What's wild is OpenAI doubling down on hyperscaling when it's obvious that the gains from pre-training are coming to an end. They seem determined to just go out in flames...
It's the inevitable peak of the venture capital pipeline, just this time it isn't individual industries (e.g. taxis with Uber, hotels with AirBnB) getting squeezed out by unsustainable pricing - it's the economy at large that's suffering this time.
And it's high time for us as a society to put an end to this madness. End the AI VC economy before it ends our economy.
We need the corporate death penalty aka forced dissolution for egregious cases of misbehavior, we need easier ways to pierce the corporate veil (and I'm more and more inclined to actually support the death penalty here as well, despite the potential for abuse), we need corporate fines to all be measured % of gross income, at least double the profit margin.
And we need all of that fast.
There's a few historical examples here....
I highly doubt that. Memory chip production takes years to scale up, which is partially a reason why the memory market (both RAM and solid-storage) is so susceptible to "pig cycles" - high prices incentivize new players to join the market (although less likely than decades ago, given just how much capital one needs and how complex the technology has gotten) and for established players to scale up their production, and then prices collapse due to oversupply.
For GPUs, the situation is even worse. During the GPU crypto mining craze, at least that was consumer GPUs so there indeed was an influx of cheap second hand gear once that market collapsed due to ASICs - but this time? These chips don't even have the hardware for rendering videos any more, so even if GPU OEMs would now get a ton of left over GPU chips they couldn't make general-purpose GPUs out of them any more.
Additionally, this assumption assumes that the large web of AI actors collapses in the next 6 months, which is even more unlikely - there's just too much actual cash floating around in the market.
Of course, if one is inclined, I'd take a wild guess and say you could try something like Steam Remote, but I wouldn't bet on that actually working out. And even if you could get it working - per an analysis of German newspaper Heise, the bloody thing has less shader compute capacity than the iGPU of AMD's Ryzen CPUs [2]. 30.000€ - and it'll probably struggle running GTA 5.
[1] https://www.nvidia.com/content/dam/en-zz/Solutions/Data-Cent...
[2] https://www.heise.de/news/Nvidias-H100-kommt-im-deutschen-Re...
Of course I'm joking, there are LLMs to check these things now.
This is a "the horse might sing" situation for the whole market that focused on breakthrough-level results (AGI, ASI, or even just "not going off the rails after the third response").