Skip to content

Directory of AI training data companies

The directory lists 50 companies that sell or produce training data, grouped into six categories, with the data types each handles and what it publishes about price. Most price by quote; where a company publishes rates, they are given with a link to the page they came from.

Last verified 7 November 2026. Every row was read from the company's own site that day.

Which companies license ready-made training data?

Existing material you license rather than commission: publisher and web content, off-the-shelf datasets, and real-world operational data gathered from companies that hold it.

Licensed data: 7 companies, alphabetical. Read 7 November 2026.
Company	What it offers	Data types	Published pricing	Also noted
AppenAlso: Annotation, Expert and RLHF	Listed on the ASX. Labeling, post-training sets for language models, and a task catalog in which coding problems ship with automatic checkers.	Text, audio, image, video, code	Not published: quote only	Its RLVR post (Feb 2026, revised Oct 2026) describes 500 coding problems, each with its own automatic checker. (Appen blog)
CentificAlso: Expert and RLHF	Sells collection, preference and SFT data, RL environments and safety testing through a platform it calls the AI Data Foundry; also runs a data marketplace.	Text, audio, image, video	Not published: quote only	Nothing further
Copyright Clearance Center	The AI Systems Training License: one voluntary, non-exclusive collective permit, covering works licensed through CCC, for organizations whose AI systems are used outside their own walls.	Text	Fee schedule not public; CCC pays royalties out to rightsholders	Nothing further
Labelbox Prism	Labelbox's channel for licensing real operating data from companies: reasoning traces, cross-system activity, support histories, robot telemetry, code edits, voice and likeness given with consent.	Text, code, audio, video, tabular	Not published for buyers	Licenses default to exclusive. (Labelbox Prism page)
RSL Collective	Licenses publishers' and creators' web content collectively under the RSL (Really Simple Licensing) Standard. Publishers join free, on non-exclusive terms.	Text	Not published	Nothing further
TELUS DigitalAlso: Annotation, Expert and RLHF	The AI data arm of TELUS Digital: ready-made datasets, generative AI post-training data, red teaming, and data for cars and robots.	Text, image, video, audio, 3D	Not published: quote only	Claims 1M+ contributors across 500+ languages. (TELUS Digital page)
TollBit	Publishers name a price for each AI bot or agent visit; developers pay it, plus a small TollBit percentage added to every transaction.	Text (web content)	Per-visit rates set by each publisher; TollBit's cut not stated	Nothing further

Disclosure: any company in this list could end up paying us for an introduction that turns into work. Entries, their sequence and our descriptions are written the same way regardless.

Which companies label training data?

Labeling platforms you run yourself and managed services that supply the annotators. Most handle images and video; several cover text, audio and 3D sensor data.

Annotation: 12 companies, alphabetical. Read 7 November 2026.
Company	What it offers	Data types	Published pricing	Also noted
CloudFactory	Supplies a managed labeling workforce on its own platform for human-in-the-loop review; includes a Vision AI tool.	Image, video, text	Not published: quote only	Nothing further
Dataloop	Software for organizing unstructured files and wiring labeling and processing steps into pipelines.	Image, video, audio, text, 3D	Not published: quote only	Claims ISO 27001 and 27701, plus a Type II SOC 2 report. (Dataloop homepage)
EncordAlso: Evaluation	One workspace to curate, manage and label image, video, document, text and audio files.	Image, video, text, audio, 3D	Not published: quote only	Carries a SOC 2 badge without saying which type. (Encord multimodal page)
HumanSignal (Label Studio)	Label Studio's maker. Community edition costs nothing; Starter Cloud and Enterprise are paid; data services available.	Text, image, video, audio, tabular	$99/mo Starter Cloud covers 12 users; each extra user $49/mo. Enterprise: quote. Community: free	Nothing further
Keymakr	In-house annotators working on Keymakr's own labeling software, for pictures and footage.	Image, video	Not published: quote only	Nothing further
Kili TechnologyAlso: Evaluation	Labeling, evaluation and post-training work on one platform; specialist annotators can be added.	Text, image, video, tabular	$0 trial (2–5 seats, 100 assets max). Grow, Enterprise: quote	Security badges claimed: ISO 27001; SOC 2 (Type II). (Kili pricing page)
Label Your Data	A managed labeling team for vision and language projects. Offers a free pilot and lists per-object and hourly rates openly.	Image, video, text, 3D, audio	From $0.015/object (keypoints), $0.02/object (boxes), $0.02/entity (NLP); hourly work usually $6/annotator-hour	States no minimum; the pilot costs nothing. (Label Your Data pricing page)
LabelboxAlso: Expert and RLHF	Labeling software metered in Labelbox Units (LBUs), with labeling labor as an optional extra. Owns the Alignerr expert network and the Prism data channel.	Text, image, video, audio, code	Metered in LBUs: 1 LBU per image row, for example. No dollar rate per LBU shown. 500 free LBUs monthly on free accounts	Nothing further
SamaAlso: Evaluation	Vision and generative AI annotation, plus validation, run on Sama's in-house platform.	Image, video, 3D, text	Not published: quote only	Shows ISO 9001, ISO 27001, TISAX. (Sama security page)
Scale AIAlso: Expert and RLHF, Evaluation	Collection, curation and labeling across text, images, video and 3D sensor fusion; also model evaluation and agent environments.	Text, image, video, 3D, code, audio	Not published: quote only	Agent environments page: 2,000+ tool-use tasks plus computer-use sets; unpriced. (Scale AI RL environments page)
Segments.ai	Labeling software built for robots and self-driving stacks: camera frames, lidar point clouds, fused sensors.	Image, 3D, video	Core from 3,600 hours/yr (works out at $2.67/hr). Fusion from 5,000 hours/yr. Enterprise 150,000+ hours/yr	Shows ISO 27001 and TISAX. (Segments.ai pricing page)
SuperAnnotateAlso: Evaluation	A platform where teams assemble annotation, evaluation and fine-tuning data; sold as Starter, Pro and Enterprise.	Text, image, video, audio	Tiers named, prices withheld	Nothing further

