In 2017, at the age of 28, Shen Yichen founded Xizhi Technology-P (01879.HK) with a focus on $NVIDIA(NVDA.US)$。
At that time, the slowdown of Moore's Law was already an industry consensus. Chip manufacturing had approached the 7-nanometer limit, and continuing to rely on electronic chips for computing power appeared increasingly challenging. His idea was to replace NVIDIA’s GPU with an entirely new optical computing chip to handle AI computing tasks.
The technology originated from a paper he published in Nature Photonics the previous year — the first experimental verification of the feasibility of using light for deep learning computations. This paper remains one of the most cited foundational works in the field of photonic computing. It was also due to this paper that Shen Yichen secured a $11 million angel round of financing from Zhen Fund, $Baidu (BIDU.US)$, Wu Yuan, and others.
Nine years later, Moore's Law is still holding strong; advanced process nodes have progressed to 1 nanometer; NVIDIA’s market value has surged dozens of times to $5 trillion, while optical computing is still searching for viable application scenarios.
Investors have also become more cautious. Each domestic GPU manufacturer in China has raised nearly 10 billion yuan, whereas in the more challenging optical computing industry, Xizhi has raised the most, but only about 2 billion yuan in total.
While the technology may foresee the future earliest, the business world tends to price only what has already materialized. Most hard-tech startups engaged in original innovation must navigate through such 'valleys of death.' ASML, Genentech, and NVIDIA all went through it.
Zhang Biao, Executive Partner of HeLi Capital and investor in Xizhi Technology, said the key is to adapt to the times. While the overall direction is important, at each stage, one must first pick the low-hanging fruit. 'Sometimes, what you fear most is spending ten years sharpening one sword.'
By the end of 2022, Xizhi’s optical computing product hit a bottleneck. The new project would require at least another 500 million yuan, while the company only had 300 million yuan left in its accounts. The technical team strongly hoped to continue investing, believing the company could secure further funding, but Shen Yichen decided to halt the project.
It was also during this time that the surge in demand for computing power brought about by large-scale model breakthroughs made high-speed interconnection between super nodes and GPU clusters increasingly critical. Xizhi shifted its focus to optical interconnect technology—using light to replace some electrical wiring, enabling faster and more energy-efficient connections between chips, servers, and GPU clusters.
"Survival is the most important thing," said Shen Yichen. He was well aware that if the company spent this money, even if it could secure additional financing, it would be placed in a passive position or even forced into a valuation bet. During that period, he witnessed more than a dozen companies founded by his MIT classmates collapse for this very reason.
By 2025, optical interconnects had contributed nearly 80% of Xizhi's revenue, amounting to approximately RMB 84 million. Among independent scale-up optical interconnect solution providers in China, Xizhi held an 80% market share. In the same year, Xizhi once again initiated the development of a new generation of optical computing chips, with the goal of integrating them into the reasoning phase of large-scale models.
Today, Xizhi Technology went public on the Hong Kong Stock Exchange, attracting$Alibaba(BABA.US)$, GIC (Government of Singapore Investment Corporation),$China Mobile(00941.HK)$, Blackrock, Temasek, Hillhouse, and 20 other cornerstone investors, with a current market capitalization exceeding HKD 80 billion.
"Now we have both the stars and the sea, as well as bread," said Shen Yichen.
Our conversation with Shen Yichen took place on the eve of the IPO. Before the interview, he had just concluded a technical meeting; after the interview, he rushed off to his next appointment right on schedule. This is his daily routine—almost no alone time, yet there is no sign of fatigue, except for his somewhat disheveled hair. A passing colleague reminded him: you need a haircut.
"Ten years have passed. Is it possible that the path of optical computing itself may not be viable?" we asked him.
"From the perspective of technological advancement, this is something that companies should attempt," said Shen Yichen.
But it is also possible to become cannon fodder.
Even if we become cannon fodder,” he said, “we were the first to propose this route. If we don’t do it, who will?”
A paper worth 22 million US dollars
LatePoint: Optical computing is considered a relatively forward-looking direction even today, but in 2017, before you graduated, you received 10 million US dollars from Zhen Fund, Baidu, Shunwei Capital, and FiveSource Capital. Did they really understand what you were trying to achieve at that time?
Shen Yichen: Most people didn’t quite understand, given that our paper in Nature (a top-tier international academic journal) had just been published. Anna (Fang Aizhi, founding partner and CEO of Zhen Fund) and I talked for 20 minutes in a coffee shop in Beijing’s Guomao district, and she immediately issued a term sheet. FiveSource Capital took longer discussions; they initially invested 200,000 US dollars and later added more.
