If you scale an instance you need to use object storage (s3/b2/etc). Fetch from object storage can occasionally have latency spikes.

5 seconds is somewhat exaggerating, I clicked through 10 or so videos on my instance to check and it's 2-3 seconds most of the time.

We can exclude rare enough outliers.

I've experienced B2 throwing a wrench into the dream of low latency, but some object stores are very fast. And more importantly you only need the first couple megabytes of each video to be on fast storage.