Who sells, and what their pages show

Each vendor, what it sells in its own terms, and the signals its own pages show. A mark in a column means a page of that vendor's, read here, shows a signal on that theme.

A model's reading of the vendors' own pages, not a measurement. An empty column does not mean a vendor does not do that work, only that no page read here shows it. Which vendors have weak evidence of growth is a call made in the research. This section was drafted by Claude, a model made by Anthropic. Anthropic buys from, sells to or competes with companies named here, so read what it says about them with that in mind. How this is made.

Sellers with evidence of fast growth 93

A page read here shows, within the past year, a revenue claim, a funding round, a named model-builder customer or an acquisition.

Vendor	Signals	Graders	Proof	Long tasks	Code	Office work	Company records	Finance	Medicine	Safety	Robots	Video	Speech	Biology
MercorGradersProofCodeOffice workCompany recordsFinanceMedicineSafetyRobots Vetted expert labour for labs (tasks, rubrics, grading); RL environments; off-the-shelf licensed datasets; APEX benchmarks and post-training research; company data licensing (Slack, email, CRM, code, documents) and wearable-camera robotics video; enterprise AI builds and evals. Vendors show a model improving on their data before they ask for an order Many vendors name law among their professional domains Vendors sell graders built to resist cheating, and audits that look for it Expert judgement is bought to write rubrics, gold answers and calibrate judge models Quality control increasingly pairs automated checks with expert auditors Vendors now capture skilled work in real workplaces as well as household chores Several robot builders pre-train on people's first-person video, some with no robot data at that stage Coding tasks must be audited for leaked answers, broken tests and gameable graders Vendors sell simulated companies: cloned work apps filled with files, chat and tasks Bought records are turned into practice environments, not handed to labs raw Vendors now buy whole companies' internal records: chat, email, CRM, code, documents Vendors sell tasks in fields where a tool, not a person, settles correctness Coding tasks now run for hours or days and rebuild whole programs The unit sold is a task with its grader, not a bare environment Coding demand is spreading to kernels, infrastructure, migrations and scientific computing Pay is often per accepted task, with referral fees and paid training on top Tasks are tuned to a pass-rate band on current models Capture is made to order and paid per accepted hour Evaluation is sold as the front end of a data sale Benchmarks saturate fast, and testers keep replacing them with harder ones Senior lawyers, bankers and doctors are offered some of the highest hourly rates A long tail of narrow jobs: trades, inspection, defence, design, language Software engineers are paid by the hour, or per task, to build coding tasks Accounting back-office work is a named target: closing the books and reconciliation Finance experts are hired to write problems and rubrics and to grade output Vendors are adding insurance, credit, tax and compliance work alongside investment banking Vendors sell replicas of named work apps for agents to act in Data is ordered to fit a named training method, including agent-recovery examples Payment follows acceptance; rejected hours, clips and files earn nothing Sellers publish benchmark gains from their data on open models Cyber is becoming a training domain, with exploit-or-patch as the reward Banking and private-equity deliverables recur across vendors: models, decks, diligence Buyers want the same patient over time, across scans, notes, labs and genomics Practising lawyers are paid to write legal tasks with case files and grading criteria Vendors recruit and collect for specialist tools such as CAD, chip-design, energy and spreadsheet software Red-team labour is hired by risk area, credential and language Vendors count coding environments by the task and quote costs per task Practising clinicians are paid to write hard cases and grade model reasoning The wanted task is files in, a finished deliverable out, scored by rubric GPU and custom-accelerator kernel writing is a sold task type Recorded sessions of developers working with coding agents are bought as data Sourcing works backwards from what labs ask for; one vendor says it can start within days Vendors claim small sets of expert tasks lift scores on professional tests Environments are priced per task Robot and first-person video data is often priced by the hour Small sets of expert-built tasks are presented as enough Machine-checked proofs are wanted; one vendor says the appeal is a grader that cannot be fooled Healthcare and finance data are bought direct from the institutions holding them A lab buys recordings of developers coding with AI assistants on their own projects, and another vendor pays developers for the same Science and engineering tasks are built around simulations and recomputable answers One vendor hires hardware engineers and semiconductor experts for judgement on real builds A staffing vendor lists work writing analyst forecasts from a fixed evidence cutoff page 1 page 2 page 3 page 4 page 5 page 6 page 7 page 8 page 9 page 10 page 11 page 12 page 13 page 14 page 15 page 16 page 17 page 18 page 19 page 20 page 21 page 22 page 23 page 24 page 25 page 26 page 27 page 28 page 29 page 30 page 31 page 32	52
Scale AIGradersProofCodeOffice workCompany recordsFinanceMedicineSafetyRobots Human-made training and evaluation data for model builders (SFT, preference data, rubrics, red-teaming, RL environments, coding and agent evaluations, robotics demonstrations), public benchmarks, enterprise and government agent applications, and since 2026 buying companies' operating records to build RL environments. Vendors show a model improving on their data before they ask for an order Many vendors name law among their professional domains Vendors sell graders built to resist cheating, and audits that look for it Pass/fail is giving way to partial credit, rubrics and hybrid graders Quality control increasingly pairs automated checks with expert auditors Vendors now capture skilled work in real workplaces as well as household chores Several robot builders pre-train on people's first-person video, some with no robot data at that stage Coding tasks must be audited for leaked answers, broken tests and gameable graders Vendors sell simulated companies: cloned work apps filled with files, chat and tasks Bought records are turned into practice environments, not handed to labs raw Vendors now buy whole companies' internal records: chat, email, CRM, code, documents Vendors sell tasks in fields where a tool, not a person, settles correctness Buyers want test sets that no model has seen and none can game Coding demand is spreading to kernels, infrastructure, migrations and scientific computing Finding and vetting experts fast is treated as a product in itself PhD-level contributors are recruited across many fields and countries Evaluation is sold as the front end of a data sale Profession-specific private benchmarks are built from real records with practising experts Benchmarks saturate fast, and testers keep replacing them with harder ones Senior lawyers, bankers and doctors are offered some of the highest hourly rates AI research itself is a named science target: agents doing ML engineering Finance experts are hired to write problems and rubrics and to grade output Vendors sell replicas of named work apps for agents to act in Data is ordered to fit a named training method, including agent-recovery examples Vendors sell diagnosis: find where the model fails, then supply the fix Cyber is becoming a training domain, with exploit-or-patch as the reward Vendors are hiring people whose whole job is buying data from companies Banking and private-equity deliverables recur across vendors: models, decks, diligence Training and evaluation need very many isolated sandboxes at once Practising lawyers are paid to write legal tasks with case files and grading criteria Labs arrive with vague goals and expect the vendor to shape the data fast Robot and biology builders publish scaling results that justify buying more data Red-team labour is hired by risk area, credential and language Practising clinicians are paid to write hard cases and grade model reasoning Independent evaluation on real patient records is becoming a product of its own Some vendors publish what they will pay for one company's records Sourcing works backwards from what labs ask for; one vendor says it can start within days One vendor says it built finance environments with most leading closed labs; others sell finance environments to labs Public finance tests are saturating; vendors sell harder tasks with partial-credit grading Vendors screen contributor work to show it was really made by a human The buyers' own job posts say this market has no settled price Private code is wanted because public code no longer gives a clean measurement Buyers want a forecast of what an environment will teach before spending compute Contributor terms at several vendors assign all rights in the work to the vendor page 1 page 2 page 3 page 4 page 5 page 6 page 7 page 8 page 9 page 10 page 11 page 12 page 13 page 14 page 15 page 16 page 17 page 18 page 19 page 20 page 21 page 22 page 23	44
Labelbox (and Alignerr)GradersProofCodeOffice workCompany recordsFinanceSafetyRobotsSpeech Horizon: RL environments and evaluations for frontier labs, authored and graded by its Alignerr contractor network. Terra: robotics data recorded on Labelbox-supplied rigs. Recursion: managed, graded agents for enterprises. Legacy annotation software. Vendors show a model improving on their data before they ask for an order Many vendors name law among their professional domains Pass/fail is giving way to partial credit, rubrics and hybrid graders Expert judgement is bought to write rubrics, gold answers and calibrate judge models Several robot builders pre-train on people's first-person video, some with no robot data at that stage Coding tasks must be audited for leaked answers, broken tests and gameable graders Vendors sell simulated companies: cloned work apps filled with files, chat and tasks Speech in under-served languages and dialects, Indic and Arabic singled out Buyers want test sets that no model has seen and none can game Environments and tasks are increasingly generated by software, not hand-built Private codebases with full history are the wanted raw material for coding Tasks are tuned to a pass-rate band on current models Robot, video and audio buyers grade data on variety, spec and labels Finding and vetting experts fast is treated as a product in itself Capture is made to order and paid per accepted hour PhD-level contributors are recruited across many fields and countries Vendors publish website terms but not their lab contracts; one says fees sit in separate order forms Variety, rare cases and failures are valued above more hours Testers measure what agents do unprompted: lying, collusion, overstepping, self-modification Software engineers are paid by the hour, or per task, to build coding tasks Accounting back-office work is a named target: closing the books and reconciliation Finance experts are hired to write problems and rubrics and to grade output Vendors sell replicas of named work apps for agents to act in The record of how work happened is worth more than the finished product Scientists and engineers with doctorates are paid by the hour Production code is bought with its commits, reviews and tickets, not as snapshots Recorded sessions of developers working with coding agents are bought as data Tool-use environments expose very many callable tools Machine-checked proofs are wanted; one vendor says the appeal is a grader that cannot be fooled Acted, emotional speech and fine-grained emotion labels are in demand; one marketplace pays voice actors most Voice actors and rare-language transcribers are listed at higher rates than ordinary speakers A lab buys recordings of developers coding with AI assistants on their own projects, and another vendor pays developers for the same Code reaches buyers through dead start-ups, non-exclusive licences and cleared rights page 1 page 2 page 3 page 4 page 5 page 6 page 7 page 8 page 9 page 10 page 11 page 12 page 13 page 14 page 15 page 16	33
AppenGradersProofCodeCompany recordsMedicineRobotsVideoSpeech People-made data for training and testing AI: alignment feedback, agentic and coding data, RL environments, speech and multimodal, physical-AI data, model evaluation, research services and an off-the-shelf catalog. Appen China serves Chinese model builders; Appen Global serves US hyperscalers and labs. Vendors show a model improving on their data before they ask for an order Many vendors name law among their professional domains Vendors sell graders built to resist cheating, and audits that look for it Expert judgement is bought to write rubrics, gold answers and calibrate judge models Quality control increasingly pairs automated checks with expert auditors Several robot builders pre-train on people's first-person video, some with no robot data at that stage Many vendors license data non-exclusively and keep the underlying asset Vendors now buy whole companies' internal records: chat, email, CRM, code, documents Video is delivered with step labels, hand pose and recovered geometry Speech in under-served languages and dialects, Indic and Arabic singled out Buyers want test sets that no model has seen and none can game The unit sold is a task with its grader, not a bare environment Private codebases with full history are the wanted raw material for coding Robot operators are paid more by the hour than people who film at home PhD-level contributors are recruited across many fields and countries The buyer set is widening to Chinese, state and open-model builders A long tail of narrow jobs: trades, inspection, defence, design, language Vendors publish website terms but not their lab contracts; one says fees sit in separate order forms Sellers of first-person and robot data name world-model builders as customers alongside robot labs Data is ordered to fit a named training method, including agent-recovery examples Payment follows acceptance; rejected hours, clips and files earn nothing Testers and vendors now grade whether code is mergeable, not only whether tests pass Private speech test sets and human ratings of voice models are being built and offered Generalist work, and work outside the United States, is still listed at far lower rates Real doctor-patient conversations and dictation are a named audio category Practising clinicians are paid to write hard cases and grade model reasoning Production code is bought with its commits, reviews and tickets, not as snapshots Real speech from named settings: clinics, call centres, courtrooms Transcripts must be verbatim, tightly timed, and mark noises and overlap Voice actors and rare-language transcribers are listed at higher rates than ordinary speakers page 1 page 2 page 3 page 4 page 5 page 6 page 7 page 8 page 9 page 10 page 11 page 12 page 13 page 14 page 15 page 16 page 17 page 18 page 19 page 20	30
Snorkel AIGradersProofLong tasksCodeOffice workCompany recordsFinance Finished datasets, benchmarks and RL environments built by paid domain experts plus Snorkel's own software and models, sold off-the-shelf (Snorkel Data Series, quarterly releases) or custom; a second line builds specialized agents for enterprises and government. Vendors show a model improving on their data before they ask for an order Many vendors name law among their professional domains Vendors sell graders built to resist cheating, and audits that look for it Pass/fail is giving way to partial credit, rubrics and hybrid graders Expert judgement is bought to write rubrics, gold answers and calibrate judge models Quality control increasingly pairs automated checks with expert auditors Vendors sell simulated companies: cloned work apps filled with files, chat and tasks Vendors now buy whole companies' internal records: chat, email, CRM, code, documents Coding tasks now run for hours or days and rebuild whole programs Buyers want test sets that no model has seen and none can game Pay is often per accepted task, with referral fees and paid training on top Tasks are tuned to a pass-rate band on current models Finding and vetting experts fast is treated as a product in itself PhD-level contributors are recruited across many fields and countries Vendors and testers build software tasks that take an agent hours, days or a week Finance experts are hired to write problems and rubrics and to grade output Vendors are adding insurance, credit, tax and compliance work alongside investment banking Vendors sell diagnosis: find where the model fails, then supply the fix Sellers publish benchmark gains from their data on open models Vendors are hiring people whose whole job is buying data from companies Practising lawyers are paid to write legal tasks with case files and grading criteria Vendors recruit and collect for specialist tools such as CAD, chip-design, energy and spreadsheet software Scientists and engineers with doctorates are paid by the hour Some vendors organise coverage by occupation across many sectors of the economy Vendors are measuring synthetic data's effect, and one data licence forbids it outright Vendors publish long professional assignments with real deliverables that take a person many hours The buyers' own job posts say this market has no settled price Long tasks need feedback along the way, not one score at the end page 1 page 2 page 3 page 4 page 5 page 6 page 7 page 8 page 9 page 10 page 11 page 12 page 13 page 14 page 15 page 16 page 17	28
Prime IntellectGradersLong tasksCodeCompany recordsSafety Not a human-data vendor: machinery for RL training — Environments Hub and open-source verifiers library, hosted training and evaluations (Lab), sandboxes, inference and rented GPU clusters; runs an open hub of community environments. Many vendors name law among their professional domains Vendors sell graders built to resist cheating, and audits that look for it Pass/fail is giving way to partial credit, rubrics and hybrid graders Coding tasks must be audited for leaked answers, broken tests and gameable graders Coding tasks now run for hours or days and rebuild whole programs Environments and tasks are increasingly generated by software, not hand-built Tasks, harnesses and delivery are converging on shared formats Vendors and testers build software tasks that take an agent hours, days or a week Vendors publish website terms but not their lab contracts; one says fees sit in separate order forms AI research itself is a named science target: agents doing ML engineering Software engineers are paid by the hour, or per task, to build coding tasks The customer's own traces, workflows and evals become the training material Logs of live agents, especially humans correcting them, are treated as prime training data Sensitive buyers want training to happen where the data already sits Training and evaluation need very many isolated sandboxes at once Day-long AI-research tasks are built to measure models improving models Test and training environments are themselves being attacked by the models Enterprises are sold small specialist models that are cheaper and faster Incidents show evaluation and training environments need to be sealed and tamper-resistant Companies want to train on their own data without giving it to a lab Games are used as long-running test and training environments for agents Some data and environment vendors now run post-training on substantial compute of their own Tool-use environments expose very many callable tools One vendor sells training by the token; several report that a single evaluation run is costly Some enterprises train on their own production traces rather than sell them Vendors say enterprise customers often need help deciding what to train and how to score it page 1 page 2 page 3 page 4 page 5 page 6 page 7 page 8 page 9 page 10 page 11 page 12 page 13 page 14 page 15 page 16 page 17	26
