Carbon as Interchangeable Integration Layer
Sanctuary AI
If general robot brains become easy to license, Carbon stops being the thing customers are buying and starts being the wiring used to connect someone else’s model to Sanctuary’s robot. That matters because Carbon sits between Phoenix hardware and customer workflows today, but Google DeepMind and NVIDIA are both pushing multi robot model stacks that can be adapted across different bodies, which would shift bargaining power toward the model owner or the robot OEM with direct access to that model.
-
Sanctuary is small relative to the companies trying to define the common robotics software layer. Sanctuary is tracked at about $149M in total funding, versus roughly $1.75B for Figure, $1.0B for Apptronik, $641M for Agility, $1.07B for Physical Intelligence, and a $14B valuation for Skild AI. That gap affects compute, data collection, and deployment scale.
-
The threat is not just better software, it is software that travels. DeepMind introduced Gemini Robotics in March 2025 as a Gemini 2.0 based robotics model designed to adapt to different robot types, then launched an on device version in June 2025 and a developer API for Gemini Robotics-ER 1.5 in September 2025. That is the path from bespoke stack to licensable utility.
-
NVIDIA is doing the same at the tooling layer. Isaac GR00T bundles open models, data pipelines, simulation, middleware, runtime libraries, and onboard compute, and NVIDIA has positioned it as a full development stack that teams can adopt in pieces. Once OEMs can assemble that stack directly, an intermediate control layer becomes easier to swap out.
The next phase favors companies that either own unique robot data and customer workflows, or own the full stack from model to machine. For Sanctuary, Carbon becomes more defensible as a tightly integrated operating system for Phoenix in live jobs, and less defensible if the market standardizes around broadly available foundation models that every humanoid maker can plug in.