Investment opportunities hinge on thorough analysis, which provides authoritative, professional, timely, and comprehensive insights to uncover potential themes.
Domestic computing power chips have progressed from inference to also handle training, taking on both roles.
At the "Token Era, Universal Intelligence" sub-forum held during the 2026 World Artificial Intelligence Conference on July 18th, Zhang Jianzhong, founder, chairman, and CEO of domestic GPU maker Moore Threads, emphasized in his speech that domestic AI models can confidently be trained on domestic chips.
"I often hear domestic model companies worrying whether using a domestic training platform will waste time or hinder their ability to catch up with global leaders. In fact, everyone can be reassured because we have better hardware platforms and advanced software technology, backed by extensive experimental results. You can truly invest confidently in training domestic models on domestic chips," Zhang Jianzhong stated. He added that training on domestic chips offers the added benefit of Moore Threads being able to provide localized support at any time, which can be better than foreign companies.
The day before, another local GPU manufacturer, Muxi, also clarified that its chips are capable of both inference and training. Sun Guoliang, Senior Vice President and Chief Product Officer of Muxi, made this point during a media interview.
"Many clients ask about our inference performance, cost-effectiveness, and how to calculate costs. But what's often overlooked is that for Muxi, our cards are capable of training. At every WAIC, we've stated we are training-inference integrated," Sun Guoliang said. He noted that due to cost-performance considerations, many products, including all-in-one machines, are designed for inference, but Muxi has the capability to cover training. The Muxi team is particularly eager to undertake large-scale model training involving tens of thousands of cards.
The statements from Moore Threads and Muxi indicate that domestic computing power cards are now capable of shouldering the heavy task of large model training, not just inference. This solidifies the foundation for China to grow its "Token" economy and even pursue "Token export."
Another key theme at this year's exhibition is the "Token" economy. GPU manufacturers, computing power service providers, and telecom operators have positioned themselves around the "Token" economy, with this year being dubbed the inaugural year of the "Token" economy.
Moore Threads proposed establishing three types of factories: a large model factory, a token factory, and an agent factory.
Zhang Jianzhong explained that the model training factory is dedicated to training various models. The token production factory generates Token services for different models. The Agent (intelligent agent) production factory creates robots and the unseen digital brains behind avatars. "Without a good factory to train the thinking patterns and methods of these Agents, it's difficult to judge whether an Agent's intelligence can meet the task requirements set by humans," he said.
Cao Peng, Chairman of JD.com's Technology Committee and President of JD Cloud, also spoke at the forum, stating that JD Cloud is collaborating with Moore Threads to jointly tackle challenges with the JoyAI large model, build a more efficient enterprise-level inference platform, and co-create full-chain embodied intelligence services. The goal is to ensure domestic computing power can not only run models well but also run business operations effectively, unleashing industrial intelligent productivity.
Regarding the "model training factory," Mao Jiming, Partner and Vice President of Jijia Shijie, mentioned that based on Moore Threads' MTT S5000, they have jointly completed the training of physical AI models such as the driving world model, driving reconstruction model, embodied world model, and world action model. Both sides will deepen cooperation focusing on model efficiency through hardware-software co-design.
Jiuzhang Yunjie, an AI infrastructure and intelligent computing cloud provider, also has two main AI factories: a training factory and a Token factory. The training factory focuses on "refining" massive computing power and data into professional models needed by various industries using advanced techniques like reinforcement learning.
The Token factory packages the professional models produced by the training factory into standardized, precisely measurable professional Tokens (categorized as consumer-grade, professional-grade, and cutting-edge-grade). This enables on-demand access, pay-per-use, and efficient circulation of intelligent value, pursuing ultimate delivery cost and scale efficiency.
A Jiuzhang Yunjie staff member noted that previously, computing power service providers emphasized "the number of cards," but transitioning to a Token factory requires professional-grade services.
The staff member also revealed that Jiuzhang Yunjie has some domestic cards and will gradually acquire more, acknowledging that card supply remains tight.
Lenovo systematically showcased its full-stack hardware and software capabilities for Token-driven computing power infrastructure at this year's WAIC. In a prominent position at its booth, Lenovo displayed its recently launched Wanquan Heterogeneous Intelligent Computing Platform V5.0 and Super Node solution. The Lenovo exhibition area also featured acceleration cards from several domestic computing power providers like Muxi and Moore Threads, which are used in Lenovo servers.
A staff member at the Lenovo booth stated that the overall market is experiencing a computing power shortage, and Lenovo servers are selling well. "Previously, we shipped goods to distributors. Now, customers are coming with cash to get the goods. The situation has completely reversed," they said.
Next-generation AI infrastructure service provider Qianshi Technology also launched the wylon Q72/288 Super Node, designed specifically for Token factories. A single cabinet integrates 72 domestic GPUs, with four cabinets horizontally interconnected to scale to a 288-card cluster. The GPU high-bandwidth memory can reach 18TB, with a total interconnect bandwidth of up to 72TB/s, supporting FP8/4 floating-point precision. It features a unified DRAM cache designed for KV Cache that can reach hundreds of terabytes, providing integrated, efficient computing and storage space for large model inference to achieve efficient and stable Token production.
It was reported that the wylon New Cloud Token Factory, based on the wylon Super Node architecture, is now online in an East China cluster node and is in the invitation testing phase. The wylon Super Node system deeply collaborates with leading domestic GPU chip companies like Cambricon, Biren, Muxi, Xiwang, Hygon, and Moore Threads, building a fully integrated domestic AI infrastructure from chips to systems and from computing power to ecosystem.
Since entering July, stock prices of listed companies in the AI hardware supply chain have experienced significant volatility, with memory chip companies seeing particularly sharp movements.
When asked, on-site staff from several GPU manufacturers remained optimistic about GPU prospects. Although rising costs of various raw materials have created some pressure, demand remains strong. However, they held a dimmer view of the memory chip market outlook.
A server industry insider stated that the price of DDR memory chips has increased by up to tenfold, essentially absorbing the entire industry's profits. Currently, such high prices are suppressing demand, making the涨价 logic unsustainable. Supply and demand are expected to gradually balance in the second half of the year.
An MLCC (Multi-layer Ceramic Capacitor) industry professional believed that the memory shortage might persist until next year, but prices have likely peaked and are no longer rising.
Comments