In September this year, a company established less than a year ago was reported to be negotiating a funding round of approximately USD 300 million. Rumored lead investors include Alibaba, with Tencent and Sequoia China participating, bringing the post-money valuation to USD 2.5 billion.
Two years ago, Moonshot AI, now a major player in China's large language model (LLM) sector with a valuation reaching USD 50 billion, secured roughly similar pricing during its arduous search for Series B financing.
This company is called UniPat. Even many people within the AI industry are unfamiliar with it. In fact, it does not need to be household name. Information regarding its founder, Li Kuan, varies across sources. One report claimed: "Alibaba and Tencent invested in an intern." In reality, at Tongyi Qianwen, Li Kuan's work included model training and training material creation, and he was responsible for several related tasks.
Companies with similar levels of mystique are being intensely discussed within the industry. Intelligent Knowledge, Copula Lab, and Humanlaya have unique names and have rapidly appeared on the radars of various VCs in recent months. Reports indicate that some of these companies, which have aggressive business progress, are seeing their valuations swiftly approach USD 2 billion.
Just as fierce competition emerged when model factories first came into being, the race for their funding is also heating up. What they do seems largely unrelated to the "genius, algorithms, and architectural breakthroughs" that people imagine models still rely on. Based on previous understanding, they could even be considered a group of "outsourcing" companies—
that is, data companies.
This wave arrived in Silicon Valley about half a year later. In Silicon Valley, the numbers are already far more exaggerated. Mercor, the most well-known among them, raised USD 350 million in October last year with a USD 10 billion valuation. This August, it was reported to be negotiating new funding at a USD 20 billion valuation, with NVIDIA considering an equity stake. Scale AI, a more familiar name, had a valuation exceeding USD 29 billion when it accepted investment from Meta last year.
In external perceptions, algorithmic breakthroughs and genius researchers remain the accelerants driving AI models to chase the frontier. In reality, the daily routine of model training increasingly resembles a data factory. These "outsourcing" data companies play a decisive role in model competition, commercialization, and even the evolution of new concepts such as RSI (Recursive Self-Improvement). This has made today's AI landscape somewhat dominated by these "data nouveau riches."
This also means that understanding data companies, as well as the true state of the data industry behind large models, is the most critical entry point for comprehending today's AI evolution and what comes next.
1When Data Is Worth More Than Products
"Stop making products; come out and sell data."
A founder whose business progress generated substantial AI application behavioral data was discussing technical trends with friends from DeepSeek. He hoped to get advice on model development, but the conversation quickly turned into cooperation discussions. They directly suggested he "sell" his data.
Even though this application company was already profitable, he felt tempted when he saw the prices offered by model companies like DeepSeek for data.
According to multiple industry insiders: Some data containing user behavior trajectories can be quoted at CNY 1,000 to CNY 3,000 per record. Work scenario data including complete operational processes can sell for up to CNY 10,000. If 10,000 records are purchased at CNY 1,000 to CNY 3,000 each, the bill amounts to CNY10 million to CNY 30 million. For data companies, single orders reaching millions or tens of millions of CNY are now commonplace.
It should be noted that data suppliers were not always this lucrative in the model training industry.
For a long time, "data outsourcing" meant labeling images and text. Chirag Mahapatra, Vice President of Mercor, recalled to us: Tasks in the Mechanical Turk era included spam classification and marking stop signs in autonomous driving images. A diligent batch of workers could handle this. This is the "data labeling" model that many people still think of immediately today.
Later, people began selecting the better answer from several responses provided by the model, using these preferences to guide training. This is Reinforcement Learning from Human Feedback, abbreviated as RLHF.
Subsequently, Mercor hired PhDs from MIT and Harvard, as well as medalists from mathematical olympiads. They wrote solution steps and defined scoring criteria. Chirag told us: He saw clear changes on the demand side but did not expect the supply side to catch up so quickly. "These people are willing to take this job, earning tuition fees or extra income."
"I think part of this is due to the impact of the pandemic. These people had time and also needed to increase their income."
Models read massive amounts of text, code, and other materials during pre-training to learn language and knowledge. During post-training, demonstrations, feedback, and repeated practice are used to adjust behavior. The materials provided by these experts are primarily used in the latter stage. Many trade-offs made by lawyers reviewing contracts or financial analysts auditing accounts are not fully preserved on public web pages; they must be reorganized by people who have done this work.
Using such materials for training also makes it easier to see results relatively quickly.
An algorithm researcher explained to us: Model scales are already very large. Researching new algorithms and proving their effectiveness requires more effort, time, and money. Moreover, recent model releases from top Labs show that generational leaps like those seen previously are becoming increasingly difficult. Model improvements are beginning to stem from other areas.
