micro1, Handshake, Turing, Mercor: Data Programs Compared

micro1, Handshake AI, Turing and Mercor all pay companies for their operating records, and on price they look alike: micro1 publishes tiers of $100K+, $500K+ and $1M+, Handshake a range of $100K to $4M+, Turing a single payment of up to $1M, and Mercor no figure at all. What separates them is everything after the price. micro1's published master terms take a licence that is perpetual, irrevocable and paid in full once, with the right to pass it on through multiple tiers. Turing acquires "defined rights" while the holder keeps ownership. Handshake binds itself by contract to training and evaluation only. Mercor publishes no licence term but pays by wire typically 2 to 4 weeks after signing. And two of the public estimators run the same formula, so the first number a holder sees says more about the door it knocked on than about its data.

This guide is for the founder, CFO or counsel of a company that has been approached by one of these programs, or that is deciding which to answer. It compares every buyer that publishes a program for buying or licensing company records, including the newer ones (Mode's BizBuyData, Labelbox's Prism, Polyshares, Appen, Centific and others), on four things: the price bands each publishes, who it takes, the licence it asks for and how and when it pays. Then it explains what those terms do to a holder's leverage, the questions to put to any of them before signing, and why a holder that puts its lot before every fitting buyer at once usually does better than one that answers a single program. For the step-by-step process of a sale, read the guide to selling company data to AI labs; this guide goes deeper on the programs themselves.

AIData.Broker acts as the holder's agent: the holder itself licenses or sells its data to the buyer, and we do not take ownership of it. Only the holder pays us, a fee fixed in the mandate, only if a sale closes. Every term below is the buyer's own published wording or named press, cited to its record; where a line is our reading rather than the buyer's words, it says so.

This guide is research from public sources, not legal, tax or investment advice.

Contents

1. The programs at a glance
2. Program by program
3. The licence each program asks for
4. Payment timing and process
5. What the terms mean for your leverage
6. One program or every fitting buyer at once
7. Questions to ask before signing
8. Questions people ask
9. What this means for you

The programs at a glance

A year ago almost nobody published a way to sell a company's internal records to the AI industry. By October 2026 our buyer board lists 37 buyers with a public way in, 24 of which take a company's records, and 25 publish some figure for what they pay. The programs below are the ones a holder of company records is most likely to meet: the four large vendors named in this guide's title, the two newer programs that most often appear beside them (Mode's BizBuyData, launched in October, and Kled, a consumer app that pays individuals), and the dozen smaller programs that publish enough to compare.

Most of these buyers are principals: they license or buy your records for their own account, turn them into training data or training environments (simulated workplaces where an AI model practises tasks), and sell the result to AI labs. A few are marketplaces that license on your behalf for a share of each sale. The difference matters for everything that follows, because a principal's published figure is the opening of its negotiation, not a measure of what labs will pay for your records.

The six programs this guide compares in depth

The other programs that publish terms

Read these tables as dated readings, not as a price list. Every buyer evaluates a sample before it names a price, and several of these programs changed their pages more than once in the month before this guide was written: Ooak raised its headline from $300K to $1M, and Handshake removed the sentence that once said who carries the legal risk of the transfer. The live version of both tables, with each buyer's source and the date it was read, is the buyer board; the buyer match tool runs your company's size, region, language and status against each buyer's published rules.

Program by program

The four programs in this guide's title are run by four of the largest suppliers of human data to AI labs, which is why holders meet them first. Each started by paying experts and contractors to write tasks, answers and judgments for labs (see the guide to AI training data companies for that business), and each added a line that buys records companies already hold. That history shapes their terms: they are building inventory to resell, and their contracts read like it.

The smaller programs are often more specific about the terms a holder cares about, and some publish more holder-friendly ones. They also carry more counterparty risk, which is why each profile below says what someone other than the buyer has confirmed about it.

micro1: the clearest tiers and the strongest licence

micro1 publishes the most structured price ladder in the market: $100k+ for a qualified partnership, $500k+ for large-scale datasets or participation across several teams, and $1M+ for highly unique, proprietary data. It wants SOPs, knowledge bases, internal documents, CRM data, project histories and decision records, from operationally mature companies of 30 or more employees that work and document in English, the US first. A separate posting of 8 October 2026 takes companies that have shut down, with 50+ peak employees and substantial Teams or Slack history.

