The Ultimate Guide to AI Data Buyers (2026): Every Company Compared

The short answer#

Five AI data companies now publicly invite ordinary businesses to license their operational records: Mercor, micro1, Scale AI, Handshake AI and Turing.¹²³⁴⁵ Four of them publish a payout range, and each page treats its range as indicative.²³⁴⁵ Beyond them are marketplaces and brokers that match sellers to buyers, wind-down services that sell the data of closing startups, and a long list of well-known expert-data suppliers that, on their own pages, describe building new data and not buying yours.

What is missing matters as much as what is there. We found no program page that publishes a full contract, and no named business that says it was paid under any of the five programs. The documented completed sales are mostly from failing companies and one bankruptcy auction, which are unusual situations.⁶⁷⁸ A business that is operating normally and holds a few years of well-organized, legally clean records is the person these programs are written for, but the public record does not yet show how often such a business gets a deal, or on what terms.

This guide covers how each buyer buys, grouped by how they work, and what each does not tell sellers. For every company we keep to what is publicly known: what it says it buys, how it pays, any figures it publishes, and what it leaves out. It then explains the contract terms that move value more than price does, shows which data types buyers say they want, and ends with red flags, diligence questions and the case for never letting one buyer set your price.

How to read this guide#

What counts as a buyer here#

We include a company if its own public materials, or a credible news report, show at least one of the following as of October 2026:

1. It says it buys, licenses or acquires business or enterprise data, or runs a program inviting companies to submit it.
2. It runs a marketplace, exchange or brokerage where companies can list or sell data to others.
3. It is a major AI data supplier that companies commonly consider, even if its pages do not describe buying data. These are included so you can see where they stand, and are labeled as such.

We do not count a company as a buyer of a data type unless a cited source says so. A page that lists a data type shows stated interest. A news report of a sale shows that a transaction was reported, usually from people who spoke without being named, and it does not establish the terms.

Sources and dates#

Every claim in this guide links to a source we opened and read in October 2026: company program pages, company announcements, filings and regulator texts, law-firm commentary, and news reports. Company pages change without notice, so statements about them are as of October 2026. Where a number comes from a news report citing anonymous sources, we say so. Where a company states a figure about itself, such as revenue or expert counts, we label it as the company's claim and do not treat it as audited.

We label every price figure as one of three kinds. Published means a company states it on its own page. Reported means a news outlet or filing relayed it, often citing people familiar with the deal. Illustrative means the page itself says the number is an example or an estimate and not an offer.

What this guide does not do#

It does not value your data, recommend a buyer, or say that any company will buy what you have. It is also not a directory or an application guide: it describes how buyers buy, and it leaves out how to approach them. Prices vary by unit, term and rights, and almost none of the public figures are rate cards for ordinary business data. For background on rights, privacy and deal structure, see our guide to selling data to AI companies.

At a glance: every buyer compared#

The table covers every company in this guide. "Not published" means we found no figure on a page we opened. "None found" means the company's public pages describe no such program. Figures are as of October 2026.

Group 1: Companies with published data-buying programs#

Five companies publish a page that invites companies to submit operational data for payment. All five are also expert marketplaces or data vendors, so most of their revenue and public attention comes from other work. Each program is recent. We found no page in this group that names a business paid under the program.

Mercor#

What they buy. Mercor's data page lists workflow data from connected tools: Slack workspace messages and threads, Gmail messages and metadata, Salesforce accounts, contacts, opportunities and reports, GitHub repositories, commits, pull requests and issues, and Notion pages and databases, among more than 100 supported tools. It also lists industry examples, including contracts and case files in legal, clinical notes and claims in healthcare, underwriting files in insurance, work orders and QA logs in manufacturing, and order records in retail.¹ The page describes the buyers only as AI labs.¹

What they pay. The page gives no dollar amounts. It says compensation depends on data volume, the tools connected and the depth of the data, with more tools generally meaning higher payout. It says there are no fees, setup costs or deductions, and that Mercor covers extraction, anonymization and transfer costs.¹ A referral bonus exists, with no stated amount.¹

Process and terms. Mercor says it extracts and de-identifies the data into an asset before payout. It says data leaves your systems only after you approve it, access is audited and logged, data is encrypted in transit, keys are managed in AWS KMS, and it holds SOC 2 Type II. It says extraction and anonymization are automated, with no lab buyer or third-party vendor seeing raw data. Payment is by wire when the anonymized dataset is delivered to the buyer, typically two to four weeks after a signed agreement.¹ The page does not say who owns the data after licensing or what license terms apply. Its homepage says automated anonymization masks more than 60 categories of sensitive identifiers.⁵²

Track record. TechCrunch reported in October 2025 that Mercor raised $350 million at a $10 billion valuation, with more than 30,000 experts on its roster earning above $85 an hour on average.⁹ A separate TechCrunch piece named OpenAI, Anthropic and Meta as customers and said most revenue comes from a few labs. In the same piece Mercor's chief executive said labs hire former employees of banks and consultancies because those firms' customers do not want to give them data, and the article said Mercor had sought a startup CTO who could authorize access to a production codebase, with a few accepting.⁵³ In the Spirit Airlines bankruptcy auction, Time reported that Mercor bid $7.5 million and lost to Google's $10 million.⁸ We found no public statement from a named company that it was paid under the data program.

