In 2026, WAIC will bring together over 1,100 companies exhibiting 4,486 products, setting a new record. The focus of the exhibition has shifted from large models to physical AI, with a surge of new innovations such as Unitree’s manned mecha GD01, SenseTime’s convenience store robot “Xiao Mai,” and the rapidly evolving soccer robot. This growth is underpinned by the simultaneous maturation of foundational models, embodied intelligence models, and computing infrastructure: Moonshot’s Kimi K3 has surpassed 3 trillion parameters, and China’s domestic AI computing clusters with tens of thousands of GPUs have been deployed. The industry is exploring dual pathways—VLA end-to-end and world models—and China, with its complete manufacturing ecosystem and rich application scenarios, is poised to define global standards for physical AI deployment.Article author and source: Singularity Research Society
Each WAIC is like a rehearsal for the future of AI.
In recent years, the stars of this event were large models, multimodal systems, and agents; this year, AI is stepping off the screen and into the real world.
SenseTime set up a future laboratory on-site, where the robot "Xiao Mai" autonomously handled picking, preparation, and delivery—with future capabilities to automatically restock and organize shelves. He Shizhihang brought a robotic circular wiring assembly line to the booth, demonstrating autonomous production line operations by a robot swarm. Accelerated Evolution transformed half the booth into a mini soccer field, where robots under one meter tall dribbled, tackled, shot goals, and even performed feints to outmaneuver defenders.
These robots, with their diverse forms and high intelligence, have entered consumer scenarios and production lines, becoming partners in our daily lives.
The realization of numerous concrete products has brought unprecedented attention to this year's WAIC. A total of over 1,100 companies participated in the exhibition, showcasing nearly 4,486 products, setting new records for both exhibition scale and foot traffic.
However, behind the buzz, there are many important industry questions worth exploring.
What is driving AI’s rapid integration into the real world? Where does physical AI obtain its data? Between world models and VLA, which is the key to breaking through physical AI’s bottlenecks? Smartphones, smart glasses, earbuds, robots— which will be the ultimate gateway for physical AI? As AI penetrates every industry, will China define the global standard for physical AI deployment?
The 2026 WAIC has concluded; I tried to find, amidst the noise, the answers that might be written into the future.
01
Spotlighting the hottest new AI innovations at WAIC
The on-site traffic at this year's WAIC was overwhelmingly dominated by AI robots that have been comprehensively upgraded and iterated.
A surge of innovative robot products with practical functionalities has debuted, challenging the public’s traditional perceptions of robots.
The GD01 manned transforming robot on the Unitree booth was the star of the show. This brand-new product, launched in May this year, stands approximately 3 meters tall and weighs 500 kilograms when manned. It can seamlessly switch between bipedal and quadrupedal walking modes, with a unit price of 3.9 million yuan, designed primarily for high-risk operational scenarios.
GD01 Manned Transformable Mecha
Beyond Unitree, several other reconfigurable robots also stand out. The Upwind Qiyuan T1 supports seamless switching between wheeled and quadruped modes, earning playful online comments calling it "human-like but dog-like." The ZhiJi Power TRON2 can be assembled like LEGO blocks into various configurations—bipedal, dual wheeled-leg, dual-arm, or even a "centaur" form—and can carry 30 kg while dragging a tire across a carpet.
It Shi Zhihang Robotics features the most prominent "industrial vibe," directly bringing a 1:1 automotive wiring harness production line to its booth, where a cluster of robots with "wheeled chassis and dual arms" demonstrates real-time production line operations. Lujue’s Kuafu robot, paired with Ant Lingbo, focuses on logistics and warehousing scenarios, efficiently performing tasks such as depalletizing, transporting, and feeding materials.
From industrial production and logistics warehousing to everyday consumer services, the application scenarios of AI robots continue to expand.
At the SenseTime booth, an immersive robotic convenience store called "Shaomai Buy" was set up, where the robot "Xiao Mai," wearing a baseball cap, scans codes to place orders, retrieves items, delivers them, and organizes inventory—all without any human intervention.
