Usually because a fast buffer has filled: memory absorbing the first part of the transfer, or a solid-state drive’s fast write cache. What you see afterwards is the sustained rate, which is the honest one.
Almost every case is one of four, and they are distinguishable by when the slowdown happens.
Memory absorbing the start
An operating system reports a copy as making progress once the data has been handed over, before it has necessarily reached the drive. For the first part of a large transfer, the figure on screen can therefore describe memory rather than storage.
When that buffer is full, the reported rate falls to whatever the storage is really doing. Nothing has gone wrong; the meter has stopped flattering.
The signature is a very high figure for a few seconds followed by a settled one, and it happens on the local machine rather than on the network.
A solid-state drive’s write cache filling
Solid-state drives absorb incoming writes in a fast region and move them into their main storage afterwards. While that region has room, writing is quick; once it is full, writing continues at the drive’s sustained rate.
Makers who publish both figures label them, and the sustained one is what governs a long copy. On a drive that is nearly full there is less room for the fast region, which is why the same drive can behave differently when it has little space left.
The signature is a copy that runs at one rate for a while — often minutes rather than seconds — and then steps down and stays there.
Heat
A drive on the later PCI Express generations steps back when it gets too hot, and recovers when it cools. That is protective behaviour rather than a fault, and it happens in machines with no airflow over the socket and in portable enclosures with plastic housings.
The signature is a rate that drops after a sustained period, recovers if you pause, and drops again — which distinguishes it from a cache, since a cache does not refill in thirty seconds.
The other end
A copy finishes at the speed of the slowest element in the whole path, and that is frequently the destination rather than the source.
A mechanical drive at the far end has a sustained rate of its own, and it is lower at the inner tracks of a full drive than at the outer ones.
A network enclosure may be doing more than storing: parity to compute, encryption to apply, or another job running at the same time.
A USB bridge in an enclosure has its own ceiling, and a fast drive behind a modest bridge never shows its own figures.
How to find out which it is
Copy the same large file locally on each machine, source to source and destination to destination. That takes the network out of the picture entirely.
Then copy over the network and compare. If the network copy matches the slower of the two local ones, the network is innocent and the storage is the limit — which is the usual finding, and the one that stops people buying a faster switch to fix a drive.
The questions that come up before an order
Is the slower figure the real one?
For a long copy, yes. Makers who publish both label them: a burst rate while a fast cache absorbs the data, and a lower sustained rate once it is full. The first describes the start of a transfer and the second describes the rest of it.
Why do thousands of small files behave completely differently?
Because most of the time goes on per-file work rather than on moving bytes — creating each file, writing its metadata, and the round trip that each of those takes over a network. That transfer never approaches the rate a single large file reaches over the same link, and no faster link changes it.
Last reviewed 10 September 2026