Installed Base as Humanoid Moat

The real moat in humanoids is not the model, it is the installed base that keeps generating failure cases from live work. Indoor manipulation data cannot be borrowed from self driving or the internet alone, so every company has to start with a narrow job, watch where the robot gets stuck, use teleoperation to recover the task, and turn those interventions into training data. That makes deployment speed and customer access more important than lab demos.

• Tesla has an enormous road dataset, but that does not solve indoor autonomy. Factory tasks involve hands, shelves, bins, tools, tight turns, and cluttered spaces, so the useful data only appears when robots are actually working inside those environments.
  2 sacra 4 sacra
• Teleoperation is not just a safety backup. It is the labeling machine. When a robot mispredicts and a human steps in, that correction shows exactly what went wrong and what the right action looked like, which is why early commercial deployments matter so much.
  1 sacra 2 sacra
• This favors companies with credible deployment channels over companies with only capital. Apptronik has Google as its exclusive Gemini Robotics humanoid partner and pilots with Mercedes-Benz and GXO, while Figure has BMW plant runtime. Those relationships create the real world reps that compound into better models.
  3 sacra 4 sacra 5 sacra

The next phase of competition is a race to lock in repetitive industrial workflows before anyone reaches broad generality. The winners will be the teams that place robots into real shifts, accumulate intervention data across many sites, and expand from one proven task into a larger fleet and a wider action model.