However, this was also related to the broader environment. The previous year, AlphaGo had just defeated Lee Sedol,$Cambricon (688256.SH)$had just been established, and many people saw AI progressing, so the AI computing sector was relatively hot at the time.
LatePoint: How did you calculate that you needed to raise 10 million US dollars?
Shen Yichen: For the angel round, we could only raise that much because we only had one paper and no team yet. Everyone calculated that the paper was worth over 100 million yuan, so we raised 11 million US dollars based on a pre-money valuation of 22 million US dollars.
To the best of my knowledge, there may not have been any paper in history that was worth this much. Of course, later on, many people who authored the Transformer paper went on to start ventures, and many of those companies achieved high valuations.
LatePoint: Why can a research paper be priced so highly?
Shen Yichen: First of all, that paper did have a relatively large impact at the time. Most articles published in Nature have an average citation count of 50, whereas our paper has been cited over 500 times in recent years.
The technology investment market in the United States experiences cycles like this: A certain direction suddenly gains attention, people realize it hasn’t been achieved for many years but now seems close to realization, and expectations rise rapidly.
Moreover, during that period, a classmate who co-authored that paper with me also founded a photonic computing company called Lightmatter, which itself became quite a topic of discussion.
LatePoint: You didn’t choose to start a venture together?
Shen Yichen: Our personalities and philosophies differ. For instance, in the U.S., there are two types of entrepreneurs—visionaries like Elon Musk and more pragmatic and grounded individuals like Larry Page. Personally, I admire the latter more.
LatePoint: What were your expectations at the time?
Shen Yichen: At the time, I believed photonic computing could be realized within three to five years, with a focus on the AI domain, replacing traditional AI chips powered by electricity, offering computational speeds approximately ten times faster than theirs. We positioned ourselves as$Tesla(TSLA.US)$。
LatePoint: So you wanted to challenge NVIDIA, which sounds like a big 'story' in itself. How is it essentially different from what he said about replacing Intel?
Shen Yichen: A decade ago, NVIDIA's market capitalization was one-ninth of Intel's; combined with Tesla's, it was roughly equivalent to Xiaomi's, so it wasn't yet a giant.
Moreover, large models did not exist in 2017, and the AI that people talked about mainly revolved around computer vision. Therefore, the scenario I envisioned was to build a photonic computer capable of rapidly identifying and analyzing massive video streams, addressing issues like content security, detection of adult content, and identification of prohibited materials.
LatePoint: Why didn’t you choose to continue your research at school or first gain experience by working at a large company?
Shen Yichen: I received offers from Apple and Meta. At the time, Apple was interested in using silicon photonics to develop a watch capable of detecting blood sugar levels since silicon photonics can penetrate the skin without causing damage. Meta wanted to use silicon photonics to solve the transparent display issue for Oculus. While large companies offer abundant resources, you would have no choice but to focus on their specific areas of interest rather than pursuing something seemingly 'unrealistic' like optical computing.
The advantage of academia is greater autonomy, but the resources are extremely limited. Typically, you only get startup funding of one to two million US dollars and work with one or two Ph.D. students at a slow pace. Entrepreneurship, however, is different. After our company was established, we secured an initial round of financing amounting to 11 million US dollars and quickly assembled a team of over 20 people. Within a year, we had redone what took me five years as a Ph.D. student, and we achieved even better results. The greatest advantage of entrepreneurship is the significant acceleration in speed.
LatePoint: You managed to attract Maurice Steinman (Mo), who was AMD's highest-level technical expert (Senior Fellow), shortly after starting your venture. But I heard that you didn’t know each other beforehand. Did you directly email him?
Shen Yichen: I sent emails to more than forty semiconductor industry experts in the United States, contacting almost everyone I could find, including figures like Jim Keller (one of the most renowned chip architects in the semiconductor industry). Most people didn’t reply, but some did, and Mo was among them.
Later, he recounted to me that on the evening he received my email, he happened to be dining with his wife near our company. He had just completed AMD’s Infinity Fabric chip interconnect architecture project and was contemplating his next move. His ideal next step was to join a venture as the Vice President of Engineering, and coincidentally, my email invited him to take on that exact role. He later told me that this felt divinely orchestrated.
LatePoint: How did you come up with this approach? It sounds highly inefficient.
Shen Yichen: At the time, MIT had a strong entrepreneurial atmosphere, and we had heard many similar startup stories. Some might think, 'How ridiculous is this student?' But what if someone replies? Anyway, I had nothing to lose.
