Synthetic data stopped being a research topic some time in 2025 and became plumbing. It now sits in the post-training of open language models, in the world models built for robots and vehicles, and in the procurement plans of data teams. The interesting question in 2026 is no longer whether to use it. It is what the synthetic data is generated from, and what it is checked against. The answer to those two questions decides whether it adds information to a model or quietly takes it away.
The year it became infrastructure
Three moves in 2025 show how far the practice has travelled.
In March 2025, NVIDIA bought Gretel, a San Diego company whose platform generates synthetic tabular and text data for training, in a deal reported at a price above Gretel's last valuation of $320 million. A GPU company buying a data generator says where the bottleneck is thought to be.
NVIDIA also opened a post-training set for its Llama Nemotron reasoning models that holds about 30 million synthetic training examples across mathematics, code, reasoning, function calling and instruction following, every response generated by public and open models. Post-training on generated data, once a careful trick, is now something a lab publishes as a dataset card.
At CES 2025 the same company released Cosmos, a family of world foundation models that generate physics-aware video for robotics and driving, trained on 20 million hours of real-world video of human interaction, industry, robotics and driving. Here the synthetic data is the output of a model that has itself learned from a very large body of recordings, and its purpose is to cover situations those recordings do not.
Analysts have moved with it, and with some caution. In June 2025 Gartner predicted that by 2027, 60% of data and analytics leaders will face critical failures in managing synthetic data, with the risk landing on AI governance, model accuracy and compliance. That is a forecast about management, but it points at something technical: generated data is easy to make in volume and hard to account for.
Where it breaks: the tails go first
The failure everyone now names is model collapse. A model trained on what models made loses the unusual parts of the real distribution first, the rare event and the long tail, and each round of generation narrows what is left.
Two results from 2025 sharpen the picture, and they pull in usefully different directions.
The first is Strong Model Collapse by Dohmatob, Feng, Subramonian and Kempe at ICLR 2025. In their theory and experiments, even a very small fraction of synthetic data in the training set, as little as one part in a thousand, can still produce collapse, and the harm does not fade as long as the synthetic share does not shrink to zero as the dataset grows. Simple fixes such as down-weighting the synthetic portion do not, in general, prevent it. Larger models can soften the effect beyond a certain point but do not remove it.
The second is Collapse or Thrive? by Kazdan, Schaeffer and colleagues at ICML 2025. They compared workflows in which each generation of synthetic data replaces the real data with workflows in which it accumulates alongside the real data. Replacement degrades test loss steadily. Accumulation contains it. And the value of synthetic data depends on how much real data there is: when real data is scarce, a small amount of synthetic data improves test loss; when real data is plentiful, adding synthetic data only makes it worse.
Read together, the two papers say something precise. Synthetic data generated from a model of the data carries that model's blind spots, and a little of it is enough to carry them in. It earns its place where the real data is thin, and only while the real data stays in the mix.
Generated from what?
That is why the useful distinction in 2026 is the source of the synthetic data.
One source is a learned model of the data itself: a language model writing more text, an image model drawing more images. It produces data that looks like what it saw. Its strength is breadth. Its weakness is exactly the one collapse describes, because what it knows about the tails is only what the training set happened to contain.
The other source is a representation of the thing the data is about: the room, the fabric, the body, the feeder, held with its measured properties and run forward under known physics or known rules. Data generated this way can reach situations nobody recorded, such as a material in a light it was never photographed in or a load a part has not yet carried, because the behaviour comes from the representation, not from the frequency of past examples. It also has a clear limit. It holds only inside the range its measurements support, and the honest thing to do is to say where that range ends.
Cosmos sits between the two: a learned model, trained on recordings of the physical world, generating new video of it. Most of the practical progress in physical AI in the last two years has come from making the representation side cheaper, through better capture, open formats for scenes and materials, and GPU simulation fast enough to run inside a training loop.
What checking looks like
Whatever the source, synthetic data needs a test it did not write. The collapse results are, in the end, results about evaluation: a model can look better on data like its own while getting worse on the world. The discipline that answers this is old and simple. Fix the question and the bar before looking. Hold out real measurements no model has seen. Score once. Publish the misses as well as the hits, so the boundary of what works is visible.
It also changes what counts as a good result. The useful claim is narrow and measurable: with this much real data, generated data from this source reduced error by this much, on measurements held out from everything, and here is the point at which it stopped helping.
Where Livgens fits
Livgens starts from the representation. It represents the thing exactly, in sound, light, matter and motion, held as the thing itself: how a room carries sound, how a fabric falls, how a part takes a load, how light lands on a street. Every thing carries its measurements and the range they hold for, so a simulation stays inside what is known. Living Worlds runs that representation forward through its lifecycle, and models train and test on what comes out.
The checking follows the discipline above. Each result is registered before its data is read and scored once on measurements no model has seen, and the misses are published too. Livgens puts its own position in one line: simulation makes a little real measurement go much further, and the results say exactly how far. A registered study of unmetered power feeders shows the shape of it: learning from a checked simulation first helped most when real measurements were fewest, the advantage was gone by eighty, and a deliberately wrong simulation did worse than either. That is the collapse literature's lesson put to work.
Health is live today; Looper, Render, Twin and Programmer are in early access.
Sources
- Nvidia reportedly acquires synthetic data startup Gretel (SiliconANGLE, March 2025)
- Build custom reasoning models with advanced, open post-training datasets (NVIDIA Technical Blog, 2025)
- Nvidia releases its own brand of world models (TechCrunch, January 2025)
- Gartner announces the top data and analytics predictions (June 2025)
- Dohmatob et al., Strong Model Collapse (ICLR 2025)
- Kazdan et al., Collapse or Thrive? Perils and Promises of Synthetic Data in a Self-Generating World (ICML 2025)
