The speaker says limited NVMe capacity can force cross-server transactions to fetch from object storage, eventually slowing the database.

“There's a limit to the amount of NVMe disk space that you can have, so if you need to split transaction across two servers, you're basically then going to be fetching it from object storage.”

“Eventually, that will slow down your database. So we think that sharding is by far the best way to do scale up into the petabytes, and then you should treat your analytics differently.”