Well, if you buy a bunch of GPUs and don't have buildings to put them in (a very real situation right now), then you're just .... holding hardware and not making a return off it and *the building (and related power infrastructure)* becomes much more important than how many GPUs you have.

There's one weird trick to preventing your GPUs from losing value. Buy all the ram.

It is called market collusion, and Jensen Wong confessed on air to "advising" SK Hynix CEO to align with his corporate projects a year before the "AI" supply chain take over.

https://en.wikipedia.org/wiki/Collusion

Technically against the law, even if folks reject the fact AMD & Nvidia are Monopolistic competitive firms in the GPU markets. Thankfully the LLM crowd are rapidly moving away from the CUDA ecosystem. =3

The trouble that that idea is that there comes a point where the generated AI needs machines to test the code that is generated... and it actually puts a limit on the growth of AI.

Another weird trick (that seems to be in play) - don't unpack them yet, so they are not yet depreciating assets! Listen to https://open.spotify.com/episode/1QjM5Ud5kjHzMIY5G7hjhW?si=o...

Compute hardware is ~50% of the cost of a datacenter. If we can assume a 5 year lifespan for the GPUs, that puts the ammortization timeframe closer to ~10 years.