More high-quality data from real-world scenarios, complex situations, and information not present on the internet, combined with post-training methods entering the model's capabilities, has become one of the most important ways to achieve progress.
This is also a direction discovered earlier by companies like Anthropic. These more advanced model teams identified bottlenecks sooner and acted sooner.
They broke down tasks by industry earlier. Last year, OpenAI conducted GDPval, a vocational capability evaluation, first selecting nine industries contributing most to US GDP, then drilling down to 44 occupations. From their actions, a clear logic emerges: Anthropic and OpenAI will push forward directions such as finance, law, cybersecurity, and science, ordered by economic value from high to low.
The method of advancement includes a format known as FDE (Forward Deployed Engineer). Anthropic enters enterprises as resident engineers, i.e., FDEs, sending personnel to streamline processes and deploy systems. Financial institutions do not easily open up their operations, so teams individually approached leading institutions. Some institutions worried that peers would use the technology first, leaving them behind, and thus agreed to try it early.
After finance, there are law, healthcare, and science. It is difficult for model companies to cover all specialties with internal teams alone. Mercor estimates: Frontier labs currently meet only 2% to 5% of their overall data needs. A significant portion of the work undertaken by data suppliers involves finding people and tasks across various industries, while model companies retain internal quality control gates.
The outsourcing flavor remains strong, but unit prices have risen significantly.
The same trend occurred in China, roughly half a year to a year later. But once understood, the actions of various companies were rapid.
Some model training teams told us: In certain vertical domains, 10,000 to 100,000 sufficiently good data records are enough to train a model to a quite high level.
Moreover, today's model companies have basically mastered the methods of integrating capability improvements brought by data from different domains into the same model. Different domains can undergo separate post-training, and then capabilities are merged into the final version. Some in the industry call this "version merging." For example, one public route taken by DeepSeek involves teacher models from multiple domains teaching capabilities to the same student model, which shares similarities with the logic of distillation.
A researcher involved in post-training and evaluation described a multimodal team he had contacted. The team decomposed the model into over twenty basic capabilities and eighty to ninety vertical scenarios, drawing a "capability tree." Researchers added data along nodes, adjusted ratios, and ran experiments to check which capabilities had emerged and which remained unchanged.
According to informed sources, 90% of DeepSeek's researchers are engaged in data-related work. "Actually, very few people are researching algorithms." These model manufacturers internally arrange rhythms like factory workers. When a group is waiting for their GPUs and has nothing else to do, they can all participate in data labeling.
Public information shows that ByteDance has integrated multiple procurement and quality inspection teams into a first-level department called "AI Data & Security." Tencent also appointed Yao Shunyu, who proposed that data will be the most important competitive direction in the second half of AI, to concurrently manage the AI Data Department. After a series of turmoil and adjustments in Alibaba's model team, the话语权 (discourse power) over data-related businesses completed a "decentralization," with internal responsibilities handed over to the most promising researchers.
These teams' data demands cannot obviously all be solved internally during this fiercely competitive period. External data companies have become the most critical partners. One data supplier mentioned that DeepSeek has set up dedicated data procurement positions. Kimi rarely procures externally, while MiniMax recently started building procurement pipeline processes. ByteDance, Tencent, and Alibaba increased external procurement this year.
Intelligent Knowledge, founded earlier, recalls that during the startup phase, they had to explain this business to investors everywhere. But after the Silicon Valley template became more visible, the situation changed. Data has become the most important investment direction this year.
Some practitioners revealed that founders of some data companies may not even reply to investors' messages anymore. Given the current macro environment of the AI industry, where one can both participate in determining the direction of intelligence evolution, playing a key role in this main line, and simultaneously run through the business model to "make big money" due to sudden changes in supply and demand, where else can such targets be found? Why would they necessarily need so much financing?
In a sea of gloom in the AI venture capital circle, these data nouveau riches naturally have the confidence to be arrogant.
2Good Data Companies Must Be Research-Oriented
What exactly are model companies buying with such huge sums?
In essence, this data is reinforcement learning data. Broken down, it often consists of tasks, environments, and verifiers.
Suppliers must deliver something the model can attempt to complete, a system for it to operate within, and standards to judge how well it performs. The model reads information and calls tools within the environment, leaving operational traces. The verifier scores the output, and training utilizes the feedback to adjust the model.
These tasks vary in difficulty.
Simple ones, such as having the model enter several invoices into a reimbursement system, can be judged by matching amounts and dates. Complex ones, such as a boss leaving a note saying "Want to go to Shanghai next week to meet someone, help me arrange it," require the model to browse chat history, calendars, and emails to figure out who to meet and when, before making arrangements.
