Following computing power and foundation models, data is becoming the next critical input in the competition for large models. As publicly available pre-training corpora gradually deplete, model training is shifting more towards post-training, reinforcement learning, and reasoning capability enhancement. The data required by model companies has also expanded from general text corpora to chain-of-thought, task trajectories, and synthetic data. Particularly in scenarios such as AI Coding, multimodal, and embodied intelligence, data is no longer just preparatory work before model training, but a crucial variable directly impacting model capability improvement. This is also reopening the market ceiling for the data services market.
In the past, data services were mainly concentrated in a few industries such as finance and marketing, with data annotation focusing primarily on standardized, labor-intensive tasks. With the emergence of large models, data needs have begun to cover more stages such as aggregation, processing, synthesis, and compliant circulation, while the technical barriers and delivery methods of data service companies are changing. TalkingData is a representative vendor in the field of AI data services. Since its establishment in 2011, TalkingData has accumulated profound experience in data processing and compliance across industry cycles including mobile internet, big data applications, and marketing technology.
Entering the era of large models, the company has strategically focused all its resources and capabilities on AI data services, positioning itself as a "next-generation AI data service platform" for large models and vertical applications. According to Cui Xiaobo, CEO of TalkingData, the data services market brought by large models could be more than ten times the size of the traditional data services market. Model companies no longer just need data collection and annotation, but high-quality data that can directly serve training, evaluation, and capability enhancement. The market space and bargaining power of data service companies are steadily rising.
In this interview, Aifenxi and Cui Xiaobo had an in-depth exchange on topics such as why the AI data services market is growing rapidly, what kind of data model companies need, how data service companies can enhance their technical capabilities, how FDE (Field Data Engineer) is changing the delivery methods of AI projects, and what new data needs will arise from enterprises applying large models.
Core Insights
- Large models are creating a new market ten times the size of traditional data services.
Traditional data services are mainly concentrated in a few scenarios such as finance and marketing, with relatively limited market boundaries. With the emergence of large models, data has become a core input on par with computing power and foundation models. The demand has also expanded from pre-training corpora to reasoning data, trajectory data, evaluation data, and synthetic data, further covering scenarios such as Coding, multimodal, healthcare, and embodied intelligence.
- Data is like a "mine"; the closer a company is to the data source and the stronger its ore-refining capability, the stronger its bargaining power.
Tech giants possess models, computing power, and engineering teams, but they may not have sufficiently rich external trajectory and error-correction data. Third-party data companies, through continuous aggregation, cleaning, and refining of data, are transforming from simple suppliers into key collaborators for model companies.
- FDE (Field Data Engineer) is not just a rebranded pre-sales/project manager, but a role directly responsible for outcomes and creating new value.
Large models have lowered the learning threshold for professional knowledge, also enabling technology companies to directly deliver business outcomes for the first time. In the future, customers will be more willing to pay for business outcomes rather than for a platform whose value is difficult to measure.
- Vertical models have limited value; new opportunities exist more in the connection layer between enterprises and models.
Large enterprises will internalize knowledge bases and core data capabilities, but they will still need external companies to complete vertical applications, data processing, and model connection. In the future, more intermediate data companies connecting enterprise data with model platforms will emerge.
The following is the transcript of this interview, slightly modified without changing the original meaning.
01. Large Models Expand the Data Services Market by At Least Ten Times
Aifenxi: How do you view the market opportunity for AI data services in this round, and what are the differences compared to past data collection and annotation?
Cui Xiaobo: I think we can first sort out the context of data applications. The first generation was before 2017, mainly used by internet companies, but basically within their own "walled gardens." External procurement was just a supplement, used in scenarios like gaming, e-commerce, and marketing. The second generation was the financial data services that emerged around 2018. Internet finance and loan facilitation companies lacked licenses and could not directly use the traditional credit reporting system, so a massive demand arose in scenarios such as risk control, credit evaluation, and anti-fraud, giving birth to a batch of data companies. Later, as policies changed and the licensed institution system was established, the business ceiling and capital ceiling of this market were restricted. The third generation was after 2020, when retail and consumer goods companies moved a large amount of business online due to the pandemic, leading to the emergence of data scenarios in marketing, supply chain, e-commerce, and content. The fourth generation is the wave of opportunities brought by large models after 2023. Its momentum is completely different from the previous ones.
