Having twice graced the cover of *Nature*, Yang Zheyu, a PhD graduate from Tsinghua University, has secured 500 million yuan in financing.

At present, physical AI has become one of the most crowded tracks in the primary market, with massive capital pouring into solutions that "enable machines to get closer to human judgment and behavioral capabilities". More than 80% of the foundation of human decision-making and action comes from the information provided by vision. In this sense, if AI is to truly enter the physical world, vision is an unavoidable threshold.

However, today's mainstream image sensors are inherently designed for "human viewing": exposure frame by frame, full output, and each frame of picture is processed as a photo that needs to be saved. This logic works well in the era of mobile phone photography, but when it is embedded in robots and autonomous driving systems, defects begin to emerge — data redundancy, tight transmission bandwidth, and a large amount of backend computing power is consumed by invalid information, while the truly critical spatiotemporal changes are easily overwhelmed. It can be said that the efficiency of visual perception is becoming the most realistic bottleneck for the implementation of physical AI.

In this context, a new idea is trying to break this deadlock, which is "brain-inspired vision". Instead of imitating cameras, it imitates the information processing mode of the human eye and optic nerve: it disassembles information such as color, edge, and motion changes at the perception front end, outputs them in a sparse and complementary form, so that the subsequent decision-making system only obtains valuable information.

In China, a team from Tsinghua University has taken the lead on this path and has attracted the attention of capital. ChinaVenture learned that Xijian Technology, a brain-inspired vision chip company incubated by the Tsinghua University Research Center for Brain-inspired Computing, has completed three consecutive rounds of financing with a total amount of nearly 500 million yuan just one year after its establishment.

Its list of investors is also very impressive — Xuhui Capital, Matrix Partners China, Huaying Capital, Innoangel Fund, Shuimuhuaqing, leading enterprises in the embodied intelligence and automotive industry chains, etc., which almost cover the most active capital types at present, including government industrial funds, leading market-oriented institutions, university-affiliated capital and industrial capital.

Behind the intensive layout of capital, in addition to the recognition of the cutting-edge direction, the more core confidence comes from the strength of the team itself. In the scientific research field, the gold content of the cover of *Nature* is self-evident. In the two major fields of AI and chips, there are only three achievements made by Chinese mainland institutions as the first completion unit that have been featured on the cover of the main issue of *Nature*, and the team of Tsinghua University Research Center for Brain-inspired Computing accounts for two of them — the "Tianjic" chip in 2019 and the "Tianmou" chip in 2024. Xijian Technology is the only industrialization platform for the technological achievements of the "Tianmou" chip.

The "all-in" venture born out of Tsinghua University's laboratory

The story of Xijian Technology starts from the Tsinghua University Research Center for Brain-inspired Computing.

Established in 2014, the center was jointly founded by Tsinghua University relying on the Department of Precision Instrument, in cooperation with 7 departments including Computer Science, Microelectronics, Electronics, and Automation. It is the earliest team in China to conduct comprehensive brain-inspired computing research. For more than ten years since its establishment, a series of original achievements coming out of here have been published in top international journals for many times. The most well-known and most significant ones are the brain-inspired computing chip "Tianjic" and the brain-inspired complementary visual chip "Tianmou" mentioned above.

Yang Zheyu, CEO of Xijian Technology, is a doctoral student of this center. During his PhD study, he participated in the core work of the two projects successively, and is the co-first author of the "Tianmou" chip paper. It is this scientific research experience that made him see the broad prospects of "brain-inspired vision". In his view, the reason why large language models exploded before vision models is that language itself is the product of intelligent compression, which is naturally more likely to generate emergent intelligence. However, vision accounts for more than 80% of the human perceived information volume, and it is the key threshold that really needs to be overcome in the next era. Today's image sensors are essentially designed "for human viewing", and forcing them to be "viewed by AI" does not adapt to the way large models perceive the world — "defining vision for AI" is undoubtedly a broad blue ocean.

