AI Industry Evolution Enters the 'Differentiation Zone' as Dominant Design Falls Behind

This article is approximately 18,000 words and follows the core structure below:

• Intuitive or abstract judgments about AI development may be difficult to effectively guide decision-making

• Two classic industry evolution analysis tools (Utterback + Klepper)

• Analyze using the consolidation framework: What stage is the current AI industry in?

• A new question that must be asked: How will the pace unfold in the future?

• The timeline of AI industry evolution: from the manufacturing era to the digital era

• Specific timeline for the AI transformation (projection)

• Verify against the latest trends in reverse

Open your phone, and updates about AI flood in almost daily: a new model has just been released, outperforming its predecessors again; a company has secured new funding, pushing its valuation to new highs; a platform has broken new user records; on the latest rankings, yesterday’s leader has already been overtaken. The density and speed of information far exceed any previous technological revolution.

This feeling continued at the WAIC conference, which just concluded in July. The venue was filled with constantly updating data, ever-improving performance, and newly refined technical specifications and product forms; an increasing number of diverse organizations—from cutting-edge labs and major corporations to startups, traditional industry players, end-device manufacturers, and research institutions—are actively joining in, becoming part of this vast technological and industrial revolution.

With such dense information and so many participants, two common reactions naturally arise.

One is anxiety—Am I falling behind? Is my company missing out on something? If I don’t act this week, will it be too late next week?

The other is an impulsive judgment—who has already won? Who has already been eliminated? Has the industry’s landscape more or less been settled?

Underlying both reactions is the same question: What stage is the AI industry currently in, and how can we make a rational assessment of it?

There are also popular claims today—such as “AI has entered a new phase.” From various perspectives, these judgments have their own logic and basis. However, for corporate decision-makers who need to make rigorous strategic assessments and decisions, more fundamental questions must first be clearly answered: On what grounds do we determine that the previous phase has ended? How can we be certain that AI has entered a new phase? What are the specific characteristics? What underlying patterns exist? And how many distinct stages make up the complete cycle? Answering these questions requires a more rigorous and scientific analytical framework to arrive at more reliable conclusions—an issue that is easily overlooked in today’s AI transformation but could carry significant consequences. Misjudging the industry’s current phase directly affects the direction, timing, and prioritization of resource allocation; if the phase is misunderstood, all actions may veer off course.

In current AI discussions, several scholars have begun analyzing and assessing the stage of the industry. Among these, Harvard Business School professor Karim Lakhani’s lengthy article, “From Ferment to Dominant Design,” published in March 2026 on his personal Substack, represents a notable effort. He applies analytical frameworks developed over decades by two highly respected scholars in the field of industrial innovation—James Utterback (MIT) and Steven Klepper (Carnegie Mellon University)—who built their tools using comprehensive historical data from multiple industries. Lakhani uses these frameworks to assess the stage of competition in large AI models and presents several original observations (which will be elaborated later). His central conclusion is that AI is currently in the late fluid stage, with no dominant design yet established; the industry’s structure should be viewed as an open question rather than a settled outcome. This aligns with the direction of this paper and is one of the key sources of insight that informed the development of this framework.

This article will fully elaborate on the classic frameworks of two scholars, explore an extended analytical framework for the digital/AI era, and integrate the latest industry data and developments from 2026. It will first rigorously assess the current stage of AI industry evolution, then address a question left unanswered by Professor Lakhani but critical for business decision-makers: According to general patterns of industry evolution, what will be the pace toward the establishment of a dominant design and the emergence of deep market reshuffling?

This is what the article refers to as "the timescale of AI industry evolution"—where we are now and how the pace will unfold in the future.

These two classic tools are the most influential public academic assets in the field of industrial innovation research. Their value lies not in abstract theories, but in patterns derived from extensive historical data and validated through repeated empirical testing.

Utterback’s Three Stages of Industry Evolution: From the Perspective of the Product Dimension

James Utterback is a professor at the MIT Sloan School of Management and one of the pioneering scholars in the field of industrial innovation and technological evolution. In his seminal 1975 study co-authored with William Abernathy, he examined the historical evolution of technology across dozens of manufacturing industries, and in his 1994 comprehensive work, Mastering the Dynamics of Innovation, he systematized a remarkably consistent pattern: the technological evolution of nearly all industries follows a recognizable three-stage sequence.

Fluid Phase — Everyone is experimenting. There are no clear answers yet on what the product should look like, what customers truly want, or what “good enough” means. Multiple designs coexist and compete, with low entry barriers and innovation focused on the product layer. The core question is: What should the product be?

Transitional Phase — Certain design choices begin to stabilize, and a dominant design emerges. The focus of innovation shifts from "what the product should be" to "how to make the product cheaper, more reliable, and more scalable"—from product innovation to process innovation. The central question becomes: What do customers value most?

Specific Phase — Product design is largely finalized, and competition shifts to efficiency—who can deliver more cheaply, reliably, and at scale. Innovation becomes incremental, and the industry becomes highly concentrated. The core question becomes: Who can deliver more reliably, more cheaply, and at greater scale?

The key insight of this model is that the qualities of the winner differ entirely across the three stages. The capabilities that define success in the fluid stage—experimentation, speed, imagination, and tolerance for failure—are fundamentally different from those that define success in the specialized stage—scale, cost efficiency, discipline, and reliability. A company leading in the fluid stage does not necessarily have the capacity to continue leading in the specialized stage—this is the underlying reason why many industry disruptors, once at the forefront, ultimately fall behind.

One additional point is that Utterback, later with Suárez (1993), further noted that industry evolution does not always progress through all three stages—some industries ultimately settle on a dominant design (Path A), while others may remain in a state of multiple competing designs (Path B) for extended periods.

2.1.1 Key Concepts: What Is a "Dominant Design," and Will It Still Exist in the AI Era?

Among the three stages, dominant design is the most central concept and warrants further explanation.

Dominant design is not about "one company's product winning out," but about "a specific combination of technologies and product form becoming the industry's de facto standard." Once established, newcomers seeking to compete must build upon this design—products that deviate from it are almost inevitably marginalized.

The classic example is the 1908 Ford Model T—it didn’t establish "Ford as the winning brand," but rather cemented the combination of "gasoline internal combustion engine + front-mounted engine + rear-wheel drive + assembly line production" as the industry’s default standard for decades to come. Afterward, automakers competed within this dominant design framework to see who could do it better, rather than challenging the fundamental question of what a car should look like.

