Overview
On-device LLMs and embodied AI are shifting from showcases to shipments. On October 14, Shenzhen MaiteXin disclosed at the Bay Semiconductor Expo that its 3D-TPU chip has entered tape-out and is expected to mass-produce in Q4, using 3D-IC to lift bandwidth utilisation above 80% and cut power to about 7 W, with a four-tier roadmap from on-device language inference to embodied-action generation. The same day, Fudan University unveiled “Danbao”, an errand robot with a vision-only on-device brain at the 26th CIIF, completing spatial understanding and autonomous decisions on a sub-20 W chip; its VSIN navigation model has just ~120 million parameters, under a tenth of peers.
Background & Interpretation
1. Technical Background and Evolution
Robots have long relied on remote control, pre-programming and pre-built maps, requiring rework when environments or tasks change, hindering scale. On-device LLM breakthroughs free robots from cloud and presets, pushing “understanding the world” down to local silicon — critical for privacy, latency and offline usability, and a prerequisite for embodied AI to reach real scenes.
2. Core Drivers and Mechanisms
MaiteXin replaces the traditional unified bus with “3D direct interconnect”, shortening data paths and raising bandwidth use so power drops from 40 W-plus to ~7 W at equal bandwidth; its four tiers (EdgeLLM → Edge-MLA/MoE → EdgeVLM → EdgeVLA) progressively move language, multimodal and embodied action on-device. Fudan’s VSIN abstracts motion and manipulation into a “navigation” problem, running a light model in real time to address the “floor” of scalable physical intelligence.
3. Market Structure and Industry Impact
Shenzhen’s robot exports hit 11.13 billion yuan in the first eight months, up 12.2x, as embodied AI moves from demo to volume shipment. Synergy between on-device chips and robot OEMs (e.g. Rockchip RK182X, D-Robotics S600) is forming a domestic “chip-algorithm-robot” chain, reducing reliance on high-end GPUs and the cloud.
Implications & Outlook
As on-device compute costs fall, the embodied-AI tipping point may arrive within one to two years. Key variables: model-compression and thermal solutions, robot bill-of-materials control, and real willingness to pay in industrial and commercial scenes.
Takeaways for Industry Participants
Chip firms should focus on low-power, high-efficiency on-device routes and co-build ecosystems with OEMs; robot makers can prioritise high-frequency scenarios like inspection, delivery and cleaning; adopters should weigh offline on-device solutions for data-privacy and disconnect-resilience advantages.
This column compiles industry information and shares technical perspectives; it does not constitute investment advice.