WAIC 2026: AI Shifts from 'Answering' to 'Accomplishing', Embodied Intelligence and Edge Computing Take Center Stage

From July 17 to 20, 2026, the World Artificial Intelligence Conference (WAIC 2026) was held simultaneously across three venues in Shanghai: Expo, Zhangjiang, and West Bund. The exhibition area exceeded 100,000 square meters for the first time, attracting over 1,100 enterprises and showcasing more than 3,000 exhibits, of which over 300 were global debuts. The academic lineup included nine Turing Award and Nobel Prize laureates. The core trend shifted from 'generating content' to 'accomplishing tasks', with embodied intelligence, edge computing power, and AI agents emerging as the three most-watched directions.

Embodied Intelligence: From Demonstrations to Real-World Scenarios

Yuanli Lingji: 6 Robots Build a Great Wall Model in 15 Hours

Yuanli Lingji, in collaboration with Stepfun, challenged WAIC attendees: 6 robots (4 desktop + 2 wheeled) autonomously assembled a 3.5m long, 1.5m wide, 1.1m high Great Wall model using over 80,000 micro building blocks (precision 0.1-1mm) within 15 hours. No remote control or pre-set scripts were used; the process was driven by their self-developed embodied general foundation model DM0.5. DM0.5 adopts a VLA and world model integrated architecture, introducing a context abstraction layer (60-second memory), embodied CoT (11 reasoning tasks), and trajectory alignment layer. It achieved SOTA with a 43% overall success rate on the RoboChallenge Table30 V2 real-machine evaluation. The post-training framework DFOL2.0 incorporates the world model DW0.5, reducing real-machine trial-and-error costs through virtual environment pre-simulation.

Qianxun Intelligence: Long-Horizon Tasks and Multi-Robot Collaboration

Qianxun Intelligence showcased Moz1 and Moz2 robots performing a 'tidy up the living room' long-horizon task in a simulated living room: autonomously identifying objects, planning paths, executing grasping and placement, and handling new tasks (e.g., a crumpled paper ball) without restarting. Their Spirit v1.6 model adopts a VLA and world model integrated architecture, reducing long-sequence computation costs through abstract action tokens, and introduces execution monitoring and dynamic replanning capabilities. Moz1 has been deployed on CATL's power battery production line for high-voltage test plug insertion, achieving over 99% stable operation success rate.

Pudu Robotics: Global Deployments Exceed 130,000 Units

According to a Frost & Sullivan report, Pudu Robotics ranks first globally in commercial service robot revenue, shipments, cleaning robot revenue, and Chinese commercial service robot exports. Its products have been deployed in 85 countries and regions, with over 130,000 units across 16 industries including catering, hospitality, and manufacturing. The self-developed embodied intelligence large model PuduFM includes a physical prediction model PIM, multimodal alignment VLA, and world model simulation engine, supporting a 'one brain, multiple forms' architecture that requires only 50 expert demonstration data points for new task adaptation.

Ant Lingbo: Launches LingBot 2.0 Full-Stack Models

Ant Lingbo released and open-sourced six LingBot full-stack 2.0 foundation models before WAIC, culminating in the industry's first embodied native world action model, LingBot-VA 2.0. The team chose pre-training from scratch rather than fine-tuning to build a full-stack layout. On the evening of July 19, the core team held a closed-door sharing session near the WAIC venue, making their first public appearance.

Edge Computing: New Infrastructure for the Agent Era

Cix Technology: Launches AGX Agentic Compute Strategy

On WAIC's opening day, Cix Technology announced the AGX Agentic Compute strategy, unveiling the desktop supercomputer AGX Station, capable of locally inferring 70B to 150B parameter models, supporting multi-machine cascading into a desktop cluster with 160-320 TOPS total computing power. Its self-developed 6nm chip, Cix P1, integrates CPU, GPU, and NPU, claiming to be the only domestic SoC+NPU solution achieving over 50% reduction in single-token energy consumption. AGX Station adopts a modular open architecture, supporting M.2, MXM, PCIe, and other AI accelerator cards, and has been adapted to multiple domestic AI chips from companies like Tianshu, Haimo, and Yuanli.

TsingMicro: Reconfigurable Architecture and 4K Supernode

TsingMicro showcased its reconfigurable computing architecture chip, claiming that traditional architectures have transistor utilization below 40%, while its reconfigurable dataflow engine can boost it to over 70%. Its 4K supernode solution organizes 4096 chips via a Mesh network, reducing interconnection costs by approximately 90% compared to foreign alternatives. The RAISA software stack supports nearly 1,000 mainstream operators, has adapted over 200 large models, and collaborates with the Zhiyuan FlagOS ecosystem to achieve operator-level adaptation on the day of DeepSeek-V4 preview release.

AI Applications: From Tools to 'Experienceable'

Yu'ai Weiwu: AI Learning Agent

First-time exhibitor Yu'ai Weiwu, an AI-native company, built an open 'AI Learning Lab' at the Expo venue. Its 'Ai Xue Pa Jiang Agent' transforms photo-based question answering from direct answers to guided reasoning, using follow-up questions, error correction, and step-down prompts for process-oriented learning. In English writing classes, the AI teacher 'Xinyu' dynamically adjusts blackboard content and pace based on student feedback, handling unexpected responses like 'I don't remember' or 'I used the wrong word on purpose'. The self-developed Ai Xue large model serves as the teaching decision hub, using a dual evaluation system to ensure instruction following and answer quality.

WeChat Mini Program Development Competition

WAIC officially announced the 2026 WeChat Mini Program Development Competition, themed 'Symbiosis with AI', open to global developers across five competition zones, evaluated on product practicality, innovation, user experience, and completeness.

Industry Ecosystem and Future Outlook

International Embodied Intelligence Skills Competition Launched

During WAIC, it was announced that the 2026 Global Developers Pioneer Summit (GDPS) and International Embodied Intelligence Skills Competition will be held on October 23-25 in Zhangjiang, Shanghai, focusing on nine key scenarios including manufacturing, warehousing and logistics, and healthcare. The competition adopts a 'challenge-based' mechanism where scenario owners set tasks and global teams compete, with a million-yuan prize pool and tens of millions in investment matching. For the first time, an international competition zone at the Oriental Hub will be established, with facilitated entry policies for overseas participants.

Computing Power Competition Logic Shifts

Multiple industry observers noted that competition among domestic AI chips has shifted from single-chip parameters to system capabilities—including software stacks, cluster scheduling, model adaptation, and industry cases. The inertia of the CUDA ecosystem remains the biggest barrier, and domestic computing power needs to reduce migration costs through open architectures and unified software stacks (e.g., FlagOS).

Conclusion

WAIC 2026 showcased a paradigm shift in AI from 'answering' to 'accomplishing'. Embodied intelligence achieved breakthroughs in long-horizon tasks, multi-robot collaboration, and fine manipulation; edge computing lowered the barrier for agent deployment through architectural innovation and open ecosystems; and AI applications focused more on real-world interaction and experience. Concurrent closed-door exchanges, developer events, and competitions further promoted industry collaboration and ecosystem building.