What are the best training data companies in 2026?
Last updated: November 9, 2026
The strongest AI training data companies in 2026 are Scale AI, which covers the most ground (LLM, robotics, public sector), Surge AI, known for premium expert post-training data, Mercor, a professional-expert network and evals provider, micro1, which offers RL environments and expert data, and Labelbox, which combines a platform, an expert network and RL environments. If you need a self-serve human panel with published prices, choose Prolific. If you want licensed off-the-shelf datasets, look at Appen. If you need specialist computer vision, medical or sensor work, consider iMerit or Sama.
This is an independent research guide. We did not test these services hands-on. Our rankings come from each company's own published offerings and from current public reporting. Read the methodology.
The ranked list
1. Scale AI
What it is: A data and AI-applications company, founded in 2016. Its Data Engine covers data generation, RLHF, red teaming, evaluation and annotation of text, images, video and 3D/LiDAR. It also runs private benchmarks (leaderboards) and physical-AI data programs.
Best for: Frontier labs, large enterprises and government buyers that want one vendor for many modalities, including robotics and defense.
Strengths: It covers more modalities than almost any rival. Scale says 25% of its contributors hold advanced degrees. Its published customer stories include Physical Intelligence and Universal Robots (robotics), Mayo Clinic, and the U.S. CDAO.
Drawbacks: In 2025 Meta took a reported 49% minority stake. Meta's 10-K records the investment at $13.80B; the widely reported deal figure was $14.3B. Some labs that compete with Meta reacted: Forbes reported in May 2026 that OpenAI left, while Google paused and then resumed work. Leadership has also changed. Francis deSouza became CEO in August 2026. There is no public price list.
Pricing: Custom quotes through a demo request.
Official site: scale.com
2. Surge AI
What it is: A human-data company for post-training. It offers an expert workforce, off-the-shelf (OTS) datasets, a Training Lab that post-trains open models on Surge data, and public benchmarks such as DAYJOB, Riemann-bench and Hemingway-bench.
Best for: Model builders who want very high-quality coding, reasoning, writing and agentic data, and who care more about quality than unit price.
Strengths: Surge publishes before-and-after training results for its datasets. For its OTS data it states you can “train on the full dataset first, and pay only if it improves your metrics.” Industry compilations report $1.2B in 2024 revenue, which would make it one of the largest players.
Drawbacks: Its public site says very little about process, security or how engagements work. It is built for well-funded AI teams, not small projects.
Pricing: Custom. OTS data is sold on a pay-if-it-improves basis, according to Surge.
Official site: surgehq.ai
3. Mercor
What it is: An expert network that supplies human data, evals and RL environments. Its pool includes physicians, lawyers, engineers and consultants. It also runs the APEX benchmark family and a program that lets companies sell their own workflow data.
Best for: Labs and enterprises that need professional-domain data (legal, finance, medicine, engineering, software) and evaluation against real professional work.
Strengths: Mercor says it has 5M+ experts, an average contracted rate of $90/hr and $4M+ in daily payouts. It builds benchmarks with industry partners, including Harvey and Cognition. CNBC reported a $10B valuation in October 2025. Secondary sources cite reporting by The Information of a $2B+ gross run rate in June 2026.
Drawbacks: Its run-rate figures are gross and self-reported or reported by the press, not audited. It is moving into enterprise AI deployment, which may compete with buyers' own priorities.
Pricing: Custom, through sales.
Official site: mercor.com
4. micro1
What it is: A “data lab” offering expert human data, Realm (RL environments), Cortex (contextual agent evaluation) and real-world robotics data.
Best for: Teams training agents that need realistic, long-horizon environments and expert trajectories, and U.S.-centric buyers.
Strengths: In October 2026, micro1 said it had passed a $1B annual run rate while staying profitable, with a core team of about 180. It also says it priced a Series B at a $4B valuation, with Microsoft among the investors. It has committed to spend $1B on acquiring enterprise data over the next 12 months. It states a U.S.-first policy and says it will not work with adversarial nations.
