11/17 2026
354
Today, Alibaba released Qwen4-Max. The release materials showed the model not only answering questions and generating text, but also producing financial research drafts, organizing legal due diligence materials, and planning automotive production lines. The team also announced the Work360 industry benchmark to evaluate how well models complete complex professional tasks.
These demonstrations still need to be tested in real-world scenarios, but they raise a common question: to make AI deliver professional results, what should it be trained on, and how do we judge whether it got the job right?
A deal on the other side of the ocean put a concrete price tag on that question. Google bid $10 million to acquire part of the internal data left behind by grounded airline Spirit, including emails, chats, documents, and some software. The airline could not continue operating, but the records its employees left over years of processing orders, coordinating business, and solving problems found a new buyer.
This is exactly the material AI lacks when learning to work. However, buying emails does not equal buying experience. Which records can be legally used, how to extract tasks from them, and how to judge whether the model did it correctly—all of this needs to be reworked.
The internet is the books humanity has read, but what Agents need to learn is what humanity has done.
Over the past few years, tech giants have spent enormous sums buying GPUs and building computing clusters. Now they also have to find new learning materials for machines. Computing investment has no end, but stronger computing power alone is no longer enough to answer the next question: how do we make AI truly learn to work?
The story of the gold rush has been used to explain too many tech businesses. The gold diggers may not have gotten rich, but those selling shovels and jeans made money. Yet no matter how well shovels sell, what ultimately sustains this business is still that there is gold in the mine to dig.
After the era of individuals taking a gold pan to try their luck passed, large mining companies with mining rights, capital, and equipment gradually took over the business. More than a century later, Newmont, founded in the early 20th century, has become one of the world's largest gold producers. In 2025, the company's net profit reached about $7.2 billion. The gold rush will eventually pass, but the gold mine business can continue indefinitely.
Today, NVIDIA has already proven how much money can be made selling shovels. What AI data companies—third-party companies that specifically provide training data, training tasks, and evaluation services to model companies—want to do is control the source of the mine that model training cannot do without.
In 2020, OpenAI published its Scaling Laws research: as parameters, data, and compute scale together, model performance rises along a predictable curve.
This brought a previously rare certainty to the AI industry: as long as there are enough funds and computing resources, there is an opportunity to improve model capabilities along a relatively clear direction. This is the new vein of ore in the AI era. In the years since, NVIDIA—with GPUs, CUDA, and its supporting software ecosystem—became one of the most important beneficiaries. Last October, NVIDIA became the first company whose market value reached $5 trillion, and it is therefore often called the shovel seller of the AI era.
But how long shovels can be sold ultimately still depends on how long the gold mine can be mined.
In 2022, DeepMind's Chinchilla had only 70 billion parameters and consumed 1.4 trillion tokens, yet it beat the 280 billion-parameter Gopher. This means that having more GPUs does not necessarily mean higher training efficiency. Insufficient data volume and unreasonable data ratios can likewise waste expensive computing resources.
In 2024, Epoch AI estimated that, based on the expansion trend at the time, model training could fully use existing public human text resources between 2026 and 2032. At the end of 2024, Ilya Sutskever called data AI's fossil fuel at NeurIPS, and there is only one internet.
In 2026, the internet still produces new web pages, code, video, and audio every day. Models can reuse existing data and can also continue to improve through synthetic data and new training methods. But public text that is easy to obtain, high enough in quality, and can significantly improve frontier model capabilities cannot keep increasing at the past pace forever.
The Chinese-speaking world is even narrower. Most good content is trapped in the respective account systems of WeChat, Douyin, and Xiaohongshu. Crawlers cannot get in, and obtaining authorization, cleaning, and organizing data all require extra investment. According to industry insiders, in the early years, to gather enough high-quality text, some people bought old books in batches, removed the bindings, and scanned them page by page.
