DeepSeek Fired the First Shot, but AI Democratization Is What Really Matters | A Conversation with Yu Ji of Xingyun Integrated Circuit

About ten days ago, we published Seven Core Questions About DeepSeek, Explained | FreeS Report, analyzing the reasons behind DeepSeek's explosive popularity and its multi-dimensional impact.

At the hardware ecosystem level, the tremors DeepSeek set off are still reverberating. NVIDIA keeps emphasizing that "DeepSeek's emergence proves we need more AI chips." For other large model vendors, the "off-peak discount program" DeepSeek launched on February 26 may trigger a new round of price cuts. At the application level, institutions like Morgan Stanley and Goldman Sachs believe that the proliferation of low-cost, high-performance models will bring explosive usage and broader commercial deployment to the AI industry.

So what does DeepSeek actually mean for entrepreneurs in the AI industry? Not long ago, Gang Li, Vice President at FreeS Fund, and Dr. Yu Ji, founder and CEO of Xingyun, had an in-depth discussion about the impact of DeepSeek and how to pursue technological innovation.

Dr. Ji holds a bachelor's degree in physics from Tsinghua University and a Ph.D. in computer architecture from Tsinghua's Department of Computer Science. He was selected for Huawei's "Genius Youth" program and worked at HiSilicon on AI chip compiler design and optimization, tackling complex technical challenges. In 2023, Ji founded Xingyun to develop ultra-high-memory GPU chips tailored for large model demands. Xingyun aims to drive AI technology democratization and bring the market back to a golden age where people cost more than machines.

We've edited portions of their conversation into this article. Topics they covered include, but are not limited to:

• Most easily accessible data has already been used for AI training. Will AI eventually generate its own data to train itself?
• Where does the moat lie for large model companies?
• Is DeepSeek's emergence an accidental breakthrough or an inevitable direction for industry development?
• What opportunities exist for restructuring and innovation in the AI chip industry?
• How should we understand Wenfeng Liang's statement that "technical advantages are fleeting; the real moat is culture and organization"?

We hope this offers fresh perspectives and food for thought. We've only excerpted portions of their conversation — welcome to search for and subscribe to "High Energy" on the Xiaoyuzhou app or Apple Podcasts to listen to the full episode.

Engagement Giveaway

How has DeepSeek changed your life or work? Share with us in the comments! By 5:00 PM on March 6, we'll randomly select three readers to receive the updated industry research handbook written by the FreeS Fund team.

/ 01 / Will AI Generate Its Own Training Data in the Future?

Gang Li: Most easily accessible data on the internet has probably already been used to train AI. As available data for large models dwindles, will we eventually see AI generating its own data to train itself?

Yu Ji: I think the key question isn't where data comes from — what's more important than data is language itself. Because language is the underlying foundation that gives intelligent agents their capacity for thought.

Large models certainly need data during the pre-training phase, but during inference, what matters more is reward.

Reward is a generalized supervision signal that tells the model which direction to adjust in order to improve performance. In reinforcement learning, for example, the model randomly generates large numbers of outputs, receives external signals that clarify what's good and what's bad, then adjusts to produce better results. As long as this kind of signal input exists, optimization can proceed.

Previously, the self-play method was fully applied when the AI program AlphaGo trained its Go skills. Self-play is a strategy in AI training that simultaneously employs a generator and a discriminator to improve output quality — like a person playing chess against themselves to steadily improve.

But in practice, self-play hasn't been adequately defined within the conceptual space of reinforcement learning or large models. So there's still vast room for exploration in how to enhance large models' reasoning capabilities.

Thinking is a process of self-verification and iteration. The power of natural language lies in our ability to ponder problems alone — even without conversation, we can perform logical deduction in our minds, find approaches, or reach deeper conclusions. This comes from language itself, and represents an even larger space for exploration.

Of course, "language" and "reward" aren't mutually exclusive. If we can leverage language to discover more reward signals, that can also improve model capabilities. For instance, according to industry analysis, one of ChatGPT's breakthroughs was training code and natural language in the same model, which improved output quality.

If large models can engage in reinforcement learning across more domains today, the differing data characteristics across domains might spark new capabilities.

Gang Li: Indeed. After all, AlphaGo's outputs are typically "black and white" because winning and losing are objective. But in the language model domain, some outputs are evaluable — math and programming have definite answers. More often, though, many questions have no standard answers. In the next phase, the data and evaluation methods that large models need may undergo new changes.

/ 02 / Where Is the Moat for AI Model Companies?

Gang Li: Under DeepSeek's open-source ecosystem, individual and enterprise users can relatively easily deploy or replicate models at low cost. So where does the moat lie for large model companies?

Yu Ji: For any company, technology itself doesn't form a moat. The real barrier lies in business model or ecosystem.

Competitive barriers built on resources are easily broken. DeepSeek shattered the resource barriers that many cloud providers and internet giants had constructed around GPUs and "10,000-card clusters."

A very practical issue is that despite high expectations for AI, it hasn't fully penetrated the world's economic system to deliver massive commercial value. Let's not set artificial thresholds for the AI industry — when the industry develops, viable business models will naturally emerge.

After all, before technology is truly mature, so-called barriers may prove fragile. Only when technology and application scenarios are tightly integrated to form a sustainable economic cycle does a company possess a genuine moat.