Moreover, as we are a student team, the people we can recruit will most likely also be students. However, the company urgently needs a seasoned industry veteran who has actually worked on chips and led engineering teams. Their inclusion will significantly assist us in attracting talent in the future.
LatePoint: What do you think was the key reason that ultimately attracted Maurice to join?
Shen Yichen: From the time I sent the email to his actual joining, there was actually an interval of about half a year. During that period, I almost tried to meet with him every week, sharing our latest progress and continuously discussing our assessment and vision for this endeavor. It was not until a certain point that he finally decided to join.
Moreover, before he joined, I asked him to help us interview candidates. So at that time, I was not only recruiting him but also trying every means to bring these truly industry-savvy individuals to our side.
LatePoint: I heard that after he joined, he took a 50% pay cut?
Shen Yichen: This is quite interesting. Actually, when he joined us, AMD's stock price happened to be relatively low. Coming to our company, his cash income was roughly halved, but I gave him stock options and the vision of building a company worth tens of billions of dollars, which far exceeded what he could have earned at AMD at that time.
Mo had previously worked only at large companies, with AMD being the smallest company he had ever worked for. Similarly, our current COO, Wang Long, whose previous company was Zheku (OPPO's self-developed chip subsidiary), was also the smallest company he had ever worked for. After working at a large company for twenty to thirty years, one may want to transition to a more dynamic and imaginative startup in the latter part of their career. Of course, their compensation would certainly far exceed mine.
There is a vast gap between technical feasibility and true commercialization.
LatePoint: You once mentioned that you were rather conservative, believing it would take three to five years for optical chips to replace traditional electronic chips. But today, nearly ten years later, electronic chips remain the mainstream.
Shen Yichen: At that time, the industry generally believed that Moore’s Law would reach its end in a few years; beyond 7 nanometers, the potential for further improvement in electronic chips was limited, and the next opportunity might lie in optics. However, no one expected Moore’s Law to be so resilient, progressing to 5 nanometers, 3 nanometers, 2 nanometers, and even pushing forward to 1 nanometer. Later, new technologies such as HBM, 3D stacking, and Chiplet also matured, further enhancing the competitiveness of electronic chips.
Another issue is that we ourselves have encountered numerous challenges in engineering, the optical supply chain, and various other aspects.
At the time, I believed that optical computing was the industry consensus, but looking back now, it may not be the case at all.
LatePoint: You are probably the person who understands optical computing the most in the world. The core technology roadmap was also proposed by you, yet you underestimated the difficulty of its implementation?
Shen Yichen: I still remember during our Series A round of financing in 2018, Zuo Lingye, a partner at Matrix Partners, asked us how long and how much money it would take for optical computing to truly achieve industrialization. I provided what I thought was already an aggressive estimate: three to five years, and 300 million US dollars. As a result, his only response after hearing this was: So cheap? Just this little amount?
I didn't understand what he meant at the time, but I do now. If I had come from Intel and had fully developed a CPU, I might have been more aware of how much it actually costs to commercialize a chip. But I came from academia, where my previous projects cost no more than 100,000 US dollars to produce results. Only later did I realize that developing a truly commercial GPU might require funding on the order of billions of US dollars; let alone creating an optical computing product aiming to change the computing paradigm, which would naturally require even more funding.
LatePoint: Your Vice President of Product Engineering, Maurice, comes from AMD and is a highly experienced veteran in the industry. Shouldn’t he know how many resources your team needs?
Shen Yichen: There’s another important reason. At the time, we estimated that if we were only making an optical computing chip applicable in a single scenario, 300 million US dollars would actually be sufficient, or perhaps even less. In fact, we succeeded in doing so. The Pace chip from Xizhi outperformed traditional electronic chips by nearly 1,000 times in specific scenarios.
However, we overlooked two key issues. First, the development cycle for chips typically takes two to three years. The scenario you envision when starting the project may no longer be relevant by the time the chip is completed.
For instance, computer vision, which we initially targeted, is no longer a focus; everyone is now paying attention to large models. The requirements for storage bandwidth and power distribution bandwidth in large models far exceed those needed for computer vision calculations, rendering our previous generation of products incompatible.
Secondly, while we demonstrated that optical chips could be 1,000 times faster than electronic chips in a specific scenario, the market demands a general-purpose chip that offers a tenfold speed increase and can be applied across more than a dozen real commercial scenarios. This represents an enormous challenge.
LatePoint: After the large model boom emerged at the end of 2022, the opportunities were clearly greater than those in computer vision. At that time, did your decision not to immediately initiate an optical computing chip for large models indicate internal hesitation about optical computing?