Watch-outs. Mercor says it was affected in late March 2026 by a supply chain attack on the open-source LiteLLM tool, and in a June 25 update said a very limited subset of its nearly five million experts had sensitive information affected, with no evidence of fraudulent use.¹⁰ TechCrunch reported that a hacker group claimed 4TB of stolen data, that Mercor did not comment on whether the claim was authentic, that Meta had paused its contracts indefinitely according to Wired, and that contractors had filed suits over the exposure.¹¹ These are incidents about Mercor's expert platform, not about its data licensing program, but anyone connecting business systems needs current security evidence. The homepage average expert rate read $103 an hour when we opened it on October 21, 2026, and our own earlier review in October recorded $107, so treat advertised rates as moving.⁵²

What sellers should watch. Mercor publishes no amounts, so there is no anchor at all, only a pay basis of volume, tools and depth. The page states no ownership, license or exclusivity terms, and payment follows delivery to a buyer it describes only as AI labs. A price with no named buyer and no stated scope cannot be judged.

micro1#

What they buy. micro1's data page lists standard operating procedures, knowledge bases, internal documentation, CRM data, project histories, QA processes, decision-making patterns and human feedback on AI outputs, across software and engineering, finance and accounting, customer support, sales and CRM, legal and contracts, and operations and logistics.² The company said on October 9, 2026 that it would spend $1 billion over 12 months acquiring and licensing enterprise operational data, financed by Citi and Hercules Capital, and that it uses de-identified company data to build reinforcement learning environments that reflect real business operations.¹²

What they pay. The page publishes three tiers: $100,000 or more for qualified enterprise data partnerships, $500,000 or more for large-scale operational datasets or ongoing participation across teams, and $1 million or more for highly unique proprietary operational data. Compensation depends on dataset size, workflow complexity, data quality, domain expertise and uniqueness. A referral reward of up to $50,000 is paid when the referred company reaches revenue milestones with micro1.²

Process and terms. The page describes an assessment, an evaluation and a discovery call before an agreement. Before participating the parties define approved scope, confidentiality standards and redaction requirements. Personal information is removed when relevant, access is limited to a few authorized staff, customer data runs in isolated pipelines, some datasets may be rewritten synthetically to reduce direct ties to original records, and companies can review representative samples before prepared data is used. Data is kept for agreed periods and deleted on request or at the end of the engagement under the agreed terms, and the page says companies retain ownership of their underlying data.² The Next Web reported the same tiers and eligibility, and said the output serves AI research companies and enterprises without saying who pays for the environments.⁵⁴

Track record. TechCrunch reported in August 2026, citing a person familiar, that micro1's gross annual run rate had grown from $100 million to $500 million over eight months, and that micro1 also sells off-the-shelf datasets to multiple customers. Its founder said micro1 does not sell data to Chinese model makers.⁵⁵ Time and CSO Online reported that micro1 made a late $12.5 million offer for Spirit Airlines data after Google won the auction.⁸⁵⁶ We found no named company that says it was paid under the program.

Watch-outs. The program is days old as of this writing. A $1 billion budget is a ceiling on spending, not a count of deals or a promise that any one company will be paid. The tiers are published as ranges tied to qualitative factors, and the page does not publish payment timing. The run rate figure is reported from an unnamed person and micro1 did not respond to TechCrunch's request for comment.⁵⁵

What sellers should watch. The three tiers are anchors set by the buyer and tied to qualitative factors it judges itself. One published floor says nothing about what a second buyer would pay for the same records. Exclusivity and usage scope are not stated, and payment timing is not published.

Scale AI#

What they buy. Scale's data partnership page asks established companies to share de-identified records from six kinds of systems: data warehouses such as BigQuery, Snowflake and Databricks; chat tools such as Slack, Teams and Google Chat; email and calendar; documents with metadata and permissions; work and knowledge tools such as Jira, Confluence, Asana, Notion and Linear; and business platforms such as Salesforce, HubSpot, Workday, ServiceNow, SAP and Zendesk.³

What they pay. Scale gives an illustrative value of $10K to $1M or more per data partnership, scaling with cadence, and says it designs partnerships so there is opportunity for recurring revenue. A referral bonus of up to $50,000 is paid only if Scale signs a partnership agreement, and only one referral per deal is paid.³

Process and terms. Scale describes scoping the value of one workflow and locking governance before an ongoing partnership. It says it anonymizes everything inside your environment before extraction, with encryption in transit and at rest, per-package approval gates, a full audit trail and on-premises or VPC options. It says nothing is shared without your explicit approval of each package, and partnerships can begin with a limited scope and are not one-time transfers.³ We found no statement on that page of license terms, data ownership, exclusivity, or which companies receive the data.

Track record. In June 2025 Scale confirmed a significant investment from Meta that values it at $29 billion, reported at about $14.3 billion for a 49% stake, and its founder moved to Meta.¹³ Scale appointed Francis deSouza as chief executive effective August 10, 2026, and named BP and Mayo Clinic as recent enterprise customers.¹⁵ We found no public count of how many companies have signed data partnerships.

Watch-outs. After the Meta investment, OpenAI said it was winding down work with Scale, Reuters reported Google was discussing dropping it, and Scale's general counsel argued Meta would get no preferential treatment and would not receive other customers' confidential information.¹⁴ For a seller, the practical question is who the downstream recipients of your data are and whether any of them is a competitor of yours. The page does not answer it.

What sellers should watch. The $10K to $1M range is described by Scale as illustrative, so it is an anchor and not a quote. Recurring revenue depends on a cadence the buyer sets. License scope, ownership and exclusivity are not stated, and neither are the downstream recipients, which matters if one of them competes with you.

Handshake AI#

What they buy. Handshake's data partnership page lists financial models, process documentation, underwriting workflows, design documentation, production records, lab protocols, procurement workflows, product catalogs and technical documentation across eleven sectors, including finance and insurance, hardware and engineering, healthcare, legal, logistics, manufacturing, pharmaceuticals, professional services, real estate, retail and software.¹⁶ Its license page names sources such as AutoCAD, Dropbox, Figma, Gmail, Google Drive, Outlook, Monday.com, NetSuite, Notion, Paychex, QuickBooks, ServiceTitan and Slack.⁴ A Handshake blog post says Handshake purchases data directly from companies, with the aim of building real-world environments where AI models learn to execute tasks.⁵⁷

What they pay. Handshake says qualified partnerships typically range from $100K to $4M or more depending on the data, and pays a $10,000 reward to a referrer whose referred company becomes a partner.⁵⁷ The license page's estimate form says: "This is an indicative estimate only and is not an offer to purchase your data." Final payout depends on data sources, volume and quality, the outcome of due diligence and acceptance of the data licensing terms.⁴

Process and terms. Handshake says data is scoped, de-identified and kept confidential, each engagement is covered by an NDA, and the data owner can understand what is being licensed, how it can be used and how payment works before committing. Its license page produces an estimate from basic company details.⁴⁵⁷¹⁶ The data partnership page says you can review who may receive the data, how it may be used and the safeguards required. We found no published statement on exclusivity, term, retention or ownership.

