WAIC 2025 Observation: The Computing Power Competition Upgrades, and Models Seek Paths to Real - world Application

On July 28, the 2025 World Artificial Intelligence Conference (WAIC) concluded in Shanghai. With an exhibition area of over 70,000 square meters, more than 800 enterprises, and a single - day ticket price that was once speculated up to 3,000 yuan, all these demonstrated the unprecedented popularity of this grand event.

Compared with previous years, this year's WAIC presented two increasingly clear industrial paths. On the one hand, hardware manufacturers represented by Huawei, Sugon, and Digital China were pushing the competition in computing power infrastructure to new performance peaks. On the other hand, model and application manufacturers represented by Baiwang Co., Ltd. and Jieyue Xingchen collectively shifted their focus from technical parameters to the in - depth implementation of business scenarios.

This "two - sided" situation could be seen everywhere at the conference. For example, reporters learned at the Huawei booth that the Ascend 384 SuperNode (Atlas 900 A3 SuperPoD), which was exhibited offline for the first time, had a single - cluster computing power of up to 300 PFLOPS (30 quintillion floating - point operations per second). This surging computing power within the cold cabinets constituted the "hottest" side of this conference. However, at the other end of the exhibition hall, the calm thinking of downstream practitioners provided another dimension of annotation for this upsurge.

"The value of large models lies not in technical tools, but in the depth of scenario implementation." On July 27, Fu Yingbo, CEO of Baiwang Co., Ltd., emphasized this during an interview with a reporter from Economic Observer. His view represented another voice in the industry. Beyond the computing power competition, how to transform AI into real business value was equally important.

The "New Infrastructure" of Computing Power Continues to Upgrade

"For the AI field, we have invested a total of 8 billion yuan to support infrastructure such as data centers and computing power leasing." On July 27, Xia Yuan, President of Core - Xin Leasing, said at a sub - forum themed "Industrial Finance Joins Hands with New AI Forces".

According to Xia Yuan, as the only financial service platform in China dedicated to serving strategic emerging industries such as integrated circuits, Core - Xin Leasing has invested a total of 210 billion yuan in the integrated circuit industry in the nearly ten years since its establishment. The 8 - billion - yuan investment in AI infrastructure this time was an extension of its strategy to "build an independently controllable AI industrial chain".

The industrial financial power represented by Core - Xin Leasing was providing continuous capital support for this increasingly expensive computing power competition. At the conference site, reporters also noticed a clear trend: this computing power competition was evolving from the pursuit of a single performance indicator into a "systematic project" that must solve a series of chained challenges such as performance, compatibility, storage capacity, and even energy efficiency in sequence.

In the pursuit of extreme performance, Huawei provided the most intuitive example at this conference. Reporters learned on - site that the Ascend 384 SuperNode (Atlas 900 A3 SuperPoD), which was exhibited offline for the first time, efficiently interconnected 384 NPUs and 192 Kunpeng CPUs through a fully peer - to - peer (Peer - to - Peer) UB bus and a non - blocking Clos architecture. Its single - cluster computing power output of up to 300 PFLOPS all pointed to one goal - to build the current strongest computing power engine.

If the pursuit of extreme performance defined the height of the computing power competition, then the in - depth understanding of different application scenarios determined the breadth of this competition.

"The market demand is evolving. We have observed two obvious trends. First, due to data security concerns, enterprises' demand for the 'private deployment' of large models is increasing. Second, the scenarios of 'training' and 'inference' are differentiating, and different scenarios require computing power solutions with different cost - performance ratios." Zhou Chuan, Vice President of the Xinchuang Business Group and General Manager of the R & D Center of Digital China, said during an interview with a reporter from Economic Observer.

Based on this judgment, KunTai, a subsidiary of Digital China, launched the industry's first large - model training and inference integrated server, KunTai R624 K2, and the inference server, KunTai R622 K2, based on the Kunpeng technology route at this conference. Sun Yanan, a product expert of the Xinchuang Business Group of Digital China, introduced to reporters that the KunTai R624 K2 could support up to 10 mainstream AI accelerator cards, and its computing efficiency was twice that of previous products. The KunTai R622 K2 achieved the highest computing power density in the industry within a 2U space.