Once the dominant design is established, the rules of the industry subtly shift: the focus of innovation moves from "what the product should be" to "how to make the product cheaper, more reliable, and more scalable"—from product innovation to process innovation. This is the most critical dividing line between the transitional and specialized phases.

A natural follow-up question: Will there still be dominant design in the AI era?

I believe the answer is yes, but the form will differ from that of the manufacturing era. As long as industries require scalable delivery, ecosystem collaboration, and stable user expectations, some form of "de facto standard" will inevitably emerge—this is an economic requirement of industrial operation. Although the pace has accelerated in the internet era, dominant designs such as TCP/IP, HTTP, HTML, and the Chromium kernel have still emerged layer by layer.

However, the dominant design in the AI era is more likely to be "layered, multiple, and composable"—no longer will a single product form dominate; instead, de facto standards will emerge at different layers of the technology stack and be connected through interfaces and protocols.

In other words, the dominant design of the AI era may not be "the next Model T," but rather something closer to a layered protocol stack like "TCP/IP + HTTP + HTML"—where each layer remains stable and interacts with others through standardized interfaces. Early forms of such a next-generation dominant design are already emerging, including MCP (Model Context Protocol), A2A (Agent-to-Agent) networks, and de facto API calling conventions. Dominant designs have not disappeared; they have simply changed their form.

2.1.2 Historical Data Validation

The reliability of the Utterback three-stage model stems from its accurate characterization of multiple dominant design establishment processes:

• Automotive industry: After the Ford Model T established the dominant design in 1908, the automotive industry rapidly transitioned from the fluid stage to the transitional stage.

• Television industry: Early black-and-white TVs coexisted with various color systems; after the establishment of the NTSC color TV standard in 1953, the industry entered a transition period.

• Personal computers: In the late 1970s, various architectures emerged, such as the Apple II, TRS-80, and Commodore, until the IBM PC architecture combined with the Wintel alliance established the dominant design in 1981.

• Smartphones: Before the release of the iPhone in 2007, the market featured various designs such as BlackBerry, Nokia Symbian, and Windows Mobile. After the iPhone with iOS/Android became the dominant design, other approaches quickly faded from the mainstream.

Common characteristic: All involve an initial phase of coexisting multiple pathways, followed by a specific combination becoming the dominant design. The consistency of this pattern forms the empirical basis of the Utterback model.

2.2 Klepper's Industrial Population Dynamics: From the "Firm-Level" Perspective

Steven Klepper is a professor of economics at Carnegie Mellon University and a leading scholar in industrial organization and the evolution of firms. As early as 1982, Klepper, in collaboration with Michael Gort, developed the first formal mathematical model of the industry life cycle from the perspective of industrial economics—commonly known in academia as the G-K model. He subsequently published a series of independent studies further deepening this research. Based on systematic analysis of decades of historical data across multiple industries, Klepper discovered a consistent pattern in the entry and exit of firms during the process of industry development:

Enter upward trend → Number of companies peaks → Major shakeout → Few leading companies dominate.

The empirical basis for this pattern is his complete tracking of the origin, location, and year of production for every firm that entered the U.S. automobile industry from 1895 to 1966. The data are as follows:

Here is an astonishing paradox:

It took about 50 to 60 years from the peak to deep integration in the 1960s, but during that time, annual U.S. automobile production increased by hundreds of times on a logarithmic scale. In other words, the industry became immensely large, while the number of companies became extremely small. The industry did not struggle—it thrived; the weak perished, not the demand.

Klepper also verified this same pattern across multiple industries, demonstrating that the "rise-peaks-fall" shape is not unique to the automobile industry—industries such as televisions (which had over 80 participants in the early 1950s before undergoing a relatively rapid consolidation), typewriters, vacuum tubes, tires, and penicillin all exhibited the same fundamental pattern. This provided Klepper’s theory with a solid empirical foundation.

2.2.1 A Key Theory: The Theory of Composite Ability Differences

Klepper explained the above phenomenon with a theory called the Composite Ability Differences Theory:

Shuffling is not a sudden judgment by the market, but the inevitable result of compounded differences in corporate capabilities over time.

Differences among companies in their ability to innovate, grow, and reduce costs will be exponentially amplified over time until the weaker ones fall below the survival line. This theory has three key points:

First, capability differences often originate from early chance events. In the initial years after entering an industry, companies’ capabilities are largely similar; however, one or two different choices regarding key decisions, key talent, or key opportunities lead to the emergence of capability gaps. At first, these differences appear almost negligible.

Second, capability differences self-reinforce and amplify exponentially. Slightly more capable companies gain greater market share → access to more resources and learning opportunities → further widen their capability lead—this is a positive feedback loop. Less capable companies, by contrast, fall into a negative feedback spiral. Over the years, initially "negligible" differences compound into an "unbridgeable" gap.

Third, when discrepancies accumulate to a "critical threshold," the weaker participants are squeezed out. The seeds of elimination were planted early in the liquidity phase—on the surface, everyone is running, but in reality, their underlying compound rates are vastly different.

The above study reaches a sobering conclusion for corporate leaders: waiting until financial reports deteriorate to prioritize capability building is already too late; waiting for visible signs of industry disruption to initiate capability upgrades means the window of opportunity has closed. What truly determines which companies will survive the shakeout is whether they treated "capability compound rate" as their most critical strategic variable from the outset.

2.2.2 Another counterintuitive but important observation: the entry window closes earlier than the shuffle signal

In addition to the capability heterogeneity theory, Klepper made another counterintuitive but crucial observation: the "entry window" for an industry closes earlier than the "shakeout signals."

Returning to the automotive industry data: after the number of firms peaked between 1907 and 1910, the rate of new company entry dropped sharply after 1910—yet there were no clear signs of shakeout at the time; the industry was still growing rapidly, and the media had not yet begun reporting on "company exits." The clear shakeout—characterized by a sustained decline in the total number of firms over decades—did not begin until after 1920. In other words:

• Around 1910—externally, it still appeared to be a time of prosperity, with everyone involved in the automobile industry, but the window for new entrants had already begun to close (fewer and fewer genuine new players were able to enter).

• 1920–1960 — Signal of consolidation began to emerge, media started reporting on integration, and only then did ordinary observers realize, “Oh, the industry is consolidating.”

By the time observers realized it, the window had already closed for a full 10 to 15 years.

