"Alibaba is fully committed to building China's 'AI factory.'" Hong Kong's South China Morning Post recently reported that at the Alibaba Cloud Summit, held in Hangzhou on May 20th, Alibaba Cloud's Senior Vice President Liu Weigang stated that Alibaba Cloud is undergoing a full-stack technology refresh, with a comprehensive upgrade of its underlying chips, Agentic Cloud, models, and reasoning platforms, all aimed at "building China's largest AI factory."
Meanwhile, from Huang's "factory narrative" in recent years to Baidu's new benchmark of "daily active intelligent agents" proposed by Li, and then to Deloitte's prediction that the United States will operate "AI factories" on a large scale in 2028, a new era of "manufacturing intelligence" is coming. What does the "AI factory" actually produce? What changes have occurred in the industry logic from the "Token (token) factory" to the "AI factory"? How far has China progressed in this "factory" race and what unique challenges does it face?
The underlying logic of the "AI factory"
The South China Morning Post reported that Liu Weigang defined AI as a brand-new manufacturing form, creating revenue through "training and reasoning factories," and stated that Alibaba is the only AI and cloud company in China with a layout in all five layers of the "AI full-stack five-layer architecture," which runs through chips, agent cloud, AI models, model service platforms, and agent applications.
In fact, as early as 2022, Huang Renxun mentioned the concept of "AI factory", "AI data centers process massive and continuous data to train and improve AI models, raw data comes in, is refined, and then intelligent output - enterprises are manufacturing intelligence and operating large 'AI factories'".
Since then, Huang has gradually explained the "AI factory" in layers, replacing the "chip narrative" with the "factory narrative". Nvidia's layout in multiple fields also proves what Huang said, "Nvidia has evolved from a GPU company to an 'AI factory' company." Nvidia's official blog shows that in 2025, Huang said publicly: "AI is now a kind of infrastructure like the Internet and electricity. What we are building today is an 'AI factory'." He further explained, "Injecting energy into the AI factory will produce very valuable products, which are called Tokens."
In March this year, Huang Renxun once again wrote an article, comparing the AI industry image to a "five-layer cake", which is composed of energy, chips, infrastructure, models and applications from bottom to top. The infrastructure layer includes land, power supply, cooling systems, construction engineering, network communication, and the system that arranges thousands of processors into a machine, which is the "AI factory". He believes that chip factories, computer assembly factories and "AI factories" are being built on an unprecedented scale, which is becoming the largest infrastructure construction in human history.
Mei Dangqing, an assistant professor of finance at the China Europe Business School, offered a more detailed perspective on the logic of the "AI factory" during an interview with the Global Times: The bottom layer consists of power, cooling, data centers, and networks; above that are chips and accelerators; then come cloud platforms, reasoning engines, scheduling systems, and security governance; above that, there are general-purpose foundation models, as well as industry-specific adaptions, knowledge bases, and Agent (smart agent) orchestration built around them; and finally, various real business scenarios. Mei Dangqing believes that Nvidia's "AI factory" also emphasizes that it is an infrastructure for scaling up the "manufacture of intelligence," with the core being the conversion of computing power into usable, measurable, and billable intelligent output.
Among the above various "AI factory" concepts, Xiong Wei, an analyst at UBS Securities covering the Chinese internet industry, revealed the most obvious bottleneck in the development of "AI factories" in China, which is the computing power. "When facing the explosive growth of demand, global companies all say that they lack computing power. China's computing power also faces some restrictions on the import of high-end chips," Xiong Wei told reporters from Global Times. Under the pressure of external forces, the development of domestic chips has been very rapid in recent years, and some chips in the upstream infrastructure layer are benefiting from the huge investment of midstream model providers, which currently gathers a lot of industrial chain value.
"Once the chip link is filled, our full-stack technology innovation will be a very perfect storyline," Liu Weiguang said in an interview with multiple media outlets after the press conference. Compared to last year, this year's mention of Alibaba as "China's leading AI full-stack service provider" is more confident because Alibaba released the Pangu AL128 super-node server based on the new generation of AI chips, Zhuanwu M890, which is mainly used to solve the massive concurrent reasoning and large model training needs in the Agent scenario.
Liu Weigang described that behind each model training, there is powerful computing power. Computing power changes with the change of model parameters, and it is also iterating faster. The two are tied together, like gears biting each other and continuously spiraling up. "If every chip can run more Tokens and higher quality Tokens than competitors in the future, then we will win." Data shows that Alibaba's GPU chip has started mass production. At present, the True武 series chips have shipped a total of 560,000 pieces, serving more than 400 enterprises and institutions such as China Telecom.
Inexpensive or luxury goods?
What does the "AI factory" produce? Bi Qi, chief scientist of China Telecom Group and a fellow of Bell Labs in the United States, explained it this way: previous factories produced products in the physical world, which had physical value. With the development of information technology, virtual products have also been shown to have huge economic value, and data can also become an asset.
The core of the "AI factory" is the production of virtual products, just like a physical factory, there are two key elements: one is the "equipment", namely computing power; the other is the "product" and its pricing method. Before the measurement of computing power was FLOPS, which stands for floating-point operations per second, and its pricing method was difficult to be accepted by the public. But after the emergence of large models, the concept of "Token" is very similar to traffic, and it has become a recognized and billable unit of computing power value.
