AMI's Open-Core for Physical AI
AMI Labs
This business model says the product is unlikely to be the raw model alone, it is the packaged system that makes a risky model usable inside a real hospital, factory, or robot fleet. AMI can give away enough research and base capability to become the default starting point for developers, then charge when an enterprise needs the hard parts, on premises deployment, adaptation to its own sensor data, auditability, guardrails, uptime, and engineering support.
-
The product design already fits this split. One tier is a hosted API for buyers that want managed access. The other is a downloadable version for regulated or air gapped environments. That is the same playbook open core infrastructure companies use, free adoption at the edge, paid control and reliability in production.
-
This matters more in AMI's market than in chatbots, because physical world AI has to be tuned to a specific environment. A hospital monitor, robot arm, or factory line each produces different sensor streams. The money is in adapting the base model to that local reality, then wrapping it with safety and monitoring so a customer can trust it.
-
The pressure is that open research can commoditize the base layer fast. Meta has already published V-JEPA 2 research built on over 1 million hours of video, and Physical Intelligence open sourced π0 weights. That pushes AMI toward monetizing the surrounding enterprise stack faster than the model itself.
The likely end state is a market where base world models spread broadly, while the winning companies own the deployment layer inside specific verticals. If AMI turns early partnerships like Nabla into repeatable templates, it can evolve from a research lab into the control plane for safety critical physical AI.