Skild AI says it passed $100 million in annual recurring revenue as its S1 model learns tasks from one video

Skild AI has put a nine-figure revenue run rate on the record, a disclosure none of its best-funded humanoid and robot-brain peers has matched in their own announcements. The company said on September 10, 2026 that it crossed $100 million in annual recurring revenue ten months after its first commercial deployment, and that $50 million has already been recognised as revenue over that period.

Skild said it now has more than 60 paying customers, using its robots to move goods, make deliveries, inspect sites, provide security and prepare food, and to work inside warehouses, factories and data centres. Mobility accounts for 10% of revenue, including autonomous mobile robot (AMR) solutions that make up 4%. The remaining roughly 90% comes mainly from manipulation tasks. The company did not disclose the number of robots in service, gross margin or profitability.

Three named deployments anchor the claim. With NVIDIA and Foxconn, Skild said it is deploying its software, the Skild Brain, on dual-arm manipulators for high-precision assembly of NVIDIA Blackwell systems. At Sumitomo Wiring Systems it is working towards deploying S1 in wire-harness manufacturing. With Mitsui & Co., whose supply chains serve 1.4 million meals a day across Japan according to Skild, it is piloting general-purpose robots in commercial kitchens.

The revenue update followed the launch of S1 on August 18. Skild describes S1 as an in-context learning model: instead of being retrained for each job, it watches one video of the task and then performs it. The company said the model handles previously unseen tasks lasting up to ten minutes, citing pancake flipping, pour-over coffee, plant repotting and kit assembly. In one example, eleven minutes passed between recording a plant-potting demonstration and the robot doing the job autonomously.

Skild's own benchmark compares S1 with a conventional language-prompted vision-language-action (VLA) policy trained on the same data, at dataset sizes from 1,000 to 100,000 hours. On unseen tasks, the in-context model reached a 66% average cumulative per-step success rate, while the language-prompted model reached 9%. Skild said one demonstration in context is worth roughly 380 post-training examples, and that for tasks longer than four minutes, collecting 380 demonstrations takes 50 to 100 hours of teleoperation. A post-trained model eventually overtakes S1, reaching 86% with 2,000 demonstrations, by the company's account. On tasks seen in training, S1's accuracy rose to 96% as pre-training scaled.

The training data mixes teleoperation, data from Universal Manipulation Interface (UMI) handheld grippers, egocentric human video and simulation. Skild said it spends three dollars on quality control for every dollar spent on collecting data. In a September 10 blog post, NVIDIA said the S1 pipeline uses its Cosmos world models and Cosmos Curator for data, Isaac Sim and Isaac Lab with the Newton physics engine for simulation, and TensorRT for inference.

Skild was founded in 2023 and has raised heavily. Its $300 million Series A in July 2024 valued it at $1.5 billion. In January 2026 it announced a $1.4 billion Series C led by SoftBank at a valuation of more than $14 billion, with NVentures, Macquarie Capital, Bezos Expeditions, LG, Schneider Electric, Salesforce Ventures and Mirae Asset among the backers. At that time it said it had made about $30 million in a few months of 2025. In April 2026 it bought the robotics business of Zebra Technologies, formerly Fetch Robotics, and its Symmetry Fulfillment orchestration platform.

The figures give Skild an unusual position among peers. Figure AI raised $1 billion at a $39 billion post-money valuation in September 2025, and Apptronik closed a $520 million Series A extension in February 2026, but neither has published a recurring revenue number. Physical Intelligence, which raised a Series B in November 2025, publishes research rather than revenue. Google DeepMind's Gemini Robotics 2, released on July 30, is still limited to partners for its action models.

The appeal of in-context learning is economic. Every fine-tuning cycle for a new task costs data collection, compute and engineering time. If a customer can record one video and get a working policy in minutes, the cost of adding a task falls towards the cost of filming it. Skild's own numbers also show the limit: post-training on 2,000 demonstrations still beats S1 on success rate, so high-volume tasks may still justify conventional fine-tuning.

Several questions remain open. ARR is a forward-looking run rate, not audited revenue, and Skild has not said how much comes from software licences versus hardware and services, including the Zebra robotics business it acquired in April. The 66% figure is a per-step average from the company's internal evaluation, not an independent benchmark. Skild has not named the robot hardware used in the S1 tests.