So, obviously, this looks hilariously dated in 2020. But even in 1990, when The Hunt for Red October was something of a phenomenon, if you were into computers this looked pretty dated. What I've subsequently come to realise, of course, through friends whose jobs involve maintenance of military equipments, is that most of it is pretty dated. I really start to wonder if, like Google with its servers back in the day, one couldn't build military systems that were capable and credible from off the shelf components, rather than paying billions to the military-industrial complex.
They often do. But the long lifetime and extremely focused role of this equipment means it rarely changes.
And I'm not sure the ballistic missile submarine computing market is ripe for disruption, or that less than a few billion is a reasonable cost target for a nuclear powered stealth multi megaton missile silo.
But a lot of the cost is due to making sure that things work in harder conditions than office/home, that the parts can be repaired quickly, etc. It also usually has to work for longer than typical lifespan targeted by server vendors (3-5 years for normal server, 20-30 years for industrial/military hw)
All of those requirements also mean that there's less made, so you have less exploitation of economies of scale - even if you use COTS chips, the market for, let's say, OpenVPX SBCs with specific connector configuration is pretty small. So even efforts to commoditise parts run into problems like that.