Shen Yichen: In fact, we had already prepared to start R&D in 2022, with the internal codename Vanguard. However, at that time, the company only had 300 million yuan in its account, and this single project would cost 500 million yuan. Moreover, the company did not yet have a product capable of true commercialization. After various assessments, the risks were deemed too high, so I decided to halt the project.
LatePoint: Couldn’t you continue through financing? Many people believe that only big gambles lead to big wins.
Shen Yichen: That was indeed a very difficult decision. At the time, many people on the team urged me to continue spending the money. Later, we even held an internal debate, dividing supporters and opponents into Red Team and Black Team, each given two months to substantiate their viewpoints.
But my core belief remained that the survival of the company was the most important thing. It turned out that this decision was correct, as the capital winter arrived in 2023 and 2024. If we had spent that money, our next round of financing would have been extremely challenging. Our past financing had always been relatively smooth, largely because we never had less than two years' worth of cash reserves, which kept us in the driver's seat and ensured we never signed any valuation adjustment agreements.
LatePoint: It seems you weren't entirely confident that investing this 500 million yuan would yield the desired product?
Shen Yichen: I believe that from a technical direction standpoint, this path is correct. However, given our engineering capabilities at the time, creating an entirely new optoelectronic integrated, vertically co-packaged chip posed a significant challenge, let alone succeeding on the first attempt.
At that time, there were very few companies globally working on this—mainly us and Lightmatter. Even now, they have not released a product. Precisely because no one in the world had done this before, no one fully understood just how difficult it would be.
LatePoint: Why can you proceed with it now?
Shen Yichen: We now have a business that generates continuous revenue—optical interconnects. After our IPO, our financing channels have also expanded, giving us a much stronger financial buffer. For example, last year, the company’s actual cash expenditure was only over 200 million yuan. Through this IPO, we raised 3 billion yuan, and with nearly 1 billion yuan in cash reserves, theoretically, even without considering revenue, we could sustain operations for up to twenty years at this rate of expenditure.
LatePost: With sufficient funding now in place, what are the major challenges from your current perspective?
Shen Yichen: The biggest challenge today remains cost and engineering complexity. For example, our PACE 3 was initially designed for larger-scale AI computing—not just single-card capability, but with consideration for how it will form larger systems in the future, or even enter hyperscale scenarios. This means it’s not as simple as making a single chip; it also involves advanced packaging, memory, and interconnects, resulting in high system complexity. Additionally, we are now using a fully domestic supply chain, where the maturity and yield of every link in the chain will impact the final cost.
However, even if the final cost does not reach the most ideal state, as long as it demonstrates performance advantages and enables small-scale deployment, that would still represent an important milestone for us. We can then gradually work toward further improvements.
LatePost: Will the performance advantages be sufficient to offset the cost disadvantages?
Shen Yichen: Based on our current design—whether it’s product specifications, system architecture, or specific engineering solutions—we believe that the performance of PACE 3 will be very strong, and its cost-performance ratio should ultimately be quite favorable.
But honestly, this is an extremely complex system. Even on a global scale, there are very few instances of optoelectronic chip systems of this complexity being successfully developed. Delivering it completely will indeed take time. The final results will only materialize after the chips return and testing is completed.
LatePost: What are your expectations for this chip?
Shen Yichen: My goal is to first identify at least one killer app—or a commercially valuable market—that proves optical chips can outperform electronic chips, whether it’s a $100 million market or a $1 billion market. This is highly achievable. From the current perspective, large model inference seems to be the application.
LatePost: Fully replacing electronic chips in large model inference?
Shen Yichen: Collaboration will definitely still be necessary. At this stage, optical computing has not yet reached the point where it can operate independently of electronic chips. There are still too many unresolved engineering issues, such as memory problems and non-linear issues. The relationship between them is somewhat analogous to that between CPUs and GPUs—it’s not about replacement but rather each handling tasks they excel at.
This is the best time to engage in optical computing in China.
LatePost: Your entrepreneurial journey initially began in the United States, but early-stage funding primarily came from domestic venture capital firms. Did you plan from the start to eventually return to China for development?
Shen Yichen: On one hand, domestic investors were more enthusiastic and competitive in their pursuit; on the other hand, U.S. funding came with many conditions that we found difficult to accept. For example, by 2019, when we sought financing again on Sand Hill Road (located in Silicon Valley), as a Chinese entrepreneur, I could already feel a noticeable shift in their attitude.