Aifenxi: When did you first realize that the market opportunity in this round would be significantly different?
Cui Xiaobo: In 2023, large model companies and platform vendors were already approaching us, hoping to provide data collection and annotation services. But at that time, I felt the demand was still leaning towards BPO (Business Process Outsourcing), which is labor outsourcing business, and I didn't realize the market would be this large. The real change probably happened in 2024, when I started to see the market demand surging significantly. By 2026, especially after April, we saw the data procurement budgets of first- and second-tier model companies rising significantly. The three basic elements of AI are computing power, foundation models, and data. In the past, the capital market looked at computing power first, then focused on foundation models, and next, it will definitely pay more attention to data.
Aifenxi: Why has data become more important at this stage?
Cui Xiaobo: A direct reason is the emergence of the Data Wall, meaning publicly available data for pre-training is gradually depleting. The pre-training approach is still mainly Transformer and scaling laws, continuing to stack computing power and data, but there is less and less high-quality corpus that can be directly used. At the same time, model training methods have changed. Model competition has shifted from simple pre-training to post-training, reinforcement learning, and reasoning capability enhancement. This stage requires a massive amount of chain-of-thought, reasoning processes, and trajectory data. Many truly valuable data do not exist on public web pages; they come from real tasks: where mistakes were made, how errors were discovered, and how they were corrected step by step. Once the training method changes, the data needs of model companies change accordingly.
Aifenxi: Where is the value difference between trajectory data and traditional corpora?
Cui Xiaobo: Traditional corpora mostly tell the model "what the correct answer is," while trajectory data also retains "how to reach the answer." Especially in AI Coding scenarios, successful code is valuable, and errors along with the error-correction process are equally valuable. The model can see why a path failed and how to adjust it later, which in itself is excellent reasoning data. Tech giants naturally have a large number of programmers and development tools internally, which can also generate data. But internal data is more concentrated in the giants' own business scenarios, inherently lacking diverse trajectories from enterprises of different industries and scales. Third-party platforms aggregate richer scenarios, so the failure patterns and error-correction paths seen by the models are more diversified.
Aifenxi: How large a market has this demand growth formed?
Cui Xiaobo: According to our estimates, looking only at trajectory, reasoning, and evaluation data related to AI Coding, the global market has already exceeded 10 billion US dollars. Of course, specific numbers will fluctuate with model versions and procurement cycles, but the direction is very clear: the proportion of data costs in the overall model training costs continues to increase. From a broader perspective, the data services market brought by large models could be more than ten times the traditional data services market. In the past, financial data services were already considered a relatively large stock market, and marketing data was also huge, but they were often tied to traffic investment and operational businesses, making it difficult to separate them out. Now, scenarios such as Coding, multimodal, healthcare, and embodied intelligence are emerging simultaneously, and the market boundaries have suddenly opened up.
Aifenxi: Will the data needs of model companies grow continuously and at a constant rate?
Cui Xiaobo: No, it is more like leapfrog growth. If a model suddenly takes the lead in a vertical scenario, other companies will quickly catch up, forming a round of arms race, and the procurement demand for related data will burst out in a concentrated manner. When everyone's capabilities are leveled, or a new leading direction emerges, the demand will shift again. Therefore, the catching up among models will continuously create new data demands.
Aifenxi: How do you view the difference in model capabilities between China and the US?
Cui Xiaobo: I don't quite agree with simply saying "China is a few months behind the US." China and the US are more like asymmetric competition. The US has a stronger foundation in scenarios like language models and programming, with deeper accumulation of corpora and technology; China has its own advantages in multimodal, because the previous generation of mobile internet infrastructure was different, and China's short videos, short dramas, content production, and distribution are more developed, accumulating a large amount of related data. Therefore, models do not develop at a constant rate, nor do all capabilities advance along a single line, but break through one vertical scenario at a time. Today it might be Coding, tomorrow it might be video and music, and the day after tomorrow it might be assisted diagnosis and treatment. Every breakthrough in a scenario will bring a new round of data demand.
Aifenxi: Besides Coding and multimodal, which incremental scenarios are you more optimistic about?