This observation has significant real-world implications: companies cannot wait for visible signs of disruption before deciding whether to enter an industry or a layer—because by the time those signals appear, the window of opportunity has already closed. Truly wise entry decisions must be made while the surface still appears prosperous and everyone is rushing in, with an independent assessment of whether the window is already shutting.

This observation will be revisited when discussing AI time coordinates—the window of entry for large AI models may close earlier than most observers expect.

2.3 Each tool has its own blind spots; using them together facilitates a more scientific assessment of the industry stage.

Each tool alone has significant blind spots—Utterback tells you how product design evolves, but not which companies will survive; Klepper tells you how companies are eliminated, but not how product development will change. Using only one tool reveals only half the picture of industry evolution.

The complete evolution of an industry encompasses two processes: "convergence of technological design" and "shuffling of the corporate population"—the former can be addressed by Utterback’s framework, the latter by Klepper’s. These two processes are related but not perfectly synchronized—the establishment of a dominant design typically occurs after the peak in the number of firms and before the acceleration of shakeout.

The automotive industry followed this pattern: the number of firms peaked at 275 between 1907 and 1910; Ford’s Model T established the dominant design in 1908; entry rates plummeted after 1910; and the total number of firms began to decline sharply after 1920. These two processes influenced each other and together defined the complete rhythm of the industry’s evolution.

Using both tools together means examining the industry from both the product dimension and the enterprise dimension—any judgment regarding the stage of industry evolution is more reliable and scientific only when observations from both dimensions corroborate each other.

3.0 Scope of Analysis

The "AI Revolution" is a vast concept. This article’s analysis focuses on the interconnected and co-evolving ecosystem of "large AI models and their applications and related products"—including the foundational large models themselves, applications and product forms built around them (such as conversational, programming, agent, co-pilot, multimodal, and embodied intelligence powered by models), as well as closely related supporting assets (cloud infrastructure, computing power, chips, toolchains, and evaluation systems). These three components are not separate industries, but rather three interdependent layers of a single, evolving industry.

Not covered in this analysis: components involving independent physical forms or distinct industry logic—such as the mechanical/hardware body of humanoid robots, or vertical industries like AI + healthcare, law, or education—each of which follows its own industry dynamics and cannot be directly mapped to the timeline presented in this paper.

3.1 From Utterback’s perspective: Each layer is still in the fluid phase

Viewing current large AI models and their applications through Utterback’s three-stage model reveals that nearly every significant aspect remains contested.

Architecture Layer: Will Transformer be the final outcome, or will one of the alternatives—Mixture of Experts (MoE), State Space Models (Mamba family), World Models, inference-time computation, or agent-native architectures—ultimately prevail? While leading companies are making large-scale investments in the MoE pathway by 2026, alternative directions such as World Models and agent-native architectures are also receiving significant advances from both academia and industry. Public discussions on diminishing marginal returns from scaling laws have prompted the industry to reconsider whether other fundamental possibilities exist. The architectural roadmap is far from converged.

Product Form Layer (the final product forms through which AI reaches users and enterprises)—currently, at least six competing forms coexist:

Six forms overlap and compete with each other, and none can claim victory—this is a typical characteristic of the fluid phase, where multiple product forms coexist.

Competitive Dimension Layer: Reasoning, Coding, Cost, Latency, Context Length, Multimodality, Agent Capability / Autonomous Task Duration, Reliability / Hallucination Rate, Security / Controllability—Which dimension will the market ultimately define as the "most important"? Each vendor claims leadership in its own strengths, yet the market has yet to reach consensus on which dimension matters most. The emergence of "how long an agent can operate independently" as a new dimension—from a fringe topic in 2025 to a core metric at the 2026 WAIC keynote in just one year—is itself a sign that competitive dimensions remain undefined.

From Utterback’s perspective, AI large models and their applications are clearly in a fluid phase—multiple pathways coexist, product concepts remain uncertain, and competitive dimensions have yet to be defined.

3.2 From Klepper’s Perspective: High Entry Phase, but Capability Differentiation Is Already Taking Shape

From Klepper’s industrial demography, the AI large model industry reveals a phenomenon of coexisting contradictions.

On one hand, global AI large model companies are still in a high-entry phase—OpenAI, Anthropic, Google, Meta, xAI, Mistral, along with China’s Baidu, Alibaba, and Tencent. DeepSeek Moonshot, Zhipu, ByteDance, and countless startups focused on specific use cases. New entrants continue to emerge, characterized by intense experimentation.

On the other hand, capability asymmetries among leading companies are quietly widening—computing power (who can afford the cost of ultra-large-scale training), distribution (who controls the entry points), enterprise customers (who have built industry trust), and organizational learning speed (who faster translates each wave of technological advancement into products and business capabilities). This is the ongoing reality of "compounding capability gaps"—the gap observed today between leaders and others is largely the result of compounding over the past three to four years.

From Klepper’s perspective, the AI large model industry is in a phase of “high entry, but capability differentiation has already begun”—similar to the automotive industry between 1900 and 1910, when the number of companies was still rising toward its peak, but some players had already quietly pulled ahead.

3.3 Rigorous Validation of Leading Design: Five Contributing Factors

Previously, based on Utterback and Klepper, a conclusion was drawn: AI large models and their applications are in the late fluid phase. This conclusion warrants further verifiable validation—by synthesizing the research and arguments of scholars such as Utterback, Suárez, and Christensen, the author has identified five key factors that enable a dominant design. Historically, the establishment of a dominant design typically requires at least three or four of these factors to be present simultaneously:

Factor One: Technical Feasibility—A particular design demonstrates clear convergence in terms of overall performance and engineering viability, leading the industry to collectively gravitate toward it as the widely accepted solution in the technical pathway.

Factor Two: Maturity of Complementary Assets—A supporting ecosystem of supply chains, standards, toolchains, training systems, and developer communities has been established around this design, creating ecosystem lock-in.

Factor three: Economies of scale emerge—this design’s scalable delivery creates a positive flywheel of “lower costs, positive unit economics, and a validated business model,” initiating a cycle of investment, returns, and reinvestment.

Factor Four: Driving Force from Leading Manufacturers—A few market-influential companies adopt and promote this design, collectively establishing it as a de facto standard through their market share and ecosystem influence.

