Handwaving this layer away as "very boring" isn't exactly fair, either. What does boring even mean here? I mean, this layer solves problems that are both essential to performance scaling of RDBMS and have been proven time and again to be hard to reliably solve in a general case. And it has furthermore been built from the ground up tailored towards the specific needs/use cases of the company.
By the aforementioned handwaving the presented successes are implicitly attributed to SQLite to a degree that isn't justified IMO.
1U of cohost with 100mbps in a cheap Eastern European DC will cost a few hundred euro per month... and my info is a few years old. It's more expensive now.
It is very specific case, that we should not extrapolate to mean it's general use case.
AWS/GCP/Azure are still a better places to start for most people.
(a) They implemented a very boring transaction/caching/replication layer that is like any other DB except they borrowed the idea that "longest chain" should be used for conflict resolution.
(b) They worked with upstream to get a few patches that were unique to their use-case. Once you're in deep with any DB this really isn't that uncommon.
(c) They used a dedicated (lol non-ephemeral) white-box server that has a lower amortized cost than EC2.
(d) Bedrock isn't bound to the hardware. You could run it on EC2 and reap the benefits just the same except you'd pay more.