Bot "Xiao Mai"
According to staff from Da Xiao Robot (spun off from SenseTime), the convenience stores have already opened 10 locations in Shanghai and 20 nationwide, achieving an average of 400 orders per day in real stores, with plans to open 100 stores by the end of the year.
In addition to large industrial and commercial robots, numerous lightweight, highly interactive biomimetic robots have won over countless fans. Songyan Power’s Xiao Bumi, standing under one meter tall, is a true social butterfly—expertly capable of autonomous walking, precise handshakes, following interactions, and rhythmic dancing, all at a price under ten thousand yuan, making it incredibly accessible. The biomimetic robot Xiao Yue focuses on emotional interaction, responding in real time to audience movements through changes in eye expression and facial gestures, giving AI robots a sense of emotional awareness.
Xiao Bumi
The evolved soccer robot can autonomously move into position, pass, shoot, and perform feints; Zhi Ji Power’s full-sized humanoid robots, Oli and Luna, are designed for general-purpose tasks and enhanced interaction, respectively.
The new products from AgiBot, including the Expedition A3 Ultra,精灵G2 Max, and the dexterous hand, were also unveiled together. The精灵G2 Max, designed for "safety-certified heavy-duty" applications, directly addresses the labor gap in warehouses operating on two-shift schedules, and is already in operation at JD Logistics warehouses.
At this year's WAIC, the new AI species go beyond diverse and powerful robots, featuring a wide array of newly upgraded consumer-grade AI hardware making their debut.
AI phones have become a new hot topic in consumer electronics. Honor’s Robot Phone, with its unique design, has emerged as the star device of this year’s exhibition, featuring an innovative rotating mechanical gimbal that gives traditional smartphones a dynamic mechanical interaction form. The device is now open for pre-order across all channels and is expected to be officially launched in August.
Robot Phone
Nubia, in collaboration with ByteDance, unveiled the second-generation DouBao phone at WAIC. The device comes with the DouBao Mobile Assistant pre-installed, transitioning from a GUI to an MCP interaction framework. Functionality is only accessible after corresponding apps explicitly grant data and control permissions, with all core processing performed on-device to ensure data never leaves the device—delivering intelligent experiences while prioritizing privacy and security.
Nubia NaviX Ultra
The first AI terminal device from Jieyue Xingchen’s STEPX was also prominently unveiled, deeply integrated with the Alipay ecosystem. Staff demonstrated the “one-line query for services” feature on-site, such as locating the nearest charging station, with plans to expand into various daily life service scenarios, making AI an efficient and convenient life assistant.
StepAI Phone
In addition to showcasing star Agent products such as Wenku and Netdisk, Baidu also unveiled its AI hardware ecosystem, among which the home monitoring AI camera caught my interest.
The Xiao Du brand lead told us that this AI camera has been on the market for over a year, powered by Baidu’s proprietary “Care Agent” as its core, which dynamically invokes ERNIE Bot or other third-party models based on scenario needs. It supports setting different monitoring scenarios for children, elderly individuals, pets, or custom roles, with a minimum price of just over 300 yuan. Currently, all major AI care features—such as facial tracking and anomaly behavior detection—are offered free of charge.
The model with a screen supports two-way WeChat video calls and features design elements such as one-touch emergency求助 and live human assistance on its senior-friendly health display. This AI camera, which is both attractive and practical, makes me feel this is what AI hardware should truly be like: affordable, useful, and无需学习.
In addition, the event showcased Alibaba’s first AI clip-on earphones and AutoNavi’s intelligent guided tour robot, offering innovative solutions for campus and commercial district navigation; NetEase Youdao’s OpenPods earphones support real-time bilingual simultaneous interpretation, making them suitable for office and learning scenarios; and updated products such as iFlytek’s AI glasses were also on display.
Gaode Intelligent Guiding Robot
A review of the hardware iteration trends at this year’s WAIC reveals that while the industry focused on entry-level AI products capable of “conversing and interacting” over the past two years, this year has seen a full-scale surge in practical AI hardware designed to “deliver real-world results, get things done, and boost efficiency,” moving toward tangible productivity and life services.