Factor five: Government or regulatory intervention—mandating or guiding uniformity through standardized legislation, certification systems, and regulatory frameworks

The Ford Model T became the dominant car design in 1908 because five factors aligned almost simultaneously: technical feasibility (the optimal combination of internal combustion engine and front-mounted engine), complementary assets (rapid development of gas stations and repair networks), economies of scale (assembly line reduced costs), a dominant manufacturer (Ford once held over half the market share), and government regulation (road standards and driver’s license systems).

Applying these five factors to evaluate the current state of AI large models and their applications, only one criterion is met, while the other four are still far from being realized—this is the specific basis for this article’s judgment that "AI large models and their applications remain in the late fluid phase": not based on intuition, but derived through systematic, point-by-point verification.

In other words, AI is still far from meeting the threshold established by the dominant design, which requires at least three to four factors to be present simultaneously.

3.4 Comprehensive Phase Judgment: The "Divergence Zone" Moving from the Late Liquidity Phase to the Early Transition Phase

Based on the above analysis and the five-factor assessment, a comprehensive judgment can be made: AI large models and their applications are currently in the late liquidity phase and have begun transitioning toward the early transition phase—an area the author refers to as the "divergence zone."

The core meaning of "differentiation zone" can be summarized as: product layers remain open, while capability differentiation has begun.

The "divergence zone" is not an independent new stage; it describes the transitional area between the fluid phase and the transitional phase in Utterback’s three-stage theory. It does not predict that AI will inevitably complete the transition to the specialized phase. As a descriptive concept, it captures the current objective dual structure:

On the product side, the architecture, form, and competitive landscape remain open—multiple approaches coexist, and a dominant design has yet to emerge; this is still part of the fluid phase.

On the other hand, at the corporate level, the capability gap between leading and non-leading players has already begun to emerge—the disparities in computing power, distribution, enterprise customers, and organizational learning speed are quietly widening, marking the early stages of this transition.

When both sides overlap simultaneously, they form a "divergence zone." This is not chaos, as visible patterns and accumulated asymmetries already exist; nor is it settled, as the criteria for determining who ultimately prevails have not yet hardened. In one sentence: external appearances still seem prosperous, but compounded capability differences have already begun.

Further worth incorporating are three original insights proposed by Professor Karim Lakhani in his March 2026 long-form article, based on the same analytical framework—these can be viewed as a more detailed refinement of the mechanisms within the "divergence band."

Lakhani’s Observation One: Open architecture and asymmetric competition can coexist—openness at the architectural level and differentiation at the enterprise level are compatible. Lakhani coined the precise term “Structured Ferment” to describe this state, which points to the same phenomenon as the “differentiation band” in this paper—though this paper emphasizes the dynamics of “capability differentiation,” while Lakhani highlights the sense of multiple concurrent pathways in “fermentation.”

Lakhani’s Observation Two: The Hierarchical Concentration Hypothesis—The future industrial structure of AI may not converge to a “single winner takes all” model, but instead will see leading players emerge across different layers of the technology stack: frontier models, enterprise integration, open ecosystems, and specialized enablers (cloud, chips, toolchains, security).

Lakhani’s Observation Three: Multi-model orchestration is an emerging industry paradigm—real-world tasks are being broken down into a chain: ideation, retrieval, drafting, coding, reviewing, formatting, delivery—where different models excel at different stages. What truly determines success may not be “which model is the strongest,” but rather “who can invoke the right intelligence at the right moment for the right task.”

3.5 Summary of Specific Characteristics at the Current Stage

Based on the above analysis, the specific characteristics of the current "divergence zone" phase can be summarized as follows:

So, if large AI models and their applications are in the "diffusion zone" (moving from the late fluid stage toward the early transition stage), according to general industry evolution patterns, what will the future pace be like as dominant design emerges and deep consolidation becomes apparent?

This is a critical question for business decision-makers, as the sense of timing directly determines the pace of action—if the window is 20 years, planning can be deliberate; if it’s only 3 to 5 years, many actions must be initiated immediately.

A very natural inference is to refer to the historical data from Utterback and Klepper: the automotive industry took approximately 15–20 years to move from the fluid phase to the establishment of a dominant design, and 50–60 years from peak to deep integration—if AI follows a similar timeline, it seems we still have a long way to go.

But here’s a critical question—Utterback and Klepper’s two tools are both based on data from manufacturing. Can the time scales of industrial evolution in manufacturing be directly applied to the digital/AI era? If not, what are the actual time scales? This is the question we’ll address next.

Structural differences between manufacturing and the digital age—and why they are accelerating

The evolution of the manufacturing industry is naturally slow because it has several common characteristics:

Physical products require factories, supply chains, and distribution channels—expansion is constrained by physical limitations.

Establishing economies of scale requires substantial capital and time—factory construction and supply chain development are long-term endeavors.

• User adoption is limited by physical distribution speeds— it can take weeks or even months for a vehicle to go from production to reaching the end user.

• Technological evolution is constrained by the pace of iteration in processes, materials, and production equipment—each improvement requires rebuilding the production line.

Precisely because of these characteristics, the automotive industry experienced a congestion period of about 20 years, followed by several more decades of transition from dominant design to deep integration. These time scales are not "universal laws" of industry evolution, but rather the result of physical and capital constraints specific to manufacturing.

When these constraints are digitally broken, all four characteristics are undergoing fundamental changes, directly compressing the timescale of industry evolution:

• Distribution has near-zero marginal cost—an app can reach a global audience within months; ChatGPT reached 100 million users in just about two months, making it one of the fastest-adopted consumer products in history.

• Economies of scale have been replaced by network effects and data flywheels—while manufacturing economies of scale require years of factory accumulation, digital industries can establish exponentially growing self-reinforcing cycles within 1–2 years.

User adoption is limited by bandwidth and endpoints, not physical distribution—reducing time from weeks or months to just minutes for download and use.

• Technological evolution is constrained by talent and computing power, not by manufacturing processes or equipment—leading companies have accumulated massive capital, computing power, and talent in an extremely short time (in 2026, the top five U.S. tech companies have collectively committed to investing up to $700 billion in AI infrastructure), a pace of capital accumulation unimaginable in the manufacturing era.

In addition to these four fundamental changes, the digital era has additionally introduced two accelerating factors:

• Feedback loop acceleration—digital products enable real-time A/B testing, reducing feedback cycles from "years" to "days or even hours," thereby accelerating learning speed and significantly compressing the time required for capability compounding.

• Global Synchronized Competition — The digital industry has been a global market from the start, with more participants, faster competition, and more abrupt market shifts.