Google also reached out to us, stating clearly they would not invest in Lightmatter. However, after thoroughly questioning us about our technical details, they ended up investing in Lightmatter.
But personally, I truly believed that this technology would be most successful if realized in China.
LatePost: When did you decide to fully transition the company to a domestic structure?
Shen Yichen: Around 2023 to 2024.
In fact, my initial plan at the beginning of 2020 was still to expand the U.S. team to at least 200 people. At that time, the leasing market was very active, and securing office space for three years was challenging, so we directly rented an entire floor in the tallest building in downtown Boston, signing an eight-year lease. This office was carefully selected after much consideration. It had a great environment, no subway underneath, minimal vibrations, making it suitable for hardware R&D. To this day, the lease term for that office has not yet expired.
LatePost: What was the most direct impact of geopolitical changes on you?
Shen Yichen: Many previous collaborations could no longer move forward. Fortunately, the local governments of Beijing and Shanghai strongly supported the construction of domestic silicon photonics production lines, and we encountered benefactors who selflessly helped us. As a result, this initiative ultimately succeeded.
LatePoint: With the rise of large models today, the demand for computing power, interconnection bandwidth, and energy consumption pressures are rapidly increasing. Optical computing appears more attractive than it did a few years ago. However, there are still not many companies in China genuinely engaged in optical computing. Why is that?
Shen Yichen: Last year, over a dozen optical computing companies were established domestically, but they all face the same issue: the technical difficulty of optical computing may be even higher than that of developing GPUs, yet the resources capital markets are willing to provide are far less than those given to GPU companies. There are many GPU companies, each having raised nearly 10 billion yuan; however, among optical computing companies, we might still be the one with the most funding. Excluding this round of IPO financing, we've only raised around one to two billion yuan in the past. This amount is far from sufficient.
LatePoint: Because GPUs represent a higher-probability opportunity?
Shen Yichen: Yes. Investors may also recognize that optical computing could have a market potential worth tens or hundreds of billions of yuan in the future, but they perceive the probability of success as only 5%, which results in a valuation of merely several billion yuan.
The high valuations received by China’s large models and commercial rockets can largely be attributed to the benchmarks set by OpenAI and SpaceX in the United States, making it easier for domestic investors to price them. However, there is no such benchmark for optical computing. The U.S. has not yet seen any company achieve significant success, so investors calculate valuations based on very low probabilities of success. Low valuations lead to limited funding, and insufficient funding makes it challenging to develop an actual optical computing product. That was the deadlock at the time.
LatePoint: The U.S. has abundant capital. Why hasn't there been a surge in optical computing startups or successful products emerging to date?
Shen Yichen: One important reason why the U.S. hasn’t succeeded yet is my classmate’s company, Lightmatter, which is trying to catch up with an even larger giant—NVIDIA.
In China, initiating the development of a competitive optical computing product would require approximately 500 million yuan. This amount is roughly equivalent to the cost of a trial production run for a 12-nanometer or 7-nanometer electronic chip. In contrast, if you claim in the U.S. that you are going to create a product to outperform NVIDIA, this isn’t something 500 million yuan could solve—it may require an investment of 500 million U.S. dollars or more to develop a product capable of competing with cutting-edge chips like Rubin. This situation is quite discouraging.
Therefore, I believe this is a golden era for optical computing in China. Over the past five years, there has been no progress in domestically produced electronic chips due to restrictions on advanced processes, which, paradoxically, creates a more realistic opportunity for new approaches like optical computing to emerge faster.
LatePoint: Compared to ten years ago, how has your assessment of the prospects for optical computing changed?
Shen Yichen: In fact, it is still the same. Moore's Law will eventually reach its limit. The current process has advanced to the nanometer scale, and the size of an atom is only a few tenths of a nanometer. Progress beyond this point will undoubtedly become increasingly difficult.
Of course, no one knows whether new packaging, architecture, or material technologies might emerge, extending the path of electronic computing for a few more years.
Reinvesting all earnings into research and development is the most responsible action towards investors.
LatePoint: Xizhi initially started with optical computing, but why does your current revenue primarily come from optical interconnect business? Many outsiders also consider you a company specializing in optical interconnects.
Shen Yichen: By the end of 2022, the company reached a critical juncture. Although we still had funds for two to three years, we had yet to see a product capable of generating real revenue before our funds were depleted. Additionally, Hummingbird (Xizhi’s early optoelectronic interconnect chip), which was highly anticipated at the time, failed to achieve commercial success. Developing a new product would cost another 500 million yuan and require waiting another two to three years. That period was quite desperate.
