Skip to content
Slow Variables
Ask
Who profits · the data market

What are the people who train AI buying?

AI models learn from examples. The public internet supplied the first of them for free. What comes next has to be made, bought or recorded, and a trade has grown up to supply it.

Each mark is one kind of record that work leaves behind and that could teach a model how the work is done: 803 of them, by family. The fill says whether anyone is yet selling it, giving it away or using their own to train AI. Every kind of record, with what was found.

  • Openly sold or released for training AI 133
  • An incumbent uses its own 327
  • Only nearby players 229
  • Nobody found 110
  • Not searched 4
  • Examined in a dossier 84
Insolvency and custody events 66
Health and veterinary 58
Construction and property 46
Manufacturing and maintenance 43
Insurance and claims 38
Government decisions and public registers 37
Agriculture and food grading 34
Courts, tribunals and disputes 34
Video and physical capture 33
Work traces and software logs 29
Retail, hospitality and consumer services 28
Audio and speech 27
Mining, energy and resources 25
Law, accounting and professional services 24
Media, creative and publishing work 24
New capture products/engines 24
Transport, marine and aviation 24
Education, training and exams 22
Finance, lending and advice 19
Games, sport and simulators 19
Labs, testing and science 18
Valuation, auctions and second-hand goods 16
Market infrastructure, tools and evals 15
Vendor-made data, environments and synthetic 15
Emergency services, policing and defence 14
Sales, office and back-office operations 14
Employment, HR and workplace safety 13
Environment, water and waste 9
Freight, logistics and customs 9
Published, licensed and public-web content 8
Payments, fraud and identity 7
Consumer, platform and AI-product usage data 4
Forecasting, betting and prediction 4
Enterprise company internals (general) 3

The largest families are shown. See every family.

How this is read

Model builders do not publish what they buy. The firms that sell to them do, without meaning to. A vendor's job listings say which experts it needs. Its research posts say which problems its customers have set it. Its terms say what a customer may do with the goods.

This section reads those pages, each fetched and dated. Nothing comes from social media. What a page says is the company's own account of itself; the site reports that it says so and does not vouch that it is true.

A model's reading of the vendors' own pages, not a measurement. A model read the pages and wrote every line here; further runs of the same model checked it, and no person reviewed each line. 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 it was made, and what is wrong with it.

Vendors read
93
with evidence of fast growth, and 76 more on weaker evidence
Model builders
21
whose pages say what they train on or buy
Signals
230
things the pages show buyers asking for
Pages cited
726
each fetched under this site's rules

Who pays whom, and which prices are published

Money runs from model builders to vendors, and from vendors to the people and companies whose work or records become the data. The number on each arrow is how many published figures this site could quote for it; the faint arrow is the one almost nobody publishes. Every figure, with its passage.

The signals

Each row is one thing the pages show buyers asking for: 230 in all, the most widely shown first. The bar and the figure beside it count the vendors with good evidence of growth whose own pages show it; vendors with weak evidence are counted apart, hatched. Open a row for what it implies and for each vendor's page.

  1. Vendors show a model improving on their data before they ask for an orderProof that the data improves a model20+5 weak evidence

    What it implies buyers want: Evidence of uplift attached to each dataset, and trial terms that shift the risk to the seller.

  2. Vendors sell graders built to resist cheating, and audits that look for itGraders and quality control17+11 weak evidence

    What it implies buyers want: Rewards designed so a high score means the work was done, with graders attacked before delivery.

  3. Pass/fail is giving way to partial credit, rubrics and hybrid gradersGraders and quality control17+10 weak evidence

    What it implies buyers want: Graders that score how far the agent got on decks, spreadsheets, legal work and long builds, not only whether it finished.

  4. Expert judgement is bought to write rubrics, gold answers and calibrate judge modelsGraders and quality control16+7 weak evidence

    What it implies buyers want: Senior professional judgement converted into reusable scoring rules and judge-calibration sets.

  5. Quality control increasingly pairs automated checks with expert auditorsGraders and quality control14+5 weak evidence

    What it implies buyers want: Delivered data that has already been through independent checks and audit.

  6. Vendors now capture skilled work in real workplaces as well as household choresRobots: recordings of people14+4 weak evidence

    What it implies buyers want: Footage of paid skilled work filmed where it happens, alongside household capture.

  7. Several robot builders pre-train on people's first-person video, some with no robot data at that stageRobots: recordings of people13+5 weak evidence

    What it implies buyers want: Human first-person video of manual tasks in very large volume, to pre-train robot models before any robot data is added.

  8. Many vendors license data non-exclusively and keep the underlying assetPrices and terms12+1 weak evidence

    What it implies buyers want: Vendors offer use-limited training rights rather than ownership, with paperwork attached.

  9. Coding tasks must be audited for leaked answers, broken tests and gameable gradersCoding11+1 weak evidence

    What it implies buyers want: Coding tasks that arrive pre-audited, with scrubbed history and fair, deterministic graders.

  10. Vendors sell simulated companies: cloned work apps filled with files, chat and tasksOffice work11+2 weak evidence

    What it implies buyers want: Whole-company practice worlds, in volume, organised by job and industry.

  11. Bought records are turned into practice environments, not handed to labs rawCompanies' own records10+4 weak evidence

    What it implies buyers want: Environments, tasks and verifiers built from real records, with the natural mess left in.

  12. Vendors now buy whole companies' internal records: chat, email, CRM, code, documentsCompanies' own records10+4 weak evidence

    What it implies buyers want: Records of how real companies do work, across many industries, that are not on the public internet.

  13. Vendors sell tasks in fields where a tool, not a person, settles correctnessGraders and quality control10+4 weak evidence

    What it implies buyers want: Tasks in maths, chips, cyber, science and data science whose rewards come from running a checker.

