**NVIDIA releases Cosmos 3 open physical AI foundation model with hybrid transformer architecture.**
The AMW Read
New product release from a dominant compute player entering a new vertical as a model provider; updates robotics player map and structural compute dynamics.
**NVIDIA releases Cosmos 3 open physical AI foundation model with hybrid transformer architecture.**
NVIDIA has released Cosmos 3, an open physical AI foundation model featuring a hybrid transformer architecture. The model is designed to advance physical AI — the application of AI to understand and interact with the physical world — by enabling better reasoning about motion, spatial relationships, and real-world dynamics.
This release matters because it extends NVIDIA's platform strategy deeper into robotics and embodied AI. By open-weighting a foundation model purpose-built for physical world understanding, NVIDIA creates a new substrate layer that robotics developers, simulation engineers, and autonomous system builders can build upon. This mirrors the hyperscaler-distribution pattern seen in general-purpose foundation models but applied to the physical AI vertical. It also signals NVIDIA's intent to claim the model-level moat in the robotics segment, complementing its existing hardware and simulation (Isaac, Omniverse) assets.
NVIDIA's move updates the Robotics/Physical AI player map: it positions the company not just as the silicon and simulator provider, but as a foundational model provider for physical intelligence. The hybrid transformer architecture suggests a deliberate design to balance real-time inference latency with multimodal understanding — a key pain point in production robotics. While open-weight availability lowers the barrier for experimentation, enterprise adoption will depend on how well Cosmos 3 integrates with NVIDIA's existing toolchain and whether it can outperform specialized robotics models from players like Covariant, Physical Intelligence, and others.