Track record. TIME reported that Handshake's chief executive said the AI unit reached more than $1 billion in gross annualized revenue after 16 months, working with eight frontier labs and hundreds of thousands of contributors. TIME presented this as his claim, not independently verified.¹⁷ The data partnership page names no AI lab partners.¹⁶

Watch-outs. A payout estimate from a web form is not an offer, and Handshake says so. The program page does not name buyers, and the FAQ answer to which labs it works with was not shown on the page as we retrieved it.¹⁶ Handshake's revenue figure concerns its expert-labor business, not data purchases.

What sellers should watch. The range is an indicative estimate that the page says is not an offer, and final payout turns on due diligence the buyer runs. Exclusivity, term and ownership are unpublished, and the buyers are unnamed, so the number cannot be compared with any other offer.

Turing (Project Lazarus)#

What they buy. Project Lazarus asks for communication threads and channels where decisions were made; engineering history including source code, reviews, tickets and iteration history; wikis, runbooks, schemas and institutional knowledge; and the reasoning behind decisions. It lists work areas of software and engineering, finance and accounting, customer support, sales and CRM, legal and contracts, and operations and logistics.⁵

What they pay. The page advertises a single payout of up to $1 million, scaled by how complete, rare and well preserved the record is, and confirmed after sampling, usually within a week. The pricing section on the page shows "Up to $0.00M," which conflicts with the headline and looks like a display error. A referral reward of up to $10,000 is listed.⁵

Process and terms. Turing says it starts under NDA, reviews a representative sample, and then issues one valuation and one contract, with no retainer or recurring commitment. It says it works from a copy, removes personal data and secrets before anything reaches a model, processes in an air-gapped environment, and holds SOC 2 Type II. "Clean rights transfer" is listed as a feature, but the FAQ entry on who owns the data showed no answer.⁵

Track record. TechCrunch reported in March 2025 that Turing raised $111 million at a $2.2 billion valuation, with about $300 million in annualized revenue and a statement from its chief executive that it had been profitable for around a year. It began supplying code to OpenAI around 2022.¹⁸ Turing's company page lists 5M+ domain experts and a "9/9 frontier labs" statistic without naming the labs or defining the denominator.⁵⁸ We found no named Lazarus seller.

Watch-outs. Eight FAQ questions on the program page, covering ownership, pricing and eligibility, showed no answers in the version we read. "Rights transfer" and a "single payout" suggest a sale and not a license, which is a different legal act. Confirm in writing whether you are selling, licensing, or granting something in between, and what you keep.

What sellers should watch. A single payment paired with a "clean rights transfer" points to a sale and not a license. If it is a sale, the seller could no longer license the same records to anyone else. The advertised ceiling and the on-page zero conflict, the ownership answer is blank, and the price is one valuation from one buyer after sampling.

Group 2: AI data suppliers that mainly hire experts#

These companies are the best known in AI data. Their pages describe producing new data with paid experts, running benchmarks and environments, or selling datasets. None publishes a program to buy a business's existing records. We include them because business owners often ask, and so that the absence is visible. An expert marketplace is also a different transaction from a data sale: when your staff do contract work, you are selling labor, and the contractor terms, not a data license, govern.

Surge AI#

What they buy. Surge's homepage offers off-the-shelf frontier data and RL environments, post-training runs, an expert workforce and benchmarks. It mentions no data partnerships and names no customers.¹⁹ MIT Technology Review quoted Surge's head of reinforcement learning environments saying Surge looks for archives that cover disagreements and tricky judgment calls, which makes Surge a reported buyer through brokers, not through a public program.⁶

What they pay. None found. Process and terms. None found.

Track record. A Surge representative told Time that buying a failed company's data makes sense only if one believes an agent will generalize across the economy, and that flawed decisions in the data can be turned into training tasks.⁸

Watch-outs. If a broker says Surge will buy your data, ask for evidence, because Surge publishes no intake channel.

What sellers should watch. Surge publishes no program and no price, so any figure attributed to it comes from a third party. Its reported interest in archives with hard judgment calls shows what kind of data buyers value, but no amount is attached to it.

Appen#

What they buy. Appen's dataset site advertises over 300 audio, image, video and text datasets in over 80 languages as licensable. In the filtered view we reviewed, 36 text and image datasets were listed and every one showed its source as "Third Party Vendor."²⁰

What they pay. None found. The page gives no license terms or pricing, and buyers add datasets to a quote list and contact Appen.²⁰

Seller channel. None found. The page does not say how a vendor is added, and "Join our crowd" links to contributor work, not data sales.²⁰

Process and terms. Quote request only.

Track record. The catalog shows that Appen resells third-party data, which is the closest thing to a seller channel in this group, but it does not show that Appen buys proprietary business data.²⁰

Watch-outs. Resale of academic and reference corpora is not evidence of demand for your records.

What sellers should watch. Pricing is set per quote request and not published. A vendor dataset listing is resale, and exclusivity and license scope are not on the page.

Toloka#

What they buy. Toloka describes itself as a provider of expertly curated data for AI agents and models, with services such as preference collection, evaluation tasks and red teaming. Its page links to "Become a data partner" without saying whether Toloka purchases data.²¹

What they pay. None found.

Process and terms. Not stated.

Track record. Nebius announced on May 7, 2025 that Bezos Expeditions led an investment in Toloka, that Nebius would keep a significant majority economic stake but give up majority voting control, and that customers included Amazon, Anthropic, Microsoft, poolside, Recraft and Shopify.²² Toloka's own page says it works with data workers in more than 100 countries and was founded in 2014.²¹

Watch-outs. The "data partner" link is ambiguous. It is unclear whether it means supplying data, experts or tools.

What sellers should watch. The data partner link has no stated terms, so there is no price, scope or exclusivity to evaluate. It should not be assumed to be a purchase channel.