The money behind it is real but smaller than the headline. On 9 October micro1 announced a commitment to spend $1 billion over the next 12 months on company data, financed with capital from Citi and Hercules Capital whose size and form neither micro1 nor the press states. The one piece of that financing anyone else has filed is Hercules' own quarterly report, which lists a $10.875 million senior secured term loan to micro1. micro1 itself reports passing a $1B annual run rate and priced its Series B at a $4B valuation. In the Spirit Airlines bankruptcy it told the court it had closed more than 50 data transactions in 45 days, a self-reported figure no counterparty has confirmed.

micro1's payout estimator is worth reading for what it reveals. For a US company of 100 employees and 10 years it shows $379,290 to $888,088, and the same inputs in Europe show $284,468 to $666,066. Handshake's estimator returns almost exactly the same numbers for the same inputs, because the two run the same formula: a fixed $140,000, plus $2,333 per employee and $11,667 per year of records. Two estimators that agree are one anchor, not two valuations.

The terms are where micro1 differs most. Its published data licence master terms take a worldwide, royalty-free, fully paid-up, non-exclusive, irrevocable, perpetual licence, with the right to sublicense through multiple tiers. The holder warrants it has not already licensed the same data on similar terms to anyone. If micro1 finds the data manifest materially inaccurate, nothing is payable and the holder refunds what was paid, even after acceptance. Deleting the data does not require micro1 to delete or retrain any model. In the holder's favour, micro1 may not resell the data as received as a stand-alone data product. It removes personal information itself "when relevant" and may rewrite some datasets synthetically. TechCrunch reported gross margins as high as 80% to 90% on the off-the-shelf datasets micro1 resells, which is the economics a perpetual, sublicensable licence makes possible.

Handshake AI: the widest range, fast pricing, a narrow use

Handshake AI is best known for paying experts to train models; its AI business passed $1 billion in gross annualised revenue in 16 months, by its founder's account to TIME. Its data-partnership page says "Payouts range from $100K to $4M" and takes companies with 20+ full-time employees and 3+ years in operation, in any industry. It states it has paid $250M to more than 1K partner companies, with no period given, a figure no data holder has publicly confirmed.

Handshake publishes the fastest pricing promise of the four: an initial review and price within 12 to 24 hours of the details it needs, and a full process that typically takes 7 to 14 business days, with payment tied to delivery milestones set in each agreement. Its use clause is the narrowest of the large programs: Handshake is "contractually restricted to using the data for AI model training and evaluation only", and the holder keeps ownership. It removes direct and indirect identifiers itself, before use, and says plainly that no de-identification process eliminates the risk of re-identification.

Two things are missing. Handshake publishes no licence term, no exclusivity rule and nothing on who pays for de-identification. And its page no longer says who carries the legal risk of the transfer, so the contract has to. Separately, Handshake has paid individuals $6 per accepted page, up to $30,000, for professional documents they own, which is a reminder that a buyer may reach a company's knowledge through its employees as well as through the company.

Turing (Project Lazarus): one payment, defined rights, code first

Turing runs Project Lazarus, which buys "defined rights" to a company's prepared working record under a single agreement for "a single payout of up to $1M". It takes operating, winding-down and closed companies, with the strongest demand in the US, UK, Canada and EU. Its process is the quickest to start: an intake call under NDA within one business day, a sample review, then a final offer and one payment, with no retainer and no recurring commitment.

Turing started with code and still prices it in the most detail. Its process page says repos typically pay $10k to $100k+ depending on quality and complexity, for repositories of 50,000+ lines, human-written, with tests. An earlier newsletter priced a modern codebase at $1,000 per pull request, up to $10K per codebase, and company documents such as tickets and specs at $50 per artifact when bundled with a codebase. Its CEO told The Information it had bought five to ten failed-startup codebases at an average in the tens of thousands of dollars each, which shows how far below the ceiling ordinary deals settle. If code is most of what you hold, read the guide to selling a codebase to AI companies.

