Synthetic Data Companies

Pick the vendor that can turn your real tables, text or images into data you can share, test with and train on. Tell us the data type and where it must stay; we name a fitting vendor before anyone is contacted.

Get a quote Should you use synthetic data?

Database tables
Free text and files
Images
Video and 3D
SAR and infrared

$5,000a month, Rendered.ai's Teams plan
$15,000a month, its Organizations plan
0 of 7vendor pages we track state an epsilon value
4ownership changes since late 2024: Hazy, Gretel, YData, MOSTLY AI

Prices as Rendered.ai prints them; counts from the vendors' own pages. Nothing here is an offer.

One row in, one row out

A generator learns the shape of your data, then writes new records that keep it. The test is whether the new row is useful and whether it gives a real person away.

customers.parquetrow 18,204

Customer age: Real: 47; Synthetic: 52
ZIP prefix: Real: 941; Synthetic: 940
Plan: Real: Business, annual; Synthetic: Business, annual
Monthly spend: Real: $1,284.50; Synthetic: $1,197.20
Tickets, last 90 days: Real: 3; Synthetic: 4

Privacy test: no exact copy of any real row; closest real row above the agreed distance

Invented example. Neither row is a real person. Changed values are highlighted.

De-identified is not the same as private

De-identification masks direct identifiers, which lowers risk but proves nothing. Differential privacy limits, by a stated epsilon, how much any one record can change what the generator learns. None of the seven vendor pages we track states an epsilon.

Privacy guarantees, compared

Three tests on every delivery

FidelityDistributions, correlations and joint patterns match the source
UtilityTrain on synthetic, test on real held-out data
PrivacyExact copies, closest-row distance, membership inference

A worked test plan

Four brands that changed hands

Nov 2024HazySoftware assets to SAS, now SAS Data Maker
Mar 2025GretelBought by NVIDIA, as TechCrunch reported
Oct 2025YDataTechnology assets agreed to KPMG
Jun 2026MOSTLY AIBrand bought by Syntho

When not to bother

Rare events are the point, the source table is small, you need an audit trail to real people, masking would do, or the world has moved since the data was collected. Five questions, one verdict.

Two markets that rarely overlap

Tabular vendors generate rows for databases, analytics and test environments. Image and simulation vendors render labeled scenes for vision, robotics and driving.

Tables and text

5 tabular vendors

Buy when production data cannot leave the team, or you need more rows than you have. Ask where the software runs.

Tonic.aiStructural for databases, Textual for free text and files, Fabricate for data from a schema or prompt.
Syntho (and MOSTLY AI, powered by Syntho)Tabular platform based in Amsterdam; now also runs the MOSTLY AI brand.
AindoPlatform from Trieste aimed at regulated sectors. Lists a Europrivacy certification.
K2viewMixes AI generation, business rules, cloning and masking. No prices; offers a free start.
SAS Data MakerWhere the Hazy generator lives on: SAS folded those assets into this tabular product.

Tabular synthetic data

Images, sensors and simulation

2 image and simulation vendors

Buy when real scenes are rare, dangerous or slow to label. Add one test: the gap between synthetic and real scenes on a real validation set.

Rendered.aiLabeled synthetic imagery, including SAR, infrared, multispectral and X-ray sensors.
Parallel DomainSimulation and synthetic data for perception in autonomous vehicles and robotics.

From $5,000 a month Rendered.ai is the only vendor here that prints a price.

Synthetic images and simulation

Disclosure: any vendor named on this page may pay an introduction fee when a team we put in touch signs with it. Whether a vendor pays has no effect on the list, its order or our notes. Full vendor list.

A vendor trial in five moves

Agree the pass marks before the trial starts, not after you see the data.

1
Write the brief
Data type, row or image counts, where the software must run, the privacy promise you need and what the data will feed.
2
Shortlist two or three
Only vendors that support your data type and deployment.
3
Trial on a slice
Data you would be comfortable sharing, inside your own environment where the vendor allows it.
4
Score the output
Fidelity, utility and privacy against the marks you agreed.
5
Read the contract
Who owns the generated data and the trained generator, retention of your source data, and how pricing scales.

Four questions for the vendor's privacy team

Is differential privacy applied during training, and with which epsilon and delta?
Which privacy tests run on every output?
Does any of our data leave our own cloud account or servers?
Who owns the generated data and the generator after the contract ends?

Short answers before you shortlist

What do synthetic data companies sell?

Mostly software. Tabular vendors such as Tonic.ai, Syntho and SAS Data Maker sell platforms that learn from your tables and generate new rows. Image and simulation vendors such as Rendered.ai and Parallel Domain generate labeled scenes for computer vision. A few also sell services around the software.

How much does synthetic data cost?

Most vendors do not publish prices. Of the seven we track, Rendered.ai is the one that does: platform plans from $5,000 a month (Teams) and $15,000 a month (Organizations). K2view shows no prices but offers a free start. The rest quote, usually on data volume, seats and where the software runs. The vendor list has each one's wording.

Is synthetic data reliable enough to train a model?

Sometimes, for part of the job. It works best for filling rare classes, testing pipelines and sharing data that cannot leave a team. Train on synthetic data, test on real held-out data, and compare the result with a model trained on real data before you rely on it.

Can synthetic data be traced back to real people?

It can. A generator can memorize and repeat rare records, so synthetic is not automatically anonymous. Only a formal method such as differential privacy, with a stated privacy budget (epsilon), gives a guarantee you can put a number on. Test every delivery for exact copies, distance to the closest real row and membership inference. See privacy guarantees.

Synthetic data or data masking: which do I need?

For a test database, masked or subsetted production data is often cheaper and good enough. Synthetic data earns its cost when masked data would still identify people, when you need more rows than you have, or when the data has to leave the team. Some tools, K2view's among them, mix both. Synthetic data or masking, in detail.

Who bought Gretel, Hazy and MOSTLY AI?

TechCrunch reported in March 2025 that NVIDIA acquired Gretel. SAS bought Hazy's principal software assets in November 2024 for SAS Data Maker. Syntho bought the MOSTLY AI trademark and related assets in June 2026. KPMG also agreed to take YData's technology assets in October 2025.

Where these facts come from

More in the write-ups. How this site works.