WAIC 2026: AI Shifts from Showmanship to Pragmatism, Physical World Becomes Main Battlefield

From July 17 to 20, 2026, the World Artificial Intelligence Conference (WAIC 2026) was held concurrently across four venues in Shanghai (World Expo, Zhangjiang, West Bund). The exhibition area exceeded 100,000 square meters for the first time, attracting over 1,100 exhibitors and showcasing more than 3,000 exhibits, with over 300 products making their global debut. The conference is expected to draw 300,000 visitors. It gathered 11 Turing Award, Nobel Prize, and Fields Medal winners, along with over 1,400 guests, and hosted nearly 200 forums.

Core Trend: From Showmanship to Pragmatism, AI Accelerates into the Physical World

The most notable change at this year's WAIC is AI moving from screens to real-world scenarios. Tracks such as embodied intelligence, AI film and television, and AI terminals no longer compete on flashy demos but focus on stable delivery, reliable deployment, and commercial viability.

Embodied Intelligence: Robots Start 'Working'

Over 200 embodied intelligence companies exhibited, with robots no longer just performing dances or fights but executing real tasks in factories, supermarkets, and homes.

Yuanli Lingji, in collaboration with Stepfun, used 6 robots to autonomously assemble an 80,000-piece Lego Great Wall model in 15 hours, demonstrating long-horizon tasks, multi-robot coordination, and sub-millimeter precision. Its general-purpose foundation model DM0.5 achieved SOTA with a 43% success rate on the RoboChallenge Table30 V2 real-world benchmark.
Qianxun Intelligence showcased Moz1 autonomously organizing items in an unstructured living room, including object recognition, dynamic replanning, and long-horizon execution, using the VLA and world model integrated architecture Spirit v1.6.
Pudu Robotics released the embodied intelligence large model PuduFM, supporting 'one brain, multiple forms,' with over 130,000 robots deployed in 85 countries and regions, generating 36.5 million hours of navigation data annually.
Mech-Mind displayed 'eye-brain-hand' components, with over 27,000 products deployed globally, serving more than 100 Fortune 500 clients.
Sodu Technology made its debut, showcasing over 10 skills including mobile manipulation and bimanual coordination, using a simulation-heavy data training approach.
RoboScience released the world's first cloud-based embodied large model Visics, enabling 'one brain controlling multiple hands,' with a 30-second handover for immediate use.

AI Film and Television: From Flashy Generation to Stable Delivery

AI film and television became a new crowd magnet. Wondershare's AI short drama tool 'Wondershare Drama Factory' improved storyboard creation efficiency by 10x, accelerated single-episode production by 800%, and reduced labor costs by 72%. DataEye data shows that in June 2026, 23,746 new AI short dramas were released natively on Douyin alone, averaging 792 per day. The industry trend has shifted from 'being able to generate' to 'being able to deliver stably,' with expert knowledge (e.g., director thinking) becoming a new barrier, and going global evolving from an elective to a mandatory course.

AI Terminals: Agent Phones and Desktop Supercomputers

Stepfun launched the world's first large-model-native AI terminal, STEPX Neo, equipped with the agent operating system Step AOS and personal agent Amoo, enabling 'result interaction' rather than 'process interaction.'
Cixtech released the AGX Station desktop supercomputer, capable of locally inferencing 70B-150B parameter models, with multiple units cascading into clusters.
ZTE unveiled the OEX supernode, integrating 128 GPUs per cabinet, expandable to 16,000 cards, and jointly launched the Matrix supernode solution with multiple domestic chip manufacturers, winning the SAIL Award.

AI Infrastructure: Domestic Computing Ecosystem Accelerates

Moore Threads and Biren Technology demonstrated an IBGDA solution with domestic GPUs directly connected to domestic RDMA NICs, achieving significant improvements in small message bandwidth and nearly halving latency.
Tsingmicro showcased reconfigurable chips and 4K supernodes, with cumulative orders exceeding 40,000 computing cards, compatible with over 200 models.
China Telecom TeleAI proposed the 'AI Flow' technology, enabling lightweight drones to directly transmit tokens via satellite, achieving voice calls at an extremely low bitrate of 0.266 Kbps.

Academia and Talent: Cloud Award and Youth Power

The WAIC Cloud Award announced winners in 'Shining Stars' and 'Rising Stars' categories, all top AI talents under 35, half holding Best Paper awards at top conferences, with individual citations reaching nearly 90,000. The first cohort of 32 students from Fudan F-LAB Genius Youth Science Innovation Camp presented projects at WAIC, covering embodied world models, AIGC content tracing, emotional companion robots, and more.

Expert Views: AI Enters the Era of Experience

Turing Award winner Richard Sutton pointed out that large models lack native intelligence, merely fitting historical human data, and the core future path of AI is reinforcement learning-driven experience.
Turing Award winner John Hopcroft emphasized that talent is the core competitiveness in the AI era, and universities should focus on cultivating the next generation.
Fudan University Professor Su Hao believes that solving large model hallucinations requires physical intelligence, with AI needing to interact with the real world.
SenseTime CEO Xu Li proposed that the core value of AI is amplifying individual capabilities, and billing will shift from Token to Task.

Impact and Outlook

WAIC 2026 marks the transition of the AI industry from a 'model competition' to a 'deployment competition.' Tracks such as embodied intelligence, AI terminals, and AI film and television have all reached a turning point from demo to product. Data flywheels, real-world deployment, and hardware-software synergy have become key competitive factors. The industry consensus is that the 'ChatGPT moment' for robots still requires 100 million hours of real interaction data, but it may arrive within the next two years.