Overview
Embodied intelligence saw dense commercialization in September. At Huawei Connect, the Ascend 960 supernode debuted with near-package optics (NPO), targeting ten-trillion-parameter training and inference, scaling to 4,096 cards and 8 EFLOPS FP8—pushing optical interconnect inside the supernode. Unitree open-sourced UnifoLM-WLA-1.0, a ~6B general robot foundation model covering 64 whole-body and tabletop physical tasks, moving “one model for the whole body” from concept to open practice.
In parallel, Huarui Zhipu released the “Smart Superbrain” edge-compute platform on a 3nm SoC with a “single-chip brain-cerebellum fusion”; Shangwei’s Qiyuan robot went on sale from 19,999 yuan; and AGIBOT–Chimelong’s first embodied-intelligence theme park opens September 24 in Hengqin with over 300 robots on duty.
Background & Interpretation
1. Technical Background and Evolution
Embodied AI is crossing the “demo-to-delivery” threshold, backed by two lines: compute moving from cloud wan-card training to edge single-chip real-time control (cutting cross-chip latency), and models moving from modular control to unified foundations covering many tasks at lower migration cost. An IDC–Inspur report projects China’s intelligent compute at 2,576.5 EFLOPS in 2026 (+87.9% YoY), with agent count CAGR of 151.2% through 2030.
2. Core Drivers and Underlying Mechanism
Edge-compute breakthroughs free robots from high-latency cloud dependence, enabling local multimodal inference, VLA models, and task planning; open foundation models lower R&D barriers via public weights, accelerating the data–train–deploy loop. Huarui’s “brain-cerebellum on one chip” reduces communication loss and complexity—an engineering validation of the edge route.
3. Market Structure and Industry Chain Effects
Consumer sales and theme parks signal embodied AI moving from labs and industry into high-interaction public scenes—also stress tests of multi-robot coordination, stable O&M, and real-data loops. On capital, Xinchí, Xingyun, and Chaoswei recently raised nearly 2 billion yuan, showing domestic compute and robotics entering commercial兑现.
Implications & Outlook
Near term, edge chips and open models lower the R&D and mass-production barrier for robot makers; scaled operations in parks and convenience stores will accumulate high-quality interaction data. Key variables are edge-chip metrics, production cadence, software ecosystems, and humanoid success rates versus cost curves in real tasks.
Takeaways for Industry Participants
Robot makers and integrators should bind mature edge compute and open models to shorten iteration; scenario players (tourism, retail, industry) can build data assets through real deployment. Investors must separate “demo hype” from “delivery ability,” favor firms with chip–model–hardware synergy, and watch for concept speculation and order risks.
This column compiles industry information and shares technical perspectives; it does not constitute investment advice.