On terms, Turing says "You retain ownership of your underlying data", removes personal data and secrets itself in isolated, air-gapped processing after the holder approves the scope, and publishes nothing on licence length or exclusivity. One sentence in its intake form deserves a question: it asks for your industry because it "Helps us match your submission with relevant buyers", which suggests the rights it acquires are meant to travel. Turing raised a $111 million Series E at a $2.2 billion valuation and reported around $300 million in ARR in March 2025.

Mercor: no price, the clearest payment

Mercor runs the opposite model to micro1. It publishes no figure, only that "Compensation is based on data volume, the tools you connect, and the depth of your data", and names eight target industries rather than a size rule. Its payment term is the most concrete of the four: a wire "upon delivery of the anonymized dataset to the buyer", typically within 2 to 4 weeks of the signed agreement. Extraction and anonymisation are automated and paid for by Mercor.

Mercor's published silence on the licence is the term to press. Its page says "We license it, package it, and pay you for it", its own catalogue offers every dataset to labs for licensing, and its intake asks the holder to confirm the materials are ones it may share "without restrictions". Nothing published says how long the licence runs, whether it is exclusive or whether Mercor may sublicense. Mercor is the $7.5 million alternate bidder for Spirit Airlines' records, and in that auction it bid $5.2 million with an independent scrub or $7 million if it could de-identify with its own tools, which tells you what control of de-identification is worth to it. It raised $350 million at a $10 billion valuation in October 2025 and was reported at over $2 billion in gross revenue run rate by mid-2026. Its record on our board notes that no seller has yet confirmed a sale of company records to it.

Mode (BizBuyData): the newest, with a public filer behind it

Mode (BizBuyData) opened on 13 October 2026. Mode Mobile, which files with the SEC, buys an agreed, filtered copy of an organisation's operating records and licenses it on for AI research and training. Its home page says a company's operating history "could be worth $100K-$5M", level with the highest figure any buyer on the board publishes ( $5M). Its estimator shows $290K to $570K for a US company of 100 desk employees with 10 years of records, and $130K to $270K for 20 employees and three years.

Its eligibility rule is unusually fine-grained: 20+ office employees, accounting firms 10+, law firms 6+, teams primarily in the US, nonprofits included. Its process is published step by step: a nonbinding letter of intent within 14 days, collection and de-identification by Mode over 15 to 30 days once the final offer is accepted, and payment expected in about 60 days. Mode publishes no licence or warranty terms yet.

Because Mode files publicly, a holder can check its counterparty in a way it cannot with most buyers. Its first-half 2026 filing reports $26,495,467 of revenue from its other business, a net loss of $3,799,464 and cash of $15,503,444 at 30 June, and it discloses doubt about continuing as a going concern, which management says is alleviated. Its launch, a paid wire release, headlined a $5 billion initiative; read that as the release's own description, not as capital set aside.

Kled: a consumer app, and the strictest licence on the board

Kled is not a company-records program: it is an app where individuals upload photos, video, documents and files, and are paid per task, for example $5 for a video, $150 for a tax return, $1,000 for 500 selfies. It belongs in this comparison for two reasons. First, its terms show the far end of what a buyer can ask: a transferable, worldwide, perpetual and irrevocable licence, sublicensable through multiple tiers, exclusive to Kled, and a duty to refund what you were paid if consent is withdrawn. Second, it buys the kind of document a company holds, from people rather than companies: it announced an $8 million request for anonymised personal financial documents at up to $50 to $150 per form.

Kled says it signed a $12 million non-exclusive licence over two years, a claim that names no buyer. It has raised a $5.5M seed and $3 million from The DATA Foundation. For a company, the lesson is internal: a consumer app that pays for work documents is a channel through which an employee could license what belongs to the employer. Your staff policy on uploading work files should say so; the guide to employee emails and chats in AI training covers the employer's side.