This judgment does not come entirely from subjective perception, and the market has long given signals. Yang Zheyu shared a detail: as early as before the company was incubated, many customers came to the Tsinghua University Brain-inspired Center with demands, hoping that the team could help them customize chips. Among them, Hikvision Nanhu Research Institute clearly expressed its willingness to promote the application of the Tianmou chip in many subsidiaries under the group. This industrial demand has been further verified at the national project level: in the central-local coordinated major brain-inspired special project led by Hikvision Group, the Tianmou chip is clearly listed as one of the two core industrial pillars of brain-inspired technology. In January 2026, the project was further upgraded and included in the sequence of national key R&D programs.

The double confirmation from the front line of the industry and national projects made Yang Zheyu more convinced — the market has long been waiting for a visual solution that can take into account both high resolution and high frame rate.

In addition to the clear market prospect, there is another driving force. "More than 80% of the high-end visual chip market has long been monopolized by overseas companies. From the perspective of patriotism, we also don't want all high-end visual solutions to come from overseas." Yang Zheyu said.

But to realize the transformation of industry-university-research achievements, a market-oriented carrier is ultimately needed. So when he graduated with a doctorate in 2025, he gave up the faculty offers from universities at home and abroad, with the determination of "going all in", he resolutely devoted himself to the industrialization process of the Tianmou chip brain-inspired vision technology, and officially founded Xijian Technology.

It is worth mentioning that in order to push this entrepreneurship from vision to reality, Yang Zheyu started to build a composite team that connects scientific research and industry from the very beginning. Among them, Dr. Wang Taoyi, CTO, has long participated in the R&D of the "Tianmou" chip and has published 13 papers in top journals such as *Nature* and *Science Robotics*. Professor Zhao Rong, Chief Scientist, is a national high-level talent, long-term associate professor in the Department of Precision Instrument of Tsinghua University, deputy director of the Research Center for Brain-inspired Computing, has long been responsible for the research direction of brain-inspired visual perception, and serves as the project leader of the major project "Heterogeneous Fusion Brain-inspired Computing Research Platform" of China Brain Project, and has published more than 100 SCI journal papers.

Up to now, the team of Xijian Technology has about 40 people, of whom more than 80% of the R&D personnel hold doctorate or master's degrees. Most of the core engineering members come from Sony, NVIDIA, ON Semiconductor and leading domestic CIS enterprises, with 10 to 20 years of working experience in image sensors and chips. The members of the industrialization and marketing team have been deeply engaged in the industry for more than 20 years, and have been responsible for related businesses with an annual scale of hundreds of millions of US dollars.

From original research, chip design to mass production of engineering and market implementation, the capabilities of this team have covered the entire industrial chain of visual perception.

Equip AI with a pair of "retinas" that can encode

To understand the value of Xijian Technology and the "Tianmou" chip, we must first clarify the current bottlenecks faced by AI at the level of "visual perception".

At present, most AI hardware still uses traditional CMOS cameras as the front end of visual collection. Its working principle is to convert optical signals into digital signals, output images frame by frame at a fixed rhythm, and then through encoding and compression, finally become the familiar video format. But the problem lies precisely in the underlying logic of this "frame-by-frame output". Imagine an autonomous driving test vehicle with more than a dozen cameras running at the same time, which can accumulate terabytes of data in one day. However, if you look closely, the sky, road surface, and guardrail in adjacent frames hardly change, but are repeatedly collected, transmitted and analyzed, which forms massive data redundancy. In addition, the three core indicators of resolution, frame rate and precision often restrict each other in traditional solutions, and improving one means sacrificing the other two, thus forming a long-standing unsolvable "impossible triangle" in the industry.

The bigger problem lies in computing power. The data volume of vision is naturally huge — a 1080P picture corresponds to about 8000 Tokens. Converted at 30 frames per second, it is hundreds of thousands of Tokens per second. However, the best edge GPU currently can only process hundreds of Tokens per second. This means that using traditional methods to make AI "see" the world will infinitely amplify the delay, which is why applications such as virtual reality and world models are difficult to implement for a long time.

