GPU Depreciation?

Viewing this from a total supply standpoint: The aggregate shortage of GPUs ensures that "legacy silicon" (older chips) retains significant economic value.

We are currently in a deficit of GPU resources; there is no oversupply of next-gen chips that renders the previous generation obsolete. Furthermore, because GPUs are high-value capital assets, old chips cannot simply be swapped out immediately for new ones. Instead, the strategy is to extend their service life as long as possible.

An extension of the first point: The demand for tokens is growing exponentially. Behind the general shortage of GPUs lies a world where the demand for total token volume is skyrocketing.

This implies that the utility value of every single GPU will be squeezed until the very last moment.

Old GPUs remain highly effective. Although the single-chip compute power of a legacy GPU pales in comparison to a modern one, the same computational results can be achieved by stacking them in larger quantities, thanks to the nature of parallel computing.

There is a mismatch between power resources and GPU resources. If a company has secured power capacity but possesses only legacy GPUs—or lacks the capital/access to acquire the latest hardware—their only viable option is to continue operating the old hardware.

In this scenario, pragmatism trumps the theoretical pursuit of maximizing "tokens per kilowatt-hour."

There is a tiered distribution of GPU resources globally. While top-tier US firms deploy the absolute latest 1-2 generations of GPUs, older hardware (or its compute capacity) can be sold or leased to second-tier domestic players, or cascaded to other nations, including Europe and China ("The East").

A complete, mature secondary market supply chain exists within the semiconductor industry, and GPUs are the crown jewel of this trade. The only lingering concern is whether, due to export bans, China will maintain sufficient channels to digest this massive scale of future secondary chips.

Legacy chips can deliver performance that exceeds their original specs. If "neocloud" providers or other users possess superior software technical support, they can optimize legacy chips to maximize efficiency and token generation.

This allows old hardware to command a premium price. Referring to the CoreWeave (CRWV) example mentioned earlier, I believe such companies provide the necessary assurances for high-efficiency chip utilization.