Ugreen's $20K Smart Home Hub Runs on NVIDIA's Jetson Thor

What Ugreen actually shipped

Ugreen — better known for phone chargers, power banks, and NAS boxes — used IFA Berlin to launch HomeAgent, its first smart home platform. It bundles security-camera storage, home automation control, and on-device AI into a single hub, managed through a companion speaker/voice assistant called Uliya. The pitch is that video analytics, facial recognition, and file management run entirely on local hardware, not a vendor's cloud.

• HomeAgent HA100 ($1,799): Rockchip RK3588 chip, 16GB LPDDR5 RAM, 26 TOPS — rule-based automation and basic object recognition.
• HomeAgent HA100 Pro ($5,999): CIX CP8180 SoC, 32GB RAM, 205 TOPS — motion detection and behavioral-intent inference for voice-driven automation.
• MasterAgent MA100 ($20,000): NVIDIA Jetson Thor (a Blackwell-based system-on-module) with 128GB unified memory and roughly 2,070 TOPS of AI compute.

All three keep the NAS heritage — SATA bay, M.2, and SD card slots for local storage — and support Matter and ONVIF out of the box, with Zigbee and Thread planned. Pre-orders opened September 3 and run through October 27, 2026, with shipping targeted for January 2027.

Why the silicon spread is the actual story

The interesting part isn't the smart speaker or the camera lineup — it's that Ugreen shipped *three different compute tiers* instead of one chip and a software toggle. That's a tell. Local facial recognition, natural-language home automation, and continuous video analytics have genuinely different compute footprints, and the company is pricing that difference directly: 26 TOPS for basic rules, 205 TOPS for intent inference, 2,070 TOPS if you want a data-center-class module doing real-time reasoning at the edge. Jetson Thor is the same Blackwell-generation platform NVIDIA positions for robotics and industrial edge AI — seeing it repurposed for a consumer/prosumer home hub is a sign of how far Blackwell-class inference has already trickled down from the data center.

For anyone architecting an AI-in-the-loop product — trading signal generation, fraud scoring, KYC checks — this is a live, priced example of the cloud-vs-edge trade-off: how much silicon does 'runs locally, no round-trip to a third party' actually cost once you account for the model, not just the marketing checkbox.

Local-first isn't the same as open

HomeAgent's protocol support (Matter, ONVIF, eventually Zigbee/Thread) suggests real interoperability, but the platform currently has no Home Assistant integration — the open-source hub most serious smart-home tinkerers already run. AES-256 protects local data and encryption covers remote access, but the automation logic itself stays inside Ugreen's stack. It's a familiar pattern: privacy-by-architecture (data never leaves the device) paired with a closed automation layer that still locks you into one vendor's roadmap.

The builder's takeaway

Strip away the camera and speaker SKUs and HomeAgent is a case study in matching compute tier to inference workload — exactly the kind of architecture decision that shows up anywhere an LLM or vision model needs to run somewhere specific for latency, cost, or data-residency reasons. Whether that's a Jetson module in a hallway or an inference endpoint behind a trading system, the question is the same: what's the minimum silicon that makes 'local' true rather than aspirational.