more and more roles. And the demand for both AI inference and AI hardware in general shows no signs of stopping.
In 5 years from now every device. Even your oven or washing machine has somekind of local LLM running for sensor reading and smart decision models. I can already think of quite some useful use cases for those devices. So I do believe that AI becomes magnitudes bigger than today. It's a logical step in our digital journey.
Why would you need the LLM to be running on the device itself? Wouldn't it be more efficient for the intelligence to be more centralized? Maybe in one central computer per home, or (probably even more desirable for corporations) for the devices to phone home to a centralized server at the service provider in the cloud somewhere.
1 reply →
Why do you need an LLM to read a sensor? This is truly AI pilled delusion.
That’s not exactly true, a lot of GPUs are sitting in warehouses because they are lacking prerequisites to being installed, in some cases this is even just power.
https://www.tomshardware.com/tech-industry/data-centers-in-n...
I don’t think the article you cite supports your point with confidence you convey. Yes other dimensions can complicate deployment, but incentives to work around those issues are high, given the cost of GPUs
That’s not the same. Those GPUs are allocated to real projects with real demand. If they could turn them on today they would do it and they would have customers.
It’s not the same as dark fiber capacity being overbuilt for demand that wasn’t there yet.
> If they could turn them on today they would do it and they would have customers.
Is this really true? Other than the occasional outages, the frontier labs don't seem to be getting overloaded by API pricing customers at peak times, and with Opus 5.5 using even less tokens (which translates into less GPU time and less money spent on the same tasks), it seems actual capacity isn't literally strained right now. If current needs are met by currently available capacity, where is this demand going to come from?
There appear to be a lot of uninstalled GB200s (looking at you Azure), but the power efficiency of those is ~10x worse than a Vera Rubin. Does it really make sense to install them, when your data center is power/cooling limited, and it's not built yet?
Maybe somebody will use them, but not the original owners.
Yes. I have 3 GPUs in the other room, sitting in boxes, waiting to be powered on, because I dont have the electrical circuit installed in the room yet.
I can directly confirm this fact, and I'm just some random dude out here in suburbia, who can only imagine what the Indiana Jones warehouse of glowing GPUs looks like in a data center.
We're all salivating over this imagined future where an AI toilet seat greets us every morning.
Who here has already asked AI how to make your own DRAM ? I know you have. Thats a sign of peak interest in a subject or market.
> We're all salivating over this imagined future where an AI toilet seat greets us every morning.
I'm not, but I can certainly believe the surveillance industry vectoralists are salivating over a future where their AI toilet seats greet each of their data subjects every morning.
1 reply →
… and cooling
That’s one big difference. The other big difference is that capital back then was mainly spent by startups with no or little traction.
Today, the "old" corporations like google, meta, amazon are also spending a lot of money.
This does not seem to be the case. The very first VeraRubin NVL72 systems are barely installed and won't be running real workloads until beginning of '27. Colossus II appears to be the largest user of Blackwell Ultra NVL72s and those are also only recently on-line in September of this year with the remaining 3/4 of them not until end of the year. The delays in datacenter deployment are very long. Even their GB200s didn't come on-line until January of '26. Hopper H100/200s still dominate.
The other hyperscalers are even further delayed. CoreWeave has some VeraRubin, but not in volume (as I understand).
https://en.wikipedia.org/wiki/Colossus_(data_center)
https://www.moduledge.com/blog/nvidia-hopper
We’re doing a similar thing now too, building tons of extra power capacity and links to data centers.
After the crash no one will need to build a new data center for 20 years, and there will be plentiful power connections
> ai capacity is being used right now today.
A lot of its current use is just basic questions that could be handled by local llms just as well, yet they are being processed on GB200s on the other side of the world.