Snorkel AI#

What they buy. Snorkel's homepage describes expert-curated datasets, custom data development, specialized agents, benchmarks and environments. It does not say it buys or licenses data from companies. One environment type is described as "your tools, codebase, corpus, data and permissions," which points to working on a client's own resources.²³

What they pay. None found. Process and terms. Not applicable.

Track record. Snorkel announced a $350 million round at a $3.5 billion valuation on September 22, 2026 and said it surpassed $375 million in annualized run-rate revenue, a company claim. The announcement named no customers.²⁴

Watch-outs. Its growth is in commissioned expert data, which is not the same market as a licensed archive.

What sellers should watch. Its public model is commissioned work, where a client pays for data to be built. That is the reverse of a data sale, so its figures say nothing about what licensed records are worth.

Labelbox and Alignerr#

What they buy. Labelbox's homepage describes Recursion agents for enterprises, Horizon environments and evals, Terra robotics data and Alignerr expert judgment. It does not mention acquiring outside datasets.²⁵

What they pay. None for data. Alignerr's site advertises role rates of $25 to $30 an hour for some language roles and $40 to $120 for most others, an "average pay" of $80 an hour and 100,000 or more experts, while Labelbox's homepage claims 2.6 million or more skill-tested experts. The figures differ and are company claims.⁵⁹²⁵

Process and terms. Not applicable.

Track record. Labelbox announced a Series E of more than $100 million, including equity and a credit facility, at a $1.25 billion pre-money valuation on October 14, 2026, and claimed 400 or more customers and profitability.²⁶

Watch-outs. Expert-count figures conflict across Labelbox's own pages.

What sellers should watch. The pay figures here are hourly expert rates, which are labor prices and not data prices. They say nothing about what your records are worth.

Invisible Technologies#

What they buy. Invisible's AI training page lists expert domain training, model evaluation, localization in 80 or more languages, RL environments that simulate business workflows, multimodal data and red teaming. It does not say it buys or licenses data from businesses.²⁷

What they pay. None found. Process and terms. Not applicable.

Track record. Invisible announced $100 million in growth funding in September 2025, bringing total raised to $144 million, reported $134 million of 2024 revenue and about 350 staff, and did not disclose a valuation.²⁸

Watch-outs. Its public materials describe services and software, not a data acquisition channel.

What sellers should watch. Its pages describe services and software. No purchase terms or prices are published, so nothing here anchors a data price.

Group 3: Marketplaces, exchanges and intermediaries#

This group is the widest. It includes platforms that aggregate data for AI developers, brokers and resellers, wind-down services, and general cloud data exchanges that are not specific to AI. Fee models differ more here than anywhere else in this guide, and so do the roles. Some take your data, some only list it, and some only introduce you to a buyer.

Protege#

What they buy. Protege is an AI data platform that connects private data holders, such as those in healthcare and media, with AI developers. Its chief executive described licensing medical imaging, clinical notes and patient histories, video and footage, audio and motion capture.²⁹ Coverage of its January 2026 round lists clinical records, medical images, media, audio conversations and motion capture.³⁰

What they pay. No payout amounts are published. A company-stated claim in coverage of the round says Protege pays revenue-share payouts to partners when their data is used, and gives no percentage.³¹ The chief executive said one provider generated seven-figure net new licensing revenue in under a year, a claim not independently verified.²⁹

Process and terms. The chief executive described a governance layer that includes de-identification, and the January coverage says de-identification is offered where appropriate along with curation.²⁹³⁰ Exclusivity is not addressed in the sources we read.

Track record. Protege raised a $30 million Series A1 led by Andreessen Horowitz in January 2026, extending a $25 million Series A from August 2025, for about $65 million total, according to the company.³⁰ A vendor directory lists its funding as $70 million, so reported totals differ.⁶¹

Watch-outs. Terms, fees and eligibility are not public, and the only named customer is a medical imaging provider.²⁹

What sellers should watch. A revenue share with no stated percentage cannot be priced, and income depends on how often a buyer uses the data. Terms, fees and exclusivity are not public, and one intermediary that controls both the buyer list and the terms is a single point of pricing.

Troveo#

What they buy. Troveo's homepage lists business, video, audio, gaming, text and robotics data types. It advertises permissioned business data across more than 30 systems of record, in IT service management, Salesforce and CRM, customer support, financial operations and inventory, and invites companies to turn email, chat, CRM, tickets and files into recurring revenue.³²

What they pay. Not published. The homepage mentions price points and a floor in teaser text but gives none, and for your own data it asks for a "Get Estimate" request.³²

Process and terms. The page says Troveo sources data from its licensor network to the buyer's spec. It claims 350 or more companies, 7,000 or more data providers, and 8 million or more hours of video, all self-reported.³²

Track record. Bloomberg reporting relayed by The Decoder in January 2025 said OpenAI, Google and others were buying creators' unpublished video at $1 to $4 per minute through intermediaries including Troveo AI and Calliope Networks.⁶²

Watch-outs. Troveo's track record is documented in video. We found no public report of a business being paid for email, chat or CRM data through Troveo.

What sellers should watch. The estimate is a number the platform sets from its own buyer pool. Business-data pricing and exclusivity are not published, and the documented track record is in video.

Defined.ai#

What they buy. Defined.ai runs an AI training data marketplace where enterprises buy or commission data, with ready-made datasets in speech, text, image, video and multimodal. The footer lists a "Data Partner" track with no details.³³

What they pay. Not published by Defined.ai. Reuters, relayed by The Decoder in April 2024, reported that Defined.ai charged $1 to $2 per image, $2 to $4 per short video, $100 to $300 per hour of longer film and $0.001 per word of text to large technology buyers. Those are what it charged, not what it pays suppliers, and they are dated.³⁴

Process and terms. The page does not say how suppliers are paid.³³

Track record. The same Reuters report says it licenses to large technology companies.³⁴

Watch-outs. Supplier terms are not public, and the unit prices are for media and generic text.

What sellers should watch. The published unit prices are what it charged buyers for generic media and text, not what suppliers are paid. Supplier terms and exclusivity are not public.