Key question: By how many times do these differences compress the timescale of the digital era compared to the manufacturing era?

Comparison of the Evolution Speeds of Six Digital Era Industries

Let’s look at several digital industries that have already completed the transition from the liquidity phase to the transition phase:

Smartphones (2007–2012) — The iPhone was released in 2007, Android joined in 2008, and by around 2010, the dual dominance of iOS and Android became clear. By around 2013, Nokia, BlackBerry, and Windows Phone had largely exited the mainstream market. Transition from the fluid phase to a stabilized phase: approximately 5 years.

Social Networks (2004–2012)—Facebook launched in 2004; multiple platforms competed between 2004 and 2008 (MySpace, Facebook); by around 2010, Facebook had established global dominance. Transition from the fluid phase to a stabilized phase: approximately 6–8 years.

Web browsers (1994–2000)—Netscape launched in 1994; the browser wars (Netscape vs. IE) occurred from 1995 to 1998; IE dominated the market from 1999 to 2000. Transition from the fluid phase to a stabilized phase: approximately 5–6 years.

Cloud Computing IaaS (2006–2015)—AWS launched in 2006; multiple competitors emerged between 2010 and 2013; by 2016–2018, the market was clearly dominated by three major players (AWS, Azure, GCP). Transition from the fluid phase to a stabilized phase: approximately 8–10 years.

Electric vehicles (2010–?) — Led by Tesla around 2010, with numerous new entrants emerging between 2015 and 2020, and a preliminary market leadership structure taking shape around 2025, though still in transition. Expected to stabilize: approximately 12–15 years (involving physical manufacturing, slightly slower than purely digital industries—this also reminds us that AI components involving physical forms, such as the mechanical/hardware bodies of embodied intelligence, will be significantly slower than purely digital components).

Quantitative judgment of time scale compression

How is the usable compression ratio derived from the above five industry duration data? The author has designed a logical deduction method.

Step 1 · Select the Benchmark Industry

The benchmark industry must meet three conditions:

Structurally complete: Go through the entire process from the liquidity phase to the transition phase and then to the specialized phase.

Verifiable data: Enterprise entry and exit, establishment of leading design, and deep integration with verifiable historical data.

Sufficient time span: Provides native time scale without compression.

The data on the U.S. automotive industry from 1895 to 1966 tracked by the Klepper system satisfies all three criteria.

But we must still ask: the industrial life cycles within manufacturing itself are not uniform (TVs have a T1 of about 6–8 years, cars have a T1 of about 15–20 years—a difference of nearly 2x). Why then use cars as the representative benchmark for the "manufacturing era"?

This can be explored through the analytical framework in Section 5.1—the six structural factors (physical distribution, economies of scale, user adoption, technological evolution, feedback loops, and global competition). This framework not only explains the differences between "manufacturing" and "digital," but also the variations within manufacturing itself. The television (which began in 1946) achieved adoption 2–3 times faster than the automobile (which began in 1895) because it inherited assets accumulated by the previous industry (radio) across these six factors—the distribution networks, vacuum tube manufacturing processes, user awareness, and technological foundations did not need to be built from scratch.

This insight reveals a deeper pattern: the compression of time scales did not arise solely from the leap from manufacturing to the digital age, but is instead a continuous, multi-generational, multi-factor acceleration process. The core reason for choosing the automobile as a benchmark is not merely the three superficial criteria—“structural integrity, available data, and long duration”—but rather that the automobile represents the most complete case of “building from scratch”: all six factors had to be initiated from nothing, with no prior industries to build upon, making it a pure example of an original time scale untouched by prior accumulation. The compression ratio derived using it as a benchmark can provide a rough estimate of the maximum cumulative acceleration achieved in the digital age.

Step 2 · Compare the five digital industries

Based on a car T1 (15–20 years):

Step 3: Categorize AI using the six-factor framework

Returning to the six-factor framework—the extent to which an industry experiences accelerated pre-accumulation across six factors directly determines its compression multiplier. Viewing the data in the table through this logic:

Electric vehicles: Minimum compression (1.3–1.5x), as they still require independent development of physical manufacturing, supply chains, and hardware iteration.

Pure digital industries such as web browsers and cloud computing: compressed by 2-3 times, as they build upon the accumulated advancements from the PC era and the early internet;

Smartphone app ecosystem: Greatest compression (3–4 times), as it emerged after highly mature digital infrastructure—nearly all preceding constraints among the six factors had already been resolved.

So, to which category do large AI models and their applications belong? The author analyzes and believes that:

It emerged at the latest stage of the digital era—cloud computing, the internet, mobile devices, developer ecosystems, and data infrastructure were all highly mature; at the same time, AI itself introduced two unprecedented self-reinforcing feedback loops—network effects and data flywheels (the more it’s used → the more data it generates → the stronger the model becomes → more people use it).

Based on the six-factor framework, AI’s compression multiple is estimated to be between 3x and 5x: the lower bound of 3x is derived from previously confirmed "late-stage acceleration" industries (such as web browsers and smartphone app ecosystems); the upper bound of 5x accounts for AI’s unique dual self-reinforcing feedback loops—a phenomenon without historical precedent—making this a reasonable but still unverified projection.

Cross-generational Overview and Comprehensive Estimation

Summarizing the above analysis, a comprehensive cross-generational comparison overview can be derived:

Comprehensive estimate: From the manufacturing era to the digital era, the timescale of industrial evolution has been compressed by approximately 3 to 5 times; the actual compression factor for AI large models and their applications is estimated to fall in the upper-middle range of this interval.

[Required Honest Disclosure]: The above comparative sampling only includes digital industries that have already established dominant designs, excluding VR/AR, the metaverse, Web3/blockchain applications, and other industries that remain undefined—all of which far exceed the 3-5x compression expectation and are still in flux. In other words, the "3-5x compression" assumption presumes that AI will follow the path of already-converged digital industries. If this underlying assumption changes, the 3-5x compression estimate must be correspondingly adjusted.

After estimating that the timescale has been compressed by approximately 3 to 5 times, we can proceed to analyze the specific timelines for AI large models and their applications. This is not a precise prediction, but a reasonable estimate based on historical comparisons. The basis for such an estimate is that AI large models will follow Path A (achieving a dominant design) rather than Path B (perpetual diversification).

