From Demos to $100M in Ten Months: Skild AI Declares the Era of Robot Deployment Open

Ten months is all it took. On September 9, 2026, Pittsburgh-based Skild AI announced that it had crossed $100 million in annual revenue run rate, ten months after its first commercial deployment. The software now runs on hundreds of robots at more than 60 companies — up from eight at the start of the year — spanning manufacturing, logistics, inspection, security, and food preparation. For a company selling not a task-specific automation program but a general robot brain, the milestone lands as the clearest signal yet that physical AI has begun its transition from research demos to recurring revenue.

The numbers behind the milestone

The figures come from Skild’s own announcement, confirmed by Bloomberg: $100M in recurring revenue run rate, 60+ paying customers, and hundreds of deployed robots across warehouses, factories, and data centers. Skild’s president told The Next Web last month that 2025 revenue was around $30M — meaning the company has roughly tripled its top-line pace inside a year. Mobility accounts for about 10% of revenue and “Fetch” solutions 4%, with the remainder dominated by manipulation-heavy work.

The commercial traction validates the company’s $1.4B raise in January 2026 at a $14B-plus valuation — reportedly the largest robotics-AI funding round in history — led by SoftBank with backing from Amazon and, reportedly, Nvidia. At the time, plenty of observers asked whether a pre-revenue-scale robotics lab could justify that price. Ten months of deployments later, the question has softened.

What customers are actually buying

Skild sells one thing: the Skild Brain, a foundation model meant to control many robot bodies. Its newest release, S1, learns a previously unseen task from a single video demonstration — no fine-tuning, no new dataset collection, no retraining run. An operator records a person doing the job; the robot interprets the intent, objects, and sequence, and executes, including tasks that stretch past ten minutes and dozens of manipulation steps.

The flagship deployments show how generalist that claim is in practice:

• NVIDIA and Foxconn — Skild Brain is deployed on dual-arm manipulators performing high-precision assembly of NVIDIA Blackwell systems. In one demonstrated workflow, a robot installs a busbar and limit block, fastens 16 screws, and adapts to disturbances mid-task. The work changes with every product cycle — precisely the kind of variability that used to mean reprogramming every robot on the line.
• Sumitomo Wiring Systems — deploying S1 to automate wire-harness manufacturing processes long considered “impossible” to automate, because the work defies fixed programming.
• Mitsui & Co. — piloting S1-powered general-purpose robots in commercial kitchens across supply chains that serve 1.4 million meals a day in Japan.

That spread — GPU assembly, wiring harnesses, food service — is the point. A specialist automation vendor would need three separate product lines. Skild is shipping one model.

Why in-context learning changes the economics

The reason this matters commercially, not just academically, is the cost structure of robot deployment. Traditional industrial robots are built for fixed jobs: every new product, layout change, or process tweak requires fresh data collection, retraining, and revalidation. Skild quantifies the alternative starkly — in its tests on new multistep tasks, S1 succeeded about 66% of the time at each step versus 9% for a comparable AI system, a more-than-sevenfold improvement. And a single video demonstration, the company estimates, does the work of roughly 380 hands-on training examples — collecting which manually would take 50 to 100 hours of teleoperation.

In one plant-potting test, the team went from recording the demonstration to autonomous execution on hardware in 11 minutes. That is the difference between a robot as a capital project and a robot as a tool.

The NVIDIA stack underneath

S1 is built and trained on NVIDIA AI infrastructure, and the collaboration — detailed in a September 10 NVIDIA blog post — spans the full development cycle. NVIDIA Cosmos world foundation models diversify training data and convert video into structured descriptions; Cosmos Curator annotates and filters data at scale. Training and validation happen in NVIDIA Omniverse and Isaac Sim, with reinforcement learning in Isaac Lab powered by the Newton physics engine to model forces, contact, and collision — narrowing the sim-to-real gap. On the deployment side, TensorRT optimizes inference latency so robots can react quickly enough for the physical world.

Notably, the two companies are jointly developing new GPU-accelerated simulation solvers for contact-rich manipulation — modeling how robots touch, grip, and manipulate solid objects — which will be made available to all developers as part of Newton. That is an open-source contribution embedded inside a commercial partnership, and it signals how much of physical AI’s future tooling is being built in public.

The deployment flywheel

The most interesting argument in Skild’s announcement is cultural, not financial. “The era of demos is over; the era of deployment has begun,” write co-founders Deepak Pathak and Abhinav Gupta. They describe demo culture as actively corrosive: a 10%-accurate robot and a 99%-accurate robot can produce identical-looking highlight clips, and if demos are what a company celebrates, demos are what its engineers will work on.

Their alternative is what they call physical RSI — recursive self-improvement through deployment. Deploy a generalist model; let each deployment specialize it; bring the specialist data back into the general base model; deploy the stronger generalist. Experience that adds nothing to a task-mastering specialist can still teach the generalist a great deal. Their analogy: imagine sending your high-school self through parallel PhDs in chemistry, physics, and mathematics, then distilling all of it back into one student. Every deployment makes the next deployment require less specialization.

The European footnote

One wrinkle worth watching: Europe’s Machinery Regulation, which replaces the machinery directive on January 20, 2027, explicitly covers machinery with “self-evolving behavior” — machines that change what they do after shipping. Such systems will require assessment by a notified body rather than a manufacturer’s self-declaration. A model whose entire selling point is performing tasks absent from its training data sits close to the center of that definition. No such system has been assessed yet, and Skild has not said whether it sells into Europe; its named customers are in Japan and the United States. But as physical AI commercializes, the regulatory regime that arrives in sixteen months will be the first to grapple directly with robots that learn on the job.

What to watch

Skild’s trajectory now has three open questions. Can S1’s 66% unseen-task step success rate be pushed toward deployment-grade reliability, and how fast? Will the deployment flywheel compound — does specialist data genuinely strengthen the generalist model, or does it dilute it? And will competitors like Physical Intelligence and Generalist AI, who are racing toward the same generalist-brain thesis, match Skild’s commercial pace or beat it on model capability? For now, the company that promised a “BERT-to-GPT-3 moment” for robotics has something more tangible than a benchmark: a nine-figure run rate, earned one deployment at a time.