"Data stops being something you collect. It becomes something you compute." At #SIGGRAPH2026, Ming-Yu Liu, VP of NVIDIA's Cosmos Lab, laid out why world models are the data engine of physical AI. Plus, he announced Cosmos-Dreams, a neural closed-loop simulator, and showed off a full demo of it in... action. You can watch the full keynote here:show more

@liu_mingyu This is how robots scale faster

@liu_mingyu Yes

@liu_mingyu 'Data you compute' inherits every blind spot of the model computing it. Closed-loop sim is the right architecture, but Cosmos-Dreams shifts the sim2real gap from collection cost to model bias — cheaper to scale, much harder to audit.

Shifting from passive data scraping to active synthetic data generation through closed-loop simulators like Cosmos-Dreams solves the physical AI wall ⚙️. ​But how do we guarantee zero sim-to-real gap when training physical robotics agents in neural environments? What’s your take? 👇

@liu_mingyu Nvidia down 11% in 5 days. Jensen should speak some more and post more. Super helpful 🤡 Dude is a clown on steroids. @JensenHuang @nvidia

Closed-loop is the load-bearing word in that sentence. Cosmos-Dreams generates what the sensors would see in response to the policy's own actions, so a wrong decision changes the next frame - which replaying logged drives can never do. NVIDIA says it builds a whole world from a single frame and runs on one RTX PRO 6000. That last part is what decides whether it's a keynote demo or something a team iterates on daily.

@liu_mingyu simulating a closed-loop universe just to train a robot feels like sci-fi becoming an engineering spec

@liu_mingyu Cosmos is nvidia's data moat. data stops collecting, starts computing. physical ai wins on inference, not training

@liu_mingyu The next leap in physical AI won't just come from more data. It'll come from better world models.

@liu_mingyu World models as the data engine is the right bet. The bottleneck in physical AI has always been data coverage, not model architecture.