So, what exactly is this wave of revolution?
The AI revolution is not only transforming production relations but also driving a transformation in productivity.
- The Computing Paradigm Revolution: Today’s computing system has shifted from being CPU-centric to GPU-centric. Once the computing unit changes, everything else—storage, networking, architectural systems, scheduling systems, operating systems, and upper-layer applications—changes too. This isn’t just a breakthrough in a single technology; it’s a shift in an entire technological system. Ten years ago, Nvidia’s market value was one-sixth of Intel’s; today, Nvidia’s market value is 30 times that of Intel.
- The Revolution of Cognitive Collaboration: What’s the essence of the internet and mobile internet? It’s a business model built on precise information matching enabled by connectivity. Companies like Didi, Google, Toutiao, and Taobao are all focused on one thing: accurately matching supply-and-demand information. But for AI, the core isn’t about precise information matching—it’s about finding the shortest path to answers for problems. If we want to place today’s large-scale AI technology on a historical timeline, we could compare it to the internet or mobile internet, but I don’t think that’s entirely accurate. A more precise historical parallel might be the integrated circuit. Today’s AI is more akin to the integrated circuit of its time—it’s not just transforming production relations; it’s driving a revolution in productivity.
- The Revolution in Human-Machine Interaction: Fifty years ago, you had to know assembly language to interact with machines. Later came so-called high-level languages like C and Basic. Today, languages have evolved into natural ones like English and French, generating software code to optimize, control, and influence the physical world.
So how do we understand this wave of change? It’s actually bringing about three revolutions.
Everything Can Be Tokenized
What is the most important technical foundation of today’s large-scale AI models? It’s the vector representation of textual semantics—a leap from processing the form of information to processing its content.
Let’s take a step back to a few decades ago. That generation of internet pioneers grew up with a book called Being Digital. In fact, it talked about just one thing, introducing a single concept: what does “digital” mean, and what is “digitalization”? (This is a textbook example.) When we open the first page of any computer science book today, what’s the core idea it always explains? It’s about turning human-recognized information into a series of 0s and 1s. In other words, digitalization is the process of converting atoms into bits. We call this process Digitalization.
Now, let’s ask a question: if Digitalization is about transforming atoms into bits, then what’s the key term for the intelligent era of artificial intelligence we’re talking about today? Let’s look back again. Fifty years ago, communication billing was based on time—one minute of a long-distance call cost one yuan and ten cents. Then, with 3G and 4G, the unit of measurement for communication became bits. Today, when you use Tongyi Qianwen or ChatGPT, the unit of measurement is called a Token. So, here’s the question: what is a Token, and why do we use Tokens? This is a very important perspective I’d like to discuss with you today: “Everything Can Be Tokenized.”
If the keyword of the digital era was Digitalization, then what’s the keyword for the intelligent AI era today? It’s Tokenization.
Jensen Huang, CEO of Nvidia, recently said something striking: we are experiencing an unprecedented industrial revolution, and at its core is the mass production of something entirely new for the first time. That something is the Token. It can be recombined and transformed into language, proteins, chemicals, graphics, images, videos, and even robots.
At the Yunqi Conference, Wang Jian, Chairman and CEO of Alibaba Cloud, said that AI has given the world a unified language: the Token. It can represent any text, code, image, video, or sound.
Then there’s an incredibly significant statement from Zhang Bo, a towering figure in artificial intelligence at Tsinghua University. What’s the most important technical foundation of today’s large-scale AI models? It’s the vector representation of textual semantics—a leap from processing the form of information to processing its content. When you translate Chinese into French or English, what does “form” mean? It’s analyzing the subject, predicate, and object. But today’s approach isn’t like that anymore. AI has turned it into a matter of content, essentially transforming a language problem into a mathematical one.
So, what can be tokenized today? We already know that text, images, sounds, and videos can be tokenized. After tokenization, they can be trained—using Transformer models—or used as input prompts to generate answers. But what I want to say today goes beyond that. More importantly, what else can be tokenized? Anything that contains knowledge can become a Token. It turns a molecular problem or a language problem into a mathematical one. So, the core shift from digitalization to intelligence is a move from bits—from digitalization—to mathematization, breaking everything down into mathematical problems. That’s why we now have autonomous driving, like Tesla’s FSD version 12, and protein prediction tools like AlphaFold. Their common trait with today’s AI is the tokenization of knowledge-containing content.
Whether in the U.S. or China, Only 3-5 Foundational Large-Scale Models Will Remain
Whether it’s the United States or China, the number of foundational large-scale models that ultimately survive will be just 3 to 5. In China, there’s no “battle of a hundred models,” nor even a “battle of ten models.” The U.S. doesn’t have this either. Today, the U.S. foundational model landscape follows a “three-plus-two” pattern (OpenAI, Anthropic, xAI, Meta). In China, the number of surviving foundational large-scale models is also likely to settle at 3 to 5. Among open-source models, the most important global ranking is the capability leaderboard on America’s Hugging Face, where, over the past six months, Alibaba’s Tongyi Qianwen open-source model has held a leading position.