  14. Video is delivered with step labels, hand pose and recovered geometryRobots: recordings of people10+2 weak evidence

    What it implies buyers want: Time-stamped action text, hand and body pose and scene geometry delivered ready to train on, not raw footage.

  15. Speech in under-served languages and dialects, Indic and Arabic singled outAudio and speech9+1 weak evidence

    What it implies buyers want: Conversational audio and transcripts in dialects and languages beyond the few that commercial systems already speak.

  16. Coding tasks now run for hours or days and rebuild whole programsCoding9+2 weak evidence

    What it implies buyers want: A steady supply of new, very long, product-scale software tasks the best model still fails.

  17. Labs hand unreleased models to outside testers for scores to publish at launchPrivate tests9

    What it implies buyers want: Independent, fast, confidential assessment timed to release, credible to customers and regulators.

  18. Buyers want test sets that no model has seen and none can gamePrivate tests9+2 weak evidence

    What it implies buyers want: Fresh, sealed evaluation data with contamination controls, often held and graded by a third party.

  19. The unit sold is a task with its grader, not a bare environmentGraders and quality control9+7 weak evidence

    What it implies buyers want: Ready-to-run task packages that drop into the lab's training harness and score deterministically.

  20. Environments and tasks are increasingly generated by software, not hand-builtMachine-made data9+5 weak evidence

    What it implies buyers want: Large numbers of machine-made environments, with automated checks on their quality.

  21. Private codebases with full history are the wanted raw material for codingCoding8+3 weak evidence

    What it implies buyers want: Never-public, team-built code with its history, to seed training tasks and clean tests.

  22. Coding demand is spreading to kernels, infrastructure, migrations and scientific computingCoding8+1 weak evidence

    What it implies buyers want: Expert-built tasks in systems, performance and ML engineering, well beyond Python bug fixes.

  23. Company data is de-identified and rights-checked before use; one licence also bars scraped and machine-made contentCompanies' own records8+3 weak evidence

    What it implies buyers want: Training-safe records: personal details removed, rights confirmed, and in some cases the source company hidden.

  24. Testers are paid to measure dangerous abilities and bad behaviour, not just skillPrivate tests8+2 weak evidence

    What it implies buyers want: Independent risk and propensity scores alongside capability scores, including AI-research acceleration.

  25. Non-exclusive licences are the default; the same data is sold to several buyersExclusivity and first use8+1 weak evidence

    What it implies buyers want: Labs accept shared data and buy access or speed rather than ownership.

  26. Pay is often per accepted task, with referral fees and paid training on topWhat experts are paid8+2 weak evidence

    What it implies buyers want: Output-priced expert labour, and vendors paying to recruit scarce experts.

  27. Robot operators are paid more by the hour than people who film at homeWhat experts are paid8

    What it implies buyers want: Robot-native demonstrations at US wages, plus cheap bulk first-person video.

  28. Tasks are tuned to a pass-rate band on current modelsGraders and quality control8+6 weak evidence

    What it implies buyers want: Hard-but-solvable tasks delivered with measured pass rates, and re-measured as models improve.

  29. Robot, video and audio buyers grade data on variety, spec and labelsGraders and quality control8+3 weak evidence

    What it implies buyers want: Diverse, densely annotated, spec-compliant sensor data with metadata, not raw hours.

  30. Tasks, harnesses and delivery are converging on shared formatsInfrastructure8+2 weak evidence

    What it implies buyers want: Deliverables that plug into the lab's existing harness, evaluation framework and data loaders.

  31. Finding and vetting experts fast is treated as a product in itselfOther8+3 weak evidence

    What it implies buyers want: Vetted specialists on demand, matched to whatever labs want this month.

  32. Builders split on buying versus collecting in-house, and on exclusivityRobots: recordings of people8

    What it implies buyers want: Some labs want outside collection at large weekly volumes; others want only the capture hardware, or nothing.

  33. Capture is made to order and paid per accepted hourRobots: recordings of people8+3 weak evidence

    What it implies buyers want: Custom collection to each lab's schema, with acceptance thresholds and proof on the lab's own benchmarks.

  34. Touch and force, recorded by gloves, are being added to videoRobots: recordings of people8+3 weak evidence

    What it implies buyers want: Contact, force and finger-level motion time-locked to video, which cameras alone cannot give.

  35. PhD-level contributors are recruited across many fields and countriesScience and mathematics8

    What it implies buyers want: Vetted graduate scientists for problem writing, review and audit.

  36. Evaluation is sold as the front end of a data salePrivate tests7+1 weak evidence

    What it implies buyers want: A vendor that measures the gap, then sells the tasks that close it.

  37. Profession-specific private benchmarks are built from real records with practising expertsPrivate tests7+1 weak evidence

    What it implies buyers want: Evaluations in law, finance, tax, insurance, medicine and voice grounded in real working documents.

  38. Benchmarks saturate fast, and testers keep replacing them with harder onesPrivate tests7+1 weak evidence

    What it implies buyers want: New expert-built tests that still separate the top models, refreshed as each one is passed.

  39. Senior lawyers, bankers and doctors are offered some of the highest hourly ratesWhat experts are paid7

    What it implies buyers want: Small volumes of top professional judgement, turned into tasks, rubrics and graded answers.

  40. Vendors and testers build software tasks that take an agent hours, days or a weekLong tasks7+2 weak evidence

    What it implies buyers want: Very long, exactly graded software projects, and training set-ups for orchestrator-plus-sub-agent work.

  41. The buyer set is widening to Chinese, state and open-model buildersOther7+1 weak evidence

    What it implies buyers want: Launch-driven, custom data for a broader set of model builders.

  42. A long tail of narrow jobs: trades, inspection, defence, design, languageOther7+2 weak evidence

    What it implies buyers want: Practical know-how and specialist media that sit outside the big professional domains.

  43. Vendors publish website terms but not their lab contracts; one says fees sit in separate order formsPrices and terms7+5 weak evidence

    What it implies buyers want: Custom paperwork per deal: data rights, usage limits, service levels and publicity rights.