02
Why are the "AI New Species" experiencing a collective surge?
The simultaneous maturation of foundational models, embodied intelligence models, and computing infrastructure is key to supporting the explosion of these new AI entities.
In recent years, the AI industry has focused more on increasing model parameter scales. This year, WAIC has shown a shift: large models are moving beyond merely pursuing "larger" sizes and entering a deeper competitive phase centered on "stronger capabilities, deeper applications, and higher efficiency."
On the eve of WAIC's opening, Moonshot AI released Kimi K3, becoming the world's first open model with 2.8 trillion total parameters and a 1 million-token context window, further exploring the boundaries of larger models in complex reasoning and long-text understanding tasks.
MiniMax prominently showcased its flagship model, M3, and teased its next-generation multimodal generation model, H3, at its booth. Through real-world ecosystem examples such as robotic dogs, AI glasses, smart headphones, and AI toys, it demonstrated M3’s comprehensive capabilities in long-context processing, code generation, and agent-based tasks.
MiniMax Booth
Mianbi Intelligence has bet on the edge side, with its new MiniCPM5-1B maintaining leadership among models under 2B parameters, and has partnered with Samsung to explore the deployment of large models on edge devices such as smartphones. Parameters remain important, but what determines industrial value is the model’s ability to enter specific scenarios and solve real problems.
Mianbi Intelligence Booth
SenseNova U1 Pro, the latest flagship multimodal model released by SenseTime, is focused on complex task闭环 and emphasizes "production-ready" capabilities. The model supports 8K ultra-high-definition output and can understand, plan, execute, inspect, and correct around a target, advancing multimodal AI from single-content generation to system-level delivery for specific tasks.
While general large models address how AI understands information in the human world, embodied intelligence models address how AI understands and interacts with the physical world.
At this year's WAIC, a large number of robots are now able to operate in industrial, retail, and logistics scenarios, thanks to the rapid advancement of embodied intelligence models. In the past, robots typically consisted of a machine paired with a dedicated algorithm, requiring separate development for each robot form factor and application scenario.
However, as the number of robots increases, this model becomes increasingly difficult to scale. The future direction is to enable a general intelligent system to adapt to different physical forms.
Ant Lingbo’s “one brain, multiple machines” is a prime example, with its LingBot Full-Stack Brain 2.0 compatible with over 20 robot configurations from 17 manufacturers, supporting everything from bipedal and wheeled robots to single- and dual-arm robots.
Ant Spirit Wave Robot
SenseMart OS by SenseTime explores enabling wheeled robotic arms, grippers, dexterous hands, and humanoid robots to share a unified retail task system.
Both were developed in response to real-world demands, reverse-engineering a universal intelligent system capable of coordinating diverse robots. As robots truly enter factories, warehouses, and stores, the era of “one robot, one model” will inevitably come to an end. The core of future competition will not merely be the robots themselves, but whether one can possess an intelligent system capable of connecting multiple robots.
In addition to embodied models, world models have also been an unavoidable keyword at this year’s WAIC. The biggest challenge for robots lies in making consistent, accurate decisions in dynamic environments. Picking up a cup in a fixed position isn’t difficult—but what if the cup is moved? What if the table tilts? What if someone suddenly walks by? This requires robots not only to know “how” to act, but also to understand “why” they are acting that way.
The value of a world model lies in helping AI develop the ability to predict the operating principles of the real world. AgiRobot has launched the embodied foundation model GO-2 and the world model GE-2. GO-2 connects task planning with action execution through action chain-of-thought and an asynchronous dual-system architecture, while GE-2 aims to create a closed loop from "world simulation" to "world learning," enabling robots—not the world model itself—to be trained and iterated within a simulation system that more closely resembles reality.
Da Xiao Robotics has released the Kaiwu World Model 3.1, which continues the integrated "Understanding—Generation—Prediction" architecture, deeply combining world cognition, scenario generation, and action decision-making; meanwhile, Shi Zhihang has introduced AWE 3.5, an embodied native base model "grown" on factory production lines, leveraging over a million hours of human-centric real-world data to enable robots to move from "being able to work" to "working well."
