AI data companies are fierce: UniPat reportedly backed by Alibaba and Tencent, valued at $2.5 billion

In September of this year, a company less than a year old was reported to be in talks for a funding round of about $300 million, with the legendary lead investor said to be Alibaba, with Tencent and Sequoia China participating, at a post-money valuation of $2.5 billion.

Two years ago, Moonshot AI—today a major player in China's large language models and already valued at as much as $50 billion—managed to secure only roughly the same price when it was laboring to raise its Series B.

This company is called UniPat, and even many people in the AI industry find it unfamiliar. In fact, it doesn't need to be a household name. As for its founder Li Kuan, accounts of his background also differ. One report said: Alibaba and Tencent invested in an intern. In fact, at Tongyi Qianwen, Li Kuan's work included model training and the production of training materials, and he was the person in charge of many related efforts.

Companies with a similar air of mystery are being discussed intensively within the industry—Intelligent Knowledge, Copula Lab, Humanlaya—their names are all distinctive, and all have rapidly appeared on various VCs' desks in recent months. Reports say that some of them, whose business progress is relatively aggressive, are seeing valuations rapidly approach $2 billion.

Just like the fierce competition when model companies first emerged, the pursuit of their financing is also entering a white-hot phase. And what they do seems to have little to do with the "genius, algorithms, architectural breakthroughs" that people imagine models still rely on. By previous understanding, they could even be counted as a group of "outsourcing" companies—

That is, data companies.

This wave, too, came over from Silicon Valley about half a year later. In Silicon Valley, the numbers are already even more exaggerated. The best-known, Mercor, raised $350 million last October at a $10 billion valuation, and in August of this year was reported to be in talks for a new round at a $20 billion valuation, with Nvidia considering taking a stake. The more familiar Scale AI had already surpassed a $29 billion valuation when it accepted Meta's investment last year.

In outsiders' understanding, algorithmic breakthroughs and genius researchers are still the accelerant driving AI models toward the frontier. In fact, however, the daily routine of model training is looking more and more like 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, which has made today's AI, to some extent, the domain of these "data upstarts."

When data is worth more than products

"Stop making products and come out and sell data."

A founder whose business progress has generated a large amount of AI application behavior data was discussing technology trends with friends at DeepSeek, hoping to get some advice on model development, but it quickly turned into a kind of partnership discussion, and the other party directly suggested he "sell" the data.

Even though this application company is already profitable on its own, when he saw the prices that model companies such as DeepSeek were offering for data, he was still a little "tempted."

According to several people familiar with the industry, some data with user behavior traces can be quoted at 1,000 to 3,000 yuan per item, and work-scenario data with a complete operation process can even sell for about 10,000 yuan. If you purchase 10,000 items at 1,000 to 3,000 yuan each, the bill is 10 million to 30 million yuan. For data companies, single orders of millions or tens of millions of yuan are already the norm.

You should know that data suppliers haven't been this sought-after in the model training industry from day one.

For a long time, "data outsourcing" meant labeling images and text. Mercor Vice President Chirag Mahapatra recalled to us: tasks in the Mechanical Turk era included spam classification and marking stop signs in autonomous-driving images, and a group of diligent people could do them. This is the "data labeling" model that many people still think of first today.

Later, people began choosing a better one from several answers given by a model, using these preferences to guide training—that is, reinforcement learning from human feedback, or RLHF.

After that, the people Mercor brought in were already PhDs from MIT and Harvard and medalists in math olympiads, who wrote out solution steps and set scoring standards. Chirag told us: he saw a clear change on the demand side, but did not expect the supply side to keep up so quickly. "These people are willing to do this work to earn tuition or extra income."

"I think part of it is the impact of the pandemic—these people had the time and also the need to increase their income."

In pretraining, models read massive amounts of text, code and other materials to learn language and knowledge; then in post-training, demonstrations, feedback and repeated practice are used to adjust behavior. The materials provided by these experts are mainly used in the latter stage. Many trade-offs made by lawyers when revising contracts and by financial analysts when auditing accounts are not fully preserved on public web pages, so people who have done this work must be found to reorganize them.

Training with such materials also makes it easier to see the results of the investment relatively quickly.

An algorithm researcher explained to us: models are already very large, and researching new algorithms and proving they work requires more energy, time and money. Moreover, from the recent model releases of top labs, one can see that generational leaps like those of the past are increasingly hard to achieve, and model progress is beginning to come more from other places.

