HyT Capital Portfolio | "China's First DPU IPO" — Yunbao Intelligent's IPO Application Accepted by Shenzhen Stock Exchange
Editor's Note: Recently, the IPO application of Yunbao Intelligent, a portfolio company from HyT Capital's angel round, has been formally accepted by the Shenzhen Stock Exchange, marking a key step toward becoming "China's first DPU IPO." As one of the few independent vendors in China with full-scope DPU R&D, productization, and scaled deployment capabilities, Yunbao Intelligent's products have achieved commercial deployment of over 100,000 units in real-world data center scenarios.
The following article is sourced from the WeChat public account "雷锋网" (leiphone) by Bao Yonggang.
Over the past six years, domestic GPU companies have ridden the AI wave, with valuations continuously resetting upward — while DPUs have been largely overlooked.
That doesn't match the industrial reality.
After Nvidia completed its acquisition of Mellanox in 2020, it had already laid out its "GPU + CPU + DPU" three-chip strategy. Over the past few years, Nvidia has continued to strengthen its networking capabilities. When Jensen Huang showcased a "six-chip combination" at CES 2026, four of the six chips were networking-related.
An increasingly clear trend is emerging: the bottleneck in AI infrastructure is shifting from computing power itself to networking and scheduling.
Especially with the arrival of the Agent era, AI systems are moving from training to high-frequency inference and continuous operation, and GPU utilization is increasingly dependent on network efficiency. DPUs are gradually evolving from a data center optional component to a critical role in AI infrastructure.
But here's a thought-provoking question: if Nvidia had already bet on DPUs years ago, why has the industry still underestimated them for the past six years?
It wasn't until Yunbao Intelligent's IPO prospectus — as it pushes to become China's first listed DPU company — was filed and made public after being accepted by the Shenzhen Stock Exchange that the market began to realize that a full-featured DPU capable of supporting low-latency, high-bandwidth, and high-performance data scheduling might be the most underestimated piece of the AI infrastructure puzzle.
Over the past few years, the focus of competition in the AI industry has always been GPU. Larger model parameters, stronger single-card performance, and more expensive HBM have captured the industry's attention almost entirely.
But as model scale continues to grow, AI clusters have moved from thousands of cards to tens of thousands. More and more companies are discovering that GPU is no longer the scarcest resource in AI systems. What is truly becoming expensive is low latency, high bandwidth, and data flow efficiency.
Many algorithm engineers feel this firsthand. In the current AI infrastructure landscape, single-node compute power is relatively easy to obtain, storage capacity comes second, and the most difficult and expensive part is bandwidth and low latency. Especially in large-scale training and inference scenarios, GPU utilization is often suboptimal. Even with deep optimization, system bottlenecks frequently appear in networking and data scheduling.
This is also why Nvidia has continuously strengthened its networking capabilities over the past few years. The signal it sends is very clear: competition in AI infrastructure is shifting from single-chip performance to system-level efficiency.
In this process, the role of DPUs has also changed.
In the past, in the era of cloud computing dominated by general-purpose CPUs, DPUs mainly handled infrastructure task offloading for networking, storage, and security, and were seen as auxiliary chips in data centers. In the Agent era, as AI infrastructure shifts from training to high-frequency inference, resource orchestration, and continuous scheduling, DPUs are becoming a system-level core node connecting compute, networking, and storage.
Especially in Scale-Up scenarios, DPUs can optimize memory sharing and data flow between CPUs and GPUs within a single node, reducing data movement latency and improving heterogeneous compute collaboration efficiency. In Scale-Out scenarios, DPUs handle data scheduling and network offloading across large clusters, directly impacting GPU utilization.
The explosion in inference demand has further amplified the importance of DPUs.
As the context window of large models continues to grow, GPU memory capacity has become a critical bottleneck for inference costs. DPUs can expand the effective available memory capacity of AI systems without increasing the number of GPU hardware units.
At GTC 2026, Jensen Huang demonstrated the capability evolution of next-generation DPUs in KV-Cache tiered storage. In the latest Vera Rubin system, the BlueField-4 series DPU handles KV-Cache management and hardware acceleration, building a "warm data layer" between GPU high-speed HBM and external storage. It dynamically allocates 16TB of dedicated context space per Rubin GPU, breaking through the hardware bottleneck of context processing and reducing per-token inference costs by 90%.
