Selling Pretrained Intent Models

The core bet is that the highest value layer in physical AI may be the model that learns why an action was taken and what happened next, not the robot itself. General Intuition is trying to sell that layer like software. Customers bring their own robot, simulator, or game, then pay for pretrained world and action models that reduce how much task specific data they need to collect and label before deployment.

• The practical edge comes from Medal data. Gameplay clips include the actual button presses behind each movement, so training data is not just video of an outcome, it is a record of intent, action, and consequence. That is much closer to robot control data than scraped internet video.
  1 sacra 2 medal
• This makes the business look more like an API vendor than a robot maker. A robotics team can use the models to pretrain in a learned simulator, fine tune on a small amount of real world data, then deploy on its own hardware. General Intuition keeps the model revenue without carrying fleet, manufacturing, or field service costs.
  1 sacra
• The main comparison is with Physical Intelligence, Skild AI, Figure, and NVIDIA. Horizontal labs also sell general control intelligence, but vertical players like Figure capture deployment data directly from robots in the field. NVIDIA is pushing open world model infrastructure through Cosmos and Omniverse, which can compress margins at the standalone model layer.
  1 sacra 3 coreweave 4 nvidia

If this approach keeps cutting the amount of real world data needed to reach useful robot behavior, the model layer can become a recurring toll booth across games, simulation, and robotics. The next phase is a race to turn pretraining advantage into default infrastructure before vertical fleets and open platforms absorb that value.