The model gave the robot a "brain," but no matter how intelligent that brain is, without computing power as its "backbone," it can only be a castle in the air.
One of the most noticeable changes at this year’s WAIC is that computing power has moved from behind the scenes to the forefront. The Zhangjiang Science Hall has for the first time established a dedicated 10,000-square-meter exhibition space for integrated chip and computing technologies—the first-ever standalone exhibition hall for computing power in WAIC’s history.
Over a hundred enterprises and more than 200 exhibits, including 67 products making their domestic debut, were showcased, with leading Chinese AI chip companies such as Huawei Ascend, Suanguan, TianShu Intelligent Chip, Birun, Moxi, Kunlun Chip, and Moore Threads all making an appearance.
The unit of measurement for computing power competition has shifted from individual chip performance to super nodes and ten-thousand-chip clusters, and from parameters to actual deployment costs. Huawei's Atlas 950 SuperPoD super node has completed its global debut, supporting up to 8,192 chips connected at high speed within a single super node, with potential for further expansion to a 500,000-chip cluster.
Huawei Atlas 950 SuperPoD Super Node
Alibaba's Pingtouge showcases the Zhenwu M890 chip and Panjiu AL128 super node server, integrating 128 chips per cabinet, delivering three times the performance of the previous generation, and currently supporting models such as Qwen, DeepSeek, and Kimi.
Sunway's scaleX ten-thousand-GPU computing cluster has been deployed and is in operation; ZTE, in collaboration with multiple domestic chip manufacturers, has launched the OEX open super-node; Dongfang Suanchip has released the high-compute DF1000 chip and demonstrated a 512-GPU cluster solution; Moore Threads has trained a 236-billion-parameter MoE model using its self-developed Kuai'e intelligent computing cluster and supported world model training. Domestic computing power has entered a new phase of "whether it can support industrial-scale applications."
Sugon scaleX Ten-Thousand-GPU Computing Cluster
Looking back at the captivating AI innovations at WAIC: humanoid robots priced under ten thousand yuan, AI cameras under a thousand yuan, AI phones ready for mass production... Their ability to offer affordable pricing and boldly scale up is made possible by the maturity of underlying foundation models and domestic computing power, ensuring stable supply.
After all, the more solid the underlying foundation, the greater the room for innovation at the application layer; the deeper the roots of the industrial chain, the richer the product forms at the front end.
03
The Endgame of Physical AI: Routes, Entry Points, and the Struggle for话语权
When models, computing power, and hardware simultaneously reach their tipping point for real-world deployment, the industry begins to confront more fundamental questions: Where is the ultimate destination of physical AI?
We didn’t find the final answer at WAIC, but every robot moving into real-world scenarios, every divergence and convergence in technological pathways, and every emerging foundation of the industry are collectively shaping the future of physical AI.
Millimeter-level errors on production lines, sudden operational conditions in warehousing, complex crowds in retail environments, non-standard demands in home settings... these "real-world noises" that simulation systems cannot fully replicate are the essential fuel for the iteration of embodied intelligence.
This also explains why manufacturers are racing to deploy their products in real-world scenarios: Shi Zhihang’s move to bring production lines onto exhibition floors is backed by real deployment data from automotive factories that feed back into model refinement; SenseTime’s robotic convenience store achieves 400 daily orders by refining its scheduling system and operational precision using actual order data; Ant’s Lingbo “one brain, multiple robots” can adapt to over 20 configurations, thanks to a core foundation of 60,000 hours of pre-training data gathered from real-world scenarios.
It’s not about creating a perfect model first and then finding scenarios; rather, it’s the vibrant data generated from real-world scenarios that continuously trains the model to become more powerful. Within this logic, the breadth and depth of scenario coverage directly determine the speed of the data loop, and ultimately, the rate at which AI evolves in the physical world.
China’s market, with its comprehensive manufacturing system and rich service industry scenarios, is naturally positioned at the forefront of this data competition.
