Skip to content

Research report ยท Live data and sources

What AI Labs Are Buying

To 11 Oct 2026, the evidence shows labs buying more coding (production-like environments), computer-use app and website clones and egocentric and robotics demonstration data and less generalist annotation and microtasks and static maths and verifiable coding tasks.

Covers
20 Jun 2023 to 11 Oct 2026
Events by the date they happened
Computed
18 Oct 2026, 15:24 UTC
Live figures daily; sources as reviewed
Source
The public job boards of the AI training platforms this site tracks, read every day. Evidence: research, filings, company statements and reporting, each linked.
Method
Version 1. Definitions
CC BY 4.0, how to cite

Every figure on this page as CSV and JSON, sent to your work email.

kinds of data in rising demand
12
cooling
2
evidence items, each linked
51

Key findings

  1. 1

    Rising: coding (production-like environments), computer-use app and website clones, egocentric and robotics demonstration data, enterprise workflow RL environments and data, expert judgment (creative and non-professional domains), expert judgment (professional domains), in-house human data and environment teams at labs, licensed media and content, off-the-shelf and synthetic data, RL environments (all domains), voice and conversational audio and wet-lab and experimental science data.

    Figure 2
  2. 2

    Cooling: generalist annotation and microtasks and static maths and verifiable coding tasks.

    Figure 2
  3. 3

    Enterprise workflow RL environments and data has the most evidence (8 items).

    Figure 3

Figure 1

Which kinds of work platforms are listing more of (live)

In the last four complete weeks, which kinds of AI training work drew more new listings than in the four weeks before?

Kind of workNew listings, 4 weeks beforeNew listings, last 4 weeksChange
Multilingual & Localization9394
+1.1%
Law, Medicine & Finance234234
0.0%
AI Evaluation & Rating (RLHF)114106
-7.0%
Transcription, Audio & Speech7768
-11.7%
Math & STEM3328
-15.2%
Coding & Software Evaluation6948
-30.4%
Data Annotation & Labeling3927
-30.8%
Writing & Content139
-30.8%
Research Studies & Surveys148
-42.9%
Robotics & Physical-AI DataSmall sample95
-44.4%
Red-Teaming & AI Safety4417
-61.4%
Agentic Tasks & RL Environments4512
-73.3%

Sample. New listings on the platforms tracked since 17 Aug 2026 (a fixed panel), in the 8 complete ISO weeks to 11 Oct 2026. Recomputed weekly.

What labs are buying shows up first as new listings for the people who make the data.

Source. AITraining.jobs Research, What AI Labs Are Buying, live data.

Figure 2

What labs are buying more of, and less of

Which kinds of training data are in rising demand, and which are cooling?

Kind of dataDirectionLatest evidence
Coding (production-like environments)Epoch AIRisingCoding remains a major source of demand and environments are moving beyond SWE-bench-style tasks toward productionised setups.
Computer-use app and website clonesThe Financial News (Korea), summarising The InformationRisingPopular items in corporate-data deals include millions of hours of software demonstration videos.
Egocentric and robotics demonstration dataBetaKitRisingMecka (body-sensor and iPhone human-activity data for humanoids) raised a $60M Series B; it passed a $100M run rate in June and expects $300M by year end.
Enterprise workflow RL environments and dataMercorRisingMercor runs a program paying companies for workflow data and says frontier models are increasingly trained to work inside companies.
Expert judgment (creative and non-professional domains)SemiAnalysisRisingSemiAnalysis: coding is the largest demand, but spending on photography, music and design has picked up as labs seek breadth.
Expert judgment (professional domains)TechCrunchRisingSnorkel's run rate rose eighteenfold in 12 months to $375M, which TechCrunch attributes to labs' appetite for high-end training data.
In-house human data and environment teams at labsEpoch AIRisingA founder told Epoch there has been "substantially more in-housing" as labs build their own data teams, to avoid vendor margins and keep priorities confidential.
Licensed media and contentTechCrunchRisingWirestock moved from selling its library off the shelf to custom content requests; it supplies six of the largest foundation model makers at a $40M run rate.
Off-the-shelf and synthetic dataTechCrunchRisingSnorkel shifted to selling completed datasets ("data-as-a-service"), generating data synthetically alongside subject-matter experts.
RL environments (all domains)TechCrunchRisingSurge CEO Edwin Chen said demand for RL environments inside labs had seen a "significant increase"; Anthropic discussed spending over $1B on them.
Voice and conversational audioContrary ResearchRisingContrary: as voice agents move to full-duplex conversation, the bottleneck has shifted to training data; all major dialogue datasets combined total about 3K hours.
Wet-lab and experimental science dataContrary ResearchRisingAnthropic set up a wet lab in the San Francisco Bay Area to run physical biology experiments (confirmed by its head of life sciences): labs generating experimental data in-house.
Multilingual and low-resource languagesSpeechData.aiSteadyA speech-data vendor prices off-the-shelf low-resource-language audio at $60 to $150 per hour versus $50 to $100 for major languages.
Generalist annotation and microtasksAmazon Mechanical TurkCoolingAmazon Mechanical Turk, the original crowd microtask marketplace, permanently closed on 30 September 2026.
Static maths and verifiable coding tasksGround Level AICoolingWill DePue (ex-OpenAI) argued that abundant, auto-gradable maths and coding datasets that RL relied on are being exhausted.

