I wonder what the cost differential is (any given rack, not just an underwater one) for a system which is designed to NEVER be maintained by people over its service life (or duty). Hard drive goes bad? Board fails? Just push the load elsewhere and shut off power to it. No idea what sort of attrition rate these systems have.
Sure, you might be at 50% capacity after 3 years, but that's just like lithium ion batteries.
Some of the big cloud providers (Google and I think Facebook) do exactly this in their DC's, when hardware dies they just leave it dead until enough has failed in that rack to make it worth swapping out the whole thing.
When you have a 100,000 of anything electrical MTBF will hit you every day.
I'm guessing that someone has done an equation and shown that to be cost effective. Does that calculation apply when the cost of building the data centre and running it are considerably higher due to the location you built in?
Interesting question. Perhaps you should look at the cost of spacecraft engineering for an example. You can't replace the hard drive on a probe billions of miles from earth!
I think what GP is proposing is significantly different than space craft: in space, you have limited space for redundancy, so you overbuild to avoid failure of key parts.
I think GP is proposing you start with, say, 11000 servers when you have a planned capacity of 10000, and just cycle off bad ones for spares (or just diminished capacity) until you redo the whole system.
I believe that's how Microsoft operates their server farms, but I could be mistaken.
Space is craaazy expensive for a lot more reasons though (ionization and lack of mission-redundancy are the big ones. Every launch is expensive, so every mission/satellite must have a bazillion nines).
Sure, you might be at 50% capacity after 3 years, but that's just like lithium ion batteries.