My overall assessment is that the likelihood of AI large models following trajectory A is significantly higher than following trajectory B—but their form will not be a "single T-shaped vehicle," but rather a "layered protocol stack with multiple dominant designs."

In addition to the argument made in Section 2.1.1 from the perspectives of "economic requirements of industry operations" and "internet layered protocol stack," the following three currently observable signals strongly support this assessment:

• The technical toolkit has largely converged—Transformer architecture, pre-training plus post-training (SFT + RLHF/DPO), multimodal fusion, and tool calling/agentization are now commonly adopted by all leading models. This is typical of the period just before a dominant design emerges in history.

• Dominant designs are emerging at the sub-layer level—the programming agent layer has converged toward the Claude Code/Cursor/Copilot/Windsurf model; multimodal-native products are converging on the "diffusion model + long-sequence fidelity" approach; within the enterprise Copilot layer, Microsoft 365 Copilot has become the de facto standard. There’s no need for the entire AI ecosystem to converge together—sub-layers are independently forming their own dominant designs.

• The leading design for the protocol layer has taken shape—MCP (proposed by Anthropic at the end of 2024 and gradually adopted by OpenAI, Google, and Microsoft in 2026) and A2A networks—leading protocol layer designs tend to be more enduring than product layer ones (just as TCP/IP has outlasted any single browser).

[Required Honest Disclosure]: Once two opposite signals are confirmed, the time coordinates in this article need to be recalibrated:

Reverse Signal One: The "bigger is better" formula may no longer hold. Over the past five years, AI capabilities have primarily been driven by the blunt formula of "larger models + more data + stronger compute" (Scaling Law). If this formula reaches its limit by 2027–2028, with no strong alternative pathway emerging in the short term—the industry will be forced to explore entirely different directions, such as world models, Mamba, and agent-native architectures. The current perception of "technological convergence" could then unravel.

Reverse Signal Two: Evolution from “Parallel Camps” to “Dominant Designs Within Separate Camps.” The current parallel camps are a normal manifestation of “capability differentiation”—the real risk lies in this differentiation deepening to the point of “complete interoperability breakdown.” Two forces are driving this trend: first, the “brain” of embodied intelligence (academically known as VLA models)—due to the need for real-time perception and processing of physical feedback, their underlying architecture may evolve along a path entirely distinct from chat large models; second, the U.S. and China continue to widen their gaps in architecture, protocols, ecosystems, and regulation. If the evolution shifts from a “unified dominant design” to “separate dominant designs within distinct camps,” the timeline may change.

Three key timelines (scenario projections)

Continuing the previous analysis, if we define the industrialization starting point of large AI models as the 2020 release of GPT-3—a widely accepted industry milestone marked by the maturation of the Transformer architecture, the first emergence of emergent capabilities, and the first enterprise-grade commercial API—we can trace a timeline from 2020 to 2035, which the author has divided into four phases:

Expanding in detail from the overall diagram, enterprise decision-makers should focus on three key timelines for further analysis and attention (analysis scope: AI large models and their applications; components with independent physical forms or distinct industry logic must be evaluated separately based on their respective rates of evolution).

Timeline One: Leading the Design of the Floating Window—Layered Floating, Expected to Roll Out in Phases from 2026 to 2032

The convergence rates across different sub-layers are out of sync and vary significantly. Based on currently observable signals, a tiered estimate (rather than a precise timeline) can be made:

• Emerging earlier (around 2026–2028) — Protocol layer (MCP / A2A / API call specifications) and programming agent layer. Iteration costs at the protocol layer are relatively low, and cross-vendor consensus is forming rapidly, with the beginnings of a de facto standard already in place; the programming agent layer has already shown clear product convergence (Claude Code / Cursor / Copilot / Windsurf / Tongyi Lingma etc.

• Medium-term emergence (around 2027–2029)—Core product forms consist of three layers: general-purpose conversational assistants, multimodal native products (diffusion models + long-sequence fidelity approach), and embedded co-pilots (designed for enterprise workflows). Clear signs of market consolidation are emerging at the top, with leading products rapidly increasing their customer penetration; however, full stabilization will still require 2–3 years.

• Later emergence (around 2028–2030) — Office/Desktop Agent Workbench layer (represented by WorkBuddy, Coze, Bailian, etc.). Primarily driven by ecosystem integration, leveraging entry points, collaborative workflow connections, and ecosystem positioning to rapidly catch up; however, enterprise-level production deployment is still in the early stages of transitioning from "individual productivity enhancement" to "organizational productivity enhancement," with full maturation expected around 2028–2030.

• Earliest emergence (2030–2032 and beyond)—General autonomous Agent layer (e.g., Manus, Operator-type), multi-Agent collaboration paradigms, fundamental architectural innovations, and embodied intelligence "brain" layer (VLA models). General autonomous Agents and multi-Agent collaboration face deep challenges in reliability, cost structure, liability boundaries, and cross-domain protocols, converging significantly slower than purely digital layers. Current architectural layers exhibit only superficial convergence ("toolkit convergence": Transformer + MoE + post-training + memory architectures); true fundamental innovation is not expected until after 2030.

[Required Honest Disclosure]: The tiers above are merely reasonable estimates based on current signals—they provide a sense of timing, not a schedule. Before 2030, the architecture, form, and competitive dimensions are still evolving; after 2030, the rules of the game may begin to reset in a layered manner.

Timeline Two: Window Closure—Expected 2026–2030 (Frontier Model Layer to Close First)

Klepper’s key observation is that “the entry window closes earlier than the shakeout signal”—in the automotive industry, the decline in entry rates occurred in 1910, while shakeout became evident between 1920 and 1960, with a lag of 10–15 years. Compressed by a factor of 3–5 in the digital era, this gap shrinks to 2–5 years—meaning the entry window likely closes 2–5 years before deep shakeout, i.e., around 2026–2030.

However, it is necessary to view the different stages of closure hierarchically:

• Frontier Model Layer: Earliest to close; entering this layer as a pure newcomer will become very difficult after 2027–2028—the combined barriers of compute, talent, and corporate trust are nearing their limit. By 2026, global venture capital investment in the frontier model layer is already consolidating among a few players.

• Protocol and Toolchain Layer: Closed earlier; once protocol standards such as MCP and A2A become dominant, late entrants must integrate into the existing ecosystem rather than building anew.

• Application and product form layers: Relatively open; the programming agent layer is seeing its window of opportunity close rapidly due to the swift commercial integration of leading products. However, there remains considerable room for entry in application-layer products such as multimodal native applications and office/desktop agents, with opportunities still viable through 2028–2032.