SimpleClosure Asset Hub#

What they buy. SimpleClosure's Asset Hub sells source code and workspace data, meaning communications, documents, designs, tickets and workflows from tools such as Slack, Drive, email, Notion and Figma, to a vetted network of buyers that it describes as AI foundation labs, RL environment builders and other technology companies. It says every dataset is processed for personal information before it is licensed to a buyer.³⁵

What they pay. Forbes, relayed by Gizmodo, reported payouts of $10,000 to $100,000 per company across nearly 100 deals in a year.⁷ MIT Technology Review reported more than 150 deals, that 47% of SimpleClosure's deals now include code or data sales, and that one crypto startup received about $25,000 for its codebase with about $15,000 more expected for workspace data.⁶ A SimpleClosure release said Asset Hub is free for dissolution customers and that access now extends to companies at any stage. Fees for others are not stated.³⁶

Process and terms. The page describes a viability assessment followed by buyer interest. It says SimpleClosure does not offer legal or financial advice, and its FAQ answers on cost and value were not shown.³⁵

Track record. The same release says AI buyers grew ninefold compared with 2025 and that the service supported more than 350 companies, which counts participants and not completed sales.³⁶

Watch-outs. MIT reported privacy objections, including a lawyer who called such data "legally radioactive," and noted there is no federal law banning such sales.⁶ Gizmodo quoted a privacy advocate saying "It's identifiable people."⁷

What sellers should watch. The reported $10,000 to $100,000 payouts come from companies that were closing, which is a different bargaining position from an operating business, so they are a weak anchor. Buyers are unnamed, exclusivity is not stated, and privacy objections are a real legal risk for the seller.

Sunset#

What they buy. MIT Technology Review described Sunset as a rival of SimpleClosure with a fast-growing data resale business, whose chief executive said it values a client's assets by headcount, years in operation, volume of data and how the team used its tools.⁶

What they pay. Not published. Process and terms. Not stated.

Track record. We found no primary Sunset page describing its data business, so this entry rests on one news report.⁶

Watch-outs. Treat claims about Sunset as unverified until you see its terms.

What sellers should watch. A value set by headcount, years and data volume is the broker's formula, not a market price. Its terms are unpublished.

Petrarch#

What they buy. Petrarch's Y Combinator profile says it sources internal company data, including codebases, project files, payment records, construction plans and autonomous driving recordings, de-identifies it, certifies legal clearance and resells it to applied AI startups and labs, starting with manufacturing and industrial companies.³⁷

What they pay. The profile does not say how suppliers are paid.

Process and terms. Not stated.

Track record. Petrarch is in the Summer 2026 batch, was founded in 2026, and has a team of three.³⁷

Watch-outs. Very new and unproven, with no completed transactions or references published.

What sellers should watch. Because Petrarch resells, the gap between what a buyer pays it and what a supplier receives is not disclosed. How suppliers are paid and whether the data is exclusive are not stated.

Korra#

What they buy. Korra is a marketplace where you describe your records and several approved buyers make offers. It says buyers want records that show work from start to finish, such as Slack threads, Jira tickets, support cases, sales notes and email, and advises leaving out direct messages, HR and pay records, legal advice, health details and client-held material.³⁸

What they pay. Offers come from buyers, and the page gives no figures. It says you pay Korra nothing and the buyer pays Korra's fee.³⁸

Process and terms. Korra says nothing is uploaded when records are described, that it removes names, emails, phone numbers and towns from the listing card, that the seller sees every offer and approves one or none, and that the license is between the seller and the buyer. It says it does not take, open, clean or deliver the records.³⁸

Track record. None published.

Watch-outs. Because Korra does not clean or deliver the records, de-identification and delivery remain your job, or a third party's.

What sellers should watch. Several competing offers is the structure that protects a seller, but the license is between seller and buyer, so scope, exclusivity and de-identification terms still have to be negotiated. Korra says it does not clean or deliver the records.

AI26#

What they buy. AI26 connects businesses with AI research labs interested in licensing historical operating records: company email, Slack, Teams and business WhatsApp conversations, documents, spreadsheets, work records and technical knowledge. Personal conversations and employee-owned accounts are not automatically eligible.³⁹

What they pay. The page advertises "upto $500k in payouts" and states in the FAQ that there is no fixed payout. Any payment depends on buyer approval and a licensing agreement, and the opportunity is described as a one-time license.³⁹

Process and terms. The page lists a 60-day turnaround. It does not mention fees charged to the business, and says the buyer and permitted uses are set out in the licensing agreement.³⁹

Track record. None published.

Watch-outs. The headline number and the FAQ disagree, which is the pattern to watch for in all marketing pages. The geographic restriction is narrow.

What sellers should watch. A published ceiling with no fixed payout leaves the number to the approving buyer. A one-time license with a 60-day turnaround also means one buyer's valuation and one chance to get the terms right.

AWS Data Exchange, Snowflake Marketplace and Databricks Marketplace#

What they buy. None of the three buys data. They are venues where providers list data products for customers to subscribe to. AWS says its catalog holds more than 3,500 third-party data sets from more than 300 providers.⁴⁰ Snowflake describes free, limited trial and paid listings.⁴² Databricks describes public listings through its Data Partner Program and private exchanges for members.⁴³

What they pay. The providers' revenue comes from customers, not from the platform. None of the three pages we read states fees to providers.⁴⁰⁴²⁴³

Process and terms. AWS says providers publish products and offers, optionally verify subscribers and create custom offers, and receive funds distributed by AWS Marketplace.⁴¹ Snowflake lets providers define pricing plans and offers with payment terms and schedules.⁴²

Track record. These pages do not say how many listings are bought for AI training, so they are infrastructure and not evidence of AI demand.

Watch-outs. A listing is a shop window. Nobody is obliged to buy, and the work of packaging, rights review and de-identification falls on you.

What sellers should watch. A listing lets the provider set a price, but nobody has to buy, and buyers see a catalog entry and not a negotiated offer. Exclusivity is not part of the model and provider fees are not stated on the pages we read.