The smaller programs, in brief

Scale AI prefers 40+ employees, shows $10K to $1M+ as an illustrative value that scales with how often data is delivered, anonymises inside your environment before anything is extracted, and designs its partnerships for recurring revenue; its own job post says it is still defining how these transactions get priced. Labelbox's Prism publishes datasets starting around $50K, a written quote within 48 hours of scoping and first call to first wire in under 21 days, and says it has distributed $100M+ to over 1k companies; its licences are exclusive by default, the one program that says so up front.

Polyshares says it is the buyer, not an agent, quotes typical licences of $100K to $2M and above and lists four anonymous licences on its own track-record page, from a $110,000 marketing agency to a $450,000 roofing company, each a perpetual exclusive licence; it pays net 30 to 60 days after the data is approved, and nobody else has confirmed those licences. Nyne takes US businesses with 20 full-time W-2 employees and three years, wants accounting, CRM and dispatch records, and is the only buyer that commits to beat a qualifying written offer for the same data and rights. Ooak licenses for training only, with your exclusions annexed, pays to the company, in 2 to 4 weeks, deletes the raw archive within 30 days, and reports closed deals from $20K to several hundred thousand dollars.

Appen, listed in Australia, and Centific both put terms in writing before extraction, pay on acceptance rather than submission and cover extraction and anonymisation at no cost to the holder. SuperAnnotate pays on delivery and again on each relicense, and licenses only to buyers the holder approves. Collinear takes a one-time export, with ownership staying with the company, and says what raises its price: 15 or more people at peak, several years of history, pull requests linked to tickets and human-written code. Troveo is a marketplace that pays a share of every sale and grants exclusivity only when it is time-limited and earned. Grepped is a broker that pays within 7 days of the data passing review, but its record rests on its own claims only. HUMXN Research Labs buys from skilled-trades and manufacturing businesses at deals that typically start around $250,000, and its $50 million programme is a buying target, with no backer named.

The licence each program asks for

The price is one number; the licence decides what that number buys. A licence is permission to use data on stated terms while the holder keeps ownership; a sale transfers the data or the rights outright. In this market the line blurs, because a licence can be drafted so broadly that it works like a sale. Four words in a grant clause do most of that work. Perpetual means it never ends. Irrevocable means the holder cannot end it, even for breach (the remedy becomes damages). Sublicensable means the buyer may grant the same rights to others, and "through multiple tiers" means those others may grant them on again. Exclusive means the holder may not license the same data to anyone else for the term.

Put the first three together for one payment and the holder has, in economic terms, sold its records: the buyer can resell access to them for ever, and the holder is paid once. That is not wrong in itself. It means the price should be measured against what a sale is worth, not against a licence that would pay again. The licence terms page explains each term with where a holder should stand, and the guide to data licensing agreements walks through a full agreement.

Three patterns stand out. First, the programs that publish the least are not necessarily the most generous: a missing licence term is a term the buyer will write, in its own standard agreement, after you have shared a sample. Second, the only large program that publishes its full grant, micro1, takes the broadest one, and the perpetual, sublicensable licence is a resale right by another name. Third, exclusivity runs both ways across the board: Labelbox and Polyshares take it by default, Kled takes it outright, Troveo makes it conditional, and micro1 explicitly does not.

What exclusivity, term and use are worth

Exclusivity is the largest single lever on price. Two founders of training-environment companies told Epoch AI that exclusive deals cost roughly 4 to 5 times non-exclusive ones, and Troveo repeats the multiple; our licence-terms figure puts an exclusive licence at 4 to 5 times a non-exclusive one, and an evaluation-only licence at 0.2 to 0.5 times a full commercial one. A holder asked for exclusivity by default, as Labelbox and Polyshares do, is being asked to give away its largest lever without a separate price for it. A holder offered a non-exclusive licence keeps the right to sell again: Troveo advertises an average of 4 sales per dataset (labelled illustrative), and SuperAnnotate pays again on each relicense.

Term and the life of the model are a second lever. Disclosed content licences with large platforms run two to three years or at least three years, and licensing counsel recommends pricing a perpetual right to models already trained separately from a finite licence to the dataset. A clause like micro1's, where deleting the data does not require deleting or retraining any model, is standard in substance; what matters is whether the price reflects that the use never ends.

