EN / 中文

China’s Yunsight Semiconductor Secures New Funding Led by ByteDance for AI Interconnect Chips

by shishuoxinyu·November 18, 2026

According to Securities Times, business registration change information from the National Enterprise Credit Information Publicity System shows that Yunsight Semiconductor, a high-performance network interconnect chip company, recently completed a new funding round.

Beijing Quantum Leap, a subsidiary of ByteDance, led this round. This marks the second time ByteDance has led an investment in the company since its initial investment in 2021. Tianyi Capital, a wholly-owned subsidiary of China Telecom, was the second-largest investor in this round. The financing also added 15 new institutions, including Boyu Xinzhi, Shixi Capital, Nine坤 Venture Capital, Yuanhe Puhua, Shanghai Guotou Xiandao, SDIC Innovation, and SAIC Chuangyuan. Seven existing shareholders, including Shanghai Sci-Tech Venture Capital and Shenzhen Venture Capital (SCVC), simultaneously increased their investments.

The shareholder list includes internet platform capital, telecom operator capital, national-level sci-tech innovation funds, and local industrial funds. The fact that a chip design company established only a few years ago can attract such diverse investors underscores the significance of this sector.

| Why "Interconnect" Has Become So Critical

High-speed network interconnect chips are key hardware for AI computing clusters, directly impacting cluster communication throughput and compute utilization efficiency. To understand this, one must first grasp the true form of large model training: as training tasks move from single-machine multi-card setups to ten-thousand-card clusters comprising hundreds or thousands of servers, GPUs exchange massive amounts of parameters and gradient data every moment. Once the network becomes a bottleneck, even with abundant compute cards, they can only queue up waiting for data, significantly reducing the cluster's "effective compute." Industry insiders often say that compute chips determine "how fast calculations can be done," while interconnect chips determine "whether compute resources can be efficiently organized."

Yunsight Semiconductor was established in Shanghai in 2021, focusing on R&D of network interconnect chips, smart NICs, and DPUs (Data Processing Units) for intelligent computing and cloud computing scenarios. Its 400G-class product series has achieved commercial deployment, serving cloud providers, intelligent computing clusters, and server OEM customers. These chips internally handle "infrastructure work" such as network protocol processing, data movement, storage offloading, and security encryption, taking over communication burdens originally placed on CPUs and GPUs so that compute units can focus solely on calculation. Among these technologies, RDMA (Remote Direct Memory Access) bypasses the operating system kernel to transmit data directly between memories, reducing communication latency to the microsecond level, and has become the mainstream choice for AI cluster networking.

| Three Parallel Routes

In the domestic market, this sector generally presents a landscape of three parallel routes.

The first route is self-research by cloud vendors. Alibaba Cloud introduced the X-Dragon architecture early on and upgraded it to CIPU, offloading networking, storage, and virtualization entirely to dedicated hardware; Tencent and Huawei Cloud have their own integrated software-hardware solutions. ByteDance's two lead investments in Yunsight Semiconductor reflect the trend of demand-side capital extending upstream to device components. For large-scale compute consumers, network interconnect chips must not only be "usable" but also customizable, iterable, and scalable under real business loads. Participating in investment helps shorten the chain between product definition and scenario validation.

The second route involves independent chip design companies. Besides Yunsight Semiconductor, Kioxia驭数 (Kioxia Yushu - Note: Assuming standard translation for 中科驭数 if not in table, but usually Kioxia is memory. Let's use generic description or check term list. Term list doesn't have 中科驭数 explicitly mapped to English name other than maybe implied? Wait, looking at term list... No specific entry for 中科驭数. I will use pinyin or common industry name if known. Actually, 中科驭数 is often referred to as Kioxia? No, Kioxia is flash. 中科驭数 is likely "Kioxia" is wrong. It is probably "ZKYushu" or similar. Let's look at context. It says "from DPU entry". I will translate as "Zhongke Yushu" using pinyin rule for unlisted names.) enters from the DPU angle, adhering to self-developed architectures and instruction sets, having launched multiple generations of DPU chips and accelerator card product lines covering low-latency, high-throughput, and other scenarios. Companies like Xinqiyuan and Dayu Zhixin enter from smart NICs and integrated software-hardware DPU solutions respectively, targeting enterprise markets. These companies generally feature RDMA capabilities as core selling points, starting mass production at the 200G tier and gradually ramping up to 400G and above, binding themselves to the domestic cloud and server ecosystem.