Datarade#

What they buy. Datarade is a data marketplace where registered businesses list data products. It does not buy data.⁴⁴

What they pay. Providers pay. Datarade's provider page lists a commission-only tier at $0 a year plus a 30% commission, Bronze at $6,000 a year plus 20%, Silver at $12,000 plus 15%, and a custom Gold tier.⁴⁴

Process and terms. Track record. We did not find a source on AI-training sales through Datarade.

Watch-outs. This is one of the few places where a fee is public, and it flows from you to the platform. Compare that with models in which the buyer pays.

What sellers should watch. Subscription tiers are paid up front and commission applies to sales, so compare net proceeds and not list price. The platform is paid whether or not a lab ever buys.

Human Native and Created by Humans#

What they buy. Both are creator and rights-holder marketplaces, not business-records buyers. Cloudflare acquired Human Native, a marketplace where publishers, creators and rights holders prepare and price content for AI use, on January 15, 2026, for an undisclosed price.⁴⁵ TechCrunch reported in 2024 that Created by Humans lets authors license works to AI companies, takes a cut of each deal and had raised $5 million.⁶³

What they pay. Not published. Created by Humans did not state its cut.⁶³

Process and terms. Not stated.

Track record. Human Native raised $3.56 million before the acquisition.⁴⁵

Watch-outs. They are relevant only if your business owns publishable content, not internal records.

What sellers should watch. These apply to content libraries only. Cuts and rates are not published.

Group 4: AI labs licensing directly#

Here we include only what is documented. The labs below have signed publisher and platform deals, and one lab announced a data partnership program in 2023. We found no AI lab that currently publishes an intake page inviting ordinary businesses to sell internal records, and we did not find such a program from Anthropic, Meta or xAI. A lab's reluctance to publish does not mean it never buys: Google won the Spirit Airlines data auction, and MIT Technology Review reported that SpaceX's AI unit was reportedly weighing purchases from struggling startups, according to Bloomberg.⁸⁶

OpenAI#

What they buy. In November 2023, OpenAI announced Data Partnerships, seeking large-scale datasets that reflect human society and are not easily accessible online, especially data that expresses human intention such as long-form writing or conversations. It said it would help digitize data and remove sensitive or personal information when necessary, and offered open datasets anyone can use and private datasets for organizations that want confidentiality.⁴⁶

What they pay. TechCrunch's report described no compensation terms.⁴⁶

Process and terms. Not stated.

Track record. Named partners at launch were the Icelandic Government and Miðeind, and the Free Law Project.⁴⁶ OpenAI has also signed publisher and platform deals, including Stack Overflow's OverflowAPI partnership in May 2024 with no disclosed financial terms, and News Corp, which the Wall Street Journal said could be worth close to $250 million over five years, a reported figure neither company confirmed.⁴⁸⁴⁷

Watch-outs. The 2023 program page is not a current intake form, and we could not verify its present status.

What sellers should watch. The 2023 announcement states no price and we could not confirm current intake, so it gives no anchor. The reported publisher deals were priced by rights holders with leverage, not by a posted program.

Google, Microsoft and Amazon#

What they buy. Google: Reuters reported in February 2024, citing three people familiar, that Reddit's AI content licensing deal with Google was worth about $60 million a year, and Reddit's own filing disclosed $203.0 million of aggregate contract value for arrangements signed in January 2024 with terms of two to three years, without naming buyers.⁴⁹⁶⁴ Google also won the Spirit Airlines data auction with a $10 million bid.⁸ Microsoft: its Publisher Content Marketplace pays publishers on delivered value through usage-based reporting for grounding scenarios, with publishers setting their own terms, and the post describes grounding only.⁵⁰ Amazon: TheWrap, citing the Wall Street Journal, reported that Amazon will pay The New York Times between $20 million and $25 million a year for news, NYT Cooking and The Athletic, for training and Alexa summaries, a figure not confirmed by the companies in that report.⁵¹

What they pay. See above. All figures are reported, not published.

Process and terms. Publishers define their own terms in Microsoft's marketplace.⁵⁰

Track record. Large sums go to large rights holders with distinctive, continuously updated content. The documented examples are Reddit, news groups and a bankrupt airline, not typical operating companies.

Watch-outs. These deals show that large buyers pay for content and data, but they are not comparable to a small business's records, as the pricing papers in this series argue. See comparing offers.

What sellers should watch. The large numbers belong to big rights holders and a bankruptcy auction, and they are reported and not published. They show what competition and unique content can bring, not what a typical operating company would be offered.

Comparing offers#

Published, reported and illustrative figures side by side#

No two figures below use the same unit, term and rights. Read the "basis" column first, and then the amount.

What it costs to list#

How buyers reason about value#

Deloitte describes three approaches to valuing data assets: income, cost and market. It says the market approach is hard to apply because reliable transaction data are limited and many deals do not disclose value drivers, so it works best as a reasonableness check.⁶⁷ Reporting on failed-startup sales fits that logic. MIT Technology Review quoted a Surge executive saying valuations rest on research theses that estimate how many useful tasks a body of data could yield.⁶ Sunset says it prices by headcount, operating history, data volume and how the team used its tools.⁶ For you, that suggests that data a buyer could cheaply commission from experts is worth less than data it cannot reproduce. This is an interpretation of the sources, not a stated rule.

Terms that matter more than headline price#

A $500,000 offer on terms that hand over everything forever can be worse than $100,000 on narrow terms. These are the terms to check, and what the public record says about each. Several sources are law-firm commentary written from a buyer's side, so read them as checklists, not as market norms.

Exclusivity#

Whether you can license the same data to others changes both price and your future. A Malaysian law firm advises settling exclusivity up front and, for licensors, keeping the right to license to third parties or use the data to train their own AI.⁶⁸ Epoch AI's interviewees estimated exclusive RL environment deals cost roughly four to five times as much as non-exclusive ones, from two founders.⁶⁰ None of the five program pages publishes an exclusivity term.