The clauses that cost holders after signing

Two clauses on the board can claw money back. micro1's refund right lets it pay nothing and recover what it paid if the data manifest is materially inaccurate, and that right survives acceptance. Kled's makes the person who uploaded content refund what they were paid if consent is withdrawn. A holder signing either should know exactly what its manifest promises, and should ask for the refund to be limited to the inaccurate part, for a stated period. The prior-licence warranty is the third: micro1's looks backwards, promising that no earlier transaction granted rights to the same data on the same or substantially similar terms, so a holder planning several non-exclusive licences has to narrow it before signing any of them.

Payment timing and process

Payment terms are where one-off deals go wrong, because most of these buyers pay on an event the buyer controls. Of the ten buyers the selling guide scores, only 5 publish how soon the holder is paid. The table below sets out what each program in this guide publishes about the steps between first contact and money in the bank.

The useful distinction is between payment on delivery and payment on acceptance. Mercor pays when the anonymised dataset reaches the buyer; micro1, Nyne, Appen and Centific pay only after they accept it, and micro1 gives itself 30 business days to decide. Acceptance-based payment is not unreasonable (the buyer cannot judge records it has not seen), but it moves the risk of delay onto the holder. Ask for a fixed acceptance window, written acceptance criteria tied to the datasheet, and deemed acceptance if the buyer is silent, which micro1 already gives.

The largest public sale shows the same mechanism at scale. In Spirit Airlines' sale agreement with Google, the $10M is paid only on delivery of de-identified data, and delivery requires an independent agent's certification. Payment waits for de-identification, so whoever controls de-identification controls the clock. Most programs here do the scrub themselves, at their own cost, which Google, micro1 and Mercor have each done without deducting it from the price. That is good for the holder's cash, but it also means the buyer decides when the data is ready. Our de-identification standard explains what a holder can do before any buyer sees a file.

What the terms mean for your leverage

Leverage in a data sale comes from three things: knowing what the data is worth to more than one buyer, controlling what leaves your systems and when, and not needing to sell. The published terms above bear on all three, and several of them are designed, reasonably from the buyer's side, to reduce the first.

The price you see depends on the door

Public estimators do not agree with each other. For a US company of about 100 people and about 10 years, the estimators read on 16 October showed Grepped $37.5K to $54K, Ooak $99k to $139k, micro1 $379,290 to $888,088 and Troveo $800K to $1M: the low ends differ about 21 times. Meanwhile the micro1, Handshake and Nyne estimators share one set of location weights, discounting Europe and Canada to 0.75 of the US figure and everywhere else to 0.3. A holder that fills in one form sees one buyer's anchor and has no way to know whether it is high or low.

Resale margins show the room

A principal buys to resell. micro1's resold datasets were reported at gross margins as high as 80% to 90%; Appen, which reports its margin after paying its crowd, kept about 37 to 40 percent. Neither figure is a criticism. Both mean that the first offer leaves room, and that the buyer knows a holder who has seen only one offer cannot measure it. Nyne's promise to beat a qualifying written offer (calculator estimates and "up to" amounts do not qualify) is the clearest public admission that a rival's written offer moves price.

Some terms reward going first

micro1's prior-licence warranty rewards selling to micro1 before anyone else, in our reading, because a holder that has licensed elsewhere on similar terms cannot sign it unchanged. Exploding offers do the same work with a deadline: Business Insider reported six-figure offers to small businesses that expire in 48 hours, with one broker describing buyers who say "I'm offering you double market. I need an answer by Friday." A deadline is a term like any other. A buyer that wants your records this week will usually still want them in three.

Control of de-identification is priced

In the Spirit auction, Mercor offered $5.2M with third-party de-identification or $7.0M if it could use its own tools, and the estate chose the lower, independently scrubbed bids anyway. One bidder valued the same records 35% more on different scrub terms. The lesson for a holder is that who de-identifies, where, and to what standard is a term worth money, and that the buyer who scrubs inside your environment (Scale says it does) is offering something different from the buyer who takes a raw copy and scrubs it on its side.

