Why does copying files start fast then slow to a crawl?
Because the first portion of the transfer is being absorbed by a fast cache, and once that cache fills you see the device's real sustained speed.
On SSDs: the SLC cache. Most consumer SSDs use TLC or QLC flash, storing three or four bits per cell. Writing multiple bits per cell is slow. To hide this, drives reserve a portion of the flash operated in fast single-bit (SLC) mode as a write buffer. Data lands there quickly, and the drive migrates it to the slower dense storage afterwards.
Write less than the cache size and everything is fast. Exceed it and the drive must write directly to TLC or QLC while simultaneously flushing the cache — and speeds can collapse dramatically, sometimes from thousands of megabytes per second to figures worse than a hard drive. QLC drives suffer worst.
The cache is often dynamic, shrinking as the drive fills — which is why a nearly full SSD slows sooner and more severely.
On the operating system: the write cache. The OS buffers writes in RAM and reports completion before data reaches the device. Small transfers appear instant; large ones eventually saturate the buffer and drop to real device speed. This is also why safe ejection matters.
Other contributors:
Many small files are far slower than one large file of the same total size, because each incurs filesystem metadata operations. A folder of 50,000 small files can take many times longer than a single file of equal size.
Thermal throttling. NVMe drives get hot under sustained writes and reduce speed to protect themselves.
Fragmentation and near-full drives on mechanical disks, where the outer tracks are faster than the inner ones.
Network or USB negotiation dropping to a slower mode.
Practical implication: benchmark figures reflect cached performance. For large sustained transfers, look for sustained write figures instead.