WAIC has changed: stepping out of parameter competitions, AI begins to work in the mud with two feet

The artificial intelligence (AI) industry is no longer lacking in stunning technological displays, what is truly scarce are real scenarios that are willing to pay for them.

This year, the WAIC exhibition area has exceeded 100000 square meters for the first time, with over 1100 companies filling the venue to capacity. If you put together the most capable companies in the venue, you will find that the background of Chinese AI is shifting from “science fiction films” to “business operations”.

At the meeting, Huashang Taolue visited a group of representative enterprises in various tracks. They are distributed in different fields, but show a common trend: either they have successfully transformed AI into commercial value, or they are actively exploring feasible paths for AI commercialization, and engineering capabilities have become more important than ever.

From computing power bases to humanoid robots, from large models to beauty giants, the concentrated appearances of these enterprises have pieced together a real progress chart of China’s AI industrialization.

01
The Controversy of Computing Power: From “Chip Manufacturing” to “System Manufacturing”
In the past, discussions about domestic computing power often focused on whether a certain chip could compete with a certain generation of products from international giants. At this year’s exhibition, almost all computing power companies are conveying the same message: the performance comparison of a single chip has become a thing of the past, and system level efficiency is the key to the future. From a 100000 card cluster to a “computing power cube” in the office, the form of computing power is accelerating differentiation.

Zhongke Shuguang has directly launched a “treasure of the town hall” – Shuguang 8000 (Peak Climbing). This thing is China’s first nationwide AI super cluster that produces 100000 calories. What is the concept of 100000 calories? This is not just a number, but also an extremely violent system engineering. The entire system has 3 billion components that work together with precision, and the installation accuracy at the end is stuck at the millimeter level. If all its computing power is used for inference, it will be the largest token factory in China, capable of carrying 5% to 10% of the country’s token traffic.

Since advanced manufacturing processes are getting stuck, don’t be obsessed with single card performance. Zhongke Shuguang connects thousands of cards and stacks computing power using its self-developed scaleFabric non-destructive interconnect technology and phase-change immersion liquid cooling. This is a typical “miracle with great force” style overtaking on a bend. Now this system has been implemented in Zhengzhou. It is said that when you open an AI application on your phone, the computing power behind it may come from the data center in Zhengzhou.

As one of the representative enterprises of domestic GPUs, Muxi set up an open source zone at the exhibition and attracted the participation of the global mainstream AI open source framework PyTorch. The subtle changes in the subject object relationship reflect that Chinese chip manufacturers are undergoing a transition from “seeking ecology” to “building ecology”. The enhancement of ecological discourse power is no less significant than the pursuit of chip performance.

Super fusion has simply turned computing power into a ‘home appliance’. The TokenBox they brought ™️ The enterprise token production platform has moved supernodes from the data center to the office. This device has the performance of an intelligent computing server, liquid cooling technology to achieve library level quietness, and can support local operation of models with large parameter quantities. Its emergence fills the market gap between cloud APIs and self built data centers, accurately meeting the needs of medium-sized industries such as finance, government, and enterprises that have high requirements for data privacy.

Mole Thread founder Zhang Jianzhong presented the “Three AI Factories” framework on the WAIC sub forum – model training, word element production, and intelligent agent factory, connecting the entire process of AI from training to implementation. The Kuai’e Wanka cluster can support 236B parameter model training, and the MTT S5000 achieves the same inference performance as international mainstream products with about half of the hardware specifications. It also achieves throughput improvement of “3 S5000 plus 2 international mainstream cards equals 9” through heterogeneous inference. The MUSA ecosystem is fully integrated with mainstream frameworks such as PyTorch and Triton, and the self-developed MusaCode has surpassed Claude Code in KernelBench evaluations, with developer migration barriers approaching zero.

▲ Zhang Jianzhong, founder, chairman, and CEO of Moore Thread, delivers a speech

Lenovo is the exhibitor with the largest span of this WAIC event, from the computing power base to the FIFA World Cup. The Wanquan Heterogeneous Intelligent Computing Platform V5.0 and Wentian Super Node form a production base for word elements, with a single node capable of carrying 40 GPUs. On top of computing power, the Qingtian AI engine embeds intelligent agents into the entire process of enterprise “research, production, supply, marketing, and service”. The enterprise’s super intelligent agent “LeXiang” has a monthly activity of 7 million and cumulative sales exceeding 5 billion yuan. 42 products on the terminal side have passed the AI terminal national standard L3 level test, and Tianxi AI 4.0 has driven 80% of the word element consumption to be completed locally.

02
Second half of the big model: no rolling parameters, rolling landing
The main theme of the big model track has changed. Last year, the industry was still enthusiastic about competing in parameter scale and ranking scores; This year, almost all big model companies are exploring how to transform models into actual products and revenue. The standard of competition is shifting from ‘making it’ to ‘selling it’.