Sutton saw further than most. In 2019, Richard Sutton, the father of reinforcement learning, wrote The Bitter Lesson, urging peers to stop manually instilling human knowledge and instead bet on general compute. In 2025, he co-authored Welcome to the Era of Experience with David Silver of DeepMind: the well of human data is almost dry, and the next stage of intelligence must grow from AI's own actions and feedback in environments.
o1 and DeepSeek-R1 first conquered math and code because math problems have answers and code can run tests; models score and improve through repeated trial and error. Andrej Karpathy put it more bluntly last November: traditional software automates things for which rules can be written, while neural networks automate things whose results can be verified.
But experience does not arise out of thin air.
For AI to practice repeatedly, there must first be operable environments, clear task goals, and standards for judging whether results are good. Math and code are relatively easy, but enterprise work is hard to organize into this form.
For example, when a procurement manager receives a delay notice from a supplier, they need to judge whether it affects the production plan, whether to change suppliers, and whether to adjust payment terms. For the same contract, legal, finance, and sales departments may have different requirements.
Leading labs are currently filling this gap. On October 6, OpenAI announced that GPT-6 Astra was the first frontier model trained with Ironclad tasks. Contract management software company Ironclad provided a hosted environment for the product, and processes such as contract approval and clause configuration were turned into tasks that the model can repeatedly practice and score according to rubrics.
These data are processed into tasks, environments, and scoring standards that models can repeatedly practice. The next question is, who will mine this mine? There are two paths: one is for model companies to build it themselves—in-house—and the other is to buy from third-party data companies—external procurement. The latter is exactly the new industry that has heated up rapidly in the past two years.
Recommendation platforms can continuously obtain training feedback from users' clicks, dwell time, and transactions. Every time a user scrolls through Douyin or searches on Baidu, they label an interest for the platform, and the flywheel turns by itself. In the recommendation algorithm era, ByteDance was almost the company that least needed to buy data externally.
But in the AI era, this approach no longer always works. An office Agent product manager told New Position that users leave many task records, but very few can actually be used for model training: sometimes only a few out of tens of thousands may be usable.
This is especially true for complex professional tasks. A single successful operation may not be enough to teach the model. It also needs to know what judgments were made in the middle, where errors occurred, and why the final result was correct.
If data casually left by users is not enough, companies have to pay people to do it.
In 2009, Li Fei-Fei's team at Stanford published ImageNet research. To teach computers to recognize objects such as cats, dogs, and cars, researchers collected images from the internet and then used crowdsourcing channels such as Amazon Mechanical Turk to hire people to check and label them one by one. The project mobilized nearly 50,000 crowdsourcing workers over time, and the dataset later expanded to about 14 million images.
At that time, humans were responsible for telling machines what was in an image. More than a decade later, what AI companies pay people to do has become teaching machines how to complete a job.
In the past, articles, images, and code on the internet could provide large amounts of training material. Now, AI is beginning to learn how to complete professional work. How employees handle exceptions, how experts make judgments, and how tasks should be accepted after completion rarely appear completely on public web pages.
Robotics is another front. What is learned from books is superficial; the internet can teach a robot to describe an image, but it cannot teach it to tighten a screw. It needs to understand object states, action sequences, operation results, and how to retry after failure, and these materials are hard to obtain in batches like ordinary web pages.
Once there is demand, money is naturally spent.
U.S. expert data platform Mercor connects AI labs with professionals across industries and undertakes training and evaluation tasks. It once spent more than $500,000 to have professionals from Goldman Sachs, JPMorgan Chase, and McKinsey complete 200 tasks, averaging more than $2,500 per task.
Surge AI, also in the U.S., mainly provides high-quality human annotation and evaluation data to leading model labs and maintains an expert advisory group numbering in the hundreds. According to reports, at peak demand, some model factories could spend more than $10 million a day on programming data.
However, this business currently has no unified pricing unit, and the price gap is enormous.
Simple image classification or text annotation is worth only a few cents per item, but a mathematician can earn thousands of dollars for posing a difficult problem to AI. Mercor hires radiologists at $200 to $400 per hour, some oncology experts at $450, and the emails and chat records of a bankrupt airline were quoted at $10 million.
