> You are looking at maybe a millisecond of latency overhead [for 1 megabyte]

Considering the data transfer alone for 1 MB / 1 msec requires 8 Gbps, I have doubts. But for fun, I created a small table in Postgres 18 with an INT PK, and a few thousand JSONB blobs of various sizes, up to 1 MiB. Median timing was 4.7 msec for a simple point select, compared to 0.1 msec (blobs of 3 KiB), and 0.8 msec (blobs of 64 KiB). This was on a MBP M4 Pro, using Python with psycopg, so latency is quite low.

The TOAST/de-TOAST overhead is going to kill you for any blobs > 2 KiB (by default, adjustable). And for larger blobs, especially in cloud solutions where the disk is almost always attached over a network, the sheer number of pages you have to fetch (a 1 MiB blob will nominally consume 128 pages, modulo compression, row overhead, etc.) will add significant latency. All of this will also add pressure to actually useful pages that may be cached, so queries to more reasonable tables will be impacted as well.

RDBMS should not be used to store blobs; it's not a filesystem.