AI data prices: 1.7 cents per chat, $2.60 per filmed hour, 100 yuan per face

What's going on

Language models have largely picked the open internet clean. AI labs now want something different: real workflows like the ones in company inboxes and chats, footage of hands at work for robots, and real faces for AI films. In 2026, a separate market sprang up for each of these data types within a few months, each with its own brokers, platforms and prices.

We compiled a price list of unit costs from news reports, court filings and vendor statements. A pattern emerges. Brokers and platforms make the most money, and so do creditors when a company goes bankrupt. The people whose emails, hands or faces are being sold get little or nothing. We also look at whether this kind of trade would even be legal in the EU.

Market 1: The bankrupt airline's inbox

The best-known case is Spirit Airlines. The US budget carrier ceased flight operations on May 2. In August, Google won the bankruptcy auction for the company's internal archive with a $10 million bid. According to Time, it contains about 100 million emails, 500 million Microsoft Teams messages and roughly 30 million lines of code. MIT Technology Review also cites 176,000 employee records. The runner-up bid of $7.5 million came from AI data vendor Mercor. After the deadline had passed, startup Micro1 offered $12.5 million.

Why would anyone pay for this? According to Time, labs use it to build reinforcement learning environments. These are simulated workplaces where AI agents practice tasks and get rewarded for success. Nick Heiner of data provider Surge AI told the magazine that such environments are only as good as their data. Even bad decisions by real employees can be turned into training tasks.

The sale isn't final yet. The court's consumer privacy ombudsman, Lucy Thomson, recommended approval on October 5. Before that, booking and passenger data on roughly 97 million customers had been carved out, and the company Tonic.ai is supposed to de-identify the remaining data, meaning it will strip out names and other direct identifiers. According to The Next Web, Thomson explicitly did not assess the risks to employees. The flight attendants' and pilots' unions have filed objections, and more than 100 members of Congress want payroll records, timekeeping data and parts of the communications excluded. According to Judge Sean Lane's calendar, the hearing has now been pushed to November 17.

Union attorney Charles Rubio warned, according to Gizmodo:

“A pseudonymized dataset can still reveal which crew bases complaints came from.”

This has also created a business around smaller companies. According to MIT Technology Review, wind-down service SimpleClosure has brokered more than 150 deals through its “Asset Hub” since April. In 47 percent of them, code or data was sold. By the end of 2026, SimpleClosure aims to return more than $10 million to dissolved startups. One founder got about $25,000 for his crypto startup's code and expected roughly another $15,000 for data from Google Workspace, Notion, Figma and Slack. Competitor Sunset sets prices based on factors such as headcount, company age and data volume. According to MIT Technology Review, the proceeds usually go to senior creditors and investors. Law professor Ifeoma Ajunwa calls this “captured capital.” No federal US law bans such sales. According to Bloomberg, SpaceX's AI division, formerly xAI, also considered purchases of this kind in September.

Market 2: The filmed work hour

Robots learn manual tasks from egocentric video, meaning first-person footage shot with a camera on the head, chest or wrist. The demand is huge, as Memories Upload shows. Fortune featured the platform today. It comes from Memories.ai, a San Francisco startup founded by former Meta researchers Shawn Shen and Ben Zhou. Participants wear the “LUCI Pin,” a camera weighing less than 45 grams. The company says more than 1 million hours were collected in the first beta month, and it aims for more than 10 million by March 2027. Memories.ai says three of the five leading AI labs are customers but doesn't name them. According to Crypto Briefing, Memories.ai pays $6 to $18 per task, such as chopping vegetables or fixing a bike. How long a task takes isn't known, so you can't work out an hourly rate.

Other vendors' numbers are more precise:

• Human Archive (India) pays a base rate of $1 per hour, according to TechCrunch. Home services customers get a discount if they agree to be recorded.
• Objectways in Karur (India) pays about $2.60 per hour, according to NPR. A textile worker said no one had explained to her what the footage was for.
• In the Philippines, one collector earned $203 for about 71 approved hours, or $2.86 per hour, according to the Globe and Mail. Many of his recordings were rejected, for example because they didn't show “meaningful work.”
• DoorDash Tasks pays drivers in select US regions for jobs like filming themselves with a body camera while washing at least five dishes. California, New York City, Seattle and Colorado are excluded, all places with stricter regulation. DoorDash doesn't publish fixed rates.

And what do buyers pay? Data vendor DexSet lists its own prices at $15 to $40 per raw hour. Between 15 and 30 percent of the footage gets discarded. An analysis by Spark Science, which we know only from secondary sources, puts raw video from India at $1 to $10 per hour. Data broker Flikforge, on the other hand, describes a glut: generic footage sells in bulk for $2 to $5 per hour, below production cost. All of these figures come from vendors and haven't been independently verified.