DPUs are shifting from a data center optional component to a critical piece of AI infrastructure — and the DPU market is rapidly expanding as a result.
According to a 2026 special report by Frost & Sullivan, the global DPU market has grown from RMB 64.99 billion in 2021 to RMB 196.49 billion in 2025, and is expected to reach RMB 436.24 billion by 2030. The Chinese market alone is projected to hit RMB 129.09 billion by 2030, making it one of the fastest-growing segments in AI infrastructure.
But here's a thought-provoking question: why hasn't this market become a real industry focus over the past few years?
As models grow larger, networking increasingly becomes the bottleneck of the entire system. With the arrival of the Agent era, the demands on AI infrastructure are changing. Systems are shifting from training ever-larger models to enabling higher-frequency, lower-cost, and longer-running inference. In this process, the importance of system capabilities such as resource orchestration, task scheduling, KV-Cache management, and storage pooling is rising rapidly.
This shift is driving demand for both CPUs and DPUs in AI systems.
When CPUs take on more inference scheduling and system management tasks, infrastructure functions like network offloading, security isolation, virtualization, and storage acceleration need to be offloaded to DPUs. So the Agent era not only elevates the importance of CPUs, but also amplifies the value of DPUs.
But there is another, more important reason why DPUs have been underestimated: very few companies can actually build a full-featured DPU.
A DPU is not just a more complex network card. It involves networking, compute, storage, virtualization, security isolation, and more — essentially a system-level chip. What truly determines the moat of a DPU is not just the chip itself, but also data-plane processing capability, the software stack, cloud-native adaptability, and stability in large-scale data centers.
DPU Functional Diagram, Source: China Academy of Information and Communications Technology (CAICT)
Even for a company as strong as Nvidia, the DPU product evolution cycle has been a long one. Nvidia spent heavily to acquire Mellanox to fill its DPU capabilities. Mellanox's early BlueField series (BF1, BF2) were not widely adopted in the market. It was not until BF3 — developed after Nvidia's acquisition of Mellanox — that a truly successful DPU product emerged.
In China, even fewer companies possess full-featured DPU R&D and mass-production capabilities. Apart from Huawei, Yunbao Intelligent is one of the few independent vendors that has achieved productization and scaled deployment,and is the only one capable of delivering a full-featured DPU product at 400Gbps, comparable to Nvidia's BF3.
Yunbao Intelligent's ability to break into this high-barrier sector is closely tied to its team background.
Founder Xiao Qiyang received his Ph.D. in Electrical Engineering from Stanford University at the age of 24. His doctoral dissertation tackled a classic theoretical problem that had remained unsolved in the AI field for over thirty years. His research findings were compiled into the book Discrete Neural Computation: A Theoretical Foundation, which received a foreword endorsement from Marvin Minsky, the father of AI. This groundbreaking theory in early AI neural networks earned him the Young Investigator Award from the National Science Foundation. He later focused on networking and distributed computing research and served as an endowed-chair associate professor at MIT. Before founding Yunbao Intelligent, he had already gained experience in the large-chip startup space — his previous venture, a networking processor company in Silicon Valley where he was a co-founder, was later acquired by Broadcom for USD 3.7 billion. In addition, Yunbao Intelligent's core team comes from Broadcom, Intel, ARM, HiSilicon, Alibaba, and other companies, covering networking chips, cloud computing, and system architecture.
A strong team enabled Yunbao to design a system-level DPU chip. According to sources close to the company, its first-generation DPU product achieved customer deployment and mass production at the A0 tape-out stage — a rare feat in the high-end chip space, both in China and globally. In other words, Yunbao Intelligent was able to go directly into real-world data center environments after its very first tape-out.
Launching a product is only the first step toward success for a large-chip company. Large-scale deployment is the most stringent test.
Yunbao Intelligent's self-developed DPU product is the first in China to reach 400Gbps. More important than the spec itself is the fact that the product has already entered real-world data center environments, with over 100,000 units deployed at scale in leading customers' scenarios, supporting applications such as high-performance computing, storage, and network offloading.
