36Kr Exclusive | The core founding technical team of Hygon has launched a startup dedicated to developing AGI architecture chips, raising hundreds of millions of yuan in financing from Tsinghua-affiliated funds and other institutional investors.
Image source / Visual China
This article contains approximately 3400 words, with a recommended reading time of 8 minutes
Author | Ou Xue
Editor | Yuan Silai
Hard Krypton learned that Tongxin Artificial Intelligence Technology (Beijing) Co., Ltd. (hereinafter referred to as "Tongxin Intelligence"), a developer of native general artificial intelligence (AGI) architecture chips, has consecutively completed hundreds of millions of yuan in seed round and angel round financing only half a year after its establishment. The financing was jointly participated by financial investors including Tsinghua Alumni Seed Fund, Beijing General Artificial Intelligence Research Institute, Solux Technology, Pingzhi Information, Dark Matter AI and Yuzi Zhangquan, gathering industrial capital, leading technology VC and financial capital. The funds will be mainly used for core architecture design and R&D before chip tape-out.
Tongxin Intelligence was founded in April 2026, initiated by Zhu Songchun, an artificial intelligence scientist, dean of the General Artificial Intelligence Research Institute, and chair professor of Peking University and Tsinghua University, and incubated relying on the Beijing General Artificial Intelligence Research Institute.
The core founding team of the company has an average of more than 20 years of experience in the chip industry, and its backbones come from leading chip and AI teams such as Hygon, Baidu, ByteDance and Alibaba. The core founding technical team graduated from universities and research institutes including Tsinghua University, Peking University, Chinese Academy of Sciences and UCLA, and has gone through the full process of architecture definition, front-end design, back-end design, tape-out and mass production delivery. The scientist team covers the full AI technology stack including large models, world models and physical intelligence. Wang Haiyang, the founder, studied under academician Xia Peisu, one of the founders of China's computer cause, is a founding member of Loongson CPU, once served as the architecture lead of 4 generations of Hygon chips, and co-founded the GPU unicorn. He has 25 years of chip R&D experience and 15 years of SoC architecture design experience, holds nearly 100 domestic and foreign invention patents in total, and has rich R&D and mass production experience in large chip fields such as CPU, GPU and GPGPU.
At present, traditional architectures represented by GPGPU and Tensor Core are mainly oriented to the optimization of statistical large models, and have inherent mismatches on loads such as causal reasoning, environmental interaction and physical intelligence. At the same time, the imbalance of computing power, storage power and transmission power pushes up the production cost of Token and the implementation cost of robot physical intelligence.
Wang Haiyang said that AGI competition has shifted from simply comparing peak computing power to comparing effective reasoning capability oriented to physical world interaction. Industries such as Token factories, physical intelligence and robots are highly sensitive to hardware cost and power consumption, and the industry needs a new chip architecture that natively adapts to cognitive reasoning.
Therefore, at the current industrial node where large language models are transitioning to AGI, Tongxin Intelligence has chosen a differentiated route of "not making an improved version of GPU, but directly designing chips around the CUV cognitive architecture".
The chip design concept of Tongxin Intelligence is derived from the three-dimensional joint space of Cognition, Utility and Value proposed by Professor Zhu Songchun, namely the CUV architecture. Based on this theoretical foundation, Tongxin Intelligence takes the self-developed three-unit architecture of Vector+Tensor+AGI Core as the basis to create a cost-effective chip that not only supports vector and tensor computing, but also provides native support for physical and social intelligence. Professor Zhu Songchun said: "The competition of general artificial intelligence is essentially the competition of architecture. What Tongxin wants to do is to implement this route on silicon: adapt chips to the cognitive architecture of AGI, and cooperate with agents such as 'Tongtong' and large models to form an end-to-end solution."
