Li Feng in Conversation with Ji Yu: Understanding NVIDIA, Deconstructing NVIDIA, Challenging NVIDIA

On the global tech stage, NVIDIA is undeniably a blazing star. Over the past two weeks, it briefly claimed the top spot in U.S. market capitalization, only to suffer three consecutive days of declines shortly after — this volatility has drawn widespread market attention.

While public discussion of NVIDIA and AI chips remains as heated as ever, opinions on future trends diverge. Not long ago, Feng Shu (Li Feng) sat down with Dr. Ji Yu, founder and CEO of Xingyun, for an in-depth conversation covering NVIDIA's monopoly, the opportunities and challenges of AI chip entrepreneurship, and the boom and potential bubble in technology startups.

Dr. Ji holds a bachelor's degree in physics and a Ph.D. in computer architecture, both from Tsinghua University. He has published dozens of papers in top-tier journals and conferences, and received the CCF Outstanding Doctoral Dissertation Award. After graduating, he was selected for Huawei's "Genius Youth" program, where he served as compiler architect at Huawei HiSilicon Ascend, leading R&D on multiple cutting-edge projects. He is also an active Zhihu contributor under the handle mackler. In 2023, Ji embarked on his own AI chip venture — put simply, he aims to use NVIDIA's own methods to deconstruct and challenge NVIDIA.

Their conversation covered, but was not limited to, the following topics:

• What mainly accounts for the currently inexplicable aspects of both the human brain and large models?
• What paths will large model development take going forward?
• Is NVIDIA's stock overheated?
• How should we view NVIDIA's runaway dominance? Why has no one been able to fully replace it so far? Given this, how can startups break through in this space?
• Why is it that the more versatile a chip becomes, the more likely it is to be deconstructed, yet the less likely it is to be replaced?
• What factors might determine success or failure along possible paths for chip entrepreneurship?
• Why is the most important competition in the semiconductor industry not technology itself, but rather how one defines the form of computer systems?
• Amid the many realities of U.S.-China competition, how should we view China's chip industry in the coming years? How can we find underlying certainties beneath surface-level uncertainties?
• Which causes more damage: supply cutoff risk or replacement cost?
• Can Huawei's Ascend 910B serve as a substitute for NVIDIA's products?
• How should we view the opportunities for "straight pursuit" versus "curve overtaking" in chips?

We hope this offers fresh perspectives and food for thought. You're also welcome to search for and subscribe to "High Energy" (高能量) on Xiaoyuzhou, Apple Podcasts, or Ximalaya to listen to this episode.

/ 01 / Beware of Getting Lost in Technical Details While Missing the Underlying Logic of Technological Development

Li Feng: You got into Tsinghua through the physics olympiad — physics is probably one of the toughest majors out there. What was your experience like at Tsinghua?

Ji Yu: My undergraduate classmates fell into two broad categories: those who came in through competition保送, and those who got in through the gaokao and were academic top performers. The two types were very distinct. The latter kept studying hard after entering, with excellent grades. The former were more "free-spirited" — their grades might not have been top-tier, but they loved tinkering with all sorts of problems all day.

Li Feng: You were the former.

Ji Yu: Yes.

Li Feng: Tsinghua's physics department offered many opportunities to go abroad. You chose to stay in China and join Huawei after graduation — what was your thinking?

Ji Yu: After finishing my physics bachelor's, I went to the computer science department. In high school, with limited exposure, you tend to pick majors based on what you think you like.

But once in university, I gradually realized that continuing down the physics path probably wouldn't offer that many big opportunities. The frontier directions people pay attention to today — condensed matter, high-energy physics, and so on — often require enormous teams to push forward, which diverged from my earlier assumptions. So after undergrad, I switched to computer science.

Computer science and physics are two very typical fields, representing two extremes. Physics abstracts the most concise laws from the messy real world, while computer science starts from the most concise laws to create a complex edifice of algorithms. Having both backgrounds gave me a cross-disciplinary perspective. In the industry, people's thinking easily gets trapped in specific technical details, causing them to miss the developmental thread behind technology and some underlying logic. By contrast, my physics training allowed me to perceive the fundamental patterns underlying computer science, a highly engineering-driven industry.