• Vertical Industry Integration Layer: Open long-term; the window for integrating AI with vertical industries such as healthcare, legal, and education may extend beyond 2035.

Key takeaway: The urgency of entry windows varies significantly across different strategic layers—the window for betting on frontier model layers is the shortest, while the window for application layers and vertical integration is relatively more ample.

Timeline Three: Deep Consolidation Begins · Expected 2030–2035 (primarily occurring at the frontier model layer)

Klepper's capability compounding mechanism operates faster in the digital era—8 to 10 years may achieve the compounding effect that took 30 years in the manufacturing era. The capability gap between firms will be exponentially amplified within this window to an irreversible degree.

However, the intensity of deep shuffling is also stratified by layer:

• The frontier model layer faces the deepest reshuffling: composite capabilities, barriers of computational capital, and concentrated corporate trust—likely converging toward a few leading players.

• Lighter reshuffling at the application and product form layers: each layer may retain several key players, with a longer tail.

• Shuffle at the protocol layer rarely occurs: once stabilized, a standard remains a standard.

• The reshuffling of the vertical industry integration layer is tied to traditional industry logic: AI is just one of many variables; the pacing of industry reshuffling will be dominant.

Urgency of the enterprise entry window

Looking at these three timelines together, a key inference is:

Most companies focusing on AI large models and their applications as a strategic priority may have only 2 to 4 years for strategic adjustment—to establish their position before dominating design (especially at the product form layer) by 2030, they must complete direction assessment and entry selection by 2028; conversely, 2026–2027 are the most critical years.

However, this timeline assumes that economies of scale and business models will achieve critical breakthroughs by 2027–2030. The reverse validation in Part Seven will show that “economies of scale and business models remain underdeveloped” is one of the most accurate observations for 2026. If the challenges surrounding business models persist beyond 2028 without resolution, the window for dominant design emergence may be delayed by 2–3 years—meaning the above timeline hinges on two assumptions: “AI follows Path A” and “economies of scale achieve breakthrough within 3–4 years.” Any change to either assumption requires a recalibration of the timeline.

For components involving independent physical forms or distinct industry logic (such as embodied intelligent hardware and AI + vertical industries), timelines must be calibrated according to their own rhythms—but the underlying principle that "differences in composite capability growth rates are exponentially amplified" applies to all layers.

This is the "time scale of AI industry evolution"—providing a layered, conditional, and precise sense of time derived from an analytical framework.

A rigorous analysis must withstand scrutiny by the latest data. Using observable trends in the AI industry in 2026, conduct a reverse validation of the earlier judgment on the "divergence zone" across four dimensions: products, companies, business models, and on-the-ground developments—select just six representative signals.

Verification One: The architecture is indeed still open.

The architectural evolution in 2026 continues along multiple pathways: under the dominance of Transformers, Mixture of Experts (MoE) in Gemini 3.1 Pro, DeepSeek V4 and similar models are being fully adopted; inference-time computation has become a core competitive dimension for new versions like GPT and Claude; state space models (Mamba family) and native agent architectures continue to see new explorations in 2025–2026; the world model direction is actively advanced by leading researchers such as Fei-Fei Li and Yann LeCun. Context windows are rapidly expanding from hundreds of thousands to two million tokens; the diminishing marginal returns of Scaling Laws began to be publicly discussed in 2026—indicating that the technological path is still being reevaluated.

Conclusion: The architecture is clearly still in flux. The technical feasibility factors driving the design are far from being met—this is direct evidence of the "product dimension remains open" side of the "divergence zone."

Verification two: Enterprises are indeed diverging—at present, the landscape features multiple factions coexisting.

The real landscape of the frontier model layer in 2026 is a multi-party structure without an absolute leader:

• Leading U.S. group: Primarily OpenAI, Anthropic, and Google, followed closely by Meta (Llama series) and xAI

• China's Open Path Faction: DeepSeek Qwen, Kimi Zhipu, Doubao Hunyuan and others have developed independent competitive advantages in dimensions such as cost, open source, multilingual support, and vertical use cases.

• Europe/Other Camp: Mistral holds a place in the open-source ecosystem

At the same time, clear signs of divergence are evident—ChatGPT’s share of user traffic has significantly declined from its previous dominant position, Gemini’s usage has grown several-fold year over year, and Anthropic has achieved rapid growth in enterprise and developer segments, with its ARR now entering the tens of billions of dollars and quarterly revenue reaching the billions of dollars.

Conclusion: This fully aligns with the judgment of the "divergence zone"—on the product level, an open landscape with multiple competing camps remains (architecture and roadmap still undecided), but on the enterprise level, the gap in capabilities between leaders and non-leaders has become clearly evident. It’s important to note that "capability divergence" does not equate to "a few winners take all"—it could instead mean "multiple camps pulling away from each other."

Verification Three: The industrial structure is evolving toward "layered concentration + multi-model orchestration."

Lakhani's predicted "layered centralization" and "multi-model orchestration" were strongly validated in practice by 2026:

• Frontier Model Layer: Multiple leading vendors and multiple camps are simultaneously positioned (see 7.2)

• Enterprise Integration Layer: Microsoft (+OpenAI), Google Workspace AI, Salesforce Einstein, and others have each made their own deployments.

• Open Ecosystem Layer: DeepSeek (Low-cost inference), Llama, and Mistral are respectively anchored

• Specialized accelerator layer: NVIDIA (GPU), AMD/Intel, and Cerebras/Groq (new architecture chips) each occupy their position—barriers to entry in this layer are being pushed to their limits by capital concentration, as described in Section 5.1.

Meanwhile, multi-model orchestration has shifted from an "emerging pattern" to a mainstream practice—users no longer ask, "Which model is the strongest?" but instead ask, "Which model is best suited for each scenario?" Using Claude for writing and programming, Gemini for searching and analysis, and ChatGPT for image generation and general conversation has become standard.

Conclusion: Both the hierarchical centralization hypothesis and multi-model orchestration hold true—and the evolution of hierarchical centralization is occurring faster than expected.

Verification Four: The time scale is indeed being significantly compressed.

The primary evidence for time-scale compression comes from four curves: adoption speed, capital inflow, cost reduction, and capability iteration.

• Adoption rate: ChatGPT reached 100 million users in approximately two months, making it one of the fastest-adopted consumer products in history.

