Loongson and Hygon Veterans Assemble, AGI Chip Company Incubated by BAAI Emerges! Raising Hundreds of Millions in Consecutive Rounds Within Half a Year

What 杨立昆 and 李飞飞 are both asking about, Tongxin Intelligent wants to answer with a single chip.
Yunchong, from Wafei Temple, Quantum位 | Public Account QbitAI
Quantum位 | Public Account QbitAI
Before a single chip is even built, the AI meant to run on it may have already gone through several rounds of updates.
This is the time gap that next-generation AI leaves to the chip industry: models can iterate quickly, but a chip must go through a long engineering cycle from definition to mass production. Waiting until computing demand is fully settled before setting out may already mean being a step too late.
Tongxin Intelligent chose to enter at this moment.
This company, initiated by renowned AI expert Professor Zhu Songchun and incubated by the Beijing Institute for General Artificial Intelligence, was founded in April this year. In just half a year, the company has completed consecutive seed and angel funding rounds totaling hundreds of millions of yuan, with the funds to be mainly invested in core architecture design and R&D before tape-out.
What this company aims to build is a new type of inference chip for AGI algorithms: compatible with current mainstream models, while adding dedicated hardware support for computations such as logic, causality, and value reasoning.
The question behind it is quite challenging: what kind of computing does next-generation AI actually need, and can we start defining it now?
To this end, Tongxin has scientists researching general artificial intelligence and engineers with experience in mass-producing large chips participate in the design together. They need to identify, from constantly changing algorithms, those computations that will be called repeatedly and are worth specially accelerating, and then decide how the chip should be organized.
Models will continue to evolve, but Tongxin has already begun making choices about which capabilities are worth building into hardware ahead of time.
AI Moves Toward the Real World, Chips Face a New Proposition
The boundaries of AI are gradually moving from interaction after interaction toward the real physical world.
As early as the first half of this year, LeCun, who left Meta to found AMI Labs and completed $1 billion in funding, pointed out that large language models are "helpless" when facing the physical world. LeCun's proposed solution is the world model: given the current state and a conceived action, predict the next state of the world, and then use that prediction for planning.
Just last month, Google DeepMind announced the world model Genie 4. Users can change the weather, add characters or objects in a simulated environment, and the model generates the corresponding changes in the environment in real time. DeepMind explicitly lists building training environments for robots and agents as an important application direction of this model.

These explorations make the tasks AI faces more concrete: it must not only understand what is in front of it, but also predict what will happen after taking an action, and provide a basis for the next decision. How agents evolve is also influencing underlying chip design. This matter has already entered the view of global giants.
In September this year, AMD announced the signing of an agreement to acquire World Labs, founded by 李飞飞, for approximately $8.2 billion. When explaining this deal, 苏姿丰 emphasized that building the next-generation AI computing platform requires a deep understanding of how models evolve.
What is noteworthy here is that chip makers need to figure out earlier what next-generation AI will actually compute, and where efficiency will be bottlenecked. World models and robot algorithms are still evolving, but hardware design needs to make choices in advance, which also brings model research and chip design closer together.
李飞飞 put it more directly: "Without dedicated hardware investment, both the efficiency and scale of AI will be constrained." At the same time, she also revealed that World Labs and AMD have already launched technical cooperation around model training and inference optimization on GPUs.
At the recent CoRL, 苏姿丰 explained this logic more concretely: hardware optimization takes years, and the deeper the understanding of what computations models need to perform and where the technology is heading, the greater the chance of achieving good co-optimization of hardware, software, and systems.
The demand for interactive response is also pushing 英伟达 to adjust its inference system. In response to agents' need for low latency, 英伟达 pairs GPUs with Groq's dedicated accelerators, letting different computing stages be executed by more suitable hardware, while balancing overall throughput and interactive response speed.
These changes point to a common question: when AI needs to make timely judgments and take actions based on environmental changes, how should chips support it?
Modern GPUs excel at large-scale parallel computing, combined with dedicated matrix operation units, which is the main reason they dominated the large language model era.
But an agent that needs to act in a real environment also has to handle another set of work: repeatedly querying relationships between objects, updating environmental states at any time, and simulating the possible consequences of actions.
When these tasks are implemented using graph structures, dynamic branching, and other methods, computation may become fragmented, and data access becomes more irregular. GPUs can execute these algorithms, but parallel resources may not be fully utilized, and part of the time will be spent finding data, moving data, and waiting for the results of the previous step.
At this point, looking only at how many operations a chip can perform per second can no longer explain how quickly a task is actually completed. Software optimization, storage organization, computing units, and scheduling methods all need to be reconsidered together.
The computing power advantage of the large model era cannot be directly converted into action efficiency in the real world.
Behind the efficiency problem, there is also an expanding inference market. CIC Consulting predicts that by 2030, China's AI inference chip market will reach approximately 1.17 trillion yuan.
The opportunity Tongxin sees lies in the efficiency gap between new tasks and existing computing methods. From the cloud to the robot edge, the different requirements tasks place on throughput, response speed, and power consumption also leave room for Tongxin to start from tasks and redefine the computing architecture.
The trend of AI moving toward the real physical world is already clear, but the implementation path is still evolving, so judgment needs to return to more basic questions: Are the pain points real? How will demand change? Which computations are worth handing over to dedicated hardware? Can theory guide product definition? And can the engineering team actually build it?
It is worth mentioning that from the cooperation between 李飞飞 and 苏姿丰, it is not hard to see that defining the next-generation AI chip requires not only a theoretical understanding of how intelligence works, but also the engineering capability to translate that understanding into hardware.
Only by following this line of thinking can one understand why Tongxin chose to start from Professor Zhu Songchun's theory of intelligence to design a chip.