During my Ph.D., I did a lot of research on neuromorphic computing. After graduating, I joined Huawei HiSilicon to work on AI chips, specifically on software compilers — hoping to solve software ecosystem problems through compiler approaches. After large models emerged, I wrote quite a few analytical pieces on my Zhihu column (ID: Mackler), dissecting things from top-level demand down through software, algorithms, and chip layers, as well as strategies for domestic chips to break through.

In 2023, I started my own company, wanting at this juncture to explore an approach that truly aligns with both market principles and the developmental laws of the computer industry, gradually deconstructing the high-end AI chip market monopolized by NVIDIA.

/ 02 / Whether It's the Human Brain or Large Models, the Inexplicable Parts Stem Mainly from Micro-Level Unmeasurability

Li Feng: Your research direction was neuromorphic computing. I've always been curious about something. Right now, people can't fully attribute the path of algorithm evolution, and our understanding of how the human brain processes, remembers, and translates various types of information is similarly incomplete. If we compare these two forms of intelligence as two branches, will they eventually point in completely different directions?

Ji Yu: I tend to think of the brain as an extremely complex system. Looking purely from the neuron perspective, it's hard to explain how such a vast, complex system gets constructed. Of course, this is because our understanding of the brain is still incomplete — it's difficult to explain its working principles in simple terms.

Similarly, although large models are human-designed, people currently struggle to fully explain how they achieve their current results. We can try to roughly understand that a large model's capabilities are a stack of several layers: one layer is a computer with strong computational power; deep learning is another layer — a method where computers mimic how the human brain learns, with working principles similar to neural network operations; language models provide yet another layer of support. Each layer serves as the foundation for the one above.

Of course, in my view, the brain's complexity far exceeds that of deep learning, and deep learning's complexity exceeds that of computers. Overall, both large models and the human brain are difficult to explain through simple principles.

Li Feng: So you believe both have unknowable aspects.

Ji Yu: A more precise term would be "unmeasurable." "More is different."

Take a simple example. At the macro level, we can measure a gas's temperature, volume, and pressure. But at the micro level, these concepts don't exist. Inside the gas are countless molecules engaged in extremely complex motion. Once the system is large enough and its complexity reaches a certain threshold, macro concepts like temperature, volume, and pressure emerge. The human brain and large models work similarly. At the macro level, we see many remarkable capabilities, but at the micro level, they're unmeasurable.

So while today's large models have all sorts of problems, that doesn't mean we should start from scratch and颠覆ively develop some entirely different algorithm. The various issues with current large models likely stem from their complexity not yet reaching a specific level. As certain dimensions of large models grow larger during their development, people may devise new methods to manage and solve these problems, leading to an intelligence form that builds upon current large models but can solve more problems and exhibit more brain-like traits.

/ 03 / Is NVIDIA's Stock Overheated?

Li Feng: The three companies people have discussed most lately are NVIDIA, Microsoft, and Apple. Over the past month, the three have practically taken turns sitting atop U.S. market cap. Do you think NVIDIA's stock has a bubble?

Ji Yu: Two main factors support NVIDIA's stock price: the development of the large model industry, and NVIDIA's monopoly position within that industry.

Of course, the large model industry itself has some bubble elements, but considering NVIDIA's monopoly, I don't think its stock price is bubbly. The underlying logic supporting NVIDIA's stock is that unless some genuine disruptors emerge to shake up the market, NVIDIA will likely maintain its monopoly.

Li Feng: NVIDIA dominates AI chips. Historically, there are some comparable companies — Intel during the fastest-growing era of personal computers, and Qualcomm during the fastest-growing era of wireless communications. But neither Intel nor Qualcomm ever achieved market cap multiples that so thoroughly crushed other chip companies to this degree.

Intel is probably the closer parallel. Starting from IBM-compatible machines, Intel established itself as the most foundational "bricks and mortar" layer of the industry. Whether software or operating systems, everyone built around Intel's framework. Reaching this level wasn't simply because Intel's CPU performance was unmatched, but because it became an industry standard — this monopolistic effect is extremely powerful.

One thing Intel did back then was particularly notable and impressive: its "Intel Inside" marketing program. OEM partners could accrue advertising credits toward Intel chip purchases by putting the Intel logo in their ads and on their computers. Although Intel was a not-easily-visible B2B brand, it had both strong technology and a successful marketing scheme that captured consumer mindshare, allowing Intel to reinforce its PC market share and leadership. Even today, this monopoly continues to support Intel's important position in the computing industry.