More high-quality data from real scenarios, complex and never before seen on the internet, combined with post-training methods, entering the model's capabilities, has become one of the most important ways to make progress.

This is likewise a direction that Anthropic and others spotted first. These more frontier model teams saw the bottleneck earlier and acted earlier.

They broke tasks down by industry earlier: last year OpenAI built GDPval, an evaluation of occupational capabilities, first selecting the nine industries that contribute most to US GDP, then narrowing down to 44 occupations. A clear logic can be seen in their moves: Anthropic and OpenAI advance sector by sector—finance, law, cybersecurity, science and so on—in order of economic value from high to low.

The way they advance includes a form called FDE. Anthropic enters enterprises as resident engineers, that is, FDEs, sending people to map out processes and deploy systems. Financial institutions do not readily open up their business, so the team approaches leading institutions one by one. Some institutions worry that peers will adopt it first and leave them behind, so they are willing to try it first.

After finance come law, healthcare and science, and it is hard for model companies to cover every specialty with in-house teams. Mercor estimates that, of the data frontier labs need overall, only 2% to 5% has been completed so far. A large part of what data vendors do is find people and tasks across industries, while model companies keep the internal quality gate.

The outsourcing flavor is still strong, but the unit price has risen far too much.

The same trend happened in China, roughly half a year to a year later. But once they figured it out, everyone moved fast.

A model training team told us: in some vertical domains, 10,000 to 100,000 sufficiently good data points are enough to train a model to a fairly good level.

Moreover, today's model companies have basically mastered the method of "integrating" the capability gains brought by data from different domains into the same model. Different domains can be post-trained separately, and then the capabilities merged into the final version; some in the industry call this "merging versions." For example, one public route at DeepSeek is to have teacher models from multiple domains teach capabilities to the same student model, which in a sense is also very similar to the logic of distillation.

A researcher involved in post-training and evaluation described to us an omni-modal team he had encountered. The team split the model into more than twenty basic capabilities and eighty to ninety vertical scenarios, drawing them as a "capability tree." Researchers fill in data along the nodes, adjust ratios, run experiments, and check which capability has grown and which has not changed.

According to people familiar with the matter, 90% of DeepSeek's researchers are doing data-related work; "actually very few people are researching algorithms." These model vendors also schedule the pace internally like factory workers: when a group is waiting for its cards and has nothing to do, everyone can go participate in labeling data.

Public information shows that ByteDance has already integrated multiple procurement and quality inspection teams into a first-level department, "AI Data and Security"; Tencent has also put Yao Shunyu, who proposed that data will be the most important competitive direction in the second half of AI, in charge of the AI Data department; and after a series of turmoil and adjustments in Alibaba's model team, the say over data-related business has also completed a "separation of powers," with internal control handed to the most promising researchers.

And at this fiercely competitive moment, these teams' data needs obviously cannot all be solved in-house. External data companies have become the most critical partners. A data vendor mentioned that DeepSeek has set up dedicated data procurement positions, while Kimi rarely buys externally, and MiniMax only recently began building procurement pipeline processes; ByteDance, Tencent and Alibaba have increased external procurement this year.

Zhizhi Zhishi Huiyi, founded relatively early, still had to explain this business to investors everywhere in its early days. But later, after Silicon Valley's model was seen more widely, the situation changed. Data has already become this year's most important investment direction.

Some practitioners even revealed that some data company founders now may not even reply to investors' messages. After all, in today's AI industry environment, where else can you find a target that can both participate in deciding the direction of intelligence evolution and play a key role on this main line, while also, because of the sudden change in supply and demand, being able to get a business model working and "make big money" right out of the gate—so why would they necessarily need that much financing?

A good data company must be a research-driven company

Having spent so much money, what exactly are model companies buying?

In essence, this data is all reinforcement learning data. Broken down, it often needs to consist of a task, an environment and a verifier.

What the vendor must deliver is something the model can attempt to complete, a system for it to operate, and a standard for judging how well it did. The model reads information and calls tools in the environment, leaving a trajectory of operations; the verifier scores it, and training then uses the feedback to adjust the model.

These tasks also differ in difficulty.

Simple ones, for example, have the model enter several invoices into an expense reimbursement system; if the amounts and dates match, it can be judged whether it did well. Complex ones, for example, have the boss leave only one sentence: "I want to go to Shanghai next week to meet someone, help me arrange it." The model has to go through chat records, calendars and emails, first figure out who to meet and when, and then make arrangements.