Disclosure: any company in this list could end up paying us for an introduction that turns into work. Entries, their sequence and our descriptions are written the same way regardless.

Which companies supply expert and RLHF data?

Networks of domain experts who write demonstrations, rank model outputs and build tasks with graders, plus builders of reinforcement-learning environments for agents.

Expert and RLHF: 11 companies, alphabetical. Read 7 November 2026.
Company	What it offers	Data types	Published pricing	Also noted
Bespoke Labs	Designs RL environments and the plumbing to run them (an environment engine, sandboxes, agent optimization) for frontier labs, newer labs and enterprises.	Text, code	Not published: quote only	Raised $40M over seed and Series A, announced 6 Jul 2026. (Bespoke Labs blog)
HalluminateAlso: Evaluation	YC company building hosted copies of tools like Salesforce, Slack and ticket queues where computer-use and browser agents can train and be tested.	Text, image	Not published: quote only	Nothing further
Handshake AIAlso: Evaluation	Handshake, the careers network, recruits students and specialists into paid AI lab projects that produce and grade model output.	Text	No buyer rate shown; contributor pay varies by project	Nothing further
InnodataAlso: Annotation, Evaluation	Trades on Nasdaq as INOD. Dataset labeling, task-specific model tuning and safety testing, as managed services.	Text, image, audio, video	Not published: quote only	Nothing further
MercorAlso: Evaluation	A marketplace of vetted experts who build training data, grade model output and write benchmarks and RL environments for AI labs.	Text, code	Not published: quote only	Said on 9 Jul 2026 that it is buying Deeptune, an RL environment builder. (Mercor blog)
micro1Also: Evaluation	An AI interviewer screens the experts; micro1 then manages them on post-training and RL data projects for labs.	Text, video, code	Not published: quote only	Nothing further
Prime Intellect	Hosts the Environments Hub, where anyone can publish RL and eval environments; also open-source RL training code and a hosted training service.	Text, code	Not published: quote only	Hub opened 27 Aug 2025. (Prime Intellect blog)
Snorkel AIAlso: Evaluation	Expert-built datasets, benchmarks and evals sold as a service to labs and enterprises, along with bespoke agent systems.	Text, code	Not published: quote only	$350M Series E announced 22 Sep 2026. (Snorkel AI press release)
Surge AIAlso: Licensed data, Evaluation	Writes RLHF and SFT examples, rubrics and verifiers, runs human evals and RL environments, and sells ready-made expert datasets to AI labs.	Text, image, audio, video, code	Not published: quote only	Works in 70+ languages, by its own count. (Surge AI products page)
TolokaAlso: Evaluation	Amsterdam HQ. Supervised fine-tuning examples, RLHF and DPO comparisons, self-grading eval and RL tasks, red teaming.	Text, image, audio, video, code	Not published: quote only	Nothing further
TuringAlso: Evaluation	Post-training data and evals for labs, plus RL environments: copies of Jira, Salesforce and Zendesk screens and MCP back ends.	Text, code, image	Not published: quote only	Nothing further

Disclosure: any company in this list could end up paying us for an introduction that turns into work. Entries, their sequence and our descriptions are written the same way regardless.