Drawbacks: Its financial figures are self-reported. A very lean core team and fast growth mean you should check account coverage and QA depth for your project.
Pricing: Custom.
Official site: micro1.ai
5. Labelbox
What it is: Originally annotation software. It now also offers Horizon (RL environments and evals), Terra (robotics data), the Alignerr expert network, and Recursion (managed agents).
Best for: Teams that want platform tooling and managed expert data from the same vendor, and labs buying RL environments.
Strengths: In October 2026 it announced a $100M+ Series E at a $1.25B pre-money valuation. In the same announcement it said it is profitable, growing more than 5x year over year, has 400+ customers and works with 90%+ of leading U.S. AI labs. Alignerr lists 2.6M+ skill-tested experts. The Series E post cites an expert network of 4M+ people.
Drawbacks: The company is now also pushing an enterprise-agents product, so its focus is split. Its expert-network figures differ between pages.
Pricing: Custom, through sales.
Official site: labelbox.com
6. Turing
What it is: A company that helps frontier labs improve models and helps enterprises deploy agentic systems, drawing on a global expert network. It is best known for coding and reasoning data.
Best for: Labs that need large volumes of software-engineering and reasoning data from vetted developers.
Strengths: It has a large developer talent network and offers dataset samples on request. Compilations report about $300M ARR as of March 2025.
Drawbacks: The public website gives very little detail on its offerings, so you will need to ask for samples and specifications. Its most recent comparable scale figure is from 2025.
Pricing: Custom.
Official site: turing.com
7. Snorkel AI
What it is: A company that grew out of the Stanford AI Lab and now sells “expert data development”: specialized training data, environments and benchmarks for frontier labs. Its benchmarks include Terminal-Bench and enterprise sets such as SnorkelFinance and SnorkelLegal.
Best for: Hard, specialized domains where frontier models fail at the edges, and buyers who want research-grade evaluation.
Strengths: It has deep research credibility and publishes many benchmarks. It also funds open-benchmark grants. Compilations report a $375M run rate and a $3.5B valuation.
Drawbacks: The company has shifted away from the weak-supervision software it was known for, so older reviews may describe a different product. Pricing is not public.
Pricing: Custom. You can request dataset samples.
Official site: snorkel.ai
8. Invisible Technologies
What it is: A company that provides AI training and RL environments for model builders, and also sells enterprise AI products: Neuron (data), Synapse (evals), Axon (agents) and Meridial (an expert network).
Best for: Model builders and enterprises that want training data and deployment help together.
Strengths: It says it has trained over 80% of leading AI models. That is a company claim. Its Meridial network lists 24,000+ vetted experts. Published case studies include Cohere.
Drawbacks: Much of the business is enterprise automation, not pure data work. Compilations report a smaller revenue base: $134M self-reported for FY2024.
Pricing: Custom, through a demo.
Official site: invisibletech.ai
9. Prolific
What it is: A platform for collecting human data and running evaluations, with 300,000+ active verified participants. It also has a Domain Experts pool and managed AI data services.
Best for: Researchers and AI teams that want self-serve human evaluation or preference data with transparent prices and no contract.
Strengths: Pricing is published. Its quality system, Protocol, runs 40+ checks. 2026 releases added agent-trajectory capture and authenticity checks that detect AI agents or synthetic responses.
Drawbacks: The platform fee is a large share of spend. You design and manage tasks yourself unless you pay for managed services.
Pricing: Pay as you go. You pay the participant reward you set plus a platform fee, usually 42.8% for corporate customers and 33.3% for academic or non-profit customers. Prolific recommends paying participants at least £9/$12 per hour; the minimum is £6/$8. Managed services are quoted per project.
Official site: prolific.com
10. Appen
What it is: A human-data company working since 1996. It offers custom collection, frontier-model alignment data and an off-the-shelf catalog of 646 datasets across nine categories.
