Proprietary Training Data Flywheels

The core risk for Skild is that the best robotics data is being created inside deployments that model vendors do not control. Figure, Tesla, and Apptronik are not just training models, they are putting whole robots into factories, watching every pick, carry, pause, and failure, then feeding that data back into their own stack. That makes the robot, the software, and the customer workflow improve together, which is a stronger loop than licensing a general brain to outside hardware makers.1 2 3 4 5

• In practice, this flywheel starts with narrow work cells. A humanoid loads parts, moves bins, or handles assembly steps. When it fails, teleoperation or human intervention keeps the line moving while also labeling the exact edge case that retrains the system. The deployment itself becomes the dataset factory.2 3 5
• Horizontal model companies still have a role, but it looks more like shared infrastructure than category ownership. The market is already splitting into foundation model providers, hardware platforms, and solution companies that win by owning customer deployments and collecting task specific data inside warehouses, factories, and logistics sites.2 4 5
• The capital gap reinforces the data gap. Figure sits at a $39B valuation, Apptronik at $5.5B, and Physical Intelligence at $5.6B, while Skild is at $14B. That funding buys fleets, sensors, compute, and customer access, which matters because physical AI still depends on expensive real world data collection rather than cheap internet scale pretraining alone.1 3 4 5

Going forward, the winners are likely to be the companies that turn commercial robot uptime into proprietary training supply. If vertical players keep scaling fleets inside auto, logistics, and factory environments, horizontal providers like Skild will be pushed toward supplying models for uncovered niches, cross platform tooling, or partners that cannot justify building their own closed loop stack.1 2 3 4 5