Cui Xiaobo: Healthcare is a direction we are quite optimistic about. Healthcare data itself has high value density, and scenarios such as new drug R&D, diagnosis and treatment, imaging, and medical records all require professional data. I judge that the market size of healthcare data services will soon surpass traditional financial data services. Embodied intelligence is also an incremental market, but at this stage, the scale is not as large as the outside world imagines. It requires world models, physics engines, and simulation environments. The direction is valid, but real shipment volumes, standards, and application scenarios are still forming. We will not make blind judgments just because the capital market is hot, but look at where the data is ultimately used and who is willing to pay continuously.
02. Companies Closer to the "Mine" Have More Obvious Advantages
Aifenxi: As data service demands change, how will the capability structure of AI data service companies adjust?
Cui Xiaobo: In the past, when people mentioned AI data services, the first thing they thought of was annotation companies. But now the supply side (AI data service companies) has clearly stratified. The first category is BPO (Business Process Outsourcing) companies, suitable for undertaking standardized, labor-intensive tasks, such as a large amount of repetitive bounding, classification, and review. The second category is algorithm and model companies, which not only deliver data to clients but also understand model training, capable of data cleaning, transformation, evaluation, and synthesis. The third category is data circulation platforms, solving the traceability, compliance, and circulation issues of data procurement. In the future, truly valuable companies will often not just occupy one of these layers, but combine data aggregation, algorithm processing, model verification, synthesis environments, and compliant circulation.
Aifenxi: Can traditional BPO companies upgrade to algorithm and model companies?
Cui Xiaobo: It's very difficult. Traditional annotation companies are mainly operational organizations, good at managing a large number of personnel, controlling costs, and ensuring standardized delivery. Algorithm and model companies need people who have actually done pre-training, post-training, or reinforcement learning; the talent structure and cost structure are completely different. If the team does not do model training themselves, many trajectory and reasoning data will still be unusable after processing, and clients will have to conduct large-scale secondary cleaning after taking them back, which is very inefficient. Now, tech giants are increasingly tending to require suppliers to first establish their own models and evaluation standards, run the data through them, and then enter the client's quality control process. At the same time, it is best for suppliers to be able to provide synthetic data algorithms and simulation environments.
Aifenxi: Why don't tech giants keep all these capabilities in-house?
Cui Xiaobo: In the past, tech giants did spend a lot of budget on constructing data internally, and the proportion of external procurement was not high. But now an obvious trend is that internal data generation models, simulation environments, and processing tasks are also starting to be outsourced. The reason is not simply calculating labor costs, but that external companies are closer to data sources. If tech giants could train models well relying only on data generated from their own scenarios, they certainly wouldn't outsource. The problem is that practice has proven that multi-source trajectory and error-correction data from third parties are very effective for model improvement. Tech giants have model capabilities, but they cannot continuously cover countless fragmented external scenarios, nor are they suitable for doing this aggregation, cleaning, and processing work in the long term.
Aifenxi: Companies with data sources look like "mining companies"?
Cui Xiaobo: Yes, this business is increasingly like mining and refining ore. First, you need "raw ore," which means being able to continuously aggregate data; second, you need "ore-refining" capabilities to clean and process messy data into high-quality data that models can use; third, you also need to be able to synthesize data and construct scenarios that are difficult to obtain directly in the real world. In this chain, companies closer to the "mine" have the advantage. They collect scattered ore and refine it into high-purity "rare earths." Will model companies buy it? As long as this kind of data can significantly improve model performance, data companies will no longer be the passive order-receiving party, and their bargaining power will become stronger and stronger. In this chain, TalkingData's positioning is an "ore-refining platform," possessing massive multimodal data assets and able to continuously aggregate multi-source data, refining them into high-quality datasets that models can directly use through algorithm processing, synthesis, and compliant circulation.
Aifenxi: What is TalkingData's core capability in this chain?
Cui Xiaobo: Data service companies still need to have their own genes. TalkingData has been in the data business for so many years; how to handle compliance, how to aggregate data, and how to clean and process it—these basic capabilities have always been there. It's just that in the past, the scenarios were not large enough, and having capabilities alone didn't leave much room for revenue. Now large models have amplified these capabilities. A few years ago, we started concentrating resources on data and model-related matters, and also participated in infrastructure construction such as trusted data spaces. When model companies procure data, they not only require data quality but also need to clearly explain the source, authorization, processing, and circulation links. Our positioning is not simply to pass a transaction, but to complete processing, evaluation, and delivery while data circulates compliantly.