Best for: Buyers who want licensed, documented datasets quickly (especially speech and multilingual), or who need large-scale human evaluation.
Strengths: Every catalog dataset comes with a spec sheet covering consent, provenance and inter-annotator agreement. The license is perpetual and non-exclusive with commercial rights. Appen holds SOC 2 Type II and ISO 27001 certification and covers 500+ locales.
Drawbacks: Because catalog licenses are non-exclusive, other buyers can train on the same data. Compilations report $249M in trailing revenue through June 2026, which is below the fastest-growing expert-data firms.
Pricing: Quotes for each dataset or custom program. Samples are available under NDA.
Official site: appen.com
11. TELUS Digital
What it is: The global technology arm of TELUS, wholly owned by TELUS since October 31, 2025. Its AI data offering covers collection, annotation, validation, SFT, RLHF, red teaming, agentic trajectories and off-the-shelf datasets.
Best for: Enterprises that need multilingual data at very high volume from a large, established provider.
Strengths: TELUS Digital says it has 1M+ contributors, 500+ languages and dialects, and over 2B labels a year. It expanded delivery in Asia-Pacific and Argentina in 2026.
Drawbacks: It is part of a large company that also sells customer-experience services, so AI data is one line of business among several. Its website blocked our automated reader, so these details come from TELUS Digital's pages as summarized by our research search and from TELUS's press release.
Pricing: Custom.
Official site: telusdigital.com
12. Toloka
What it is: A provider of training data for AI agents and LLMs. It offers RL environments, specialized datasets, evaluation and red teaming, off-the-shelf datasets, robotics and egocentric video, and the public Toloka Arena leaderboard. Experts join through Mindrift.
Best for: Teams building computer-use, coding, research or corporate-assistant agents that want both environments and data.
Strengths: It is focused on agent data and offers a self-service agreement as well as managed projects.
Drawbacks: It is smaller than the leaders; compilations report $26.4M in FY2024 revenue. Check delivery capacity for very large programs.
Pricing: Custom. Self-service terms are available.
Official site: toloka.ai
13. iMerit
What it is: An expert data company, now part of EXL. It runs the Ango Hub platform and iMerit Scholars, a network of 25,000+ domain experts in 60+ countries. Its work covers generative AI, autonomous mobility, medical AI and robotics.
Best for: Domain-heavy work such as DICOM and medical imaging, LiDAR and sensor fusion, and math or chain-of-thought reasoning data.
Strengths: It combines tooling with an in-house workforce, has strong sensor-fusion and medical capability, and advertises guaranteed quality.
Drawbacks: EXL completed its acquisition of iMerit, so priorities may shift during integration. Current scale figures are not public.
Pricing: Custom, through a demo.
Official site: imerit.ai
14. Sama
What it is: A managed data annotation, validation and model-evaluation company working since 2008. Its products are Sama Annotate and Sama Curate, and it covers computer vision, multimodal and sensor data.
Best for: Production computer-vision programs such as autonomous vehicles, retail, agriculture and robotics, and buyers who care about impact sourcing.
Strengths: Sama is a recertified B Corp. It reports a 99% first-batch acceptance rate and says customers stay an average of 8 years; both are company claims. Case studies include Walmart.
Drawbacks: It puts less emphasis on frontier-LLM reasoning data and RL environments than the leaders. No current scale figures are public.
Pricing: Custom. Scope and pricing are discussed on an expert call.
