The experience of in-vehicle AI agents depends critically on the allocation of compute between the device side and the cloud: relying entirely on a cloud-based large model will cause degradation to basic voice commands in no-network or weak-network environments, while also generating network latency and traffic costs.