They highlight the exact reliability constraint I was thinking of: that replacing failed TPUs is trivial on Earth but impossible in space. Their solution is redundant provisioning, which moves the problem from "operationally impossible" to "extremely expensive."
You would effectively need custom, super-redundant motherboards designed to bypass dead chips rather than replace them. The paper also tackles the interconnect problem using specialized optics to sustain high bitrates, which is fascinating but seems incredibly difficult to pull off given that the constellation topology changes constantly. It might be possible, but the resulting hardware would look nothing like a regular datacenter.
Also this would require lots of satelites to rival a regular DC which is also very hard to justify. Let's see what the promised 2027 tests will reveal.