Runtime, Data, Ecosystem Decide Robotics

This market is likely to be won by the company that becomes the default runtime for real robots in production, not the company with the flashiest demo. In robotics, better data comes from more hours on live machines, faster inference means fewer pauses and safer control, and ecosystem adoption determines whether OEMs, integrators, and developers keep feeding tasks, tools, and training data back into one shared stack.

• Training data quality matters because robotics data is narrow and physical. Covariant has years of warehouse pick trajectories and reports 100 plus robotic arms in the field, while Skild is building a cross robot data flywheel from deployments in security, construction, delivery, and warehousing. The model that sees the most real failures and recoveries learns the fastest.
  1 sacra 2 sacra 3 sacra
• Inference speed matters because robots cannot wait around for a large model to think. Physical Intelligence describes a runtime that predicts short action chunks in about 100 milliseconds so motion can continue while the next commands are computed. That kind of latency directly affects whether a robot feels smooth and usable on a factory floor.
  3 sacra
• Ecosystem adoption can outweigh model quality because platform distribution compounds. NVIDIA already offers Isaac, simulation tools, Jetson hardware, and a broad robotics partner network, and Skild is one of the companies building on that stack. Google has also folded Intrinsic more tightly into Google, pairing robotics software with Gemini and Cloud distribution.
  4 nvidia 5 intrinsic

The next phase looks less like a pure model race and more like a platform land grab. Foundation models will keep improving across the field, so durable advantage will come from owning deployment loops, latency tuned infrastructure, and developer workflows. That pushes the category toward a small number of operating layers that sit between robot hardware and every real world task.