ASUS UGen300 Puts 40 TOPS of AI on a USB-C Stick

An NPU You Can Move Between Machines

ASUS has put the UGen300 on public sale, a fanless USB-C accelerator built around the Hailo-10H that delivers 40 TOPS at INT4 or 20 TOPS at INT8, backed by 8GB of onboard LPDDR4 running at 4266 MT/s. It costs $299.99, draws about 2.5W, weighs 145 grams, and works with Windows, Linux, and Android on both x86 and Arm hosts. CNX Software covered the launch on August 3, 2026.

• 40 TOPS (INT4) or 20 TOPS (INT8) from a Hailo-10H, connected over a USB 3.1 Gen2 10 Gbps Type-C port
• 8GB of dedicated LPDDR4 at 4266 MT/s on the accelerator itself, so models don't compete for host memory
• $299.99 for the 8GB stick, with a $259.99 M.2 variant and a lower-capacity 2GB version in development
• 2.5W typical draw, fanless, 105 x 50 x 18mm, supporting 150+ pre-trained models across Keras, TensorFlow, TensorFlow Lite, PyTorch and ONNX

Why the Onboard 8GB Is the Spec That Matters

Most consumer AI accelerators are inference engines with almost no memory of their own. They stream weights across the bus from the host, which is fine for a small vision model and miserable for anything language-shaped, because the weights have to keep coming and the interconnect becomes the bottleneck.

Putting 8GB of LPDDR4 on the device changes what fits. A quantized small language model or a vision-language model can be resident on the accelerator rather than shuttled to it, and the host's own RAM stays free for the application. That is the difference between an add-on that speeds up one camera pipeline and one that can hold a working model while your machine does other things.

What Kind of Host Does This Make Sense For?

The one you already own that has no NPU. That describes an enormous installed base: older laptops, small-form-factor desktops, Arm single-board computers, and industrial boxes that will not be replaced for years but could usefully run local inference.

USB is the point here. An M.2 accelerator is cheaper and neater, but it requires an available slot, a teardown, and a machine you're willing to open — and it stays in that machine. A stick moves between a development laptop, a test bench, and a deployed unit without a screwdriver. For anyone prototyping edge AI who does not yet know which host will ship, that flexibility is worth the premium over the $259.99 M.2 version.

The trade-off is honest: 10 Gbps over USB is real bandwidth, but it is not PCIe. Workloads that constantly move large tensors between host and accelerator will feel it. Workloads where a model lives on the device and you send it frames or tokens will not. Our edge AI dev board buyer's guide walks through the same decision from the other direction — when to buy a board with the NPU built in instead.

How Does 40 TOPS Compare to What Else Is Out There?

Favourably for the form factor, and it helps to have the reference points. Copilot+ laptop NPUs land in the 40-50 TOPS range. The Sixfab AI HAT+ with a DEEPX NPU brings similar acceleration to a Raspberry Pi 5 in HAT form. Dedicated edge AI boxes like the InHand Mo 68A on TI silicon sit lower at 8 TOPS, while Jetson-class modules go far higher and cost accordingly.

The number to hold onto is that the INT4 figure — 40 TOPS — is the one that applies to quantized language models, which is exactly where this device is aimed. Vision workloads running INT8 get 20 TOPS, still ample for multi-stream detection and classification.

Availability and the Rest of the Line

The UGen300 is on sale now through the ASUS online store and Newegg. ASUS lists supported use cases as large language models, vision-language models, computer vision, and speech-to-text — a broader spread than the camera-analytics framing most accelerator sticks get.

The M.2 sibling at $259.99 is the better buy if you know the host and it has a spare slot. The 2GB variant still in development will be the interesting one for volume deployments where the model is small and the price matters more than the headroom. For everyone else browsing mini PC and edge AI hardware, the USB version is the one that can be tested against three machines in an afternoon.

Sources: CNX Software — August 3, 2026; ASUS UGen300 product listing — August 2026; Hailo-10H product page — 2026.