The Zhitong Finance App learned that CITIC Securities released a research report stating that DPU undertakes infrastructure tasks such as server networking and storage management, and is widely used in eight-card servers, supernode servers, high-performance computing servers, and storage servers. In the AI era, its usage continues to increase as the number of computing power chips increases, and the price continues to increase as communication speed is upgraded; in a horizontal comparison, the market is relatively large among all connected devices. Currently, domestic replacement of DPU is gradually being promoted; among them, independent third-party suppliers, which are gradually emerging, are rapidly growing in revenue due to deep ties with customers and advanced technical capabilities. It is recommended to focus on leading third-party vendors.
CITIC Securities's main views are as follows:
Interconnection layer: There are often different levels of interconnection networks in computing power infrastructure, which are used to achieve different interconnection goals and form different network planes to differentiate traffic.
1) Inside the cabinet: ① GPU-GPU (scale-up): Enables high-speed GPU interconnection within the cabinet. ② CPU-peripherals: Connect the CPU and peripherals to manage scheduling.
2) Inter-cabinet: ① GPU-GPU (scale-out): enables high-speed GPU interconnection across cabinets. ② CPU-storage (business-side FrontEnd network): Enables data management for external storage. ③ GPU-GPU (second-layer scale-up): Cross-cabinet group scale-up network. Among them, the GPU-to-GPU (scale-out network) and CPU-storage (front-end network) between cabinets require networking devices, and the component corresponding to inflation is DPU.
DPU: Connects the server to the switch and offloads some CPU tasks.
DPU is divided into AI DPU and full-function DPU.
AI DPUs are connected to the server through PCIe, package host-side data into network frames, and connect to switches through an Ether/IB interface to achieve high-speed interconnection; in the AI era, they are mainly used for scale-out networks with GPUs across cabinets, and generally have hardware acceleration capabilities such as RDMA.
Compared with AI DPU, full-function DPU has added CPU cores, which can take on the CPU's network, storage, security, virtualization, etc., and adapt to a separate storage and calculation architecture; full-function DPU generally has independent DDR and can be integrated with PCIe Switch as the connection center; in the AI era, it is mainly used for intelligent computing servers to connect to storage servers and front-end networks for external business.
Currently, DPU has become one of the pillars of data centers, and together with CPU and GPU, it is called by Nvidia's “three main chips” in data centers.
Market space: Domestic demand is expected to support the market of nearly 100 billion yuan in 2028, nearly doubling compared to 2026.
Volume: DPU is mainly used in four types of scenarios: eight-card servers, supernode servers, total computing servers, and storage servers. According to various product solutions: 1) In an eight-card server, XPU: AI DPU = 1:1, XPU: full-function DPU = 8:1. 2) In the supernode server, XPU: AI DPU = 1:1, XPU: full-function DPU = 4:1. 3) In total computing servers, high-performance computing servers (used by cloud service providers, Internet vendors, and operators) usually need to remove the CPU burden, and are generally equipped with a full-function DPU. 4) In storage servers, high-performance storage servers required for large-scale intelligent computing are generally equipped with an AI DPU.
Price: Gradually upgraded with the rate, the full-function DPU is impacted by the storage price. Overseas situation: 1) In terms of AI DPU, the overseas ConnectX-7 (400Gbps rate) manufacturer takes about 1,000 US dollars, while the price of ConnectX-8 (800Gbps speed) is about 1,500 US dollars. It can be seen that the price of the intergenerational upgrade has increased by 50%. 2) In terms of full-function DPU, the price of overseas BlueField-3 (400Gbps rate) storage is estimated to be between 2000 and 3,000 US dollars before the surge, and close to 4,000 US dollars after the surge. The price of BlueField-4 is unknown. Domestic situation: The price of domestic benchmark products is usually 70 to 80% off compared to mainstream products from major international manufacturers.
Spatial estimates: In 2028, the domestic demand supporting AI DPU market space may reach 41.4 billion yuan, and the full-function DPU market space may reach 49.8 billion yuan. In 2026, China will be dominated by 400Gbps AI DPU and full-function DPU, and will be fully upgraded to 800Gbps AI DPU and full-function DPU in 2028. Combining assumptions about domestic computing power chips, supernode penetration rate, total usage of computing servers and storage servers, and the expected price of domestic DPU, it is estimated that in 2028, China's total DPU space will reach 91.2 billion yuan.
Pattern: Domestic alternatives are on the rise, and we are optimistic about the prospects of third-party independent suppliers.
Looking at overseas markets, most cloud vendors have been promoting self-development plans for a long time; however, with the exception of Amazon, most vendors still focus on third-party supply, especially AI data center scenarios, with Nvidia (BlueField, ConnectX) and Broadcom (Thor) as the core, and AMD's Pensando is catching up fast. As an alternative to domestic production in the domestic market, some cloud service providers such as Huawei, Alibaba, and Baidu promote self-research, while third-party vendors cooperate to provide products to cloud vendors and operators; moreover, some third-party vendors are technologically advanced in terms of full-function DPU. Deep customer bonding by third party vendors guarantees the lower revenue limit, while their technological leadership may further expand customers and open up revenue ceilings.
Risk factors:
Macroeconomic development fell short of expectations; AI capital expenditure fell short of expectations; domestic substitution progress fell short of expectations; cloud vendor procurement fell short of expectations; and the competitive landscape of the industry deteriorated.