CITIC SEC: DPU undertakes network and storage management, with a simultaneous rise in volume and price in the AI era.

CITIC SEC released a research report stating that the DPU takes on infrastructure tasks such as server networking and storage management, widely used in eight-card servers, ultra-node servers, high-performance general-purpose servers, and storage servers. In the AI era, its usage increases with the number of computing power chips, and its price rises as communication speeds enhance; comparatively, it has a relatively large market among all interconnect devices. Currently, domestic substitution of DPU is gradually advancing; among these, the emerging independent third-party suppliers are experiencing rapid revenue growth due to deep integration with customers and relatively leading technological capabilities. It is recommended to pay attention to leading third-party suppliers. CITIC SEC's main points are as follows: Interconnection Level: There are often different levels of interconnection networks within computing infrastructure to achieve various interconnection objectives, creating different network planes to distinguish traffic. 1) Within the Cabinet: GPU-GPU (Scale-up): Achieving high-speed interconnection of GPUs within a cabinet. CPU-Peripheral: Achieving interconnection between CPU and peripherals for management and scheduling. 2) Between Cabinets: GPU-GPU (Scale-out): Achieving high-speed interconnection of GPUs across cabinets. CPU-Storage (Business Front-End Network): Managing data from external storage. GPU-GPU (Layer 2 Scale-up): Scale-up network for cross-cabinet groups. Among these, the inter-cabinet GPU-GPU (Scale-out network) and CPU-Storage (Front-End network) require networking devices, corresponding to the inflation component as DPU. DPU: Enables connection between servers and switches and offloads some tasks from the CPU. DPUs are divided into AI DPU and full-function DPU. AI DPU connects to servers via PCIe, packaging data from the host side into network frames and connecting to switches through Ethernet/IB interfaces for high-speed interconnection; in the AI era, it is primarily used for cross-cabinet GPU Scale-out networks and typically has hardware acceleration capabilities such as RDMA. Compared with AI DPU, the full-function DPU adds CPU cores and can undertake networking, storage, security, virtualization, and other tasks of the CPU, adapting to a disaggregated computing and storage architecture; full-function DPUs generally have dedicated DDR and can integrate PCIe Switch as a connection hub; in the AI era, they are primarily used for connecting intelligent computing servers to storage servers and external business Front-End networks. Currently, DPU has become one of the pillars of data centers, alongside CPU and GPU, collectively referred to by NVIDIA as the "three major chips of data centers." Market Space: It is estimated that the domestic demand-supply market will approach 100 billion yuan by 2028, nearly doubling compared to 2026. Volume: DPUs are mainly applied in four types of scenarios: eight-card servers, ultra-node servers, general-purpose servers, and storage servers. According to various product plans: 1) In eight-card servers, XPU:AI DPU=1:1, XPU:Full-function DPU=8:1. 2) In ultra-node servers, XPU:AI DPU=1:1, XPU:Full-function DPU=4:1. 3) In general-purpose servers, high-performance general-purpose servers (used by cloud service providers/internet manufacturers/carriers) usually require offloading from the CPU and typically include one full-function DPU. 4) In storage servers, high-performance storage servers required for large-scale intelligent computing generally incorporate one AI DPU. Price: As rates gradually improve, full-function DPU is impacted by storage price surges. Overseas situation: 1) For AI DPU, the purchasing price for the overseas ConnectX-7 (400Gbps rate) from major manufacturers is about $1,000, while for ConnectX-8 (800Gbps rate) it is about $1,500, indicating a 50% price increase with generational upgrades. 2) For full-function DPU, the estimated price before the surge for overseas BlueField-3 (400Gbps rate) was around $2,000 to $3,000, surging to nearly $4,000 afterward, with BlueField-4 prices still unknown. Domestic situation: Prices for domestically comparable products usually come with a discount of 70% to 80% compared to mainstream international products. Space Calculation: By 2028, the domestic demand-supply AI DPU market space may reach 41.4 billion yuan, and the full-function DPU market space may reach 49.8 billion yuan. By 2026, domestic AI DPUs and full-function DPUs will be primarily at a 400Gbps rate, and by 2028, they will fully upgrade to 800Gbps rates for both. Based on assumptions regarding the penetration rate of domestic computing power chips, ultra-node usage, general-purpose and storage server usage, and expected prices for domestic DPUs, it is estimated that the total space for DPUs in China will reach 91.2 billion yuan by 2028. Pattern: Domestic substitution is in full swing, and there is optimism about the prospects for independent third-party suppliers. From the overseas market perspective, most cloud manufacturers have long promoted self-research programs; however, apart from Amazon, which maintains full self-supply, most manufacturers still primarily rely on third-party supplies, especially in AI data center scenarios, centered around NVIDIA (BlueField, ConnectX) and Broadcom (Thor), with AMD's Pensando catching up. In the domestic market, some cloud service providers, like Huawei, Alibaba, and Baidu, are advancing self-research, while third-party manufacturers provide products to cloud manufacturers and operators through partnerships; moreover, some third-party manufacturers are leading in full-function DPU technology. The deep integration of third-party suppliers with their clients ensures a revenue bottom line, and their technological lead may further expand the customer base and open up revenue ceilings. Risk Factors: Macroeconomic development is below expectations; AI capital expenditures are below expectations; domestic substitution progress is not as expected; cloud manufacturers outsourcing is less than anticipated; industry competition dynamics may deteriorate.