  44. Variety, rare cases and failures are valued above more hoursRobots: recordings of people7+1 weak evidence

    What it implies buyers want: Coverage of many tasks, objects, tools and places, plus mistakes and recoveries, rather than repeats of one scene.

  45. Robot builders and vendors feed data from deployed robots back into trainingRobots: deployment7

    What it implies buyers want: Data from robots at work, obtained through deployment and partnership.

  46. Testers measure what agents do unprompted: lying, collusion, overstepping, self-modificationSafety and security7+3 weak evidence

    What it implies buyers want: Propensity tests in realistic deployments, plus monitors and environments built with lab safety teams.

  47. AI research itself is a named science target: agents doing ML engineeringScience and mathematics7+2 weak evidence

    What it implies buyers want: Open-ended ML-research tasks with hidden tests, scored against human records.

  48. Sellers of first-person and robot data name world-model builders as customers alongside robot labsVideo and world models7+1 weak evidence

    What it implies buyers want: First-person video, robot logs and scans of places with physical properties, sold once to robot and world-model teams alike.

  49. Software engineers are paid by the hour, or per task, to build coding tasksWhat experts are paid6+3 weak evidence

    What it implies buyers want: Production-experienced engineers writing, reviewing and auditing agent coding tasks.

  50. Accounting back-office work is a named target: closing the books and reconciliationFinance6+1 weak evidence

    What it implies buyers want: Tasks that must be right every time, inside accounting systems.

  51. Finance experts are hired to write problems and rubrics and to grade outputFinance6+1 weak evidence

    What it implies buyers want: Practitioners with finance credentials producing ground truth.

  52. Vendors are adding insurance, credit, tax and compliance work alongside investment bankingFinance6+1 weak evidence

    What it implies buyers want: The same task-and-rubric data for neighbouring financial professions.

  53. Vendors sell replicas of named work apps for agents to act inUse of named software6+5 weak evidence

    What it implies buyers want: Frozen, faithful copies of real software, because live sites change and block bots.

  54. Data is ordered to fit a named training method, including agent-recovery examplesOther6+2 weak evidence

    What it implies buyers want: Demonstrations, preference data, reward data and environments, each to specification.

  55. Payment follows acceptance; rejected hours, clips and files earn nothingPrices and terms6+5 weak evidence

    What it implies buyers want: Buyers pay for verified quality and push reject risk down to suppliers and workers.

  56. Buyers often sample or pilot first, and some vendors describe orders arriving in burstsPrices and terms6+5 weak evidence

    What it implies buyers want: Try-before-buy samples, agreed acceptance tests, and campaign-style purchasing.

  57. Vendors sell diagnosis: find where the model fails, then supply the fixProof that the data improves a model6+3 weak evidence

    What it implies buyers want: A supplier that names the model's gap and proves the remedy against the lab's own numbers.

  58. Sellers publish benchmark gains from their data on open modelsProof that the data improves a model6+3 weak evidence

    What it implies buyers want: Data aimed at the tests that rankings weight most, with transfer to tests the seller did not write.

  59. Cyber is becoming a training domain, with exploit-or-patch as the rewardSafety and security6+7 weak evidence

    What it implies buyers want: Security tasks, environments and practitioner labour for training models to find and fix vulnerabilities.

  60. Testers and vendors now grade whether code is mergeable, not only whether tests passCoding5

    What it implies buyers want: Preference data, rubrics and rewards that capture engineering judgement and reliability, not only correctness.

  61. Vendors are hiring people whose whole job is buying data from companiesWho does the buying5+1 weak evidence

    What it implies buyers want: A standing supply of licensed private company data, sourced by the vendor on labs' behalf.

  62. The customer's own traces, workflows and evals become the training materialCustom models for enterprises5+6 weak evidence

    What it implies buyers want: Pipelines that convert a company's live usage into tasks, graders and training signal.

  63. Banking and private-equity deliverables recur across vendors: models, decks, diligenceFinance5+2 weak evidence

    What it implies buyers want: End-to-end deal work product in real file formats.

  64. Buyers want the same patient over time, across scans, notes, labs and genomicsHealthcare and medicine5

    What it implies buyers want: Linked, de-identified records broad enough for general medical models.

  65. Logs of live agents, especially humans correcting them, are treated as prime training dataPeople working with AI5+3 weak evidence

    What it implies buyers want: Correction events and tool-use sessions from real deployments, turned into environments.

  66. Sensitive buyers want training to happen where the data already sitsInfrastructure5+3 weak evidence

    What it implies buyers want: Train-in-place access and on-premises delivery in place of file transfer.

  67. Training and evaluation need very many isolated sandboxes at onceInfrastructure5+3 weak evidence

    What it implies buyers want: Cheap, isolated, production-like execution environments at reinforcement-learning volume.

  68. Day-long AI-research tasks are built to measure models improving modelsLong tasks5

    What it implies buyers want: Environments and evaluations for agents doing AI research, with graders that cannot be passed without doing the work.

  69. Task length and difficulty must keep rising as models improveLong tasks5+1 weak evidence

    What it implies buyers want: Each model generation needs a new, longer and harder set; last year's tasks lose value.

  70. Vendors recruit and collect for specialist tools such as CAD, chip-design, energy and spreadsheet softwareUse of named software5+3 weak evidence

    What it implies buyers want: People who can demonstrate and grade work inside specific professional applications.

  71. Labs arrive with vague goals and expect the vendor to shape the data fastOther5+1 weak evidence

    What it implies buyers want: Research-partner relationships with the post-training team, delivered in days.

  72. Vendors say environment supply, not lab budget, is the constraintOther5+1 weak evidence

    What it implies buyers want: A continuing supply of new, harder environments, mainly for training.

