China Shifts Focus to DPUs as AI Compute Demands System-Level Synergy

China's artificial intelligence infrastructure strategy is expanding beyond individual computing chips to address system-level network and data transmission bottlenecks. According to state media reports, relevant institutions project that new investment in China's national computing power network will reach 4 trillion yuan during the 15th Five-Year Plan period. As large AI model clusters expand, industry analysts emphasize that simply adding GPUs is no longer sufficient without high-performance Data Processing Units (DPUs) to offload infrastructure tasks and manage data flows between computing and storage nodes.

As large-scale AI models require increasingly massive data center clusters, single-chip performance gains are easily eroded by network latency, memory bandwidth limits, and data congestion. DPUs serve as the vital link between computing power and network efficiency, determining how effectively theoretical computing capacity translates into actual output across data centers.

China's computing power infrastructure is entering a phase where system-level coordination outweighs standalone chip performance, bringing Data Processing Units (DPUs) to the forefront of domestic technology planning, according to a report by People's Daily. As artificial intelligence clusters expand to train and run increasingly complex models, computing bottlenecks are shifting from raw calculation to data transfer, interconnect latency, and memory access. Industry analysis cited by People's Daily notes that simply increasing the number of GPUs is no longer sufficient to resolve compute bottlenecks. Without efficient data scheduling and network transport, raw computing power is dissipated through latency, congestion, and data movement overhead. In modern AI architecture, computing hardware is increasingly organized into a triad: CPUs handle general computing and overall system coordination, GPUs execute large-scale parallel processing, and DPUs manage high-speed data transmission between storage and network layers while offloading infrastructure tasks from host processors. Analysts compare the setup to an urban system where the CPU serves as administrative management, GPUs act as production centers, and the DPU functions as the intelligent transport and logistics hub. This architectural transition comes amid massive infrastructure spending. The national computing network has been integrated into China's strategic 'Six Grids' deployment alongside water, next-generation power grids, telecommunications, urban underground pipelines, and logistics networks. Relevant institutional estimates project that new investment in computing network construction will reach 4 trillion yuan during the upcoming 15th Five-Year Plan period. Policy support has steadily aligned with these structural needs. In October 2023, six government departments led by the Ministry of Industry and Information Technology issued the Action Plan for High-Quality Development of Computing Power Infrastructure, explicitly calling for accelerated research, development, and deployment of DPUs and lossless networking technologies. While domestic semiconductor efforts have established notable footprints in CPUs—such as Hygon, Phytium, and Huawei's Kunpeng—and in AI GPUs through vendors like Cambricon and Huawei's Ascend, DPUs remain a critical area requiring further domestic development to ensure resilient, end-to-end computing infrastructure.