I'm curious — how do you view how NVIDIA arrived at its current monopoly?

Ji Yu: NVIDIA entered through 3D graphics accelerators. In the 1990s, demand in this area rose, and NVIDIA's 3D graphics accelerator, as an expansion card compatible with Intel's architecture, successfully secured the discrete graphics niche and avoided being integrated by Intel.

In 1999, NVIDIA launched its era-defining product — the GeForce 256 graphics card, the first product built around the "GPU" concept. From there, NVIDIA began expanding its ecosystem, aiming to push its GPU — which had carved out an independent niche in 3D graphics acceleration — into general-purpose computing, competing with CPUs in a different form. You could say NVIDIA spent nearly a decade laying the groundwork. While it didn't make much visible progress during that period, it did get all the necessary preparations in place.

It wasn't until 2012, when deep learning made major advances, that NVIDIA was able to leverage AI to secure a critically important position in the computing industry. Moreover, as deep learning itself grew in importance, NVIDIA gradually became an enormously powerful presence in data centers.

Li Feng: People outside the chip industry might be curious — NVIDIA entered through 3D graphics accelerators. Why did it have such inherent advantages in adapting to artificial intelligence?

Ji Yu: Hardware and algorithms have a mutually influential relationship. First, hardware shapes algorithm development — meaning, if you have better hardware, algorithms may evolve in directions that better suit that hardware.

Broadly speaking, both 3D graphics acceleration and today's AI depend heavily on high compute power. After NVIDIA developed its 3D graphics processors, it realized such high compute power must be applicable elsewhere, so it generalized that compute capability.

The question was: what industry could sustain such high compute power? NVIDIA placed many bets — on supercomputing, physics simulation, and many other fields, including AI. Ultimately, AI developed rapidly under its impetus.

/ 04 / Countering Industry Inertia: Join the Inertia, Become Part of It

Li Feng: The AI investment boom in China roughly started after 2014, focused mainly on autonomous driving, security, and speech recognition. Since AI was already showing promise then, and NVIDIA was already pulling ahead, why couldn't major companies building underlying chip architectures — including Intel — catch up to it in the subsequent 5 to 10 years?

Ji Yu: When we talk about "catching up," we usually think in technical terms — how to build a chip that meets today's AI needs in compute power and other dimensions. But I see it differently.

Looking at computer history, the emergence of IBM-compatible PCs in 1981 was a critical watershed. Before compatible PCs, the computer industry followed a vertically integrated model — IBM made one system, DEC (Digital Equipment Corporation) made another.

At that stage, for end consumers or downstream industry customers, whoever's product better met your needs, you used theirs.

After compatible PCs emerged, the computer industry rapidly shifted to a horizontal layered architecture. Microsoft focused on operating systems, Intel on CPUs, software companies on software. This vertically integrated technology stack formed by horizontal layering proved more competitive than any single company's standalone product, quickly displacing the old vertically integrated model.

This horizontal architecture also brought profound consequences. The computer industry's broad structure entered a very stable state. No matter how market demands changed, you couldn't casually replace any middle layer, because such replacement would cause massive disruption to that layer's upstream and downstream. Replacing Intel x86 would mean throwing out all of Microsoft's historical accumulation; replacing Microsoft would require Intel and all upper-layer applications to be discarded as well.

The result was that industry inertia became enormous. NVIDIA's strategy as a newcomer against this inertia was to actively join the inertia, become part of it.

Similarly, today, as various AI applications emerge, NVIDIA has become everyone's default mainstream product. At this point, crudely replacing it is extremely difficult, because this isn't merely a technical competition. The industry doesn't say: NVIDIA's compute is stronger today, so I'll buy NVIDIA; Intel's is stronger tomorrow, so I'll switch to Intel, oscillating between the two. With industry inertia already so massive, market-supported replacement cycles become very slow and prolonged.

Precisely because of this, whether Intel or other chip companies, what they must consider isn't merely how to catch up on compute power — that's just a baseline condition. More importantly, they need to figure out how to overcome inertia.

Li Feng: How much did NVIDIA's CUDA system contribute to its current industry position?