  73. Robot and biology builders publish scaling results that justify buying more dataProof that the data improves a model5

    What it implies buyers want: More hours and more varied data, with human video and licensed footage as proven supplements.

  74. Builders run teleoperation in-house with hourly operators, in one case around the clockRobots: remote operation5

    What it implies buyers want: Trained operators and robot uptime are the constraint; labs staff this themselves or pay a vendor to.

  75. Test and training environments are themselves being attacked by the modelsSafety and security5+1 weak evidence

    What it implies buyers want: Sandboxes and graders proven to contain a model that is actively trying to get out or cheat.

  76. Red-team labour is hired by risk area, credential and languageSafety and security5

    What it implies buyers want: Adversarial prompts, refusal-boundary sets and policy grading from credentialed specialists, in many languages.

  77. Robot-model builders lean on simulation and generated video, mixed with real dataMachine-made data5

    What it implies buyers want: Physically accurate simulation assets and generated video to stretch scarce real robot data.

  78. Private speech test sets and human ratings of voice models are being built and offeredAudio and speech4+2 weak evidence

    What it implies buyers want: Unseen conversational test data and paid human judgement of naturalness, empathy and timing.

  79. Vendors count coding environments by the task and quote costs per taskCoding4+1 weak evidence

    What it implies buyers want: Coding environments built by senior engineers, with cost tracked per task.

  80. The record of how work happened is worth more than the finished productCompanies' own records4+4 weak evidence

    What it implies buyers want: Linked, time-ordered records showing decisions and reasoning, not only outcomes.

  81. Enterprises are sold small specialist models that are cheaper and fasterCustom models for enterprises4+3 weak evidence

    What it implies buyers want: Data, environments and training runs that make an open model good at one company's job.

  82. Generalist work, and work outside the United States, is still listed at far lower ratesWhat experts are paid4+2 weak evidence

    What it implies buyers want: Bulk evaluation and non-English coverage, bought cheaply.

  83. Scientists and engineers with doctorates are paid by the hourWhat experts are paid4+1 weak evidence

    What it implies buyers want: Graduate-level problems, formal proofs and engineering judgement, at lower rates than law or finance.

  84. Vendors pay their own task builders and researchers lab-level salariesWhat experts are paid4+1 weak evidence

    What it implies buyers want: Environment and benchmark design by senior researchers, not contractors.

  85. Game footage is sold with the player's keyboard, mouse or controller inputsGameplay4+1 weak evidence

    What it implies buyers want: Video time-locked to the inputs that produced it, in bulk, for world models and game-playing agents.

  86. Real doctor-patient conversations and dictation are a named audio categoryHealthcare and medicine4

    What it implies buyers want: Real clinical speech with checked transcripts, where drug and lab terms must survive.

  87. Practising clinicians are paid to write hard cases and grade model reasoningHealthcare and medicine4+3 weak evidence

    What it implies buyers want: Clinical scenarios, multi-rater ground truth and corrected step-by-step reasoning.

  88. Independent evaluation on real patient records is becoming a product of its ownHealthcare and medicine4

    What it implies buyers want: A third party that holds real clinical test data and does the grading.

  89. Incidents show evaluation and training environments need to be sealed and tamper-resistantInfrastructure4+1 weak evidence

    What it implies buyers want: Network-isolated sandboxes, tamper-resistant logging and security review of test set-ups.

  90. Screen recordings with aligned mouse and keyboard events are a purchased inputUse of named software4+1 weak evidence

    What it implies buyers want: Long expert workflows in real software, with inputs, reasoning and a checkable result.

  91. The wanted task is files in, a finished deliverable out, scored by rubricOffice work4

    What it implies buyers want: Spreadsheets, slide decks and memos as gradable outputs.

  92. Some vendors organise coverage by occupation across many sectors of the economyOffice work4

    What it implies buyers want: Breadth: tasks for ordinary operational jobs, not only elite professions.

  93. Production code is bought with its commits, reviews and tickets, not as snapshotsPrivate code4+3 weak evidence

    What it implies buyers want: Unflattened private repositories with the surrounding paper trail.

  94. Failed, sloppy and autonomous robot runs are kept as training dataRobots: remote operation4

    What it implies buyers want: Robot data across quality levels, labelled as success or failure, including human take-overs when a robot fails.

  95. Some builders report needing only minutes to hours of robot data per new taskRobots: remote operation4

    What it implies buyers want: Small, exact sets of robot demonstrations for the last training step, not bulk teleoperation as a base.

  96. Vendors are measuring synthetic data's effect, and one data licence forbids it outrightMachine-made data4

    What it implies buyers want: Synthetic data used where its benefit is tested, and kept out of some human-origin sets.

  97. Video must arrive labelled: captions, camera moves, pose and action boundariesVideo and world models4+1 weak evidence

    What it implies buyers want: Per-clip labels for action, subject, camera motion and timing, tightly aligned to the footage.

  98. Licensed, rights-cleared video is bought, with brokers holding exclusive AI rightsVideo and world models4+1 weak evidence

    What it implies buyers want: Training-only licences with clear provenance, bought through whichever broker holds the catalogue.

  99. Orders have shifted from bulk libraries to narrow, curated, commissioned clipsVideo and world models4

    What it implies buyers want: Scene-level clips chosen or shot for one model objective, delivered fast and repeatedly.

  100. Real speech from named settings: clinics, call centres, courtroomsAudio and speech3

    What it implies buyers want: Domain dialogue for voice agents in customer service and medicine, with transcripts.

  101. Natural two-way conversation with each speaker on a separate channelAudio and speech3+1 weak evidence

    What it implies buyers want: Unscripted, wide-band, channel-separated dialogue with overlap kept, for speech-to-speech models.

  102. GPU and custom-accelerator kernel writing is a sold task typeChips and hardware3

    What it implies buyers want: Performance-engineering tasks scored against a measured reference on named hardware.