/ 03 / Is DeepSeek's Emergence an Accidental Breakthrough or an Inevitable Industry Direction?

Gang Li: You previously mentioned that two factors underpin NVIDIA's stock price: the development of the large model industry and NVIDIA's monopoly position within it. (Read: Li Feng in Conversation with Yu Ji: Understanding NVIDIA, Deconstructing NVIDIA, Challenging NVIDIA)

▲ Scan the QR code to listen to Feng Shu and Yu Ji discuss innovation opportunities in the AI chip space.

During DeepSeek's viral moment, some in the U.S. even proposed tightening chip and GPU export controls on China to restrict Chinese AI large model development. Their first reaction to challenge was to circle the wagons. Do you think DeepSeek is an accidental disruptor, or is this an inevitable direction for industry development?

Yu Ji: Within such a massive system and ecosystem as AI, DeepSeek is a very important player. Its breakout into mainstream awareness has significantly advanced the AI industry.

Previously, OpenAI, Microsoft, NVIDIA, and others formed an interest alliance, wanting everyone to use their infrastructure and models. Scaling Law drove the formation of thousand-card and ten-thousand-card clusters, essentially building a large-scale computing system. The result: hardware increasingly resembles mainframes from the 1980s, requiring massive investment to run the best models.

DeepSeek was the first step in prying open this system — "firing the first shot." More importantly going forward is how AI-democratized industrial structures will contend against the small circle of monopolies built on resources.

The further AI develops, the more innovation it needs. Innovation isn't limited to the model layer — it includes infrastructure (Infra) transformation, chip restructuring, and the complete system built around AI democratization. And chip restructuring is precisely the direction that Xingyun has been committed to.

For the hardware industry, the real path to breakthrough lies in shaping a new system more attractive and competitive than the CUDA ecosystem.

If going forward we can build something around DeepSeek, domestic chips, and the goal of AI democratization — similar to the ecosystem of X86 clusters and personal computers of the past — we may gradually erode the core value of this "large computing system."

We might be able to use chips costing thousands or tens of thousands of yuan to access good large models, even building clusters with service capabilities exceeding previous expensive hardware.

Historically, most revolutions in PC and internet domains followed this pattern: innovative companies replaced traditional giants' high-price strategies with more economical approaches. This transformation not only reduced hardware costs but also formed more efficient computer foundation systems.

/ 04 / The Innovation Process Itself Needs Efficiency Gains

Gang Li: As an entrepreneur, how do you view Wenfeng Liang's statement that "technical advantages are fleeting; the real moat is culture and organization"?

Yu Ji: I strongly agree. Wenfeng Liang has led by example — he's an excellent role model.

Technical advantages matter, but they're hard to sustain. When OpenAI first emerged, no one could catch up to its technology, but that was only temporary. With so many outstanding talents worldwide, it's highly probable that this technology can be replicated and surpassed.

DeepSeek's key to success lies in its joint innovation across infrastructure, algorithms, and models — not breakthroughs at single points.

The innovation process itself also needs efficiency gains. This kind of joint innovation places extremely high demands on organizational capability.

Organizational capability isn't about gathering smart people and expecting value to emerge automatically. The role of organization is to set a coherent, sensible direction so everyone advances together.

The prerequisite for proposing such an integrated approach is going deep into the details, knowing what needs to happen at each layer, to effectively break through boundary conditions at every level.

I previously spoke with a junior colleague working at DeepSeek who said Wenfeng Liang knew every technical detail of his project extremely well. When the overall leader is deeply involved in projects and understands the technical details, they can better organize the big picture and truly let the talent in the team create value.

Culture and organization are the real moat, but putting this into practice is extraordinarily difficult.

I've observed that some people, both in investing and entrepreneurship, heavily believe in empiricism — trusting industry experts and benchmark figures who have achieved success. This empiricist approach did play a role in China's economic rise; it solved the problem of how to improve efficiency. Anti-empiricism isn't about abandoning experience — it's about building new organizational methods to efficiently drive innovation.

China isn't short of anti-empiricist innovation attempts — non-Transformer architectures, non-GPU chip routes, and so on. But part of why these attempts haven't had high success rates is the lack of holistic thinking and methodology.

Gang Li: I also strongly agree with Wenfeng Liang's view. From an investor's perspective, we want startups to improve capital efficiency and create greater value. And the efficiency question is fundamentally about talent and team organization — how to find the best people and combine them effectively.

Just like each expert in an MoE (Mixture of Experts) architecture, a company needs to think through: how do you organize every expert in your team to form a company that is both large and capable?

Engagement Giveaway How has DeepSeek changed your life or work? Share with us in the comments! By 5:00 PM on March 6, we'll randomly select three readers to receive the updated industry research handbook written by the FreeS Fund team.

▲ The Right Timing, Right Place, and Right People Behind Ne Zha 2's Blockbuster Success | FreeS Research

▲ Seven Core Questions About DeepSeek, Explained | FreeS Report

▲ 2025: Three Directions Where AI Is Reshaping Drug R&D

▲ Outlook 2025: What Innovation Opportunities Exist in AI? | FreeS Report

▲ AI for Science: At the Inflection Point of a New Research Paradigm | FreeS Report

Star the FreeS Fund WeChat official account for timely business insights