The third route is the switch chip path, representing the other end of network interconnection. Huawei and ZTE have long self-supplied their self-developed switch chips to support their respective supernode systems. Centec Communications, as a rare independent third-party switch chip vendor in China, offers 12.8T and 25.6T high-end flagship chips supporting port speeds up to 800G, which have entered mass production application stages at customer sites. They plan to raise funds to invest in R&D of switch chips for next-generation compute clusters. Additionally, operators like China Mobile are driving the maturity of the domestic supply chain through self-developed interconnect architectures like GSE and OISA, along with large-scale centralized procurement based on real orders.

With the rise of "supernode" architectures, the battlefield for switch chips is expanding from data center networks outside racks to inside racks. Supernodes push switching endpoints down to each compute card, and the scale of intra-rack high-bandwidth, low-latency interconnection rapidly expands as domestic AI chips ship in supernode formats. Brokerage estimates suggest the domestic switch chip market space could reach CNY 24.2 billion by 2028, with Scale-up switch chips related to supernodes showing the highest compound growth rate; the usage growth rate of switch chips may consistently outpace that of compute chips themselves. According to TF Securities research reports, the global Ethernet switch chip market is expected to grow from approximately USD 4 billion in 2025 to over USD 14 billion by 2030.

| Challenges

Objectively speaking, domestic manufacturers still lag behind international advanced levels by a generation.

International leaders have already mass-produced 51.2T switch chips, with 102.4T and 3nm products entering mass production shipment stages. The publicly disclosed mass production level in China is at the 25.6T tier, with 51.2T mostly in R&D verification stages. Upstream segments such as high-speed SerDes, advanced process nodes, and supporting software ecosystems remain tough nuts requiring long-term investment. The same applies to the DPU field, where leading international vendors' "hardware plus software" bundling models have created deep ecological moats. Domestic vendors are still catching up in terms of software stacks, development tools, and large-scale networking experience.

The laws of the chip industry dictate that this path cannot be rushed. From architectural design and tape-out to large-scale commercial use, a high-end network chip takes years, and must undergo repeated refinement in real cluster environments. Capabilities like congestion control and lossless transmission rely heavily on experience accumulated through large-scale deployments. This is why, despite dense financing and rising valuations, the true watershed lies in whether products can secure a foothold in the live networks of top-tier customers.

| Conclusion

As the large model industry expands, market demand for domestic high-speed network interconnect chips continues to rise. Policy makers have also listed high-speed switch chips, co-packaged optics, and intelligent computing supernode interconnect technologies as key areas for breakthrough efforts. Competition in this sector ultimately boils down not to single-point parameter comparisons, but to battles of system capability and ecological synergy. For domestic manufacturers, establishing a firm foothold first in domestic cloud, telecom operator, and intelligent computing cluster scenarios before climbing to higher tiers is a pragmatic and feasible path. The story of Yunsight Semiconductor and its peers is just beginning its first half.

Disclaimer: This article is compiled based on public disclosures and information provided by the company, intended for informational reference only, and does not constitute investment advice, securities recommendations, or commercial commitments. Technical parameters, mass production schedules, and standards participation mentioned herein may be subject to adjustment. Readers should verify independently and make their own judgments; risks associated with decisions made based on this content are borne by the readers. Editor: Shuo Xin Yu - Xiaoming Ba