• Speed of capital inflow: Anthropic reached a nine-figure ARR in just about four years since its founding—something that would have taken decades in the manufacturing era.

• Cost reduction rate: The price of tokens for major large models has decreased by approximately two orders of magnitude over the past three years, with inference costs for leading open-source and low-cost models now entering the "cent-per-million-token" range.

• Speed of capability iteration: From GPT-3 to the GPT-5 series, less than six years—during the manufacturing era, a typical product generation cycle took 5 to 10 years.

Conclusion: Time-scale compression is not only valid; on certain specific curves (adoption, cost, capability iteration), the actual compression can reach 5 to 10 times. However, note that the compression rate of these specific curves exceeds the overall industry evolution rate, as the broader industry evolution is still constrained by slower variables such as business model validation, regulatory stabilization, and ecosystem maturity—remaining within the median range of 3 to 5 times as determined in Section 5.3. In other words, the "signal side" (curves) of AI industry evolution has clearly accelerated, while the "structural side" (establishment of dominant design) requires a longer period of consolidation.

Verification Five: Economies of scale and a stable business model have not yet been established.

Among the five enabling factors for dominant design, "the emergence of economies of scale" is a necessary condition—dominant design can only truly solidify when the cost of scaled delivery is sufficiently low, a stable profit model is clear, and a positive flywheel of "investment-output-reinvestment" is established. Applying this standard to AI large models and their applications in 2026 reveals that economies of scale and business models are far from being established:

• Training and inference costs are still undergoing significant fluctuations—token prices have dropped by approximately two orders of magnitude over the past three years, and inference costs for leading open-source and low-cost models have entered the "cent-per-million-token" range; the cost curve continues to shift rapidly and is far from stable.

• Leading companies continue to make large-scale, sustained investments—frontier model companies such as OpenAI and Anthropic are generally in phases of high investment and high fundraising; although Anthropic’s ARR has reached the tens of billions of dollars, it has not yet moved beyond the stage where investment far exceeds returns relative to training and compute costs.

• AI's return on investment (ROI) for enterprises has been slower than expected—many corporate AI initiatives remain in the proof-of-concept (POC) or pilot phase, and the transformation of AI into measurable gains in total factor productivity and positive ROI/ROE is still largely exploratory; according to 2025–2026 industry surveys such as McKinsey’s “State of AI” and MIT NANDA’s “GenAI Divide,” most enterprise AI investments have yet to yield clear positive financial returns.

• A sustainable business model is still under exploration—subscription fees, token-based billing, agent performance-based pricing, enterprise customization, API revenue sharing, open-source plus services… Each model has representative players, but none have yet achieved “scalable profitability + compounding growth.”

Conclusion: The positive flywheel of "investment-output-reinvestment" for large AI models and their applications has not yet been established—this is a direct manifestation of the failure to meet the conditions for economies of scale. If economies of scale and business models have not yet stabilized, the industry must still be in the differentiation phase, as this is a prerequisite for converging toward a specialized phase.

Verification Six: WAIC 2026 — "The more elite the individuals, the greater the divergence."

The recently concluded WAIC 2026 in July 2026 provided an on-site validation from the perspective of "human" involvement.

The four cutting-edge topics of the conference (physical intelligence, scientific intelligence, computing infrastructure, and agent governance) are also regarded as the direction of the "next phase"—a typical characteristic of this transitional period is the coexistence of multiple branches, with no single path emerging as dominant. A telling moment that illustrates this is that Richard Sutton, the 2024 Turing Award laureate, and Yin Qi, Chairman of Step Intelligence, nearly simultaneously offered two entirely opposing judgments:

• Sutton: "AI is still weak and unreliable," "the path of transferring human knowledge" has "reached its limits," and the future must move into the "era of experience."

• Yin Qi: "Model capabilities are crossing a critical threshold in 2026," "the industry is standing at the foot of the AGI summit," and agents will become "the smallest unit of productivity."

In the same venue, at the same moment, two highly representative figures offered nearly opposite answers to the question, "Where are we really?" Meanwhile, Yoshua Bengio warned that "society is not yet prepared for systems that are more capable and more autonomous." Xue Lan from Tsinghua University and Mark Nitzberg from Berkeley each proposed boundaries around "value judgments" and "life-and-death decisions"—while regulators and governance bodies are still debating principles, not implementation standards.

Conclusion: If AI has truly entered the "second half" or the "execution phase after dominant design has been established," top researchers and top entrepreneurs would not exhibit such stark disagreement on the most fundamental question of "Where are we really?" This itself is strong on-the-ground evidence that we are still in the "divergence phase."

Comprehensive verification: This is the "divergence zone."

Look at the first six sets of verification together—they are not six isolated events, but six independent perspectives corroborating the same conclusion. In one sentence:

Large AI models and their applications are currently in a "divergence phase"—product stratification remains open, capability differentiation has begun, and business models have yet to be validated. This conclusion is corroborated three times: in Part Three (Stage Analysis), Part Six (Timeline), and this section (2026 Trends), and is not based on a single perspective.

Back to the first two questions—where we are and how the pace will unfold in the future.

By combining the framework of two classic analytical tools, rigorously testing the five dominant design factors, and conducting reverse validation against the latest 2026 trends, we can provide a clear answer regarding current AI large models and their applications.

We are currently in the "divergence zone," where AI large models and their applications are transitioning from the late fluid stage to the early transition stage. The factors shaping dominant design have only partially satisfied one criterion, and the industry structure may evolve into a "layered concentration" rather than a "single winner."

What lies ahead—according to a layered timeline, dominant designs may emerge across multiple sub-layers around 2027–2030, with major reshuffling becoming evident between 2030 and 2035, leaving businesses approximately 2 to 4 years for strategic adjustment.

Layering means that the urgency varies across different strategic positions (frontier model layer / application layer / vertical integration layer), but the underlying principle of "compounded capability differentials" applies to all layers. What is provided here is a layered, conditional sense of timing—not a precise schedule. The specific pace of evolution will be calibrated based on changing conditions such as scaling laws and factionalization trends.

For business decision-makers, these two judgments together form a timeline. Above this timeline lies a more fundamental question that must be answered: At this stage and within this window of opportunity, how should a company think, act, and recalibrate itself?

The author will discuss this issue in the next article…

This article is from the WeChat public account "Tencent Research Institute" (ID: cyberlawrc), author: Li Hongsheng.