DAWs need to be set up for the best worst case. In other words, having some reserve beyond its worst likely PEAK performance latency.
Just like peak meters are required to make sure full-scale is not reached, even though the RMS might be much lower, the same holds for performance, where the VST Performance is like a peak meter, and the OS CPU performance figures are closer to RMS meters in function.
You can have a system only using 20% on average, but a short peak CPU usage spike may allow the buffers to empty, resulting in hiccups. So, unless something is wrong, one would still need to increase the sample buffer size (or overclock the CPU) to prevent such spikes from creating problems in future.