To enable the model to complete this itinerary, suppliers must prepare the "company's" email, calendar, and ticket booking systems. The model must find meeting targets from communications and check time conflicts from the calendar. After sending an email, the system must reflect corresponding changes. This operating space, isolated from real business and allowing for repeated trial and error, is typically called a sandbox.
Sent emails and modified records must be restored before the next round of practice, and multiple batches of tasks running simultaneously must not cross-contaminate data. It is understood that last year, some teams encountered frequent underlying system failures during evaluations, preventing models from completing tasks; high concurrency could even crash the environment. This year, such issues have decreased significantly.
Business materials built into the systems are also directly purchased by some. For instance, internal materials from companies that have gone bankrupt, liquidated, or frequently pivoted business models—there are no shortage of such AI startups to find today.
However, data suppliers' purchasing standards impose higher requirements. For example, when Humanlaya seeks materials from bankrupt companies, it requires that the liquidated enterprise had approximately 100 employees at its peak. Emails, chats, and business records provide clues about how the company operated in the past. Suppliers must also rebuild the relationships between ticketing systems, internal communications, and records in simulated environments, allowing the model to operate within them.
The ability to provide consequences based on the model's operations further widens the gap between suppliers. Many companies in the market still deliver static materials, where traces are walked by humans and every step is pre-calibrated. If the model tries a different operation path, the environment may not provide the subsequent state and feedback. If the model attempts different routes itself, suppliers must also generate these variations.
Next, even if the environment restoration is highly realistic, it lacks part of the feedback from real business. Resident engineers can continue observing how enterprises use systems and whether budgets are spent more effectively. Frozen training environments cannot continue to obtain this feedback from real business. Historical records can show what employees did at the time, but it is difficult to fully explain why they made those trade-offs. This part requires finding people to supplement.
This is the scoring phase. For example, an email showing "sent successfully" is not enough; the model might send it to the wrong person or miss attachments. Invoice amounts can be verified, but whether an itinerary is arranged well also involves the experience of the boss and the meeting target. Here, data companies need to build expert networks, allowing these experts to break requirements into rubrics, i.e., itemized scoring conditions. According to our understanding, some tasks can be broken down into forty to fifty items, or even over a hundred. These naturally correspond to price increases.
Experts capable of breaking down these problems clearly are hard to find. Organizing experts and producing materials also requires a system.
For example, "Intelligent Knowledge" uses "Yimian Qianshi" to find and organize experts. Copula Lab, spun out of MiniMax, is developing Chinese financial long-process tasks. It similarly requires experts to provide reference materials, office document delivery requirements, and scoring rules. Founder Liang Li previously oversaw Agent business, and technical co-founder Cai Jiaren worked on open source and post-training.
Mercor spends great effort finding such people. Entrepreneurs in Silicon Valley know how "involved" (competitive/intense) Mercor is. Often, they and their competitors receive sudden demands. They are almost never just working 996 schedules; when the expert network is a new field, they sometimes live directly at the client's location.
Additionally, for some difficult-to-secure experts, they use various methods. Rumors include side agreements, though this cannot be confirmed. In our exchange, Chirag admitted that building expert networks has become very difficult today. It is no longer sufficient to just give them an engineer skilled in Python programming; they must also see how this person completes integrations in customer management systems and makes different tools work together. Tasks and scoring must be derived from this work. This directly impacts orders and unit revenue.
It is understood that the net profit retained from each order is calculated starting from production costs. When quoting for such data, the typical value of an expert doing similar work and the time invested are important bases. Delivery may also include organizing workflows, writing scoring guides, building environments, and debugging. Buyers pay according to agreed tasks or projects, and suppliers then compensate experts and production teams.
Upfront costs may also be amortized across multiple orders. Chirag told us: Some datasets are jointly designed by labs and suppliers for exclusive use by one party. Others can be delivered to multiple parties as agreed, covering the invested expert time, software, and review costs across different orders.
Profitable templates have emerged in this business. When Forbes interviewed Surge AI founder Edwin Chen in 2025, it disclosed that the company's revenue in 2024 was approximately USD 1.2 billion. According to Chen, the company was profitable almost from day one. It organizes experts and part-time staff to provide questions, answer evaluations, and scoring requirements for model companies.
But to continue selling to the same customer, suppliers must constantly refresh content.
It is understood that when model factories buy 100,000 data records, they transform the formats upon return, potentially expanding them to 300,000 records. Already delivered materials are difficult to sell again. Once a branch of law is delivered, the next batch must seek another branch.