Starting from Understanding Intelligence, Redefining AGI Chips
The biggest problem in designing a chip for a workload that has not yet taken shape may be that there is no reference point.
When GPU R&D first began, there was graphics rendering to benchmark against, and AI accelerators had matrix multiplication to optimize. When it comes to AGI chips, model routes remain divergent, computing demand has not yet taken shape, and observing the current situation cannot yield an answer.
Demand is still changing, and what should be placed on silicon must be based on a set of criteria. Tongxin's starting point is the CUV framework proposed by Zhu Songchun's team.
Professor Zhu Songchun is a winner of the Marr Prize and the Helmholtz Prize and has conducted research at the intersection of vision, cognition, and robotics for thirty years. The CUV framework published by his team this year in Engineering incorporates three aspects of agents into a unified framework: cognitive architecture (C), capability system (U), and value system (V). Among them:
- C is the cognitive architecture, organizing the internal modules of an agent and the way they collaborate, connecting perception, memory, reasoning, and other stages;
- U is the capability system, covering callable skills such as recognizing objects and planning paths;
- V is the value system, describing goals, preferences, and constraints, and providing the basis for choosing actions. Applied to robots, this can mean completing service tasks or balancing safety and the needs of others.
The CUV framework places what is seen, what can be done, and what should be done into the same decision-making process. What drives the growth of intelligence is "what is wanted," not "how much data has been fed."
These questions were already discussed by Zhu Songchun's team in a paper as early as 2020. In an era when deep learning was at its peak, Professor Zhu Songchun and his team wrote "small data, big tasks" into a paper. Five things invisible in pixels—function, physics, intention, causality, and utility—were called by them the "dark matter" of vision.
Six years later, world models and spatial intelligence have become prominent, and the industry has begun asking how AI understands what consequences an action will bring. But Zhu Songchun and his team had already highlighted this question.
It is worth mentioning that "small data, big tasks" emphasizes using prior knowledge such as physical laws and common sense to handle a broader range of tasks with less training data. This knowledge is equivalent to a manual that helps machines understand the world.
Take pouring water for a guest as an example. According to Tongxin's design approach, a robot can organize the positional relationships among the cup, the water source, and the guest into a parsing graph, then plan actions around the goal of delivering water, invoking skills such as picking up the cup, filling it with water, and moving, and finally deliver the water into the guest's hands.
On the way to deliver water, if the guest changes position, information that remains valid is retained, affected states and paths are updated, and repeated computation is reduced.
In this process, the chip needs to complete more specific work: storing the current state and goal state, accessing relationships between objects, performing probabilistic calculations, comparing different plans, and scheduling corresponding skills. These processes all need to be implemented as executable algorithms and then broken down into basic operations, that is, operators.
Tongxin's work is to identify the parts suitable for acceleration from these operators and further define computing units, storage, and data paths.
Vector and tensor operations are an important computational foundation of large language models, and this is where GPUs excel. An agent that needs to act in the real world must repeatedly answer "what will happen if I do this, and is it worth doing": reasoning about relationships between objects on a graph, performing probabilistic inference for uncertain situations, simulating physical consequences, and then scoring each candidate action.
Therefore, in Tongxin's definition, an AGI chip is not a faster GPU. It adds a new layer of operation types beyond vectors and tensors; what is added is structure, not merely more computing power.
In terms of Tongxin's chip plan, the company has planned three types of units: Vector, Tensor, and AGI Core. Vector and Tensor handle vector and tensor computation and support mainstream model architectures such as CNN and Transformer; AGI Core is aimed at logic, causal, and value reasoning based on parsing graphs as well as probabilistic computation, with dedicated acceleration capabilities designed for them.