"The computing power center, in coordination with large models, continuously and endlessly produces tokens with economic value, just like factories produce products with machines. The former sells tokens, and the latter sells goods," Bi Qi told reporters from the Global Times. "AI factory" is the combination of computing power centers and large models, and its product is tokens, just like the product of the Internet is traffic.
Bici further explained that different Tokens have widely varying values. "It's like the clothes produced by a factory, some cost a few yuan per piece, while others cost tens of thousands of yuan each as luxury brands. The value depends on its application, the industry it solves problems for, and the actual value it creates."
On May 13th, Baidu's founder, Li Yanhong, proposed the "measuring standards" of the AI era - Daily Active Agents (DAA) - at Baidu's AI Developers Conference. He believes that "as humans enter the era of agents, the measure of a platform and ecosystem's prosperity should focus more on the DAA metric, paying attention to how many agents are working for humans and delivering results. This is closer to value and essence than mindless Token consumption."
Beech believes that "daily active intelligent agent number" is a higher quality "AI factory" pricing method. "It is more primitive to have a computing center that produces Tokens like a sewing machine. The key is to produce high-quality, high-value-added 'clothes' and price them according to their value in the industry." Beech said that the economic value of Token used in which industry depends entirely on the quality and direction of the large model and the problems it can solve for whom. "The computing center is like a sewing machine, and the design and intellectual property rights of the large model, as well as the form of services provided to provide value, determine whether the clothes are cheap or luxury goods."
The concept of "AI factory" is easily misunderstood as "compute worship". It seems that whoever has more GPUs, more data centers, more power, wins. Mei Dangqing believes that the competition in the first half may focus on the above factors, but the real second half is whether enterprises can establish an evaluation system to guide AI products to develop in a better and higher quality direction, otherwise, "AI factory" is prone to become an expensive Token production machine.
China and the United States are the two major players in the "AI factory" field.
In its March 2023 report, "Deloitte Enterprise AI Infrastructure Survey: Outlook to 2028," Deloitte surveyed 515 U.S. enterprise leaders from five industries (Consumer, Energy & Resources, Industrial, Financial Services, Life Sciences & Health Care, Technology, Media, & Telecommunications) with annual revenues of over $500 million, and from November to December 2025, the results showed that more than 70% of respondents expected to operate "AI factories" on a large scale by 2028.
"Many countries and regions are talking about 'AI factories' these days, but the underlying logic is quite different," said Mei Danqing, telling reporters that the United States is more like a frontier model and a global AI service factory; Europe tends to build sovereign AI and public innovation infrastructure, its focus is not entirely on commercial profits, but hopes that European enterprises and research institutions do not completely rely on American platforms in the AI era; the Middle East is more like an energy and capital-driven computing hub; South Korea and Japan focus on developing industry AI and manufacturing upgrading platforms.
Overall, China and the United States are the two major players in the "AI factory" field. Mei Dangqing believes that China has three advantages in developing "AI factories": First, China has a high density of application scenarios, and enterprises have a fast speed of trial and error. Manufacturing, logistics, e-commerce, finance, government affairs, medical and other scenarios are not just simple chats, but also a large number of repetitive decision-making, system calls, and exception handling, etc.; Second, China has strong engineering and cost control capabilities. Chinese enterprises are good at quickly productizing new technologies, making them low-cost and scaled. There will be strong motivation in low-cost reasoning, engineering optimization, model compression, and domestic chip adaptation; Third, China has strong infrastructure construction capabilities. The underlying layer of the "AI factory" is infrastructure projects, which need electricity, land, cooling, network, data center construction and operation and maintenance, etc. China has rich experience in large-scale infrastructure, energy scheduling and industrial park supporting aspects.
"The construction of China's 'AI factory' has clearly begun," said Meདānqīng. Alibaba's latest move is not to release a single model or a single chip, but to present a package deal that includes models, cloud infrastructure, AI chips, and the Agent service platform. This indicates that China's leading cloud vendors have realized that future AI competition is not about single-point capability competition, but about full-stack capability competition.
Xiong Wei analyzed that the Token pricing that Chinese model factories can provide is far lower than that of their American counterparts. Although there is a certain gap in model performance, the gap in performance is far less than the gap in pricing. "The developers of Chinese models invest more resources and energy in R&D, and the strategy is more focused, making the model's reasoning efficiency more extreme, and making the cost advantage of the Chinese side more obvious."
Bici believes that current large model technology is still in the rapid iteration stage, and there may be multiple "lanes" in the future, while the United States currently occupies the dominant position in basic science and disruptive innovation. The United States may always be able to take the lead by constantly "laning", and China must follow closely, otherwise there is a risk of falling behind.
But Bi also said that there is a high degree of complementarity between China and the United States in the AI industry, and the space for cooperation is far greater than competition. Although the United States is good at creating cutting-edge science from "0 to 1", if it cannot find applications, the technology will "rot in the pot". China is also committed to "0 to 1" original innovation, while also focusing on "1 to 100" engineering application and execution. If the two countries can cooperate, it will definitely be a "1+1>2" result. In addition, although the current public opinion is rendering competition, the two countries' national conditions may still ultimately lead to a situation of "competition, but cooperation is greater than competition".
SunSirs has been continuously tracking price data for over 200 commodities for nearly 20 years, please contact support@sunsirs.com for subscription.