  103. Recorded sessions of developers working with coding agents are bought as dataCoding3

    What it implies buyers want: Real developer-and-agent trajectories, with expert preferences, for imitation and preference training.

  104. Coding data is sold with a measured gain on someone else's benchmarkCoding3+2 weak evidence

    What it implies buyers want: Evidence, on an open model and an outside test, that a coding set moves scores before the lab buys.

  105. Vendors pitch ordinary businesses on licensing their own data as a new revenue lineCompanies' own records3+2 weak evidence

    What it implies buyers want: Buyers want fresh company data badly enough that vendors now court the firms that hold it

  106. Some vendors publish what they will pay for one company's recordsCompanies' own records3

    What it implies buyers want: Published ranges rise for long, rare, proprietary operating histories; referral bounties are offered for finding sellers.

  107. Sourcing works backwards from what labs ask for; one vendor says it can start within daysWho does the buying3+1 weak evidence

    What it implies buyers want: Labs issue written data requirements; vendors find the holders and sign them.

  108. Drug makers license biological and clinical data to train their own modelsCustom models for enterprises3

    What it implies buyers want: Licensed patient, perturbation and assay data as training input for in-house models.

  109. Companies want to train on their own data without giving it to a labCustom models for enterprises3

    What it implies buyers want: Post-training that leaves data, traces and weights under the customer's control.

  110. Vendors sell private tests to help enterprises choose models and judge agent readinessPrivate tests3+2 weak evidence

    What it implies buyers want: Evidence on the buyer's own work, not public scores, before large model or agent spending.

  111. One vendor says it built finance environments with most leading closed labs; others sell finance environments to labsFinance3

    What it implies buyers want: Long-horizon financial reasoning environments whose rewards need scarce domain experts.

  112. Public finance tests are saturating; vendors sell harder tasks with partial-credit gradingFinance3+1 weak evidence

    What it implies buyers want: Process-level rewards and tasks an agent cannot score on with a fake deliverable.

  113. Games are used as long-running test and training environments for agentsGameplay3

    What it implies buyers want: Very long sequential-decision tasks with a built-in score, used to test and train agents.

  114. Labels are increasingly made by machines, by execution or simulation, not by human annotatorsGraders and quality control3

    What it implies buyers want: Buyers want labels that are cheaper and denser than people can produce, where a machine can be trusted

  115. Vendors screen contributor work to show it was really made by a humanGraders and quality control3+1 weak evidence

    What it implies buyers want: Provenance checks: screening for machine-written content and automated submissions.

  116. Video and robot builders are sizing pipelines for very large volumes of footageInfrastructure3+1 weak evidence

    What it implies buyers want: Intake, storage and capture capacity for very large volumes of video and robot data.

  117. Run logs are too large to read, so trace tooling is being builtInfrastructure3

    What it implies buyers want: Tools that turn huge volumes of agent transcripts and raw data into findings.

  118. Some data and environment vendors now run post-training on substantial compute of their ownInfrastructure3

    What it implies buyers want: Suppliers that can train on their own environments and host a customer's training run.

  119. Vendors publish long professional assignments with real deliverables that take a person many hoursLong tasks3+1 weak evidence

    What it implies buyers want: Whole-job, under-specified professional tasks as tests and possible training material.

  120. Tool-use environments expose very many callable toolsUse of named software3+1 weak evidence

    What it implies buyers want: Multi-step API tasks with checkable end states, across terminal, browser and business tools.

  121. Vendors claim small sets of expert tasks lift scores on professional testsOffice work3

    What it implies buyers want: Small, expensive, expert-built task sets that transfer across domains.

  122. Consent records, licence trails and security certificates are now a gate to dealsOther3+2 weak evidence

    What it implies buyers want: Documented, human-origin, lawfully licensed data from suppliers who pass audit.

  123. One vendor sells training by the token; several report that a single evaluation run is costlyPrices and terms3+2 weak evidence

    What it implies buyers want: Budgets for many repeats of expensive runs, with cost per task tracked beside score.

  124. Environments are priced per taskPrices and terms3+1 weak evidence

    What it implies buyers want: Complete, gradable tasks bought by the unit.

  125. Rarity and length of history set the price of a company's recordsPrices and terms3+1 weak evidence

    What it implies buyers want: Old, proprietary, hard-to-find domains rather than bulk volume.

  126. Robot and first-person video data is often priced by the hourPrices and terms3

    What it implies buyers want: Hourly rates vary with content value and volume.

  127. Small sets of expert-built tasks are presented as enoughProof that the data improves a model3+1 weak evidence

    What it implies buyers want: A small number of hard, expert-built tasks in each field, in place of mass labelling.

  128. One seller gives away plain factory video; top buyers ask for depth, stereo and a higher frame rateRobots: recordings of people3+2 weak evidence

    What it implies buyers want: Synchronised stereo, depth, motion-sensor and wrist-camera capture with calibration; plain single-camera video has little price left.

  129. Testers and vendors are hardening security to hold models and dataSafety and security3

    What it implies buyers want: Suppliers with hardened systems and tighter access controls.

  130. Machine-checked proofs are wanted; one vendor says the appeal is a grader that cannot be fooledScience and mathematics3+3 weak evidence

    What it implies buyers want: Formal proofs and statement-matching checks, usable directly as a reward signal.

  131. Synchronised multi-camera footage and raw-versus-edited pairs sell as their own productsVideo and world models3+3 weak evidence

    What it implies buyers want: Paired before/after video and time-locked multiple views of one action, to teach editing and cutting between angles.

  132. Vendors say video buyers want native high resolution, recent footage, steady cameras and high frame ratesVideo and world models3

    What it implies buyers want: True high-resolution, high-dynamic-range source footage with camera metadata, not up-scaled or graded archive material.

  133. Vendors offer world-model builders video joined to the inputs that caused itVideo and world models3+1 weak evidence

    What it implies buyers want: Footage paired frame by frame with inputs and scene state, so a model can learn what follows an action.

