HyT Capital Portfolio | Fangqing Technology Raises Over 500 Million RMB in Three Rounds Within Six Months
"An impressive shareholder lineup was quickly formed."
"暗涌Waves" exclusively learned that Shanghai Fangqing Technology, an AI chip and system architecture developer, has recently completed its Pre-A round financing. Industry investors include a major internet company, Xinlian Capital, and Hundsun Electronics, along with its affiliated industry fund Yima Capital. Financial investors include GF Xinde and a leading VC firm. Existing shareholders, including Shanghai Lingang Science and Technology Innovation Investment Management Co., Ltd. and 37 Interactive Entertainment Network Technology Group Co., Ltd., also made additional investments. The specific amount and valuation of the Pre-A round have not been disclosed.
This marks the third financing round for Shanghai Fangqing Technology within the past six months. Previous investors include: Xiaomi Strategic Investment, NIO Capital, Future Capital, Shanghai Lingang Science and Technology Innovation Investment Management Co., Ltd., HyT Capital, ECC Capital (Elevation China Capital), 37 Interactive Entertainment Network Technology Group Co., Ltd., ENC Digital Technology Co., Ltd., and Great Filter Venture, among other industrial and financial institutions. According to "暗涌Waves," the total amount raised in recent rounds has exceeded 500 million RMB. The funds will be used for core technology R&D, productization, and ecosystem and market expansion.
Fangqing Technology was founded in late 2022 and is registered in the Lingang area of the Shanghai Pilot Free Trade Zone. In August 2024, Liang Jun, former CTO of Cambricon and former chief architect of the HiSilicon Kirin SoC at Huawei, joined Shanghai Fangqing Technology as CEO.
Furthermore, Fangqing Technology has proposed a new technical direction: a distributed computing architecture that decouples "context-aware" and "context-free" components. Specifically, it decouples the Feed-Forward Network (FNN) and the Attention mechanism into two independent modules, assigning them to the most suitable hardware architecture for distributed processing — rather than connecting them in series within the same layer as in the traditional Transformer — thereby improving overall computational efficiency.
From a capital perspective, Liang Jun's name alone, in the chip field who represents a global capability to see through and connect the entire industry chain.
As a system architect with rare experience in both top-tier general-purpose SoC and high-performance AI chips in China's chip field, Liang Jun's career almost spans the entire golden twenty years of China's chip design — from catching up to breaking out. He spent 17 years at HiSilicon (Huawei), serving as the lead architect of the Kirin SoC chip, and personally created the Kirin 970, the world's first mobile SoC with an integrated NPU, which truly brought the concept of "on-device AI" to life. In 2017, he moved to Cambricon as CTO, leading the early-stage technical planning and product development of this AI chip unicorn, and experienced the full cycle from unicorn to IPO on Shanghai's STAR Market (Science and Technology Innovation Board).
Liang Jun's experience — "having seen the mountaintop and led the ascent himself" — is undoubtedly a scarce asset in today's primary market. The heavy bets that capital has placed on Fangqing Technology are not just backing the founder, but also backing the possibility of breaking the existing monopoly on computing power.
2025 is seen as a decisive year for the breakout of AI applications. With domestic large models like DeepSeek sparking a new wave, and giants such as ByteDance, Tencent, and Alibaba announcing hundreds of billions of RMB in AI infrastructure investment, the anxiety over computing power has never been more tangible. Yet within this wave, the logic of capital is also undergoing a subtle shift — from simply looking for an "alternative" to Nvidia, to gradually searching for "new species" that can break through the efficiency bottleneck of the Transformer architecture.
The traditional Transformer architecture connects the Attention mechanism (responsible for memory and context) and the FNN (responsible for logic and knowledge) in series within the same chip layer, resulting in significant efficiency waste. When you only want the AI to perform simple logical reasoning, it has to drag along a heavy memory module. When you need to process extremely long texts, the massive memory throughput demand clogs up the computing unit.
This is why entirely different solutions have emerged, such as GPGPU, RVV (RISC-V Vector Extension), compute-in-memory, and SRAM-based approach that emphasizes speed over capacity. Fangqing Technology's decoupled architecture is also a new attempt.
Liang Jun told "暗涌Waves" that the first priority of chip design should shift from pursuing the performance and integration of a single chip to pursuing a scalable system design. From a practical application perspective, Fangqing's new approach can liberate computing power from a single SoC-centric model, allowing edge devices such as smart glasses and earbuds to become computing nodes on equal footing with smartphones.
"When the first priority is changed to pursuing a scalable system design, it will change the existing approach to AI hardware design, open up opportunities to create new system form factors, and generate new markets."
This may also imply that Fangqing Technology's ceiling is not just that of an AI chip and system architecture design company, but rather that there exists a platform-level opportunity to define the next-generation AI hardware system. This kind of growth potential partly explains why, in such a short period of time, it has been able to quickly assemble a premium roster of shareholders, including major internet companies, hardware giants, automakers, state-owned capital, and leading VC firms.
