Meta Description: Compare the top synthetic data companies in 2027 for privacy-safe tabular data, computer vision imagery, autonomy simulation and LLM training text sets.
Most machine learning projects stall for one of three reasons: there is not enough data, the data that exists cannot leave a secure environment, or the rare cases that matter most barely appear in the records. Synthetic data companies address all three. They generate tabular records that keep the statistical shape of a production database without exposing real customers, render labeled images and sensor data for scenes nobody has photographed yet, simulate driving and flight conditions that would be dangerous to stage, and produce structured text for training and evaluating language models.
Synthetic data is not a full replacement for real, human-labeled examples. Many teams still need people to label a seed set, review edge cases or validate what a generator produces. If that is your main need, our list of the top data annotation and labeling companies covers the human-labeling side of the market, and the two approaches are often used together.
The market has also consolidated. NVIDIA acquired Gretel in 2025, SAS bought the principal software assets of Hazy in late 2024, KPMG agreed to acquire the IP and technology assets of YData in October 2025, and Syntho acquired the MOSTLY AI brand in June 2026. The ten companies below are all operating as of October 2026, either independently or as a product clearly continuing under an acquirer. Each profile notes ownership changes where they apply, so ML and data teams can judge the roadmap behind the tool they buy.
Synthetic Data Companies At A Glance
Company
Headquarters
Best For
Core Services
NVIDIA
Santa Clara, California, USA
LLM training data and private tabular data
NeMo Data Designer, NeMo Safe Synthesizer
Tonic.ai
San Francisco, California, USA
Test data and de-identified training data
Tonic Fabricate, Tonic Structural, Tonic Textual
Syntho
Amsterdam, Netherlands
Privacy-preserving tabular data in Europe
Syntho Engine, MOSTLY AI platform and open-source SDK
K2view
Yokneam, Israel
Enterprise test data across many systems
Synthetic data generation, test data management, data masking
SAS
Cary, North Carolina, USA
Regulated analytics teams
SAS Data Maker with Hazy technology
Applied Intuition
Sunnyvale, California, USA
Automotive and defense autonomy programs
Sensor Sim, Synthetic Datasets, simulation and validation tools
Synthetic data platform, synthetic data as a service
Best Synthetic Data Companies In 2027
1. NVIDIA — Apache 2.0 NeMo Data Designer Built On The Gretel Acquisition
NVIDIA entered synthetic data as a vendor in its own right when it acquired Gretel in March 2025. TechCrunch, citing Wired, reported the price in the nine figures, above Gretel’s last valuation of $320 million. Gretel was a San Diego company founded in 2019 with about 80 employees, and its team and technology were folded into NVIDIA’s generative AI services for developers. The old gretel.ai address now redirects to NVIDIA’s synthetic data page, so buyers who used Gretel should treat NVIDIA’s NeMo tools as the continuing product line.
Two tools matter here. NeMo Data Designer is open source under the Apache License 2.0 and generates text, structured data and images either from scratch or from seed datasets. It supports sampler columns for controlled diversity, LLM-generated columns, dependency-aware field relationships, multimodal context, connections to external tools through MCP servers, and validation through Python, SQL and LLM judges. That makes it a strong fit for teams building instruction, reasoning or evaluation sets for language models.
NeMo Safe Synthesizer covers the privacy side. NVIDIA’s documentation describes it as an early access API that creates private synthetic versions of sensitive tabular datasets, with PII detection and replacement before training, differential privacy as a training parameter, and quality and privacy scores in its reports. NVIDIA’s documentation points teams with little or no real data toward Data Designer and teams with sensitive real data toward Safe Synthesizer. Teams combining synthetic text with human feedback should also review our list of RLHF and LLM training data companies.
2. Tonic.ai — Three Products For Structured, Unstructured And From-Scratch Data
Tonic.ai was founded in 2018 in San Francisco by Ian Coe, Karl Hanson, Andrew Colombi and Adam Kamor, and its team also works from Atlanta, New York, Washington, D.C. and London. Its about page references an $8 million Series A in 2020 and a $35 million Series B in 2021, with Insight Partners among the investors shown. The company started in test data management and now positions its products for model training as well as software QA.
The product line is split by data type. Tonic Structural de-identifies, subsets and synthesizes structured and semi-structured data from relational databases, data lakes, NoSQL stores and flat files. Tonic Textual de-identifies, redacts and synthesizes unstructured text and files, using named entity recognition, which is useful for teams that want to train on support tickets, clinical notes or contracts without exposing personal data. Tonic Fabricate generates relational data, free text and mock APIs from scratch when there is no source data at all.