  134. Acted, emotional speech and fine-grained emotion labels are in demand; one marketplace pays voice actors mostAudio and speech2

    What it implies buyers want: Performed speech labelled by emotion and tone, to make synthetic voices expressive.

  135. Transcripts must be verbatim, tightly timed, and mark noises and overlapAudio and speech2+1 weak evidence

    What it implies buyers want: Human-checked, word-for-word transcripts with speaker turns, timing and non-speech events.

  136. Vendors are starting to collect and list face and voice recorded togetherAudio and speech2+1 weak evidence

    What it implies buyers want: Synchronised face, gesture and voice for models that see and talk.

  137. Two lab-data vendors say AI buyers want clean, well-documented data, ordered in campaignsBiology made in the lab2

    What it implies buyers want: Low-noise, standardised results with failures kept, returned quickly enough for model-in-the-loop work.

  138. Vendors are adding primary cells, organoids and animal models alongside cell linesBiology made in the lab2

    What it implies buyers want: Perturbation data from samples nearer to patients, with readouts over time.

  139. Labs and datasets are pitched as ways to check general models' hypothesesBiology made in the lab2

    What it implies buyers want: Physical validation of model predictions, and lab data organised as environments and evaluations for science agents.

  140. Drug makers and institutes commission single-cell perturbation data built for model trainingBiology made in the lab2

    What it implies buyers want: Interventional data — many drugs or gene edits, measured cell by cell under one protocol — rather than observational leftovers.

  141. Some enterprises train on their own production traces rather than sell themCompanies' own records2

    What it implies buyers want: Enterprises want training on logs, corrections and QA labels that never leave their control.

  142. Deals come as one-off file exports or as continuing collectionCompanies' own records2+4 weak evidence

    What it implies buyers want: Both one-off exports and recurring supply are offered; some vendors seek recurring agreements.

  143. Chinese model builders are hiring staff to buy training data from overseas suppliersWho does the buying2

    What it implies buyers want: Buyers want access to foreign vendors and data they cannot get from domestic suppliers.

  144. Chinese model builders keep standards and evaluation in house and buy corpora and labour outsideWho does the buying2

    What it implies buyers want: Buyers want suppliers that deliver data and labour to a specification the buyer writes and checks itself.

  145. The buyers' own job posts say this market has no settled priceWho does the buying2

    What it implies buyers want: Deal structures and prices for private data are being invented now, with several buyers bidding.

  146. Healthcare and finance data are bought direct from the institutions holding themWho does the buying2

    What it implies buyers want: Licensed sector datasets with compliance and data-rights terms attached.

  147. Vendors say enterprise customers often need help deciding what to train and how to score itCustom models for enterprises2

    What it implies buyers want: Scoping, success criteria and trusted graders supplied along with the training.

  148. Evaluation runs now cost real money, and human baselines are often missingPrivate tests2+1 weak evidence

    What it implies buyers want: Funded compute and expert human baselines for serious evaluation.

  149. Brokers take exclusive AI rights from owners, then license non-exclusively to labsExclusivity and first use2

    What it implies buyers want: Rights-cleared media with one seller per catalogue, so labs must come through the broker.

  150. One robot-data vendor reports buyers moving to a single supplier; another vendor offers isolated collection per customerExclusivity and first use2

    What it implies buyers want: Consistent data from a single source, collected in a set-up no other customer shares.

  151. Biology data funders and contributors get a timed head start, not permanent exclusivityExclusivity and first use2

    What it implies buyers want: Early access for those who pay or contribute, then wider release.

  152. Voice actors and rare-language transcribers are listed at higher rates than ordinary speakersWhat experts are paid2+1 weak evidence

    What it implies buyers want: Performed emotion and scarce languages; plain conversation is cheap.

  153. Real financial paperwork is bought as raw material, from firms and individualsFinance2+2 weak evidence

    What it implies buyers want: Authentic, de-identified financial records to seed tasks and environments.

  154. Healthcare back-office work is wanted: billing codes, scheduling and operationsHealthcare and medicine2

    What it implies buyers want: Medical coding and operations data; models score much higher on clinical notes than on billing codes.

  155. A lab buys recordings of developers coding with AI assistants on their own projects, and another vendor pays developers for the samePeople working with AI2

    What it implies buyers want: Real developer-agent sessions with step-by-step expert preferences.

  156. Long agent runs need snapshot, resume and a durable record of every runInfrastructure2+1 weak evidence

    What it implies buyers want: Stateful environments that can be saved, branched, replayed and audited after the fact.

  157. Testers measure whether agents stay coherent and honest over very long runsLong tasks2

    What it implies buyers want: Long-running, adversarial tests of sustained judgement, plus data on recovery from mistakes.

  158. Long tasks need feedback along the way, not one score at the endLong tasks2+1 weak evidence

    What it implies buyers want: Dense, localised reward signals and new scoring methods for long, real-world work.

  159. Long runs need pause, resume, checkpoint and replay built into the environmentLong tasks2+1 weak evidence

    What it implies buyers want: Environments that survive day-long runs and can be stopped, inspected, branched and replayed.

  160. Harder office tasks are long, vague and buried in noise, and models still failOffice work2

    What it implies buyers want: Tasks that need the agent to work out the job, in messy context, over hours or days.

  161. Step-by-step records of people doing office work are a listed productOffice work2+3 weak evidence

    What it implies buyers want: Authentic traces of routine work, including mundane hand-offs between co-workers.

  162. Some health and biology data vendors sell governed access rather than downloadable filesPrices and terms2

    What it implies buyers want: Buyers may train where the data sits and cannot always take raw files away.

  163. Some content licensors are testing pay-per-use terms; one says recent content is worth morePrices and terms2

    What it implies buyers want: Usage reported by the buyer under standard terms, with recent content possibly priced higher.