Which companies generate synthetic training data?

Software that generates tabular records, text or rendered imagery, usually from seed data or a schema, often with privacy masking built in.

Synthetic: 7 companies, alphabetical. Read 7 November 2026.
Company	What it offers	Data types	Published pricing	Also noted
Aindo	Synthetic tables for regulated industries, from a company in Trieste, Italy.	Tabular	Not published: quote only	Holds Europrivacy (GDPR) certification, per its security page. (Aindo security page)
K2view	A data management vendor whose generator mixes AI, rules, cloning and masking to produce test, analytics and AI data.	Tabular	No prices; can start free	Nothing further
MOSTLY AI (Syntho)	Synthetic tables and text. The MOSTLY AI name now belongs to Syntho, an Amsterdam platform, and trades as MOSTLY AI, powered by Syntho.	Tabular, text	Not published: quote only	Brand sale announced by Syntho on 9 Jun 2026. (Syntho announcement)
Parallel Domain	Simulated sensor data for training and testing the perception of cars and robots.	Image, 3D, video	Not published: quote only	Nothing further
Rendered.ai	Renders labeled, physics-accurate images for vision models, including radar (SAR), infrared, multispectral and X-ray.	Image, 3D	$5,000/mo Teams; $15,000/mo Organizations; services: quote	Nothing further
SAS Data Maker	The SAS synthetic data product built on the software it bought from Hazy in November 2024.	Tabular	Not published: quote only	Hazy asset purchase: 12 Nov 2024. (SAS press release)
Tonic.ai	Three products: Structural for databases, Textual for documents and free text, Fabricate for data generated from schemas or prompts.	Tabular, text	Not published: quote only	Bought Fabricate on 22 Apr 2025. (Tonic.ai press release)

Disclosure: any company in this list could end up paying us for an introduction that turns into work. Entries, their sequence and our descriptions are written the same way regardless.

Where can you buy datasets on a marketplace?

Platforms where independent providers list data and set their own prices, plus a marketplace for recruiting paid human participants.

Marketplaces: 6 companies, alphabetical. Read 7 November 2026.
Company	What it offers	Data types	Published pricing	Also noted
AWS Data Exchange	Third-party data products sold through AWS Marketplace; sellers set prices and pay AWS a listing fee.	Tabular, text, audio, video	Seller-set. AWS takes 3% of public offers; private offers 1.5–3% depending on contract size	Nothing further
Databricks Marketplace	Data products from Databricks Data Partner Program members, delivered inside Databricks.	Tabular	Not published	Nothing further
Datarade	A catalog of commercial data products. Sellers price their own listings and Datarade keeps a commission.	Varies by provider	Seller-set. Seller plans: $0/yr + 30% commission, up to $12,000/yr + 15%	Nothing further
Defined.aiAlso: Licensed data	Marketplace for licensed speech, images, video, music, sound effects and text supplied by contributors.	Audio, text, video, image	Not published	Nothing further
ProlificAlso: Expert and RLHF, Evaluation	Recruits paid participants for studies and for AI human data; buyers fund participant pay and a platform fee.	Text, audio, video, image	Participants: $8/hr floor, $12/hr suggested. Platform fee extra; its rate is not on the page we read	Nothing further
Snowflake Marketplace	Paid data listings delivered inside Snowflake, priced per query, per billable event, monthly, by subscription or in installments.	Tabular	Seller-set. Usage billed after the fact; subscriptions billed in advance	Nothing further

Disclosure: any company in this list could end up paying us for an introduction that turns into work. Entries, their sequence and our descriptions are written the same way regardless.

Which companies evaluate and red-team models?

Platforms for running evals and tracing model behavior, and specialists in adversarial testing. Several vendors in other sections sell human evaluation too.

