中文
← Back to news
IndustryJul 20, 2026

WAIC 2026: AI Moves from Screens to the Physical World, Super Individuals and Industry Closure Become Keywords

From July 17 to 20, 2026, the World Artificial Intelligence Conference (WAIC 2026) was held simultaneously at three venues in Shanghai: Shanghai World Expo Exhibition & Convention Center, Zhangjiang, and West Bund. The exhibition covered over 100,000 square meters, with more than 1,100 enterprises participating, over 3,000 exhibits on display, and over 300 products making their global debut. The number of visitors was expected to reach 300,000. With the theme "Intelligence · Symbiosis," the conference focused on cutting-edge fields such as large models, embodied intelligence, world models, and AI terminals, and for the first time set "One-Person Company, Super Individual" as an official topic.

Embodied Intelligence: From Performance to Work, Robots Enter Real Scenarios

At this year's WAIC, the robot exhibition area featured over 200 enterprises competing, but the industry focus shifted from "backflips and dancing" to "stable work capability."

  • JD.com showcased the full JoyAI series of large models and an embodied data collection system, proposing a three-stage roadmap of "digital intelligence—attached intelligence—embodied intelligence," and open-sourced the industry's largest first-person-view dataset, EgoLive (2,000 hours, 65,866 episodes).
  • Daxiao Robotics released the Kairos 3.1 world model, which topped four global embodied intelligence benchmarks, and launched a complete set of solutions covering retail, hospitality, and inspection. Its fulfillment robot W1 is already operating in real forward warehouses.
  • Yuanli Lingji, in collaboration with Stepfun, used six robots to autonomously assemble a 3.5-meter-long Great Wall model from 80,000 building blocks in 15 hours, driven entirely by the embodied general foundation model DM0.5.
  • Qianxun Intelligence demonstrated Moz1 performing long-horizon tidying tasks in an unstructured living room, autonomously handling unexpected objects. Its Spirit v1.6 adopts an integrated VLA and world model architecture.
  • Sudo Technology made its debut, showcasing over 10 skills including robot mobile manipulation, bimanual collaboration, and precision assembly, emphasizing a training approach that prioritizes simulation supplemented by real robots.
  • RoboScience released the world's first cloud-based embodied large model, Visics, achieving "one brain controlling multiple hands" with a 30-second handover, and launched the EaaS (Embodied Intelligence as a Service) model.
  • Mech-Mind displayed its "eye-brain-hand" components. Its 3D vision system and Mech-GPT multimodal large model have been deployed on a large scale in industrial scenarios, with over 27,000 units globally.
  • Pudu Robotics released the PuduFM embodied intelligence large model, supporting a "one brain, multiple forms" strategy. Over 130,000 robots have been deployed globally, generating 36.5 million hours of navigation data annually.
  • Qizhi Openmind released a universal technical foundation for intelligent robots, including the HumanGPT world model and the Modou IDE, aiming to become the "Android" of the robotics industry.

World Models: From Generation to Action, the Underlying Infrastructure of Physical AI

World models became one of the hottest technical topics at this year's conference, with multiple companies showcasing integrated solutions spanning understanding, generation, prediction, and action.

  • SenseTime released SenseNova U1 Pro, achieving native 8K output and interleaved text-image reasoning, capable of autonomously completing delivery-level designs from sketches to finished products. Its Raccoon AI correctly predicted the dark horse Cape Verde in the World Cup.
  • Extreme Vision exhibited a product matrix of general world models, covering content creation (YiSu), autonomous driving (DriveDreamer), embodied intelligence (GigaWorld), and home services (Shiguang S1), forming a closed loop of "generation—understanding—action."
  • Tezign proposed a Subjective World Model (SWM), achieving 85% accuracy in AI Persona behavior simulation through four-layer heterogeneous data collaborative training, with over 300,000 AI Personas accumulated.
  • TeleAI released AI Flow, achieving the first direct satellite transmission of tokens from lightweight small drones, enabling voice calls at an extremely low bitrate of 0.266 Kbps, and winning the SAIL Empowerment Award.

AI Terminals and Computing Power: Agent-Native Systems and Domestic Computing Power Ecosystem

AI terminals are evolving from "OS+AI" to "agent-native," and competition in domestic computing power is shifting from single-chip parameters to system capabilities.

  • Stepfun launched the world's first large-model-native AI terminal brand STEPX, the agent-native operating system Step AOS, and the personal agent Step Amoo. The STEPX Neo phone received L3-level AI terminal national standard certification.
  • ZTE released the OEX orthogonal architecture supernode, integrating 128 GPUs per cabinet, scalable to 16,000 cards. It launched the Matrix supernode solution in collaboration with multiple domestic chip manufacturers, winning the SAIL Award for the second consecutive year.
  • Cix Technology announced the AGX Agentic Compute strategy and the desktop-level supercomputer AGX Station, capable of locally inferring 70B-150B parameter large models and supporting multi-machine cascading.
  • Tsingmicro showcased its reconfigurable computing architecture and 4K supernode solution, with interconnection costs reduced by approximately 90% compared to foreign solutions. Its RAISA software stack has been adapted to over 200 large models.

Super Individuals and Youth Power: Post-00s Geeks and AI-Native Entrepreneurs Take the Stage

For the first time, WAIC set "One-Person Company, Super Individual" as an official topic, with platforms like Xiaohongshu and Fudan F-LAB becoming incubators for new AI creators.

  • Xiaohongshu showcased RED Skill and Vibe Coding features at WAIC. Over 160,000 developers have been active on the platform in the past year, with post-00s and post-05s being the main publishers. More than 7,300 original Skills have been published, and Vibe Coding works are released at a rate of 40,000-50,000 per month.
  • Fudan F-LAB Genius Youth Science and Innovation Camp: 32 students presented at WAIC, with post-00s accounting for 72%. Projects included the embodied world model ROBOTOPIA, AIGC traceability (Moyin Qianyin), and emotional companion robot Cutawa.
  • WAIC Cloud Sail Award announced the 2026 winners. Among the winners under 35, half hold Best Paper awards at top conferences; one has nearly 90,000 citations, and another's startup achieved a first-round valuation exceeding $4.5 billion.

Key Insights and Industry Consensus

  • Richard Sutton (Turing Award winner): Large models do not possess native intelligence; AI will enter the "era of experience," and reinforcement learning is the core path to general intelligence.
  • John Hopcroft (Turing Award winner): The core competitiveness in the AI era is talent, and the fundamental mission of universities is to cultivate the next generation of talent.
  • Xu Li (CEO of SenseTime): The core value of AI is to amplify individual capabilities rather than replace human labor, and billing will shift from tokens to tasks.
  • Su Hao (Professor at Fudan University): The key to solving large model hallucinations is physical intelligence, and AI evaluation criteria will shift from impressive demonstrations to reliable operation.
  • Yin Qi (Chairman of Stepfun): The next explosion in the AI industry will come from the combination of agent models and next-generation terminal products.

Impact and Outlook

WAIC 2026 marks the systematic migration of AI from the digital world to the physical world. Embodied intelligence has entered a phase of "reliability" competition, world models have become the infrastructure of physical AI, and super individuals and AI-native terminals are reshaping the industry landscape. The industry consensus is that in the next 2-3 years, robots will experience their "ChatGPT moment," and the ability to close the loop among data, models, terminals, and scenarios will be the key to determining success or failure.

Also available in 中文.