These computing needs targeted by Tongxin's AGI Core have also appeared, as if by coincidence, in the research routes of other teams in the industry.
Values written like {{NAME_ab12cd34}} are placeholders for real values. Use them exactly as written, in replies and in tool calls alike: they are filled in with the real values before anything runs. Don't ask to verify them, and don't replace them with examples.
Article title: Loongson and Hygon Veterans Assemble, an AGI Chip Company Incubated by BAAI Emerges! Hundreds of Millions Raised in Six Months
Body: 李飞飞's spatial intelligence focuses on how machines understand three-dimensional environments, while DeepMind's graph networks organize information using entities and relationships. When faced with incomplete information, DARPA-supported probabilistic programming attempts to combine existing knowledge and complete inference with less data; Pearl's causal theory further asks what result a change in one condition will bring. Silver, Sutton, and others proposed in "Reward is Enough" that an agent pursuing rewards through trial and error may develop capabilities such as perception, learning, and reasoning.
In Tongxin's view, the storage and processing of graph structures, probabilistic and causal inference, and the evaluation of action outcomes all have parts worth specifically accelerating, and AGI Core has accordingly planned corresponding compute units. Professor Zhu Songchun's theoretical system provides a clear starting point, from which Tongxin hopes to extract computing capabilities that can be used by more models.
Different parties may understand intelligence differently, but as the foundation of intelligence, what chips need to find is the greatest common divisor among them.
Turning Theory into Silicon: What Tongxin Relies On
Having a systematic theoretical framework and understanding intelligence and algorithms, the next step is to actually build the chip.
Tongxin's founder and CEO Wang Haiyang studied under Academician Xia Peisu, one of the founders of China's computer cause. Wang Haiyang himself is a founding member of the Loongson CPU and once served as the chief architecture leader for Hygon's first to fourth generation CPU chips, and is also a co-founder of a GPU unicorn company. He has 25 years of chip R&D and 15 years of SoC architecture design experience, and has participated in the entire process from R&D to mass production of three types of large chips: CPU, GPU, and GPGPU.