However, this does not mean that real-world data will replace simulation. The maturation of physical AI requires a combination of “simulation-based pre-training” and “real-world fine-tuning.” Answers in the digital world can be infinitely reproduced, but the truths of the physical world can only be repeatedly verified through real-world interactions. This is the most rigorous barrier to physical AI—and also its most fundamental allure.
Data determines the soil for growth, while the roadmap determines the ceiling of growth. Over the past two years, the industry has continuously debated the core roadmap breakthrough for physical AI: should we pursue the end-to-end VLA approach, or bet on the foundational cognitive roadmap of world models? This roadmap debate remains unresolved at the WAIC conference.
Currently, robots and intelligent hardware that can be deployed and generate commercial value are almost all built on the VLA approach. Galaxy General’s pharmacy robot directly maps visual perception of shelves to nighttime picking actions, eliminating the need for manual inventory checks. Star Epoch’s STAR1 humanoid robot integrates whole-body coordination control and visual navigation into an end-to-end model, enabling autonomous material handling and palletizing in warehouse environments. Together, they demonstrate that VLA’s core value lies in shortening the link between “seeing” and “doing” to the minimum, allowing robots to start working and generating revenue right away.
Meanwhile, leading players such as Zhìyuán, Dàxiǎo, and Tāshí Zhìháng are simultaneously developing world models to build foundational technological capabilities for the next generation of robots. Zhìyuán positions its world model as the underlying engine for physical common sense, enabling robots to first understand gravity and friction before learning to manipulate objects. Dàxiǎo is constructing an industrial world model, using digital twins of production lines to simulate fault recovery and dynamic scheduling. Tāshí Zhìháng asserts that the world model is the sole stepping stone from specialized robots to general-purpose robots.
In the short term, VLA is the optimal solution for implementation; in the long term, world models are an inevitable path to intelligent advancement. The two are not mutually exclusive, but rather dual tracks of industrial evolution.
The technological roadmap determines the height of capability, while the entry form defines the boundaries of commerce. Undoubtedly, smartphones are currently the most widely adopted and ecosystem-matured portable intelligent nodes; the simultaneous entry of Honor, Nubia, and Jieyue Xingchen confirms that smartphones remain the core entry point for AI to reach end consumers. Robots are the terminals with the strongest physical-world mobility, serving as the primary vehicles for extending human physical capabilities and executing on-site tasks. AI cameras, smart headphones, and intelligent learning devices anchor vertical scenarios such as home, office, and education, becoming specialized intelligent touchpoints for these niche contexts.
It is unlikely that a single entry point will dominate all scenarios in the future; intelligence will permeate various devices like water, adapting on demand.
As the technical roadmap and entry formats become clearer, the deeper competition lies in the struggle for industrial influence. In the past competition around general-purpose large models, global industry technical standards and evaluation systems were largely dominated by overseas companies. But with the advent of the physical AI era, China’s industrial advantages are beginning to emerge.
We have the world’s most comprehensive hardware supply chain, enabling us to bring humanoid robot prices below ten thousand yuan. We offer the richest array of real-world applications, providing ample testing grounds for the technology. Additionally, we have progressively mature domestic computing power and foundational models that underpin the entire industry’s autonomous iteration.
Few markets worldwide can simultaneously combine the three advantages of manufacturing, application scenarios, and computing power, and even fewer can enable AI to integrate so rapidly into real-world business cycles.
When China's robots are deployed first across countless production lines and ordinary households, we will naturally develop our own standards for implementation, efficiency, and safety, thereby defining the global paradigm for physical AI deployment.
Standing inside the WAIC venue, robots and intelligent hardware are everywhere, easily giving the impression that “the future has arrived.” But beneath the bustling surface, we are witnessing a pivotal shift in the AI industry—from an information revolution in the digital world to a physical revolution in the real world.
WAIC acts like a prism, revealing a comprehensive snapshot of China’s current AI industry: models are rapidly advancing from linguistic intelligence to spatial and embodied intelligence, while underlying computing power is building a domestic closed loop from cloud superclusters to edge-side chips. There is no single ultimate answer, but everyone is steadily moving toward their goal.
The curtain has just risen on physical AI, and China is destined to be among those standing center stage, writing the rules.