Data prices vary widely, but the value of companies doing data business is generally rising.
Domestic expert data platform Humanlaya has gathered more than 100,000 experts across industries and disciplines. This year it received investment from CDH, Sequoia China, and Capital Today, with a valuation close to $1 billion; evaluation and synthetic data company UniPat, less than a year old, has already reached a valuation of $2 billion to $3 billion.
Overseas examples are even more extreme. Last June, Meta acquired about 49% of Scale AI for approximately $14.3 billion. Founded in 2016, Scale is one of the largest data annotation companies in the U.S. It started with autonomous driving image annotation and later became an important data supplier to leading model companies.
Capital is willing to pay this price not only because of the demand changes mentioned earlier, but also for another reason: what models need to learn has changed.
Today, it is increasingly difficult for companies to open obvious gaps in technical routes, data processing, and quality, and model release cadence is fast. Whoever can obtain training materials that others do not have will undoubtedly gain a bit more advantage.
In the era of search and recommendation, data was a byproduct of traffic and was mostly placed under a middle platform. Now, more and more companies are letting model leaders directly participate in data decisions.
From July to September, data leaders at several major domestic tech companies were adjusted in succession. In July, Xiaohongshu hired Chen Guanting, an early core member from DeepSeek, as head of pretraining data. In August, ByteDance upgraded AI data to a first-level department parallel to Seed, Flow, and Douyin, overseeing sourcing, evaluation, and delivery, with a separate Pretrain Data department under Seed.
On September 23 at the Yunqi Conference, Alibaba introduced four leaders for training data, expert ecosystem, evaluation, and embodied data, and its Cognitive Origin Plan opened 10,000 expert collaboration seats in medicine, advanced manufacturing, and financial compliance. On September 28, Tencent's chief AI scientist Yao Shunyu also became head of the AI Data department, putting models, infrastructure, and data into one person's hands.
Yao Shunyu wrote The Second Half in 2025, whose core point is that after the reinforcement learning recipe matures, what separates the field in the second half is how to define tasks, build environments, and design evaluations.
DeepSeek took another extreme. According to LatePost, Liang Wenfeng once had half of the core researchers personally participate in data cleaning, synthesis, and annotation. He calculated: if the model architecture is half as good and data efficiency is also half as good, together that requires 4 times more compute.
Zhipu also put data production into its internal process. As early as 2024, Zhipu had configured a data procurement team overseas, and internally Zhang Xiaohan led the construction of an annotation platform called Kuangbiao. During rotation training between language model and multimodal model teams, the team waiting for training also participates in data annotation. Before each model release, it must pass two tests: one by the AI institute checking general capabilities, and another by the delivery team evaluating according to real customer scenarios.
ByteDance's orientation is in-house research. At the Seed all-hands meeting at the end of July, Zhang Yiming clearly stated that the company would not distill competitors' models and would develop everything in-house. Not relying on distilling competitors' outputs means Seed needs to rely more on its own R&D, real tasks, and data production system to accumulate capabilities.
However, in-house does not mean not buying.
While insisting on in-house development, ByteDance is also a major buyer externally: world models, film and television master tapes, programming data, and expert networks have become several main external procurement tracks. According to data suppliers, ByteDance's investment in world model data is 3 to 4 times that of peers.
What major companies outsource is precisely third-party data providers. By the type of data they supply, these providers can be roughly divided into three categories: traditional training data service providers, expert and professional task data companies, and embodied data companies.
The first category: traditional training data service providers. These companies started out in businesses such as speech, image, and text annotation and serve a broad customer base. Speechocean was founded in 2005 and has long provided AI training data services in speech, vision, and natural language. In the first half of 2026, the company generated operating revenue of 192 million yuan, up 22.45% year on year. Its disclosed customers and partners include Alibaba, Tencent, Baidu, ByteDance, and several large overseas tech companies.
