A Small Player in the Data Underworld

@Wu Ruirui

Before National Day, elsewhere published a piece about data companies acting way too godlike. The star data vendors featured in that article were just the cherry-picked few. Far larger is what's happening beneath the surface. Over the past while, what I've learned about the scope of the data industry and its current state has exceeded even my own expectations.

If AI is, in your imagination, a place where no ordinary people tread, the data industry will disabuse you of that notion. It's more like Beijing's Sanyuanli market in the early morning — there are rare ingredients selling for thousands of yuan, and celery going for a few kuai per jin. The shoppers come from downtown, the embassy district, luxury compounds, and immovable old neighborhoods alike. All walks of life, buying and selling amid the same din.

Just before I started writing, I met Xiao Wu. He is practically the distillation of every archetype in this industry.

A year ago, Xiao Wu was still a fringe AI app entrepreneur. Three years into his startup, he had little to show for it: users wouldn't pay, VCs wouldn't invest, and his savings of several million yuan had evaporated. To make ends meet, he started taking contract work, helping companies deploy models. In the process, through dealings with the motley crowd of token resellers, he unexpectedly became a data broker.

Over the past year, Xiao Wu has scraped data from underground overseas websites, visited assembly-line thousand-person labeling factories, and tracked expert networks on Surge in Silicon Valley. He's seen middle-station operators making tens of millions annually selling data, and others who crossed red lines and vanished without a trace. He himself was lucky — not long after starting, he'd earned back everything he'd lost.

Perhaps because of his telling — sometimes animated, sometimes disillusioned — the silent, vast masses of this industry came into vivid focus before me.

The data industry rewrote Xiao Wu's view of entrepreneurship, yet also kept him lingering. He even has an extreme formulation: there are only three paths in AI — sell your company (raise funding), sell tokens (model factories / relay stations), or sell data. For someone like him who can't raise funding or join a model factory, selling data is the sweetest deal.

The data industry seems distant, but in fact concerns everyone. It extracts what's in everyone's mind to feed models, and will ultimately affect everyone's life. The top data vendors lean closer to research — producing questions, environments, and trajectories. But beneath them, all manner of people hustle, circulate, and trade.

And Xiao Wu's story will let you more truly understand how all of this happens.

"Xiao Wu" is a pseudonym. The protagonist's story somehow reminded me of the film Xiao Wu. Beyond being primarily based on his own account, this piece also incorporates multiple real stories from my interviews. To prevent identification, certain details and narrative threads have been obfuscated.

Everything Is Data

Xiao Wu first realized the data business existed this spring.

A former schoolmate who sourced data for model factories asked if Xiao Wu could scrape a batch of specified videos from overseas sites. Having worked in broad entertainment for years, he did know where to scrape — but this counted as data?

Still running his startup, he found a technically skilled employee. In a few days they'd scraped dozens of terabytes. The cost was essentially nothing — electricity, internet, employee wages, working out to a few dozen yuan per terabyte. But the schoolmate actually bought it at 40,000 yuan per terabyte. That first data deal brought in over a million yuan. Far easier than app entrepreneurship.

Of course not every deal was this profitable. Xiao Wu later tried selling to other companies with video products. This time, the price was negotiated down from 40,000 to 600, ultimately closing at 1,000 yuan per terabyte.

This taught Xiao Wu that relationships are the prerequisite for everything: knowing in time who wants what data, who can afford it, who to sell to, and building sufficient trust with that person.

He discovered that the broker hierarchy in this industry has many layers, each subcontracting and taking a cut. The closer one's connections are to the top, the higher the prices. The layer closest to model factories is fixed — those few star data vendors. Besides producing themselves, they also serve as integrators for model factories. Every small supplier wants to know them.

Xiao Wu shared neither school nor work history with them, so he decided to cultivate himself. He received all manner of requests from his schoolmate — menu recipes, drug instructions, handwritten medical records, game source code, console game screen recordings, film studio outtakes... The schoolmate told him: for models to replace all labor, they must learn all industries' knowledge. Real, high-quality data and trajectories are now in short supply, and model factories will certainly buy, as long as he's willing to search every corner of the world.

So a month later, Xiao Wu shut down his app project entirely. He decided to become a data nomad, for the time being.

He went to Silicon Valley, to the people doing gig work on Surge. Platforms like Surge — many domestic model factories want their data. He infiltrated student groups, recruiting well-connected people who also took Surge assignments as subordinates. Once they'd thoroughly mapped Surge's expert networks, he found them one by one and had them download copies of their Surge work for him, paying the same compensation as Surge. Earning extra for zero additional effort — almost no one refused.

