Synthetic data companies sell AI developers either software that generates artificial training data or generated tasks and simulated environments, and DataPacts records seven of them with legal entity, ownership changes and a dated source. Syntho announced on 2026-06-09 that it acquired the MOSTLY AI brand, trademark and related assets, with the value not disclosed. No filing or buyer page in this reference records an AI developer buying a synthetic dataset from one of the seven.
This directory sits within AI training data licensing, is alphabetical and carries no ranking. DataPacts sells no data, no generation software and no placement in the list. The fact-check note follows.
Fact-check note: every entity, date and claim below was read from the company's own page, a registry, a filing or a named announcement and checked on 2026-11-09; the method is on the methodology page.
What a Synthetic Data Company Sells to AI Developers
A synthetic data company sells AI developers the means to produce artificial records, as software or as generated tasks, and only rarely a finished dataset in the public record. Three supply routes appear in the sources checked:
- Generator software: a product or toolkit the developer runs to produce artificial records, such as MOSTLY AI's software development kit or NVIDIA's NeMo tools.
- Generated tasks and simulated environments: tasks, checks and simulations delivered to labs, as Eigent AI, Andromède and Veris AI describe their offers.
- In-house generation by labs: a lab produces synthetic data with its own models, which is not a vendor sale.
"Synthetic data vendor" also means a seller of software test data for databases and applications, a different market that this reference does not cover. The definition of the data itself, its four generation methods and its limits sit on the page about what synthetic data is in AI training and its limits.
Seven Synthetic Data Companies on Record, With Source and As-Of Date
The table lists every synthetic data company this reference records, alphabetically, in two segments, with what it sells or offers, its legal entity and ownership status, checked November 9, 2026. Three terms are used in the cells:
- Not stated: no primary source gives the field.
- Segment: generator software, or generated tasks and environments.
- Profile: a page on this site where one is planned; rows without one are covered here only.
| Company | Segment | What it sells or offers (its own page) | Entity and registry on record | Ownership or status | As of | Source |
|---|---|---|---|---|---|---|
| Andromède | generated tasks and environments | describes itself as a data lab for frontier agents; generates RL environments and tasks with scoring checks from real-world data | Andromede AI, Inc.; no registry number found | private beta, access by request | 2026-11-09 | andromede.ai |
| Eigent AI (CAMEL-AI) | generated tasks and environments | custom agent RL environments (enterprise plan); SETA terminal environments, 400 initial synthetic terminal tasks (CAMEL-AI, 2026-01-09) | EIGENT UK LTD (Companies House 15147455, incorporated 2023-09-19, London) | active | 2026-11-09 | eigent.ai; camel-ai.org; Companies House |
| Gretel | generator software | none on its own domain: www.gretel.ai 301-redirects to NVIDIA's synthetic-data page, which does not mention Gretel | not stated | redirect observed; no ownership record checked | 2026-11-09 | nvidia.com (redirect target) |
| MOSTLY AI (powered by Syntho) | generator software | synthetic data software for tabular and textual datasets; SDK under Apache 2.0 | brand of Syntho since 2026-06-09 | brand, trademark and related assets acquired by Syntho, announced 2026-06-09, value not disclosed | 2026-11-09 | mostly.ai; syntho.ai |
| NVIDIA | generator software | NeMo Data Designer and NeMo Safe Synthesizer (named on its synthetic-data page) | NVIDIA Corporation (SEC CIK 0001045810, Delaware) | active; also profiled as a frontier lab | 2026-11-09 | nvidia.com; sec.gov |
| Syntho | generator software | synthetic data software; owner of the MOSTLY AI brand | not stated | acquirer of the MOSTLY AI brand (2026-06-09, Amsterdam) | 2026-11-09 | syntho.ai |
| Veris AI | generated tasks and environments | simulation platform, "a simulated copy of your systems, data, APIs, and users"; priced "per bench or an annual licence" (Veris's claim) | Veris Technologies Inc.; no registry number found | active | 2026-11-09 | veris.ai |
Each row rests on the company's own page, a registry or the named deal announcement; no row records a sale of a synthetic dataset to an AI developer. The three companies that generate tasks are RL environment builders, and the full directory with funding is the list of RL environment companies with legal entity and funding.
Synthetic Data Generators: Software Sold to Developers, Not Finished Datasets
A synthetic data generator is software, or a model, that produces artificial records resembling real data, and the generator vendors on record sell it as a product or publish it under an open license. MOSTLY AI publishes its software development kit under the Apache 2.0 license, and NVIDIA's synthetic-data page names NeMo Data Designer and NeMo Safe Synthesizer. Gretel has no page of its own in the record: its domain redirects to NVIDIA, as observed on 2026-11-09, and a redirect alone is not a record of who owns Gretel. No price for a generator license or a synthetic dataset is on record, so none is given here, and this page offers no feature comparison, review or setup steps.
Rules reach the vendors' outputs and their use. Art. 50(2) of the AI Act, Regulation (EU) 2024/1689, requires providers of systems that generate synthetic audio, image, video or text to mark outputs in machine-readable form, and systems on the market before 2026-08-02 comply by 2026-12-02 under Art. 111(4), inserted by Regulation (EU) 2026/1744. Recital 111 counts synthetic data generation in the cumulative compute used to train a model.