At that moment, large-scale models surged. We quickly realized that future large model training could not rely solely on a single GPU but must depend on large-scale clusters. As clusters grow larger, data transmission between GPUs is more likely to become a bottleneck. As long as this bottleneck exists, optical interconnects have a chance. Fortunately, we had previously developed relevant technical reserves for optical computing, so we rapidly transformed optical interconnects into an independent business line.
LatePoint: It sounds like a decision the company had to make after hitting a wall?
Shen Yichen: There were not many options at the time. If I were in the United States, I could have relied on the technological lead in optical computing to secure another round of financing. Investors there would certainly prefer that I focus solely on R&D until one day being acquired by a major company. However, in China, especially under the capital environment at that time, no one would believe in just a 'story.'
However, I did not expect the demand for computing power to grow to this extent today. Recently, when meeting with a potential client, he mentioned the number of domestic chips undergoing tape-out next year. I was greatly shocked and am constantly being educated by the rapidly changing market.
LatePoint: I heard that during your recent IPO, the pricing logic was mainly based on being considered as an optical interconnect company?
Shen Yichen: I believe this is fairer to investors and easier to explain. At the time, I directly told the investors that if we only look at the optical interconnect business, the company is worth more than this price; as for optical computing, if it succeeds in the future, it will be an added bonus for you.
LatePoint: Some of your customers are GPU companies like Muxi and Biren. What are the biggest concerns for these traditional GPU manufacturers when collaborating with them?
Shen Yichen: The biggest concern for customers is stability. Copper wiring has been used for many years and is sufficiently mature. If something goes wrong, they know how to fix it. However, optical interconnects are different. Directly connecting two computing chips or GPUs using light has never been done on a large scale before. People worry about whether the signal transmission might result in errors: sending 01 here but receiving 10 there, which would disrupt normal system functionality. Additionally, factors such as temperature, dust, and vibration might affect transmission efficiency, which were all unknowns at the time.
Over the past two years, we have been working on compatibility with these GPU manufacturers, during which some instability issues arose. However, it was not the kind of problem where you spend a year adjusting and only achieve a 1% improvement. Typically, within one or two months, we can resolve stability issues to a level acceptable to customers. Now, even large-scale clusters of thousands of cards can be implemented, indicating that this solution has been largely validated.
LatePoint: Since all problems can be solved, why hasn't optical interconnect been widely adopted in hyperscale nodes yet?
Shen Yichen: It's because optical interconnect is still more expensive than electrical interconnect. Optical fibers themselves are cheaper than copper wires, but once you add the optoelectronic conversion chips at both ends, the overall cost increases. Therefore, as long as copper wires can be used, customers generally prefer them due to their affordability, maturity, and stability. However, when the distance increases, copper wires struggle to maintain performance. For example, beyond one meter, bandwidth and power consumption become problematic, making optical solutions necessary.
The core issue is how many computing cards you need to connect. Within a single server or a single cabinet, people still try to use copper as much as possible; however, once connections span across servers or cabinets, the necessity for optical interconnect becomes increasingly important.
LatePoint: When do you think optical interconnect will become fully widespread?
Shen Yichen: There are already clear signals domestically. Huawei’s CloudMatrix 384 has started large-scale implementation, connecting 384 cards across 14 cabinets with a significant number of optical interconnect modules, demonstrating that this path has been validated.
Major internet companies should begin large-scale adoption starting next year. We are already discussing design integration with them.
LatePoint: NVIDIA doesn't seem to be in a hurry? Its current strategy still focuses on fitting more GPUs into a single cabinet, using NVLink and copper backplanes to maximize scale-up within a single cabinet.
Shen Yichen: That's because it has already invested $10 billion in high-power consumption cabinet systems like NVL72, with liquid cooling, power supply, backplanes, and electrical interconnects forming an entire system. Given such a significant investment, it will certainly try to extend the lifecycle of electrical interconnects in the short term to make this system compatible with multiple generations of chips.
Google, on the other hand, is different. From the beginning, it concluded that computing clusters would ultimately need to move towards optical solutions, so it chose to invest earlier in optical interconnects, and even optical switching. Although adopting optics early on is more expensive, in the long run, there is an opportunity to recover system efficiency and scalability advantages.
LatePoint: When do you predict NVIDIA will fully transition to optical interconnects?
Shen Yichen: It is highly likely that the next-generation product will adopt optical interconnects. With each generation, interconnect bandwidth increases, and the higher the bandwidth, the shorter the distance copper can transmit. Problems solvable within dozens of centimeters today might become challenging in the next generation.
