The ASUS UGen300 is 40 TOPS of INT4 inference in a 105x50x18mm dongle drawing 2.5 Watts from USB-C. At $300, it attaches to any host with a USB 3.2 port -- no PCIe slot, no HAT interface, no custom carrier board. The Hailo-10H inside already appeared in the Raspberry Pi AI HAT+ 2; the UGen300 cuts the host dependency further.
The form factor shift matters for embedded development workflows more than for deployment. Prototyping AI vision or LLM inference on an industrial SBC, an existing product, or a testbench that lacks PCIe now has a $300 path that requires no board redesign. A developer can attach inference capacity to any Linux host in a cable swap, run the production model, validate latency and output quality, and only then commit to a PCIe or M.2 integration. That compresses the evaluation cycle -- the decision about which inference hardware to integrate gets made with real data instead of benchmarks from a different platform.
The pricing floor matters too. Raspberry Pi AI HAT+ 2 with PCIe is a similar TOPS number at a similar price but requires a Pi 5 host and the physical HAT form factor. UGen300 removes both constraints. When inference decouples from the host bus architecture, the hardware decision tree gets shorter. Teams shipping physical AI into embedded products should benchmark UGen300 before locking a platform.