What was even more noteworthy was its "open" strategy. Zhou Chuan emphasized to reporters, "While adapting to the powerful Huawei Kunpeng and Ascend ecosystem, our products are also compatible with mainstream domestic and foreign AI accelerator cards. Our original intention is to use products with higher performance and better cost - performance to break down hardware compatibility barriers and enable every industry to truly use and afford large models."

If CPUs and AI accelerator cards are the "brains" of computing power, then the storage system is its "blood - supplying heart". No matter how powerful the computing power is, if data cannot be "fed" in, it's all in vain.

"We have observed that there are currently three major problems in data centers: 'incomplete data view, disordered data management, and poor data utilization'," He Zhen, President of Sugon Storage, told reporters. "The systems of different manufacturers are incompatible, and data forms isolated islands, which is 'incomplete data view'. The value of massive data is difficult to quickly locate and circulate across domains, which is 'disordered data management'. The high latency and poor experience in data calls between the eastern and western regions are 'poor data utilization'."

As the manufacturer with the first - place market share in the AI storage market for two consecutive years, Sugon Storage was trying to unblock this "data lifeline".

At this conference, China Mobile and Sugon Storage jointly launched the first global unified file storage system in China and initiated an intelligent storage capacity scheduling platform. This platform first covered four major national hub nodes in the Yangtze River Delta, Chengdu - Chongqing, etc., aiming to serve the national strategy of "Eastern Data, Western Computing".

"Don't let the GPU wait for data." Zhang Xinfeng, Vice President of Sugon Storage, used a vivid metaphor to explain the importance of storage - computing synergy. She told reporters that through technologies such as context storage and GDS (GPU Direct Storage), allowing data to bypass the CPU and directly reach the GPU could significantly improve inference efficiency, increasing the number of tokens (which can be understood as the unit of words processed by AI) generated per second from hundreds to 2,000 - 3,000.

"We have calculated that for every one - yuan investment in storage capacity, the computing power cost can be reduced by ten yuan," Zhang Xinfeng added.

Yang Zhilei, Vice President of Sugon Storage, emphasized the technical challenge of "seamless cross - domain operation" to reporters: "In the process of cross - domain data scheduling, it is necessary to ensure that business operations continue and that the data has strong consistency during concurrent access by multiple nodes. This poses a huge challenge to the metadata management ability and network latency of the storage system." He said that Sugon Storage's solution had accelerated from the underlying chips to the upper - layer software and considered the requirements of privacy computing and trusted computing, laying out for future data security.

Another direct physical challenge brought about by this computing power competition was the soaring power consumption along with the performance. At the conference site, this challenge also gave rise to new solutions.

Reporters learned at the SuperCloud Digital Technology Group booth that the company launched its full - stack liquid - cooled intelligent computing server, R8429 L13, for the first time at this conference. According to relevant staff, when the power of a single cabinet exceeded 20 kW, liquid cooling had become a key heat dissipation solution, and this product was designed to address the increasingly severe energy consumption problem of data centers in the era of large models.

From the injection of capital, to the surging performance, and then to the return to openness, synergy, and efficiency, a three - dimensional and multi - dimensional competition around the "new infrastructure" of computing power was fiercely taking place on the track of China's AI industry.

Large Models Enter the Deep Water Area of the Industry

If computing power infrastructure is "building roads", then models and applications are the "cars" running on the roads.

At the conference site, reporters observed that model manufacturers were collectively "coming down from the clouds", bidding farewell to simple parameter comparisons and technical "show - offs", and diving into the "deep water area" of the industry to find scenarios, build ecosystems, and start an exploration to transform technology into business value.

Xu Li, Chairman and CEO of SenseTime, pointed out in his speech at the main forum that after learning human - annotated data (perceptual AI) and text data on the Internet (generative AI), the evolution of artificial intelligence was facing a new data bottleneck.

"When the knowledge in books and on the Internet is exhausted, where will the next - generation intelligence come from?" Xu Li asked. He believed that the answer lay in moving towards the physical world. Through "embodied intelligence" and the "world model", AI could grow like humans through interactive exploration with the real world.

