Yup, I kind of think that GPU/TPU obsolescence won't be a cliff, but rather a long descent.
Sure, the A100 Nvidia GPUs won't be used for training LLMs in 2026, but they have other uses and for more limited use cases could even be used for LLM training (smaller datasets, industry specific use cases).
Finally toward the end of life, the GPUs could be used for powering inference. AI search results, consumer grade LLMs, etc.
The ROIC will have tiers and a step-wise progression lower but likely longer than a 5-6 year depreciation life.


Hmm, I lean a bit the other way.
I think a useful analogy might be all those nice older aircraft parked in the desert. They work perfectly well, but nobody wants to pay to run them because they aren't as fuel efficient as newer beasts. Most of the value is in the core engine which is easy to sell, but there aren't any buyers because that's precisely where the inefficiency lies.

The electricity to run data centres adds up to a *very* large expense, and chip supply bottlenecks will ease at some point along with their pricing. Between the two of those, I would not be at all surprised to see a lot of current top-of-the-line chips being mothballed long, long before they stop working, simply because it's cheaper to plug in something newer and more efficient. (or sold to smaller firms--or gamers!--who are not running enough of them to care that much about the power costs). I can see top-of-the-line processing cards having a "leading data centre" residency half life of maybe only 3 years.

Purest speculation, of course. But I personally wouldn't risk any money predicated on their lasting a long time.

Jim