"When it comes to new technologies, we often overestimate their short-term benefits but underestimate their long-term impact."

In December 2023, during a public address, Xiao Qiyang, the founder and CEO of Yunbao Intelligent, delved into the topic of tech bubbles. "The same holds true for the internet and smartphones, and AI is no different. The advancement of new technologies heavily relies on the creation of bubbles and the infusion of capital," he remarked.

Xiao further emphasized, "AI will undergo a complete transformation in the next decade."

Now, Yunbao Intelligent stands on the brink of its own significant transformation.

On June 30, with aspirations to become the first domestic DPU listing, Yunbao Intelligent submitted its IPO application to the ChiNext board, aiming to raise 3.035 billion yuan. After deducting issuance fees, the proceeds are earmarked entirely for next-generation DPU R&D and industrialization projects.

However, during its IPO journey, the close ties between financial controller Lu Yao and sponsor CITIC Securities have raised significant market concerns regarding the independence of the intermediary.

To put it simply, Lu Yao, the financial controller of Yunbao Intelligent, spent 21 months within the CITIC Securities system before departing in November 2025 and joining Yunbao Intelligent just a month later.

Concurrently, CITIC Securities also took on the responsibility of providing IPO tutoring and sponsorship services to Yunbao Intelligent, with the signatory sponsor representative having previously worked alongside Lu Yao in the same securities firm environment.

In essence, this scenario involves the former employer sponsoring the IPO while a former employee oversees the finances. As the gatekeeper of oversight, CITIC Securities' independence in this IPO has come under scrutiny from the market.

Secondly, Lu Yao allegedly became involved in and supervised IPO planning activities even before officially joining Yunbao Intelligent.

Public event records indicate that on October 25, 2025, Lu Yao, an employee of CITIC Capital at a public event, was identified as being responsible for the listing and capital operations of Yunbao Intelligent, a domestic DPU unicorn. This occurred more than two months before Yunbao Intelligent officially initiated its IPO tutoring registration process.

This suggests that Lu Yao may have been overseeing Yunbao Intelligent's undisclosed IPO project prior to leaving his previous position.

Public regulatory information reveals that as early as September 2023, the Shenzhen Stock Exchange issued a written warning to Lu Yao, then a sponsor representative at China Merchants Securities.

The warning stemmed from Lu Yao's failure to adequately verify several issues during the sponsorship of the Dacheng Precision IPO project, including irregular financial internal controls, cross-period revenue recognition, and inaccurate accounting entries. These lapses led to the project's swift withdrawal after being selected for on-site inspection.

Now, as the financial controller of Yunbao Intelligent, whether Lu Yao can avoid repeating past mistakes has become a focal point of concern for the market.

Setting aside compliance controversies, Yunbao Intelligent has already established itself at the forefront of domestic DPU development based on its achievements.

Revenue surged from 173,200 yuan in 2023 to 370 million yuan in 2025, marking a more than 2,000-fold increase over two years. Its valuation soared past 14 billion yuan, supported by dozens of top-tier institutions, including Tencent, Sequoia China, IDG Capital, and Temasek.

However, following the release of the prospectus, the market perceived a different reality.

Over three years, Yunbao Intelligent's operating cash flow remained negative, and accounts receivable skyrocketed from zero to 180 million yuan.

Behind this, a chip unicorn once courted by institutions now faces commercial dilemmas far more intricate than technical challenges on the eve of its IPO.

Xiao Qiyang, the founder and CEO of Yunbao Intelligent, is widely recognized in the industry for his exceptional background.

At the age of 24, he earned a PhD in Electrical Engineering from Stanford University, solving a theoretical problem in AI that had remained unsolved for over 30 years in his doctoral dissertation. His work earned praise from Marvin Minsky, the "father of artificial intelligence." Prior to co-founding RMI in Silicon Valley in 2002, which later facilitated its acquisition by Broadcom, he taught at the University of California, Irvine, and MIT.

In 2020, this serial entrepreneur in his fifties embarked on a second venture in Shenzhen, targeting the DPU—the third main chip for data centers dominated by NVIDIA.

