Zhang Yiming tops the list of the richest people in Asia: Why is ByteDance getting smarter the more it expands?

On September 16, the Bloomberg Billionaires Index was updated: Zhang Yiming, founder of ByteDance, overtook Indian billionaire Gautam Adani with a net worth of 105 billion U.S. dollars, topping the list of Asia's richest people for the first time. Back in March 2019, this figure was only 13 billion U.S. dollars.

In the same week, another piece of news seemed somewhat "out of place". According to foreign media reports, ByteDance's net profit in the first half of this year declined year-on-year due to increased investment in artificial intelligence; at the same time, its revenue increased by 30%, and the proportion of overseas revenue hit a new high.

Profits are falling, while personal worth is rising. Obviously, what the capital market is optimistic about is not a single hit application, but ByteDance's ever-accelerating growth mechanism.

Attributing ByteDance's development to "powerful algorithms" is the most convenient explanation, but it is like building a "black box" that cannot bring people any inspiration. From Toutiao to Douyin, TikTok, and then to Lark and Volcano Engine, ByteDance has not only pushed short videos to the extreme, but also continuously expanded its boundaries outward. However, in strategic studies textbooks, deep cultivation of scale and diversified expansion are precisely two things that are difficult to achieve at the same time.

A latest study published in the *Strategic Management Journal* spent six years and conducted 41 interviews to deconstruct this paradox. The researchers found that under appropriate conditions, AI will not be "diluted" when reused across fields, but will become stronger the more it is used: every time it enters a new market, it gains a new set of data; the model trained by the new data makes the next expansion easier. But to achieve such an effect, organizational design always needs to evolve synchronously with AI capabilities, from the AI middle platform and business partner mechanism to the restructuring of six major business units in 2021.

The richest list will change, but the logic behind ByteDance's "the more it expands, the smarter it gets" is worth serious reading for every manager who is thinking about AI strategies.

In 2021, TikTok once broke the internet traffic pattern dominated by Google. According to the annual global traffic report of Cloudflare, an internet infrastructure service provider, TikTok was the most visited website of the year, pushing Google, which had dominated the search field for more than 20 years, off the top spot. In the same year, ByteDance, TikTok's parent company, also quietly completed an important restructuring. The company, which was founded less than ten years ago, was split into six major business segments:

Douyin, TikTok, Lark (office collaboration), Volcano Engine (enterprise-level AI services), Dali Education, and Nuverse (games).

Looking at these two pieces of news together, they are somewhat counterintuitive. On the one hand, Douyin and TikTok have pushed short video applications to the extreme; on the other hand, ByteDance is like a commercial empire that keeps expanding its boundaries. In classic strategic studies textbooks, these two things are usually difficult to hold true at the same time. The more diversified an enterprise is, the more likely it is to encounter cost inflation, coordination failure and management overload. But ByteDance has not significantly slowed down its growth while continuously entering new fields. How did it achieve this? Could it be that the textbooks are wrong?

A latest case study published in the *Strategic Management Journal* attempts to answer this question. The research team spent six years studying ByteDance, conducted 41 interviews, and collected a large number of internal and public archives.

Their conclusion challenges a basic assumption in strategic studies: valuable resources enable enterprises to diversify. In ByteDance's case, the logic is exactly reversed. Diversification does not only consume the resource of AI, but also makes AI itself more valuable.

Why Does Growth Often Come at a Cost?

For most of the 20th century, making a business bigger often meant "choosing a battlefield". A company can deeply cultivate a certain product or service, build operational advantages, continuously reduce costs, and lock in customers through scale; it can also enter multiple markets, hoping that its existing capabilities can be reused in different fields.

But most strategic scholars believe that it is difficult to do both things well. Factories, sales teams, managers' time and energy, and physical capital are all limited resources. Investing in one market means that you cannot invest in another market at the same time.

The rise of digital business has made this trade-off less stringent than in the past, but it has not really disappeared. Software, algorithms and data have the characteristics of "low marginal cost": once the system is built, the cost of serving one million users will not be much higher than serving one thousand users. But when digital enterprises try to diversify, they will still encounter the problem of resources being diluted.

A well-known example is Meituan. When Meituan expanded from food delivery to ride-hailing, operational resources, platform traffic and management energy needed to be redistributed between different businesses, and the ride-hailing business has never been able to shake Didi's position. When Uber launched its food delivery business, platform traffic, delivery capacity and brand attention needed to be redistributed between the two business lines, and some drivers would also switch between the two types of orders.

ByteDance's core bet is to make AI the key to breaking this deadlock. AI is not just another digital resource, it has a feature that traditional strategic theory has not fully anticipated: self-learning. Every time the system processes an interaction — a user stays a few more seconds on a certain video, skips a short video, clicks on a certain title, forwards a certain post — the model will get new feedback. When the same architecture is deployed to a new business field and can obtain data with a similar structure to the original field, the model is not just "moved over" for use, it will become better.