TuringGradersLong tasksCodeOffice workCompany recordsFinanceMedicine To labs: training datasets, RL environments and benchmarks in software engineering, enterprise knowledge work and advanced STEM; off-the-shelf data packs; Turing Frontier expert network; Project Lazarus, which buys companies' internal records. To enterprises: forward-deployed engineers building agent systems. Many vendors name law among their professional domains Pass/fail is giving way to partial credit, rubrics and hybrid graders Expert judgement is bought to write rubrics, gold answers and calibrate judge models Quality control increasingly pairs automated checks with expert auditors Vendors sell simulated companies: cloned work apps filled with files, chat and tasks Bought records are turned into practice environments, not handed to labs raw Vendors now buy whole companies' internal records: chat, email, CRM, code, documents Vendors sell tasks in fields where a tool, not a person, settles correctness Coding tasks now run for hours or days and rebuild whole programs Private codebases with full history are the wanted raw material for coding Coding demand is spreading to kernels, infrastructure, migrations and scientific computing Company data is de-identified and rights-checked before use; one licence also bars scraped and machine-made content Pay is often per accepted task, with referral fees and paid training on top Tasks are tuned to a pass-rate band on current models Vendors and testers build software tasks that take an agent hours, days or a week Software engineers are paid by the hour, or per task, to build coding tasks Accounting back-office work is a named target: closing the books and reconciliation Vendors sell replicas of named work apps for agents to act in Logs of live agents, especially humans correcting them, are treated as prime training data Practising lawyers are paid to write legal tasks with case files and grading criteria Vendors count coding environments by the task and quote costs per task The record of how work happened is worth more than the finished product Practising clinicians are paid to write hard cases and grade model reasoning Production code is bought with its commits, reviews and tickets, not as snapshots GPU and custom-accelerator kernel writing is a sold task type Rarity and length of history set the price of a company's records page 1 page 2 page 3 page 4 page 5 page 6 page 7 page 8 page 9 page 10 page 11 page 12 page 13 page 14	26
AfterQueryGradersProofCodeOffice workCompany records Training and test data made with paid professionals: worked examples with reasoning, grading rubrics, tool-using and computer-using environments, ready-made datasets, custom benchmarks. For enterprises, custom model training and on-site engineers. Vendors show a model improving on their data before they ask for an order Many vendors name law among their professional domains Pass/fail is giving way to partial credit, rubrics and hybrid graders Expert judgement is bought to write rubrics, gold answers and calibrate judge models Vendors now buy whole companies' internal records: chat, email, CRM, code, documents The unit sold is a task with its grader, not a bare environment Private codebases with full history are the wanted raw material for coding Company data is de-identified and rights-checked before use; one licence also bars scraped and machine-made content Tasks are tuned to a pass-rate band on current models Finding and vetting experts fast is treated as a product in itself Profession-specific private benchmarks are built from real records with practising experts The buyer set is widening to Chinese, state and open-model builders Vendors publish website terms but not their lab contracts; one says fees sit in separate order forms Sellers publish benchmark gains from their data on open models Vendors are hiring people whose whole job is buying data from companies Enterprises are sold small specialist models that are cheaper and faster Screen recordings with aligned mouse and keyboard events are a purchased input The wanted task is files in, a finished deliverable out, scored by rubric Coding data is sold with a measured gain on someone else's benchmark Small sets of expert-built tasks are presented as enough Legal benchmarks are private, built from real matters, with long documents and dense rubrics Private code is wanted because public code no longer gives a clean measurement Demonstrations can come from a stronger model instead of from people page 1 page 2 page 3 page 4 page 5 page 6 page 7 page 8 page 9 page 10 page 11 page 12 page 13 page 14 page 15	23
HalluminateGradersProofLong tasksOffice workCompany recordsFinance RL environments ("gyms") for finance knowledge work: a simulated desktop with data room, email, chat, spreadsheets and documents; expert-written tasks; graders; plus benchmarks showing where models fail. Sold to frontier AI labs. Vendors show a model improving on their data before they ask for an order Vendors sell graders built to resist cheating, and audits that look for it Pass/fail is giving way to partial credit, rubrics and hybrid graders Expert judgement is bought to write rubrics, gold answers and calibrate judge models Bought records are turned into practice environments, not handed to labs raw Vendors now buy whole companies' internal records: chat, email, CRM, code, documents Company data is de-identified and rights-checked before use; one licence also bars scraped and machine-made content Pay is often per accepted task, with referral fees and paid training on top Evaluation is sold as the front end of a data sale Profession-specific private benchmarks are built from real records with practising experts Senior lawyers, bankers and doctors are offered some of the highest hourly rates Finance experts are hired to write problems and rubrics and to grade output Vendors are hiring people whose whole job is buying data from companies Banking and private-equity deliverables recur across vendors: models, decks, diligence Task length and difficulty must keep rising as models improve Vendors say environment supply, not lab budget, is the constraint The wanted task is files in, a finished deliverable out, scored by rubric One vendor says it built finance environments with most leading closed labs; others sell finance environments to labs Public finance tests are saturating; vendors sell harder tasks with partial-credit grading Healthcare and finance data are bought direct from the institutions holding them Real financial paperwork is bought as raw material, from firms and individuals Vendors test assistants on real phones and build replicas of consumer apps and sites Simulated environments must show they transfer to the real product page 1 page 2 page 3 page 4 page 5 page 6 page 7 page 8 page 9 page 10 page 11	23
micro1Company recordsFinanceRobotsSpeech Realm: evaluations and RL environments for AI labs built by contracted experts; Cortex: agent evaluation and monitoring for enterprises; Robotics: filmed demonstrations, annotation, a robot lab; a government line; buys company operating data to build RL environments. Many vendors name law among their professional domains Vendors now capture skilled work in real workplaces as well as household chores Many vendors license data non-exclusively and keep the underlying asset Bought records are turned into practice environments, not handed to labs raw Vendors now buy whole companies' internal records: chat, email, CRM, code, documents Speech in under-served languages and dialects, Indic and Arabic singled out Company data is de-identified and rights-checked before use; one licence also bars scraped and machine-made content Robot operators are paid more by the hour than people who film at home Profession-specific private benchmarks are built from real records with practising experts Senior lawyers, bankers and doctors are offered some of the highest hourly rates A long tail of narrow jobs: trades, inspection, defence, design, language Vendors are adding insurance, credit, tax and compliance work alongside investment banking Payment follows acceptance; rejected hours, clips and files earn nothing Vendors recruit and collect for specialist tools such as CAD, chip-design, energy and spreadsheet software The record of how work happened is worth more than the finished product Generalist work, and work outside the United States, is still listed at far lower rates Vendors pay their own task builders and researchers lab-level salaries Vendors are measuring synthetic data's effect, and one data licence forbids it outright Some vendors publish what they will pay for one company's records Rarity and length of history set the price of a company's records Deals come as one-off file exports or as continuing collection Stand-in data lets some firms train without exposing customer records One vendor makes sellers warrant the data has not been licensed this way before page 1 page 2 page 3 page 4 page 5 page 6 page 7 page 8	23
ProximalGradersProofLong tasksCodeCompany recordsSafety Hard, long-running coding tasks packaged as RL environments (task, sandbox, grader), evaluations showing where a model is weak, and post-training data aimed at the weakness; also post-trains open models itself to prove the data works. Vendors show a model improving on their data before they ask for an order Vendors sell graders built to resist cheating, and audits that look for it Quality control increasingly pairs automated checks with expert auditors Many vendors license data non-exclusively and keep the underlying asset Bought records are turned into practice environments, not handed to labs raw Coding tasks now run for hours or days and rebuild whole programs Coding demand is spreading to kernels, infrastructure, migrations and scientific computing Non-exclusive licences are the default; the same data is sold to several buyers Evaluation is sold as the front end of a data sale Benchmarks saturate fast, and testers keep replacing them with harder ones Vendors and testers build software tasks that take an agent hours, days or a week AI research itself is a named science target: agents doing ML engineering Sellers publish benchmark gains from their data on open models Cyber is becoming a training domain, with exploit-or-patch as the reward Testers and vendors now grade whether code is mergeable, not only whether tests pass Logs of live agents, especially humans correcting them, are treated as prime training data Training and evaluation need very many isolated sandboxes at once Vendors are measuring synthetic data's effect, and one data licence forbids it outright Coding data is sold with a measured gain on someone else's benchmark Long agent runs need snapshot, resume and a durable record of every run Long tasks need feedback along the way, not one score at the end Long runs need pause, resume, checkpoint and replay built into the environment Science and engineering tasks are built around simulations and recomputable answers page 1 page 2 page 3 page 4 page 5 page 6 page 7 page 8 page 9	23
Surge AIGradersProofLong tasksCodeOffice workFinanceMedicine Human-made training and test data for AI labs (practice environments, rubrics and verifiers, demonstrations, human evaluation, expert fields, many languages, several media); now also ready-made datasets sold with before-and-after benchmark scores, and a service that evaluates, improves and trains models for enterprises. Vendors show a model improving on their data before they ask for an order Many vendors name law among their professional domains Vendors sell graders built to resist cheating, and audits that look for it Pass/fail is giving way to partial credit, rubrics and hybrid graders Expert judgement is bought to write rubrics, gold answers and calibrate judge models Vendors sell simulated companies: cloned work apps filled with files, chat and tasks Labs hand unreleased models to outside testers for scores to publish at launch Private codebases with full history are the wanted raw material for coding PhD-level contributors are recruited across many fields and countries Profession-specific private benchmarks are built from real records with practising experts Benchmarks saturate fast, and testers keep replacing them with harder ones Senior lawyers, bankers and doctors are offered some of the highest hourly rates Software engineers are paid by the hour, or per task, to build coding tasks Vendors are adding insurance, credit, tax and compliance work alongside investment banking Data is ordered to fit a named training method, including agent-recovery examples Payment follows acceptance; rejected hours, clips and files earn nothing Sellers publish benchmark gains from their data on open models Scientists and engineers with doctorates are paid by the hour Coding data is sold with a measured gain on someone else's benchmark Vendors publish long professional assignments with real deliverables that take a person many hours Vendors claim small sets of expert tasks lift scores on professional tests Healthcare back-office work is wanted: billing codes, scheduling and operations Harder office tasks are long, vague and buried in noise, and models still fail page 1 page 2 page 3 page 4 page 5 page 6 page 7 page 8 page 9 page 10 page 11 page 12 page 13 page 14 page 15 page 16 page 17	23
Handshake AILong tasksOffice workCompany recordsFinanceSafetyBiology Hourly expert contractors ("fellows") writing prompts, grading model answers, red-teaming and building tasks for AI labs; de-identified company records bought from companies and passed to labs; agent building and tuning for enterprises. Vendors sell simulated companies: cloned work apps filled with files, chat and tasks Vendors now buy whole companies' internal records: chat, email, CRM, code, documents Testers are paid to measure dangerous abilities and bad behaviour, not just skill Pay is often per accepted task, with referral fees and paid training on top PhD-level contributors are recruited across many fields and countries Senior lawyers, bankers and doctors are offered some of the highest hourly rates A long tail of narrow jobs: trades, inspection, defence, design, language Finance experts are hired to write problems and rubrics and to grade output Vendors are hiring people whose whole job is buying data from companies Banking and private-equity deliverables recur across vendors: models, decks, diligence Vendors recruit and collect for specialist tools such as CAD, chip-design, energy and spreadsheet software Red-team labour is hired by risk area, credential and language Generalist work, and work outside the United States, is still listed at far lower rates Scientists and engineers with doctorates are paid by the hour Some vendors organise coverage by occupation across many sectors of the economy Some vendors publish what they will pay for one company's records Sourcing works backwards from what labs ask for; one vendor says it can start within days Vendors publish long professional assignments with real deliverables that take a person many hours Rarity and length of history set the price of a company's records Deals come as one-off file exports or as continuing collection One expert-labour vendor publishes a biology lab-image benchmark One vendor lists safety red-teamers and policy graders as their own paid line page 1 page 2 page 3 page 4 page 5 page 6 page 7 page 8 page 9 page 10 page 11 page 12 page 13 page 14 page 15	22
ProtegeProofOffice workCompany recordsFinanceMedicineRobotsVideoSpeech Broker of licensed, non-public real-world data: signs data owners (hospitals, labs, broadcasters, sports bodies, game and scanning companies), curates and delivers datasets to AI developers, pays owners per use; also builds evaluation sets and benchmarks, healthcare first. Vendors show a model improving on their data before they ask for an order Vendors now capture skilled work in real workplaces as well as household chores Several robot builders pre-train on people's first-person video, some with no robot data at that stage Many vendors license data non-exclusively and keep the underlying asset Bought records are turned into practice environments, not handed to labs raw Speech in under-served languages and dialects, Indic and Arabic singled out Buyers want test sets that no model has seen and none can game Non-exclusive licences are the default; the same data is sold to several buyers Evaluation is sold as the front end of a data sale Profession-specific private benchmarks are built from real records with practising experts Sellers of first-person and robot data name world-model builders as customers alongside robot labs Accounting back-office work is a named target: closing the books and reconciliation Vendors sell diagnosis: find where the model fails, then supply the fix Buyers want the same patient over time, across scans, notes, labs and genomics Labs arrive with vague goals and expect the vendor to shape the data fast Game footage is sold with the player's keyboard, mouse or controller inputs Real doctor-patient conversations and dictation are a named audio category Independent evaluation on real patient records is becoming a product of its own Orders have shifted from bulk libraries to narrow, curated, commissioned clips Synchronised multi-camera footage and raw-versus-edited pairs sell as their own products Courtroom audio appears as a named legal data category Step-by-step records of people doing office work are a listed product page 1 page 2 page 3 page 4 page 5 page 6 page 7 page 8 page 9 page 10 page 11 page 12 page 13 page 14	22
Invisible TechnologiesGradersCodeOffice workFinanceRobots To model builders: RL environments, agent and tool-use training, custom evaluations, off-the-shelf datasets, multimodal and multilingual data, SFT demonstrations, RLHF comparisons and red-teaming, staffed by experts from its Meridial marketplace. To enterprises: forward-deployed engineers, platform and staffing. Many vendors name law among their professional domains Expert judgement is bought to write rubrics, gold answers and calibrate judge models Quality control increasingly pairs automated checks with expert auditors Coding tasks must be audited for leaked answers, broken tests and gameable graders The unit sold is a task with its grader, not a bare environment Coding demand is spreading to kernels, infrastructure, migrations and scientific computing Pay is often per accepted task, with referral fees and paid training on top Tasks are tuned to a pass-rate band on current models Finding and vetting experts fast is treated as a product in itself Capture is made to order and paid per accepted hour A long tail of narrow jobs: trades, inspection, defence, design, language Accounting back-office work is a named target: closing the books and reconciliation Vendors are adding insurance, credit, tax and compliance work alongside investment banking Payment follows acceptance; rejected hours, clips and files earn nothing Labs arrive with vague goals and expect the vendor to shape the data fast Vendors say environment supply, not lab budget, is the constraint Generalist work, and work outside the United States, is still listed at far lower rates Some vendors organise coverage by occupation across many sectors of the economy One vendor says it built finance environments with most leading closed labs; others sell finance environments to labs Harder office tasks are long, vague and buried in noise, and models still fail One vendor says labs build the interface themselves and buy the domain work page 1 page 2 page 3 page 4 page 5 page 6 page 7 page 8 page 9 page 10 page 11	21
