in the case of large language models, the wire is the communication of your parameters between your layers of memory that is often the bottleneck. To do a forward pass, you need to use all parameters once, and so the communication between the compute and the storage is the bottleneck, and that bottleneck is also a bunch of wires.