Ji Yu: The role was crucial. CUDA is what allows NVIDIA's monopoly to persist. Many people fall into a thinking inertia, believing that for startups to break through, they either catch up or surpass through different technical paths. But whether catching up or surpassing, the essence is hoping to replace the incumbent. However, with the computer industry's horizontal layering already possessing such massive inertia, replacement simply won't happen.

**/ 05 / ** No Matter How Great System Inertia, Evolution Doesn't Stop — and Evolution Creates Opportunities for Metabolism

Li Feng: Given this, what opportunity did you see when you started your company?

Ji Yu: No matter how great system inertia, evolution doesn't stop. As long as there's evolution, there are opportunities for metabolism. The market doesn't reject improving certain metrics while maintaining system structure unchanged. NVIDIA itself is an example — because people had needs for greater compute power and stronger interconnectivity, GPUs seized the opportunity and found their opening.

What must be emphasized is that this metabolism isn't inevitable. It may not be an overwhelming trend, but rather something that depends on human action.

Within the industry's default inertia, certain demand changes will occur, creating opportunities to reshape what computers should look like. For instance, today's computers are by default understood as Intel's or AMD's x86 CPU as the processor, plus NVIDIA's GPU, plus some network cards, or plus some NVLink communication protocols. But as demands evolve in the future, beyond quantitative changes within existing horizontal layers, could a new "organ" emerge? This is a demand we can design and cultivate.

Compared to past deep learning, current large language models have created massive changes in requirements for computer systems. In this process, we can either follow industry inertia with band-aid solutions — improving whatever needs improving — or we can redesign, gradually splitting off a new form from today's existing architecture, guiding computer systems toward entirely new forms. This is also an important reason why NVIDIA was able to carve out a position in computer systems in the past.

**/ 06 / ** How New Demands Get Met: The Contest Between Improvement and Deconstruction

Li Feng: If large language models trigger an opportunity to "grow a new organ" as a new type of demand at a different level, as large models evolve going forward, will there emerge needs that NVIDIA cannot fully accommodate? And what type of "organ" might grow from that? Could NVIDIA become the Intel of the previous cycle at some future point?

Ji Yu: One thing needs clarification. NVIDIA was able to enter the market not because Intel couldn't support new demands, or because the system had problems that Intel refused to solve, thereby yielding an opportunity to NVIDIA. Rather, the GPU as a "new organ" was essentially something NVIDIA actively pushed forward.

Whether this new demand gets met through Intel gradually improving its CPU, or through someone else building a GPU to better satisfy it — this is fundamentally a contest between these two routes. NVIDIA's victory came from winning this contest.

In the past, Intel's CPU handled everything in computers. NVIDIA stepped forward and said: our GPU only handles computation. Ultimately, it completely surpassed Intel's CPU in the computation dimension, so this task was successfully deconstructed out. Today, NVIDIA's GPU has also become extraordinarily massive in the computation dimension. The more responsibilities a single chip assumes, the greater the opportunity to deconstruct it.

Similarly, it's not that there are currently new demands that NVIDIA cannot satisfy. It will do its utmost to improve GPU capabilities in all dimensions, trying to meet all market demands. We need to think about how to win this contest between improvement and deconstruction.

In this context, our starting a company is essentially hoping to prove that computer systems don't have to consist only of CPUs and GPUs — there can be a new coprocessor (a chip that offloads specific processing tasks from the system's microprocessor). A new system combining these three would be more efficient than a CPU-plus-GPU combination alone.

To achieve this, the most critical step is deconstructing the tasks currently handled by GPUs — determining what belongs to the GPU and what belongs to the coprocessor.

Our current deconstruction approach is to further split computation into compute-intensive computing and memory-access-intensive computing. The compute-intensive portion is what NVIDIA's GPUs excel at; after deconstruction, this can continue to be handled by GPU products. The memory-access-intensive portion can be handled by coprocessor chips.

We're not saying we make two chips to beat one NVIDIA chip. Rather, we use one of NVIDIA's highest-end chips plus one of our supplementary chips as a new combination, to compete against NVIDIA's highest-end chip alone.

This logic is something we must actively pursue — the industry won't spontaneously move in this direction.

So-called "industry spontaneity" is actually everyone trying their utmost to satisfy all visible demands. The result is that industry monopoly keeps strengthening. NVIDIA keeps releasing new products with ever-stronger performance metrics and ever more comprehensive dominance. But just as NVIDIA was previously able to deconstruct the seemingly impregnable Intel, its breakthrough method may itself be used to deconstruct it.

