China’s AI infrastructure race is usually described in terms of graphics processors, data centers, and power supply. Mucang is betting that another layer will become equally important: the network chips that allow thousands of accelerators to exchange data fast enough to work as a useful cluster.
Wuxi Muchuang Integrated Circuit, known as Mucang, announced that it had completed a Series B financing round worth hundreds of millions of yuan. The company says the money will support research, industrialization, and deployment of 400G and 800G smart network-interface chips for scale-out clusters, as well as high-speed GPU interconnect chips for scale-up systems.
The report by Ijiwei describes an investor group that includes strategic backers Muxi and Lenovo Capital, alongside local state-linked funds and other investors. The exact round size was not disclosed, so it is more accurate to describe the financing in the broad range published by the company’s coverage rather than attach a precise number.
Mucang’s proposition rests on a simple fact about large-model computing: adding more accelerators does not automatically produce more useful computing capacity. A cluster must move parameters, activations, checkpoints, and other data between devices. If that movement is too slow or unreliable, expensive chips can spend too much time waiting for one another.
AI Clusters Need Fast Data Movement, Not Only Fast Chips
A modern training or inference cluster is a distributed system. Large workloads are split across many processors, which must remain synchronized while they exchange enormous volumes of data. The processors receive most of the public attention because they execute the calculations. Yet the interconnect determines how efficiently the system behaves as it grows.
Mucang is targeting this layer with smart network-interface cards and GPU interconnect technology. The company says its N30 400G smart NIC supports RoCEv2 and GPU Direct RDMA, technologies designed to move data between devices with low overhead. It also says it is working with domestic GPU firms on a high-speed multi-card interconnect intended to fill a gap in local AI clusters.
Those are company and report claims, not independent performance findings. The technical direction nevertheless explains why investors would focus on an apparently less glamorous part of the AI stack. In a cluster with many accelerators, bandwidth and latency can shape the real value of each processor. Faster networking can improve utilization, reduce waiting, and make a system more practical for large-model training or serving.
China’s effort to build domestic AI capacity has made this component layer strategically important. EastFrontier’s report that state-funded data centers were ordered to use domestic AI chips illustrated how local substitution is becoming a system-level issue. Building a domestic accelerator is one part of that challenge. Building the networking, software, storage, and packaging around it is another.
A Reconfigurable Architecture Offers a Different Route to Scale
Mucang describes its technology as a reconfigurable architecture that combines time-division multiplexing and spatial parallelism. The company says this approach reduces reliance on manufacturing processes below 12 nanometers while still allowing it to address high-speed networking needs. Such claims should be read as the company’s explanation of its engineering strategy, not as proof that it has matched every imported alternative.
The idea is important because China’s AI hardware ecosystem cannot assume it will always have the newest manufacturing tools or unrestricted access to the highest-end foreign components. A firm that can improve capability through architecture, design, and software control may be able to compete without following the most advanced-process route at every stage.
Ijiwei reported that Mucang has developed network chips spanning from gigabit products to 400G systems, and that it has built up near-million-chip cumulative shipments. That provides a commercial backdrop for the Series B, although the report does not independently detail customer names, revenue, or current shipment volumes by product line. The safer conclusion is that Mucang is presenting itself as a firm with both an existing networking portfolio and an ambition to move upward into AI-cluster interconnects.
The company has also associated itself with co-packaged optics, a technology that seeks to bring optical connections closer to switching and compute silicon. Mucang and Lankun Micro announced a domestic CPO smart NIC earlier this year, according to the reports. That is a long-term infrastructure bet. As bandwidth rises, traditional electrical connections can become more difficult to scale efficiently, making optical integration a potential route to lower power use and faster links.
The Financing Tests China’s Full-Stack Compute Strategy
The Series B is significant less for its exact size than for what it says about where capital is moving. Investors are not only backing companies that make an AI model or a GPU. They are also looking at the connective tissue that turns separate processors into an operating system for large-scale computation.
Mucang’s financing comes as Chinese companies are searching for practical ways to scale workloads under constraints on advanced chips and manufacturing equipment. That search involves software optimization, cluster design, domestic accelerators, and high-speed data movement. Any weak layer can limit the value of the others.
The startup still has much to prove. It must turn stated product plans into commercially deployed 400G and 800G equipment, establish compatibility with domestic GPUs and servers, and demonstrate reliability in environments where an interconnect failure can interrupt an entire training job. It will also face established foreign and domestic competitors in a market that demands rigorous validation.
But the focus itself is revealing. China’s AI infrastructure conversation is moving beyond whether it can obtain enough chips. It is also becoming a question of whether those chips can be connected, coordinated, and utilized efficiently. That is why the recently published Enflame IPO timetable matters as context: accelerator supply and cluster networking are complementary, not competing, parts of the same buildout. Mucang’s Series B is a financing event inside that broader shift, where the network increasingly determines how much of an AI cluster’s theoretical power can be turned into work.