Another company founder, Shi Feng, studied under Professor Zhu Songchun, received his PhD from UCLA, is responsible for the technical route and architecture definition of Tongxin's AGI Core, and also serves as director of the AGI Chip Center at BAAI.
Both founders once participated in Loongson R&D. Later, Wang Haiyang brought large chips to mass production generation after generation in the industry, while Shi Feng combined intelligent algorithms with chip design and was responsible for translating model and algorithm requirements into compute units in AGI Core.
The founding core team behind them has an average of more than 20 years of industry experience, with key members from companies such as Hygon, Baidu, ByteDance, and Alibaba, and has gone through the complete process of more than ten large chips from architecture definition and front-end/back-end design to tape-out and mass production delivery.
These experiences turned into several concrete trade-offs.
The first is storage. Wang Haiyang is one of the earliest people in the industry to propose the idea of hybrid compute-in-memory. Tongxin does not intend to make a compute-in-memory chip; it treats compute-in-memory as a means to alleviate the "memory wall": different tasks have different requirements for access speed, capacity, and cost, so Tongxin pairs them with different storage and computing methods—use a kitchen knife to cut vegetables, a fruit knife to peel fruit—saving the overhead of moving data back and forth and compute units waiting idly for data.
Further down is area. Reconfigurable design allows some hardware resources to be reorganized according to the task, with general-purpose units ensuring compatibility and dedicated units improving efficiency. Facing the different needs of cloud, edge, and device, Tongxin plans to adopt Chiplet design: based on the same architecture and software stack, different configurations meet the cloud's throughput requirements and the robot device side's power consumption and size requirements.

In terms of product cadence, Tongxin has not jumped straight to the future. Because a chip designed only for tomorrow has no one to use it today.
Tongxin's first-generation product is positioned as a high cost-performance inference chip, targeting cloud large models, world models, and embodied intelligence device-side computing. Existing models can run directly, AGI Core is onboard at the same time, and new computing capabilities are verified in stages.
AGI chips and GPUs are not an either-or relationship. Tongxin positions itself and GPU-centric solutions as symbiotic: large language models remain an important component of the AGI system and one of the ways agents understand the world, walking on both legs of compatibility and specialization.
Progress is faster than most people think. Only three months after its establishment, Tongxin completed architecture design and project initiation, has now entered the chip design stage, and will carry out FPGA prototype verification in early December. Commercialization covers three directions: cloud inference, general agents, and device-side physical intelligence, and preliminary large-scale cooperation intentions have been formed.
In addition, Tongxin plans to adapt to BAAI's general agent "Tongtong" and related large models, and to jointly build joint laboratories with Peking University, Tsinghua University, the Institute of Computing Technology of the Chinese Academy of Sciences, and leading large-model industry partners.
Wang Haiyang mentioned that for Tongxin, architecture definition is only the starting point; after that come design, verification, tape-out, and mass production. The computing capabilities chosen today must, after this round of R&D, still be able to adapt to the models and tasks of that time.
This also gives Tongxin's exploration greater observational value: as AI extends into the real world, chip teams need to understand algorithms earlier, break down key computations into operators, and then implement the parts worth accelerating into hardware design.
This path will ultimately prove its value through robots' more timely responses, agents' lower operating costs, and developers' more convenient deployment.
From theory to algorithms, from operators to silicon wafers, and then to real tasks, Tongxin chooses to advance these links together.
The competition for next-generation AI chips has already begun before the birth of the silicon wafer: understanding how intelligence works and defining what kind of computing it needs.
For Tongxin, entering early is only the first step. Next, it needs to use a chip that truly runs to prove that today's judgment about intelligence can become tomorrow's efficiency advantage.
As Professor Zhu Songchun said: "The chip is the heart of the agent. To establish a heart for the machine, we must start with this chip."
*This article is republished by QbitAI with authorization. The views belong only to the original author.
- 2.48x faster with no loss in image quality, AI video "cost reduction and efficiency improvement" has a new solution 2026-11-16
- High-difficulty blind test surpasses GPT-6-Astra and Opus-5! Wujie Frontier open-sources Fx-Bio-1031, bringing RSI into life sciences 2026-11-13
- openJiuwen releases RSI-native AgentOS, allowing Agents to transform freely and iterate continuously! 2026-11-13
- Made in Shanghai, connected in Hong Kong, Chinese AI accelerates toward the world 2026-11-11
Related Reading

The team incubated by BAAI has also set its sights on RSI
When AI understands you more and more, how can intelligence truly belong to you?

Zhu Songchun team's latest research: robots can have heart-to-heart talks with humans! Also said the next step is to build "AI Baymax"
The research results have been published in Science magazine and its sub-journal headlines

Peking University takes the lead in officially announcing Zhu Songchun's joining, appearing at Yanyuan, serving as dean of the AI Research Institute