Provider terms can also limit what a customer trains on a model's outputs, and each vendor or lab sets its own. Anthropic's Commercial Terms Section D.4, effective 2025-06-17, bar access to its services to "build a competing product or service, including to train competing AI models". This is information, not legal advice.
Who Buys From Synthetic Data Companies, and What the Record Does Not Show
No filing, regulatory summary or buyer page in this reference records an AI developer buying a synthetic dataset from a vendor; the labs that disclose synthetic data say they generate it with their own models. The statements on record, as of 2026-11-09:
- Anthropic, PBC, Form S-1, 2026-10-26: training data sources include synthetic data generated by its own models.
- xAI (SpaceXAI), Grok 4.7 public summary of training content, 2026-09-21: synthetic data generated with its own models, such as Grok 4.6, is answered "Yes".
- Apple, Foundation Models tech report, 2025: models are trained on responsible web crawling, licensed corpora and synthetic data.
Generated tasks reach labs as RL environments, sold by the builders in the table, and the buyer evidence for them sits on the page about RL environments, with the wider picture under how frontier AI labs acquire data.
Buyers of real company data go the other way. Mercor's listings, checked 2026-11-09, name synthetic or sample charts and synthetic or scripted traffic as "not a fit", and micro1's Data License Master Terms s.2.2, checked 2026-11-09, give the licensee ownership of synthetic data derived from licensed records. The detail on both sits on the parent page, so it is not repeated here.
Checking a Synthetic Data Company Before Buying From It or Supplying Source Records
A developer buying from a synthetic data company, or a data holder supplying it source records, can check five things, starting with the legal entity behind the brand.
- Legal entity and registry: the table shows which companies give a registry number and which do not, for example Companies House for Eigent AI and the SEC for NVIDIA. Syntho's site states no legal form.
- Ownership changes: the MOSTLY AI brand moved to Syntho on 2026-06-09; a domain redirect, such as Gretel's to NVIDIA, is not an ownership record.
- Rights in the source records and in derived synthetic data: a buyer may own what it derives, as in micro1's s.2.2.
- Personal data: output generated from personal records is outside the GDPR only if it is anonymous under Recital 26 of Regulation (EU) 2016/679, and the EDPB Guidelines 02/2026, adopted 2026-07-07 for consultation and not final as of 2026-11-09, name inference from synthetic data as a risk.
- Marking duties and provider terms: Art. 50(2) marking and terms such as Section D.4 above.
The contract clauses to read are derived data, ownership of outputs and deletion, set out in the guide to AI data licensing agreements. Four labeled lists summarize the position of a data holder supplying source records.
- Benefits: a holder can share data with less personal data if the output is anonymous; generator software under an open license can run in-house.
- Risks: the buyer may keep derived synthetic data; the brand owner can change; inference from synthetic data can still breach the EDPB test.
- Downsides: no public price for a synthetic dataset or generator license is on record; no customer record exists to check a vendor's claims against.
- Alternatives: license real records under a narrow derived-data clause; generate synthetic data in-house for internal use only; or not supply or license at all, as set out under alternatives to selling data.
How This Company List Is Built: Inclusion, Order and Gaps
A company enters this list only when a primary record shows it selling synthetic data, generator software or generated tasks, the order is alphabetical, and no company's place depends on referral fees. Tonic.ai appears in this reference only as the de-identification agent in the Spirit Airlines data sale (Dkt. 1684, filed 2026-10-05, sale not approved as of 2026-11-09), so it is not counted. Names on vendor round-ups are added when a primary record is checked.
The list carries no ranking and no score, and a row can be removed or moved between segments when a dated record changes, and it holds no price, customer or market-size figure because no checked record gives one. A reader can see how we source, date and check every figure before relying on any row.
Where Synthetic Data Companies Meet Privacy Law, Model Collapse and Real-Data Buyers
This page lists the companies; whether their output is personal data, what training on it does to models, and how it compares with real records sit on their own pages. Model collapse (Shumailov et al.) was published in Nature on 2024-07-24.
- Topics: synthetic vs real training data, model collapse and synthetic data and privacy law, with the hub on what synthetic data is in AI training and its limits.
- Related: RL environments, anonymization vs pseudonymization, PII removal and de-identification, AI data licensing deals, data labeling companies and US bankruptcy data sales.
Frequently Asked Questions
Which are the top synthetic data companies?
DataPacts does not rank synthetic data companies; the list is alphabetical, and none of the seven publishes a customer list in this reference. Veris AI states it prices "per bench or an annual licence", as its own claim, and no synthetic dataset price is on record.
Do synthetic data companies sell datasets or software?
Mostly software and generated tasks: the generator vendors on record sell or publish generation tools, and the task builders sell environments to labs. No finished synthetic dataset sale to an AI developer is on record.
Is there an open-source synthetic data generator?
Yes, MOSTLY AI publishes its software development kit under the Apache 2.0 license, as its site shows on November 9, 2026. DataPacts does not review tools or give setup steps.
Do synthetic data companies buy real company data?
No synthetic data company in this list runs a published program to buy company data; the buyers that do are profiled separately.
The programs are compared under AI training data buyers compared.