Official site: sama.com
Comparison table
| # | Company | Core offering | Best for | Public pricing? |
|---|---|---|---|---|
| 1 | Scale AI | Multimodal data engine, RLHF, evals, physical AI | Breadth, robotics, government | No |
| 2 | Surge AI | Expert post-training data, OTS datasets | Premium coding and reasoning data | No (pay-if-improves OTS) |
| 3 | Mercor | Professional expert network, evals, RL environments | Legal, finance, medicine, engineering | No |
| 4 | micro1 | Expert data, RL environments, agent evals | Agent training, U.S.-first buyers | No |
| 5 | Labelbox | Platform, Alignerr experts, Horizon environments | Tooling plus managed data | No |
| 6 | Turing | Developer and expert network | Coding and reasoning data | No |
| 7 | Snorkel AI | Expert data development, benchmarks | Hard specialized domains | No |
| 8 | Invisible | AI training plus enterprise AI platform | Training plus deployment | No |
| 9 | Prolific | Self-serve participant platform | Human evaluation, research | Yes (reward + 42.8%/33.3% fee) |
| 10 | Appen | Custom data, 646-dataset OTS catalog | Licensed datasets, speech | No |
| 11 | TELUS Digital | Full-lifecycle AI data services | Multilingual high volume | No |
| 12 | Toloka | Agent data, RL environments, evals | Agent builders | No |
| 13 | iMerit | Ango Hub plus Scholars experts | Medical, LiDAR, sensor fusion | No |
| 14 | Sama | Managed annotation and validation | Production computer vision | No |
How to choose a training data company
- Name the data type first. Pre-training corpora, SFT demonstrations, preference/RLHF data, RL environments, evaluations and computer-vision labels each lead to different vendors. Expert networks (Surge, Mercor, micro1, Turing) suit post-training. Managed annotation firms (Sama, iMerit) suit vision and sensor data.
- Ask for a paid pilot or samples. Most vendors offer dataset samples, often under NDA. Score them against your own rubric before you commit volume.
- Check expert verification. Ask how credentials are checked, how contributors are tested and how AI-generated submissions are detected.
- Insist on measurable QA. Ask for inter-annotator agreement metrics, review layers and how disagreements are resolved.
- Review ownership and licensing. Is custom data exclusive to you? Off-the-shelf licenses, such as Appen's, are usually non-exclusive.
- Consider conflicts and data security. Ask about ownership, other customers and certifications such as SOC 2 and ISO 27001.
- Compare pricing models. Most vendors quote per project, per task or per expert-hour. Self-serve platforms such as Prolific publish their fees.
Frequently asked questions
What is a training data company?
A training data company creates, collects, labels or licenses the data used to train and evaluate AI models. In 2026 that includes expert-written demonstrations, preference rankings for RLHF, RL environments, agent trajectories, red-teaming and evaluation sets, as well as classic annotation of images, video, audio and LiDAR.
Which are the biggest AI training data companies?
By reported revenue or run rate, the largest in 2026 include Mercor (a reported $2B+ gross run rate), Surge AI ($1.2B reported 2024 revenue), micro1 ($1B+ self-reported run rate) and Scale AI ($870M reported 2024 revenue). These figures use different measures and are not directly comparable.
How much does AI training data cost?
Most vendors quote custom prices based on task complexity, expertise required and volume. Expert work can be expensive: Mercor reports a $90/hr average contracted rate for its experts. Prolific publishes its model: the participant reward you set plus a platform fee, usually 42.8% for corporate customers.
Is Scale AI owned by Meta?
No. Meta holds a reported 49% minority stake, bought in 2025. Meta's 10-K says it does not have significant influence over Scale. Scale's founder, Alexandr Wang, moved to Meta but stayed chairman, and Francis deSouza became CEO in August 2026.
What is the difference between data labeling and expert human data?
Data labeling usually means tagging existing data, such as drawing boxes or classifying text, at volume. Expert human data means domain specialists (doctors, lawyers, engineers, mathematicians) writing, grading or correcting model outputs and building tasks and environments that frontier models learn from.
Can I buy off-the-shelf training datasets?
Yes. Appen licenses a catalog of 646 datasets. Surge AI, Toloka and Mercor also offer off-the-shelf data. Check whether the license is exclusive and what provenance documentation comes with it.
Which training data companies handle robotics data?
Scale AI, micro1, Labelbox (Terra), Mercor, Toloka, iMerit and Sama all advertise robotics or physical-AI data.
Related guides: Best AI datasets · Robotics data companies · Best inference providers