Aifenxi: Does this mean a "data circulation platform" cannot just be a transaction channel?
Cui Xiaobo: Yes. The problem with a pure transaction channel is that there is no service, just passing accounts. Model companies will feel that apart from providing compliance certificates, you haven't helped them reduce internal processing costs. A truly valuable platform must have both compliant circulation capabilities and model capabilities. After the data comes in, it needs to help clients with cleaning, processing, and quality verification, and preferably construct a synthesis environment. What the client gets is not a hard drive, but a traceable, verifiable, high-quality dataset that can directly enter the training process.
Aifenxi: In highly sensitive scenarios like healthcare and finance, data is inconvenient to circulate directly. How is this solved?
Cui Xiaobo: The main approach adopted is "data usable but invisible." For example, in healthcare scenarios, we strictly follow the hospital's data security management system and cooperate with the hospital to complete data organization and annotation within a compliant framework. At the same time, without touching the original medical records, we construct a synthetic data layer based on desensitized real data to simulate more statistically valid data samples. Model training and verification can be completed within the scope of data authorization, and what is ultimately delivered is model parameters or evaluation results, rather than original medical record data. This way, we can utilize data value while reducing the risk of sensitive data leakage. Simulation in healthcare, finance, and embodied intelligence will widely adopt similar approaches in the future.
Aifenxi: Why do you emphasize that TalkingData must be very close to model companies?
Cui Xiaobo: Cognition comes from the front line. If you just stand outside and watch, it's hard to know what the next-generation models are actually attacking and what data they need. We have been very low-key in the past two years, but we maintain high-frequency communication with the algorithm, engineering, and platform teams of leading model companies. Their first-hand needs are directly raised to us, and we also discuss the direction and strategy of next-generation models. Being close to clients is not just to get orders, but to see how demands change in advance. After seeing deeper, the organization and resources can be more firmly piled in one direction, forming a positive cycle.
03. FDE Is Not On-site Pre-sales, But Creating New Value
Aifenxi: TalkingData is a company with a long history. Facing this round of AI changes, will the historical baggage of an old company be heavier?
Cui Xiaobo: There will be some. Last year, when we communicated with the capital market, whenever people mentioned TalkingData, they felt you were still an old company doing marketing technology or traditional data services. Later, we spent a long time adjusting our business. Now our external positioning is very simple: we are an AI data service platform, with the vision of providing all the data needed by AI large models, and our core products are high-quality datasets and related services. A few years ago, we gradually stopped many original software and marketing technology businesses, concentrated resources on data and models, and adjusted the organization to the current two to three hundred people. This process was very painful because old clients, contracts, and teams had to be gradually digested. But after the adjustment, it's more like a new company without historical baggage, while retaining compliance, licenses, infrastructure, brand, and core talent.
Aifenxi: Why do you think many companies, even if they figure it out now, find it hard to turn around immediately?
Cui Xiaobo: Figuring out the cognition is just the first step; if organizational capabilities cannot be adjusted, you can only watch. Especially for companies with more than 500 people, originally hundreds of people were doing software, projects, and delivery. You cannot change the talent structure overnight; existing clients and contracts cannot just be stopped. Many founders will struggle: the original market is still there, the story has always been told this way, should we go all out for the new business? We have already dealt with these most headache-inducing matters a few years ago, so when new opportunities come, we don't have too much baggage and can directly throw money, people, and brand into it. This kind of advance in organization is very important.
Aifenxi: How do you view the endgame of foundation models? Will general AGI (Artificial General Intelligence) arrive soon?
Cui Xiaobo: I think AGI is still some distance away. The competition for general models will continue for a long time because there is no once-and-for-all moat. What has truly been proven to work now are still a few vertical scenarios. Later, it will also be conquering one scenario at a time. Agents will probably form three categories: the first is life assistants, where entry points and ecosystems are very important; the second is office and efficiency agents, where the US market is more likely to run the model, and China will have many followers, but the market and payment basis are different; the third is highly vertical agents. It is difficult for tech giants to invest resources in every segmented industry in the long term, which instead leaves opportunities for startups, and the more vertical it is, the clearer the opportunities may be.
