Model Value Shifts to Deployment

The biggest value in physical AI is likely to sit closest to the data loop and deployment surface, not in a model API by itself. A standalone model company can win if its pretraining is so useful that robot builders and game studios plug it in everywhere, but once a company owns the robot, the sensors, the customer workflow, and the stream of real failures from production, it usually gets the best feedback for improving performance and the most leverage over pricing.

• General Intuition is making the strongest horizontal case. It sells world and action models through an API and feeds them with Medal gameplay data, billions of action labeled clips that teach what button press caused what outcome. That is valuable training fuel, but it still sits one layer above the robot operator, OEM, or simulator platform that owns live deployment data.
  1 sacra
• Skild AI and Physical Intelligence show the middle path. They aim to be robot brains for many embodiments, but both are pulling closer to deployment by mapping robot hardware into their stack, fine tuning on partner robots, and turning customer usage back into shared training data. That makes them less like pure model vendors and more like operating systems with attached data flywheels.
  2 sacra 6 sacra 7 sacra
• Figure and NVIDIA represent the two clearest ways the standalone layer gets squeezed. Figure keeps model, hardware, and field data in one loop with Helix running on its humanoids. NVIDIA pushes the opposite direction, offering world models, containers, simulation, and Omniverse tooling so developers can build on NVIDIA controlled infrastructure instead of any single startup API.
  3 sacra 4 figure 5 nvidia

The market is heading toward a stack where model companies have to climb down into tooling, simulation, and data collection, or risk becoming interchangeable. The winners will look less like a raw model endpoint and more like a control point in the workflow, where every deployment generates proprietary data and every new robot or developer makes the platform harder to replace.