For the model to complete this trip, the vendor has to prepare the "company's" email, calendar and ticketing systems. The model must be able to find the meeting counterpart from communications and check time conflicts from the calendar; after an email is sent, the system must also change accordingly. This kind of operating space, isolated from real business and allowing repeated trial and error, is usually called a sandbox.

Emails that were sent and records that were changed must be restored before the next round of practice, and multiple batches of tasks running at the same time must not leak data into one another. It is understood that last year, when some teams ran evaluations, models often failed to complete tasks due to underlying system failures, and high concurrency could even crash the environment; this year, such problems have become much less common.

But data vendors' purchasing standards have raised higher requirements: for example, when Humanlaya was looking for information on bankrupt companies, it required that the bankrupt and liquidated companies whose data was acquired had about 100 people at their peak. Emails, chats and business records can leave clues about how a company operated in the past. Vendors also have to rebuild the relationships among the ticketing system, internal communications and records in a simulated environment, so that the model can go in and operate them.

Whether consequences can be given as the model operates also widens the gap between vendors. Many companies on the market still deliver static materials, with trajectories walked through by humans and every step pre-labeled. But if the model operates in a different way, the environment may not be able to give the next state and feedback. The model tries different routes on its own, and vendors have to produce these changes as well.

Next, no matter how closely the environment is reproduced, it still lacks part of the feedback from real business. On-site engineers can continue to observe how enterprises use the system and whether budgets are spent better; but a frozen training environment cannot continue to obtain this feedback from real business. Historical records can show what employees did at the time, but it is hard to fully explain why they made that trade-off. This part still needs people to fill in.

This is the scoring stage. For example, an email showing "sent successfully" is still not enough; the model may have sent it to the wrong person or missed an attachment. Invoice amounts can be checked, and whether a travel arrangement is good also involves the experience of the boss and the person being met. Here data companies need to build an expert network and have these experts break requirements down into rubrics, that is, item-by-item scoring conditions. According to our understanding, some tasks can be broken down into forty or fifty items, or even hundreds. These of course also correspond to an increase in price.

Experts who can break these problems down clearly are not easy to find. Organizing experts and producing materials also requires a system.

For example, "Intelligent Knowledge" uses "One Face, a Thousand Recognitions" to find and organize experts, and uses Xpert Studio to configure tasks, distribute work, and then do review and quality inspection. Copula Lab, which came out of MiniMax, is developing long-process Chinese financial tasks and likewise needs experts to provide reference materials, office document delivery requirements and scoring rules. Founder Liang Li previously was in charge of the Agent business, and technical co-founder Cai Jiaren has worked on open source and post-training.

Mercor spends a great deal of effort finding such people. Entrepreneurs in Silicon Valley all know how "intense" Mercor is; often the demands they and their competitors receive are very sudden, and they are almost not just working 996 across the board—when the expert network is a new field, they sometimes simply live at the client's location.

And, for some hard-to-win experts, they use various methods, which in some rumors include drawer agreements and the like, but this cannot yet be confirmed. In our conversation, Chirag admitted that building an expert network has become very difficult today; just giving it an engineer who is very good at writing programs in Python is no longer enough—it also has to see how this person completes integration in the customer management system and gets different tools to work together. Tasks and scoring must both be derived from this work. And this directly concerns orders and per-unit revenue.

It is understood that how much money each order can leave starts from the production cost. When quoting prices for this kind of data, how much experts are usually worth for similar work and how much time they invest are important bases. Delivery may also include organizing workflows, writing scoring, building environments and troubleshooting. Buyers pay according to agreed tasks or projects, and vendors then pay experts and the production team.

Upfront costs may also be shared across multiple orders. Chirag told us: some datasets are designed jointly by labs and vendors and used by only one party; others can be delivered to multiple parties as agreed, so the expert time, software and review costs invested can be covered by different orders.

This business has already produced profitable examples. When Forbes interviewed Surge AI founder Edwin Chen in 2025, it disclosed that the company's 2024 revenue was about $1.2 billion; according to Chen, the company was profitable almost from day one. It organizes experts and part-time workers to provide model companies with questions, answer evaluations and scoring requirements.

But to keep selling to the same customer, vendors have to constantly change to new content.

It is understood that when model factories buy 100,000 pieces of data and transform the format back home, they may expand it to 300,000 pieces. Materials already delivered are hard to sell to them again. Once one branch of law has been delivered, the next batch has to find another branch.

But for these data companies, they often pursue the attribute of "research" even more. Or at least present themselves that way. Tasks, environments and scoring have already been built into these companies' public products and research directions.

