From Foundation Models to Operating Systems

The center of gravity in robotics is moving from inventing a unique base model to owning the messier system around it. Google DeepMind is pushing robot intelligence through APIs and cross embodiment control, NVIDIA is turning simulation and synthetic data into shared infrastructure, and Amazon has already absorbed Covariant talent and licensed its models. That makes standalone model IP less rare, and shifts value toward field data, safety layers, integrations, and deployed customer workflows.

• Physical Intelligence and Skild show how horizontal robot brains are becoming easier to replicate. Physical Intelligence open sourced π0 weights and lets developers fine tune with small amounts of robot data, while Skild sells a cloud API that maps many robot types into one control layer.
  7 sacra 8 sacra
• NVIDIA is lowering the cost of catching up. Cosmos ships open world models, synthetic data engines, and Isaac and Omniverse based simulation, so a rival does not need Covariant scale real world data on day one to train and test manipulation policies.
  2 deepmind 3 nvidia 9 nvidia
• Amazon now has the strongest version of the warehouse data flywheel that Covariant once sold into. It hired Covariant co founders and about a quarter of the company, licensed the technology, and can improve those models using one of the world’s largest robot fleets inside live fulfillment operations.
  1 aboutamazon 4 sacra 5 sacra

From here, the winners are likely to look less like pure model labs and more like operating systems for real deployments. The durable advantage will come from who can turn every robot task, failure, and recovery into better software, faster validation, and easier rollout across customer sites at industrial scale.