There's a very counterintuitive conclusion here: the more all-encompassing a chip becomes, the more likely it is to be deconstructed — but the less likely it is to be replaced.

Li Feng: It was precisely because people's demands for entertainment and media grew increasingly complex, and requirements for PC chips' graphics processing capabilities kept rising, that a foundation was laid for NVIDIA to emerge as an independent player.

Currently, as AI finds broad application across industries, the information and demands it must process are also growing more numerous and complex. NVIDIA's compute power may also hit ceilings, and the fervor around large model applications and related investment may subsequently cool. How would you assess future trajectories?

Ji Yu: Actually, large language models themselves already encompass all the various possibilities we just discussed. Compared to deep learning, which mainly consumed compute power, today's large models consume much more comprehensively — they need very high video memory, very strong interconnectivity, and of course very high compute power. So it's not a matter of future demands becoming richer and therefore causing certain changes; large models themselves represent an enormous change.

Of course, when we speak of the possibility of a "new organ" emerging, the computer industry won't spontaneously trend toward ever-greater deconstruction. There must be certain deliberate actors — like us — who consciously shape boundary conditions that make computer system architectures potentially develop in a direction we prefer to see, and that is also more beneficial for the industry.

If no one can accomplish this deconstruction, the industry will continue along the inertia of GPUs satisfying all demands.

**/ 07 / ** Can What Computers Should Look Like Be Redefined?

Li Feng: What large model application scenarios is your coprocessor mainly targeting?

Ji Yu: We're targeting all current large language model needs, primarily the text generation phase, because this stage is a very typical memory-bound phase — one that requires massive memory bandwidth.

Li Feng: Could you break that down a bit?

Ji Yu: Computing today comes in many forms. Some tasks are extremely compute-intensive, while others may not require that much computation but need to read vast amounts of data. The latter corresponds exactly to where text generation stands today.

The current situation is that NVIDIA's chips have strong compute power and fast data read speeds. They efficiently handle the first type of task and are also highly efficient at the second type — seemingly flawless.

But if we deconstruct this, what if some chips specialized purely in computation while others focused solely on massive data reads? Under this boundary, the chip responsible for the latter part could abandon certain constraints and push more aggressively to meet customer needs for data reading more efficiently.

Li Feng: If the current market consisted only of internet giants competing on large models and their applications, would this demand pattern you describe still exist? Would startups still have market space?

Ji Yu: It would still exist, because this is a universal need — not specific to any particular large model company. There's industry consensus on this; the question is how to translate that consensus into computer system architectures and technical specifics.

Everyone knows read speeds need improvement, and NVIDIA is pushing hard there too. They're still boosting compute, but also improving on this other dimension. Other companies may be thinking: "I used to compete with NVIDIA on compute; now I also need to compete on memory bandwidth."

We still hope to shape alternative forms for what computers should fundamentally look like.

Li Feng: Since there's consensus, many will see this opportunity, making competition fierce. Though architectures will differ, whose products are more likely to be adopted?

From chip design to tape-out, trial production, mass production, and finally large-scale adoption by mid-sized or larger enterprises — this takes considerable time. It severely tests chip designers' ability to anticipate market changes two or three years out.

I think this is the most challenging aspect of this industry — both the greatest pain and the greatest opportunity. It comes down to who, standing today, can accurately bet on changes two or three years from now. From your perspective, what factors might determine success or failure along these possible paths?

Ji Yu: Currently, many semiconductor companies still assume by default that computer systems should look like this: some CPUs, some GPUs, connected together to run all horizontally layered tasks.

This default leads everyone to fight for market share of a particular device. Some want to capture GPU share by trying to replace it with TPUs or XPUs; others think current interconnects are insufficient, so they push interconnect metrics higher.

What we hope to compete for is the evolution of computer system architecture toward a different direction. For instance, the need for faster data reading can actually be addressed through different approaches.

Going back 20 years, if one had fully accepted the computer system architecture of that era — one CPU plus expansion cards — the competition would ultimately have been about device replacement, fighting for that CPU portion. That, in my view, was destined to fail, because you cannot directly replace something that already exists and matters greatly.