In AI data centers, CPUs are akin to master chefs, adept at preparing a few complex dishes; GPUs are like vegetable cutters, capable of slicing 10,000 ingredients simultaneously.

DPUs, on the other hand, are the porters, responsible for transporting the prepared ingredients in and out. In short, DPUs handle the mundane tasks to free up core computing power for critical operations.

In China's semiconductor startup landscape, Yunbao Intelligent's revenue trajectory is nothing short of remarkable.

In 2023, the company recorded just 173,200 yuan in revenue, a negligible amount. In 2024, revenue jumped to 36.3557 million yuan, a 210-fold increase. By 2025, it reached 370 million yuan, marking another significant leap.

According to a Frost & Sullivan report, based on 2025 revenue, Yunbao Intelligent ranks first among domestic independent manufacturers in China's full-function DPU market.

However, the income statement tells a different story: as revenue rises, so do losses.

Yunbao Intelligent's net loss was 667 million yuan in 2023, 610 million yuan in 2024, and widened to 1.189 billion yuan in 2025—totaling 2.466 billion yuan over three years.

The majority of these funds were allocated to R&D.

From 2023 to 2025, Yunbao Intelligent's cumulative R&D expenses reached 1.758 billion yuan, 4.32 times its total revenue over the same period.

This ratio is not uncommon among Chinese chip design companies.

As a core chip for data centers, DPUs must simultaneously handle network protocol processing, data encryption/decryption, storage virtualization, RDMA low-latency transmission, and other technical stacks, posing extremely high R&D barriers.

NVIDIA's first-mover advantage in DPUs was built on its $6.9 billion acquisition of Mellanox in 2020.

But what's truly intriguing is the company's projected timeline for profitability.

Yunbao Intelligent expects to become profitable as early as 2028. This forecast is based on a series of assumptions, including continuous product iteration, successful customer expansion, and R&D investment dilution through economies of scale.

However, the prospectus fails to answer a crucial question: what revenue level is needed to break even?

Using the 2025 non-GAAP loss of 562 million yuan as a baseline, assuming constant gross margin and expense ratios, the required revenue scale would far exceed the current 370 million yuan.

Yet China's DPU market is far from a blue ocean—NVIDIA's BlueField series holds over 60% of the global market share, Intel's IPU follows closely, and domestic players like Inspur, Corigine, and Nebulon are also gaining traction in niche segments.

Yunbao Intelligent's financial paradox essentially mirrors the broader reality of domestic chip startups: technological breakthroughs can be achieved in four years, but commercial viability may take a decade.

Yunbao Intelligent relies heavily on major clients.

The prospectus discloses that Tencent was its largest end customer in 2025, accounting for 95.22% of its annual revenue through direct and indirect sales.

In other words, out of 370 million yuan in 2025 revenue, over 352 million yuan came from Tencent.

From 2023 to 2025, sales to the top five customers accounted for 100%, 99.35%, and 92.87% of revenue, respectively. This means nearly all of Yunbao Intelligent's revenue comes from fewer than five customers, with Tencent as the absolute mainstay.

Understanding this dynamic requires examining Yunbao Intelligent's shareholding structure.

The prospectus shows that before the IPO, Tencent held a 19.78% stake through multiple entities, making it the largest shareholder and nearly the sole customer.

For this six-year-old DPU chip design company in the commercialization phase, relying on Tencent's support is the most pragmatic survival strategy. Its DPU products can achieve large-scale deployment and multi-scenario commercialization within Tencent, while Tencent's bulk purchases provide a pathway to commercialization.

For Tencent, its data centers face massive expansion and computing efficiency improvement needs. Validating and mass-producing Yunbao Intelligent's DPU products in Tencent Cloud's actual business scenarios is itself proof of technical strength.

As Xiao Qiyang stated in an interview, in large-scale AI computing scenarios, each GPU must connect to thousands of others, and each GPU requires a dedicated DPU.

Tencent's cloud infrastructure provides the ideal testing ground for Yunbao.

