Config Vulnerable to Model Substitution

This points to Config’s real risk, that its highest value product can shrink from a full robot learning stack into labor and data prep work if customers treat outside models as the brain and use Config mainly to feed them task specific examples. Loop already supports NVIDIA GR00T and Physical Intelligence models, so a capable buyer can keep model choice in house and only pay Config for collecting demonstrations, cleaning them, and converting them into training ready datasets.1 2 3 4

• Config is more vertically integrated than data only peers like Lightwheel and Maniformer because it sells both Loop workflow software and its own CFG-1 manipulation model. That creates more revenue per customer when Config owns the whole workflow, but it also makes model substitution a direct threat when buyers prefer another foundation model.5 6 7
• Open model supply is getting stronger. NVIDIA positions GR00T as an open, end to end stack with foundation models, simulation, and deployment workflows, and Physical Intelligence describes pi0.5 as a broadly generalizable vision, language, action model trained across heterogeneous robot tasks. That means a well staffed robotics team can source the model elsewhere and spend its budget on proprietary task data instead.2 3 8 9
• In practice, a smaller contract would look like a factory or robot maker hiring Config to film human demonstrations, label edge cases, or convert raw trajectories into a format that works with the buyer’s chosen model. The recurring software and model revenue then compresses into project based services, which usually carry lower margins and weaker lock in.1 5 6

The path forward is for Config to make Loop valuable even when CFG-1 is not the chosen model. If it becomes the easiest way to capture, evaluate, and deploy manipulation data across GR00T, pi0.5, and future models, it can stay in the stack as the system of record for robot learning rather than being pushed down to a one time data vendor.1 2 3 4 8 9