He went to domestic film annotation factories. These were thousand-person operations, workers being film and directing students and professionals, daily annotating shot divisions, camera angle differentials, frame ratios for AI — teaching AI the film knowledge they'd learned. Their compensation was under 300 yuan per day, roughly equivalent to the cost of generating 20 standard 15-second videos with Seedance 2.0.

He went to companies specializing in gameplay videos. World models need data reflecting physical laws and complex interactions. 3D game player operation data, modeling parameters are good substitutes — complete records of health points, gravity, impact force. Some non-top-tier open-world game companies were already selling such data; as for unwilling top-tier game IPs, these companies would specially hire employees to play and record gameplay footage live. He could find hundreds of millions of hours of console game video.

He went into university laboratories. Scientists' experimental process data is precious food for AI4S. The thousands of experiments behind a single paper — what matters to models isn't just the successful runs, but the failed ones, the modified trajectories. The task given to him was to buy scholars' complete experimental records. Dealing with scholars, money isn't a panacea. They care more about academic reputation: Will word get out that I sold data? Is there evidence of academic sloppiness in my experiments? If AI proves some conjecture using my materials, does the credit go to me or the model?

In his travels, Xiao Wu saw a microcosm of society. Anyone, any industry, anything that happens — it can all become data. And models have found compradors in nearly every industry. Ordinary people with some access, mainly seizing opportunities, can all quickly monetize.

Circles were also rapidly narrowing. Easily obtainable simple data was falling in price, but high-quality expert data was rising. The overall data market wasn't shrinking, but it was getting harder for ordinary people to participate.

Company Undertaker

That summer, a friend Xiao Wu had met in Silicon Valley sent him a business opportunity: they had data from Silicon Valley companies and wanted help finding buyers domestically. This was a more efficient data collection method: "body collection" for bankrupt or moribund companies.

This opportunity list had over a hundred companies willing to sell their data, covering every industry. Consumer goods — perfume, supplements, apparel, games; healthcare — specialties, claims, pet medicine; offline services — restaurants, industrial design, real estate development, property management; tech — SaaS, software development; functional companies — e-commerce, logistics, sales; even primary markets — incubators, VC portfolio management... It seemed data from any industry, any company could be bought and sold.

When he opened the detailed description for one of them, he was shocked. This data inventory almost completely recorded the past decade-plus of this SaaS company's operations — from customer acquisition, contracting, delivery, R&D, after-sales service, internal management to daily collaboration.

Xiao Wu felt his hair stand on end. If models could learn all of this, what remained of human "expertise" or "experience" in the workplace? But for himself, trading such data was clearly far more valuable than individual personal data. So while panicking, he vigorously brokered deals. Ultimately he found a domestic model factory to buy this company's data, closing at $3 million.

After this deal, his friend forwarded more requests from model factories: seeking to buy pivot projects from AI app companies, budgets ranging from tens of thousands to a million dollars; a Chinese model factory even asked Xiao Wu if he had non-performing asset lists from banks, to find companies that appeared normally operating but were actually near bankruptcy.

Once when Xiao Wu mentioned these in casual chat with an investor, the other party was inspired: so many of their fund's moribund portfolio companies — if they could package and sell their data, wouldn't that be equivalent to a buyback?

Among all company data, AI app company data is the most valuable. Some companies do harness well, with long enough task trajectories — records of users collaborating with AI, giving feedback, making revisions, all preserved in the data, naturally becoming learning material for models.

For a period, Xiao Wu often asked app companies on behalf of model factories whether they'd sell user trajectories. As a former app entrepreneur, he of course knew how outrageous such inquiries were — essentially pronouncing death on the product. At first, the rejections Xiao Wu received were swift and decisive. But he could sense the entrepreneurs calculating that trade-off: If I don't sell, will my competitor sell? If I don't sell, will the model still learn my scenario?

Humans hesitated; models learned without pause. After some time, entrepreneurs would proactively approach Xiao Wu to sell data, but model factories had raised their standards: high-value industries only, large user bases only, users who could complete the final step within the product. The price they'd pay had also degraded from actual money to API discounts. When the entire AI app sector also fell to freezing point, some entrepreneurs were even willing to give data to model factories for free, just to build relationships. Xiao Wu guessed this was laying groundwork for acquisition, or for the entrepreneur's next job.

Xiao Wu would recall when he used to raise funding, how he was always interrogated about how he'd compete with model factories. These past few years, some app entrepreneurs avoided model factories, or confronted them, or secretly hoped to be acquired. Never did they imagine there was also this layer of existence.

