Shared Base Models, Costly Adaptation

The key business question is not whether one robot model can do many demos, but whether it can cut the repeated labor of building each new deployment. Physical Intelligence is trying to make the base model and hosted training stack a shared R&D asset across many robot types and tasks, while still charging for the expensive part that does not disappear, customer adaptation, inference, and hands on deployment work needed to get a robot from 80% success to production reliability.1 2 3 4

• The workflow already looks more like cloud AI infrastructure than packaged software. Developers upload robot data, run fine tuning on managed compute, then call the adapted model through a hosted gateway. That means usage scales with training jobs and inference volume, not with simple seat licenses, so compute remains a direct delivery cost.1 2
• Generalization still leaves a large last mile. In robotics, the value often sits with the company that integrates the model with a specific workstation, robot body, camera stack, and customer process. Anvil describes this as taking a general brain and pushing it from good enough to near perfect inside one workflow, which is where customer data collection and support concentrate.3
• There is precedent for selling robotics as an operating service rather than a pure software subscription. Locus Robotics turned warehouse picking into a recurring robots as a service model with roughly $2,000 per robot per month pricing, bundling hardware, software, and ongoing operation. That is a useful analogue for why physical AI may monetize through deployment economics, not classic SaaS margins.4

If Physical Intelligence keeps improving cross task transfer, more of each new customer rollout will look like parameter tuning instead of greenfield engineering. That would widen gross margin at the model layer, but the winners in physical AI will still be the ones that turn shared intelligence into repeatable, paid operations on real robots in real customer environments.1 3 4