Sample. 15 kinds of training data, each read from its most recent evidence (linked), to 11 Oct 2026.

Direction is what the evidence says about demand for each kind of data (more buyers, bigger contracts, rising or falling prices), as of its date.

Source. AITraining.jobs Research, What AI Labs Are Buying, live data.

Figure 3

All the evidence

What does each source say about demand for each kind of data?

Kind of dataDirectionEvidence
Enterprise workflow RL environments and dataMercorRisingMercor runs a program paying companies for workflow data and says frontier models are increasingly trained to work inside companies.
Enterprise workflow RL environments and datamicro1Risingmicro1 committed to spend $1 billion over 12 months acquiring and licensing enterprise operational data to build RL environments, financed by Citi and Hercules Capital.
Egocentric and robotics demonstration dataBetaKitRisingMecka (body-sensor and iPhone human-activity data for humanoids) raised a $60M Series B; it passed a $100M run rate in June and expects $300M by year end.
Generalist annotation and microtasksAmazon Mechanical TurkCoolingAmazon Mechanical Turk, the original crowd microtask marketplace, permanently closed on 30 September 2026.
Voice and conversational audioContrary ResearchRisingContrary: as voice agents move to full-duplex conversation, the bottleneck has shifted to training data; all major dialogue datasets combined total about 3K hours.
Expert judgment (professional domains)TechCrunchRisingSnorkel's run rate rose eighteenfold in 12 months to $375M, which TechCrunch attributes to labs' appetite for high-end training data.
Off-the-shelf and synthetic dataTechCrunchRisingSnorkel shifted to selling completed datasets ("data-as-a-service"), generating data synthetically alongside subject-matter experts.
Enterprise workflow RL environments and dataFortuneRisingmicro1 told the Spirit Airlines bankruptcy court it had completed more than 50 data transactions in the preceding 45 days.
Enterprise workflow RL environments and dataFortuneRisingSimpleClosure said its number of AI buyers for company code and workplace data grew ninefold versus 2025.
Wet-lab and experimental science dataContrary ResearchRisingAnthropic set up a wet lab in the San Francisco Bay Area to run physical biology experiments (confirmed by its head of life sciences): labs generating experimental data in-house.
Expert judgment (professional domains)TechCrunchRisingTechCrunch describes near-bottomless demand for unique training data; some researchers think data spend could rival compute.
Off-the-shelf and synthetic dataTechCrunchRisingmicro1 increasingly generates synthetic data without humans and resells off-the-shelf datasets at 80% to 90% gross margins.
Enterprise workflow RL environments and dataThe Financial News (Korea), summarising The InformationRisingThe Information (about 20 interviews) reported surging demand for corporate data (chats, emails, meeting recordings, code histories); Protege said its transaction volume rose from $30M to at least $100M.
Enterprise workflow RL environments and dataThe Financial News (Korea), summarising The InformationRisingProtege's CEO said total transaction volume rose from $30 million the previous year to at least $100 million in 2026.
Computer-use app and website clonesThe Financial News (Korea), summarising The InformationRisingPopular items in corporate-data deals include millions of hours of software demonstration videos.
Static maths and verifiable coding tasksGround Level AICoolingWill DePue (ex-OpenAI) argued that abundant, auto-gradable maths and coding datasets that RL relied on are being exhausted.
Expert judgment (professional domains)Dealroom News (summarising The Information)RisingMercor reached a $2B gross run rate in June 2026 paying domain experts (physics, finance) to create training data; growth also from AI app developers and Fortune 500 firms.
Generalist annotation and microtasksRuntime WireCoolingRuntime Wire frames Mercor's growth against MTurk's decline as labs shifting paid human labor from commodity microtasks to expert work.
Egocentric and robotics demonstration dataStartupFeedCoolingPrice signal, not volume: Indian egocentric data rates fell 30% from January to June 2026 as more suppliers entered, per an industry executive.
Multilingual and low-resource languagesSpeechData.aiSteadyA speech-data vendor prices off-the-shelf low-resource-language audio at $60 to $150 per hour versus $50 to $100 for major languages.
Licensed media and contentTechCrunchRisingWirestock moved from selling its library off the shelf to custom content requests; it supplies six of the largest foundation model makers at a $40M run rate.
Licensed media and contentCASRAIRisingWiley reported $49M of AI licensing revenue in FY2026 (year to 30 Apr 2026), up 23%, and over $110M lifetime.
Egocentric and robotics demonstration dataCNN Business (syndicated on KVIA)RisingAn industry analyst said that by late 2025 the robotics industry began using human video data as a cheaper middle ground; one supplier finds only about half of submitted footage usable.
Egocentric and robotics demonstration dataMIT Technology ReviewRisingmicro1's CEO said robotics data demand is increasing fast and robotics firms spend over $100M a year on real-world data from micro1 and peers.