Judging from Huang's attitude, he may even want to bypass traditional optical modules and directly pursue silicon photonics co-packaged optics. Otherwise, at this year’s GTC, he wouldn’t have spent so much time discussing CPO (Co-Packaged Optics), but rather introduced a new technology allowing two chips to achieve high-speed interconnection within 5 centimeters, creating a denser cabinet relying entirely on electrical connections.
LatePoint: How do you think the market for optical interconnects will evolve in the future?
Shen Yichen: Competition will intensify. One example is 'Yi Zhongtian' ($Xinyisheng (300502.SZ)$、$Zhongji Xuchuang (300308.SZ)$、$Tianfu Communication (300394.SZ)$These module manufacturers and server manufacturers, although they previously did not produce chips, will certainly expand into the chip sector now that they have the financial resources; many venture companies are also beginning to gain access to capital.
LatePost: We heard that many new ventures outside are already referring to you as the 'previous generation' of optical computing companies.
Shen Yichen: I’ve heard about it. These companies need funding, and investors will definitely ask: Luxi has been in the industry for eight years, with a strong team, sufficient funding, and is now publicly listed—so how can you surpass it? Therefore, they must tell a different story. Thus, we are honored to become the dragon everyone wants to disrupt. I think it’s fine. It also gives us more motivation to move forward.
LatePost: Although you are now generating revenue from optical interconnects, Luxi's financial statements show that your book losses are widening. How long do you think it will take to improve this situation?
Shen Yichen: To be honest, I am not overly focused on achieving profitability. If I really wanted to, there would be opportunities in the next year or two. However, I believe that this is not the most important goal at this moment, especially in such a rapidly changing and competitive environment. Running a company with the mindset of achieving quick profitability might cause us to miss out on many opportunities and also suggests that the company lacks substantial R&D efforts.
However, I will control the losses and have internally proposed narrowing the deficit as much as possible, mainly to ensure that no one overspends. Investing all earned revenue back into R&D is the best way to honor our investors.
Some things just need to be tried by someone, even if some may end up being sacrificed.
LatePost: We heard that during your Ph.D. studies at MIT, you started two ventures based on your academic research, both of which eventually failed?
Shen Yichen: The first venture came from a Science (a top-tier international academic journal) level research achievement, which focused on optical nanomaterials, or what could be understood as a special type of optical film. Its most unique feature is that it displays completely different visual effects from various angles: from some angles, it looks like a mirror, while from others, it resembles transparent glass.
We initially intended to apply it to solar panels, hoping to improve the efficiency of light energy utilization and solve energy problems. However, the film deteriorated within three days of application. After much investigation, we discovered that its best practical application was as an anti-peeping film for mobile phones. It could not even be used on computers because the film could not be made large enough. Later, 3M purchased the patent for several hundred thousand US dollars, but it has not been utilized to this day.
LatePoint: When conducting experiments in the lab, did you not discover that it would fail on solar applications after three days?
Shen Yichen: Such issues are often overlooked when publishing papers. For instance, when we published our first paper on optical computing, we did not consider what would happen if the chip ran for an entire day; as long as it could successfully operate for even a second, it was sufficient for a research paper.
The second time, we developed a type of transparent display glass. It appeared transparent under normal conditions, but when illuminated by a specific laser projector, images could be displayed, effectively turning the glass into a screen. Mass production of this technology was not difficult, but the business model was not viable — a film covering one or two square meters could only be sold for about 200 yuan, while the specialized three-color laser projector required cost 20,000 yuan. After some calculations, we realized that producing this film was less profitable than acting as an agent to sell projectors.
LatePoint: It seems like the perennial issue with technology startups, underestimating the difficulty of transforming a research paper's findings into a commercial product.
Shen Yichen: Yes. For example, in our current work on photonic chips, we are essentially standing on the shoulders of giants, leveragingTaiwan Semiconductor (TSM.US)、SMIC (00981.HK)such mature manufacturing capabilities. However, the situation was different when we worked on optical films. There were no existing nanoscale optical film production lines in the market. To transform it from a laboratory sample into a product, we had to build our own production line, which required significant investment and incurred very high costs.
LatePoint: But why couldn't you continue with the phone film? The global market for anti-peeping films for mobile phones should still be worth hundreds of millions of US dollars annually.
Shen Yichen: By the second year, I had developed my photonic chip. I simultaneously began exploring photonic chips and clearly felt that the market's interest in this technology was significantly higher.