Aifenxi: Moving from general models to vertical industries, why does FDE become important?
Cui Xiaobo: The biggest challenge in vertical fields is often not the model, but the lack of common professional knowledge and business language. For example, a new drug R&D team raises a demand, but what they need is diagnosis and treatment data from hospitals. Pharmaceutical factories and hospitals seem to be in the same healthcare industry, but they are actually two sets of knowledge structures and discourse systems. In the past, it was difficult for the two sides to communicate efficiently. Later, we dispatched a group of people with both technical backgrounds and the ability to quickly learn professional knowledge on both the pharmaceutical factory and hospital sides. Large models have lowered the threshold for knowledge learning. After two or three weeks, they can communicate professionally with R&D personnel and doctors. The two groups confirm the needs in the morning and can extract and construct data on the hospital side in the afternoon. In the past, this link might take two months to run through.
Aifenxi: What is the essential difference between this and traditional on-site pre-sales and project managers?
Cui Xiaobo: It's not just changing the name of pre-sales or project manager to FDE. The key is what new value this person creates on-site. They must simultaneously possess professional knowledge, technical background, and the ability to solve client problems, connecting the previously disconnected demands, data, and models. If what they do after being on-site is still demand communication, project coordination, and platform selling, then there is no difference from before. FDE should enable clients to obtain results they couldn't get before, such as compressing a two-month link into one day, or turning data that couldn't be directly used into trainable and evaluable datasets.
Aifenxi: You just mentioned "results." Will this change the business model of AI companies?
Cui Xiaobo: Yes. In the past, technology companies liked to sell platforms, tools, and software licenses, but clients are increasingly finding it hard to accept. Clients will ask: What's the difference between your platform and me directly calling the same foundation model? Everyone can use the underlying model; why should you charge a very high platform fee? After the emergence of large models, technology companies can directly deliver results for the first time. For example, if a client needs advertising materials, we don't necessarily sell them a material generation platform, but deliver a certain number of qualified materials every day, settling based on quality and conversion effects. What the client wants is the materials, not a platform that still requires them to operate it themselves.
Aifenxi: But many companies still hope to use FDE to do projects first and then sell the platform in. Can this logic still hold?
Cui Xiaobo: It will face two layers of pressure. The first layer comes from clients. Business is hard to do now; clients are more willing to say, you help me earn ten dollars, I'll share one dollar with you; or you help me save ten dollars, I'll share one dollar with you. Apart from this, clients are unwilling to easily pay for platforms whose value is unmeasurable. The second layer comes from model and cloud vendors. The scaffolding, orchestration tools, and middle layers you build are easily replaced by a single version upgrade of the underlying model. In the end, you are either crushed by tech giants or abandoned by clients. A more stable position is being responsible for results: treating the supply from tech giants as tools, and becoming the creator and distributor of business value yourself. As long as you can continuously create measurable results for clients, this relationship will not disappear due to a certain model upgrade.
04. Enterprises Will Internalize Core Data Capabilities; Third-Party Opportunities Lie in the Connection Layer
Aifenxi: From the perspective of national data elements, what role will the National Data Administration play in the development of the AI data services industry in this round? Cui Xiaobo: We have quite a lot of contact. I understand that the National Data Administration is now mainly focusing on three types of things. The first is data infrastructure, including the construction of trusted data spaces; the second is high-quality datasets, focusing on serving large model training, post-training, and applications; the third is gradually integrating industrial promotion, standards, and regulatory capabilities. In the past, data transactions stayed more at "whether there is a trading venue," and now it is starting to shift to "whether data can really be used around AI scenarios." High-quality datasets are not simply piling up raw data, but require expert knowledge, reasoning logic, quality standards, and traceable links.
Aifenxi: What space will these infrastructures leave for private data companies?