Egocentric and robotics demonstration dataMIT Technology ReviewRisingmicro1 says it has tens of thousands of hours of chore footage; Scale AI announced more than 100,000 hours.
Egocentric and robotics demonstration dataThe Next WebRisingDoorDash launched a Tasks app paying couriers to film chores and record other-language conversations for AI and robotics training.
Multilingual and low-resource languagesThe Next WebSteadyDoorDash's Tasks app includes recording unscripted conversations in another language.
Licensed media and contentThe Motley Fool (earnings call transcript)CoolingGetty: two large multiyear deals drove Q4 2025; management expects data licensing at low single-digit percentages of revenue and forecast no additional large-scale licensing revenue for 2026.
Enterprise workflow RL environments and dataEpoch AIRisingEpoch AI interviews (18 people at RL-environment startups, neolabs and frontier labs) found enterprise workflows to be a major growth area after maths and coding.
Enterprise workflow RL environments and dataEpoch AIRisingAn RL-environment founder predicted enterprise workflows would "explode" in 2026 because labs favour tasks that are valuable and quantifiable.
Computer-use app and website clonesEpoch AIRisingEnterprise-workflow environments often rely on clones of apps like Slack or SAP; a lab researcher warned many clones are buggy.
Wet-lab and experimental science dataEpoch AIRisingEpoch notes partnerships between labs and product companies, including Benchling and Anthropic for biology workflow environments.
Static maths and verifiable coding tasksEpoch AICoolingInterviewees said maths tasks are easy to make but may be declining and transfer poorly.
Coding (production-like environments)Epoch AIRisingCoding remains a major source of demand and environments are moving beyond SWE-bench-style tasks toward productionised setups.
In-house human data and environment teams at labsEpoch AIRisingA founder told Epoch there has been "substantially more in-housing" as labs build their own data teams, to avoid vendor margins and keep priorities confidential.
Computer-use app and website clonesSemiAnalysisRisingSemiAnalysis reports much computer-use environment work is outsourced; UI gyms cost about $20,000 per site and OpenAI has bought hundreds.
Expert judgment (creative and non-professional domains)SemiAnalysisRisingSemiAnalysis: coding is the largest demand, but spending on photography, music and design has picked up as labs seek breadth.
Generalist annotation and microtasksBusiness Insider (syndicated on AOL)CoolingMercor ended a Meta project and offered contractors a similar one at $16 an hour, $5 less; Scale cut a generalist contractor team.
Multilingual and low-resource languagesVentureBeatSteadyMeta built a 3,350-hour corpus across 348 low-resource languages from compensated local speakers for its Omnilingual ASR release.
Expert judgment (professional domains)Fortune (Bloomberg)RisingOpenAI paid 100+ former investment bankers about $150 an hour (via third-party suppliers) to build financial models for training.
Generalist annotation and microtasksBusiness Insider (syndicated on AOL)CoolingBusiness Insider: Scale AI slashed a team of contractors focused on "generalist" work in October 2025.
Voice and conversational audioContrary ResearchRisingDavid AI, a conversational-audio data vendor, said at its October 2025 Series B that it works with several Mag7 companies and most leading AI labs.
Multilingual and low-resource languagesBusiness InsiderSteadyBusiness Insider: xAI and others pay up to about $100 an hour for trainers, including for teaching languages such as Japanese.
Expert judgment (professional domains)TechCrunchRisingMercor pitched investors on building RL environments for domain tasks such as coding, healthcare and law.
RL environments (all domains)TechCrunchRisingSurge CEO Edwin Chen said demand for RL environments inside labs had seen a "significant increase"; Anthropic discussed spending over $1B on them.
Expert judgment (professional domains)Business Insider (syndicated on AOL)RisingxAI eliminated more than 500 data labellers in September 2025 to prioritise higher-paid specialist roles.
Generalist annotation and microtasksMechanizeCoolingMechanize (an RL-environment seller) predicted labs would move away from cheap procedurally generated tasks and low-paid contractors toward full-time domain experts.
Generalist annotation and microtasksTechCrunch (citing Bloomberg)CoolingScale AI cut 14% of staff and 500 contractors; its interim CEO said it had scaled its core data-labeling business too quickly.
Licensed media and contentThe Decoder (summarising Bloomberg)RisingOpenAI, Google and others bought unpublished creator video at $1 to $4 per minute; an agency executive warned the window will not stay open forever.
Generalist annotation and microtasksAppen Limited (ASX: APX), ASX announcementCoolingGoogle, Appen's largest customer (30.3% of FY23 revenue), terminated its contract, ending all projects by 19 March 2024.
Generalist annotation and microtasksThe Verge (with New York Magazine)CoolingKenyan Remotasks annotators said pay fell from about $5 to $10 an hour to $1 to $3 an hour by early 2023.