Researchers call this phenomenon "cross-fertilization": learning in one scenario will improve performance in another scenario; and the new data brought by the latter will in turn strengthen the entire system. This subverts the traditional strategic theory. We usually think that valuable resources are the premise of diversification; but in AI-driven companies, diversification can in turn amplify the value of the resources themselves.

Research Methodology

To support this conclusion, the research team conducted a very in-depth case study. ByteDance is a privately held company, and external researchers have very limited access to it. But the research team continuously collected data from 2019 to 2025, and traced back to restore the company's development trajectory since 2012. They conducted 41 semi-structured interviews with 36 ByteDance-related professionals. These respondents included engineers, product managers, advertising staff, public relations staff, international team members, and academic AI experts who have cooperated with the company. The research team held five group discussions with relevant personnel, visited ByteDance's Beijing headquarters many times, collected 741 pages of text archives and 853 minutes of video materials, and cross-compared these materials with industry reports, public interviews, and news reports from 2012 to 2025.

The research team presented ByteDance's diversified development path as a three-stage process model: first use AI to expand the scale in a single field, then extend AI to multiple fields, and finally use organizational restructuring to manage both scale and breadth at the same time. At each stage, AI capabilities and organizational design evolve synchronously. As long as one side cannot keep up, growth will hit a bottleneck.

Stage 1: From "People Looking for Information" to "Information Looking for People"

When Zhang Yiming, founder of ByteDance, launched Toutiao in August 2012, the Chinese internet market was already dominated by giants. Baidu led the search sector, Alibaba dominated e-commerce, and Tencent controlled social networking. The three companies, commonly known as "BAT" in the industry, together accounted for nearly 70% of the market value of listed Chinese internet companies. If ByteDance continued to follow the search logic and compete head-on with these giants, it would have almost no chance of winning.

ByteDance chose to subvert the logic itself.

Toutiao's AI system transformed information distribution from "people looking for information" to "information looking for people". The system no longer waits for users to actively input what they want to see, but infers their interests from their behaviors: what they clicked on, how long they stayed, what they skipped, and what they forwarded. Then the system actively pushes content that may match these interests to them.

Every interaction generates new training data. The first 100 pieces of content that new users see are actually a key test. If the system can figure out user preferences quickly enough, users are more likely to stay. Research shows that ByteDance achieved a 45% user retention rate at this stage, which laid the foundation for its subsequent growth.

However, technological innovation alone is not enough to support the company's development. In the early stage, ByteDance intentionally adopted a flat organizational structure with very few decision-making levels, and employees called each other "classmates" regardless of their ranks. The company also adopted a transparent OKR system, where employees can see the work goals of different levels, even those of executives. These organizational arrangements are important because the front-end product team and the back-end algorithm team need to iterate quickly: the product team observes user behaviors, and the algorithm team adjusts model parameters, sorting weights and testing processes. The purpose of organizational design is to shorten the distance from "user feedback" to "technical response" as much as possible.

This is the first stage, where AI becomes the engine of scale growth in a single field. But a more difficult question follows: can this engine be moved to other places?

Stage 2: New Wine in Old Bottles

ByteDance's next strategic direction was short video applications. Douyin was officially launched in 2016, and developed into TikTok after entering overseas markets.

Toutiao mainly focuses on text content. The biggest difficulty when reusing this set of recommendation engines in the short video track is to rebuild the input pipeline. The system no longer extracts meaning from text, but needs to understand videos, images, audios, music, actions, subtitles, and the interactions between users and these contents. ByteDance therefore adopted a series of AI technologies such as computer vision and speech recognition to "translate" short videos into features that the recommendation algorithm can process. As the business expanded wider and wider, repeating this work one by one became less and less cost-effective, so ByteDance built an "AI middle platform": a centralized infrastructure to precipitate algorithm capabilities, content understanding models and aggregated user data for new products to call.

When ByteDance expanded its business overseas, this middle platform was also reused in overseas markets. A ByteDance engineer used a vivid metaphor: The algorithm architecture is the "wine bottle", and the user behavior data trained locally is the "wine". Douyin contains the data of Chinese users, and TikTok in the United States contains the data of American users, but the "wine bottle" is largely the same.

The concept of "middle platform" originated from Alibaba's "large middle platform, small front desk" strategy proposed in 2015, which means precipitating general technical capabilities into the middle platform, and then flexibly calling them by front-end businesses. But ByteDance's AI middle platform has a key difference: it is not a static tool library, but a system that keeps learning. Every new piece of data sent from the front end gives the middle platform model a chance to become a little smarter.