However, for these data companies, they often pursue "research" attributes more strongly. Or at least present themselves as such. Tasks, environments, and scoring have been incorporated into these companies' public products and research directions.
For example, UniPat's ExpertEval compiles judgments from medical, financial, and legal experts into 3,213 cases, averaging about 22 scoring requirements per question. SaaS-Bench allows models to enter 23 sets of enterprise software to complete 106 work tasks. This is already somewhat close to the working style of post-training teams in large language model laboratories.
Mercor also officially announced the establishment of its own research department recently. Their explanation is: AI is reshaping almost all work, but existing evaluation and data methods cannot keep up with this change. And they are a company that can cover the complete closed loop of "Evaluation → Data → Post-training" in its research agenda. Mercor has recruited several core researchers from companies like OpenAI to join this department.
Although Mercor denies the possibility of becoming a model company, many investors believe in our exchanges that these data companies will eventually face this choice. Their actions today are more honest than their words. Who doesn't want to become a research company, or even the birthplace of the next paradigm of models?
And many Chinese data companies have founders closely tied to their former model employers. Many model manufacturers also became investors and shareholders immediately. These intricate connections bring variables.
For example, UniPat's Li Kuan previously did post-training at Tongyi and was one of the project leads for the open-source search agent WebSailor. We learned earlier that UniPat's cooperation with Alibaba focuses mainly on reinforcement learning post-training data. Besides funds, Alibaba provides computing power, with annual computing costs in the range of tens of millions of CNY. Similarly, NVIDIA purchases data from Mercor and is also negotiating an equity stake.
"Everyone understands that such opportunities are greater, whether in terms of valuation or team exit strategies," said an investor who focused on and bet on data companies early on. In his view, many institutions investing in data companies today hope for exit channels where these data companies are acquired by model companies, ultimately becoming part of or joining model teams.
The story of Meta and Scale stands before us. Today, the shadow of exits looms. Data companies becoming part of the model vendors they serve obviously provides another imaginative exit direction.
3Involution and Reshuffling
In our exchanges with upstream and downstream players in the data industry, we clearly feel that everyone is rapidly in a period of flux. Previously, because everything was fresh, model manufacturers' demands were urgent and vigorous. The quality of data submitted by many data companies was initially unconfirmed by model companies.
In the past period, the same model company bought large amounts of data from multiple data companies. This data entered the models, improving model capabilities. This formed a perfect loop when everyone was eager to compete and catch up with each other.
"As long as you have data, give it to me, and I'll look at it." A member of a major tech giant's model team once told us this. Part of the procurement was filling immediate gaps, while another part was stocking up for future experiments.
A researcher told us that for popular data, giants like ByteDance are willing to pay several times the price in exchange for exclusive usage rights for half a year to a year, avoiding data leakage into the market. Public reports also state that ByteDance sometimes phases in "buyouts" of suppliers' core personnel.
In these collaborations between data companies and model manufacturers, relationship factors are also strong. News about what data is needed for the next batch flows first within small circles of procurement and suppliers. When demands for categories like healthcare, 3D, or real-time interactive voice emerge, those who hear first go looking for experts. Directions sometimes change every month, and suppliers must constantly send samples to test against clients' standards.
Therefore, suppliers are also desperately expanding production. A "data nouveau riche" recently chased by investors has designed a new method: offering employees low base salaries with high commissions. For every data record produced, they earn more money.
Everything feels less like AI and more like the practices of outsourcing companies.
"I don't know." When asked how the delivered data proves effective, the head of one data company answered this way.
Asked further how model factories judge it, he said, "Also don't know." Another company received contradictory feedback: the same batch of data was deemed good by one model factory and bad by another.
In fact, even publicly available data subjected to layer-by-layer review contains errors. SWE-bench Verified was once one of the most commonly cited coding tests when companies released models. Questions were filtered by professional engineers item by item, with each question independently reviewed by three people. This February, OpenAI announced it would no longer use it to report scores. It spot-checked 138 questions that o3 failed to stabilize after 64 runs. 59.4% had substantive issues in test design or question descriptions. Over one-third of the questions had tests written too narrowly, judging correct solutions as wrong. The frontier models tested by various companies could also reproduce standard answers or details of some questions, having likely seen them during training.
This ambiguity amidst high growth is being questioned.
As model progress continues to slow, more and more model manufacturers are realizing that this set of methods cannot verify exactly which data played a key role and which data might be substandard.
A technical person from a data company told us that the industry has reached a stage where quality is more important than diversity. In his view, model companies must strictly guard the final gate of quality inside, otherwise there will inevitably be widespread opportunistic behavior outside.