Applied ComputeGradersProofOffice workCompany recordsSafety A platform to train open-weight models with reinforcement learning on the customer's own data and graders, serve them and keep training on production traffic; sold with researchers placed inside the customer. Vendors sell graders built to resist cheating, and audits that look for it Pass/fail is giving way to partial credit, rubrics and hybrid graders Expert judgement is bought to write rubrics, gold answers and calibrate judge models Bought records are turned into practice environments, not handed to labs raw Environments and tasks are increasingly generated by software, not hand-built Logs of live agents, especially humans correcting them, are treated as prime training data Sensitive buyers want training to happen where the data already sits The record of how work happened is worth more than the finished product Enterprises are sold small specialist models that are cheaper and faster Companies want to train on their own data without giving it to a lab Run logs are too large to read, so trace tooling is being built Some data and environment vendors now run post-training on substantial compute of their own Vendors claim small sets of expert tasks lift scores on professional tests Small sets of expert-built tasks are presented as enough Testers and vendors are hardening security to hold models and data Some enterprises train on their own production traces rather than sell them Vendors say enterprise customers often need help deciding what to train and how to score it Legal benchmarks are private, built from real matters, with long documents and dense rubrics Stand-in data lets some firms train without exposing customer records One vendor built a synthetic stand-in from public legal documents because client data could not be used page 1 page 2 page 3 page 4 page 5 page 6 page 7 page 8 page 9 page 10 page 11 page 12 page 13	20
Bespoke LabsGradersProofLong tasksCodeOffice work RL environments built to look like a whole company (large codebase, microservices, tickets, chat, logs), the platform that generates and runs them, and agent optimization including post-training small open models for enterprises. Vendors sell graders built to resist cheating, and audits that look for it Pass/fail is giving way to partial credit, rubrics and hybrid graders Expert judgement is bought to write rubrics, gold answers and calibrate judge models Quality control increasingly pairs automated checks with expert auditors Coding tasks must be audited for leaked answers, broken tests and gameable graders Vendors sell simulated companies: cloned work apps filled with files, chat and tasks The unit sold is a task with its grader, not a bare environment Environments and tasks are increasingly generated by software, not hand-built Coding demand is spreading to kernels, infrastructure, migrations and scientific computing Tasks, harnesses and delivery are converging on shared formats Vendors publish website terms but not their lab contracts; one says fees sit in separate order forms AI research itself is a named science target: agents doing ML engineering The customer's own traces, workflows and evals become the training material Day-long AI-research tasks are built to measure models improving models Task length and difficulty must keep rising as models improve Vendors count coding environments by the task and quote costs per task Enterprises are sold small specialist models that are cheaper and faster Buyers want a forecast of what an environment will teach before spending compute Demonstrations can come from a stronger model instead of from people Repositories are converted into graded tasks, sometimes for one customer's own model page 1 page 2 page 3 page 4 page 5 page 6 page 7 page 8	20
Epoch AIGradersLong tasksCodeOffice work Donor-funded nonprofit that also takes paid commissions: benchmark development (FrontierMath), model evaluation and analysis, custom research reports, data collection, data-centre intelligence, briefings, and an "Epoch on demand" retainer. Vendors sell graders built to resist cheating, and audits that look for it Expert judgement is bought to write rubrics, gold answers and calibrate judge models Quality control increasingly pairs automated checks with expert auditors Many vendors license data non-exclusively and keep the underlying asset Coding tasks must be audited for leaked answers, broken tests and gameable graders Vendors sell tasks in fields where a tool, not a person, settles correctness Coding tasks now run for hours or days and rebuild whole programs Tasks are tuned to a pass-rate band on current models Tasks, harnesses and delivery are converging on shared formats Finding and vetting experts fast is treated as a product in itself Vendors and testers build software tasks that take an agent hours, days or a week AI research itself is a named science target: agents doing ML engineering Data is ordered to fit a named training method, including agent-recovery examples Testers and vendors now grade whether code is mergeable, not only whether tests pass Day-long AI-research tasks are built to measure models improving models The wanted task is files in, a finished deliverable out, scored by rubric Environments are priced per task Machine-checked proofs are wanted; one vendor says the appeal is a grader that cannot be fooled Exclusivity is available at several times the price Labs commission and own research-level maths problems, audits and verifiers page 1 page 2 page 3 page 4 page 5 page 6 page 7 page 8 page 9 page 10 page 11 page 12	20
Vals AIGradersLong tasksCodeFinanceMedicine A testing firm, not a data vendor: independent benchmark runs on private test sets published as leaderboards, pre-release testing for labs, private validation sets licensed to companies, Vals Smith (a company's GitHub repository turned into a coding benchmark), government evaluation. Many vendors name law among their professional domains Vendors sell graders built to resist cheating, and audits that look for it Coding tasks must be audited for leaked answers, broken tests and gameable graders Coding tasks now run for hours or days and rebuild whole programs Labs hand unreleased models to outside testers for scores to publish at launch Buyers want test sets that no model has seen and none can game Testers are paid to measure dangerous abilities and bad behaviour, not just skill Profession-specific private benchmarks are built from real records with practising experts Benchmarks saturate fast, and testers keep replacing them with harder ones AI research itself is a named science target: agents doing ML engineering Vendors are adding insurance, credit, tax and compliance work alongside investment banking Day-long AI-research tasks are built to measure models improving models Independent evaluation on real patient records is becoming a product of its own Vendors sell private tests to help enterprises choose models and judge agent readiness Public finance tests are saturating; vendors sell harder tasks with partial-credit grading One vendor sells training by the token; several report that a single evaluation run is costly Evaluation runs now cost real money, and human baselines are often missing Healthcare back-office work is wanted: billing codes, scheduling and operations One tester sells enterprises private benchmarks to guide which model they pay for page 1 page 2 page 3 page 4 page 5 page 6 page 7 page 8 page 9	19
Fleet AIGradersProofLong tasksCodeSafety Simulated software environments ("training gyms") where an AI agent attempts a real job task and gets graded; plus tasks, graders, benchmarks, post-training recipes and monitoring of agent behaviour. Sold to frontier labs, hyperscalers and enterprises. Vendors show a model improving on their data before they ask for an order Many vendors name law among their professional domains Vendors sell graders built to resist cheating, and audits that look for it Pass/fail is giving way to partial credit, rubrics and hybrid graders Expert judgement is bought to write rubrics, gold answers and calibrate judge models Quality control increasingly pairs automated checks with expert auditors Coding tasks must be audited for leaked answers, broken tests and gameable graders The unit sold is a task with its grader, not a bare environment Environments and tasks are increasingly generated by software, not hand-built Pay is often per accepted task, with referral fees and paid training on top Senior lawyers, bankers and doctors are offered some of the highest hourly rates Testers measure what agents do unprompted: lying, collusion, overstepping, self-modification Vendors sell diagnosis: find where the model fails, then supply the fix Practising lawyers are paid to write legal tasks with case files and grading criteria Labs arrive with vague goals and expect the vendor to shape the data fast Some data and environment vendors now run post-training on substantial compute of their own Long agent runs need snapshot, resume and a durable record of every run Long runs need pause, resume, checkpoint and replay built into the environment page 1 page 2	18
METRGradersLong tasksCodeSafety Not a seller: a donor-funded research nonprofit producing capability scores (time horizon), pre-release evaluations of unreleased models, and embedded risk assessments inside labs (monitor red-teaming, incident investigation, risk-report reviews). Vendors sell graders built to resist cheating, and audits that look for it Coding tasks now run for hours or days and rebuild whole programs Labs hand unreleased models to outside testers for scores to publish at launch Testers are paid to measure dangerous abilities and bad behaviour, not just skill Tasks, harnesses and delivery are converging on shared formats Benchmarks saturate fast, and testers keep replacing them with harder ones Vendors and testers build software tasks that take an agent hours, days or a week Testers measure what agents do unprompted: lying, collusion, overstepping, self-modification Testers and vendors now grade whether code is mergeable, not only whether tests pass Day-long AI-research tasks are built to measure models improving models Task length and difficulty must keep rising as models improve Test and training environments are themselves being attacked by the models Vendors pay their own task builders and researchers lab-level salaries Incidents show evaluation and training environments need to be sealed and tamper-resistant Run logs are too large to read, so trace tooling is being built Testers and vendors are hardening security to hold models and data Evaluation runs now cost real money, and human baselines are often missing One outside tester is let inside labs for audits and incident investigations page 1 page 2 page 3 page 4 page 5 page 6 page 7 page 8 page 9 page 10 page 11 page 12 page 13	18
MechanizeGradersLong tasksCodeSafety Hand-built software-engineering tasks, each a prompt plus a sandbox plus an automatic grader, bundled into environments that frontier labs use for RL training or evaluation; one salaried engineer per task, about one task a week. Vendors sell graders built to resist cheating, and audits that look for it Pass/fail is giving way to partial credit, rubrics and hybrid graders Quality control increasingly pairs automated checks with expert auditors Coding tasks must be audited for leaked answers, broken tests and gameable graders Vendors sell tasks in fields where a tool, not a person, settles correctness Coding tasks now run for hours or days and rebuild whole programs Testers are paid to measure dangerous abilities and bad behaviour, not just skill Non-exclusive licences are the default; the same data is sold to several buyers Vendors and testers build software tasks that take an agent hours, days or a week Testers measure what agents do unprompted: lying, collusion, overstepping, self-modification Testers and vendors now grade whether code is mergeable, not only whether tests pass Task length and difficulty must keep rising as models improve Vendors count coding environments by the task and quote costs per task Vendors pay their own task builders and researchers lab-level salaries Environments are priced per task page 1 page 2 page 3 page 4 page 5 page 6 page 7	15
Irregular (formerly Pattern Labs)GradersSafety Pre-release tests of how well a new AI model can hack: it builds targets (simulated networks, real phones, routers, databases, software), runs the model against them and reports scores. Also sells its SOLVE difficulty scoring, defence research and policy work. Pass/fail is giving way to partial credit, rubrics and hybrid graders Labs hand unreleased models to outside testers for scores to publish at launch Buyers want test sets that no model has seen and none can game Benchmarks saturate fast, and testers keep replacing them with harder ones Testers measure what agents do unprompted: lying, collusion, overstepping, self-modification Cyber is becoming a training domain, with exploit-or-patch as the reward Test and training environments are themselves being attacked by the models Incidents show evaluation and training environments need to be sealed and tamper-resistant Run logs are too large to read, so trace tooling is being built One vendor sells training by the token; several report that a single evaluation run is costly Testers and vendors are hardening security to hold models and data One cyber tester says labs buy its offensive-cyber tests before releases page 1 page 2 page 3 page 4 page 5 page 6 page 7 page 8 page 9	12
Preference ModelGradersProofCodeSafety RL environments (task plus sandbox plus grader) for one domain: the work of an ML engineer or researcher — kernel writing, post-training, data curation, experiment design — delivered into buyers' training runs. Gives away Karotte, an open-source framework. Vendors show a model improving on their data before they ask for an order Vendors sell graders built to resist cheating, and audits that look for it Pass/fail is giving way to partial credit, rubrics and hybrid graders The unit sold is a task with its grader, not a bare environment Coding demand is spreading to kernels, infrastructure, migrations and scientific computing AI research itself is a named science target: agents doing ML engineering Training and evaluation need very many isolated sandboxes at once Test and training environments are themselves being attacked by the models Vendors pay their own task builders and researchers lab-level salaries Incidents show evaluation and training environments need to be sealed and tamper-resistant GPU and custom-accelerator kernel writing is a sold task type Games are used as long-running test and training environments for agents page 1 page 2 page 3 page 4 page 5 page 6 page 7	12
DatoricGradersProofOffice workRobotsVideoSpeech Recordings made to order by paid contributors through private invite-only apps, plus ready-made sets: robot and first-person video, computer-use traces, speech in many languages, two-person conversation, gameplay with inputs, voice-preference rankings. Vendors show a model improving on their data before they ask for an order Touch and force, recorded by gloves, are being added to video Sellers of first-person and robot data name world-model builders as customers alongside robot labs Buyers often sample or pilot first, and some vendors describe orders arriving in bursts Private speech test sets and human ratings of voice models are being built and offered Game footage is sold with the player's keyboard, mouse or controller inputs Screen recordings with aligned mouse and keyboard events are a purchased input Vendors screen contributor work to show it was really made by a human Vendors offer world-model builders video joined to the inputs that caused it One robot-data vendor reports buyers moving to a single supplier; another vendor offers isolated collection per customer Step-by-step records of people doing office work are a listed product page 1 page 2 page 3 page 4 page 5	11
KledGradersProofFinanceVideoSpeech A phone app where ordinary people upload photos, video, voice and documents and do paid tasks; Kled filters for fraud, packages and licenses datasets to AI labs and enterprises. Also holds exclusive AI licences on TV, cartoon and drone archives. Vendors show a model improving on their data before they ask for an order Many vendors license data non-exclusively and keep the underlying asset Speech in under-served languages and dialects, Indic and Arabic singled out Buyers want test sets that no model has seen and none can game Non-exclusive licences are the default; the same data is sold to several buyers Screen recordings with aligned mouse and keyboard events are a purchased input Licensed, rights-cleared video is bought, with brokers holding exclusive AI rights Vendors screen contributor work to show it was really made by a human Consent records, licence trails and security certificates are now a gate to deals Brokers take exclusive AI rights from owners, then license non-exclusively to labs Real financial paperwork is bought as raw material, from firms and individuals page 1 page 2 page 3 page 4 page 5 page 6 page 7 page 8	11
MiFeng / 觅蜂科技 (AgiBot spin-out)GradersRobotsVideo Capture kits (a head unit with several cameras, wrist cameras, a hand-held gripper); robot teleoperation, simulation and robot-free human capture data, stock first then to order; processing software; a gig collection app; sold as a licence or outright. Vendors now capture skilled work in real workplaces as well as household chores Many vendors license data non-exclusively and keep the underlying asset Non-exclusive licences are the default; the same data is sold to several buyers Robot, video and audio buyers grade data on variety, spec and labels Builders split on buying versus collecting in-house, and on exclusivity Capture is made to order and paid per accepted hour Sellers of first-person and robot data name world-model builders as customers alongside robot labs One seller gives away plain factory video; top buyers ask for depth, stereo and a higher frame rate One robot-data vendor reports buyers moving to a single supplier; another vendor offers isolated collection per customer Where a price is shown, real-robot teleoperation data is priced above robot data overall page 1 page 2 page 3 page 4 page 5	10
