The USB-connected Edge TPU is widely adopted for edge inference, yet its performance characteristics remain poorly understood in real-world settings. Its inference latency is shaped not only by compute but also by parameter streaming over a constrained and potentially contended I/O channel. Prior work has largely relied on profiling with modified SRAM capacities, which provides only isolated insights and fails to capture runtime behavior when weights exceed on-chip memory or when multiple devices contend for USB bandwidth. In this paper, we present a streaming-aware latency model for the USB Edge TPU. In evaluation, we validate the model using several practical DNNs. Experimental results show that the predicted upper and lower bounds can effectively bound the measured latencies. Our work offers a foundation for future work on a predictive framework that generalizes to multi-device environments that require real-time guarantees, for which USB contention introduces nondeterminism.