Like NVIDIA back then: accepting Intel's CPU as irreplaceable, then at the macro level transforming computer systems into a CPU+GPU architecture, offering a path to higher performance than pure CPU systems — that's how they broke through.

This competitive approach isn't conventional semiconductor competition. Semiconductor technology itself certainly matters, but more important is how you define computer system architecture — which tests a company's deep understanding of system ecosystems and developer needs.

Of course, since you're adding something new to computer architecture, you need to know how to draw this boundary and how to make this boundary and combination achievable.

So this is highly challenging, and also means significant opportunity.

**/ 08 / ** When the Large Model Bubble Bursts, Cost Tensions Will Intensify

Li Feng: AI will certainly be ubiquitous in the future. However, large models are expensive to use today because they consume enormous compute and power. There's some consensus that large models themselves currently have a certain degree of investment bubble. Assuming this bubble bursts in the next two to three years, or competition over large model technology itself subsides, with greater focus on applying the technology — what variables would affect the emergence of this new architecture you described?

Ji Yu: First, cost factors. If the large model bubble dissipates, cost tensions become more pronounced. Our push to evolve computer systems in new directions is fundamentally about solving already prominent cost problems.

Let's use NVIDIA's past as an example.

It wasn't that CPUs lacked graphics processing capability — they had it, but you needed to buy Intel's highest-end processor to get corresponding performance. Such top-tier products were extremely expensive.

NVIDIA at that time actually offered an alternative: individual gamers who couldn't afford such expensive professional-grade products could buy the cheapest CPU, pair it with a not-very-expensive NVIDIA GPU, and still run 3D games. This was fundamentally a cost-driven proposition.

So what we're trying to do is similar.

Currently, NVIDIA's machines perform increasingly well but also sell at ever-higher prices. Cards have gone from tens of thousands to hundreds of thousands of RMB; servers from millions to tens of millions. If today we can deconstruct this need and say: "OK, you buy some relatively inexpensive NVIDIA consumer-grade cards for one portion, then supplement with our much cheaper new hardware for other parts" — this new combination can substantially reduce costs.

In fact, today's computer industry has reached a point somewhat similar to when IBM launched mainframes. IBM's mainframes were designed for banks' high-throughput data processing needs; today's NVIDIA "mainframes" may be designed for AI training or inference.

The emergence of such "mainframes" is fundamentally unfavorable for industry development, because the internet industry was previously built on very low-cost white-box systems. We hope through structural innovation to make this system's components more open and lower-priced.

/ 09 / US-China Competition Accelerates Chip Entrepreneurship in China

Li Feng: You've discussed many thoughts on AI chip entrepreneurship. Factoring in doing chip startups in China specifically, how do these considerations change?

Ji Yu: Actually, the current US-China relationship environment can accelerate our business logic.

Li Feng: Why?

Ji Yu: From a pure market perspective, given industry inertia, getting people to use less NVIDIA product and accept something new — driving this transition would be quite slow. Even getting people to buy more GPUs took NVIDIA a very long time.

For today's large model industry, beyond cost reduction, seeking supply chain security beyond NVIDIA is also an important driver. The tendency is to say: we're not trying to directly replace NVIDIA, but to reduce NVIDIA's importance in this system. This atmosphere works in our favor.

Li Feng: So there are two layers: First, current US chip restrictions have limited domestic enterprises' procurement of NVIDIA's high-end chips. If a new combination of "available NVIDIA chips + other types of coprocessors" can achieve performance comparable to higher-end NVIDIA chips, this will stimulate relevant market demand.

Second, any company needing to procure such chips at scale and deploy them to applications must consider underlying architecture security. These enterprises also have considerable motivation to find more options beyond NVIDIA to ensure supply chain security and stability.

/ 10 / Supply Cutoff Risk vs. Replacement Cost: Which Hurts More?

Li Feng: Setting aside your identity as an entrepreneur, and looking from the perspective of your Zhihu persona (ID: mackler), how do you view the development of China's chip industry in the coming years, given the practical impacts of US-China competition?

Ji Yu: China has numerous internet companies built upon horizontally layered industry structures, and the upper layers are becoming increasingly substantial. So regardless of whether or how the US and China decouple in the future, the industry's underlying logic is very difficult to violate — you'd be hard-pressed to find alternatives.