But over the past few years, Yunbao has maintained a relatively low profile — both in product development and customer deployment. As a result, although DPUs have always existed within the AI infrastructure ecosystem, they have rarely entered the public spotlight.
It was not until Nvidia continuously strengthened its networking strategy, the Agent era pushed AI systems into a new era of heavy scheduling, and Yunbao Intelligent began its push to become China's first listed DPU company, that the industry began to recognize that DPUs may be the most overlooked piece of the AI infrastructure puzzle.
The importance of DPUs is continuing to rise.
For companies with true full-featured DPU capabilities, the AI market represents not only new demand, but also greater room for capability spillover.
Leiphone has learned that Yunbao Intelligent will launch a DPU product specifically designed for AI networking scenarios this year, to further address the needs of the AI infrastructure market.
What determines the long-term value of a full-featured DPU company is not a single product, but the ability to consistently penetrate core infrastructure scenarios.
DPUs are already helping to improve GPU utilization, reduce system latency, and optimize overall resource efficiency in data centers, cloud computing, high-performance computing, and large-model inference scenarios.
DPU applications are also expanding into finance, telecommunications, energy, and other industries. In financial scenarios, DPUs can enhance the stability and security isolation of core trading systems; in the energy sector, they can support the digital scheduling needs of power grids and industrial systems.
For full-featured DPU vendors, continuous technology evolution capability is equally important.
DPU network interface speeds have already entered a phase where 400Gbps is now being deployed at scale, and 800Gbps is entering commercial use, and the demand for higher bandwidth and lower latency from AI infrastructure continues to rise rapidly.
It is also reported that Yunbao Intelligent's next-generation 800Gbps/1.6Tbps DPU products will soon be launched to further address the needs of next-generation AI data centers.
In the industry competitive landscape, DPU vendors themselves are also extremely scarce.
A recent report titled DPU Development Analysis Report published by the China Academy of Information and Communications Technology (CAICT) shows that in China's DPU market, Nvidia ranks first, leveraging its long-term accumulation in chip architecture, mature data-plane processing capabilities, and a robust software ecosystem. Yunbao Intelligent ranks second, and first among domestic DPU vendors, making it one of the few independent manufacturers in China that has achieved large-scale DPU mass production and commercial deployment.
As AI infrastructure increasingly emphasizes system-level capabilities, the competitive landscape in China's high-end networking chip market has begun to converge. In the future, there may be only two types of players: Yunbao, and everyone else.
This scarcity is also bringing Yunbao broader recognition.
At the Building a Strong Foundation for a Strong Nation – China's Manufacturing Achievements During the 14th Five-Year Plan exhibition, co-hosted by the National Museum of China and the Ministry of Industry and Information Technology, Yunbao's DPU series products were selected and displayed in the museum's "National Treasures of Technology" exhibition area, becoming one of the highlighted chip products in the exhibition.
Yunbao DPU Series Products Exhibited at the National Museum of China
As domestic AI infrastructure companies continue to attract capital market attention, the value of the DPU sector is also being reassessed.
Compared to the GPU market, which has already entered a highly crowded stage, there are still very few companies with true full-featured DPU R&D, mass production, and scaled deployment capabilities. Since its IPO application was accepted by the Shenzhen Stock Exchange, Yunbao Intelligent has moved one step closer to becoming "China's first DPU listed company," and its scarcity in the domestic AI infrastructure space has begun to draw attention from the capital markets.
Over the past two years, domestic GPU companies have generally achieved market capitalizations in the hundreds of billions of RMB after going public, prompting capital markets to reassess the long-term value potential of domestic DPU companies. As the most representative independent vendor in China's DPU sector, Yunbao Intelligent's scarcity — and its performance in the capital markets — leaves significant room for imagination on this multi-billion-yuan track.
DPUs have now been included in the national strategic framework. From a "chokepoint" technology to a "foundational infrastructure," DPUs have become the final piece in achieving independent and controllable compute infrastructure. Domestic DPU companies are now entering their own value revaluation cycle.
Back
Next article