The solution of the "Tianmou" chip is to transform the camera from a "frame-by-frame snapshot" into a set of selective visual primitive system. It no longer mechanically saves each frame of picture as it is, but draws on the multi-path mechanism of the retina to perform differentiated collection and complementary encoding of visual information such as light intensity, spatial structure and time changes at the perception front end. This not only breaks the restriction of mutual trade-off between resolution, frame rate and precision, but also greatly reduces the repeated transmission and computing burden on the backend.

According to the team, the "Tianmou" chip has now achieved high-speed perception of 10,000 frames per second and a high dynamic range of 130dB, while reducing the transmission bandwidth by 90% and the power consumption to one tenth of that of traditional high-speed sensors, with comprehensive performance more than double that of similar overseas products.

It is worth mentioning that in the business map of Xijian Technology, the "Tianmou" chip is not an isolated chip, but a complete set of "machine vision nervous system", which includes the Tianmou chip, protocols, and multi-modal large models. Among them, the chip acts as the "retina" of the machine, responsible for perceiving the physical world and completing the disassembly and encoding of visual information at the front end; the transmission protocol acts as the "optic nerve", efficiently transmitting information in a form more suitable for model processing; the visual model acts as the "visual center of the brain", understanding and reasoning the information, and further connecting decision-making and action.

The collaboration of chips, protocols and models enables machine vision to no longer stay in image collection and passive input, but rise to the native visual perception and understanding capability facing the physical world.

After a year of efforts, the achievements are being continuously delivered. In August 2026, the team's latest research achievement was featured on the cover of *Nature Sensors*, only two years after the Tianmou chip paper was on the cover. At the product level, the company's mass production process is advancing steadily: the first-generation chip has been successfully lit up, and the second-generation chip for mass production has completed tape-out, with the resolution upgraded from 200,000 pixels to 5 million pixels. At the same time, visual information processing is further moved forward to the perception end, which can directly extract high information density visual Tokens for large models, and the first Token delay is controlled within tens of milliseconds.

At the application end, the Tianmou chip has covered multiple tracks such as embodied robots, autonomous driving, industrial inspection, high-speed photography, and UAV inspection, which can accurately capture industrial abnormal events at the order of hundreds of microseconds and meet the low-power deployment requirements at the edge end.

Brain-inspired vision, an increasingly valued business

When large language models enter the "hundred-model war", cutting-edge attention is accelerating to turn to the physical world — embodied intelligence, humanoid robots, and autonomous driving are all calling for a pair of "eyes" that can understand the real world. Brain-inspired vision is exactly a feasible path provided for these eyes.

Nowadays, this field is receiving more and more attention from policy and capital levels. During the 15th Five-Year Plan period, brain-inspired computing is not only listed as a key technology research direction, but also included in the achievement transformation system of national engineering laboratories. At the same time, leading institutions, government industrial funds and industrial capital have also taken action one after another to accelerate their layout.

In Yang Zheyu's view, the reason why brain-inspired vision is highly expected is that it represents a truly underlying innovation path. Disruptive innovation never chases along the established track, but returns to the first principle, redefines the problem itself, and explores a path that has not yet been verified. His goal is grand enough and clear enough: "Wherever there is autonomous decision-making by AI in the future, there will be Tianmou's chips."

However, in the face of this huge blue ocean market, Yang Zheyu does not intend to build a closed ecosystem. "We are open, including IP licensing, inviting more people to jointly promote the development of general visual intelligence."

The era of physical AI may be an era for young people like him who dare to take risks and innovate. The story of Yang Zheyu and Xijian Technology has just begun.

This article is from the WeChat official account "ChinaVenture", Author: Wang Manhua, Editor: Wang Qingwu, published with authorization from 36Kr.