Alibaba not only showcases the technological capabilities of the Tongyi Big Model, but also the complete chain of “model+computing power+e-commerce scenarios”. The biggest challenge in monetizing large models is finding willing to pay scenarios, while Alibaba’s advantage lies in its vast e-commerce ecosystem. The Tongyi big model can be directly embedded into the operational scenarios of Taobao and Tmall, from product understanding to intelligent customer service, and then to marketing creativity, achieving seamless integration between technology and business scenarios.

Kimi, as a leading large model unicorn in terms of valuation in China, is facing the core test of commercialization. The Kimi intelligent assistant has a large number of daily active users, but how to convert its ultra long context processing capability into actual revenue is the key to its next development. The commercialization of large models is not simply about accumulating users, but requires verifying truly valuable payment scenarios.

SenseCore devices from SenseTech are undergoing a shift in computing power from training to inference. Through the comprehensive layout of Nichirin’s big models, AI infrastructure, visual AI, and embodied intelligence, SenseTime is trying to prove itself not only as a big model company, but also as a full stack AI service provider.

Face wall intelligence is making efforts in the field of end side large models. When other companies are competing to pursue billions of parameters, Face Wall Intelligent chooses to make the model smaller so that it can run locally on end devices such as mobile phones, robots, and car mounted devices. The advantages of end-to-end models in terms of cost-effectiveness and privacy protection indicate the broad prospects of the “small model era”.

Tencent’s hybrid big model has been deeply embedded in national level application ecosystems such as WeChat, Tencent Meeting, and Tencent Documents. While other big model companies are still searching for scenarios, Tencent’s scenarios have already grown on the models. In addition, Tencent’s AI programming tool CodeBuddy directly benchmarks GitHub Copilot, attempting to seize the entrance in the developer tool arena. The competition of big models is shifting from ‘whose parameters are big’ to ‘whose entry points are many’.

JD showcased the deep implementation of Yanxi’s big model in retail, logistics, and financial scenarios, but what is more noteworthy is its “world’s largest embodied intelligent data collection center” – accumulating over 10 million hours of real robot operation data. This means that JD.com is not only expanding its models, but also stockpiling data fuel for the era of robots. When most companies are still discussing where data comes from, JD’s answer is simple: from its own warehouse.

03
Embodied intelligence:
From Demo to Delivery, Robots Start Working
At this year’s WAIC event, hundreds of various robots were demonstrating synchronously, creating a lively atmosphere. But the most noteworthy thing is not how difficult the robot can perform, but who can truly achieve large-scale delivery and stable operation.

Yushu Technology’s IPO on the Science and Technology Innovation Board has been approved, with plans to raise 4.2 billion yuan, proving that the robotics industry already has enormous commercial potential. The “Robot Unmanned Factory” sand table displayed indicates that manufacturing automation is moving towards a new stage of “building robots with robots”, which is also an important indicator of industry maturity.

As a new unicorn in the field of embodied intelligence, Galaxy General Motors started with a universal humanoid robot that has both operational and mobility capabilities. Although the path is more challenging, it also means a higher ceiling for development.

Zhiyuan answered the question of “can it be made” with a set of harder numbers. As of now, Zhiyuan robots have been produced in 15000 units, with a flexible delivery capacity of over 100000 units per year. The team founded by Zhihui Jun (Peng Zhihui) follows the industrial route of “mass production first, then iteration”. In the field of embodied intelligence, which is filled with demos and concept stories, Zhiyuan is one of the few companies that can speak with “offline quantity” instead of “demonstration videos”.

Qiyuan Robot is a consumer grade embodied intelligent brand under Shangwei New Materials. It brings the world’s first deformable personal robot, Qiyuan T1, to WAIC 2026. It adopts a Transformer cross form integrated architecture and operates in a wheeled humanoid form indoors, which is quiet and flexible; When encountering complex road conditions such as grass, gravel, steps, etc., it automatically switches to quadrupedal form for stable passage. The entire form switching process is autonomously completed by environment recognition, without the need for manual intervention. T1 is also deeply linked with motion cameras, supporting voice following shooting and trajectory tracking, extending the functional boundaries of personal robots from “companionship” to “mobile image creation”. When most humanoid robot companies were still competing for industrial scenes, Qiyuan chose a differentiated path – to make robots into ordinary people’s living rooms and outdoors.

Leju Intelligence is one of the earliest companies in China to deploy open-source humanoid robots, and its Kuafu series of humanoid robots has been iterated to the fourth generation. Leju has chosen a differentiated path of “open source building ecology, ecology promoting mass production” – by using open source hardware design and motion control algorithms to attract developer community participation, and then using community ecology to support mass production. At a time when all humanoid robot companies are competing for financing and delivery, Leju attempts to answer another question with the phrase ‘Android moment of humanoid robots’: who can define industry standards.

As a leading enterprise in global warehousing and logistics robots, Jizhijia’s products have undergone long-term testing in real warehousing environments. This stable operation in real business scenarios is the most favorable proof for the commercialization of robots.

