Summary of Key Highlights
This year’s WAIC (World Artificial Intelligence Conference) was the largest in its history, attracting 300,000 visitors in person, over 1,100 companies, and more than 3,000 exhibits. The focus shifted from last year’s competition in large model parameters to three main areas: the clustering of computing power, the commercialization of large models, and the practical application of embodied intelligence. Domestic computing power manufacturers made a collective appearance, evolving from competition based on individual chips to coordinated systems using “super nodes.” Large models are no longer just theoretical concepts; they have been integrated into real products such as smartphones and office software. Robots have moved away from flashy demonstrations to practical applications that can perform tasks. At the same time, capital investment in the industry has been uneven, with core technologies like dexterous hands and algorithm foundations becoming the focus of attention, indicating that the AI industry is transitioning from technical demonstrations to large-scale commercialization.
1. Domestic Computing Power: Collaborating for Greater Efficiency
In the past, domestic computing power was measured by the performance of individual chips. This year, the focus shifted to “super nodes”—systems that connect hundreds or even thousands of chips together to form a “super brain” capable of handling the massive computational demands required for training large models.
- Huawei Ascend 950 Super Node: Comprising 16 cabinets and 1,024 chips, it achieved an incredible computing power of 10^18 operations per second, winning the conference’s highest award, the SAIL Award.
- Sunway 8000: A domestically produced cluster of 100,000 chips, with a density 20 times that of conventional nodes. Visitors had to wait in long lines to see it due to the high demand.
- Muxi S600: A single cabinet equipped with 64 GPUs, featuring a “wireless” design that enhances chip communication and can handle models with trillions of parameters.
- ZTE Matrix Super Node: Compatible with various domestic GPUs, it collaborates with companies like Wallrun and Suiyuan to promote an open ecosystem.
These super nodes represent a coordinated effort among multiple components—chips, entire systems, and interconnectivity architectures—to improve the efficiency and cost-effectiveness of large model training. However, they are quite expensive (up to 20 million yuan per unit) and are mainly targeted at large corporations and state-owned enterprises, out of reach for startups.
2. Large Models: From Showoff to Practical Products
Last year, companies competed based on the number of parameters in their models and their rankings. This year, the focus was on actual products that could function effectively.
- Kimi K3 Model: The world’s first open-source model with 3 trillion parameters, which received praise from Elon Musk for being “impressive.” Super node manufacturers are highly interested in models with large parameter counts as they require more computing power.
- AI-powered Phones: ZTE Nubia NaviX Ultra (the world’s first AI-powered phone) and Honor Robot Phone feature four-degree-of-freedom motion control systems. All three were selected as “highlight exhibits.”
- Office Intelligence: Tencent WorkBuddy is the most popular office intelligence platform in China, with a dedicated app. Jinshan Office also released versions of their products that emphasize the importance of data integration and cost-effectiveness.
Large models are no longer just chat assistants; they have become intelligent partners within phones and efficiency-enhancing tools in office software, directly solving users’ practical problems.
3. Robots: From Showy Demos to Practical Applications
Last year’s robot exhibits featured impressive dance and acrobatic performances. This year, the focus was on robots that can perform real tasks.
- Yushu GD01: A 3-meter-tall humanoid robot that can transform from a human form to a quadruped in 2.8 seconds, attracting crowds to watch its transformations.
- Yundeshu DR02: An industry-grade humanoid robot capable of military-style punches and acrobatics, already in use at substations for practical applications.
- Jieyue Xingchen + Yuanli Lingji: Six robots collaborated to build a Great Wall model using 80,000 parts in 15 hours; the smallest parts were less than 1 centimeter in size. This achievement relies on a combination of general-purpose and specialized models.
- Tencent’s “Xiao Liu” Massage Robot: Trained by traditional Chinese medicine experts, it can provide therapeutic massages. The Fourier GR-3 robot can deliver water automatically without manual path planning.
The upstream supply chain has also seen growth: companies like Yimike Technology (dexterous hand sensors) have received funding, and Songying Technology released a physical AI operating system for multi-robot collaboration. JD.com made the first first-person perspective dataset available to help robots learn from human actions. The industry consensus is that only robots that can generate stable profits are considered successful.
4. Capital Trends: Uneven Investment in Core Technologies
The presence or absence of companies at the conference reflected different investment strategies.
- Absent Companies: DeepSeek, which hasn’t attended for two years, recently raised $71 billion before the conference; Zhipu didn’t participate due to a “quiet period” related to its listing on the STAR Market.
- Returning Companies: ZeroOneWanwu and Baichuan Intelligence returned, with Li Kaifu personally promoting their products.
Investment in the AI industry is more targeted now, focusing on core technologies such as dexterous hands and algorithms. For example, Jijia Shijie’s home robots received orders for hundreds of units and will be launched in the third quarter. Songying Technology’s operating system, compatible with domestic GPUs, is a focal point for investors.
5. The Craze of WAIC: An Unprecedented Event
The event was an absolute success:
- Record-Scale: The exhibition covered 100,000 square meters, with participation from 1,100 companies and the global debut of 300 products.
- Huge Audience Interest: Traffic was congestion one hour before the opening; tickets were sold for thousands of yuan by scalpers, and hotels in the area were booked months in advance. “WAIC” became a trending phrase on WeChat.
- Clear Industry Focus: Hall H1 showcased large model applications, Hall H2 featured domestic computing power solutions, Hall H3 highlighted embodied intelligence, Hall H4 focused on startups (over half of the founders are from the post-90s generation), and Hall Zhangjiang specialized in computing power technologies, while Hall Xuhui displayed AI endpoints.
WAIC serves as a mirror reflecting the real state of the AI industry: it has moved from conceptual hype to practical applications, from individual efforts to collaborative development. Domestic AI is entering a more pragmatic phase.
(The entire analysis is written in plain language to make the latest developments in the AI industry accessible to everyone.)