DPU（Data Processing Unit，数据处理单元）是一种面向数据中心与智算基础设施的数据面专用处理器，通过支持弹性网络、云存储、虚拟化、安全、数据管控与加速、低时延传输等系统级功能，卸载计算芯片（CPU/GPU）负担、充分释放和优化算力，显著提升计算集群整体能效。
英伟达最早于2020年提出DPU是与CPU、GPU并列的数据中心“三大主力芯片”之一。在2026年CES大会，英伟达进一步指出：DPU是每个计算节点的标准配置，在其最新发布的Vera Rubin平台中6颗主力芯片中包含2颗DPU芯片，DPU逐渐成为全球科技竞争的战略制高点。
DPU为AI智算和大规模数据处理而生：在云计算场景，DPU卸载CPU原本负担的弹性网络、云存储、虚拟化、安全以及数据管控等复杂任务，充分释放CPU计算资源、减少“数据中心税”；在AI智算场景，一方面，DPU通过GDR、GDA等技术和多路径拥塞管控算法，显著降低GPU集群Scale-out域通信时延，提升集群综合算力效率；另一方面，DPU通过管理键值缓存（KV Cache），为GPU内存与传统存储之间构建一个独立、高速、可共享的“记忆层”，加速大模型上下文数据的管理与访问、显著减少数据搬移开销，从而突破传统GPU系统的存储瓶颈。
在云计算基础设施持续扩张与人工智能算力需求快速增长的共同驱动下，DPU市场规模保持高速增长态势。2025年中国DPU市场规模达到约人民币500亿元，预计到2030年将增长至超人民币1,200亿元，是云计算与AI基础设施领域中增长最为迅速、成长空间最为广阔的细分市场之一。
从竞争格局来看，中国DPU市场整体呈现较为集中的态势，国际头部独立DPU厂商占据主要市场份额。在中国独立全功能DPU市场中，英伟达凭借长期的芯片架构积累、成熟的数据面处理能力以及完善的软硬件生态体系，占据市场首位；云豹智能排名第二，并在本土独立全功能DPU厂商中位列第一。
DPU (Data Processing Unit) is a data-plane dedicated processor designed for data centers and intelligent computing infrastructure. By supporting system-level functions such as elastic networking, cloud storage, virtualization, security, data management and acceleration, and low-latency transmission, it offloads burdens from computing chips (CPU/GPU), fully releases and optimizes computing power, and significantly enhances the overall energy efficiency of computing clusters.
NVIDIA first proposed in 2020 that DPU is one of the three primary data center chips alongside CPU and GPU. At the 2026 CES, NVIDIA further stated that DPU is a standard configuration for every computing node. In its latest Vera Rubin platform, two of the six primary chips are DPU chips, and DPU is gradually becoming a strategic high ground in global technology competition.
DPU is purpose-built for AI intelligent computing and large-scale data processing. In cloud computing scenarios, DPU offloads complex tasks originally borne by the CPU, including elastic networking, cloud storage, virtualization, security, and data management, thereby fully releasing CPU computing resources and reducing the “data center tax.” In AI intelligent computing scenarios, on the one hand, DPU significantly reduces communication latency in the Scale-out domain of GPU clusters and improves overall cluster computing efficiency through technologies such as GDR and GDA and multipath congestion control algorithms; on the other hand, by managing key-value cache (KV Cache), DPU builds an independent, high-speed, and shareable “memory layer” between GPU memory and traditional storage, accelerating the management and access of large model contextual data and significantly reducing data movement overhead, thereby breaking through the storage bottlenecks of traditional GPU systems.
Driven by the continuous expansion of cloud computing infrastructure and the rapid growth of artificial intelligence computing demand, the DPU market has maintained a high growth trajectory. In 2025, the China DPU market reached approximately RMB50 billion, and it is expected to grow to over RMB1,20 billion by 2030, making it one of the fastest-growing and most promising segments within cloud computing and AI infrastructure.
From a competitive landscape perspective, the China DPU market as a whole exhibits a relatively concentrated structure, with leading international independent DPU vendors accounting for the majority of market share. In the China independent full-function DPU market, NVIDIA ranks first, supported by its long-term accumulation in chip architecture, mature data-plane processing capabilities, and a well-established software and hardware ecosystem; JaguarMicro ranks second and holds the leading position among domestic independent full-function DPU vendors.
