AI Governance & Security Vendor Wiki

7 min read

Gretel (NVIDIA)

Primary category: synthetic-masked-data.

One-liner — Formerly the best-known standalone synthetic-data startup. Since its March 2025 acquisition by NVIDIA, Gretel is no longer a product you buy: its technology survives as two open-source NVIDIA NeMo components, NeMo Data Designer (generate synthetic datasets for AI training and evaluation) and NeMo Safe Synthesizer (make a privacy-safe synthetic copy of a sensitive table).

What it does — Two successor components, which is what a buyer evaluates today:

  • NeMo Safe Synthesizer: takes a sensitive tabular dataset and produces an entirely synthetic one “with no one-to-one mapping to your original records”. The workflow is PII replacement, then model training (optionally with differential privacy), then a quality and privacy evaluation report. This is the direct descendant of Gretel’s core privacy-synthetics business. It needs an NVIDIA GPU (A100 or larger) on Linux for training and generation.
  • NeMo Data Designer: builds synthetic datasets from scratch or from a small seed, mixing LLM-generated fields, samplers and validators. It is used for fine-tuning, evaluation and agent test sets rather than for masking production data. It began as an early preview inside Gretel Navigator before the acquisition. NVIDIA open-sourced it (repo created 2025-10-16).

Pre-acquisition, Gretel sold a SaaS console and APIs (Gretel Navigator, synthetics, transforms/PII redaction, evaluation) plus open-source libraries such as gretel-synthetics.

Naming / provenance — Founded 2019 in San Diego by Alex Watson, Laszlo Bock, John Myers and Ali Golshan (CEO). Slug decision: the page keeps the slug gretel and title “Gretel (NVIDIA)”, not an NVIDIA product slug. Survey respondents and buyers still say “Gretel”. The NeMo component names have changed with NVIDIA’s packaging: NeMo Microservices was retired on 2026-10-01 in favour of “NeMo Platform” per NVIDIA docs banners. NeMo is a large framework, and synthetic data is a small part of it. If NVIDIA’s synthetic-data products later become a shopping unit in their own right, revisit with an nvidia-nemo page and make this an alias. Related NVIDIA page: nvidia-openshell, whose product lead is Gretel’s co-founder Ali Golshan per his NVIDIA bio.

Ownership & viability — acquired-closed, owner NVIDIA. Wired and TechCrunch reported on 2025-03-19 that NVIDIA had acquired Gretel, for a price “said to be nine figures”, above Gretel’s most recent $320M valuation. NVIDIA declined to comment and never issued a release, so price and exact close date rest on press reporting (announced and closed are both set to the report date). Corroboration that the deal closed: Gretel’s co-founder lists NVIDIA as his employer since 2025, gretel.ai 301-redirects to an NVIDIA page, the gretelai GitHub repos are archived (gretel-synthetics last pushed 2025-06-24), and NVIDIA ships the successor components. Pre-acquisition Gretel had raised more than $67M (Anthos Capital, Greylock, Moonshots Capital and others).

Judgement: acquired and absorbed, not shelved. The technology is actively developed (both NeMo repos had commits on 2026-10-26), but as NVIDIA infrastructure for AI developers, not as a privacy product for enterprise data teams. Former Gretel SaaS customers had to migrate. Whether NVIDIA offers a hosted equivalent outside NVIDIA AI Enterprise is unverified.

Positioning & differentiators — platform-module in synthetic-masked-data: it is now a component of NVIDIA’s NeMo AI stack, not a standalone vendor.

  • Versus tonic-ai and k2view: they mask and subset production databases for software testing, keeping real-looking rows with real referential structure. Safe Synthesizer learns a model of a table and samples new rows, which suits analytics and ML training more than application QA across many joined tables.
  • Versus MOSTLY AI (now Syntho) and SAS Data Maker (ex-Hazy): the same model-based synthetic-tabular niche. NVIDIA’s version is free and open source but GPU-bound and self-operated. The others are commercial products with UIs and support.
  • Data Designer competes more with DIY LLM-based dataset generation than with any masking vendor.

Who should choose them / anti-fit — Fits a fund with a quant or data-science team already on NVIDIA GPUs that wants synthetic versions of sensitive tabular data (client, transaction or HR data) to share with researchers, vendors or a fine-tuning job, or that needs synthetic eval sets for in-house models. Free OSS keeps the cost to compute and engineering time. Anti-fit: anyone who wants a supported, UI-driven product with a vendor contract (the standalone Gretel product is gone); dev/QA teams that need masked copies of application databases (use Tonic, K2view, or a TDM suite); and shops without GPU capacity.

Known weaknesses / gotchas —

  • The product you may remember no longer exists as sold. Survey answers of “Gretel” after 2025 likely mean NeMo components, legacy self-hosted libraries, or memory of the brand. Treat them with care.
  • Packaging churn. Gretel to NeMo Microservices to NeMo Platform within about 18 months. Pin versions and expect API changes.
  • GPU and Linux only for Safe Synthesizer training and generation.
  • Synthetic is not automatically safe. Model-based synthesis can memorise rare records. Rely on the privacy evaluation report and differential-privacy settings, and do your own review before releasing data outside the firm.

Deployment & data handling — self-hosted: both components run in the customer’s own environment (pip/uv install, containers, or NVIDIA’s platform). data_leaves_tenancy: "false" for self-run OSS. If Data Designer is pointed at a hosted LLM endpoint (NVIDIA-hosted or a third-party model provider), seed data and prompts go to that endpoint. Batch, out-of-band.

Integrations & partnerships — Part of NVIDIA NeMo, with enterprise support through NVIDIA AI Enterprise (terms not verified). Data Designer has a plugin catalogue and works with OpenAI-compatible model endpoints. MCP support unverified.

Compliance & FS traction — No certifications apply to self-run OSS. No financial-services customers verified, before or after the acquisition. NVIDIA says Safe Synthesizer defaults are intended to help meet HIPAA and GDPR. That is a design goal, not a certification.

Commercial — Open source, Apache-2.0 (both repos). Cost is GPU compute plus engineering time, and optionally NVIDIA AI Enterprise licensing for support (pricing not reviewed).

Open questions

  • Exact close date and price (NVIDIA never confirmed). Is there an NVIDIA 10-K/10-Q reference?
  • Is any hosted, Gretel-equivalent service sold by NVIDIA today, and what support terms apply to Safe Synthesizer under NVIDIA AI Enterprise?
  • What happened to Gretel’s SaaS customers and contracts at shutdown (dates of console sunset)?
  • MCP support in Data Designer or the NeMo Platform.
  • Any FS users of Safe Synthesizer.

Sources

History

  • [2026-10-26] Page created via wiki-create and researched the same day, as a seed vendor of the new synthetic-masked-data category. Slug kept as gretel (company name buyers recognise) with owner NVIDIA, title “Gretel (NVIDIA)“. Acquisition dated to the 2025-03-19 press reports; NVIDIA never announced it.

Graph View

Backlinks