Our math: If you earn $2.60 an hour and your footage is resold for $15 to $40, you get roughly 7 to 17 percent of the sale price. Rejected recordings often aren't paid at all.

Market 3: Faces as inventory

China is now churning out AI-generated short dramas. According to the State Council Information Office, more than 90 percent of the 430,000 short dramas released in the first eight months of 2026 were AI-generated. Synthetic faces quickly look artificial, so there's demand for real ones. According to the business paper Daily Economic News (每日经济新闻), a face costs as little as 100 yuan at the low end. That buys 2 to 3 photos and a signed consent form, and the whole process takes about ten minutes. The usual range is 500 to 1,500 yuan, mostly as a one-time, perpetual release. Influencers with audiences in the millions get 8,000 to 10,000 yuan. Rest of World, as cited by The Next Web, reports 99 to 500 yuan per episode, or $15 to $74. According to Sixth Tone, platforms keep about 30 percent, and the platform ActID takes a tenth, according to The Next Web.

Sixth Tone also looked at what people actually earn. Many users got few or no requests after signing up. The viral story that you can earn “almost 2,500 yuan in three days” with your face is more the exception than the rule. A report by 显微故事 (Xianwei Gushi) published on November 10 and picked up by the portal smzdm analyzes the contracts involved. It lists five typical predatory clauses:

1. a “perpetual, irrevocable” one-time payment,
2. worldwide use across all channels, “including but not limited to,”
3. a pre-checked list of permitted content, including villain roles, death scenes and intimate scenes,
4. unlimited transfer to third parties via sublicenses,
5. a “lowest price online” clause with a 10,000 yuan penalty per violation.

At 10,000 yuan, the penalty can be up to a hundred times what the face sold for. The authors point to China's data protection law, PIPL. Under it, facial data counts as sensitive information and requires separate consent, and any change of purpose requires new consent.

The price list

For comparison, a study by Hamidah Oderinwale and Anna Kazlauskas (arXiv 2510.24990) analyzes data deals from 2020 to 2025. Reddit gets about $60 million a year from Google. HarperCollins charges $5,000 per book, and authors get half. Licensed video in the study costs $1 to $4 per minute, which works out to $60 to $240 per hour. That's professional footage with cleared rights, though, so it's only partly comparable. One finding still fits our thesis: only 7 of 24 major deals pay the original creators. The rest pay only intermediaries.

Would this be legal in the EU?

Short answer: hardly, at least not in this form. If you want to use data for a purpose other than the one it was collected for, Art. 6(4) GDPR requires you to check whether the new purpose is compatible with the original one. This principle is called purpose limitation. Nobody writing work emails expects them to end up training AI agents one day. On September 11, 2024, Germany's Data Protection Conference of federal and state regulators (DSK) stated that employee data may generally be transferred only with consent unless there's a transfer of undertaking under Section 613a of the German Civil Code (BGB). A transfer of undertaking means a buyer takes over the business along with its workforce. The DSK also says that selling customer data as the sole asset generally requires consent. Law firm WBS.LEGAL considers blanket employee consent “practically impossible,” because employees aren't truly free to say no to their employer. On top of that, chats often contain health data, which gets special protection under Art. 9 GDPR.

Anonymization would be the remaining way out. According to Börsen-Zeitung, Felix Glocker of law firm CMS considers it the more realistic route. Michael Rozijn of Schultze & Braun, on the other hand, expects regulators to view a transfer of this size “extremely critically.” It would take “near-perfect anonymization.” In our research, we didn't find any published statement from a German data protection authority on AI data sales out of insolvency proceedings.

Two cases we've reported on show that Europeans have an effective tool here. Voice actor Matti Klemm learned through a GDPR data access request that Netflix had used his voice for AI training, and he now plans to sue. Makers of EEG headbands changed their privacy policies only after the journal Science started asking questions. In the US, by contrast, Rubio says even California's privacy law, the CCPA, protects customers but not personnel records.

The bigger picture

Our take: the price list largely backs up the thesis. With bankruptcies, it's clear-cut. The people who wrote the emails get nothing, and the money goes to creditors, investors and brokers. With video, collectors do get paid, but by the vendors' own figures only a small share of the sale price, and rejected footage often earns nothing. In China's face trade, the platform reliably takes its cut, while many of the people offering their faces never land a single job.

A few caveats apply. Most buyer-side prices come from data brokers who have something to sell themselves. According to MIT Technology Review, there's no public evidence yet that real company archives measurably improve AI models. And if Flikforge is right about the video glut, prices will keep falling anyway. Whether US courts will protect workers will be decided on November 17 in the Spirit case. In Germany, there's a strong case that a deal like this would fall foul of data protection law, but no court has ruled on it yet.