21st Century Business Herald reporter Deng Hao and Sun Yan report from Shanghai

Under the Sino-US trade friction, independent and controllable has become the only way for the development of local industries, especially in this round of AI wave, how to catch up with and maintain the status of the first echelon industry is particularly important.

Recently, at the HUAWEI CLOUD Ecosystem Conference 2025, Huawei launched the CloudMatrix 384 SuperNode and announced that it has been launched in the Wuhu data center.

Zhang Xiuzheng, vice president of Huawei, said that tens of thousands of HUAWEI CLOUD CloudMatrix 384 super nodes will be launched in the first half of this year, which may completely end the anxiety of computing power. In the future, CloudMatrix Super Nodes can build large clusters of more than 10,000 pieces to provide computing power.

Yuan Jinhui, founder of Silicon-based Mobility, announced at the aforementioned meeting that Silicon-based Mobility and HUAWEI CLOUD have officially launched DeepSeek-R1 with the best practices of large-scale expert parallelism based on CloudMatrix 384 super-node Ascend cloud services and the high-performance inference framework SiliconLLM. Under the premise of ensuring 20 TPS for a single user, the decode throughput of a single card exceeds 1920 Tokens/s, which is comparable to the H100 deployment performance.

From the perspective of the industry, in the context of the limited supply of NVIDIA's high-end chips such as H100 and H20, the launch of HUAWEI CLOUD CloudMatrix 384 may inject a boost into the rise of domestic AI.

Break out autonomously

In 2022, the United States launched the first round of chip bans, and Nvidia was asked by the U.S. government to restrict the export to China of two of the latest two generations of flagship GPU computing chips, A100 and H100, which are used to accelerate artificial intelligence tasks.

In order to meet export control requirements, in March 2023, NVIDIA launched the H800 for the Chinese market. Although the H800 was slightly inferior in performance to the H100, it soon faced a ban in October of the same year. On the basis of H800, Nvidia launched H20, which has significantly reduced computing power and memory bandwidth.

However, the United States is considering new control measures. There are reports that the Trump administration had considered tightening restrictions on Nvidia's sales of its H20 chips designed for the Chinese market, but the White House changed its plan to restrict Nvidia's H20 exports to China after Nvidia CEO Jensen Huang traveled to Mar-a-Lago in Florida to attend a dinner.

No matter how repeatedly the U.S. export controls jump, in the context of geopolitical fluctuations, the independent rise of domestic computing power has become the only way.

At present, the Cambrian (688256. SH), Haiguang Information (688041. SH) and other listed AI computing chips, as well as non-listed AI computing chip companies such as Muxi, Tiantian Zhixin, and Bichen Technology, help solve the hunger for computing power.

It is worth mentioning that domestic computing power chips focus on inference, and representative products such as Ascend 310 and DeepEdge10 are representative products. IDC's "China Semi-Annual Accelerated Computing Market (First Half of 2024) Tracker" also shows that in the first half of 2024, AI chips used for inference accounted for 61% of the market share.

The CloudMatrix 384 SuperNode is comparable to the NVIDIA NVL72 Supernode in terms of scale and inference performance. According to Haitong International's research, previously, NVIDIA's NVL72 system supported a maximum of 72 card clusters, but Huawei has expanded the scale of a single node to 384 cards through the "384 card non-convergence networking" technology, becoming the world's largest commercial AI computing power cluster.

In the actual test, the DeepSeek-R1 model deployed based on CloudMatrix 384 achieved a throughput of 1920 tokens/s under a load of 20 TPS for a single user, surpassing the 1850 tokens/s of international mainstream chips, and the generation accuracy was better in tasks such as knowledge quiz and code generation.

In addition, the emergence of CloudMatrix 384 supernodes is also expected to achieve a breakthrough in local computing power on the training side. According to Guotai Haitong's research, the high-speed interconnection bandwidth of CloudMatrix 384 supernodes is expected to accelerate the breakthrough of domestic computing power cards to training scenarios.

Pull the upstream volume

To a large extent, the emergence of supernodes is to achieve high-speed interconnection and communication in the form of hardware integration and improve the training or inference efficiency of AI models.

With the mass production of the super-node solution represented by Huawei, the upstream supply chain may usher in a large-scale cycle.

According to public information, Huawei's super-node network switches use 6,812 400G optical modules to achieve 2.8 Tbps inter-card interconnection bandwidth.

It is understood that Huagong Technology (000988. SZ) has laid out a full range of products from optical chips based on various compounds to devices, modules, and intelligent terminals, and 400G/800G optical modules continue to increase.

Orient Securities expects that Huagong Technology's 1.6T optical module will be commercialized within the year, and the application of 3.2T silicon photonic module will be accelerated. In addition, Huagong Technology has conducted tests of 400G, 800G and 1.6T products in a number of leading customers. 800G LPO products have a clear demand, and are currently stepping up testing and introducing as soon as possible, and 1.6T products are speeding up sample testing, and the overall progress is in the first echelon of the industry.

The actual controller of Accelink Technology (002281.SZ) is China Information Technology, as a veteran optical communication manufacturer, it has created a vertical integration platform of optical chips, optical devices and optical modules. China Galaxy Securities believes that Accelink is currently one of the few domestic manufacturers with strategic R&D capabilities for optical chips, with a very considerable annual production capacity and a high self-sufficiency rate of low-speed optical chips, and is one of the few domestic manufacturers that mass-produce DFB chips below 25G, and continues to break through high-end products such as 100G and 200G optical chips.

In addition to optical modules, due to the increase in computing power and hardware density, how to effectively dissipate heat has become a key consideration for the super-node solution.

Taking NVIDIA B200 as an example, its power consumption exceeds 1000W, which is close to the upper limit of air cooling and heat dissipation. Guohai Securities believes that with the PUE requirements of data centers such as "dual carbon" + "eastern data and western computing", and considering the TCO cost of newly/renovated data centers, liquid cooling solutions have more advantages than other heat dissipation technologies. 

According to the data, Invic (002837.SZ) is a leading provider of precision temperature control and energy-saving solutions and products in the industry, and is committed to providing equipment heat dissipation solutions for cloud computing data centers, computing equipment and other fields. At present, it has become a major customer supplier such as Huawei, ZTE, Tencent, and Alibaba.

According to the analysis of Guohai Securities, Invic has invested in liquid cooling technology platforms such as cold plate and immersion for a long time, and has launched Coolinside liquid cooling cabinets and full-chain liquid cooling solutions. In the first half of 2024, its operating income from temperature control and energy saving in the computer room will be 856 million yuan, a year-on-year increase of 85.91%.

Debang Securities also believes that air cooling is difficult to cool high-density racks, and the penetration rate of subsequent liquid cooling will increase rapidly. It believes that the large liquid cooling system launched by Shenling Environment integrates the wind-side heat dissipation and liquid-side heat dissipation of the liquid-cooled data center, simplifies the system, supports the same source of wind and liquid, and is more adaptable, and can provide products such as dry coolers or evaporative coolers on the primary side, prefabricated pipelines, CDUs on the secondary side, ring networks on the secondary side, Manifold and other products. In addition, with the close cooperation with Company H, it is expected that with the increase in the shipment of Al computing equipment of Company H, the application of Shenling environmental liquid cooling products will be further accelerated.