If it were something more impactful, such as improving solar efficiency by 10%, I would have been highly motivated to continue working on it because it addresses energy issues. However, if it merely involved making an already dim anti-peeping film slightly brighter, I would find it hard to consider it a sufficiently important matter.
LatePost: What is the most important criterion, market size or the scale of the company?
Shen Yichen: It still comes down to the intrinsic value of the matter itself—whether it’s something I would be proud to tell my parents, friends, or even my children about in the future.
Take my mentor, Marin Soljačić, for example. He became a full professor at MIT at a very young age, and one of his most famous achievements was inventing wireless power transmission, also known as wireless charging technology.
To give another example, if China could create its own NVIDIA, that would certainly be remarkable. But if I were to go back ten years, I would still choose to work on optical computing rather than GPUs because, even if optical computing only captured 10% of the market share in the end, it would still represent a more groundbreaking achievement. The traditional GPU path would have been pursued by many others regardless of my involvement; but with optical computing, if we didn’t pursue it, the industry as a whole would likely have progressed much more slowly than it has today.
LatePost: Apart from being the 'first mover,' what else draws you to optical computing?
Shen Yichen: What one chooses to do is not particularly important. It just so happens that this is what I research. You can see that I’ve also done entrepreneurial ventures in optical films. But the core issue remains whether the technology itself holds value (Technology matters).
LatePost: Is this influence something your mentor instilled in you?
Shen Yichen: I think his greatest contribution to my entrepreneurial journey was that he did not take a large controlling stake. This is actually quite rare among professors who guide student startups.
In many cases involving professors and students starting businesses in China, it is common for the professor to hold a significant proportion of shares without committing to working full-time in the venture, while the student ends up with a small stake and works full-time in the company. In the end, success stories are few and far between because, should the venture fail, the student bears the entire opportunity cost, whereas the professor incurs minimal loss and can continue teaching and mentoring the next batch of students.
LatePost: Both you, Han Bicheng of BrainCo, and Wen Shuhao of XtalPi, belong to the group of entrepreneurs with scientific backgrounds who emerged around 2017. As fellow technologists-turned-entrepreneurs, how do you think you differ from today’s wave of entrepreneurs focused on large-scale AI models?
Shen Yichen: The current wave of large model startups in China includes many founders who have already gained industry experience or have previously started businesses. For instance, Yang Zhilin and Yan Junjie did not embark on entrepreneurship immediately after graduating from school. Professor Tang Jie is somewhat different, as he has been a university professor.
However, our wave was different. Many individuals had just completed their Ph.D.s or were even still doctoral candidates when they began commercializing their research papers and lab results. That period marked the first time scientists were systematically pushed to the forefront of entrepreneurship by capital. If this had occurred five years earlier, the opportunities might have been entirely different, and investors may not have dared to back such deep-tech projects.
LatePoint: Who do you consider the benchmark entrepreneur for technology startups?
Shen Yichen: Zhang Zhongmou. However, I believe his greatest achievements are not limited to technology but also include several critical business decisions. For example, choosing to establish an independent wafer foundry and pursuing the path of semiconductor contract manufacturing; entering CPU process technologies at the right time; avoiding competition with clients by not developing proprietary chips, including refraining from entering memory production like Samsung. These decisions were later proven to be highly significant.
By comparison, we have yet to achieve such commercial validation today. Of course, I hope to make strategic decisions of a similar magnitude in the future.
LatePoint: Do you still compare yourselves to NVIDIA now?
Shen Yichen: I think it’s difficult for us to position ourselves as a counterpart to NVIDIA anymore. A more realistic path is to become a leading company in optical chips within future data centers. We estimate that optical chips could account for 30% to 50% of the data center market in the future, which represents a substantial opportunity in itself.
The two most important tasks moving forward are: securing a strong position in optical interconnects as soon as possible and identifying the first truly compelling killer application for optical computing.
LatePoint: You’ve been working on optical computing for ten years now. Have there been moments when you doubted whether the path of optical computing itself might not be viable?
Shen Yichen: From the perspective of technological advancement, it’s essential that someone attempts this. Moreover, what if Moore’s Law truly reaches its limit? If successful, the potential upside would be extremely high.
LatePoint: But is it also possible to become cannon fodder?
Shen Yichen: Even if we become cannon fodder. We were the first to propose this route; if we don’t take on this role, who will? Besides, this path deserves to have its sacrifices. Look at quantum computing—so many companies are still pursuing it.
Looking to pick stocks or analyze them? Want to know the opportunities and risks in your portfolio? For all your investment-related questions,just ask Futubull AI!
Editor/joryn