In terms of technical implementation, Tongxin Intelligence adopts a hybrid in-memory computing architecture, a general+special-purpose computing architecture, and a flexible unified cloud-edge-end architecture. Through self-developed technologies such as vector matrix unit, MoE expert mode, multi-level hybrid in-memory computing, graph traversal engine, and natively accelerated STC-PG (Space-Time Causal Parsing Graph), it solves the structural mismatch problems such as "memory wall", "sparsity and irregularity" and "graph structure reasoning" that exist when general-purpose GPUs run AGI.
According to Wang Haiyang, the company's first-generation chip is positioned as a cost-effective reasoning chip, which supports both cloud-side model factories and end-side physical intelligence, with its end-side product benchmarking NVIDIA's Thor series and cloud-side product benchmarking NVIDIA's B300 series.
"With the enhancement of large model capabilities, AI is moving from the training era to the reasoning era. The demand for training models is declining, while the reasoning demand will occupy an absolute majority, and users' pursuit of Token cost performance is constantly improving." Wang Haiyang said that defining the first chip of Tongxin Intelligence as a cost-effective reasoning chip is exactly to respond to this trend.
Hard Krypton further learned that the company completed architecture design and project establishment only three months after its establishment, and has now entered the chip design stage. It plans to launch FPGA prototype verification work in early December, and will display phased achievements at that time.
In terms of commercial implementation, Tongxin Intelligence has planned three market layouts and has initially reached large-scale cooperation intentions: first, relying on the system of the General Artificial Intelligence Research Institute and industrial strategic partners, to carry out cooperation in the field of AGI general agents; second, for large model intelligent computing clusters and Token factories, cooperate with complete machine manufacturers and local intelligent computing centers to provide cost-effective Token computing power; third, for physical intelligence and embodied robots, provide end-side large computing power reasoning chips.
In addition, the company also plans to build joint laboratories with Peking University, Tsinghua University, Institute of Computing Technology of the Chinese Academy of Sciences, leading large model and embodied robot companies in the future, to carry out joint research around CUV cognitive architecture, physical intelligence operators, and AGI chip software and hardware collaboration, so as to open up the industry-university-research implementation channel.
The following is an edited excerpt of the interview between Hard Krypton and Wang Haiyang, founder of Tongxin Intelligence:
Hard Krypton: What is the difference between the AGI chip that Tongxin Intelligence focuses on and other AI chips?
Wang Haiyang: For example, many companies working on optical computing now are essentially doing computing acceleration for current large language models. But large language models are not equal to AGI. We are not making a faster and more power-saving "calculator", but doing native design at the bottom of the hardware for the CUV cognitive architecture proposed by Professor Zhu Songchun. In addition to supporting vector and tensor computing to be forward compatible with the existing Transformer architecture, we have added dozens of hardware accelerators specially designed for AGI algorithm operators in the chip, so that future physical intelligence and spatial intelligence algorithms can run on our chip completely and efficiently. This is a brand new thing in the current chip category, which can be called a "new species".
Hard Krypton: In the current contacts with the industry, what do customers really care about?
Wang Haiyang: What everyone cares most about now is actually the reasoning cost, that is, the price of Token. In the past few years, everyone was competing for training computing power, but with the enhancement of model capabilities, there are fewer and fewer base model companies, and the real implementation is mainly in reasoning. When customers look at our chips, first, they see whether we can support the evolution to AGI in the future, and second, they see the cost performance. We define the first-generation chip as "a cost-effective reasoning chip that supports AGI algorithms", which supports both cloud-side intelligent computing clusters and end-side embodied intelligence.
Hard Krypton: What is the industry's enthusiasm for AGI chip demand at present? What trends will you focus on in the future?
Wang Haiyang: This is a gradual process. The development of large models is very fast, but the academic and industrial circles are gradually discovering its limitations, and now everyone is starting to align with AGI. We believe that AGI chips are in a transition stage similar to the period after the publication of the Transformer paper in 2017 and before the outbreak of ChatGPT. Chip R&D takes at least 20 months from project approval to mass production, and it is the right time for us to do it now.
