Open Base Models Expose Skild

Open weights push value away from the base robot model and into everything wrapped around it. Physical Intelligence is using outside capital to spread a hardware agnostic control model across many robot types, while Skild is trying to sell a similar horizontal brain through cloud APIs. If the core policy becomes widely available, the scarce piece is no longer raw model access, it is deployment data, safety layers, tuning tools, and channel access to OEMs and integrators.1 2 3 4 5

• Physical Intelligence raised a $400M Series A in November 2024, then open sourced pi0 model weights after showing a 3.3B parameter vision language action model trained on open datasets plus data from 8 robot types. That makes its go to market look more like Android for robots than a closed premium API.2 3 4
• Skild and Physical Intelligence both pitch hardware agnostic robot intelligence, but Skilds workflow runs through Skild Cloud, where customers upload robot specs and call high level APIs. That leaves Skild more exposed if base policies become cheap, because customers can swap in open models and spend budget on integration, safety, and vertical apps instead.1 2 5
• Covariant shows what durable differentiation looks like after model novelty fades. Its edge comes from real warehouse picking data, workcell integrations with ABB, KNAPP, and Bastian, and proof in production workflows like induction and putwalls. In robotics, the winner is often the stack that survives floor reality, not the lab model with the best demo.5 6 7

The market is moving toward a split structure. Open and semi open base models will set a common starting point, while the winning companies build the operating layer on top, the adapters for each robot, the safety envelope, the simulation and fine tuning loop, and the distribution network that gets software onto real machines at scale. That is where Skild will need to deepen its moat.1 2 5