Re: Centralized all flash is just an impressively bad idea.
That's making a huge assumption - that many apps really need such high levels of throughput. SSDs historically have been about the IOPS, not about the data transfer rate. For sequential workloads for quite some time even 7200 RPM disks were fine for that kind of stuff.
And even though SSDs are about IOPS, put a dozen or more in a chassis and you have a ton of IOPS capacity.
NVMe is a good example here - for many workloads it simply won't have a noticeable impact to performance(over SSDs).
As you say that is how storage experts do it, for most everyone else they have to buy that expertise in the form of pre built solutions whether it is traditional enterprise storage and/or storage from a service provider.
It's very dangerous for someone to go out and try to design their own storage system, storage is very complex, both hardware and software. Worse than that it holds state. Data problems can be easily automatically replicated to other systems before anyone has time to react. Applications and operating systems are generally very unforgiving to variable storage availability and latency.