Cui Xiaobo: Public data and important data assets can be held by state-owned entities such as local data groups. Private enterprises do not necessarily need to own data; instead, they can do applications, processing, and operations. For data companies, being able to legally use, process, and create value is more important than owning the data itself. Holding data in your own hands actually brings greater responsibility and risk. Institutional ideas such as trusted data spaces and the separation of data rights and interests are essentially gradually clarifying boundaries: who holds, who uses, who processes, and who operates. Once the boundaries are clear, startups will know what they can do, and clients will also dare to hand over their business.
Aifenxi: Will enterprises own an independently growing "enterprise AGI" in the future, and how do you view the opportunities for vertical models?
Cui Xiaobo: I think most enterprises will not train foundation models themselves. We have discussed with some global brands, and their choices are very clear. Foundation models will be accessed from Microsoft, overseas model vendors, or domestic tech giants, and they will not build a set in every country and every business unit. In the AI era, large multinational enterprises are actually more inclined to have a single set of global capabilities. In the past, informatization could have one set for the US and one for China. Now, because the investment in foundation models is too large, algorithm and model teams will concentrate towards the headquarters in Europe and the US. For the Chinese market, this is not necessarily a good thing; the autonomous space for local teams at the model and platform levels will shrink. If vertical models are just "fine-tuned on open-source models," their value is indeed limited. But if vertical models are superimposed with industry-specific data flywheels and scenario closed loops, the moat will be significantly deeper. However, from a larger opportunity perspective, we judge that the connection layer between enterprises and models—that is, data aggregation, processing, evaluation, and compliant circulation—will form a more sustainable market than the vertical models themselves.
Aifenxi: Since enterprises do not train foundation models themselves, can third-party AI companies still enter the enterprise internally?
Cui Xiaobo: They can enter, but the space may be smaller than many people imagine. Large enterprises will keep knowledge bases and core data capabilities in their own hands. They will say, I've even handed over the foundation models to tech giants; if I hand over the knowledge base to a third party again, what core capabilities are left? Therefore, knowledge bases, core business data, and unified AI platforms are very likely to be internalized by enterprises. External companies are more suitable for doing vertical applications, efficiency tools, and clearly deliverable business results. You cannot just talk about a grand enterprise AI platform, but answer in which business link, how much more money I can help clients earn, how much less they can spend, and how much time can be shortened.
Aifenxi: Then will the data service demands on the enterprise side eventually only have large clients like foundation model companies in the long term?
Cui Xiaobo: Not necessarily. The demand for training large models within enterprises may be limited, but new data service demands will arise between enterprises and models. For example, models and search platforms will stipulate a set of metadata formats, and enterprises need to organize products, content, and business information into a structure that models can correctly crawl and understand. Specialized data companies will emerge in the middle to help enterprises optimize websites and data, process them according to the requirements of different model platforms, and then submit them to models or search platforms. These companies do not run into enterprises to train models, but connect enterprises and models.
Aifenxi: How large is the market opportunity for this "connection layer"?
Cui Xiaobo: I think it will form a very large market, but it will not appear in the form of a single product or platform, but be dispersed in different enterprises, industries, and business scenarios. The stronger the foundation models, the easier it is for the general capabilities within enterprises to be completed by tech giants or the enterprises themselves. But the data structures, product information, and business semantics of different enterprises are all different. These contents will not automatically become formats that models can understand and use, and model vendors cannot solve these long-tail problems one by one. Therefore, a large number of intermediate forms will emerge in the future, including Search, GEO, data structure optimization, model access, and data processing in vertical scenarios. They connect enterprises at one end and model platforms at the other, helping enterprises organize products, content, and business information into a structure that models can correctly understand, crawl, call, and convert. The opportunity in this market does not lie in making another foundation model, but in serving a massive number of enterprises and application scenarios. The more links that need connection and processing between enterprises and models, the larger the market space of this layer will be.
Aifenxi: What are TalkingData's next key priorities?
Cui Xiaobo: We are simply an AI data service platform, providing the required data for AI large models and vertical applications. Our core products are high-quality datasets and related services. Specifically, it is about continuously doing a few things well: first, getting close to data sources and aggregating more "raw ore." Second, refining the data using algorithms and models. Third, constructing synthetic data and simulation environments, and then safely circulating the data through compliant infrastructure. In the past, scenarios were small, and these capabilities were not fully amplified; now that the scenarios are here, what we need to do is reorganize our years of accumulation and seize this round of opportunities.