I wonder if doing this to hardware would have overall been cheaper than doing the baroque Rube Goldberg shit we do to software to make it somewhat kind of fault tolerant… or at least convince ourselves it is.

(You can write real fault tolerant software but a wad of tangled crap and hidden dependencies crammed into Kubernetes is not it.)

For a compare-and-contrast, Tandem is the Stratus competitor that went more toward software FT (though it does have redundant hardware). You can also look at various n-of-m voter redundancy setup that are very common in aerospace computers. Definitely not rube goldberg, but also not cheap if for no other reason than you need 3+ of everything (e.g. the primary Space Shuttle computer ran 4 synchronized voters).