Skild AI Hit $100M Selling Robots a Brain

Skild AI crossed a $100 million annual revenue run rate on September 10, ten months after its first commercial deployment. Its software runs on hundreds of robots at more than 60 companies, up from eight customers at the start of 2026. Named deployments include Blackwell assembly work with NVIDIA and Foxconn, Sumitomo wire harnesses, and Mitsui kitchens.

Alongside the revenue number, Skild detailed S1, a robot foundation model that learns previously unseen long-horizon tasks from a single video demonstration — using the video as in-context input, without updating weights and without task-specific post-training. Skild built S1 and the underlying research on NVIDIA infrastructure, spanning synthetic data generation, training, simulation, and real-world deployment.

Ten months from first deployment to $100 million is the fastest commercial ramp in the history of robotics software. The reason is not the model. It is what the model removed.

The economics that changed

Industrial robot deployment has always had the same cost structure: the hardware is a fraction of the total, and integration is the rest. A robot cell that does one task costs a few hundred thousand dollars in equipment and takes months of systems-integration work to program, fixture, tune, and validate. Change the task and much of that work happens again.

That cost structure is why robotics automation stalled at high-volume, low-variance work. If you are making the same part a million times, amortizing six months of integration is trivial. If you are making forty variants of a wire harness, it never pays.

Single-video task acquisition attacks exactly that. If a new task costs a demonstration instead of an integration project, the economics of high-mix, low-volume work invert — and high-mix, low-volume is the overwhelming majority of manufacturing that has never been automated.

The customer list confirms this reading. Wire harnesses at Sumitomo are the canonical high-mix, dexterity-heavy, never-automated task. Kitchens at Mitsui are unstructured environments with variable objects. Neither is a job you solve by programming a trajectory.

The line worth quoting back at the industry

Skild's own framing includes an observation that most robot-learning research ignores: 99.9% accuracy is still unusable if it is 10× too slow.

This is the honest constraint, and it is the one that separates a deployment business from a demo business. Almost every impressive manipulation result published in the last three years runs at a speed that would fail an industrial cycle-time requirement. A robot that picks correctly in twelve seconds when the line allows four has not automated anything; it has produced a video.

A company at a $100 million run rate across sixty customers has necessarily solved for cycle time, because purchase orders in manufacturing are written against throughput. That is a stronger signal of genuine capability than any benchmark in the field currently offers.

The revenue quality question

Two caveats belong on the number.

Run rate is not revenue. A $100 million annualized run rate ten months in means recent monthly revenue extrapolated forward. In a business scaling from 8 to 60 customers inside a year, that extrapolation flatters growth — and some portion of those sixty customers are running pilots, not production fleets. Skild reports hundreds of robots deployed, which across 60 customers averages to single-digit-to-low-double-digit units each. That is pilot-scale density.

The NVIDIA relationship cuts both ways. Training on NVIDIA infrastructure as part of a broad collaboration is validation and it is also concentration. NVIDIA has its own humanoid reference designs, its own Cosmos physical-AI stack, and a clear strategic interest in owning the robot software layer. Today that alignment is a tailwind. The version of this story in 2028 depends on whether NVIDIA stays a platform or becomes a competitor.

The number that would resolve both concerns is net revenue retention — whether the customers who started with five robots are running fifty. Skild has not published it.

What this means for the humanoid trade

The contrast with the rest of the sector is stark and instructive.

This month alone: Neura Robotics is raising €1 billion to move from prototyping to building. XPeng lit up IRON's production line targeting mass humanoid output. SoftBank is in talks for a majority stake in 1X at about $6 billion — roughly 40% below the $10 billion 1X sought less than a year ago. IFA Berlin filled a hall with humanoids that cannot yet do useful work.

Billions are flowing into bodies. Skild sells brains, on hardware that already exists, doing work customers already needed done — and it is the one posting revenue.

There is a lesson in that ordering. The humanoid form factor is a bet that general-purpose embodiment eventually beats specialized machines. It may be right. It is also a bet that requires solving locomotion, manipulation, power, cost, and reliability simultaneously before the first dollar arrives. A foundation model that drops into an arm already bolted to a line gets paid while that bet is being settled.

What to watch

Whether S1's single-video claim survives contact with variance. Learning a multi-step task from one demonstration in a controlled setting is different from learning it when the parts arrive in a different orientation, the lighting changes, and the gripper is worn.

Robot count per customer. Sixty customers at pilot density is a good year. Sixty customers at fleet density is a category.

Whether anyone else gets close. Skild's ramp is currently unmatched, which means either it found the right wedge first or the market is small enough that one company can fill it. The next four quarters distinguish those.