Most consumer-grade NICs have a default MAC address which is retrievable with device drivers, but delegate (Ethernet) packet assembly to the OS. If the OS asks the NIC to emit a packet, then the NIC often receives the packet as a blob, DMA'd from main memory, and emits the bytes as octets. Other NICs do manage packet assembly, but allow overwriting the default MAC address. By the time I was learning Linux, we had GNU MAC Changer available in userland with the macchanger command, and many distros have configuration for randomizing or hardcoding MAC addresses upon boot.

I want to say that this is all because olden corporate network management policies could require a technician to replace a NIC without changing the MAC address, but more likely it is because framing and packet assembly was not traditionally handed to a second controller, and was instead bit-banged or MMIO'd by the CPU.