I don’t claim I know when the correction will happen but I wonder what massive alternative demand you see for the mountains of highly specialised server gpus in storage that will be obsolete in maybe 3 years time? For many of them that means likely before they will ever be turned on. The dotcom bubble created infrastructure that was, largely, not obsolete when the bubble bursted and made a lot of sense to salvage. That is a fundamental difference to inference infrastructure.
The WorldCom fiber layouts were akin to the railroad bubble in the 1840s, in that those were pathways with nothing to use them. I can see the parallels here, but the difference is GPUs aren’t nailed to the ground. They can be moved to demand, unlike railway lines and fiber lines.
GPUs process data. They don’t spoil or expire. Sure, they’ll lose value, but it’s not like they stop being useful, even if highly specialized. Hell, even selling them second hand to China with an export waiver would be a way to recoup value. So already, the premise is flawed in that, specialized or not, China will use them. Or the EU or universities looking for a deal and building out their own local processing.
Both rail and communication infrastructure lead to some useless connections but much of it was no useless, in both cases. GPUs are not bolted to the ground but they do become obsolete no matter if you deny it or not. The issue is that the real costs is in using GPUs is very different from these previous bubbles. Those obsolete GPUs will cause much higher operating costs than newer generations, to the point where they won’t be interesting to use even if you gave them away for free. To make matters worse, other infrastructure is much more flexible in its use, one can transport all sorts of things on railways, one can send all sorts of data on communication infrastructure. Those specialised GPUs aren’t very useful for anything other than a fairly narrow use case.
I think you do not fully appreciate the crazy amounts of GPUs we are talking about here. China has no massive real shortage of GPUs. They managed to get black market GPUs more or less directly from Nvidia just fine. Nor are European universities IT wastelands without compute capabilities. But even if they’d go crazy on expandig compute infrastructure with outdated power hungry GPUs, that would be barely more than a drop in the ocean. Nvidia does have to resort to circular financing to keep the boom cycle accelerating, with GPUs going just to some storage facility if they exist at all. That is not how healthy demand looks like.
If you are talking about GPUs being only a small share of the overall total sum, bad news that the supporting infrastructure is also to a large extend tailor made for that very narrow use case. No one else will need such huge data center facilities designed specifically for GPUs, that includes also the non GPU components. And the infrastructure is the only thing of substance of this bubble. The models aren’t it. Open weight models are on the heels of the closed models. As soon as they are good enough for common applications, the business case for charging billions is slowly evaporating.
You are also mistaken, I am not worried about GPUs. I am merely stating that they and their server infrastructure (which is tailor made for them) are rapidly getting obsolete equipment by their nature and while the clock is ticking they are largely not even being used. This is fundamentally different from the dotcom and railway bubble.
I don’t claim I know when the correction will happen but I wonder what massive alternative demand you see for the mountains of highly specialised server gpus in storage that will be obsolete in maybe 3 years time? For many of them that means likely before they will ever be turned on. The dotcom bubble created infrastructure that was, largely, not obsolete when the bubble bursted and made a lot of sense to salvage. That is a fundamental difference to inference infrastructure.
The WorldCom fiber layouts were akin to the railroad bubble in the 1840s, in that those were pathways with nothing to use them. I can see the parallels here, but the difference is GPUs aren’t nailed to the ground. They can be moved to demand, unlike railway lines and fiber lines.
GPUs process data. They don’t spoil or expire. Sure, they’ll lose value, but it’s not like they stop being useful, even if highly specialized. Hell, even selling them second hand to China with an export waiver would be a way to recoup value. So already, the premise is flawed in that, specialized or not, China will use them. Or the EU or universities looking for a deal and building out their own local processing.
Both rail and communication infrastructure lead to some useless connections but much of it was no useless, in both cases. GPUs are not bolted to the ground but they do become obsolete no matter if you deny it or not. The issue is that the real costs is in using GPUs is very different from these previous bubbles. Those obsolete GPUs will cause much higher operating costs than newer generations, to the point where they won’t be interesting to use even if you gave them away for free. To make matters worse, other infrastructure is much more flexible in its use, one can transport all sorts of things on railways, one can send all sorts of data on communication infrastructure. Those specialised GPUs aren’t very useful for anything other than a fairly narrow use case.
I think you do not fully appreciate the crazy amounts of GPUs we are talking about here. China has no massive real shortage of GPUs. They managed to get black market GPUs more or less directly from Nvidia just fine. Nor are European universities IT wastelands without compute capabilities. But even if they’d go crazy on expandig compute infrastructure with outdated power hungry GPUs, that would be barely more than a drop in the ocean. Nvidia does have to resort to circular financing to keep the boom cycle accelerating, with GPUs going just to some storage facility if they exist at all. That is not how healthy demand looks like.
In a $10 Trillion bubble, what percentage, exactly, are the GPUs you’re taking about?
Is it 10%? Are there a Trillion dollars worth of GPUs you’re worried about? Or less?
If you are talking about GPUs being only a small share of the overall total sum, bad news that the supporting infrastructure is also to a large extend tailor made for that very narrow use case. No one else will need such huge data center facilities designed specifically for GPUs, that includes also the non GPU components. And the infrastructure is the only thing of substance of this bubble. The models aren’t it. Open weight models are on the heels of the closed models. As soon as they are good enough for common applications, the business case for charging billions is slowly evaporating.
You are also mistaken, I am not worried about GPUs. I am merely stating that they and their server infrastructure (which is tailor made for them) are rapidly getting obsolete equipment by their nature and while the clock is ticking they are largely not even being used. This is fundamentally different from the dotcom and railway bubble.