Origin LabProofVideo Licensed gameplay video recorded inside the game engine by paid players, with keyboard and mouse inputs, depth, surface normals, audio, camera telemetry and game state on one clock; sold to world-model and video-generation labs via API. Tasks, harnesses and delivery are converging on shared formats Buyers often sample or pilot first, and some vendors describe orders arriving in bursts Robot and biology builders publish scaling results that justify buying more data Game footage is sold with the player's keyboard, mouse or controller inputs Robot and first-person video data is often priced by the hour Vendors say video buyers want native high resolution, recent footage, steady cameras and high frame rates Vendors offer world-model builders video joined to the inputs that caused it One vendor records long, scripted play sessions to cover rare events One vendor sells engine-level ground truth with the footage: depth, camera pose, physics One vendor licenses game data at per-hour rates with provenance records, player consent and evaluation-only samples page 1 page 2 page 3	10
Andon LabsGradersLong tasksCodeSafety Tests of whether an AI agent can run something alone for a long time, and how it behaves: simulated businesses and real ones run by an agent. Vendors sell graders built to resist cheating, and audits that look for it Quality control increasingly pairs automated checks with expert auditors Coding tasks must be audited for leaked answers, broken tests and gameable graders Buyers want test sets that no model has seen and none can game Testers are paid to measure dangerous abilities and bad behaviour, not just skill Testers measure what agents do unprompted: lying, collusion, overstepping, self-modification Test and training environments are themselves being attacked by the models Testers measure whether agents stay coherent and honest over very long runs One evaluator tests agents running real businesses and reports cost per run page 1 page 2 page 3 page 4 page 5 page 6	9
Encord (physical-AI line)Robots A data platform (store, search, label and review video, depth, lidar, robot logs) plus a collection service: on-robot collection, teleoperation in Bay Area pods, first-person human video, hand-held gripper capture, action labelling and remote rescue of deployed robots. Vendors now capture skilled work in real workplaces as well as household chores Video is delivered with step labels, hand pose and recovered geometry Touch and force, recorded by gloves, are being added to video Variety, rare cases and failures are valued above more hours Robot builders and vendors feed data from deployed robots back into training Sensitive buyers want training to happen where the data already sits Robot-model builders lean on simulation and generated video, mixed with real data Failed, sloppy and autonomous robot runs are kept as training data One seller gives away plain factory video; top buyers ask for depth, stereo and a higher frame rate page 1 page 2 page 3 page 4	9
WirestockVideo Photos, video, design files and three-dimensional assets made by freelance creators, sold to AI labs as training data: an off-the-shelf library and custom jobs where creators shoot or edit to a lab's brief and Wirestock labels and checks. Many vendors license data non-exclusively and keep the underlying asset Non-exclusive licences are the default; the same data is sold to several buyers A long tail of narrow jobs: trades, inspection, defence, design, language Video must arrive labelled: captions, camera moves, pose and action boundaries Licensed, rights-cleared video is bought, with brokers holding exclusive AI rights Orders have shifted from bulk libraries to narrow, curated, commissioned clips Synchronised multi-camera footage and raw-versus-edited pairs sell as their own products Vendors say video buyers want native high resolution, recent footage, steady cameras and high frame rates People are paid to rank and fault-find generated video page 1 page 2 page 3 page 4 page 5 page 6 page 7	9
DeeptuneOffice workFinance "Training gyms": simulated copies of workplace software (spreadsheets, Slack, Salesforce, ticketing, finance and monitoring tools) with tasks, data and infrastructure, where an agent practises and is scored. Sold to frontier research labs. Many vendors name law among their professional domains Vendors sell simulated companies: cloned work apps filled with files, chat and tasks Accounting back-office work is a named target: closing the books and reconciliation Vendors sell replicas of named work apps for agents to act in Banking and private-equity deliverables recur across vendors: models, decks, diligence Vendors say environment supply, not lab budget, is the constraint Some vendors organise coverage by occupation across many sectors of the economy Vendors sell private tests to help enterprises choose models and judge agent readiness page 1 page 2 page 3 page 4	8
Mecka AIGradersRobots Calibrated multi-sensor first-person capture of people doing real work on its own hardware (body sensors, tactile gloves, camera rigs, phones), reconstructed into motion and geometry; bespoke programmes for labs; a capture app, dataset browser and programming interfaces; robot deployment as integrator. Vendors now capture skilled work in real workplaces as well as household chores Many vendors license data non-exclusively and keep the underlying asset Robot, video and audio buyers grade data on variety, spec and labels Builders split on buying versus collecting in-house, and on exclusivity Capture is made to order and paid per accepted hour Touch and force, recorded by gloves, are being added to video One seller gives away plain factory video; top buyers ask for depth, stereo and a higher frame rate Some data vendors now sell post-training and a working robot, not only datasets page 1 page 2 page 3 page 4 page 5 page 6	8
microAGI (microagi GmbH) — Shift and AtlasRobots Shift: pays workers to film their own work first-person on a head-strapped phone and sells the lightly labelled footage to robot-model labs. Atlas: records factory workers, fine-tunes a partner's robot model, installs a partner's robot, rented on success. Post-training research. Vendors now capture skilled work in real workplaces as well as household chores Video is delivered with step labels, hand pose and recovered geometry Robot operators are paid more by the hour than people who film at home Capture is made to order and paid per accepted hour Touch and force, recorded by gloves, are being added to video Variety, rare cases and failures are valued above more hours Robot and first-person video data is often priced by the hour Some data vendors now sell post-training and a working robot, not only datasets page 1 page 2 page 3 page 4	8
Arena (formerly LMArena)ProofLong tasksCodeSafety A free public site where people vote between anonymous models, plus paid "AI Evaluations": labs and enterprises pay for evaluations on that community, pre-release testing under codenames, and the data behind results. Also an API/router, data products and AutoEval. Labs hand unreleased models to outside testers for scores to publish at launch Vendors publish website terms but not their lab contracts; one says fees sit in separate order forms Testers measure what agents do unprompted: lying, collusion, overstepping, self-modification Sellers publish benchmark gains from their data on open models Recorded sessions of developers working with coding agents are bought as data Testers measure whether agents stay coherent and honest over very long runs An arena tests unreleased models on real users and shares feedback with model developers page 1 page 2 page 3 page 4 page 5	7
Ginkgo DatapointsBiology Fee-for-service lab data generation by robots, customer keeps the result: antibody developability, perturbation response profiling (DRUG-seq RNA and Cell Painting imaging), specialised high-throughput screens, small-molecule ADME. Buyers often sample or pilot first, and some vendors describe orders arriving in bursts Drug makers license biological and clinical data to train their own models Two lab-data vendors say AI buyers want clean, well-documented data, ordered in campaigns Vendors are adding primary cells, organoids and animal models alongside cell lines Biology data funders and contributors get a timed head start, not permanent exclusivity Some health and biology data vendors sell governed access rather than downloadable files Drug companies pool standardised antibody measurements while keeping sequences private page 1 page 2 page 3 page 4 page 5	7
IdlerOffice workCompany recordsSafety Sells reinforcement-learning environments, evals and off-the-shelf datasets built from real production work, such as licensed codebases and de-identified business data, to frontier labs. Many vendors name law among their professional domains Vendors sell simulated companies: cloned work apps filled with files, chat and tasks Bought records are turned into practice environments, not handed to labs raw Company data is de-identified and rights-checked before use; one licence also bars scraped and machine-made content Cyber is becoming a training domain, with exploit-or-patch as the reward Vendors say environment supply, not lab budget, is the constraint Vendors pitch ordinary businesses on licensing their own data as a new revenue line page 1 page 2	7
Lightwheel / 光轮智能RobotsVideo Human first-person work recordings ("EgoSuite"); simulated assets and synthetic robot data ("AutoDataGen" on its "SimFoundry" physics solver); simulated robot-model testing ("RoboFinals"); and "RoboStack", feeding deployed-robot failures into the next data order. Vendors now capture skilled work in real workplaces as well as household chores Many vendors license data non-exclusively and keep the underlying asset Non-exclusive licences are the default; the same data is sold to several buyers Builders split on buying versus collecting in-house, and on exclusivity Touch and force, recorded by gloves, are being added to video Variety, rare cases and failures are valued above more hours Sellers of first-person and robot data name world-model builders as customers alongside robot labs page 1 page 2 page 3 page 4	7
micro1 — robotics departmentRobots Labour-and-data service for robot labs: paid-crowd first-person home and workplace video with phone motion data, expert-demonstrated capture, annotation of robot and human video, on-site US teleoperation in shifts, and a Malibu lab that runs, records and evaluates robots. Vendors now capture skilled work in real workplaces as well as household chores Video is delivered with step labels, hand pose and recovered geometry Pay is often per accepted task, with referral fees and paid training on top Robot operators are paid more by the hour than people who film at home Capture is made to order and paid per accepted hour Payment follows acceptance; rejected hours, clips and files earn nothing A few vendors run tests of robot models on real hardware page 1 page 2 page 3 page 4	7
MidcenturyRobotsVideo Sells large first-person video datasets of real manual work with hand pose, motion and action labels, handheld-gripper recordings, and a cloud simulation platform for robot policies, to robotics developers and AI labs. Vendors now capture skilled work in real workplaces as well as household chores Several robot builders pre-train on people's first-person video, some with no robot data at that stage Video is delivered with step labels, hand pose and recovered geometry Robot builders and vendors feed data from deployed robots back into training Robot-model builders lean on simulation and generated video, mixed with real data Failed, sloppy and autonomous robot runs are kept as training data Video must arrive labelled: captions, camera moves, pose and action boundaries page 1 page 2	7
Protege — video licensing arm (Protege Media)GradersVideoSpeech Licensed video from broadcasters, studios, sports rights holders and platforms under revenue-share deals with exclusive AI-training rights, pooled in one store and built into clipped, filtered, face-masked, rights-checked datasets when a lab orders. Robot, video and audio buyers grade data on variety, spec and labels Licensed, rights-cleared video is bought, with brokers holding exclusive AI rights Orders have shifted from bulk libraries to narrow, curated, commissioned clips Vendors say video buyers want native high resolution, recent footage, steady cameras and high frame rates Vendors are starting to collect and list face and voice recorded together Brokers take exclusive AI rights from owners, then license non-exclusively to labs Some content licensors are testing pay-per-use terms; one says recent content is worth more page 1 page 2 page 3 page 4 page 5	7
CuriosityStreamCodeRobotsVideo Rights to video it owns or has licensed in from other studios, packaged as labelled custom and off-the-shelf datasets for AI training; also audio and a private store of software code. Several robot builders pre-train on people's first-person video, some with no robot data at that stage Private codebases with full history are the wanted raw material for coding Buyers often sample or pilot first, and some vendors describe orders arriving in bursts Video must arrive labelled: captions, camera moves, pose and action boundaries Orders have shifted from bulk libraries to narrow, curated, commissioned clips Synchronised multi-camera footage and raw-versus-edited pairs sell as their own products page 1 page 2 page 3 page 4	6
Defined.aiCodeMedicineRobotsSpeech A marketplace of ready-made licensed datasets (speech, text, image, video, multimodal), some its own and many from outside data partners who earn licence fees; plus custom collection, annotation and evaluation using a paid crowd. Private codebases with full history are the wanted raw material for coding Touch and force, recorded by gloves, are being added to video Real doctor-patient conversations and dictation are a named audio category Production code is bought with its commits, reviews and tickets, not as snapshots Real speech from named settings: clinics, call centres, courtrooms Consent records, licence trails and security certificates are now a gate to deals page 1 page 2 page 3 page 4	6
Haitian Ruisheng / Speechocean (北京海天瑞声, 688787)Speech Training datasets sold two ways: ready-made datasets it owns and licenses repeatedly, and made-to-order collection and labelling, across speech, computer vision and text; a new application-services line (digital humans, voice-model platforms, government data platforms). Many vendors license data non-exclusively and keep the underlying asset Speech in under-served languages and dialects, Indic and Arabic singled out The buyer set is widening to Chinese, state and open-model builders Natural two-way conversation with each speaker on a separate channel Acted, emotional speech and fine-grained emotion labels are in demand; one marketplace pays voice actors most Transcripts must be verbatim, tightly timed, and mark noises and overlap page 1 page 2	6
Ocular AISpeech Studio-quality recordings of real conversation with each speaker on a separate channel, evaluations that show where a voice model fails, matched audio-plus-video recordings, and an Expert Network of paid contributors. Many vendors name law among their professional domains Private speech test sets and human ratings of voice models are being built and offered Natural two-way conversation with each speaker on a separate channel Consent records, licence trails and security certificates are now a gate to deals Speech is collected to order with consent records; one vendor licenses non-exclusively by default Vendors are starting to collect and list face and voice recorded together page 1	6
Tempus AIMedicineBiology Licences to de-identified patient data from its own cancer testing labs (DNA/RNA sequencing, pathology images, clinical notes), plus analysis and, increasingly, a place to train models on that data with GPUs attached; licensed to drug companies. Buyers want the same patient over time, across scans, notes, labs and genomics Sensitive buyers want training to happen where the data already sits Drug makers license biological and clinical data to train their own models Some health and biology data vendors sell governed access rather than downloadable files One vendor licenses patient images and sequencing data for buyers to train their own models One clinical-data company says drug makers license its patient data to train models, and one bought the model page 1	6
BugcrowdGradersSafety Sells crowdsourced security testing to organisations, and reinforcement-learning environments built from real open-source vulnerabilities, with automatic graders, to frontier labs and model makers. Vendors sell tasks in fields where a tool, not a person, settles correctness The unit sold is a task with its grader, not a bare environment Evaluation is sold as the front end of a data sale Buyers often sample or pilot first, and some vendors describe orders arriving in bursts Cyber is becoming a training domain, with exploit-or-patch as the reward page 1 page 2 page 3	5
HUDGradersCompany records An open-source SDK for defining RL environments, tasks and graders; a hosted platform that runs agents against them, records traces and can train on results; and "DataVendor", a marketplace where outside builders sell environments and training data to labs. Quality control increasingly pairs automated checks with expert auditors Environments and tasks are increasingly generated by software, not hand-built Company data is de-identified and rights-checked before use; one licence also bars scraped and machine-made content Vendors publish website terms but not their lab contracts; one says fees sit in separate order forms One vendor is hiring a legal operations lead for data-licence paperwork page 1 page 2 page 3 page 4	5
LuelSpeech A two-sided marketplace: contributors record audio, video and images by app, paid per accepted submission; AI teams order collections to a spec or license finished sets. Luel Labs also builds speech models and voices from the data. Many vendors license data non-exclusively and keep the underlying asset Speech in under-served languages and dialects, Indic and Arabic singled out Non-exclusive licences are the default; the same data is sold to several buyers Natural two-way conversation with each speaker on a separate channel Speech is collected to order with consent records; one vendor licenses non-exclusively by default page 1 page 2 page 3 page 4	5
Protege — audio and speech sideGradersMedicineSpeech Audio that already exists in the world, licensed from the organisations that hold it, then filtered and quality-checked for training (clinical, call-centre, courtroom, Indic, Arabic, low-resource languages; human-reviewed transcripts). Does not record to order. Speech in under-served languages and dialects, Indic and Arabic singled out Robot, video and audio buyers grade data on variety, spec and labels Real doctor-patient conversations and dictation are a named audio category Real speech from named settings: clinics, call centres, courtrooms Courtroom audio appears as a named legal data category page 1	5
Scale AI — physical-AI lineGradersProofRobots Teleoperated two-arm demonstrations from its own cells, a robot-free wearable rig, collection on the customer's hardware, annotation and expert review, ready-made datasets tested by fine-tuning a policy, recording software in Universal Robots' AI Trainer, a planned robot-data marketplace. Vendors show a model improving on their data before they ask for an order Robot, video and audio buyers grade data on variety, spec and labels Capture is made to order and paid per accepted hour Some sellers offer teleoperation on the buyer's own robot; one calls it its most valuable tier One robot-data unit weighs sourcing through a marketplace as well as in-house page 1 page 2 page 3	5