When a top-tier architect returns to the stage with his brand-new thinking on the "post-Transformer era," capital that firmly believes "AI is the future" indeed has little reason to miss this bet.
Below is a conversation between "暗涌Waves" and Liang Jun, CEO of Fangqing Technology —
"Waves": Domestic GPU chip companies are extremely hot right now. Moore Thread's IPO surged, MXMACA is also in the IPO process, and Huawei has also shifted to GPU architecture. What is your view on the future development of GPU chips?
Liang Jun: The GPGPU architecture (note: general-purpose graphics processing unit, used for non-graphics computing tasks, emphasizing general-purpose computing capability) is designed for high concurrency and high throughput. To achieve low latency at the same time requires paying a greater price and cost.
At the same time, compared to global competitors, Chinese companies face more restrictions on the supply chain, including constraints on manufacturing processes, making the challenges greater. In other words, Chinese companies will have to pay a greater price.
The reason many domestic companies have turned to the GPGPU architecture, aside from better compatibility with Nvidia's existing software ecosystem, is based on market feedback on products over the past few years. They have come to realize that the market is essentially still a general-purpose computing market, and any shortfall in underlying general-purpose computing design will at some point manifest as an inability to support customer needs.
From the perspective of R&D organization and management, when the organizational goal is defined as being CUDA-compatible, to a large extent it no longer becomes necessary to reason backwards from application-layer requirements to determine whether the underlying software and hardware implementation is appropriate. Using Nvidia's design as a benchmark, one can directly compare whether the software and hardware implementation at the underlying level matches. This greatly simplifies R&D management. The price paid is that, given the current state of the supply chain, the ceiling of what the product can achieve also becomes capped.
"Waves": Compared to that, what is new about Fangqing's architecture?
Liang Jun: Fangqing's goal is to become a company that defines innovative systems, so we have chosen a different path, targeting a completely different market as well.
The reason design for general-purpose computing is difficult is that a large number of underlying design details ultimately end up reflecting on the programming interface. Balancing the trade-off between maintaining programmability in a general sense and using dedicated hardware for acceleration is not just a technical issue, but also an issue of R&D organization and management.
Running a team with strong technical taste while still delivering on time is an extremely challenging task, but based on our experience, it can still be done.
"Waves": Many companies have also emerged on the new technical architecture, represented domestically by Kunlunxin, and overseas by companies such as Groq and Tenstorrent. How do you evaluate the direction of current mainstream architecture?
Liang Jun: In terms of the design of the lowest-level computing core, the options have already converged to a few limited categories —
One is GPGPU, which provides programmers with a CUDA-compatible or CUDA-like programming interface. Another is based on designing computing systems using RISC-V, following the release of the RISC-V RVV (RISC-V Vector Extension) version 1.0 in the second half of 2021. The advantage of this approach is that it uses an open-source instruction set architecture rather than a proprietary one, which significantly reduces the risk of software investment for customers. However, there is also a problem. The current issue is that almost all vendors are designing based on a CPU design mindset, when what is actually needed is only compatibility with the RISC-V instruction set — the hardware design needs to be implemented with a completely new approach.
The third category is private designs by cloud vendors. Because they are sold as a service, any shortcomings in the chip's generality can be compensated for through system design. The company that does this best is Google. After eight or nine years of iterating through seven generations of chips, it has achieved a breakthrough this year with certain large customers. Groq can be considered to fall into this category as well. Since the decode phase of LLMs is serial output, Groq's design pursues extremely low latency. The system's throughput equals parallelism multiplied by the reciprocal of latency. This provides a competitive advantage in user experience and cost per token, but the trade-off is sacrificing generality in programming, so it is sold as a service.
"Waves": Does that mean Fangqing is different from all of them?
Liang Jun: Fangqing's approach is to design systems based on a decoupled architecture. We believe that the decoupled architecture is a higher-level computing architecture and programming model, and is currently our main focus. The choice of the underlying computing core is no longer the main concern — that conclusion has already been reached. The first two approaches each have their strengths and weaknesses, and we do not have a strong preference for either. Whichever route is done well can meet market demands. However, the need to strike an appropriate balance between maintaining general-purpose programmability and using dedicated hardware for acceleration is consistent across approaches, and requires that both the software and hardware teams have a correct understanding of this.
Furthermore, in the definition of the decoupled architecture, the system is decomposed into context-aware and context-free parts, natively supporting heterogeneous computing.
"Waves": You emphasize a distributed computing architecture that decouples "context-aware" and "context-free" components. What is the thinking behind this?
Liang Jun: From a computational perspective, AI systems are large-scale parallel computing. Specifically, when processing input-output sequences, due to the unrelated nature between different sequences, there is a significant difference between the implementation method used from software down to the underlying hardware and the implementation method used for processing weight-related computations. One could also say that the decoupled architecture is a higher-level computing architecture and programming model tailored for such applications.
