WAIC 2026: AI Stops Showing Off and Starts Getting to Work

The Vibe Has Shifted

If you have been following WAIC since 2024, the first thing you would notice walking into this year's exhibition hall is: the vibe has completely changed.

Two years ago, booths were surrounded by crowds holding up phones, watching robots do backflips, dance in formation, or compose poetry with large language models. Applause and gasps everywhere—but when the conversation turned to real-world deployment, an awkward silence settled in.

WAIC 2026 took place in Shanghai from July 17 to 20. The numbers are still staggering: 100,000 square meters of exhibition space, over 1,100 exhibitors, more than 4,000 exhibits, and over 400 global debuts. But what stood out was not the scale—it was the temperament.

Robots were no longer performing. They were assembling wiring harnesses on mock production lines set up right on the show floor. Visitors stopped asking "What can it do?" and started asking "How much?" and "When can I get one?" LLM vendors stopped flexing parameter counts and instead showed how AI agents could help you make a slide deck or fill out an expense report.

After three days walking the halls, here are the four clearest industry trends I observed. They all point to one thing: AI has finally stopped being a spectator and started being a doer.

Trend 1: Computing Power—From Single Chips to SuperPoDs

Step into Hall H2, the Technology and Innovation Hall, and you are hit by rows of massive computing clusters.

In previous years, the computing power section of WAIC was all about single-chip specs: TFLOPS, memory bandwidth, paper metrics. This year, the winds have shifted entirely: every major vendor shifted their showcase focus from raw chip performance to full-system delivery capability.

Huawei publicly debuted the industry's largest SuperPoD system in its physical form—the Ascend 950 (Atlas 950 SuperPoD). This black monolith supports 1,024 compute cards interconnected at high speed, working "like a single computer." When training trillion-parameter models, the system eliminates the need for repeated data shuttling; the entire cluster operates as one coherent machine.

Sugon unveiled "Sugon 8000," China's first fully homegrown 100,000-card AI supercluster, achieving full-stack self-development across chips, compute, storage, networking, cooling, applications, and services—earning it a "Crown Jewel of the Expo" designation. ZTE's OEX SuperPoD made its public debut with an orthogonal backplane-free design that eliminates cabling entirely, cutting interconnect costs by 80%. Alibaba Cloud, MetaX, Moore Threads, Biren Technology, and Baidu AI Cloud all showcased their own SuperPoD solutions.

The Peng Cheng Laboratory, together with the Global Computing Consortium, released the SuperPoD Definition and Practice White Paper on-site, offering an authoritative definition: a SuperPoD is a computing system composed of multiple compute nodes tightly connected through efficient interconnect protocols, featuring unified memory addressing across physical nodes, and logically behaving as "a single computer."

In plain terms: previously, chips collaborating was like people shouting to each other across walls. A SuperPoD tears down those walls and builds a highway instead.

The driving logic is clear. Global AI computing has hit multiple bottlenecks: the inherent compute-memory separation in the von Neumann architecture causes data movement overhead; exponentially growing compute demand collides with physical constraints of power, cooling, and data center space; and regional compute centers lack unified scheduling standards. No matter how powerful a single chip is, it cannot solve these problems alone. The computing race has entered the "system-level" era.

One notable signal: DeepSeek disclosed in a technical report that as the Ascend 950 SuperPoD enters mass production, the price of the DeepSeek-V4-Pro model will drop significantly. This means the scaled deployment of SuperPoDs is already driving down the cost of domestic LLM services from the ground up.

Trend 2: Robots—Enough with the Show. Time to Work.

WAIC 2026 dedicated an entire hall—H3—to embodied intelligence, hosting over 200 companies, more than double last year's count.

The biggest change? Robots have stopped being performers and become workers.

Tasbot recreated a 1:1 automotive wiring harness assembly line right on the show floor. Robots continuously picked, routed, and inserted wires—harnesses are soft and deformable, and assembly precision must reach sub-millimeter levels. Magic Atom's wheeled humanoid robot MagicBot D1 navigated between mock shelves, identified materials at varying heights, grasped them precisely, and transferred them to target locations. A few steps away, a robotic arm demonstrated box-folding and glue-sealing, re-planning its path in milliseconds upon encountering deviation.

Home scenarios were equally lively. Qiyuan Robotics showcased the Qiyuan T1, the world's first transformable personal robot—autonomously switching between wheeled bipedal and quadrupedal forms: a home assistant indoors, a robotic dog for tracking shots outdoors. Some humanoid robot arms have begun understanding vague commands like "tidy up the living room" and autonomously breaking them down into multi-step execution plans.