The second category: expert and professional task data companies. Mercor and Surge AI, mentioned earlier, both belong to this category. They do more than annotation; they organize experts across industries to produce the tasks, environments, and scoring standards models need in order to practice professional work. To train a model to complete investment analysis, a data supplier must not only provide the final report but also have experts prepare original business materials, reference operating steps, and a scoring checklist for evaluating results.
The third category: embodied data companies. They target robot training and collect operational data on people and equipment in real environments. Mifeng Technology, founded only in February this year, was incubated by domestic humanoid robot company Zhiyuan. In June it mass-produced a wearable, body-free data collection device, and on August 31 the 20,000th unit rolled off the line and was delivered to JD.com. However, these data cannot be automatically crawled like web pages. For a robot to learn an action, people and equipment often need to repeatedly demonstrate and collect data on site.
But in the view of New Position, high prices and valuations do not mean this business is easy to run.
For third-party data companies, the difficulty stems precisely from the coexistence of two paths: in-house development and external procurement. As model companies begin to produce data themselves, design tasks, and build evaluation environments, the business of external data companies also faces new tests. Major customers are both a source of revenue and potential competitors.
When major companies build data themselves, orders for external suppliers may be squeezed. But getting too close to a major company also carries risks. Last June, Meta acquired about 49% of Scale AI for approximately $14.3 billion. According to reports, major customers such as Google subsequently adjusted their cooperation with Scale; Google had originally planned to pay it about $200 million that year. Even without an actual data leak, when a supplier serving multiple competitors establishes a deep capital relationship with one of them, other customers have to reconsider the risk of cooperation.
Data companies must not only prevent clients from generating their own data but also avoid losing other clients by being tied to a single one. Many annotation tasks, synthetic data, and automated scoring tools can be built by model companies themselves. If an external vendor merely organizes experts and delivers data according to client requirements, no matter how large the business grows, it may only be a more advanced staffing outsourcing firm.
By contrast, what is more worth investing in is a production system that can be reused across clients. A design methodology for professional tasks, a training environment that can be run repeatedly, proven scoring tools, and a long-accumulated expert network—these need not be rebuilt from scratch every time an order comes in. Only by accumulating these capabilities can a data company have the chance to evolve from a one-off delivery project into infrastructure that model companies rely on long-term.
Rockefeller did not win at the oil wells. When he founded Standard Oil in 1870, Pennsylvania well owners were swinging between fortunes and bankruptcy, with crude oil prices rising and falling sharply. He bet on refining and transportation—using scale to drive down costs and railroad freight rebates to crush competitors—and ten years later controlled roughly ninety percent of America's refining capacity. Drillers came and went; what Standard Oil captured was the link that turned crude oil into products.
High-priced data is not necessarily a good business either. Clients may produce it themselves, exclusive licensing limits reuse, and a batch of expensive tasks may quickly lose scarcity.
Data expires, buyers build their own, and buyers are frighteningly concentrated: Mercor's $614 million in revenue in the first half of the year, about 91% of which came from foundation model companies. Meta bought nearly half of Scale's equity, and major clients reportedly left as a result—the neutrality of a well evaporated in a single transaction announcement.
This is also why the data industry cannot simply replicate Nvidia.
Nvidia's advantage lies not only in GPUs but also in CUDA and its developer ecosystem. When CUDA was released in 2007, GPUs were still for gaming; in 2012, AlexNet was trained on two gaming graphics cards, and only then did this software wait for its moment to become the industry standard. Over the following decade-plus, every generation of developers wrote code on CUDA, and the cost of replacing it grew ever higher.
The CUDA of the data era will not be a pile of files, but a validation system composed of tasks, environments, and scoring standards. Able to rerun, able to iterate repeatedly, able to score automatically. This requires both industry knowledge and environment engineering, quality control, and independent evaluation. Automated scoring can scale up, but the scoring itself must still be subject to scrutiny.
Many can take away training data; few can decide "what counts as done right."
*The header image and illustrations in this article are sourced from the internet.