AMI's Research-First Slows Data

The key tradeoff is that AMI is optimizing first for model quality, while rivals like Skild are already optimizing for learning speed in the field. AMI is still training world models and proving them with partners, while Skild is already selling robot intelligence into live deployments and World Labs already has a public product surface. In physical AI, every shipped deployment is both revenue and fresh sensor data, so time spent in research can deepen the model but slows the compounding loop that often defines category leaders.

• AMI is still pre production as of March 2026. Its planned workflow is to ingest sensor streams, adapt the model to a specific environment, simulate candidate actions, then choose a safe action. That can produce strong science around prediction and control, but it means AMI is not yet harvesting broad field data at commercial scale.
  1 sacra
• Skild already has the opposite motion. Its Brain is sold as a software layer for many robot types, and each deployment feeds performance data back into the base model. The company said in January 2026 that live revenue reached about $30M within months, across security, construction, delivery, data centers, warehouses, and factory assembly.
  2 sacra 4 skild
• World Labs shows why product surface matters early. Marble is already public, with docs, pricing, billing, and API workflows for generating and exporting 3D worlds. That gives developers something concrete to build around now, which helps define the category before AMI has a comparable external product.
  1 sacra 3 worldlabs

The next phase of competition is likely to reward companies that can turn world models into repeatable deployment loops. AMI can still win by using healthcare and industrial partners to create high value, safety critical data channels, then layering planning, simulation, and controllability software on top. But the market is already shifting toward whoever can combine research depth with real usage data fastest.