  164. Private code is wanted because public code no longer gives a clean measurementPrivate code2+2 weak evidence

    What it implies buyers want: Unseen codebases as the basis for honest coding scores and hard training tasks.

  165. Buyers want a forecast of what an environment will teach before spending computeProof that the data improves a model2+1 weak evidence

    What it implies buyers want: A reliable per-environment estimate of training value, so compute is not wasted on tasks that teach nothing.

  166. Some data vendors now sell post-training and a working robot, not only datasetsRobots: deployment2

    What it implies buyers want: A robot post-trained for one job on the customer site, with data collected as part of the service.

  167. Some sellers offer teleoperation on the buyer's own robot; one calls it its most valuable tierRobots: remote operation2+1 weak evidence

    What it implies buyers want: Demonstrations on the exact robot to be deployed, with force and motion, run by someone else.

  168. One world-model builder is collecting real robot logs; one free set has been used as a test setRobots: remote operation2

    What it implies buyers want: Real robot logs to ground simulators, and neutral test sets; one robot maker gives its data away.

  169. Science and engineering tasks are built around simulations and recomputable answersScience and mathematics2+2 weak evidence

    What it implies buyers want: Tasks in physics, chemistry, biology and engineering whose results a script can check.

  170. Demonstrations can come from a stronger model instead of from peopleMachine-made data2

    What it implies buyers want: In some cases vendors supply tasks and graders while a strong model supplies the trajectories.

  171. Stand-in data lets some firms train without exposing customer recordsMachine-made data2+1 weak evidence

    What it implies buyers want: Stand-in or synthetically rewritten versions of confidential workloads.

  172. One expert-labour vendor publishes a biology lab-image benchmarkBiology made in the lab1

    What it implies buyers want: Hard, expert-written biology questions and working lab images, not textbook figures.

  173. Gene-sequence data from outside public databases, collected under benefit-sharing agreementsBiology made in the lab1

    What it implies buyers want: Diverse, long-read sequence data with documented consent, as fuel for biological foundation models.

  174. One vendor licenses patient images and sequencing data for buyers to train their own modelsBiology made in the lab1

    What it implies buyers want: Multimodal clinical data on the same patients over time, licensed for model training.

  175. Drug companies pool standardised antibody measurements while keeping sequences privateBiology made in the lab1

    What it implies buyers want: Purpose-designed, standardised protein measurements, with training rights on others' data but not the raw files.

  176. Labs pay a small specialist for chip-design environmentsChips and hardware1

    What it implies buyers want: Chip-design tasks graded by real tool feedback, up to a manufacturable design.

  177. One vendor hires hardware engineers and semiconductor experts for judgement on real buildsChips and hardware1

    What it implies buyers want: Expert data from people who designed, built and debugged physical products.

  178. One chip-design vendor says labs give it unreleased models and work with it frequentlyChips and hardware1

    What it implies buyers want: Private, per-model capability reports on hardware tasks, and joint work rather than a catalogue sale.

  179. One robot-data unit weighs sourcing through a marketplace as well as in-houseWho does the buying1

    What it implies buyers want: More robot data than in-house collection can supply, pulled from partners.

  180. One tester sells enterprises private benchmarks to guide which model they pay forCustom models for enterprises1

    What it implies buyers want: Proof on the buyer's own work of which model or agent is reliable enough.

  181. Exclusivity is available at several times the priceExclusivity and first use1+1 weak evidence

    What it implies buyers want: A minority of buyers will pay a large premium to keep a task set or dataset to themselves.

  182. One vendor makes sellers warrant the data has not been licensed this way beforeExclusivity and first use1

    What it implies buyers want: First use: the seller warrants no earlier grant of the same rights.

  183. One vendor lists safety red-teamers and policy graders as their own paid lineWhat experts are paid1

    What it implies buyers want: Specialist judgement on whether model outputs are dangerous, in many languages.

  184. One forecasting firm says hard questions are the scarce input and pays for themForecasting1

    What it implies buyers want: Questions in the band between solved and unknowable, answered with reasoning.

  185. Three large labs fund a live forecasting benchmark; one invested in a forecasting firmForecasting1

    What it implies buyers want: Questions whose answers are not yet known, so they cannot be memorised.

  186. A staffing vendor lists work writing analyst forecasts from a fixed evidence cutoffForecasting1

    What it implies buyers want: Point-in-time reports on markets, elections and drug launches, in the source language.

  187. One vendor records long, scripted play sessions to cover rare eventsGameplay1

    What it implies buyers want: Directed capture of unusual behaviour and long temporal context, which uploaded highlight clips lack.

  188. One vendor sells engine-level ground truth with the footage: depth, camera pose, physicsGameplay1

    What it implies buyers want: Measured, not estimated, scene state alongside each frame, for labs without their own game studios.

  189. A vendor claiming the largest action-labelled game corpus pitches transfer to robotsGameplay1

    What it implies buyers want: Action-labelled play as a base for agents that act in the physical world; whether outside labs will pay for it is not shown.

  190. One clinical-data company says drug makers license its patient data to train models, and one bought the modelHealthcare and medicine1

    What it implies buyers want: Outcome-linked data that predicts which patients respond to a drug.

  191. An arena tests unreleased models on real users and shares feedback with model developersPeople working with AI1

    What it implies buyers want: De-identified real prompts, votes and full agent sessions, plus independent behaviour counts.

  192. Vendors test assistants on real phones and build replicas of consumer apps and sitesUse of named software1+2 weak evidence

    What it implies buyers want: In-market, on-device evaluation and phone-app environments.

  193. One vendor says labs build the interface themselves and buy the domain workUse of named software1

    What it implies buyers want: Deterministic, cheat-resistant checks and tasks tuned just beyond current models.