Tahoe TherapeuticsProofBiology Large single-cell datasets showing how cells respond to drugs, made in its own lab with the Mosaic method (many cell lines pooled and treated together); also builds its own models and drug programmes on the same data. Robot and biology builders publish scaling results that justify buying more data Vendors are adding primary cells, organoids and animal models alongside cell lines Labs and datasets are pitched as ways to check general models' hypotheses Drug makers and institutes commission single-cell perturbation data built for model training Biology data funders and contributors get a timed head start, not permanent exclusivity page 1 page 2 page 3 page 4	5
Vibrant LabsGradersLong tasks Automatically generated RL environments, tasks and verifiers for tool-use and computer-use agents, tuned to the edge of a target model's ability, plus open benchmarks. Environments and tasks are increasingly generated by software, not hand-built Tasks are tuned to a pass-rate band on current models Task length and difficulty must keep rising as models improve Vendors say environment supply, not lab budget, is the constraint Tool-use environments expose very many callable tools page 1 page 2 page 3 page 4	5
EnactProofRobots Post-training data for robotics teams: it runs customers' robot policies until they fail, records targeted recovery demonstrations, and runs real-world evaluations. Vendors show a model improving on their data before they ask for an order Variety, rare cases and failures are valued above more hours Data is ordered to fit a named training method, including agent-recovery examples Vendors sell diagnosis: find where the model fails, then supply the fix page 1 page 2	4
General Intuition and MedalVideo Medal, a free app, records gameplay clips with player inputs. General Intuition trains action and world models on the clips and sells model access through an API to partners in games, simulation and robotics. It sells no data. Robot builders and vendors feed data from deployed robots back into training Video and robot builders are sizing pipelines for very large volumes of footage Vendors offer world-model builders video joined to the inputs that caused it A vendor claiming the largest action-labelled game corpus pitches transfer to robots page 1 page 2 page 3 page 4	4
Human Archive (YC W26)ProofRobots Two datasets collected mostly through Indian home-service, hotel and restaurant firms whose staff wear its kit: HA-Multi (depth video, tactile gloves, body motion sensors, wrist cameras, time-aligned) and HA-Ego (single-camera video with labels); also robot-deployment, image and audio sets. Vendors show a model improving on their data before they ask for an order Vendors now capture skilled work in real workplaces as well as household chores Video is delivered with step labels, hand pose and recovered geometry Touch and force, recorded by gloves, are being added to video page 1 page 2	4
Human Native (now part of Cloudflare)Video Before the deal: a UK marketplace turning publishers' and creators' content into licensed, structured datasets for AI developers. Cloudflare has since shipped Pay Per Use, Pay Per Crawl and AI Index for paid, metered AI access to publisher content. Tasks, harnesses and delivery are converging on shared formats Licensed, rights-cleared video is bought, with brokers holding exclusive AI rights Some content licensors are testing pay-per-use terms; one says recent content is worth more AI companies are said to pay for fresh, structured web content over crawling page 1 page 2 page 3	4
Humyn LabsRobotsSpeech Sells speech recordings across many languages and is building first-person video with hand and motion tracking for robot and physical AI builders. Video is delivered with step labels, hand pose and recovered geometry Speech in under-served languages and dialects, Indic and Arabic singled out Tasks, harnesses and delivery are converging on shared formats Buyers often sample or pilot first, and some vendors describe orders arriving in bursts page 1 page 2	4
MarkovGraders Recordings of professionals using desktop software with time-aligned mouse and keyboard events and narration: "CAD environments" (problem statement, reference inputs, gold output file, grading rubric) and plain supervised-training recordings for browser, coding, design and spreadsheet work. Pass/fail is giving way to partial credit, rubrics and hybrid graders Vendors recruit and collect for specialist tools such as CAD, chip-design, energy and spreadsheet software Game footage is sold with the player's keyboard, mouse or controller inputs Screen recordings with aligned mouse and keyboard events are a purchased input page 1 page 2	4
Ooak DataCompany records Licenses the internal records of real companies, including closed firms, anonymises them into digital twins of the workplace, and sells them as agent training environments to frontier labs and enterprise teams. Bought records are turned into practice environments, not handed to labs raw Vendors now buy whole companies' internal records: chat, email, CRM, code, documents Company data is de-identified and rights-checked before use; one licence also bars scraped and machine-made content Vendors pitch ordinary businesses on licensing their own data as a new revenue line page 1 page 2	4
Phinity LabsGraders Markets autonomous chip design; revenue today is work for AI model labs: RL environments, graders, agent infrastructure and evaluation for hardware (chip-design) tasks, including benchmarking unreleased models. Vendors sell tasks in fields where a tool, not a person, settles correctness Labs hand unreleased models to outside testers for scores to publish at launch Labs pay a small specialist for chip-design environments One chip-design vendor says labs give it unreleased models and work with it frequently page 1 page 2	4
SuperAnnotateGradersSafety Sells expert-built training data, practice environments and benchmarks, recruiting doctors, lawyers, engineers and scientists for AI labs and enterprises. Many vendors name law among their professional domains Expert judgement is bought to write rubrics, gold answers and calibrate judge models Finding and vetting experts fast is treated as a product in itself Red-team labour is hired by risk area, credential and language page 1 page 2 page 3 page 4	4
Applied IntuitionGraders Sells simulation, synthetic labelled camera, lidar and radar data, and data-curation tools to carmakers, trucking, mining and defence firms building autonomous machines. Training data is a side line next to its autonomy software. Robot builders and vendors feed data from deployed robots back into training Vendors are measuring synthetic data's effect, and one data licence forbids it outright Labels are increasingly made by machines, by execution or simulation, not by human annotators page 1 page 2 page 3	3
Chakra Labs and PangoProof Pango (2025): a Chrome extension paying people to work in Sheets, Slides, Figma and Canva while recording screen and inputs. Chakra Labs (2026): exact clones of real software as environments, with tasks, graders, evals and datasets. Vendors show a model improving on their data before they ask for an order Training and evaluation need very many isolated sandboxes at once Labs arrive with vague goals and expect the vendor to shape the data fast page 1 page 2 page 3	3
Good Start LabsCompany records RL environments, gameplay data and benchmarks built from games for frontier and open-model labs, plus a service that helps game publishers license their gameplay data to AI labs. Labs hand unreleased models to outside testers for scores to publish at launch Vendors pitch ordinary businesses on licensing their own data as a new revenue line Games are used as long-running test and training environments for agents page 1 page 2 page 3	3
HumanlayaCompany records Runs an expert platform that supplies expert data and evaluation work to most major Chinese model companies, and publishes public benchmarks. It also asks enterprises to license their anonymised workflow data. Company data is de-identified and rights-checked before use; one licence also bars scraped and machine-made content PhD-level contributors are recruited across many fields and countries The buyer set is widening to Chinese, state and open-model builders page 1 page 2 page 3	3
Illumina BioInsight / Billion Cell AtlasBiology Membership of an alliance with access to a very large single-cell dataset of genetic perturbations that Illumina is generating, hosted on its own cloud. Sensitive buyers want training to happen where the data already sits Drug makers license biological and clinical data to train their own models Drug makers and institutes commission single-cell perturbation data built for model training page 1	3
Macrodata LabsGradersRobotsVideo Sells automated processing of robot and first-person video for robotics teams: timestamped subtask labels, recovered hand motion and episode quality checks, run on its metered cloud. Video is delivered with step labels, hand pose and recovered geometry Video must arrive labelled: captions, camera moves, pose and action boundaries Labels are increasingly made by machines, by execution or simulation, not by human annotators page 1	3
Mind Lab The research arm of the Macaron AI assistant, selling benchmark building, task environments rebuilt from a customer's own work, and model post-training to pharmaceutical, financial, public-sector and gaming firms, plus a self-serve training platform. Tasks, harnesses and delivery are converging on shared formats The customer's own traces, workflows and evals become the training material Companies want to train on their own data without giving it to a lab page 1 page 2 page 3	3
Paper InstrumentsGraders Builds evaluation datasets and grading rubrics with knowledge-work professionals for AI companies, and uses rubric-based reinforcement learning to post-train open models. Pass/fail is giving way to partial credit, rubrics and hybrid graders Expert judgement is bought to write rubrics, gold answers and calibrate judge models The customer's own traces, workflows and evals become the training material page 1 page 2	3
Patronus AIGradersProof Builds training tasks, graders and simulated apps and workflows for frontier labs to use in mid-training and reinforcement learning, aimed at where their models fail. Quality control increasingly pairs automated checks with expert auditors Environments and tasks are increasingly generated by software, not hand-built Vendors sell diagnosis: find where the model fails, then supply the fix page 1 page 2	3
PolymathCodeOffice work Sells simulated work settings with tasks and graders where agents practise long software and other work, to leading model labs. Vendors sell simulated companies: cloned work apps filled with files, chat and tasks Coding demand is spreading to kernels, infrastructure, migrations and scientific computing Software engineers are paid by the hour, or per task, to build coding tasks page 1 page 2	3
RunloopGraders Sells isolated cloud sandboxes, hosted coding benchmarks and an execution layer for reinforcement-learning rollouts and fine-tuning data, to agent builders, post-training teams and enterprises, billed by plan plus metered compute. Vendors sell tasks in fields where a tool, not a person, settles correctness Vendors sell private tests to help enterprises choose models and judge agent readiness Labels are increasingly made by machines, by execution or simulation, not by human annotators page 1 page 2	3
Synthetic Sciences Sells long tests and training of whether models can do experimental research, and helps teams turn user feedback into training data for open models. The customer's own traces, workflows and evals become the training material Logs of live agents, especially humans correcting them, are treated as prime training data Science tasks test whether agents can design and run experiments with limited resources page 1 page 2	3
Trajectory LabsSafety Sells red-teaming, safety evaluations and attack data, supplying expert red teamers to frontier labs before launches. Labs hand unreleased models to outside testers for scores to publish at launch Testers are paid to measure dangerous abilities and bad behaviour, not just skill Red-team labour is hired by risk area, credential and language page 1 page 2 page 3	3
Verita AI Sources, screens, trains and pays domain and creative experts for frontier AI companies, for annotation, preference data, evaluations and RL-environment work, acting as their employer of record. Finding and vetting experts fast is treated as a product in itself A long tail of narrow jobs: trades, inspection, defence, design, language Vendors say demand has moved from large crowds to small groups of experts page 1 page 2 page 3	3
Vetto AIGradersProof Sells rubric-graded data, practice environments and ready-made datasets from its public benchmarks, built by professionals it recruits, to AI labs. Vendors show a model improving on their data before they ask for an order Pass/fail is giving way to partial credit, rubrics and hybrid graders Evaluation is sold as the front end of a data sale page 1 page 2 page 3	3
XDOFVideo Robot demonstrations made by people driving robots: teleoperation on the buyer's own robot, teleoperation on its standard two-arm rigs, and first-person human data from sensor-wearing collectors; plus cleaning, annotation, a dataset portal (Foundry), policy-training support and evaluation. Sellers of first-person and robot data name world-model builders as customers alongside robot labs Some sellers offer teleoperation on the buyer's own robot; one calls it its most valuable tier One world-model builder is collecting real robot logs; one free set has been used as a test set page 1 page 2 page 3	3
Basecamp ResearchProofBiology Drug candidates and design work from its own biology models, through pharma partnerships. It keeps its private gene-sequence dataset; no page offers the raw data for licence. Robot and biology builders publish scaling results that justify buying more data Gene-sequence data from outside public databases, collected under benefit-sharing agreements page 1 page 2	2
Besimple AIProofSpeech Sells licensed conversational audio with human transcription, expert annotation, evaluation pipelines and voice benchmarks to teams building speech and multimodal models. Vendors show a model improving on their data before they ask for an order Private speech test sets and human ratings of voice models are being built and offered page 1	2
BioStack PlatformsMedicine De-identified real patient records, imaging and clinical notes, sold as raw data or turned into healthcare RL environments, evaluations and benchmarks for AI labs and life-science teams. Buyers want the same patient over time, across scans, notes, labs and genomics Independent evaluation on real patient records is becoming a product of its own page 1 page 2	2
Deccan AIGradersCode Sells private code, expert-written practice environments with graders, and first-person video, drawing on a large contributor network in India, to AI labs. The unit sold is a task with its grader, not a bare environment Private codebases with full history are the wanted raw material for coding page 1 page 2	2
Gray Swan AI Sells pre-release red-teaming and safety research to frontier labs, and automated attack testing and runtime filtering to enterprises deploying agents. Its attack data comes from a public red-teaming arena that pays prizes. Labs hand unreleased models to outside testers for scores to publish at launch Testers are paid to measure dangerous abilities and bad behaviour, not just skill page 1	2
iMeritGraders Sells expert-led model evaluation, RLHF, red-teaming and fine-tuning data, plus image, lidar, sensor and medical annotation, to frontier labs and autonomy, healthcare and robotics teams. It is now part of EXL. Expert judgement is bought to write rubrics, gold answers and calibrate judge models PhD-level contributors are recruited across many fields and countries page 1 page 2	2
LatchGraders Sells biology benchmarks graded against real lab results and biosecurity testing to model builders, alongside software for biotech scientists. Vendors sell tasks in fields where a tool, not a person, settles correctness Testers are paid to measure dangerous abilities and bad behaviour, not just skill page 1 page 2	2
Mantic Probability forecasts of real-world events for hedge funds and companies, with a tool to create and track personalised forecasts and MCP and API access, built on other labs' models. It pays people to write hard forecasting questions. One forecasting firm says hard questions are the scarce input and pays for them Three large labs fund a live forecasting benchmark; one invested in a forecasting firm page 1 page 2	2
MedraBiology Robots plus AI that run wet-lab experiments, installed at the customer or run in its own lab; for model builders, experimental biology data on demand. Two lab-data vendors say AI buyers want clean, well-documented data, ordered in campaigns Labs and datasets are pitched as ways to check general models' hypotheses page 1 page 2	2
Vision Lab (YC)Robots First-person video from live factory lines, paired with a wider fixed view of the station and the written procedure (which step, what order, which tool, done right or wrong); also trains its own model to read factory video. Vendors now capture skilled work in real workplaces as well as household chores Video is delivered with step labels, hand pose and recovered geometry page 1	2
Arga Labs Sells stateful replicas of business services such as Slack, Stripe, GitHub and Jira, used as RL training environments for labs, sandboxes for enterprise agents and checks on code changes. Vendors sell replicas of named work apps for agents to act in page 1	1
Eigent AIGraders Sells an open-source multi-agent desktop app to enterprises, plus custom RL environments, datasets and verifiers for enterprises and AI developers, reportedly including frontier labs. Vendors sell tasks in fields where a tool, not a person, settles correctness page 1	1
FIZ Karlsruhe Licenses its inorganic crystal structure database, including bulk access for data mining licensed per named project with prices on request. It supplies the database to CuspAI for AI models. Long-established crystal-structure databases hold AI-training rights back from ordinary access, and one buyer says it holds them exclusively page 1	1
Gradient HealthMedicine "Atlas", a searchable library of de-identified medical scans with radiology reports, licensed to medical AI companies; from June 2026 adding health records, pathology, lab results and heart tracings. Also licenses through broker Protege. Buyers want the same patient over time, across scans, notes, labs and genomics page 1	1
Prolific (prolific.com) Prolific sells human data from its participants and verified experts, including made-to-order first-person video, manipulation recordings and annotation. It also lists ready-made expert datasets. Many vendors name law among their professional domains page 1	1

Sellers with weak evidence of growth 73weak evidence

Their pages show what they sell, but no page read here gives good evidence that they are growing fast.