From the perspective of trends, first, reasoning will account for an absolute majority, and various optimistic evaluations believe that the proportion of reasoning will reach more than 90%. Second, improve the computing efficiency of computing chips. At present, chips are facing three major bottlenecks of computing power, storage power and transmission power, among which the memory bottleneck is the biggest problem restricting the efficiency of computing chips. The hybrid in-memory computing we proposed is the core solution to solve the storage power bottleneck in the future. No matter how the technical schools change in the future, computing chips that solve the storage power bottleneck through in-memory computing will become the absolute mainstream, and we have a clear layout in this regard. Third, it will gradually transition to AGI chips. Now no matter the academic circle or the industrial circle is accelerating the exploration of AGI, and chip manufacturers will also accelerate their layout.
Investor Views:
Liu Xiaodong, founder of Yuzi Zhangquan Capital, said: Yuzi Zhangquan Capital focuses on the layout of technology intelligent manufacturing and disruptive science and technology innovation fields, and is committed to finding and investing in high-quality targets with real technical content at home and abroad. The investment in Tongxin Intelligence is based on our optimism about the long-term opportunity of reasoning-side computing power after the separation of training and reasoning, and we also value the underlying chip architecture that the company has laid out in advance for the AGI era. The company's team is also an important reason that attracts us: the team of Tongxin Intelligence not only has the scientific research foundation of the research institute, but also has the engineering and mass production experience of nearly 20 large chips from leading chip manufacturers, with scientists, industrial experts and managers taking charge of their respective parts. This "triangular" structure is exactly the most critical link in transforming innovation into entrepreneurship. As investors, we provide capital and patience, and are willing to work with the company to deepen the root of the architecture and ecology and move forward steadily and far.
Wang Xuehui, managing partner of Shuimu Alumni Fund, said: On one side of Tongxin, there are scientists studying general AI, and on the other side, there are engineers who have personally completed the mass production of Hygon and Loongson chips. The company is positioned as a new domestic chip specially used for humanoid robots and agents, integrating software and hardware to solve the pain point that existing chips cannot handle physical intelligence in the real world. We are optimistic about the future prospect of the large track and the team, and hope that Shuimu can continuously empower the company to realize tape-out as soon as possible.
Han Zongyu, head of strategic investment of Solux Technology, said: We are optimistic about Tongxin's core technology of integrating AGI chips and native models, and its founding team with rich mass production experience. This investment is an industrial strategic layout. Relying on Solux's global industrial foundation in medical care, logistics and family scenarios, we will form deep synergy with Tongxin to strengthen the underlying computing power of embodied intelligence. We look forward to the two sides jointly promoting chip adaptation, product verification and commercial implementation, and building a controllable and implementable domestic AGI robot industrial chain.
Li Liwei, founding partner of Puchao Capital, said: Puchao Capital continues to focus on the layout of the computing power infrastructure industrial chain in the artificial intelligence era. Tongxin Intelligence has opened up the technical link from AGI theory to chip architecture. The team has both cutting-edge AGI scientific research accumulation and long-term full-cycle mass production experience of large computing power chips, and uses a differentiated architecture to adapt to the needs of causal relationship and multi-physical field computing. We look forward to the company's software and hardware collaborative implementation, to consolidate the computing power base in the domestic AGI era.
Lai Guangning, founding partner and chairman of Qingchuang Bole, said: When every technological revolution is truly implemented, the competition is not about who has a louder concept, but about who can make scarce things cheap, reliable and scalable. Now that AGI has come to this day, whoever can make intelligent computing more cost-effective will master the key to the next stage of industrialization. Tongxin Intelligence's choice to make AGI chips is essentially participating in defining the computing power base of this era. They can not only make original judgments on the architecture, but also solidly carry out implementation work; what China's underlying innovation often lacks is exactly this perseverance to keep going. Qingchuang Bole is willing to accompany it to make this thing a success; and we hope that what Tongxin will be remembered for in the future is not just their chip products, but the foundation they have laid for China's AI industry.