Skild AI Crosses $100 Million Annual Revenue Run Rate in Ten Months

Skild AI said on September 9, 2026 that it crossed $100 million in annual revenue run rate, ten months after the robotics company’s first commercial deployment.

The milestone was disclosed in a company blog post titled “The Hidden Pillar of Robotics.” In the post, the company said it has scaled to more than 60 paying customers across moving goods, making deliveries, inspecting sites, providing security, and preparing food, as well as operating inside warehouses, factories, and data centers. Mobility accounts for 10% of revenue and Fetch solutions account for 4%, the company said.

Named Commercial Deployments

The post described three deployments with technology and industrial partners. Together with NVIDIA and Foxconn, Skild AI is deploying the Skild Brain on dual-arm manipulators for high-precision assembly of NVIDIA Blackwell systems, work the company said changes with every product cycle and previously meant reprogramming every robot on the line.

At Sumitomo Wiring Systems, Skild AI said it is working towards deploying S1, its latest foundation model, to automate processes in wire harness manufacturing that the post says were considered impossible to automate. With Mitsui & Co., whose supply chains serve 1.4 million meals a day across Japan, the company is piloting general-purpose robots powered by S1 in commercial kitchens.

S1 Model and Company-Reported Results

Skild AI introduced S1, which it calls its flagship robotic foundation model, in an August 18, 2026 post. The company said it built S1 from the ground up as an in-context learner: an operator records a single video demonstration of a task and gives it to the robot as a prompt, without updating the model’s weights. S1 is built on NVIDIA AI infrastructure, which the company said provides the accelerated computing foundation needed to train at scale across its mix of robotics data. According to the August post, S1 is the first robotics foundation model to show in-context learning on extremely long-horizon tasks of up to ten minutes that were never seen during pre-training.

In a controlled internal study described in that post, Skild AI trained in-context learning and language-conditioned VLA policies on identical data, architectures, and compute across datasets ranging from 1,000 to 100,000 hours. At 100,000 hours of pre-training, the in-context policy reached a 66% success rate on unseen long-horizon tasks of four to eight minutes, compared with 9% for the language-prompted policy, the company reported. On tasks drawn from the pre-training distribution, the respective figures were 96% and 89%. The company said it used human intervention to recover from failures during rollouts, mainly for the language-prompted baseline, so that every step could be graded cumulatively.

The August post also stated that a single in-context demonstration matched roughly 380 post-training episodes on unseen tasks, a crossing the company said it estimated by interpolating between measured points. Collecting 380 long-horizon demonstrations takes 50 to 100 hours of teleoperation, according to the post, while post-training the VLA policy on 2,000 demonstrations reached an 86% success rate. For one plant-potting task, the company logged 11 minutes from demonstration to autonomous execution: supplies arrived at 8:54 PM, a single egocentric human demonstration was recorded at 9:22 PM, and S1 began executing the task autonomously at 9:27 PM.

On data quality, the company wrote: “For every dollar we spend on collecting data, we spend three on quality control.” Every data point entering pre-training is screened for low-level precision, task coherence, and annotation fidelity, according to the post.

The September 9 post describes deployment as the hidden pillar of robotics research and argues that demo videos are an unreliable measure of progress, because a successful clip from a robot with 5%, 10%, or 99% accuracy can look the same. The company said it taught its model to make eggs last year: the first egg took a week, and making the process reliable across different eggs and setups took two months. Squeezing the last 5% of performance takes greatly more effort than the first 95%, the post says. The post also argues that a robot that is 99.9% accurate but ten times too slow is not deployable, since every station on a factory line must finish within roughly the same cycle time.

According to the post, lessons from deployments directed the company’s research toward S1, because a system that requires a new dataset and another post-training run for every supplier change, workstation rearrangement, or revised assembly sequence does not scale. The company described the resulting loop as physical recursive self-improvement: experience from specialized deployments is brought back into the general base model, so each new deployment starts from a stronger model.

The August post documents a development timeline running from LocoFormer, an in-context learner for locomotion, in September 2025 through first in-domain in-context results in February 2026 and a first out-of-distribution pancake flip in May 2026 to the S1 release in August 2026. That post stated that S1 is already at work with the company’s commercial partners. The September 9 post closes: “The era of demos is over; the era of deployment has begun.”