Tonic is one of the few vendors here with public pricing. Its pricing page lists a free Fabricate plan with $5 of monthly credits and a Plus plan at $29 per month with $25 of credits, while Structural, Textual Enterprise and Fabricate Enterprise are custom quotes. The site lists SOC 2 and HIPAA compliance and GDPR support, and its logo wall includes eBay, Philips, Qualcomm, the NHL and Blue Shield of California, with published case studies from Flexport and Everlywell.
3. Syntho — Amsterdam Platform That Acquired The MOSTLY AI Brand In 2026
Syntho is an Amsterdam-based synthetic data company led by CEO Wim Kees Janssen. On June 9, 2026 it announced the acquisition of the MOSTLY AI brand, which now continues as “MOSTLY AI, powered by Syntho.” MOSTLY AI was founded in Vienna in 2017 and was one of the best-known names in privacy-preserving tabular synthesis. The announcement describes a brand acquisition and a plan for a unified platform combining multiple generation methods; it does not detail changes to existing MOSTLY AI contracts, so current customers should confirm support terms directly.
Buyers get two toolsets. The Syntho Engine covers synthetic data masking with a sensitivity scanner, rule-based generation through formulas, patterns and subsetting, and AI-generated synthetic data with quality assurance reports, upsampling and foreign key management. The MOSTLY AI side offers a data intelligence platform that can be deployed on Kubernetes or OpenShift, plus an open-source Synthetic Data SDK under the Apache 2.0 license that runs locally, uses the TabularARGN model architecture and includes built-in differential privacy.
Syntho’s customer logos lean toward Dutch public and healthcare institutions, including KPN, Philips, Nederlandse Spoorwegen, Leiden University Medical Center, UMCG and Statistics Netherlands, while MOSTLY AI testimonials include Swiss Post and Erste Group. Syntho lists Basic, Standard and Ultimate plans with feature-based pricing on request, typically on a one-year license, and says it has won the Philips Innovation Award and the Global SAS Hackathon.
4. K2view — Four Generation Methods On An Entity-Based Data Model
K2view was founded in 2009 by Achi Rotem and Rafi Cohen and built its business on data integration and data products for large enterprises. Synthetic data sits inside its data protection and provisioning line, alongside test data management, masking and tokenization. The company was named a Visionary in the 2025 Gartner Magic Quadrant for Data Integration Tools and won gold at the AI & Data Awards 2025 for its synthetic data generation platform.
What separates K2view is its entity-based approach. Instead of generating table by table, it uses an entity data model, such as a customer, account or order, as the blueprint, which keeps referential integrity intact across many source systems, including mainframes, enterprise applications, relational and NoSQL databases, and cloud stores. Its synthetic data page describes four methods that can be blended per scenario: AI generation trained on masked production data, rules-based generation for new features and negative testing, data cloning for load and performance testing, and intelligent masking that preserves relationships.
The platform also supports subsetting, reservation, data aging and rollback, and an AI agent that can read test cases and provision matching synthetic entities. K2view’s customer logos include Walmart, Verizon, Vodafone, BBVA and Regions Bank, with a testimonial from AT&T. It is best suited to enterprises whose main synthetic data problem is producing compliant, connected test and training data from sprawling legacy estates rather than generating a single analytics table.
5. SAS — Data Maker With Hazy Technology And Differential Privacy
SAS, the analytics software company headquartered in Cary, North Carolina, first announced SAS Data Maker in 2024. In November 2024 it acquired the principal software assets of Hazy, a UK synthetic data company, and integrated that technology into Data Maker. The product reached general availability through the Microsoft Marketplace in November 2025, and SAS has said it plans to add other cloud providers and integrate the tool into SAS Viya. Data Maker was also launched in private preview as a native app on the Snowflake Marketplace in 2024.
Data Maker is aimed at business and analytics users as much as data scientists. It has a no-code graphical interface, supports multitable source data, time series data and differential privacy, and includes built-in visual evaluation metrics that compare distributions between real and synthetic data. SAS also highlights lineage, audit logs and regulatory-readiness support, which matter to auditors in banking, insurance and healthcare.
According to SAS, early adopters in health care, financial services and government tested the product in private preview. SAS reports that a UK financial services firm saw a 28% improvement in credit scoring model accuracy and that a European telecom company cut data access time from weeks to minutes. These are vendor-reported results, but they show where the product is pointed: regulated organizations that already run SAS and want synthetic data inside a governed analytics stack.
6. Applied Intuition — Autonomy Simulation Vendor To 18 Of The Top 20 Automakers
Applied Intuition was founded in 2017 by Qasar Younis and Peter Ludwig and builds software for vehicle intelligence across cars, trucks, mining, construction, agriculture and defense. In June 2025 it closed a $600 million Series F at a $15 billion valuation, led by BlackRock-managed funds and Kleiner Perkins, and said that 18 of the top 20 global automakers use its solutions. Its partner list includes Toyota, Porsche, Stellantis, Volkswagen and Komatsu, and its homepage notes a $25 million U.S. Navy contract award for AI sensor fusion.