In order to make the middle platform truly close to specific businesses, ByteDance also set up the position of business partners, who act as connectors between technology and business. They are nominally affiliated to the middle platform, but usually work side by side with front-end engineers and product managers in the business team, and collect new experiences generated in the business to feed back to the middle platform. The research points out that with this mechanism, ByteDance can incubate a new application in about three weeks.

More importantly, diversification did not "dilute" ByteDance's AI resources, but enriched the entire system. The behavior data of readers on urban-themed works in Fanqie Novel can in turn help Toutiao push news titles more accurately; the click data on Douyin can also optimize the advertising algorithm in the entire ecosystem. Every time it enters a new field, the algorithm can be exposed to more diverse signals, and the entire product matrix benefits as a result.

Stage 3: Co-evolution of Organizational Structure

In the early 2020s, the limitations of the centralized middle platform began to appear. ByteDance operates dozens of applications at the same time, and different products have different priorities: some pursue user retention, some focus on advertising monetization, some pay attention to e-commerce conversion, some need to support creators, and some are for enterprise services. No matter how sophisticated the unified middle platform is, it cannot serve all goals at the same time.

At this stage, some of ByteDance's products failed one after another, including its attempts in knowledge payment, social networking and independent e-commerce. The problem was not that AI failed, but that AI alone could no longer balance the interests of multiple businesses and departments, so the organization needed to evolve again.

In November 2021, ByteDance was restructured into six business units: Douyin, TikTok, Lark, Volcano Engine, Dali Education and Nuverse. Each business unit obtained more exclusive technical and AI support, and no longer fully relied on the unified middle platform. This decentralization is not complete, and some shared functions, such as the live streaming algorithm shared by multiple short video applications, are still retained centrally. But in most cases, the AI team can work closer to the strategic priorities of specific business units.

Such restructuring does not come without a cost. It means re-segmenting highly integrated resources, redefining decision-making power, redesigning the technical architecture, and dealing with internal organizational resistance. ByteDance did this because the old structure had hit the ceiling. The new structure is a more adaptive organizational arrangement: the middle platform can be split according to business needs, and can be re-merged when necessary.

Centralization facilitates cross-field learning, and decentralization facilitates local adaptation. A suitable organizational structure is not designed once and for all, but continuously adjusted in the process of growth. For enterprises that want to use AI as a "cross-fertilization resource", this is a key experience: enterprises must be prepared to repeatedly restructure themselves.

Can ByteDance's Experience Be Directly Replicated?

Not all companies can copy ByteDance's experience. To truly give play to the cross-domain learning advantages of AI, at least several prerequisites are required.

First of all, enterprises must have sufficient data and AI capabilities as a starting point. The reason why ByteDance can scale AI is that it has massive user interactions, high-quality feedback loops and a strong engineering team. For industries with scarce data, high customization or strict supervision, it is difficult to start from the same starting point.

Secondly, the organization must have the ability to absorb what AI has learned. If advanced AI is trapped in isolated systems or rigid hierarchies, its learning can only go around in local areas. ByteDance's advantages come not only from algorithms, but also from an organizational design that allows learning to flow across products.

Equally critical is that the organization must be willing to keep changing. ByteDance's early flat structure and OKR culture supported rapid iteration; but as the company expanded, it had to rebalance centralization and decentralization at the business unit level.

Finally, cross-domain learning will only take effect when different fields can generate data that "the same AI system can understand and utilize". ByteDance benefits because video clicks, news reading duration, and novel browsing behaviors essentially generate user interaction signals with similar structures. If the data forms generated by the two businesses are too different, even if they belong to the same group, they may not be able to strengthen each other.

Management Insights

Build a symbiotic ecosystem of strategy and organization to make AI truly work

The most important insight brought by this research is to stop using AI only as a tool.

Although the explosive growth of AI has been going on for several years, many companies still treat AI as a new tool or function when using it: adding a chatbot, automating a certain process, optimizing a certain decision-making link. These practices can certainly create value, but they only touch a small part of AI's strategic potential. The bigger opportunity is to build a system where the data and feedback generated by one business link can empower another link.

This will also change the way leaders think about diversified development. The question in the past was: are new businesses related to the company's existing businesses? In AI-driven companies, a more worthy question to ask is: Can new businesses generate data that the company's existing AI system can learn from? The so-called "relevance" no longer only refers to whether products or markets are similar, but also refers to whether different businesses can generate signals that can jointly strengthen the same AI system.

To describe this change with the simplest metaphor, it is like a "flywheel". Once it starts rotating, the newly added business is no longer just a burden, but may become a force that accelerates the entire system. But this flywheel will not rotate by itself. ByteDance's experience shows that AI strategy and organizational design must evolve together. The AI middle platform allows learning to flow between products, and the subsequent organizational restructuring leaves room for local adjustments for business units. In contrast, no matter how excellent an AI system is, once it is trapped in organizational islands, it