Vendor	Signals	Graders	Proof	Long tasks	Code	Office work	Company records	Finance	Medicine	Safety	Robots	Video	Speech	Biology
Toloka (and its contributor brand Mindrift)GradersProofOffice workCompany recordsFinanceMedicineRobots To labs: human-made training and test data, RL gyms (simulated companies and apps), off-the-shelf datasets, red-teaming, robotics video. To enterprises: fine-tuning service, private leaderboard (Arena), labelling platform. Tendem: vetted humans callable by agents. Workers recruited through Mindrift. Vendors show a model improving on their data before they ask for an order Many vendors name law among their professional domains Vendors sell graders built to resist cheating, and audits that look for it Pass/fail is giving way to partial credit, rubrics and hybrid graders Expert judgement is bought to write rubrics, gold answers and calibrate judge models Quality control increasingly pairs automated checks with expert auditors Several robot builders pre-train on people's first-person video, some with no robot data at that stage Vendors sell simulated companies: cloned work apps filled with files, chat and tasks Vendors now buy whole companies' internal records: chat, email, CRM, code, documents Vendors sell tasks in fields where a tool, not a person, settles correctness Buyers want test sets that no model has seen and none can game Evaluation is sold as the front end of a data sale Profession-specific private benchmarks are built from real records with practising experts Vendors sell replicas of named work apps for agents to act in Buyers often sample or pilot first, and some vendors describe orders arriving in bursts Sellers publish benchmark gains from their data on open models Vendors are hiring people whose whole job is buying data from companies Banking and private-equity deliverables recur across vendors: models, decks, diligence Vendors recruit and collect for specialist tools such as CAD, chip-design, energy and spreadsheet software Enterprises are sold small specialist models that are cheaper and faster Generalist work, and work outside the United States, is still listed at far lower rates Practising clinicians are paid to write hard cases and grade model reasoning Machine-checked proofs are wanted; one vendor says the appeal is a grader that cannot be fooled Deals come as one-off file exports or as continuing collection Science and engineering tasks are built around simulations and recomputable answers Vendors test assistants on real phones and build replicas of consumer apps and sites Contributor terms at several vendors assign all rights in the work to the vendor Contributor work sessions are themselves instrumented and recorded page 1 page 2 page 3 page 4 page 5 page 6 page 7 page 8 page 9 page 10 page 11 page 12	28
PlatoGradersProofLong tasksCodeOffice workCompany records Simulated copies of real web and desktop software (replica sites, VM images of business apps, an Ubuntu desktop) that agents act in and are scored on, plus a pipeline turning customers' real traces into environments; for browser and computer-use agents. Vendors sell graders built to resist cheating, and audits that look for it Quality control increasingly pairs automated checks with expert auditors Bought records are turned into practice environments, not handed to labs raw Coding tasks now run for hours or days and rebuild whole programs Tasks are tuned to a pass-rate band on current models Vendors and testers build software tasks that take an agent hours, days or a week Vendors publish website terms but not their lab contracts; one says fees sit in separate order forms Vendors sell replicas of named work apps for agents to act in Training and evaluation need very many isolated sandboxes at once Vendors pay their own task builders and researchers lab-level salaries Screen recordings with aligned mouse and keyboard events are a purchased input One vendor sells training by the token; several report that a single evaluation run is costly Evaluation runs now cost real money, and human baselines are often missing Long agent runs need snapshot, resume and a durable record of every run Long runs need pause, resume, checkpoint and replay built into the environment Step-by-step records of people doing office work are a listed product Simulated environments must show they transfer to the real product page 1 page 2 page 3 page 4 page 5	17
TroveoCodeCompany recordsFinanceVideo Broker of licensed, non-public media: signs content owners (creators, broadcasters, sports bodies, game and esports companies), indexes, labels and delivers footage to AI developers. Since Apr 2026 also audio, text, enterprise workflow data, gaming, robotics data; company records; custom sourcing. Vendors now buy whole companies' internal records: chat, email, CRM, code, documents Private codebases with full history are the wanted raw material for coding Labs arrive with vague goals and expect the vendor to shape the data fast The record of how work happened is worth more than the finished product Game footage is sold with the player's keyboard, mouse or controller inputs Production code is bought with its commits, reviews and tickets, not as snapshots Video must arrive labelled: captions, camera moves, pose and action boundaries Environments are priced per task Synchronised multi-camera footage and raw-versus-edited pairs sell as their own products Vendors offer world-model builders video joined to the inputs that caused it Deals come as one-off file exports or as continuing collection Real financial paperwork is bought as raw material, from firms and individuals Exclusivity is available at several times the price Weak-evidence sellers say price follows rarity, rigour and age, not size page 1 page 2 page 3 page 4 page 5 page 6 page 7	14
DatacurveGradersLong tasksCodeSafety Coding-centred training data and tests made by paid software engineers: RL environments, long tasks, recorded expert work sessions, worked examples for fine-tuning, ready-made datasets and benchmarks (DeepSWE). Contributors come through Shipd, a bounty site. Vendors sell graders built to resist cheating, and audits that look for it Pass/fail is giving way to partial credit, rubrics and hybrid graders Quality control increasingly pairs automated checks with expert auditors Coding tasks must be audited for leaked answers, broken tests and gameable graders Coding tasks now run for hours or days and rebuild whole programs Finding and vetting experts fast is treated as a product in itself Vendors and testers build software tasks that take an agent hours, days or a week AI research itself is a named science target: agents doing ML engineering Software engineers are paid by the hour, or per task, to build coding tasks Cyber is becoming a training domain, with exploit-or-patch as the reward Vendors count coding environments by the task and quote costs per task Coding data is sold with a measured gain on someone else's benchmark Long tasks need feedback along the way, not one score at the end page 1 page 2 page 3 page 4 page 5	13
SimpleClosure AssetHubGradersCodeOffice workCompany records Winds down startups; AssetHub takes their code and internal records (SourceCode: repositories, tickets, docs; WorkspaceData: Slack, email, Drive, Notion, Figma), strips personal data, and sells or licenses them to AI buyers. Since Sep 2026 also takes operating companies. Bought records are turned into practice environments, not handed to labs raw Vendors now buy whole companies' internal records: chat, email, CRM, code, documents The unit sold is a task with its grader, not a bare environment Private codebases with full history are the wanted raw material for coding Company data is de-identified and rights-checked before use; one licence also bars scraped and machine-made content Data is ordered to fit a named training method, including agent-recovery examples The record of how work happened is worth more than the finished product Production code is bought with its commits, reviews and tickets, not as snapshots Rarity and length of history set the price of a company's records Step-by-step records of people doing office work are a listed product Repositories are converted into graded tasks, sometimes for one customer's own model Weak-evidence sellers say price follows rarity, rigour and age, not size page 1 page 2 page 3 page 4 page 5 page 6	12
David AIGradersProofSpeech Recorded human speech licensed as datasets: two-speaker channel-separated conversation (Converse), multilingual (Atlas), three-plus speakers (Chorus), expert dialogue (Dialog); also voice-acting recordings, transcription and, new in 2026, evaluations of voice models. Contributors paid through Babel Audio. Vendors show a model improving on their data before they ask for an order Speech in under-served languages and dialects, Indic and Arabic singled out Buyers want test sets that no model has seen and none can game Robot, video and audio buyers grade data on variety, spec and labels Vendors publish website terms but not their lab contracts; one says fees sit in separate order forms Private speech test sets and human ratings of voice models are being built and offered Natural two-way conversation with each speaker on a separate channel Speech is collected to order with consent records; one vendor licenses non-exclusively by default Transcripts must be verbatim, tightly timed, and mark noises and overlap Vendors are starting to collect and list face and voice recorded together Voice actors and rare-language transcribers are listed at higher rates than ordinary speakers page 1 page 2 page 3 page 4 page 5 page 6 page 7 page 8	11
Outlier (Scale AI's contractor brand)GradersVideoSpeech Scale AI's contributor marketplace: paid freelance tasks (hard prompts, grading rubrics, rating and ranking model answers, voice conversations, image/video comparison, phone-screen recordings with an AI assistant) — the labour behind Scale's Gen AI data business. Expert judgement is bought to write rubrics, gold answers and calibrate judge models Finding and vetting experts fast is treated as a product in itself Payment follows acceptance; rejected hours, clips and files earn nothing Private speech test sets and human ratings of voice models are being built and offered Generalist work, and work outside the United States, is still listed at far lower rates Scientists and engineers with doctorates are paid by the hour Vendors screen contributor work to show it was really made by a human Vendors test assistants on real phones and build replicas of consumer apps and sites Contributor terms at several vendors assign all rights in the work to the vendor People are paid to rank and fault-find generated video Contributor work sessions are themselves instrumented and recorded page 1 page 2 page 3 page 4 page 5 page 6 page 7	11
Theta (Theta Software Inc.)GradersProofLong tasksCompany recordsFinanceSafety Custom RL environments for professional knowledge work (finance, back office, healthcare, legal, IT, software), built from a customer's own data plus Theta's hired experts; and the trained or tuned agents that come out of them, delivered hands-on. Pass/fail is giving way to partial credit, rubrics and hybrid graders Quality control increasingly pairs automated checks with expert auditors Bought records are turned into practice environments, not handed to labs raw Benchmarks saturate fast, and testers keep replacing them with harder ones Vendors publish website terms but not their lab contracts; one says fees sit in separate order forms Vendors sell replicas of named work apps for agents to act in Test and training environments are themselves being attacked by the models Public finance tests are saturating; vendors sell harder tasks with partial-credit grading Vendors publish long professional assignments with real deliverables that take a person many hours Tool-use environments expose very many callable tools Buyers want a forecast of what an environment will teach before spending compute page 1 page 2 page 3 page 4	11
CentificGradersOffice workCompany recordsMedicineRobots Sells reinforcement-learning environments, human evaluation, multimodal data and a dataset marketplace to frontier labs and enterprises, staffed by its own crowd and expert network. It also pays enterprises to license anonymised operational data for lab training. Vendors sell graders built to resist cheating, and audits that look for it Expert judgement is bought to write rubrics, gold answers and calibrate judge models Vendors sell simulated companies: cloned work apps filled with files, chat and tasks Company data is de-identified and rights-checked before use; one licence also bars scraped and machine-made content Tasks, harnesses and delivery are converging on shared formats Touch and force, recorded by gloves, are being added to video Payment follows acceptance; rejected hours, clips and files earn nothing Practising clinicians are paid to write hard cases and grade model reasoning Vendors pitch ordinary businesses on licensing their own data as a new revenue line Deals come as one-off file exports or as continuing collection page 1 page 2 page 3 page 4 page 5	10
pre.devGradersProofCode Licenses private codebases from the teams that wrote them and sells frontier labs verified RL coding tasks built from their commits, or the cleaned codebases themselves priced per token. Vendors show a model improving on their data before they ask for an order Pass/fail is giving way to partial credit, rubrics and hybrid graders Private codebases with full history are the wanted raw material for coding Tasks are tuned to a pass-rate band on current models Sellers publish benchmark gains from their data on open models Production code is bought with its commits, reviews and tickets, not as snapshots Coding data is sold with a measured gain on someone else's benchmark Private code is wanted because public code no longer gives a clean measurement Repositories are converted into graded tasks, sometimes for one customer's own model Vendors prove each task can be solved before they ship it page 1 page 2 page 3 page 4	10
PetrarchOffice workCompany records Broker of private company records: finds companies with unused internal data (including bankrupt ones), checks ownership, strips personal information on the supplier's machines (Lindo), restructures it, and resells to applied AI startups and labs. Bought records are turned into practice environments, not handed to labs raw Vendors now buy whole companies' internal records: chat, email, CRM, code, documents Company data is de-identified and rights-checked before use; one licence also bars scraped and machine-made content Buyers often sample or pilot first, and some vendors describe orders arriving in bursts The record of how work happened is worth more than the finished product Deals come as one-off file exports or as continuing collection Step-by-step records of people doing office work are a listed product Code reaches buyers through dead start-ups, non-exclusive licences and cleared rights page 1 page 2 page 3 page 4 page 5	8
Cortex AIRobotsVideo First-person workplace video with hand and body pose, depth and sub-task labels; robot trajectories collected in real workplaces; remote operators who rescue deployed robots, with the rescue recorded. Plus a marketplace paying workplaces to host sessions. Vendors now capture skilled work in real workplaces as well as household chores Several robot builders pre-train on people's first-person video, some with no robot data at that stage Video is delivered with step labels, hand pose and recovered geometry Sellers of first-person and robot data name world-model builders as customers alongside robot labs Some sellers offer teleoperation on the buyer's own robot; one calls it its most valuable tier A few vendors run tests of robot models on real hardware page 1 page 2	6
Miso MotionRobotsVideo Made-to-order first-person and multi-camera video of real workers doing paid jobs, with consent records, for physical-AI and robotics teams; a spin-out of a Korean home-services marketplace. Vendors now capture skilled work in real workplaces as well as household chores Video is delivered with step labels, hand pose and recovered geometry Tasks, harnesses and delivery are converging on shared formats Capture is made to order and paid per accepted hour Synchronised multi-camera footage and raw-versus-edited pairs sell as their own products Some vendors run a real business so its daily work becomes data to sell page 1 page 2 page 3	6
Moonvalley / Asteria (now part of Reka)RobotsVideo Marey, a text-to-video and image-to-video model for film and advertising trained only on licensed footage; it pays for footage and runs an annotator platform. Since June 2026 the team is at Reka, whose Claru line sells egocentric, robot data. Several robot builders pre-train on people's first-person video, some with no robot data at that stage Capture is made to order and paid per accepted hour Payment follows acceptance; rejected hours, clips and files earn nothing Licensed, rights-cleared video is bought, with brokers holding exclusive AI rights Video and robot builders are sizing pipelines for very large volumes of footage Contributor terms at several vendors assign all rights in the work to the vendor page 1 page 2 page 3 page 4 page 5	6