The computing paradigm is rapidly shifting from processor-centric system design to memory-centric system design. In an era where AI models are prevalent, this shift in computing paradigm is a reality, but it has not yet been widely recognized. This is also the real reason why concepts such as compute-in-memory and near-memory computing have gained such traction in the industry in recent years. However, much of the current discussion is based on a hardware-centric understanding; a shift in computing paradigm would offer a completely new explanation.
Whether it is KV Cache (key-value cache) or weights, both can be largely defined as memory. Traditional memory has two attributes: capacity and bandwidth. The new type of memory adds two new attributes: computational semantics and communication, bringing the total to four dimensions. If we draw a comparison, the part that handles input-output processing is more like the processor in the von Neumann architecture, while the part that handles weight-related processing is more like the memory in a traditional processor. At the same time, we believe that both parts can be largely regarded as a new form of memory.
Based on these insights, we have re-examined our approach to system design. For many years, the trend in system design has been toward higher integration — SoCs with higher performance and more integrated functions. Faced with the current evolution of algorithms, a single SoC, constrained by physical limitations, is increasingly showing its weaknesses in terms of bandwidth, memory capacity, and other areas.
We believe that the first priority of chip design should shift from pursuing the performance and integration of a single chip to pursuing a scalable system design. From this perspective, adopting a decoupled architecture for system design becomes a reasonable choice.
"Waves": Specifically, how does this architecture improve chip capabilities?
Liang Jun: In the definition of the decoupled architecture, the system is decomposed into context-aware parts, context-free parts, and the communication between them. System expansion then shifts from one-dimensional expansion to multi-dimensional expansion, and the boundaries between the various components in the system are clearly defined. From this perspective, we have examined various computing systems, and the conclusions are very positive. Not only can we design more forms of computing systems, create new markets, and accelerate the deployment of various applications, but also, because the combination is done at the system level, system development can be appropriately decoupled from the chip development cycle. At the same time, since there is no longer a need to design a full-featured SoC, chip costs and development costs can also be reduced, which in turn helps accelerate the pace of innovation in the industry.
"Waves": If you were to use the most common example to explain to the general public, what pain point are you trying to solve?
Liang Jun: For example, after the SoC system of a mobile phone is changed to a decoupled architecture, the phone, smart glasses, smart earbuds, smart watches, and other devices can all be connected as independent input-output processors to the weight processor. Alternatively, you could say that the traditional SoC handles the context-aware part, while the system adds a weight processor to handle the context-free part. So, as long as you believe that the capabilities of models will become increasingly stronger, various IO processors — such as earbuds and glasses — will only need to connect to the weight processor to independently perform more functions. In the existing system definition, these devices are accessory devices to the phone's SoC. In the new system, these devices are on equal footing with the phone's SoC.
Our judgment is that when the first priority is changed to pursuing a scalable system design, it will change the existing approach to AI hardware design, open up opportunities to create new system form factors, and generate new markets.
"Waves": Today, with the CUDA ecosystem still very strong, what is the relationship between Fangqing's architecture and CUDA? Will customer migration costs be an issue?
Liang Jun: Systems designed based on a decoupled architecture natively support heterogeneous computing. Both the context-free part and the context-aware part can be built on top of existing systems. From this perspective, systems based on a decoupled architecture are the most compatible with existing systems — which, in a way, is counterintuitive.
The entire industry is still evolving rapidly. Fangqing's strategy is to design systems based on a decoupled architecture. Once the various components of the system are decoupled, the pace of innovation accelerates, allowing us to define more types of computing systems and create new markets, rather than replacing existing systems.
"Waves": When you left your previous company four years ago, there was a lot of speculation. Why did you ultimately choose to join a startup?
Liang Jun: The answer is simple. For Fangqing's goal, this kind of work is better suited to a startup. Startups have no historical baggage and decision-making is simpler. On the other hand, in the face of challenges from the market and capital markets, they must innovate and create differentiated products to survive. With a small team, more energy can be devoted to technical details.
I joined in August 2024. Since then, Fangqing has also brought in several industry shareholders. These industry shareholders recognize that Fangqing's team has the capability to complete technology platform development, market definition, and product R&D in a systematic way. They approve of the progress made so far and are willing to provide us with financial support. So, as things stand, this choice of mine has proven correct.
On the other hand, I have a labor dispute involving a significant amount of money with my previous company, which is already known to the public. Against this backdrop, after a two-year non-compete period, re-entering the job market was extremely difficult. Given the realities, and in making a difficult but correct decision, I trust my own judgment.
"Waves": When people mention you now, they still refer to you as former CTO of Cambricon and former chief architect of the HiSilicon Kirin SoC at Huawei. In 5 to 10 years, how do you hope people will introduce you?
Liang Jun: There is no shortage of high-tech companies in the Chinese market. What is rare is a company that has sustained technological originality in foundational technology, strong technical taste, and at the same time a successful super product in the market. That is the future goal I have set for Fangqing. When that day comes, I hope people will recognize me this way — Liang Jun is the CEO of Fangqing.
Back
Next article