Three Key Judgments on AI Development
Looking at the present from the perspective of the next five to ten years, I believe there are three critical judgments: The first is that all intelligent hardware will be driven by large-scale AI models. The second is that all software will be restructured by large-scale AI models. The third is that all data will be activated by large-scale AI models.
What does it mean that all hardware will be driven by large-scale AI models? This includes phones, PCs, TVs in homes, various robots, cars, AGVs (automated guided vehicles), humanoid robots, and robots in traditional industrial settings. This is a competition to reclaim control over traffic distribution. In the past, it was about traffic distribution itself, but now there’s a prerequisite: intent understanding. Intent understanding could lead to a redistribution of traffic allocation rights. Intelligent hardware is a highly certain and competitive industry in China, deeply integrated with countless sectors.
What does it mean that all software will be restructured by large-scale AI models? This includes traditional software. In the past year, companies like Salesforce and Adobe have seen their valuations rise by 50%, or even double or triple. The head of SAP’s China Research Institute said that with the arrival of AI, the most complex ERP software has reached a tipping point from quantitative to qualitative change. Fifty percent of its processes should be restructured by AI, creating new methods of human-machine interaction, business workflows, and user experiences. In this process, it will also give rise to many new enterprises.
We see a significant track in the U.S.: code generation. In this niche field alone, the U.S. has at least six unicorn companies with valuations between $1 billion and $3 billion. A famous American IT influencer, Beck, made a key observation: 90% of software engineers’ skills have become worthless, but the remaining 10% need to improve by a thousandfold to meet the demands of AI-driven technological change. The commercial value in this track is already a certainty. In China, the leader in code generation—boasting the highest market share and fastest progress—is Alibaba’s Tongyi Lingma. It has evolved from a Copilot to an Autopilot.
How is AI-native development progressing in the U.S.? One institution conducted a study. First, they found that for the top 100 SaaS companies in 2018, the median company took 34 months to reach $5 million in revenue. Today, for AI companies founded in 2024, it takes just 24 months—a 10-month reduction. Second, comparing AI companies founded before and after 2020, the median revenue of the top 100 companies pre-2020 took 39 months to hit $5 million, while those founded post-2020 took only 26 months. This shows that in the development of AI-native enterprises, the U.S. has surpassed the previous SaaS industry wave. Over the past decade, China’s SaaS companies may have experienced a “lost decade,” already lagging far behind. In this round, the U.S.—already dominant in the SaaS sector—has taken another step forward, opening up new innovation space in an unstoppable manner.
The Relationship Between Large-Scale Models and Data Elements
Large-scale models are the shortest path to creating value from data elements. They will transform the way data elements are innovated and utilized. From the technical perspective of large-scale models, there are roughly four approaches: the first is prompt engineering, the second is RAG (Retrieval-Augmented Generation), the third is fine-tuning, the fourth is pre-training, and the fifth is Agents.
What Are the Challenges of AI Entrepreneurship?
Can We Overcome the Galápagos Effect?
What is the biggest challenge facing entrepreneurship today?
There are challenges on both the supply side and the demand side.
Computing power, models, and data all belong to the supply side.
But we also face issues on the demand side. In my view, the structure of market demand in China today might be another significant problem we’re confronting. We often say that China has the advantage of a large country with a vast market, but in the to-B (business-to-business) digital market, that “large country, vast market” advantage comes with a question mark. It might not be an advantage at all—perhaps it’s more of a “large country, small market” disadvantage. So today, China’s large-scale model industry stands at a crossroads. If we turn left, we might catch up with the mobile internet era. If we turn right, we could end up in the dead-end alley of SaaS.
I often refer to this phenomenon as “how to overcome the Galápagos Effect.” On that island, species evolve, iterate, and develop in a closed environment, but once they leave that system, they have no competitiveness. In the context of industrial economics, take Japan’s mobile phones and software as an example—they favor customization over standardized products. Similarly, China’s SaaS industry, industrial internet industry, and the so-called “four little AI dragons” from previous waves exhibit this pattern.
The question we need to ponder is this: as China’s AI large-scale model “vehicle” drives forward, can we break free from the Galápagos Effect? A critical observation is that the demand structure on the to-B side determines the form of supply. Are we facing a fragmented market, or a unified national market? All our companies have lofty labels on their foreheads—we call them SaaS companies, industrial internet companies, or AI large-scale model companies—but in reality, their delivery models are stuck in a primitive state. These models come in three forms: project-based, productized, and platformized. The question we need to consider is this: today, American companies deliver their products and forms efficiently and at low cost through platformization. Will our market structure drag our entire industry back to a primitive, project-based state? If the delivery model for China’s AI large-scale models remains project-based, this industry has no future. Thus, overcoming the Galápagos Effect may be one of the challenges facing entrepreneurs today.
