Actually thinking about it- I have inherited plenty of performance problems at the application layer, but that's because the joins were effectively being done at the application level (and making huge numbers of database calls) when they should have been done at the database layer.
Now a WD Blue NVMe does 95k/84k r/w iops at $215 and that's just off their website (https://shop.westerndigital.com/products/internal-drives/wd-..., may be more depending on shipping method...)
That said, it's not a fair comparison and I wouldn't want to run a big service on a single sqlite/nvme setup for more reasons than are worth mentioning, but not prematurely optimizing can take you really far - scale and money - with good design.
It could be 10 users that each views a few web pages once every month or it could be 10 heavy users for an internal system that uses the system most of the working day and creates thousands of calls per user to the system each day.
It is as a lot of people have mentioned also very different with information that can be easily cached and things like monetary transactions where you don't want to see old data in reads and therefore often can't cache or use read replica db.
For systems with heavy users I have seen the need to use a few servers even with just 1000 - 2000 users. For a web system the same can happen if you have 1000 users active on the site at the same time (within a minute or so).
I also work with fintech systems built upon .NET Core and have similar experiences regarding scaling of these solutions. You can get an incredible amount of throughput from a single box if you are careful with the technologies you use.
A single .NET Core web API process using Kestrel and whatever RDBMS (even SQLite) can absolutely devour requests in the <1 megabyte range. I would feel confident putting the largest customer we can imagine x10 on a single 16~32 core server for the solution we provide today. Obviously, if you are pushing any form of multimedia this starts to break down rapidly.