Field of use#

Sidley advises defining which model versions may be trained, whether use is internal or commercial, and whether fine-tuned derivatives are covered.⁶⁹ A practitioner memo suggests a single named model with the right to extend to others on an additional fee, and for most enterprise licensors expects one model.⁷⁰

Term and the model after the license ends#

The same memo recommends a perpetual license for training and using trained models paired with a finite license for ongoing access to the dataset.⁷⁰ Sidley notes a trained model generally keeps working after a license expires, so audit and deletion duties should be negotiated in advance.⁶⁹ The Malaysian firm warns that removing data from a trained model may be infeasible or may require costly retraining, so a licensor may have to accept that the model retains it.⁶⁸ That means "we will delete it at the end" cannot be assumed to cover the model.

Who removes personal information, and who pays#

The program pages split this work differently. Mercor says extraction and anonymization are automated and it covers the costs.¹ Scale says it anonymizes inside your environment before extraction.³ micro1 says personal information is removed when relevant and some data may be synthetically rewritten.² Handshake says you can understand who will access and prepare the data, including required de-identification.¹⁶ Korra leaves cleaning and delivery to you.³⁸ In the Spirit sale, the buyer was solely responsible for de-identification costs, which did not reduce the price, and chose the agent that did the work.⁷¹

For legal standards, HIPAA recognizes two de-identification routes, expert determination and removal of 18 identifier types with no actual knowledge that the remainder could identify someone.⁷² Under the GDPR, pseudonymised data is still personal data if it can be linked back with additional information, and anonymity is judged by all the means reasonably likely to be used.⁷³ Neither is a standard for business text, and neither replaces a privacy professional's review.

Audit and deletion#

Best Lawyers suggests clear retention and deletion rules covering backups, caches and shadow copies, a certificate of destruction, and ongoing audit rights with audit logs and reporting.⁷⁴ Sidley says deletion and audit duties may survive termination.⁶⁹ micro1 says data is kept for agreed periods and deleted on request or at the end of the engagement under the agreed terms.² Scale and Mercor describe audit trails for access, not audit rights over the buyer's later use.³¹

Ownership of derived data and models#

Vorys notes that raw facts are generally not protected by copyright and ownership between parties is mostly set by contract, so it recommends specifying permitted uses, with each data type and output treated separately.⁷⁵ Sidley says model ownership and derived datasets should be allocated explicitly because whether a trained model is a derivative of its training data is contested.⁶⁹ micro1 says companies retain ownership of their underlying data, while Turing lists "clean rights transfer" and shows no ownership answer.²⁵

Warranties and indemnity#

Sidley says provider representations that you have the rights to license, and indemnity for third-party claims, are increasingly standard in well-negotiated AI development agreements.⁶⁹ That is why the rights review has to come first. The FTC has said companies must honor privacy commitments wherever they were made, that there is no AI exemption from the laws on the books, and that it has required deletion of models and algorithms built on unlawfully obtained data.⁷⁶

Payment timing and form#

Timing is stated rarely. Mercor says wire payment typically follows two to four weeks after the signed agreement, on delivery of the anonymized dataset to the buyer.¹ Turing says the valuation follows sampling, usually within a week.⁵ AI26 lists a 60-day turnaround.³⁹ Scale describes recurring revenue scaling with cadence, and Protege claims revenue share when data is used.³³¹ micro1, Handshake and Scale tie final payment to review and acceptance.²⁴³ Ask whether the number is paid at signing, at delivery, or over time, and whether it can be reduced after due diligence.

What buyers say they want, by data type#

The matrix lists only buyers whose pages or cited reports name the data type. Blank means no cited source. A name in a cell means stated interest or a reported transaction, not a promise to buy anything.

A few cautions apply. Most of these buyers also describe de-identification as part of the process, so the rights and privacy questions come before any question of demand. Healthcare data is the most regulated category in the table. Marketing data may well have buyers, but we found no public statement of interest, so we cannot say. And the strongest public record of actual purchases concerns code, messages and email from failed companies, which are the categories that have appeared in reported sales.⁶⁷

Red flags and due diligence questions#

Red flags#

• A value quoted before anyone has seen a sample. Handshake's own estimate form says it is not an offer, and Turing says the amount is confirmed after sampling.⁴⁵
• A headline that the fine print contradicts. AI26 advertises up to $500k and says there is no fixed payout, and Turing's page shows both "up to $1M" and "Up to $0.00M."³⁹⁵ Take screenshots with dates.
• Requests for credentials or broad access before a contract. Scale describes zero bytes shared without per-package approval, and Mercor says no data leaves your systems until you approve it. Use those statements as the standard to hold any buyer to.³¹
• No ownership, license or exclusivity terms anywhere. Most program pages omit them.¹³⁴
• No named buyer, no named completed deal. We found none for any of the five programs.
• A fee you pay to list. Datarade's commission model is public and runs from 15% to 30% plus subscription.⁴⁴ Check whether any intermediary charges you before a sale.
• Rights you cannot confirm. The FTC says it may be unfair or deceptive to loosen data practices to allow AI training and tell users only through a quiet retroactive terms change.⁷⁷
• Third-party "advice" that carries a referral fee. Scale and micro1 pay referral rewards of up to $50,000, Handshake $10,000 and Turing up to $10,000, and Mercor pays an unstated bonus.³²⁵⁷⁵¹ A recommendation from someone who earns that fee is not neutral, and that includes us.
• A buyer with an unresolved security incident and no documentation. Mercor disclosed one and says it completed its investigation.¹⁰

