Robotics AI startup Skild AI has surpassed $100 million in annual recurring revenue (ARR) 10 months after commercially deploying a foundation model for robots.

Deepak Pathak (딥락 파택), CEO of Skild AI, posted the news on social media platform X (Twitter) and shared the company’s performance and philosophy.

Skild AI has secured more than 60 paid customers over the past 10 months in areas including logistics movement, delivery, on-site inspection, security, cooking, and warehouse, factory and data center operations, he said. Mobility accounts for about 10 percent of revenue.

Skild AI is working with Nvidia and Foxconn to apply its foundation model, Skild Brain, to dual-arm robots and deploy them for precision assembly work on Nvidia Blackwell systems. With Sumitomo Wiring Systems, it is pursuing automation of wire harness manufacturing processes that had been considered impossible to automate. With Mitsui & Co, it is piloting general-purpose robots running its latest model, S1, in a food service supply chain that provides 1.4 million meals a day.

Pathak stressed that from the start he saw deployment as a core element of the technology, not an output. He said robotics AI is about installing robots at real customer sites and making them run.

He also highlighted that reality differs from the common belief in robotics that making a superintelligent model and building superhuman hardware will suddenly make robots appear everywhere.

"In robotics, you cannot put off field deployment. How to build the supply chain, who will install it, who will handle integration, who will fix it when it breaks, who will update it when the process changes. You only find answers to these questions by going to the field yourself," he said.

He also made clear that demos alone make it hard to gauge capability. "Whether it is 5 percent accuracy or 99 percent, one successful scene can look the same. When we trained it to cook eggs last year, it took a week to cook the first egg, but it took another two months to make it stable across different eggs and environments," he said. "The effort to raise the last 5 percent of performance is far greater than the effort to build the first 95 percent."

He also pointed to speed. "Even at 99.9 percent accuracy, if it is 10 times slower, you cannot deploy it. If robots handle half of 10 process stations on a line, if even one station is slow, the overall throughput is set to that robot," he said.

He also said change is constant, such as suppliers swapping parts or factories rearranging workbenches. He said it is not sustainable if new data must be collected and the model retrained each time.

Skild AI’s latest model, S1, built on Nvidia AI infrastructure, is a result of that thinking, he said. S1 can perform a new task immediately without retraining if a worker shows it a video demonstrating the motion.

He also said demo-driven and deployment-driven cultures are difficult to coexist within a single company.

"Even with a robot that is 50 percent accurate, you can pick out one good scene. To deploy it, you have to handle the other 50 percent, and that is much harder," he said. "We decided to reward deployment."

He also introduced the idea of Physical Recursive Self-Improvement, or physical RSI, using deployment data for iterative improvement. He said that when a base model trained on multiple robots and tasks is deployed at each site, it adapts to new tasks through in-context learning, and the accumulated data is then brought back to the base model so the next deployment starts from a better position. "The era of demos is over. The era of deployment has begun," Pathak stressed again.