Honor’s world’s first robot phone, Robot Phone, integrates the industry’s most sophisticated four degree of freedom gimbal system into its body. Through the OpenClaw open platform, the gimbal system has evolved into an intelligent agent actuator, capable of independently completing cross application continuous tasks such as cake ordering and taxi hailing. Honor has upgraded MagicOS to the industry’s first companion based multimodal intelligent agent operating system, Agentic OS, and proposed that the evolution of the future intelligent world of all things will present a new architecture of “one master, multiple specialties, and three end collaboration”. CEO Li Jian mentioned in a conversation with Kevin Kelly that innovation will move from digital screens to the physical world. China’s engineering capabilities that integrate software and hardware, the world’s richest production and living scenarios, and the collaborative ecosystem of open collaboration with global partners will become unique environmental advantages for China’s technology industry to lead the trend in the next round of transformation.

Wuwen Zhike aims at the “data desert” of the robotics industry – embodied intelligence requires high-quality interactive data from the physical world, but has long been trapped in the “impossible triangle” of authenticity, scale, and low cost that cannot be achieved simultaneously. Its Real2Sim2Real closed-loop system does not select any side stations. Starting from real data collection, simulation generalization, and then returning to the real world for verification, it has achieved a 70% efficiency improvement and a 90% cost reduction, holding orders worth hundreds of millions of yuan. In Deqing, Zhejiang, the largest embodied intelligent data collection training ground in the Yangtze River Delta is being built, and the Boundless Platform is providing “data fuel” to partners such as Horizon, Tashi Zhihang, and Lingxin Qiaoshou.

04
AI enters various industries:
Traditional giants deeply embrace AI
The most profound industrial transformation often occurs in traditional industry giants who deeply integrate AI into their core businesses. When leading enterprises in the fields of industry, beauty, communication, etc. begin to seriously apply AI, it means that the penetration of AI has broken through the boundaries of the technology circle.

At the Hikvision exhibition area, a series of large model software and hardware products, as well as vertical large model applications, answered this question – large models not only enable machines to have the ability to intelligently perceive the physical world such as “eyes, ears, nose, tongue, and body”, but also evolve the cognitive ability to understand intentions, achieving a leap from passive response to active service: multimodal intelligent cameras can switch “work” mode with one sentence, becoming all-weather “security officers”, actively detecting anomalies and alerting; The industrial large-scale model is deeply integrated into the manufacturing front line, helping to improve the yield rate; Various intelligent agents can make complex business flows efficient and worry free with just one sentence… These all indicate that AI is accelerating its landing in the physical world and truly becoming an intelligent partner for thousands of industries.

L’Oreal, as a global beauty giant, showcased practical applications such as AI skin detection, AI assisted formula development, and generative AI marketing creativity at WAIC. This indicates that AI has evolved from a mere marketing gimmick to a substantial productivity tool in the beauty industry.

L’Oreal showcases 3CE GENBA generative beauty consultants

China Mobile has demonstrated its grand layout in the field of computing power networks. As the world’s largest mobile communication operator, China Mobile is committed to building its nationwide network and data centers as the core infrastructure for AI, driving AI from the cloud to the edge.

China Telecom, based on Tianyi Cloud, showcased the implementation of the “Xingchen” large model series in government, emergency, industrial and other scenarios. The logic of operators doing AI is different from that of Internet companies – they do not pursue the leadership of the model itself, but sell AI to existing government and enterprises as a capability increment of cloud services. This “infrastructure+AI” model, although not as sexy, has extremely strong income certainty.

China Unicom has brought the “Yuanjing” big model and industrial Internet platform, focusing on intelligent manufacturing and smart cities. The collective appearance of the three major telecom operators on WAIC sends a clear signal: the infrastructure of AI is not only the responsibility of cloud providers, but also the main battlefield for telecom operators. Whoever masters the physical nodes of the computing power network holds the distribution power of the AI era.

A clear AI industrialization map

Looking back at these exhibiting companies, we can clearly see four profiles of China’s AI industrialization:

The first layer consists of enterprises that build infrastructure, such as Zhongke Shuguang and Muxi, which are addressing the issues of computing power supply and ecological construction.

The second layer consists of “brain building” enterprises, such as Alibaba and SenseTime, which are committed to transforming the capabilities of large models into practical business scenarios.

The third layer consists of “body building” enterprises such as Yushu and Jizhijia, which are driving robots from the demonstration stage to large-scale delivery and practical applications.

The fourth layer consists of enterprises that use AI, such as Zhongkong and L’Oreal, which represent the deep transformation and empowerment of traditional industries by AI technology.

These four levels are intertwined and together form the complete cross-section of China’s AI industrialization in 2026. AI is no longer a cutting-edge technology hanging in the air, but a complete industry chain from computing power base to model capability, and then to hardware carrier and industry applications. In this process, technology is shedding its halo and truly becoming a new engine driving economic development.