I think “not viable for research” is not a thing, haha. At least, if the ratio of compute to communication is right, we can test our algorithms on old hardware using smaller problems.

Edit:

The SuiteSparse collection (formerly UF matrix collection) is really great for this. You can download matrices with their year and some context about their application. The ones smart people were trying to figure out how to solve 20 years ago on clusters. And then you can marvel at how your laptop smashes them using PARDISO or some other direct sparse solver. Or you can try to figure out a clever way to solve them with 10 year old hardware.

ime, They usually get sold, either as full packages or per rack. There are refurbishers who tend to handle the piecewise extraction of all worthwhile components. See also, the recent auctions of Cheyenne

They aren't usually re-used as full compute racks because the infra required (cooling, power) is non trivial.

For the past couple of decades at least, supercomputers are decommissioned when newer HW becomes so much more power efficient that it's cheaper to buy new hardware than keeping the old one running. Usually that lifespan is in the range of five years or so.

Of course, if you have free power (or say, the power bill is paid by a different department), sure go wild and keep that old darling running.

If my local Craigslist is any indicator the parts will be stripped down and resold. There were some very cool and very expensive, when new, SGI machines I found on my local Craigslist page.

I quickly closed the browser tab when I found out about the power requirements and noise.

These computers will probably not make it to CL but eBay and niche auctions might sell them.