This judgment also explained why many AI enterprises were sparing no effort to promote the in - depth integration of models and industrial scenarios. Only in real scenarios could AI access new, high - quality data containing physical laws and thus achieve a higher - dimensional evolution.

"The industry has entered the stage of'returning to customer needs and basing on real application scenarios'," Jiang Daxin, CEO of Jieyue Xingchen, also said at the conference site. He also emphasized that the industry must answer the question of "whether model performance is completely equivalent to model value".

Based on this, the core goal of the new - generation base model, Step 3, launched by Jieyue Xingchen, was to "reduce costs and increase efficiency". Relevant staff of Jieyue Xingchen introduced that the inference efficiency of this model on domestic chips could be up to three times that of DeepSeek - R1. At the same time, Jieyue Xingchen chose to "make friends" and jointly initiated the "Model - Chip Ecosystem Innovation Alliance" with nearly 10 chip manufacturers such as Huawei Ascend and Moore Threads, trying to unblock the software - hardware synergy at the underlying level and reduce the application cost of the entire industry.

Building platforms to empower developers was another way of "making friends".

"Our positioning is as an AI infrastructure, aiming to reduce the burden and increase the efficiency of scientists and industrial developers. By providing easy - to - use tools, we want to make AI an 'invisible assistant'," Xiao Yunfeng, the engineering leader of the Inspire Platform of Infinite Light, told reporters. Its "Inspire Platform" was committed to building a "full - link computing power foundation" to enable developers to "break away from technical trivialities and focus on the essence of science and business innovation".

When model manufacturers started to "make friends" and platform manufacturers started to "offer tools", a more specific question followed: How exactly can the value of AI be "extracted" in the specific scenarios of various industries? A group of enterprises deeply involved in vertical fields gave their answers.

"Our core competitiveness lies in the 'transaction data gold mine' accumulated over more than a decade, with a total transaction volume of 953.5 trillion yuan," Fu Yingbo, CEO of Baiwang Co., Ltd., elaborated on his business logic to reporters. He believed that general open - source models provided a technical foundation, but only when combined with in - depth industry data like Baiwang's could a "business brain" that truly understood business logic and could solve enterprises' rigid requirements such as tax compliance be trained.

Fu Yingbo also said that Baiwang Co., Ltd. was actively cooperating with model manufacturers such as Alibaba Tongyi Qianwen and Lingyi Wanwu, as well as chip enterprises such as Moore Threads, to jointly promote the implementation path of "open - source models + industry scenarios".

In creative industries such as culture and entertainment, AIGC (AI - generated content) was also reshaping the production process.

"AI is not a substitute for creative people, but a'super assistant'. Its existence allows creators to jump out of cumbersome processes and engage in core links," Dr. Sun Daqian, Executive Director of Digital Domain, told reporters. The company launched the "AI DOMAIN" one - stop imaging creation platform at WAIC, integrating seven core functions such as text - to - image, image - to - video, and intelligent image expansion.

According to Liu Hongjie, General Manager of Digital Domain in China, Digital Domain also jointly established the "Digital Visual Innovation Alliance" with the Hong Kong University of Science and Technology, Shanghai Film Group, and Moore Threads, aiming to "explore intelligent and large - scale content production channels" and promote the mass production of content such as short videos and short dramas.

When the focus shifts from the industrial end back to individual consumers, the implementation of AI is manifested in product forms closer to daily life.

According to the staff at the Mobvoi booth, the newly launched TicNote AI voice recorder of the company encapsulated complex AI capabilities such as recording, summarizing, analyzing, and creating into a hardware - software integrated product that consumers could use, aiming to become users' "portable AI thinking partner".

Whether it is the "business brain" empowering enterprise decision - making, the "super assistant" assisting content creation, or the "thinking partner" serving individual users, all the competition in computing power and exploration of models ultimately point to the same goal -

As Li Hongjing, Chairman of Autel Robotics, quoted in his speech at the conference forum, "The destination of AI is not 'looking smart', but promoting the real GDP growth of society, industries, and the country."

This article is from the WeChat public account Economic Observer. Author: Zheng Chenye. Republished with permission from 36Kr.