Due diligence questions#

1. Who are the intended buyers or categories of buyer, and can any be excluded, including competitors of ours?
2. Where does de-identification happen, who pays for it, and does any raw copy reach the vendor first?
3. Will linked records stay linked, and what is the re-identification risk that remains?
4. Is the arrangement a sale, a license or something in between? For what term, what field of use and which models?
5. Is it exclusive, and what happens to the exclusivity premium if it is?
6. What do you keep: ownership of the original data, rights to derived datasets and to models?
7. What happens at the end: deletion, certification, audit rights, backups and caches, and the model?
8. What do you warrant, and what indemnity do you give? Does our insurance cover it?
9. When and how are we paid, and can due diligence reduce the amount after signing?
10. What security evidence can you show? Ask for SOC 2 reports, incident history and the sub-processors that touch data. Mercor and Turing say they hold SOC 2 Type II, which is a vendor statement until you see the report.¹⁵
11. Which of our contracts, privacy notices and employee agreements allow this? The Data & Trust Alliance's provenance standard defines 22 metadata fields, including consent documentation location and license to use, as a voluntary way to document origin and permitted use.⁷⁸
12. If we are closing, what rules apply? In a US bankruptcy, a trustee generally may not sell personally identifiable information against the debtor's privacy policy unless the sale is consistent with it or a consumer privacy ombudsman is appointed and the court approves.⁷⁹ The FTC chair wrote in March 2025 that consumers should be able to trust that companies keep their promises, in a letter about the 23andMe bankruptcy.⁸⁰

The Spirit Airlines sale as a case study#

The Spirit sale shows these terms in action. Google opened at $5 million, Mercor countered and offered more if it could take the raw data and anonymize it itself, and Google closed at $10 million, with Google solely responsible for de-identification costs and choosing the agent.⁷¹ The sale excluded customer profiles, call recordings and chat sessions, according to The Next Web, though CSO Online described the data differently, including customer service calls, and we could not reconcile the two accounts.⁷¹⁵⁶ A union objected on privacy grounds and a court postponed its hearing to September 9.⁸¹ Court-appointed consumer privacy ombudsman Lucy Thomson recommended approval on October 5 for de-identified data, with booking and transaction records removed from the sale, Tonic.ai hired to de-identify the unstructured data, certification before every transfer and a public pledge by the buyer not to re-identify.⁸² We could not confirm a final ruling.

Why one buyer should never set your price#

Every price on a program page is one buyer's number, chosen by that buyer, tied to factors that buyer judges, and described by the buyer as indicative. It is an anchor. It is not a measure of what the market would pay, because on that page there is no market, only one bidder.

Competition moves price#

The one large documented auction shows the gap. In the Spirit Airlines sale, Google opened at $5 million, Mercor countered, and the sale closed at $10 million.⁷¹ Time reported a $7.5 million backup bid from Mercor, and Time and CSO Online reported a late $12.5 million offer from micro1.⁸⁵⁶ That was a bankruptcy sale, so it is not a forecast for an operating company. It does show that the first number a buyer names and the number a contested sale reaches can be far apart.

Non-exclusive licensing to several buyers#

A buyer that asks for exclusivity is asking for something it values. Two founders told Epoch AI that exclusive deals cost roughly four to five times as much as non-exclusive ones.⁶⁰ A Malaysian law firm advises licensors to keep the right to license to third parties.⁶⁸ None of the five program pages publishes an exclusivity term. Granting exclusivity to the first buyer without pricing it gives away the premium and the option to license the same records to others.

Contract terms are part of the price#

The same dollar figure means different things under different terms. Field of use, term, deletion, ownership of derived datasets and models, warranties and payment timing are covered above. Sidley notes that a trained model generally keeps working after a license expires, so a perpetual grant and a one-year license are different sales at the same number.⁶⁹ A single buyer's form contract gives you nothing to compare it with.

Stay anonymous until terms are agreed#

A buyer does not need your company name, system list or sample records to discuss scope, scale and terms. Korra, for example, says it removes names, emails, phone numbers and towns from its listing card, so buyers see a description and not an identity.³⁸ Sharing identifying detail early tells one buyer about your data and your leverage before it has committed to anything.

Using a broker or going direct#

Sellers deal with a buyer directly or through a broker. Going direct keeps any fee at zero and gives you the buyer relationship, and it puts the rights review, de-identification and negotiation on you or your counsel, with one buyer's offer as your only reference point. Several buyers pay a referral reward when an introduced company signs, which shows that introductions are already part of how this market works.³²⁵⁷ Our guide to selling direct versus through a broker sets out the trade-offs.

A broker can run a competitive process, screen buyers, run the rights and privacy review, prepare data and negotiate. A broker also has a conflict of interest: most are paid when a deal closes. Fee models in this guide differ. Korra says the buyer pays and sellers pay nothing, Datarade charges providers a subscription and commission, and SimpleClosure says Asset Hub is free for its dissolution customers.³⁸⁴⁴³⁶ Questions to ask any broker, including us:

• How are you paid, and what percentage? Do you also take a referral reward from a buyer?
• Who owns the buyer relationship and the right to resell the data?
• What rights and privacy review do you do, and who signs off?
• How do you de-identify, and what happens if re-identification is found later?
• May I use my own lawyer, and can I see your agreement before I send data?
• What exclusivity do you require from me?
• Can you name completed deals? If not, say so plainly.

What we could not verify#

• Whether any business has been paid under the Mercor, micro1, Scale, Handshake or Turing programs. We found no public statement from a named seller.
• Ownership, license, exclusivity and retention terms for Mercor, Scale, Handshake and Turing. The pages we read do not state them, and Turing's FAQ answers were blank.
• Handshake's FAQ answer on which labs it works with, which was not displayed on the page we read.
• The final outcome of the Spirit Airlines sale. Reports differ on hearing dates, a later report that we could not open described a further delay, and we found no ruling.
• Fees on AWS Data Exchange, Snowflake Marketplace and Databricks Marketplace, which their pages do not state.
• Sunset's terms, which appear only in one news report.
• Revenue, valuation and expert-count figures for most companies, which are company claims or reports citing unnamed people.
• Current rates: Mercor's advertised average changed between reviews, and company pages change without notice.

Where to go next#

Already talking to a buyer? We review offers free: see offer review. Otherwise, start by listing your data sources and the contracts behind them. Our free valuation asks for a short profile with no files or raw data, and how it works explains the process. For specific data types, see customer service data, communications, documents and SOPs, workflows, CRM history, code, marketing and healthcare. For background on rights, privacy and deal structure, read the guide to selling data to AI companies, and see our comparisons of several of the companies above, including Mercor vs micro1, Mercor vs Turing and Handshake AI vs Mercor.