Over the past decade, whether in China or elsewhere globally, very few companies have truly affected NVIDIA's market share.

So facing potential supply cutoff risks from US-China relationship variables, if you adopt alternative solutions that confront industry inertia head-on and overturn the industry's upper-layer stratification, the cost is borne by customers. Behind this lies a weighing of which hurts more: supply cutoff risk or replacement cost?

From my own perspective, overturning everything in the upper layers is a cost currently completely unacceptable to everyone — even more unacceptable than supply cutoffs.

Defense against supply cutoff risks has indeed given domestic chips considerable room and opportunity for trial and error. But from the domestic chip perspective, people still need to keep thinking about which opportunities can both meet market-driven demand and represent market space for Chinese chips under US sanctions.

Li Feng: Since we mentioned domestic chips, there's an unavoidable topic — your former employer, Huawei. How do you view Huawei's current stage in chips and process technology broadly, and what阶段性进展 it might achieve? As CEO of an AI chip startup, how do you view Huawei's AI chips?

Ji Yu: Huawei is a company extremely adept at end-to-end solutions, including its "customer-centric" philosophy and so forth, which constitute its foundational character. This character has contributed significantly to Huawei's great success in communications, mobile phones, and even automotive sectors.

From my own perspective, Huawei is better suited to such vertical industries. It excels at end-to-end full-stack integration because it has software people, hardware people, product people — end-to-end integration best matches its character.

This DNA has its strengths and relative weaknesses. Once in horizontally layered industries, many of its approaches actually face challenges.

In other words, today when people ask whether Huawei's Ascend 910B can replace NVIDIA? I think you can neither say it can't replace nor that it can — it depends on what kind of replacement you want.

For example, if you're facing industry customers wanting an end-to-end system that can solve their large model or other AI application needs, this we can understand as a ToB replacement.

This is achievable, because today to run large models, whether you use NVIDIA's solution plus the entire open-source software stack above it, or Huawei's chips with adapted corresponding software — neither is problematic. But from the internet and large model companies' very developer-oriented DNA, looking at replacement of horizontally layered computer systems, it's actually very difficult to replace.

Li Feng: Setting aside the AI chip area connected to you, looking three to four years ahead, what might Huawei become in chips broadly?

Ji Yu: Following its current development path, Huawei will certainly work hard to gradually resolve various supply chain issues. This serves both Huawei's interests and China's semiconductor industry interests, so it will keep pushing this forward.

That's also a good thing for the domestic chip industry as a whole, and ultimately benefits startups like ours.

From the perspective of building a chip startup in China, we certainly hope our products can create more favorable conditions for the chip supply chain through areas where we excel — computer form factors, software ecosystems, and the like. Conversely, if Huawei can drive major advances in the supply chain capabilities of China's chip industry, that gives us greater safety margins as well. These two things can reinforce each other.

/ 11 / "Straight-line pursuit" or "corner overtaking"?

Li Feng: Broadly speaking, when it comes to chip-related endeavors, there are usually two approaches to catching up. One is called straight-line pursuit — "doing NVIDIA the NVIDIA way." For instance, Huawei's Ascend follows this straight-line pursuit path. It requires enormous resources, capital, and capabilities. The other approach we're more familiar with is "corner overtaking" — choosing a different method to achieve similar long-term results.

In the AI chip race, FreeS Fund, as an early-stage investor, has so far invested in chips that are somewhat more "heterogeneous" — not using NVIDIA's approach to AI chip architecture. This includes, for example, photonic chips (using photons instead of electrons to perform AI mathematical operations, controlling photons for linear computation), as well as using RISC-V on open-source architectures, and so on. From your perspective, what opportunities and challenges do all these chips that differ from pure GPUs present?

Ji Yu: First, the compute chip industry is inherently a monopoly, so straight-line pursuit is extremely difficult. This difficulty isn't just technical — it's more commercial, because the business structure itself is monopolistic.

One thing you must be clear about when building compute chips is that you either become a monopoly, or you die. There are basically only these two possibilities. It's unlikely you can comfortably settle for being number two or three in the industry.

Second, if you hope to build a monopolistic category, how do you construct that monopoly? It's not purely a technical problem — it involves complete business logic. If you're simply developing a new technology but can't break through the horizontally layered business structure we discussed earlier, can't wedge yourself in, then it's very hard to build a highly monopolistic category.