  194. Vendors say demand has moved from large crowds to small groups of expertsOther1+1 weak evidence

    What it implies buyers want: Buyers want fewer, highly qualified contributors and pay for expertise rather than volume

  195. AI companies are said to pay for fresh, structured web content over crawlingOther1

    What it implies buyers want: Clean, de-duplicated, real-time feeds with freshness and quality signals.

  196. One evaluator tests agents running real businesses and reports cost per runOther1

    What it implies buyers want: Side-by-side real-world comparisons and human cost baselines.

  197. Contributor terms at several vendors assign all rights in the work to the vendorPrices and terms1+5 weak evidence

    What it implies buyers want: Clean ownership of the work with no residual claims from contributors.

  198. Long-established crystal-structure databases hold AI-training rights back from ordinary access, and one buyer says it holds them exclusivelyPrices and terms1+2 weak evidence

    What it implies buyers want: Buyers want explicit, separately negotiated AI training rights to experimentally verified reference records, and at least one wants them exclusively.

  199. Repositories are converted into graded tasks, sometimes for one customer's own modelPrivate code1+4 weak evidence

    What it implies buyers want: Issue-to-fix task sets with tests, cut from private repositories.

  200. Code reaches buyers through dead start-ups, non-exclusive licences and cleared rightsPrivate code1+1 weak evidence

    What it implies buyers want: Rights-cleared private code under non-exclusive licence, sourced from companies that no longer need it.

  201. Simulated environments must show they transfer to the real productProof that the data improves a model1+1 weak evidence

    What it implies buyers want: Simulators with measured transfer, in many varieties, paired with practice on real sites.

  202. A few vendors run tests of robot models on real hardwareRobots: deployment1+1 weak evidence

    What it implies buyers want: Scoring of robot models on physical robots.

  203. Where a price is shown, real-robot teleoperation data is priced above robot data overallRobots: remote operation1

    What it implies buyers want: Reported prices are highest for real-robot demonstrations; published rates are rare.

  204. One outside tester is let inside labs for audits and incident investigationsSafety and security1

    What it implies buyers want: Trusted third parties with deep access who can audit internal agent use and investigate incidents.

  205. One cyber tester says labs buy its offensive-cyber tests before releasesSafety and security1+1 weak evidence

    What it implies buyers want: Fresh, private, realistic attack scenarios scored on the raw model within the release window.

  206. Science tasks test whether agents can design and run experiments with limited resourcesScience and mathematics1+1 weak evidence

    What it implies buyers want: Buyers want models that do research, not only answer questions

  207. Labs commission and own research-level maths problems, audits and verifiersScience and mathematics1+1 weak evidence

    What it implies buyers want: Problems past expert level with checkable answers, plus third-party scoring.

  208. People are paid to rank and fault-find generated videoVideo and world models1+2 weak evidence

    What it implies buyers want: Human preference and defect labels on generated video.

  209. Some vendors scale contributor pay with the quality of the workWhat experts are paid0+1 weak evidence

    What it implies buyers want: Buyers want to pay for good data, not for hours logged

  210. Vendors sell judgement for work that has no single right answerGraders and quality control0+1 weak evidence

    What it implies buyers want: Buyers want data that teaches taste where no automatic check exists

  211. Vendors prove each task can be solved before they ship itGraders and quality control0+2 weak evidence

    What it implies buyers want: Buyers reject tasks that may be impossible; a checked solution is part of what they pay for

  212. Contributor work sessions are themselves instrumented and recordedPeople working with AI0+3 weak evidence

    What it implies buyers want: Screen-recorded, in-market use of an assistant, with session telemetry kept as data.

  213. Tools and tasks are frozen at a past date so the answer cannot leak inInfrastructure0+2 weak evidence

    What it implies buyers want: Buyers want training and test worlds that show a model only what was known on a given day

  214. Working designers and creatives are paid to judge what models makeOther0+1 weak evidence

    What it implies buyers want: Buyers of image, design and creative models want professional taste, not general raters

  215. Environments now pit several agents against each other, as rivals or opponentsOther0+2 weak evidence

    What it implies buyers want: Buyers want training and tests for agents that face adaptive opponents, not just fixed tasks

  216. Some vendors run a real business so its daily work becomes data to sellOther0+2 weak evidence

    What it implies buyers want: Buyers value data captured from real paid work over staged collection

  217. Vendors turn industry rules and operators’ unwritten know-how into training environmentsOther0+2 weak evidence

    What it implies buyers want: Buyers want agents trained on how regulated and industrial work is really done, which is not written down online

  218. Weak-evidence sellers say price follows rarity, rigour and age, not sizePrivate code0+2 weak evidence

    What it implies buyers want: Scarce-language, well-tested, human-era code from complex products.

  219. Vendors sell data built to target named public benchmarksProof that the data improves a model0+2 weak evidence

    What it implies buyers want: Buyers want data aimed at the public tests their models are judged on, even without a measured gain

  220. Large labelling firms now list first-person video for robots alongside their older linesRobots: recordings of people0+1 weak evidence

    What it implies buyers want: Demand for first-person video is large enough that general data firms build product lines for it

    • LXTread 2026-10-14
  221. Cyber tests now span whole networks and run for hours, not single puzzlesSafety and security0+2 weak evidence

    What it implies buyers want: Buyers want cyber environments that hold an agent to a long attack or defence across many machines

  222. 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 saleScience and mathematics0+2 weak evidence

    What it implies buyers want: Buyers who want experimental materials data, failures included, must build their own laboratory or license an old reference database, because the labs making new data are not selling it.

What the pages do not show

Almost no page gives a price that a model builder paid. The prices that are published are what vendors pay their experts and what they offer companies for records. Those that could be checked against their page are on prices and pay.

The vendors disagree with each other. Some say data made by people matters less with each new model; others say data made by machines helps less at the frontier.

A vendor absent from a row may still do the thing the row describes. It is listed only where one of its own pages, read here, shows it. Some vendors are missing altogether because none of their pages answered.