Synthetic data is one part of a much larger autonomy toolchain. The company’s Synthetic Datasets product, built on its Sensor Sim technology, provides physically accurate synthetic sensor data for advanced driver assistance and automated driving models. Marketplace listings describe labeled training data for parking perception tasks such as parking spot detection and depth estimation, traffic sign recognition, off-road terrain segmentation and action recognition from aerial assets.
Applied Intuition also turns real-world sensor logs into curated, labeled segments for training and evaluation, and uses neural reconstruction and behavior agents to create controllable scenario variations. It is a natural choice for OEMs and defense programs that want synthetic data, simulation and validation from one supplier. Smaller computer vision teams that only need a dataset may find a focused vendor easier to engage.
7. Parallel Domain — PD Replica Neural Reconstruction From Drive And Flight Logs
Parallel Domain was founded in 2017 by Kevin McNamara and works from San Francisco, Karlsruhe in Germany and Vancouver in Canada. In November 2022 it announced a $30 million Series B led by March Capital, with Costanoa Ventures, Foundry Group, Calibrate Ventures and Ubiquity Ventures participating, and named Google, Continental, Woven Planet and Toyota Research Institute among its customers at the time.
Its current platform has two products. PD Replica builds photorealistic neural reconstructions from real drive or flight logs using camera, lidar and GPS data, with lidar described as improving fidelity but optional. Each replica includes world physics, HD maps, scene segmentation, reconstructed dynamic agents, object insertion and matched lighting. PD Sim then configures and runs test scenarios with deterministic camera, lidar and radar simulation at scale, with controlled variations in lighting, traffic and agent trajectories.
The approach suits teams that already have hours of recorded sensor data and want to multiply its value, turning a single drive into many reusable scenarios across sensor setups, geographies and weather. Parallel Domain also publishes sim-to-real gap reports that measure how closely a replica matches the original scene, which helps teams quantify how far they can trust synthetic results. Its industries include automotive, trucking, drones, robotics, agriculture and defense.
8. Duality AI — Falcon Digital Twin Simulator Built On Unreal Engine
Duality, legally Duality Robotics, Inc., was founded in fall 2018 and is headquartered in San Mateo, California. Its co-founders bring two relevant backgrounds: CEO Apurva Shah worked on film pipelines at Pixar, and Mike Taylor led Caterpillar teams building large field robots and won the DARPA Urban Challenge with Carnegie Mellon. The company says its team holds more than 70 patents across robotics, simulation and visualization.
Falcon is Duality’s digital twin simulator, offered as a cloud service. Scenarios are built in an Unreal Engine-based editor using a catalog of simulation-ready system, item and environment twins, and virtual sensors can be tuned to match observed data. Falcon supports open-loop runs for generating synthetic training data and closed-loop runs for validating autonomous systems, with a Python API and middleware that connects to ROS and ROS2. Duality has also announced Vibe Sim, an agentic co-pilot that generates synthetic data from prompts.
Duality’s focus areas are aerial autonomy, counter-UAS and off-road autonomy. SiliconANGLE’s 2023 coverage of its cloud launch named DARPA, Honeywell, KEF Robotics and Procter & Gamble as customers, and the company’s own site features case studies with Honeywell Aerospace, KEF Robotics and P&G. It suits robotics and defense teams that want engineering-grade digital twins rather than a fixed library of prebuilt datasets.
9. Bifrost — Stardust Synthetic Data For Land, Air, Sea And Space
Bifrost was co-founded in 2020 by CEO Charles Wong, who previously worked on perception models for self-driving cars at nuTonomy, and Aravind Kandiah. TechCrunch reported in October 2024 that the San Francisco company had raised an $8 million Series A led by Carbide Ventures, bringing total funding to $13.7 million, with staff in the U.S. and Singapore. Its site now lists Sequoia, Lux Capital, Airbus Ventures, Wavemaker, Carbide and Techstars as backers.
Bifrost’s synthetic data platform, Stardust, produces perception datasets in RGB, infrared, depth and segmentation, with pixel-perfect labels and scenario metadata matched to a customer’s own sensor imagery. The company organizes its work by environment: maritime use cases such as uncrewed surface vessels, collision avoidance and port surveillance; aerial perception and counter-UAS; satellite and geospatial monitoring; planetary terrain and landing for space; and off-road scenes with degraded visibility. A second product, Manifold, runs robot policies across simulator benchmarks and grades the results.