Audience reactions shifted too. Less applause, more price inquiries. Fewer gawkers, more professionals huddled around engineers discussing deployment details.

A line from Professor Su Hao of Fudan University during the keynote stuck with me: "Today's AI can write poetry, write code, make PowerPoints—but it cannot help an elderly person turn over in bed. The value of intelligence is still mostly trapped in the digital world, while humanity's heaviest needs all reside in the physical, atomic world."

Of course, the gap between a demo and a product remains a long engineering road. In real environments, lighting changes, object deformation, human intrusion, and equipment aging all introduce interference. Making a robot complete one motion is easy; making it do so continuously, reliably, in a real setting is the hard part. Industry experts generally project two to five years—the second half of the embodied intelligence race will be defined by "high heat, high investment, high trial-and-error, and high uncertainty."

Trend 3: Models—From Content Generation to Physical Reasoning

In previous years, the LLM track at WAIC was all about text, image, and video generation. This year, the spotlight shifted to a much harder concept: world models.

"World models are the core engine driving physical AI development, and their core capability is prediction and reasoning," said Xu Kai, researcher at the CAS Institute of Industrial AI.

Why the sudden explosion of interest? Because embodied AI models hit a wall. When a digital world model makes a mistake, you get a weird-looking image. When a physical agent makes a prediction error, you get irreversible real-world damage. The industry urgently needs a "physical brain" that can minimize a robot's trial-and-error cost in complex, uncertain environments.

The Kairos 3.1 world model, unveiled at a forum during the conference, leverages a hybrid Transformer architecture to compress visual observations, language instructions, force-tactile states, and policy trajectories into a unified latent space—enabling robots not only to replicate pre-programmed actions but to predict the cascading physical consequences of their operations. Shengshu Technology's MotuBrain, meanwhile, focuses on dynamic physical motion cognition, capturing human joint movements and force changes in real time.

Yao Maoqing, partner and SVP of AGIBOT, was candid during a panel: "World models are hot this year, but most are still stuck at the visual generation level. Data type mismatch and data scarcity are the two major bottlenecks."

Academician Zhou Zhihua of Nanjing University proposed a pragmatic path forward: first build a reasonably reliable world model with a small amount of high-quality data, then use interaction with that world model as a substitute for real-world interaction to form decisions, then test those decisions in real scenarios—dramatically compressing trial-and-error costs while iteratively improving the world model.

Trend 4: Agents—From Solo Q&A to Team Collaboration

WAIC 2026 named 10 "Crown Jewels of the Expo," and Baidu's general-purpose AI agent "Baidu Dazi" was among them. A reporter told it, "Compile this week's major AI industry news in China, categorize it, and make a briefing." The agent searched, filtered, organized, and formatted—end to end, hands-free.

Kingsoft Office's new Lingxi Pro and WPS Comate no longer work in the "you write a sentence, it completes a sentence" mode. You give it a task—it handles layout, tables, and sourcing on its own. Suzhou-based PatSnap took a specialized route, building agents that help R&D teams search patents and analyze competitor technology, compressing what used to take days into hours.

Huang Wei, founder of Unisound, used a vivid analogy: the LLM is AI's "brain," and the agent gives that brain "hands and feet"—it plans steps, uses tools, interacts with the outside world, and double-checks its own work afterward.

Yu Xiaohui, President of CAICT, summarized three shifts: the competition has moved from individual agent intelligence to whole-system orchestration; agents no longer freeze after training but evolve through use; and instead of working alone, multiple agents will form collaborative teams.

Even the way visitors ask questions has changed. Last year, people asked "Can you write a poem?" "Can you translate this?"—testing knowledge. This year, they open with "Can this help me make a PPT?" "Can it auto-fill my expense report?"—caring only about saving their own effort.

But hype aside, there is plenty of hard work left. IDC predicts rapid growth in global "digital employees," but for agents to truly enter daily life, engineers still need to solve: insufficient real-world task data, lack of cross-task generalization, and missing real-time feedback loops. On the security front, the more autonomous an agent becomes and the more tools it can invoke, the more complex the risks—what if it sends messages on its own? What if multi-agent collaboration goes haywire?

Final Thoughts

After three days, my biggest takeaway is this: the AI industry is going through a "disenchantment."

Nobody is asking grand questions like "Will AI replace humans?" anymore. The questions have become: Can it help me write this weekly report? Can it reduce errors on the production line by just one more unit? Can we bring the price of LLM services down a little further?

From showing off to getting things done, from flexing muscles to delivering homework—AI has been on this journey for years. WAIC 2026 may well be a milestone at the turning point.

Looking forward to seeing more things on next year's show floor that people can actually use.