For example, to make it seem like the model was stumped, some data companies forcibly combine several irrelevant questions into long ones, relying on lengthening operation chains to lower success rates. It is understood that models trained on such materials can see fast score increases, but actual usage feels awkward. Some models endlessly try themselves in tasks. Real users rarely work according to the sequence of those questions.
Leaderboards provide a shorter customer acquisition path. Industry insiders claim that when a company releases an evaluation stating a model performs well and attracts attention, other model companies will visit to buy related materials. A set of materials can be sold to multiple buyers as agreed. Evaluation results allow chasers to see the gap and bring orders. A researcher complained that some teams treat leaderboard manipulation itself as the goal, directly procuring training question banks from the question setters.
A group of smaller-scale data merchants has grown around these leaderboards. They specialize in creating similar, harder new questions to sell to model factories wanting to improve scores, comparable to Olympiad tutoring classes. Buy questions, train, re-test; scores go up, but the workload in users' hands may not necessarily become easier.
Even the names of experts may be easier to sell than what the experts actually did. We learned that "professional data" from more than one company is synthesized internally and then shown to experts for endorsement. "The expert might not change a single line," said a practitioner.
Companies further removed from production may be selling customer relationships. Some can secure orders but lack production capacity; others can organize production but cannot contact model factories. Orders are handed to partners, then broken down to the next layer. At the bottom, the workers may not know who the final client is or what capabilities the materials need to supplement, but the price difference is retained at each layer.
Facing this chaos, model manufacturers have begun thinking of new methods.
This is first reflected in tightened procurement. Both ByteDance and Tencent are consolidating procurement authority and starting to require ROI calculations. Internally, methods are being established to clarify why a batch of materials is useful or useless.
Furthermore, model clients are beginning to trace back into the production process. Experts still use models to write code, generate tests, or format documents, but buyers are starting to request finer records: which experts provided judgments, where disagreements lay during discussions, and how consensus was finally reached.
More and more model companies, due to effects falling short of expectations, are starting to ablate suppliers' data one by one, meaning separately removing or adding batches of materials to observe how effects change. They spent a long time investigating.
Industry insiders revealed that some results are worrying. Recently, ugly situations have emerged. After discovering how chaotic the data provided by some data companies was, model manufacturers chose to blacklist data providers who treated them as suckers or "cutting leeks" (exploiting them).
A technical person from a data company said that in this industry, the number of clients (Party A) can be counted on one hand. Losing one makes it hard to replace. These companies selling data amid anxiety clearly do not intend to last long.
Data companies are also encountering obstacles with their clients. Industry insiders claim that model companies send personnel to provide enterprise services, accessing internal business materials. After desensitization, the materials return to the training environment. Enterprises paid service fees, but their contributed data was not sold separately for money. This insider judges that enterprises will soon realize the issue. In his view, the experts and companies providing this data are, to some extent, helping models "kill" themselves.
Additionally, there is a looming risk hanging overhead: the privacy of this data.
This May, Spirit Airlines, the US budget carrier that ceased operations, had its internal data auctioned off by Google for USD 10 million in August. This included approximately 100 million emails and 500 million Teams messages. The second-highest bid came from Mercor at USD 7.5 million. Flight attendant unions immediately objected, arguing that removing names does not protect confidential content left by employees at work. The approval hearing scheduled for August 19 was postponed.
In late March, hackers breached Mercor's system via a malicious version of the open-source tool LiteLLM. According to lawsuits and hacker group claims, approximately 4TB of data leaked, including about 3TB of video interview recordings and identity verification materials, involving over 40,000 contractors and clients. By mid-April, Mercor faced at least six class-action lawsuits, and Meta indefinitely suspended its cooperation with them.
Under rising financings, chaotic standards, and wide arbitrage spaces in pricing, more gray areas are being invented. An industry insider relayed to us that some data suppliers pay employees of major tech companies to recreate internal data externally, then deliver it to outside clients.
Other industry insiders revealed that after competition heated up, some data companies needed to present better-looking financial statements to investors. Thus, rumors circulated about a company generating USD 30 million in revenue within six months of establishment. Reportedly, part of this was earned by betting on Polymarket using their own predictive models, which was also counted into total revenue.
There are also new "designs" made directly on equity. Suppliers promise small shares or benefits to those who can influence procurement. After securing orders, completing financing, and seeing valuation rise, they cash out through equity transfers. Procurement supports revenue, revenue supports valuation, valuation rises, and those influencing the procurement benefit.
These practices will not stop here; they will only be exposed to more people. "Ultimately, we must return to commercial value," said a data company founder. He believes trust is ultimately the most important factor.
"Otherwise, today's prosperity may eventually turn into an illusion."