Skild AI Crosses a $100M Run Rate: A Robot Brain That Learns From One Video, and Customers Up From 8 to 60 This Year
Robotics finally has a number that didn't come from a pitch deck: $100 million.
Three key facts
Ten months after its first commercial deployment, Skild AI has reached a $100 million annualized recurring revenue run rate. Bloomberg reported the milestone on September 10. For scale, the company booked roughly $30 million in revenue for all of 2025. It had eight customers at the start of this year and now has more than 60, with hundreds of robots running its software across manufacturing, logistics, inspection, security and food preparation.
It doesn't sell robots. It sells their brains. The core product, Skild Brain, runs on the company's S1 model, which uses in-context learning to pick up a new task from a single human video demonstration, with no need to gather thousands of examples and retrain weights for every job, and it can handle action sequences longer than ten minutes. Skild's own figures: 96% success on tasks it has seen, 66% on tasks it hasn't, and one demo video worth about 380 conventional training examples.
The money arrived well before the revenue. SoftBank led a $1.4 billion Series C in January at a valuation above $14 billion, bringing total funding past $2 billion, and in April Skild bought Zebra Technologies' robotics automation business. Meanwhile, the EU's new Machinery Regulation takes effect on January 20, 2027, bringing AI safety functions and "self-evolving" machinery under regulation for the first time. A model that performs tasks it has never seen falls squarely inside that definition.
WangDou's Take
First, cool the number down. A run rate is your best recent month times twelve, not cash in the bank. And $14 billion divided by $100 million is 140 times. Software companies look pricey at fifteen. This valuation is buying the future, not the present.
Now the flashiest selling point: 66% success on unseen tasks. In a demo video, that's a miracle. On a production line, that's botching one job in three. A ruined pancake gets flipped again; a misdriven screw on an assembly line means recalling the whole batch. So what customers are really paying for is probably the 96%, reliable work on familiar tasks, while "learns from one video" is the key that gets the sales team through the door.
Hundreds of robots supporting a $100 million run rate works out, roughly, to somewhere between a hundred thousand and several hundred thousand dollars per robot per year. That's proof customers will pay real money for robots that do real work. But Europe's 2027 rules are already on the way. When a robot that teaches itself new moves causes an accident, who's liable: the person who filmed the demo, the company that trained the model, or the factory that bought the machine? That's a much harder problem than flipping pancakes.
Source: Bloomberg · The Next Web · Securities.io