Sample. 51 evidence items, newest first.

Source. AITraining.jobs Research, What AI Labs Are Buying, live data.

For labs, data vendors and researchers

Buy the data behind this report

The figures above are a summary. Behind them are records we collect ourselves, every day:

  • 9,572 roles tracked, each with its history

    Every role on 29 job boards, read daily since 13 Jul 2026: its pay and place, and the first and last day it was open.

  • 6,200 apply clicks in 28 days

    Which roles people actually apply to, by field, pay and kind of work. It happens on this site, so nobody else can count it.

  • 39 lab and vendor hiring boards

    The job boards of 25 AI labs and 14 human-data vendors, read every day: who is hiring for human data, and who is hiring to buy it.

What you can buy

  • The records

    The rows behind any figure, with their history, not only the totals.

  • A regular feed

    The same figures every week or month, as files.

  • A custom cut or benchmark

    Your countries, languages or fields, or your own platform against the market.

Method and definitions

The public job boards of the AI training platforms this site tracks, read every day. Evidence: research, filings, company statements and reporting, each linked. The live figures are computed every day from this site's own data, the same way for every reader; every sourced row links the publication it is taken from, with the figure as that publication states it. Each figure states its sample beneath it.

New listings (live)
Roles first seen in a week on platforms tracked since the start of the comparison, so a newly added platform does not read as growth.
Direction
What each source says about demand for a kind of data (more buyers, bigger contracts, rising or falling prices), as of its date.

What this data cannot tell you

  • Each report sets its sourced records beside a live figure computed from this site's own data every day; the sourced records are added as they are published and reviewed.
  • Compiled from public sources: an event that was never announced or reported is missing, and private companies' figures are what they or the press stated, not audited.
  • New listings for the people who make data are a leading sign of what labs buy, not a measure of it.

How the job feeds are read and roles are classified: site methodology. How fresh, complete and consistent the data is, checked every day: data quality and coverage.

Editions

The live report is the dataset as last reviewed, beside figures recomputed every day from this site's own data; new events are added as they are published. Each complete month is also published as an edition, frozen two days after the month ends, so a figure you cite from an edition never moves.

  • Live dataSources as last reviewed, live figures daily

Cite and reuse

Free to quote, chart and republish under CC BY 4.0 with a link to this page. Live figures change daily: cite a monthly edition for numbers that will read the same later, or keep the access date.

Citation

AITraining.jobs (2026). What AI Labs Are Buying (live data, 20 Jun 2023 to 11 Oct 2026). https://www.aitraining.team/research/what-ai-labs-are-buying (accessed 18 Oct 2026)

BibTeX

@misc{aitrainingjobs_what_ai_labs_are_buying_live,
  author = {{AITraining.jobs}},
  title = {What AI Labs Are Buying (live data, 20 Jun 2023 to 11 Oct 2026)},
  year = {2026},
  howpublished = {\url{https://www.aitraining.team/research/what-ai-labs-are-buying}},
  note = {Accessed 18 Oct 2026},
}
Data: every figure as CSV (one line per value) and JSON, sent to your work email.Pay and open-roles data

More research

All 78 research reports