The WAIC 2026 exhibition showcased the full range of Chinese chips for artificial intelligence, including Huawei Ascend and Dongfang Suanxin. This event demonstrated the development of China's computing ecosystem, which has reached a critical mass with 108 chips, 261 models, and 208 products for AI terminals, as well as over 300 real robots across three zones spanning more than 100,000 square meters.
A dedicated section for Chinese chips was set up in the Zhangjiang zone, featuring over 40 companies involved in computing technologies, including Huawei, Muxi, Moore Threads, and Tianshu Zhixin. The infrastructure displays were dominated by Supernode racks; for instance, Huawei Atlas 950 SuperPoD demonstrated the connection of 1024 Ascend cards using full utilization of 800G optical modules and high-density optics, representing a significant improvement over last year's copper-based CloudMatrix 384.
Other participants, such as ZTE, Muxi, Biren Technology, and Moore Threads, also presented their competing Supernode solutions. Some companies, including Tsingway and Dongfang Suanxin, showcased programmable and reconfigurable chip architectures that utilize control circuits to reconfigure computing blocks to enhance efficiency without reliance on leading-edge technology nodes.
The maturation of the ecosystem is reflected in reduced software compatibility issues. As one industry insider noted, in the past, Chinese chips faced implementation difficulties due to concerns about the software ecosystem and NVIDIA supply restrictions. This year, since NVIDIA cannot supply products to China, and the construction of a trillion-dollar American AI infrastructure consumes most of the advanced GPU capacity, domestic chips have become more accessible. When consumers lack alternatives, the problem of ecosystem compatibility gradually weakens as core models complete zero-day adaptation for internal computing platforms.
Among the most discussed models was Kimi K3, an open-source model with 2.8 trillion parameters. However, the key point is cost optimization: Kimi K3 uses the Stable LatentMoE framework, activating only 16 out of 896 experts during each inference, ensuring competitive performance at affordable prices for the mass market. While leading AI laboratories aim for the performance levels of Fable 5 and GPT-5.6, most WAIC participants focused on vertical applications in healthcare, life sciences, education, and tourism. For example, Bei Dian Shu Zhi demonstrated how AI-supported family medicine allows one online specialist to serve between 5,000 and 8,000 residents, while Baichuan showed the application of the M4 model in oncology and pediatrics. These examples prove that most enterprises do not require cutting-edge models to gain practical value, making cost-effective domestic computing infrastructure a true catalyst for AI industrialization in the Chinese economy.
Against the backdrop of global discussions about whether China is experiencing a deficit or surplus of computing power, China's semiconductor center in Zhangjiang, Shanghai, presents a clearer picture: stable and universal computing power remains in acute shortage, with orders for certain chips scheduled for next year and beyond.
Sun Guolian, Product Director at the GPU manufacturer MetaX (MXC series), noted that the shortage is observed regarding universal and reliable computing power. However, there is a surplus in non-universal architectures that can only operate with outdated or highly specialized models. Chips from this company's MXC600 series are supplied in large batches, and next-generation product development is in advanced stages.
Parallel to traditional electronic computing, Zhangjiang positions itself as China's national silicon photonics center. The Shanghai Zhangjiang silicon photonics cluster now unites over twenty companies covering the entire production cycle. Among them stands Lightelligence—the world's first company producing AI silicon photonic chips, which has been listed on the Hong Kong Stock Exchange.
Shen Yichen, founder of Lightelligence, who holds a Ph.D. from MIT and has previously published in the journal Nature, forecasts that silicon photonic chips will account for over 30% of the intelligent computing center market within five years, compared to less than 3% today. The company has already implemented hybrid optoelectronic solutions in three optical interconnect clusters of a thousand cards, enabling the operation of over 5000 GPUs for 44 corporate clients, including research institutions, internet companies, GPU manufacturers, and AI infrastructure operators.
The Zhangjiang 895 incubator has attracted several developing photonic startups, such as Qisuan Guangqi, specializing in optical chip R&D, and Guangjiu Technology, focusing on photonic technologies. The cluster also includes JCET (Changdian Technology) for advanced packaging and Suzhou Yite for silicon photonics testing, ensuring full coverage of the chain from design to testing.
The Zhangjiang model represents a dual approach to solving China's computing power problem: developing traditional electronic computing to ensure universal reliability while simultaneously creating a silicon photonics ecosystem. This ecosystem potentially offers a breakthrough in optical interconnects, allowing thousands of GPUs to function as efficiently as one chip by transmitting data via optical rather than electrical channels.
