Either case or both cases.
However, tx utilization's significance is a bit different from rx utilization, as the former's queuing can distort utilization.
For example, assume 10 ingress interfaces all egress 1 interface. Further, assume, 10 frames/packets all arrive concurrently, but the egress interface only can buffer 5. The tx usage will only account for the 5 frames/packets transmitted, but 50% of the traffic was lost.
Also keep in mind utilization is but a total possible capacity measurement over some time interval, actual usage is always zero or 100%. Such a measurement, alone, tells little about the interface, as all interfaces (themselves) should be able to sustain 100% usage, in and out, concurrently.