FUNDAMENTOS
Nvidia, Google and OpenAI all talk about synthetic data. What each source says, what is announcement and what is result, and what it means for vision.
"Big tech already uses synthetic data" is an easy sentence to repeat. This post looks at three public sources, one from Nvidia, one from Google and one from OpenAI, and separates what each says from what is often attributed to it. The three deal with different things, and only one of them reports results with numbers.
What each source says
- Nvidia (CES 2025, January 6, 2025). In the company’s blog post about the keynote, Jensen Huang talks about synthetic data for autonomous vehicles: Omniverse is used to test and generate synthetic data, and a “data factory” built with Omniverse and Cosmos can turn “hundreds of drives into billions of effective miles.” He also mentions generating millions of synthetic motions to train humanoid robots. It is a manufacturer that sells the platform speaking at a product event.
- Google Research (April 16, 2026). The post presents Simula, a method for generating synthetic datasets that controls coverage, complexity and quality separately. The tests are on text (cybersecurity, legal reasoning, math, multilingual knowledge), in a distillation setup: the teacher model is Gemini 2.5 Flash, the student is Gemma-3 4B, with datasets of up to 512K examples. Raising complexity gave a 10% accuracy gain on GSM8k and hurt the result on LEXam, where the teacher was weaker. The authors say there is no single recipe for all domains. The argument about generating edge cases to test systems before they happen is a motivation in the text, with no safety experiment in the post.
- OpenAI (cookbook, April 10, 2024). It is a tutorial, now marked archived, on how to use a model to generate tabular and text synthetic data, including ideas for sparse or imbalanced data. The page does not claim that OpenAI trains its own models on synthetic data.
What this shows, and what it doesn’t
None of the three sources demonstrates that a big-tech vision model was trained on synthetic data and beat a baseline. Nvidia describes a platform, Google shows text results, and OpenAI teaches a technique. For computer vision, the closest evidence comes from papers with numbers, such as Microsoft’s DAViD (depth, normals and masks for people, trained on 300K synthetic images) and SynBalance, from researchers at AWS AI Labs and the University of Washington (generated images for rare classes). Both are specific results on specific benchmarks, and we have analyzed each here on the blog.
The limit of this reading
- Sources of different nature. A keynote blog, a research blog and a tutorial don’t carry the same weight. This post did not use press reports about these companies’ internal use of synthetic data, because we could not confirm them against a primary source.
- Text versus vision. Google’s example is language, and Nvidia’s is driving and robotics. What holds for one doesn’t automatically hold for the other.
- The sources’ dates. The Google source is later than this post’s original publication (February 2025). The landscape moves fast, and the list here is a sample, not a complete survey.
What changes with a controlled scene
This section is our reading, not the sources’. What the three have in common is data generated under control: simulated driving scenes in Nvidia’s case, coverage and complexity parameters in Google’s, instructions about what to generate in OpenAI’s. In vision, the controlled scene lets you declare the conditions (light, material, camera, class proportions), get the label along with the pixel, and reproduce each sample. What it doesn’t do is replace validation on real data, which remains the test for any of these methods.
What to take to your project
- Separate announcement from result. For each claim, ask whether there is a benchmark, a metric and a baseline behind it.
- Check that the evidence is for your kind of data. A result in text or autonomous driving doesn’t answer your vision task on its own.
- Treat the real/synthetic mix as a hypothesis to measure, with a test on real data from your case.
Want to understand where synthetic data fits in your pipeline? Tell us about it in a diagnostic call. If synthetic data isn’t the answer, we’ll say so.
Sources
- Nvidia. CES 2025: AI Advancing at ‘Incredible Pace’. Nvidia blog, January 6, 2025.
- Davidson, T. R., Harkous, H. Designing synthetic datasets for the real world. Google Research, April 16, 2026.
- Almeida, D. R. Synthetic data generation (Part 1). OpenAI Cookbook, April 10, 2024 (archived).
Leave a Reply