VmaxGradersSafety Sells a self-play product in which a setter model writes verifiable tasks with environments and verifiers from a customer brief, exporting certified task packs for the customer’s own reinforcement-learning training. Vendors sell graders built to resist cheating, and audits that look for it The unit sold is a task with its grader, not a bare environment Environments and tasks are increasingly generated by software, not hand-built Tasks are tuned to a pass-rate band on current models Cyber is becoming a training domain, with exploit-or-patch as the reward The customer's own traces, workflows and evals become the training material page 1 page 2 page 3	6
hillclimbGradersProof RL environments and evals for AI-research ability, sold to frontier labs. At launch (Nov 2025): maths training data and maths RL environments written by elite mathematicians. Vendors show a model improving on their data before they ask for an order Pass/fail is giving way to partial credit, rubrics and hybrid graders Expert judgement is bought to write rubrics, gold answers and calibrate judge models Sellers publish benchmark gains from their data on open models Vendors say environment supply, not lab budget, is the constraint page 1 page 2 page 3 page 4	5
Huzzle LabsGradersFinance Sells long-horizon RL environments with programmatic graders to frontier labs, staffed by experts from the Huzzle talent platform, and post-trains small open models on single workflows for enterprises. The unit sold is a task with its grader, not a bare environment Finding and vetting experts fast is treated as a product in itself Vendors are adding insurance, credit, tax and compliance work alongside investment banking Sensitive buyers want training to happen where the data already sits Enterprises are sold small specialist models that are cheaper and faster page 1 page 2 page 3	5
ParsewaveGraders Sells hard coding and other tasks tuned to each lab model, made exclusively for one lab and never resold, starting with a small pilot. Pay is often per accepted task, with referral fees and paid training on top Tasks are tuned to a pass-rate band on current models Software engineers are paid by the hour, or per task, to build coding tasks Payment follows acceptance; rejected hours, clips and files earn nothing Buyers often sample or pilot first, and some vendors describe orders arriving in bursts page 1 page 2 page 3	5
ReasonCoreGradersProofCompany recordsSafety Sells expert-written reasoning datasets with verified solutions, exclusive or off-the-shelf, plus enterprise-style reinforcement-learning environments and adaptive evals, to frontier labs and enterprises; it also buys domain data from businesses. Vendors sell graders built to resist cheating, and audits that look for it Testers are paid to measure dangerous abilities and bad behaviour, not just skill Testers measure what agents do unprompted: lying, collusion, overstepping, self-modification Vendors pitch ordinary businesses on licensing their own data as a new revenue line Vendors sell data built to target named public benchmarks page 1	5
UlamGradersProof A team of research mathematicians selling unpublished maths problems, machine-checked proof traces, verifier-backed RL tasks and private benchmarks to frontier labs and AI developers. Vendors sell tasks in fields where a tool, not a person, settles correctness Buyers often sample or pilot first, and some vendors describe orders arriving in bursts Vendors sell diagnosis: find where the model fails, then supply the fix Machine-checked proofs are wanted; one vendor says the appeal is a grader that cannot be fooled Labs commission and own research-level maths problems, audits and verifiers page 1 page 2	5
Veris AI A simulation platform that recreates the users, tools and data an enterprise AI agent works with, so companies can test and reinforcement-train their agents; sold to enterprises and now self-serve. The customer's own traces, workflows and evals become the training material Logs of live agents, especially humans correcting them, are treated as prime training data Enterprises are sold small specialist models that are cheaper and faster Vendors sell private tests to help enterprises choose models and judge agent readiness Stand-in data lets some firms train without exposing customer records page 1 page 2 page 3	5
ApturaFinanceMedicine A London firm that hires doctors, nurses, radiologists, lawyers and finance professionals to build training data, evaluations and RL environments from real professional work for frontier labs and specialist AI companies. Pay is often per accepted task, with referral fees and paid training on top Banking and private-equity deliverables recur across vendors: models, decks, diligence Practising clinicians are paid to write hard cases and grade model reasoning Contributor terms at several vendors assign all rights in the work to the vendor page 1 page 2 page 3 page 4	4
ARIMLABSSafety A Warsaw lab building long-running cybersecurity simulations and other verifiable training environments, plus benchmarks, for frontier-model developers. Testers are paid to measure dangerous abilities and bad behaviour, not just skill Testers measure what agents do unprompted: lying, collusion, overstepping, self-modification Cyber is becoming a training domain, with exploit-or-patch as the reward Cyber tests now span whole networks and run for hours, not single puzzles page 1 page 2	4
AviroGradersProof Builds simulated enterprise worlds, with document corpora and clones of business software, and sells them as post-training datasets and reinforcement-learning environments to frontier labs and enterprises. Vendors show a model improving on their data before they ask for an order Pass/fail is giving way to partial credit, rubrics and hybrid graders Vendors sell replicas of named work apps for agents to act in Small sets of expert-built tasks are presented as enough page 1	4
Build AIGradersRobots Makes head-worn cameras, puts them on factory workers through factory management, and publishes the first-person footage free under an open licence. No paid product, price list or customer is described on any page that could be read. Vendors now capture skilled work in real workplaces as well as household chores Several robot builders pre-train on people's first-person video, some with no robot data at that stage Robot, video and audio buyers grade data on variety, spec and labels One seller gives away plain factory video; top buyers ask for depth, stereo and a higher frame rate page 1 page 2 page 3 page 4	4
DisseiGradersFinance Finance training data, sealed evaluations and long-running RL environments built from anonymised records of its own lending deals, sold to AI labs and enterprise finance teams. The unit sold is a task with its grader, not a bare environment Real financial paperwork is bought as raw material, from firms and individuals Tools and tasks are frozen at a past date so the answer cannot leak in Some vendors run a real business so its daily work becomes data to sell page 1 page 2 page 3	4
EmulatedGradersCode Ready-made post-training datasets and production-like software environments, with live services and sealed graders, for teams training agents on coding, infrastructure operations, ML engineering and research. Vendors sell graders built to resist cheating, and audits that look for it Coding demand is spreading to kernels, infrastructure, migrations and scientific computing AI research itself is a named science target: agents doing ML engineering Incidents show evaluation and training environments need to be sealed and tamper-resistant page 1 page 2	4
FirstsourceRobots A labour pool ("Gigsourcing Platform") pointed at robot data: first-person video collection in homes with touch and body-position streams, teleoperation data pipelines, matched first- and third-person video, and defect image labelling. Several robot builders pre-train on people's first-person video, some with no robot data at that stage Capture is made to order and paid per accepted hour Touch and force, recorded by gloves, are being added to video Payment follows acceptance; rejected hours, clips and files earn nothing page 1	4
IncalmoSafety Cybersecurity RL environments and evaluations, including network-scale attack ranges, for frontier labs, security vendors and enterprises; it also trains its own cyber models on them. Environments and tasks are increasingly generated by software, not hand-built Cyber is becoming a training domain, with exploit-or-patch as the reward One cyber tester says labs buy its offensive-cyber tests before releases Cyber tests now span whole networks and run for hours, not single puzzles page 1 page 2 page 3	4
LightWrkGradersRobots Expert judgements of robot behaviour, such as comparisons, scores and failure analyses, from robot operators and tradespeople, for teams training robot and world models. Expert judgement is bought to write rubrics, gold answers and calibrate judge models A long tail of narrow jobs: trades, inspection, defence, design, language Variety, rare cases and failures are valued above more hours Some vendors scale contributor pay with the quality of the work page 1 page 2	4
TraverseGradersCompany records Sells labs training data that records professionals doing real, ambiguous work in fields such as medicine and law, with their reasoning. Many vendors name law among their professional domains Expert judgement is bought to write rubrics, gold answers and calibrate judge models The record of how work happened is worth more than the finished product Vendors sell judgement for work that has no single right answer page 1 page 2	4
AbundantGraders Sells RL environments and training datasets, including automatically generated coding tasks from GitHub repos, to AI labs and large enterprises. Vendors sell graders built to resist cheating, and audits that look for it Environments and tasks are increasingly generated by software, not hand-built Repositories are converted into graded tasks, sometimes for one customer's own model page 1	3
Align AIGraders Sells expert-written training data, graded practice environments and checks of agent behaviour against laws and company rules, aimed at regulated work such as insurance, to labs and enterprises. Vendors sell graders built to resist cheating, and audits that look for it The unit sold is a task with its grader, not a bare environment Vendors turn industry rules and operators’ unwritten know-how into training environments page 1 page 2	3
CollinearGradersProofSafety Sells practice environments, task sets and graders built around the failures of frontier models, plus simulated customers that enterprises use to test agents before launch. Vendors sell graders built to resist cheating, and audits that look for it Vendors sell diagnosis: find where the model fails, then supply the fix Cyber is becoming a training domain, with exploit-or-patch as the reward page 1 page 2	3
Contra LabsVideo The AI data arm of the freelance network Contra: it pays professional creatives to rank model outputs and sells creative preference data, corrective training data, computer-use trajectories and evaluations to AI labs. Contributor terms at several vendors assign all rights in the work to the vendor People are paid to rank and fault-find generated video Working designers and creatives are paid to judge what models make page 1 page 2	3
CueBench Sells training environments for scientific reasoning and performance engineering, aimed at hard tasks where frontier models still mostly fail. Vendors publish website terms but not their lab contracts; one says fees sit in separate order forms Science and engineering tasks are built around simulations and recomputable answers Science tasks test whether agents can design and run experiments with limited resources page 1 page 2	3
e2f (now TrustScale)Graders Sells expert-written benchmarks, red-teaming and graded evaluations, and is shifting toward automated tools that check model answers with humans on hard cases. Pass/fail is giving way to partial credit, rubrics and hybrid graders Expert judgement is bought to write rubrics, gold answers and calibrate judge models Quality control increasingly pairs automated checks with expert auditors page 1 page 2	3
General ReasoningGraders A London research lab that hosts RL environments as pay-per-use endpoints for labs and startups to train and test agents, and sells a frozen web-search archive for leak-free evaluation. Pass/fail is giving way to partial credit, rubrics and hybrid graders Training and evaluation need very many isolated sandboxes at once Tools and tasks are frozen at a past date so the answer cannot leak in page 1 page 2 page 3	3
Habitat IncGraders Pays engineers to turn real open-source commits into hard coding tasks for AI training; its main site now presents it as a firm that buys businesses and runs them with AI. Tasks are tuned to a pass-rate band on current models Software engineers are paid by the hour, or per task, to build coding tasks Repositories are converted into graded tasks, sometimes for one customer's own model page 1	3
Magma A broker: a company running an AI agent connects its database; Magma strips identifying details, packages the agent logs (actions, user corrections, finished jobs) into training environments and licenses them to AI labs, sharing each licence fee with the supplier. Many vendors license data non-exclusively and keep the underlying asset Non-exclusive licences are the default; the same data is sold to several buyers Logs of live agents, especially humans correcting them, are treated as prime training data page 1	3
Monte Sells a post-training platform with embedded researchers that turns a company’s own traces and expert judgement into evaluations and trained open-weight models. It does not sell third-party training data. The customer's own traces, workflows and evals become the training material Sensitive buyers want training to happen where the data already sits Vendors sell private tests to help enterprises choose models and judge agent readiness page 1	3
NormalGraders An applied research lab that publishes CAD benchmarks, datasets and verifiers testing whether AI agents can turn engineering drawings into parts, offering fuller access to AI builders on request. Vendors sell tasks in fields where a tool, not a person, settles correctness Vendors recruit and collect for specialist tools such as CAD, chip-design, energy and spreadsheet software Vendors prove each task can be solved before they ship it page 1	3
Rise Data LabsFinance US-based expert data, evaluations and custom RL environments built from a client's own workflows, for AI labs and enterprise model teams. Accounting back-office work is a named target: closing the books and reconciliation The customer's own traces, workflows and evals become the training material Contributor work sessions are themselves instrumented and recorded page 1 page 2	3
UniPatGraders Sells customised reinforcement-learning data to Alibaba, which also supplies it with capital and compute. It has published a research-grade science dataset graded by rubric and built with expert time. Pass/fail is giving way to partial credit, rubrics and hybrid graders The buyer set is widening to Chinese, state and open-model builders Data is ordered to fit a named training method, including agent-recovery examples page 1 page 2	3
Archal Sells resettable simulated copies of common work apps so agent builders and training teams can start every run from the same state; priced per seat with usage add-ons. Vendors sell replicas of named work apps for agents to act in Training and evaluation need very many isolated sandboxes at once page 1 page 2	2
CajalGraders AI agents ("Tau") that write machine-checked mathematical proofs in Lean, pitched at quantum computing and finance; as a second line, proof datasets, evals and RL environments for frontier AI labs and research institutes. Vendors sell tasks in fields where a tool, not a person, settles correctness Machine-checked proofs are wanted; one vendor says the appeal is a grader that cannot be fooled page 1	2
China robot-data market (cross-vendor notes)GradersRobots Market notes on robot data in China: hourly prices, industry capacity and funding of independent data firms. Robot, video and audio buyers grade data on variety, spec and labels One seller gives away plain factory video; top buyers ask for depth, stereo and a higher frame rate page 1	2