Track records differ more than prices

Most of these programs are months old. Of the four large ones, our board records that no holder has confirmed a sale of company records to Handshake, none has confirmed a Project Lazarus deal and none has confirmed one to Mercor; micro1's own court filing is the strongest evidence of volume, and it is self-reported. Among the newer programs, several rest on their own claims only. SimpleClosure, which handles shutdowns, has warned founders about new buyers advertising $1 million to $5 million offers with no completed purchase behind them. Check a buyer's funding, its filed or confirmed deals and its payment terms before you share a sample: its company page on this site gathers what others have recorded.

One program or every fitting buyer at once

Every program above is built for a one-to-one conversation: you apply, the buyer samples, the buyer prices, you sign its agreement. That is fast, and when one buyer is plainly right and its terms are acceptable as published, it can be the right choice. But it is the weakest position to sell from, because the holder never learns what anyone else would pay or accept.

The one public auction of a company's archive shows the difference. In the Spirit Airlines sale, three qualified bidders took Google's opening bid of $5.0M to a winning $10.0M in about two and a half hours, and a fourth buyer, micro1, later filed $12.5M, a 25 percent premium over the winner. Had the estate negotiated only with its first bidder, it would have accepted 50% of what the archive eventually fetched. Competition did more than raise the price: it made de-identification, retention and payment terms negotiable, and the estate picked the structure it trusted over a higher number.

The same holds below the scale of a bankruptcy. Warmly, a software company being acquired by HubSpot, received four offers of up to $300,000 for its codebase and communication records and turned them all down, which is what a holder that does not need to sell can do. Buyers also price on different theses: one wants code with its tickets, another wants CRM history, a third wants a workforce's chat. The program that values your mix most is rarely the first one to email you.

So this is also how we work, stated plainly. We put a holder's lot before every buyer whose published rules it meets, at the same time, with the same datasheet and a deadline for offers, without naming the holder until a buyer is interested and has signed an undertaking. The holder then compares offers on price, licence scope, exclusivity, de-identification and payment trigger together, and licenses to the buyer it chooses. A competitive process takes a few weeks longer than answering one program, and it costs one datasheet rather than one sample per buyer: samples go only to finalists. Our page on going direct or through us sets out what each route costs a holder.

Worked example: a 60-person software company

Example. A US software company with 60 employees, eight years of history, and Slack, Jira, GitHub, Google Workspace and HubSpot in daily use, is contacted by one of the programs above. Its first question is who would take it. On published rules it qualifies for micro1 ( 30+ staff, English), Handshake ( 20+ full-time, 3+ years), Nyne if its staff are on W-2 payroll ( 20+ W-2, three years), Mode ( 20+ office employees), Scale ( 40+ preferred), Ooak ( 20+ staff), Turing, Collinear, Labelbox and the marketplaces. Its codebase meets Turing's bar only if it holds 50,000+ human-written lines with tests.

Its second question is what the estimators say near its size. None takes exactly its inputs, so it reads the nearest: micro1 for 50 employees and 10 years shows $285,896 to $604,463; Handshake for the same inputs $288,755 to $610,508 (the shared formula again); Nyne for 50 to 100 employees and 8 to 15 years $304,500 to $657,000; Mode for 50 desk staff and 10 years $220K to $440K; Ooak for a 51 to 100 person company with six tools $85k to $121k. These are indications, not offers, and every buyer says so.

Its third question is what each would take. If it answers micro1 alone and signs the master terms, it grants a perpetual, sublicensable licence and warrants no earlier similar licence, so a later non-exclusive deal with Handshake or Mode would need micro1's warranty narrowed first. If it answers Labelbox alone, it grants exclusivity by default, which the market prices at 4 to 5 times a shared licence, without a separate price. If it puts one datasheet before all of them, it learns which buyers value its Jira-to-GitHub links, which value its HubSpot history, and which will pay more for exclusivity than the sum of two non-exclusive licences. It may still sign with micro1. It will know why.

Questions to ask before signing

The questions below follow the clauses that cost holders most, in the order to raise them. The offer check runs a written offer against the published clauses of the buyer that made it, and the negotiation page explains where a holder should stand on each.