The issue lies in the information Yunbao Intelligent conceals.

The prospectus reveals that in 2024, Tencent-related entities were its largest customer, accounting for 71.31% of direct sales. By 2025, Tencent's share of direct sales plummeted to 5.04%, with more sales shifting to indirect models.

Meanwhile, Yunbao Intelligent discloses that its end customers also include China Mobile, Sina, and China Southern Power Grid, to demonstrate the applicability of its DPU products in critical infrastructure sectors like telecommunications and energy.

However, the prospectus does not disclose trends in revenue contribution from non-Tencent customers.

If the four largest end customers other than Tencent collectively accounted for less than 5% of revenue in 2025, then diversifying its customer base remains merely a strategic narrative without substantial commercial support.

The core question is whether Yunbao Intelligent can survive independently if Tencent's strategic needs change.

Currently, global cloud giants are accelerating their self-developed chip efforts—AWS has Nitro and Graviton, Alibaba Cloud has Yitian 710 and CIPU, and Google has TPU.

If Tencent's demand fluctuates or shifts to self-developed solutions, Yunbao Intelligent must prove it can independently face the market without being merely an extension of Tencent's computing infrastructure.

On October 29, 2025, at NVIDIA's GTC conference, Jensen Huang unveiled the Rubin architecture-based next-generation DPU—BlueField-4. The product integrates 126 billion transistors, 64 Arm Grace cores, and supports 800G network throughput and ConnectX-9 interconnect technology.

From 400G to 800G, from standalone chips to deep integration with NVIDIA's GPU-NVLink-CUDA ecosystem, the technological gap is widening at a visible pace.

This is not merely a performance competition but an ecosystem lock-in issue.

Huang specifically noted during the launch that one core application scenario for BlueField-4 is accelerating KV cache processing, directly serving multi-round conversational inference in large language models.

This means NVIDIA's DPUs are no longer just network cards offloading tasks from CPUs but are deeply coupled with GPUs, CUDA software stacks, and NVLink interconnect protocols.

For customers using NVIDIA GPUs, adopting NVIDIA's DPUs becomes the path of least resistance. This chip-interconnect-software trifecta bundling strategy is creating powerful network effects in the cloud infrastructure market.

Returning to Yunbao Intelligent's prospectus: the company plans to allocate 1.174 billion yuan of the 3.035 billion yuan raised to R&D for next-generation high-performance DPU core technologies.

But the prospectus fails to address a more urgent question: when NVIDIA's BlueField-4 enters mass production in the second half of 2026, can Yunbao Intelligent's next-generation products simultaneously iterate to 800G? How will it ensure that customers like Tencent and China Mobile won't switch due to performance gaps?

Upstream in the supply chain, global cloud giants are pursuing a clear vertical integration path.

AWS's self-developed Nitro DPU has undergone multiple iterations, Alibaba Cloud launched CIPU and announced its full replacement of traditional network cards in 2024, and Google's TPU evolves in tandem with its self-developed network chips.

These three also happen to be the world's largest cloud service providers. When end customers start building their own chips, the market space for independent DPU chip designers will continue to shrink.

Xiao Qiyang's response to this query is unequivocal: "Our objective extends beyond merely achieving the pinnacle of singular performance; rather, it is to forge the most economically viable domestic DPU."

When juxtaposed with its rivals, Cloudflare Intelligence places a greater emphasis on cost-effectiveness and adaptability to diverse scenarios. This approach caters to the unique requirements of domestic cloud service providers amidst the backdrop of indigenous innovation and domestic substitution efforts.

However, the realization of this vision hinges on a multitude of factors. These encompass the degree of policy constraints imposed on NVIDIA within the Chinese market, the velocity at which domestic cloud service providers undertake indigenous innovation and substitution, as well as Cloudflare Intelligence's own pace of product iteration.

For Cloudflare Intelligence, the sole aspect within its control is the tempo of its Initial Public Offering (IPO). By endeavoring to go public and secure funding at the earliest opportunity, coupled with expediting research and development cycles, it can establish a foothold before the window of opportunity closes.