CuaGraders Sells open-source and hosted desktop sandboxes for computer-use agents, plus a catalogue of computer-use environments, verifiable tasks and recorded runs for model builders. Pass/fail is giving way to partial credit, rubrics and hybrid graders Vendors recruit and collect for specialist tools such as CAD, chip-design, energy and spreadsheet software page 1 page 2	2
exponential (Exponential Security Labs)Safety Builds multi-agent security environments, trains attacker and defender agents in them, and sells adaptive security benchmarking to AI developers. Cyber is becoming a training domain, with exploit-or-patch as the reward Environments now pit several agents against each other, as rivals or opponents page 1	2
Foundry AIGradersProof Custom and ready-made RL dataset suites for coding and multimodal agents, with hidden verifiers, rubrics and expert-corrected trajectories, plus private evaluations, made with vetted senior engineers for model teams. Tasks are tuned to a pass-rate band on current models Vendors sell diagnosis: find where the model fails, then supply the fix page 1 page 2	2
hiloop Sells automated research as a service: customers bring a model and an evaluation, and it runs many experiments and returns the best verified improvement; it also generates and scores agent training environments. Environments and tasks are increasingly generated by software, not hand-built The customer's own traces, workflows and evals become the training material page 1	2
Olam LabsSafety Builds multi-agent simulated environments, including a public arena where people play games against AI models, to evaluate and train frontier models for labs and researchers. Testers measure what agents do unprompted: lying, collusion, overstepping, self-modification Environments now pit several agents against each other, as rivals or opponents page 1 page 2	2
Poindexter Labs Expert annotation and evaluation for frontier labs from Olympiad medallists and senior academics, plus a licensable expert-training platform that governments and enterprises can run privately. Sensitive buyers want training to happen where the data already sits One vendor sells training by the token; several report that a single evaluation run is costly page 1	2
Quesma Sells analytics on companies' coding-agent sessions; says it has also designed RL training environments for frontier labs and publishes open coding-agent benchmarks. The customer's own traces, workflows and evals become the training material Logs of live agents, especially humans correcting them, are treated as prime training data page 1	2
ReveloGraders A Latin American engineering network whose engineers write coding datasets, containerised RL environments with verifiers, and evaluations that it packages and sells to AI labs; it also places engineers with US companies. The unit sold is a task with its grader, not a bare environment Buyers often sample or pilot first, and some vendors describe orders arriving in bursts page 1 page 2	2
SieveVideo Sells licensed video and media datasets, including before-and-after edit pairs, to labs and companies building media, robot and agent models. Consent records, licence trails and security certificates are now a gate to deals Synchronised multi-camera footage and raw-versus-edited pairs sell as their own products page 1	2
SymbalProofFinance Sells evals, benchmarks and post-training datasets in finance, science, medicine, law, audio and robotics, built by an expert pool, to frontier labs; it also builds agents on enterprises’ own data. Finance experts are hired to write problems and rubrics and to grade output Vendors sell data built to target named public benchmarks page 1 page 2	2
Andromede Generates reinforcement-learning environments, tasks and verifiers from real-world data for post-training and testing long-horizon agents. It says it works with a small set of unnamed partners. Environments and tasks are increasingly generated by software, not hand-built page 1	1
BenchFlowGraders Builds agent benchmarks and stateful replica workplaces with verifiers, plus an open community arena for post-training task packages. Its pages show mostly open releases and no priced offering. The unit sold is a task with its grader, not a bare environment page 1	1
Cambridge Crystallographic Data Centre Licenses its crystal structure data under standard terms that bar use inside AI systems unless it expressly agrees otherwise. Long-established crystal-structure databases hold AI-training rights back from ordinary access, and one buyer says it holds them exclusively page 1	1
Defined.ai (audio pass) A marketplace of ready-made training datasets (its own and third parties'), plus made-to-order collection and annotation. Began as a speech-data company; now sells text, voice and image data. Consent records, licence trails and security certificates are now a gate to deals page 1	1
EdotEnvLong tasks Builds multi-step reinforcement-learning environments from financial market data, where agents run quant-research loops, and sells them to AI labs, researchers and enterprises. Task length and difficulty must keep rising as models improve page 1	1
GOMAX Lab Supplies expert-built data corpora and Chinese-language benchmarks for scientific and industrial AI, made with industry partners. It also runs a research project with a state information body. A long tail of narrow jobs: trades, inspection, defence, design, language page 1	1
JD.com (robot-data collection plan)Robots JD.com is collecting real-scene human video across warehouses, factories and homes for embodied AI. It uses the data for its own models, offers it through a data trading platform and opens some sets to academia. Vendors now capture skilled work in real workplaces as well as household chores page 1	1
LXTRobots Sells custom and ready-made AI training data, from collection and annotation to RLHF and red-teaming, across text, speech, image and video to tech companies and enterprises, and rents its contributor crowd directly by API. Large labelling firms now list first-person video for robots alongside their older lines page 1	1
Maingen Sells simulations of industrial operations, built with the people who run them, so labs can train and test agents on that work. Vendors turn industry rules and operators’ unwritten know-how into training environments page 1	1
Maxinsights Sells robot training data, collected and labelled to client specification across several regions, to builders of robot models. Sourcing works backwards from what labs ask for; one vendor says it can start within days page 1	1
Metaphi Builds reinforcement-learning environments, persona simulations and evaluations checked against expert human trajectories, for frontier AI companies and enterprises; it acquires proprietary environments such as a real bank’s COBOL system. Private code is wanted because public code no longer gives a clean measurement page 1	1
NoumadGraders Sells training environments built from unsolved research problems, with rewards checked by formal methods, to frontier labs. Vendors sell graders built to resist cheating, and audits that look for it page 1	1
OpenTrain AI A marketplace where AI companies hire vetted trainers, red-teamers and labellers to work inside the buyer's own tools, with an optional managed service. Vendors publish website terms but not their lab contracts; one says fees sit in separate order forms page 1	1
Pareto Recruits specialists such as scientists, lawyers and clinicians to produce expert training data for model builders, and publishes research benchmarks. Vendors say demand has moved from large crowds to small groups of experts page 1	1
PixelstwoRobots Sells tactile and dexterous manipulation data captured with its own touch-sensing hardware, collected in the United States, India and Southeast Asia, to robotics teams, made to order. Touch and force, recorded by gloves, are being added to video page 1	1
RL LabsSafety Builds long-horizon RL environments for training frontier models in software engineering, security, computer use and incident response. Cyber is becoming a training domain, with exploit-or-patch as the reward page 1	1

Model builders 21

Buyers and makers of data, not sellers: companies that build models for robots or for language, whose pages or listings say what they train on, collect or buy.

Vendor	Signals	Graders	Proof	Long tasks	Code	Office work	Company records	Finance	Medicine	Safety	Robots	Video	Speech	Biology
DynaRobots Robot-model builder. Pre-trains on human first-person video, then a little robot data per task; runs its own operation and deployed robots; also takes video from data partners. Several robot builders pre-train on people's first-person video, some with no robot data at that stage Builders split on buying versus collecting in-house, and on exclusivity Variety, rare cases and failures are valued above more hours Robot builders and vendors feed data from deployed robots back into training Some builders report needing only minutes to hours of robot data per new task Video and robot builders are sizing pipelines for very large volumes of footage page 1 page 2 page 3	6
1XRobots Robot-model builder. Trains on web video, human first-person video and a smaller amount of robot data; operators at its own site and customers' robots; a named outlet reports a data vendor as a supplier. Several robot builders pre-train on people's first-person video, some with no robot data at that stage Robot operators are paid more by the hour than people who film at home Builders run teleoperation in-house with hourly operators, in one case around the clock Some builders report needing only minutes to hours of robot data per new task Video and robot builders are sizing pipelines for very large volumes of footage page 1 page 2 page 3	5
FigureGradersRobots Robot-model builder. Trains on its own crowd video and on internet video; collects through its own app and collectors in sensor suits; says it stopped buying from vendors. Several robot builders pre-train on people's first-person video, some with no robot data at that stage Robot operators are paid more by the hour than people who film at home Robot, video and audio buyers grade data on variety, spec and labels Builders split on buying versus collecting in-house, and on exclusivity Variety, rare cases and failures are valued above more hours page 1 page 2 page 3 page 4	5
Apptronik Robot-model builder. Collects humanoid data that feeds Google DeepMind's robotics models, at its own site and at customers' sites. Robot operators are paid more by the hour than people who film at home Robot builders and vendors feed data from deployed robots back into training Builders run teleoperation in-house with hourly operators, in one case around the clock Failed, sloppy and autonomous robot runs are kept as training data page 1 page 2	4
NVIDIA (GR00T, EgoScale)ProofRobots Robot-model builder. Trained on human first-person video and a small amount of its own teleoperation, multiplied by video generation; publishes open datasets. Several robot builders pre-train on people's first-person video, some with no robot data at that stage Video is delivered with step labels, hand pose and recovered geometry Robot and biology builders publish scaling results that justify buying more data Robot-model builders lean on simulation and generated video, mixed with real data page 1 page 2	4
SkildRobots Robot-model builder. Trains on human video, simulation, teleoperation and deployment logs; says it scales all collection in house. Robot operators are paid more by the hour than people who film at home Builders split on buying versus collecting in-house, and on exclusivity Builders run teleoperation in-house with hourly operators, in one case around the clock Robot-model builders lean on simulation and generated video, mixed with real data page 1 page 2 page 3	4
Agility Agility builds the Digit humanoid robot and collects its own training data by teleoperating it in-house and at customer sites. It also trains control models in simulation. Robot builders and vendors feed data from deployed robots back into training Builders run teleoperation in-house with hourly operators, in one case around the clock Robot-model builders lean on simulation and generated video, mixed with real data page 1 page 2 page 3	3
Alibaba Model builder. Its listings show it buying base, industry, multimodal and customised corpora and labelling from outside vendors, including international ones, while task standards stay with its own teams. The buyer set is widening to Chinese, state and open-model builders Chinese model builders are hiring staff to buy training data from overseas suppliers Chinese model builders keep standards and evaluation in house and buy corpora and labour outside page 1 page 2	3
GeneralistRobots Robot-model builder. Pre-trains on human-wearable data with no robot data, then a little robot data per task; ships its own wearables; says it has several data foundry partners; named as a customer by Scale AI. Several robot builders pre-train on people's first-person video, some with no robot data at that stage Builders split on buying versus collecting in-house, and on exclusivity Some builders report needing only minutes to hours of robot data per new task page 1 page 2	3
Genesis AIRobots buyer — trains on glove data, first-person video and internet video (no hours); collects with its own tactile glove on partners' workers; 'could also pay third-party partners' (TechCrunch). Vendors now capture skilled work in real workplaces as well as household chores Builders split on buying versus collecting in-house, and on exclusivity Touch and force, recorded by gloves, are being added to video page 1 page 2	3
Physical IntelligenceGraders Robot-model builder. Trains on data from many robots, autonomous runs, human first-person video, open sets and web data; collects by teleoperation in shifts; named as a customer by Scale AI. Robot, video and audio buyers grade data on variety, spec and labels Builders run teleoperation in-house with hourly operators, in one case around the clock Failed, sloppy and autonomous robot runs are kept as training data page 1 page 2	3
ByteDance Model builder. Its pages show one department covering data standards, sourcing and procurement, with a large outside data budget; it keeps evaluation staff in house, recruits its own subject specialists to write questions, and says it will not distil. The buyer set is widening to Chinese, state and open-model builders Chinese model builders keep standards and evaluation in house and buy corpora and labour outside page 1 page 2	2
Tencent Model builder. Its listing shows it buying training data from overseas suppliers; the notes show nothing on what it keeps inside. The buyer set is widening to Chinese, state and open-model builders Chinese model builders are hiring staff to buy training data from overseas suppliers page 1	2
CuspAI Model builder. Says it holds exclusive AI training rights to the main experimental crystal-structure databases and a publisher's materials catalogues. Long-established crystal-structure databases hold AI-training rights back from ordinary access, and one buyer says it holds them exclusively page 1	1
Google DeepMind Robot-model builder. Collects its own two-arm teleoperation on its own fleet; Apptronik collects humanoid data for its robotics models. Some builders report needing only minutes to hours of robot data per new task page 1	1
Moonlake buyer — world-model builder; trains on two-arm robot logs replayed through its simulator (no hours); hiring to 'collect and curate real-world robotics datasets'; uses data in the format of XDOF's open set. One world-model builder is collecting real robot logs; one free set has been used as a test set page 1	1
Periodic Labs Model builder. Trains its own model on data from its own laboratories rather than bought data. Materials labs that make their own experimental data train their own models on it or hold it as their asset; none lists the dataset for sale page 1	1
RARA Factory A laboratory that makes and measures new materials daily and keeps the resulting dataset, failures included, as its own asset; no data sale is offered. Materials labs that make their own experimental data train their own models on it or hold it as their asset; none lists the dataset for sale page 1	1
Tesla OptimusRobots Tesla collects training data for its Optimus robot in-house, employing operators who perform movements wearing motion-capture gear and helmet cameras. Several robot builders pre-train on people's first-person video, some with no robot data at that stage page 1	1
World Labs World Labs sells a world model that generates spatial scenes, pitched as a source of robot simulation and training data. Its pages do not say where its own training data came from. Environments and tasks are increasingly generated by software, not hand-built page 1	1
World-model builders (general notes)Video World-model builders as a class: they license gameplay video with controls, and Chinese vendors count them as buyers of robot-free human data. Sellers of first-person and robot data name world-model builders as customers alongside robot labs page 1	1