Evaluation: 7 companies, alphabetical. Read 7 November 2026.
Company	What it offers	Data types	Published pricing	Also noted
Braintrust	Run evals, score outputs and trace live LLM applications in one place.	Text, code	Starter $0; Pro $249/mo with 5 GB and 50,000 scores; Enterprise: quote	Shows SOC 2 Type II. (Braintrust pricing page)
Confident AI (DeepEval)	A hosted eval service on top of DeepEval, the Apache-2.0 testing framework it maintains for Python and TypeScript.	Text	Free tier exists; paid tiers unpriced on the page we read	Nothing further
Galileo	Monitoring, evals and guardrails for agents; Cisco now owns it.	Text	Not published: quote only	Cisco's intent to buy announced 9 Apr 2026; completed May 2026, per Cisco. (Cisco blog)
Giskard	French platform for testing agents and red-teaming them continuously. Open-source Python library; paid Professional tier.	Text	Open source: free. Professional: no price shown	Nothing further
Gray Swan AI	Security for AI: Shade for red-team tests, Cygnal for filtering inputs and outputs, and Arena, a red-teaming community it puts at 23,000+ people.	Text	Not published: quote only	Nothing further
LangSmith	LangChain's tool for tracing, testing and watching LLM agents in production.	Text, code	Developer seat $0; Plus $39/seat/mo with 10,000 base traces, metered beyond; Enterprise: quote	Nothing further
Patronus AIAlso: Expert and RLHF	Builds simulated copies of websites and company systems where agents train and get stress-tested, and grades agents automatically.	Text, code	Not published: quote only	Nothing further

Disclosure: any company in this list could end up paying us for an introduction that turns into work. Entries, their sequence and our descriptions are written the same way regardless.

What changed in published terms?

Newest first. Each line records a shift in what a listed company publishes about its terms, prices or owners, dated by announcement or by when we spotted it.

Date	Change	Source
7 Nov 2026	Opening check of all 50 entries. Also seen that day: OpenAI's API pricing page describes its fine-tuning platform as winding down and shut to new users.	OpenAI API pricing page
17 Sep 2026	Beacon announced it is acquiring Haize Labs, an AI red-teaming company; it is not listed here as an independent vendor.	The Next Web
9 Jul 2026	Deeptune, maker of RL environments that mimic enterprise software, is being bought by Mercor, per Mercor's announcement.	Mercor blog
9 Jun 2026	Syntho acquired the MOSTLY AI brand; the product now runs as MOSTLY AI, powered by Syntho.	Syntho announcement
May 2026	Cisco says its acquisition of Galileo, announced on 9 April 2026, has completed.	Cisco blog
20 Oct 2025	KPMG agreed to buy the IP and technology of YData, so YData has no entry of its own here.	Pipedot report
22 Apr 2025	Tonic.ai acquired Fabricate, a tool for generating data from a schema or prompt.	Tonic.ai press release
19 Mar 2025	Reported by TechCrunch: NVIDIA bought Gretel, the synthetic data startup. Gretel has no entry of its own here.	TechCrunch
12 Nov 2024	SAS bought the principal software assets of Hazy, which now form part of SAS Data Maker.	SAS press release

How to read this directory

The index has no commercial tie to, and no blessing from, any company in these tables; editing is independent. Entries draw on each company's own pages as they stood on 7 November 2026, and ownership changes cite the announcement or the outlet that broke the news. Since nobody here tests vendors, an entry tells you what is on offer, never how good it is. Before you sign, run a paid pilot and work through the provenance checklist.

Folded into other companies, so without entries of their own: Gretel (NVIDIA, as reported), Hazy (its software lives on in SAS Data Maker), YData (technology sold to KPMG), Haize Labs (Beacon deal announced) and Deeptune (Mercor deal announced).

Sources behind this section

About the directory

How were companies chosen for this directory?

We list companies that sell training data, labeling, expert data, synthetic data, data marketplaces or evaluation to outside buyers, and whose own site we could read on 7 November 2026. Companies that have been acquired and folded into another product, or that only recruit workers, are noted under their owner instead.

Is this a ranking?

No. Inside each category the order is A to Z. Nobody here benchmarks or inspects vendors, so an entry only restates what that company says publicly.

Why do so many entries say "Not published"?

Most training data vendors price by quote after scoping a project. Where a company publishes no price, we say so rather than estimate one.

Why are Outlier and Alignerr missing?

Both recruit contributors rather than sell to buyers. Outlier belongs to Scale AI and Alignerr to Labelbox, and a buyer signs with those parents, which do appear.

How do I know an entry is current?

Every row shows the date its source page was read. When a re-read finds a change in terms, pricing or ownership, it goes into the terms changelog on this page with the date we saw it.