For example, UniPat's ExpertEval organizes the judgments of medical, financial and legal experts into 3,213 cases, with an average of about 22 scoring requirements per question; SaaS-Bench, meanwhile, puts models into 23 sets of enterprise software to complete 106 work tasks. This is already somewhat close to the way large language model labs' post-training teams work.

Mercor also recently officially announced the establishment of its own research department, and their explanation was: AI is reshaping almost all work, and 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" with its research agenda. Mercor has already recruited multiple core researchers from companies such as OpenAI to join this department.

Although Mercor denies the idea that it can become a model company, many investors have told us in conversations that these data companies will ultimately have to face this choice, and their actions today are more honest than their words. Who wouldn't want to become a research company, or even the birthplace of the next-generation paradigm of models?

And for many data companies in China, the founders were already close to their former model employers, and many model vendors also became investors and shareholders right away. These intricate connections also bring variables.

For example, UniPat's Li Kuan previously did post-training at Tongyi and was also one of the project leads for the open-source search agent WebSailor. We previously learned that UniPat's cooperation with Alibaba is mainly focused on reinforcement learning post-training data. In addition to funding, Alibaba also provides computing power, with annual computing power costs in the range of tens of millions of RMB. Similarly, Nvidia purchases data from Mercor and is also in talks to take a stake.

"Everyone understands that the opportunities brought by that are greater, whether in valuation or in the team's exit, they are all different," said an investor who paid attention to and bet on data companies early on. In his view, many institutions investing in data companies today also expect the exit channel to be the hope that these data companies can be acquired by model companies and ultimately become or join model teams.

The story of Meta and Scale is right in front of us, and today the shadow of exit hangs over this. Data companies becoming part of one of the model vendors they serve obviously provides yet another imaginative exit direction.

Involution and Reshuffling

In our conversations with the upstream and downstream of the data industry, we clearly felt that at this moment everyone is quickly in another period of change. For a while before this, because all of this was very new, model vendors' demand was urgent and vigorous, and model companies were once unable to confirm the quality of the data delivered by many data companies.

For a period in the past, the same model company bought large amounts of data from multiple data companies, and this data entered the models, improving model capabilities. This formed a perfect closed 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 take a look." A person from a major tech company's model team once told us this. Some procurement was filling immediate gaps, while another part was stocking up for later experiments.

A researcher told us that when encountering popular data, major companies such as ByteDance are also willing to pay several times the price in exchange for exclusive use arrangements of six months to a year, to prevent the data from flowing into the market. There are also public reports saying that ByteDance sometimes also "buys out" suppliers' core personnel for a period of time.

In the cooperation between these data companies and model vendors, there is also a strong relationship factor. What the next batch of data needs, the news will first circulate in the small circles of procurement and suppliers. Which type of demand begins to appear—medical, 3D, real-time interactive voice—those who hear it first go first to find experts. The direction sometimes changes once a month, and suppliers also have to continuously send samples to meet customers' standards.

So suppliers are also desperately expanding production. A "data upstart" that investors have recently been chasing has also designed a new method: giving employees a low base salary and high commissions, so that for every piece of data produced, they earn one more share of the money.

"I don't know." This was the answer from the head of a data company when asked how to prove the delivered data is effective.

When asked again how model factories judge, he said "I don't know either." Another company received the opposite feedback: for the same batch of data, one model factory said it was good, and another said it was not.

In fact, even public data that has been reviewed layer by layer can also be wrong. SWE-bench Verified was once one of the programming tests most commonly cited when various companies released models. The questions were screened one by one by professional engineers, and each question had three independent reviewers. In February this year, OpenAI announced it would no longer use it to report results. It sampled 138 questions that o3 could not do stably after 64 runs, and 59.4% had substantive problems in test design or question description. More than one third of the questions were written too narrowly in the test, which would judge correct solutions as wrong; the frontier models tested from various companies could also reproduce the standard answers or question details of some questions, having already seen them during training.

This ambiguity under high growth is being questioned.

As model progress continues to slow, more and more model vendors are beginning to realize that this set of methods cannot verify which data played a key role, and which data may even be padding.

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 hold the final line on quality internally, otherwise there will certainly be a lot of opportunistic behavior outside.

For example, in order to make it look like the model is being tested, some data companies forcibly splice several unrelated questions into long questions, relying on lengthening the operation chain to lower the success rate. It is understood that models trained on this kind of material can see their scores rise very quickly, but in actual use they feel awkward, and some will try endlessly on their own within tasks. Real users rarely work in the order of those questions.