Closed Loop

But AI app companies are ultimately limited in number, and those that reached PMF even more so. So how to get models more traces of humans using AI?

Xiao Wu once thought this could only rely on data vendors expanding output through expert networks. But later he discovered a more efficient method: relay stations.

Xiao Wu told me that many domestic tech giants need to call OpenAI and Anthropic APIs, but can't do so through official channels. They can only buy tokens overseas through specialized B2B relay stations.

This is extremely profitable business. Overseas "wool-pullers" batch-register startups to apply for AWS and Google Cloud startup credits, securing hundreds of startup slots monthly, netting tens of thousands of dollars in credits per startup per month. Domestic relay stations buy API access from these people at steep discounts, then sell to tech giants at a middle price.

For domestic relay stations, the juiciest part isn't even here — it's selling data. Once enterprise clients start calling APIs and task trajectories accumulate, they extract this data and resell it to other companies. Task trajectory data from high-talent-density organizations calling SOTA models — this itself is extremely valuable distillation material.

The sexiness of this business lies in having no additional cost, only compounding returns. Relay stations just need to periodically compress the data into a zip file, send it over, and after some time a sum recorded under some other name appears on the books.

Some relay stations even combined these two businesses into a perfect closed loop: using data sales revenue to subsidize token prices for clients, clients use more, bringing more data, which continues to be sold. By Xiao Wu's estimate, even second-tier internet companies spend $20 million monthly on relay stations. Many relay stations reached profits of tens of millions of RMB this year.

All of this began cooling this summer. Reportedly, one tech giant, when buying data, was shocked to discover they'd bought their own relay station usage trajectory data. The B2B service they'd purchased was not only scraping their data to sell, but selling it back to them. The giant's management was furious. This caused severe chain reactions in the data industry. Afterward, many who ran relay stations while selling data vanished from sight.

Though he didn't run relay stations himself, this sobered Xiao Wu somewhat from his money-making fever. A fact he well knew: ordinary data sellers, himself and all brokers, some top data vendors, even model factories themselves — many practices existed in gray zones.

Black swan events often mark industry turning points. The National Data Group established last month will undoubtedly accelerate this rhythm.

Plague

But why has this gray trade swept so broadly?

As a broker drifting among the masses, Xiao Wu could feel the undercurrent of fever and despair intertwined. In his view, those who dared use the wildest paths to collect data weren't money-seeking brokers, but some model factories.

"Any method to improve model capabilities, they'll pursue at all costs. Even if exposed, it doesn't matter." For model factories, the open secret is: intelligence is labor, is commercialization. As long as the math works out, this machine will keep devouring.

Down to each individual, there's shared helplessness: everyone knows full well that AI evolves by eliminating humans. Buying and selling data is offering up what's in your mind to the model, until one day you're no longer needed. Today even people at model factories, training models, are being laid off.

But it seems no one is willing to stop. Everyone wants to grab a lifeboat on the Titanic.

Xiao Wu remains bitter about failing to raise funding when he was building apps. Investing his entire savings into entrepreneurship at the age of standing firm — one could hardly say he lacked entrepreneurial spirit. Yet Xiao Wu still felt at a loss in this era. For someone like him who couldn't enter a model factory or carve out a living in apps, data suddenly seemed like the final battlefield. He threw himself into it with something like vengeance.

"I sell data to those people who got big results in fundraising — it's like I'm also earning money from investors."

And after that?

Xiao Wu figures that if it takes ten years for models to fully replace humans, his data business can last ten years too. I told him that many in the VC industry doing data call this a short-term business, soon to end. He said that's because they still want to do things that can raise funding, be capitalized, have high ceilings. He just wants to treat this as a business.

As for after ten years: "AI and humans must have a showdown — either AI is harnessed, or cyberpunk." But that's too distant, and has nothing to do with a small-time data vendor like him.

A data vendor friend once told Xiao Wu something: once they hired a batch of experts to label data, paying thousands of yuan hourly. One of them, labeling away, suddenly broke down and started cursing: "You keep injecting what's in our brains into machines every day — is AI saving humanity or destroying it?"

This friend thought about this for several days, finally admitting that indeed this was the fact — some people in AI are making massive profits, but many people's lives don't seem to be getting better. Are some people able to earn so much because they're smarter, work harder, are better at narrative — or simply because they chose the right major?

The conversation stopped there. Neither of them said anything more.

Cover image: Victor Gabriel Gilbert, Le Carreau des Halles, 1880, MuMa