Skild AI hits $100M revenue run rate in 10 months. Physical AI is apparently very lucrative

10 months is how long it took Skild AI to go from its first commercial deployment to $100 million in annual recurring revenue run rate, according to a recent story in Bloomberg.

Founded in Pittsburgh in 2023, the company makes what it calls a "universal brain" for robots: software that lets machines learn new tasks by watching a single video, no retraining required.

That's a significant ability. Robotics has spent decades struggling with the reality that teaching a robot something new means thousands of examples, expensive data collection, and lengthy retraining cycles. 

Skild says its S1 model breaks that equation entirely.

From 8 customers to 60 in one year

At the start of 2026, Skild had eight paying customers. 

By September, it had more than 60, with hundreds of robots deployed across manufacturing, logistics, inspection, security, and food preparation. Cafes in Japan. Warehouses in the United States. High-precision assembly lines at Foxconn building NVIDIA Blackwell systems ... tasks that include busbar installation, 16-screw fastening, and real-time error recovery.

Revenue in 2025 came in around $30 million. The $100 million run-rate figure reported today represents roughly a 3x increase in less than twelve months. 

At a $14 billion valuation — SoftBank led the $1.4 billion Series C in January 2026, joined by NVIDIA, Jeff Bezos, Samsung, Salesforce, and Sequoia — Skild is trading at about 140x ARR. Investors are clearly betting on a trajectory, not a snapshot.

"This is a major milestone for us as it signifies a new chapter for robotics, moving from an era of demos to an era of deployments," said co-founder and CEO Deepak Pathak.

What the S1 model actually does

The core insight behind Skild is that in-context learning — the same capability that lets large language models adapt to new problems without retraining — can work for physical robots too. 

Show S1 one video of a task it has never seen, and it figures out how to do it. No weight updates. No fine-tuning cycle.

The company claims a 96% success rate on tasks it has encountered before and 66% on tasks it has never seen IRL after watching a single demonstration video. 

The company says one video provides the equivalent of roughly 380 hands-on training examples under conventional approaches. It also claims S1 is the first robotics foundation model to handle in-context learning on tasks lasting up to ten minutes, well beyond the short-horizon capabilities most models have shown.

(Should somebody tell Figure AI, which just committed to spending billions on training data?)

The S1 model was built on an enormous data foundation. Skild runs 100,000 simultaneous robot instances inside NVIDIA Isaac Lab, generating what amounts to a millennium of simulated experience every few days. It also trains on internet video at a scale roughly 1,000 times larger than conventional robotics models. That data flywheel, combined with the NVIDIA partnership for GPU-accelerated simulation and physics modeling via the Newton physics engine, gives Skild a compounding advantage that is hard to replicate quickly.

The horizontal bet

Most of the attention in humanoid robotics goes to the hardware companies: Figure, 1X, Apptronik, Unitree. 

Skild is playing a different game. The company is building a software layer that abstracts away the hardware differences between robot models. ABB and Universal Robots are among the manufacturing partners. The Zebra Technologies robotics automation business, acquired in April 2026, added logistics muscle.

Deployment costs for Skild Brain run between $4,000 and $15,000 per robot. Compare that to $250,000 for conventional robot programming approaches. That is a fundamentally different economic proposition for customers weighing automation investments.

CEO Pathak put it directly: "Learning by experience, and not preprogramming, is the step change that has happened in robotics."

What the $100M milestone actually means

Revenue run-rate figures are always a little slippery: they project current momentum forward and assume nothing changes. But the trajectory here is hard to dismiss. Going from $30 million in 2025 to $100 million run rate in 2026, with a customer base that grew 7.5x in eight months, suggests Skild is not just selling pilots. 

Robots are being deployed, tasks are being executed, and customers are paying recurring fees.

The broader context matters too. Physical AI was barely a phrase 18 months ago. Today it attracts billions in capital and has a legitimate revenue base. Skild is not the only player (Physical Intelligence, Covariant, and others are building similar foundation model approaches), but hitting $100M ARR this fast puts it in a category that very few enterprise software companies reach in their first two years, let alone robotics companies.

Europe's Machinery Regulation takes effect January 20, 2027, requiring AI-based machinery with self-evolving behavior to go through formal oversight assessments. Nobody has completed that process yet. As Skild scales into European markets, that's an interesting twist.