Before any of these conversations, settle your own side. A buyer's counsel asks 26 rights questions, 15 of which block a sale until answered: customer contracts, software vendors' API terms, employee messages and old privacy promises. The readiness check works through them, and a datasheet describes the lot without exposing it, so you can show every buyer the same thing.

Questions people ask

What is the micro1 data partnership program?

It is micro1's program to license companies' operating records (SOPs, knowledge bases, CRM data, project histories) and turn them into training environments it sells to AI labs, at published tiers of $100k+, $500k+ and $1M+. It takes companies of 30+ employees that work in English, the US first, and has a separate track for closed companies. Its master terms take a perpetual, irrevocable, non-exclusive licence that micro1 may sublicense through multiple tiers.

How much does micro1 pay for company data?

Its published tiers start at $100k+, and its estimator shows $167,427 to $293,614 for a US company of 20 employees with three years of records, rising to $379,290 to $888,088 for 100 employees and 10 years. Those are indications: micro1 prices after it samples the data, and its estimator says it is an "indicative estimate only". Compare it with an independent reading on the valuation page and the guide to what company data is worth.

Does Handshake AI really pay companies for their data?

Handshake publishes payouts of $100K to $4M for companies with 20+ full-time staff and three years in operation, and says it has paid $250M to more than 1K partner companies. Its AI business is large and independently reported. What no seller has yet done is confirm in public a sale of company records to it, so treat its program as real but young, and read its agreement for the terms its page leaves out.

What is Turing's Project Lazarus?

It is Turing's program to buy defined rights to a company's working record or codebases for one payment of up to $1M, from operating, winding-down and closed companies. Repositories typically pay $10k to $100k+, and Turing says it bought failed-startup codebases at an average in the tens of thousands of dollars. The holder keeps ownership; the licence term is not published.

Does Mercor buy company data?

Yes. Mercor connects to a company's tools, extracts and anonymises the records at its own cost, and licenses the dataset to AI labs, paying by wire typically within 2 to 4 weeks of signing. It publishes no price and no licence term, and names eight target industries. It is also the $7.5 million alternate bidder for Spirit Airlines' records.

Can I sell the same data to more than one of these buyers?

Only if every licence you sign is non-exclusive and none forbids it. micro1's licence is non-exclusive but its earlier-licence warranty can block a second similar deal, Labelbox's licences are exclusive by default, and Kled's is exclusive. Plan the sequence before you sign the first.

Do I keep ownership of my data?

Most programs say so: Turing, Handshake, Troveo and Collinear all state that ownership stays with the company. Ownership matters less than the scope of the licence, though: a perpetual, irrevocable, sublicensable licence leaves you owning records you can no longer exclusively exploit. Read the grant clause, not the reassurance.

How long until I am paid?

The published timings run from 7 days after review (Grepped) and 2 to 4 weeks after signing (Mercor) to net 30 to 60 days after approval (Polyshares) and about 60 days (Mode). micro1 lets you invoice only after written acceptance, with 30 business days to reject. Rights work on your side usually takes longer than any of these.

What this means for you

The programs in this guide are real buyers with real money, and some of them pay well. But they compete on terms as much as on price, and a holder that reads only the headline gives away what the terms decide: how long the buyer may use the data, who else may receive it, whether the holder may sell again, and when the money arrives.

1. Check your fit before you call anyone. Run your size, region, language and status through the buyer match tool to see which programs' published rules you meet, and read each buyer's live terms on the buyer board.
2. Get an independent reading of value. One buyer's estimator is one anchor; value your data against published prices and sales before you hear a number.
3. Settle your rights file first. Use the readiness check, and see the rights hub for customer contracts, vendor terms and employee messages.
4. Describe the lot once. A datasheet lets every buyer see the same description, and keeps samples for finalists.
5. Read the licence before the price. Put any offer you receive through the offer check, and use the licence terms page to decide where you stand on term, exclusivity and sublicensing.
6. Compare offers side by side. Whether you run the process yourself or through an agent, ask every fitting buyer at once and compare price, licence, de-identification and payment trigger together.

If you would rather not run that process yourself, we can run it for you, on the terms set out on our pricing page. For the market these programs sit in, read how AI companies get their training data and who the AI training data companies are.