Skild AI's $100M run rate: the robot brain is software
Skild AI went from 8 customers to 60+ and a $100M revenue run rate in ten months. The 96%/66% task gap is the number that decides your automation budget.
Skild AI crossed a $100 million annual revenue run rate, Bloomberg reported on September 10, ten months after its first commercial deployment. The company does not build robots. It sells the model that runs them — hundreds of machines across more than 60 companies, up from eight earlier this year. Physical automation just produced its first believable software revenue curve, and the shape of that curve is the part worth reading.
What actually happened
Skild's product is a robot foundation model. Its S1 model, launched in August, learns a task by watching a single video of a person doing it, including sequences that run past ten minutes. Skild says one demonstration does the work of roughly 380 post-training examples. Co-founder Deepak Pathak has cited the model doing things absent from its training data — flipping pancakes, in the example that made the rounds.
The success rates are published and uneven: 96% on tasks the model has seen, 66% on tasks it has not. That thirty-point gap is the entire procurement conversation. A system that is right two times in three on novel work is a capable research result and a staffing problem, depending on what happens when it is wrong.
The financials: roughly $30 million in revenue in 2025, a $100 million run rate now, and a $1.4 billion Series C led by SoftBank in January 2026 at a valuation above $14 billion. The research and training ran on Nvidia infrastructure, per Nvidia's writeup of the collaboration.
Why a robot foundation model matters for your business
You are probably not buying a robot this year. You should still read this release, because the business model is the news, not the hardware. Skild sells a brain that runs on machines the customer already owns, priced like software, improved centrally, updated for everyone at once. That is the same pattern that ate every other category of capital equipment software, and it sets the terms you will see quoted when a vendor does come to your warehouse or kitchen.
The number to carry into that meeting is the 66%. Ask any physical-automation vendor to split their accuracy claim into seen and unseen tasks, and to define which of your workflows falls on which side. A pick-and-place station with a fixed SKU set lives near the 96%. A returns bench, where the next item is genuinely unknown, lives near the 66%. Those are different purchases with the same brochure.
And note what one video replacing 380 examples actually changes: the cost of teaching, not the cost of being wrong. Cheap teaching means more tasks get attempted, which means more novel tasks, which means you spend more time in the 66% band, not less. Budget for the exception handling, not just the deployment.
Key takeaways
- Skild AI hit a $100M annual revenue run rate ten months after first commercial deployment, up from about $30M in 2025
- Hundreds of robots at 60+ companies, up from 8 earlier this year
- S1, launched in August, learns a task from one video demo - roughly the value of 380 post-training examples
- Published success rates: 96% on seen tasks, 66% on unseen - ask every vendor for that split
- SoftBank led a $1.4B Series C in January 2026 at a valuation above $14B; training ran on Nvidia infrastructure
- The model is sold as software on hardware you already own - expect that pricing shape from the next vendor who calls
- Cheaper teaching raises the share of novel tasks, so budget exception handling alongside deployment
Automation earns its keep at the exception, not the demo. We build the workflow layer around automated systems - the queue that catches the 34%, the audit trail, the human handoff - so a 66% success rate becomes a working process instead of a stalled pilot. See what we've shipped, or bring us a process you want measured before you automate it.
Sources: Bloomberg: Robotics Startup Skild AI Hits $100 Million in Revenue Run Rate, Nvidia: Skild AI S1 physical AI.
- #skild-ai
- #robotics
- #physical-ai
- #foundation-models
- #automation-roi
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
Persimmon simulates users so you can test the agent
humans& released Persimmon, a 550B user simulator that fools AI judges 19.8% of the time. A practical way to load-test a support agent before real customers do.
Read itMoonshot AI hit $1B ARR — the cheap model tier grew up
Moonshot told investors annualized revenue topped $1B in August, up from $300M in June. What a durable low-cost model vendor does to your routing math.
Read it