/ 12 / Seeing through surface uncertainties to find underlying certainties

Li Feng: The next question is also investment-related. I recall that about half a year ago, when you were just preparing to start your company, the industry was still in a relatively cold phase. People had considerable doubts about these AI chip startups with different architectures. But in the past two months, this direction seems to have started heating up. Does this impression match what you're sensing?

Ji Yu: First, industries certainly have cycles, and investment is also something that fluctuates rapidly.

From our own perspective, when there are more opportunities to secure more resources, we'll definitely work hard to get them. However, the chip development cycle is very long, especially when you're trying to change the underlying logic of the computer industry — the cycle becomes even longer. It's hard to deliver results in one or two years; you can only change things bit by bit.

Entrepreneurship isn't just about technology and products — it inevitably involves fundraising. So whether the industry heats up or cools down, everything has two sides.

When it's hot, you might be able to get more resources. But when it's cold, there's actually less noise in the industry overall.

So I tend to take a relatively conservative view of everything. Too conservative isn't good either — that only leads to despair. Being relatively conservative lets you see both sides of things, and figure out how to seize opportunities and continuously amplify the side that works in our favor.

Many people look at NVIDIA's dominance today and feel despair, thinking it's strong in every aspect and utterly unbeatable. It is indeed very strong, but the flip side of that strength is having to attend to everything, which gives us opportunities to break through. This duality can be found in anything, and it's one way we find certainty amid all uncertainty.

Li Feng: From deciding to start your company until now, have you noticed changes in yourself?

Ji Yu: A trait I've always had is probably relatively deep industry insight, and secondly, the ability to see through surface uncertainties to find underlying certainties. For entrepreneurship, this is quite important — that is, facing much uncertainty, how do you keep yourself sufficiently stable internally.

Frankly, my decision to start a company in 2023 was quite impulsive. Many things weren't fully prepared, but I had thought through the most critical thing: the business thinking.

As for fundraising, team building, and other aspects, I wasn't fully prepared when I started, but these are things you can learn and accumulate through exploration going forward. Only the essence of business — no one can teach you that; you have to continuously develop that insight yourself.

Li Feng: I very much agree with that. From an investor's perspective, from 2014 to now over this decade, the AI field has seen at least four interesting waves of enthusiasm — autonomous driving, data middle platforms, AI for science (such as AI plus biopharma, AI plus materials), and AI plus large models.

A technology goes through many bubbles and downturns, a process accompanied by technological progress and application iteration. Typically, technological progress happens in the early bubble phase, while application iteration occurs as the technology bubble gradually subsides. So technology and application promotion go through many cycles of bubble formation and dissipation.

If we focus on large models, this is the first wave of enthusiasm for large models, but the fourth wave for AI. Autonomous driving alone went through three waves. I believe large models will see several more waves of enthusiasm in the future, and more interesting industry changes will emerge.

The AI we're discussing here includes the data layer, the algorithm layer, the chip layer, and the underlying energy layer. Each layer will likely generate new opportunities through such changes and the rise and fall of bubbles. We also hope that AI chip-related startups can, as Ji Yu described, continuously extract the most important long-term certainties from uncertainty, and achieve success.

---

▲ The Private Cloud Era Arrives: How AI NAS Reshapes Your Digital Life | FreeS Research

▲ Congratulations to XtalPi on Successful IPO: FreeS Fund's Research-Driven Approach, Learning to Invest Through Investing

▲ The Road to Embodied Intelligence | FreeS Report 37

▲ Toward 2024: How We Think About AI Startup Investment

▲ Li Feng in Conversation with Wenzhao Lian: Imagination and Bubbles of Large Models, the "Impossible Triangle" of Robots and the Future | FreeS VC Dialogue

▲ Li Feng in Conversation with Zhang Wei, Founder of LimX Dynamics: Humanoid? Robot? | FreeS VC Dialogue

▲ From Lithography to Nanoimprint: Challenges and Opportunities at the Semiconductor Industry's Turning Point | FreeS VC Dialogue

▲ U.S. Escalates Chip Controls, What Changes Will Occur in China's Semiconductor Industry? | FreeS Research · Chip Series

Star the FreeS Fund WeChat Official Account for timely business insights delivered to your fingertips