Platform Control of Robot Intelligence
Generalist
The real threat is distribution, not just model quality. NVIDIA and Google DeepMind can turn robot intelligence into infrastructure, which lets them spread a baseline stack across many builders at once while companies like Generalist still have to win each customer, collect each dataset, and assemble more of the system themselves. Amazon shows the same pattern in logistics, where the biggest robot buyer is pulling frontier models in house and training against its own fleet and workflows.
-
NVIDIA is collapsing layers of the stack into one package. Its new open reference humanoid combines a robot body, dexterous hands, Jetson Thor compute, and Isaac GR00T software and models, while GR00T N2 is ranked first on generalist robot policy benchmarks. That makes more of the humanoid baseline cheap and reusable.
-
Google DeepMind can ship embodied intelligence through existing cloud channels. Gemini Robotics adds physical action to the Gemini model family, Gemini Robotics-ER is available through Google AI Studio and the Gemini API, and the system is already designed to transfer across different robot forms. That reduces the need for buyers to bet on a single startup platform.
-
Amazon highlights what competing from above looks like in a vertical market. It hired key Covariant researchers, licensed Covariant's models, has deployed 1 million robots, and is building new agentic AI capabilities for its fleet. When the operator owns the robots, data, and deployment sites, an outside robotics AI startup has much less room to wedge in.
The market is moving toward a split. Startups will still matter where they own unique workflows, customer relationships, or hard to gather real world data, but the default intelligence layer is getting standardized by platform companies. That will push Generalist to differentiate through proprietary data loops, deployment speed, and tight integration with specific robot use cases.