The organizations shown on Bifrost’s site include NASA, Saronic, Shield AI, Honda, NTT Data, ClassNK and ST Engineering. That mix points to the company’s sweet spot: teams training detectors for rare objects in hard environments, such as small vessels at sea or drones against clutter, where collecting and labeling real footage is slow or impossible.
10. Rendered.ai — SAR, Thermal And Hyperspectral Synthetic Imagery From $5,000 A Month
Rendered.ai is based in Bellevue, Washington, and its platform has been in production since late 2021. Co-founders Duane Harkness and Ethan Sharratt lead the company alongside Executive Chairman Nathan Kundtz, with Ario Sanchez as CEO. The company has built much of its business in earth observation and the geospatial industry, and lists AvaSure, Kallisto AI, COVAR, Vizgard, SatVu and OneAI among its customers, with a testimonial from Orbital Insight.
Rendered.ai stands out for sensor coverage beyond ordinary cameras. Its platform generates synthetic data for synthetic aperture radar, infrared, thermal, multispectral, hyperspectral and X-ray sensors, as well as RGB and custom or emerging sensors. That matters for satellite analytics, security screening and defense teams, where labeled real imagery is scarce, expensive or classified. It has also partnered with Kallisto AI on military camouflage across the electromagnetic spectrum.
Unusually for this market, Rendered.ai publishes subscription pricing. Its site lists the platform as a service from $5,000 per month for the Teams tier and $15,000 per month for the Organizations tier, with self-managed deployments quoted separately. It also offers synthetic data as a service on a project basis, scoped by dataset size and sensor modality, plus model development and auto-labeling engagements. Teams comparing it with human labeling of real imagery can see how that side of the market works in our list of computer vision annotation companies.
How To Choose A Synthetic Data Company
What Kind Of Data Do You Actually Need?
Start by separating the four segments in this list, because the vendors rarely overlap. Tabular and relational privacy data comes from Tonic.ai, Syntho, K2view, SAS and NVIDIA Safe Synthesizer. Language model training and evaluation text is best served by NVIDIA Data Designer, with Tonic Textual for de-identified real text. Labeled imagery and sensor data for perception models comes from Bifrost, Rendered.ai and Parallel Domain. Full autonomy simulation and validation is the domain of Applied Intuition, Parallel Domain and Duality. Buying a tabular tool for a vision problem, or the reverse, is the most common and most expensive mistake.
Are You Starting From Real Data Or From Nothing?
Some generators learn from a real dataset and reproduce its statistics, while others build data from rules, prompts or 3D scenes. If you hold sensitive production data you cannot share, a learned approach with privacy controls such as differential privacy fits best. If you have little data at all, or need cases that have never happened, rules-based generation, LLM pipelines or simulation are the better route. Several vendors, including K2view, Syntho and NVIDIA, offer both paths, so ask which method each use case will actually use.
How Will You Measure Quality And Privacy?
Ask every vendor how they prove the output is useful and safe. Good answers include fidelity reports comparing distributions, downstream model accuracy against a real holdout set, privacy metrics or membership-inference testing for tabular data, and sim-to-real gap analysis for imagery. Insist on running your own evaluation before signing a multi-year contract, and keep a slice of real, human-labeled data as the benchmark.
Where Must The Data And Models Run?
Privacy-sensitive teams often need generation to happen inside their own environment. Check whether the product runs on premises, in your cloud account, through a marketplace or only as a hosted service. Open-source options such as NeMo Data Designer and the MOSTLY AI SDK run locally, SAS Data Maker runs through cloud marketplaces, and Syntho and K2view support self-hosted enterprise deployment.
Who Owns The Roadmap After An Acquisition?
Consolidation is reshaping this market. Gretel now lives inside NVIDIA, Hazy inside SAS, the MOSTLY AI brand at Syntho, and YData’s technology at KPMG. Before you commit, ask how long current products will be supported, whether pricing or licensing will change, and whether the acquirer sells the tool to outside customers or mainly uses it internally. For budgeting the wider program around synthetic data, our guide to AI development cost covers how data work fits into overall project spend.
Conclusion
The synthetic data market in 2027 is less a single category than four adjacent ones. NVIDIA, Tonic.ai, Syntho, K2view and SAS solve privacy and access problems for structured and text data, while Applied Intuition, Parallel Domain, Duality AI, Bifrost and Rendered.ai generate the imagery, sensor data and scenarios that perception and autonomy models need. The right choice depends far more on your data type, deployment constraints and evaluation plan than on any vendor’s headline claims.
Most mature teams end up blending synthetic and real data. Use generators to fill gaps, protect sensitive records and cover rare events, then validate against a real, human-labeled benchmark. Shortlist two or three vendors from the right segment, run a pilot on your own model, and judge them by the accuracy and privacy results you can measure.