The leaderboards also offer a shorter path to customer acquisition. Industry insiders say that when one company publishes an evaluation showing that a certain model performs well and attracts attention, other model companies will come knocking to buy the relevant materials. A single set of materials can, by agreement, be sold to multiple buyers; the evaluation results let laggards see the gap, and also bring in orders. One researcher complained that some teams treat gaming the leaderboard as the goal itself, going straight to the question-setters to buy training question banks.

Around these leaderboards, a group of smaller data vendors has also grown up. They specialize in producing new questions of the same kind but harder, and sell them to model makers that want to raise their scores—comparable to Olympiad tutoring classes. Buy questions, train, test again—the scores go up, but the work in users' hands does not necessarily get any easier.

Even an expert's name may be easier to sell than what the expert has actually done. We have learned that the "professional data" of more than one company is synthesized internally and then handed to an expert for a quick look as an endorsement. "The expert may not change a single line," one practitioner said.

Companies further from production may instead be selling customer relationships. Some people can get orders but have no production capacity; others can organize production but have no access to model makers. Orders are handed to partners and then split down to the next layer. At the very bottom, the people doing the work may not know who the end customer is, or what capabilities the materials are meant to supplement, yet the price spread remains at every layer.

Faced with this chaos, model makers have begun to think of new approaches.

This is first reflected in tighter procurement: both ByteDance and Tencent are centralizing procurement authority and have begun requiring ROI accounting. Internally, they are starting to establish methods to figure out why a batch of materials is useful or useless.

In addition, model customers have begun to trace back into the production process. Experts still use models to write code, generate tests, or tidy up formats, but buyers have started to demand more detailed records: which experts provided judgments, where disagreements arose during discussions, and how consensus was finally reached.

More and more model companies, because results fall short of expectations, have begun to take each supplier's data and run ablations—removing or adding a batch of materials separately to see how the results change. It took a long time to investigate.

Industry insiders reveal that some results are worrying, and ugly situations have recently begun to emerge: after discovering how chaotic the data provided by some data companies is, model makers have chosen to blacklist some data providers that treated them as suckers and "harvested" them.

A technical person at a data company said that in this industry, the clients can be counted on one's fingers, and losing even one is hard to make up for. These companies selling data on the wave of anxiety clearly do not intend to last long.

And data companies have also begun to encounter obstacles with their customers. Industry insiders say model companies send people to do enterprise services and access internal business materials; after de-identification, the materials return to the training environment. The enterprise pays a service fee, yet the data it contributed itself does not earn it a separate payment. This person judges that enterprises will soon realize the problem. In his view, the experts and companies providing this data are, to some extent, helping models "kill" themselves.

In addition, there is another looming risk: the privacy of this data.

U.S. budget airline Spirit, which was grounded in May this year, had its internal data auctioned off by Google in August for $10 million, including about 100 million emails and 500 million Teams messages; the second-highest bidder was Mercor, at $7.5 million. The flight attendants' union immediately objected, arguing that removing names does not protect the confidential content employees left behind in their work, and the approval hearing originally scheduled for August 19 was postponed.

At the end of March, hackers breached Mercor's systems through a malicious version of the open-source tool LiteLLM. According to the complaint and the hacker group's account, about 4TB of data leaked, of which about 3TB was video interview recordings and identity verification materials, involving more than 40,000 contract workers and customers. By mid-April, Mercor faced at least six class-action lawsuits, and Meta indefinitely suspended cooperation with it.

Amid ever-rising financing, chaotic standards, and pricing with ample room for arbitrage, more gray areas are being invented. One industry insider relayed to us that some data vendors pay employees of big companies to redo a copy of the same data as the internal version outside the company, and then deliver it externally.

Other industry insiders reveal that as competition became white-hot, some data companies needed better-looking financial statements for investors, so word spread of a company that booked $30 million in revenue within six months of founding—reportedly part of it earned by using its own prediction models to bet on Polymarket, which was also counted in total revenue.

There are also new "designs" done directly on equity: suppliers promise a small amount of shares or profits to people who can influence procurement, and once orders are won, financing is completed, and the valuation rises, they cash out through equity transfers. Procurement props up revenue, revenue props up valuation, and the rising valuation in turn benefits the people who influenced that procurement.

These practices will not stop; they will only be put in front of more people more and more often. "Ultimately it still has to come back to business value," said the founder of one data company. He believes that in the end, trust is still the most important thing.

"Otherwise, today's prosperity may ultimately turn into an illusion."

English text is being prepared.