How to price a physical-AI ramp: Skild AI's 8-to-60 jump
A three-point test for pricing early physical-AI companies, using Skild AI's customer and deployment jump as the worked case.
When a physical-AI startup claims it is close to commercial scale, the question is whether the control layer is broad enough, the paying base is widening fast enough, and the raised capital is turning into deployments. Skild AI's customer and deployment ramp gives a worked case for that test.
Skild AI was founded in 2023 in Pittsburgh by Carnegie Mellon Robotics Institute professors Deepak Pathak and Abhinav Gupta. Its bet is a shared control layer that can move across sites.
A general-purpose control layer is the first gate
Inspect the control layer. A physical-AI company needs a model that can run across robot types and tasks, not a narrow skill tuned to one machine. Done looks like a general robot control model that can be licensed to multiple operators without rebuilding the core for each site.
Ask what changes when the robot changes. If the company must retrain a new model for every arm, chassis, or warehouse layout, the revenue curve will be slow and uneven. When the same model can be deployed with site-specific tuning, the cost of adding a customer should fall as the base grows.
Skild's S1 model showed 96% success on previously seen tasks, 66% on unseen tasks, and 9% for a language-prompted baseline. The gap between familiar work and new work is where deployment cost appears.
Customer and robot growth is the second gate
Track paying customers and robot units. The commercial ramp should show both breadth and volume: a customer count that expands quickly while the same software runs on more robots in different industries. One large customer is less useful than several smaller ones.
Do not count pilots as revenue. Pilots can prove the model, but they do not prove that the operator will keep paying, expand, or refer other sites. Ask how many deployments are under contracts, how many are renewing, and how many robots each customer is running. Renewal rates are the clearest sign of operator trust.
Skild AI had 8 customers at the start of 2026 and later had more than 60, while hundreds of robots used its machine-learning software in manufacturing, logistics, inspection, security, and food preparation; its disclosed $100 million annual revenue run rate came ten months after initial commercial use.
Capital should show up in deployments, not just valuation
Follow the capital into deployments. Post-raise capital is only useful if it converts into paying customers, robot units, and operating revenue. The signal is a funding round followed by a visible increase in commercial use, not just a higher valuation.
Check what the new money is actually buying. It should fund field engineering, integration, safety review, and customer success, not only research headcount. Field work is where physical AI gets expensive. When the company cannot say which deployments the capital is enabling, the valuation is based on a promise.
Skild AI had raised over $2 billion, including a $1.4 billion Series C in January led by SoftBank, with NVentures, Bezos Expeditions, Samsung, LG, and others participating, that put its valuation above $14 billion, more than three times its price seven months earlier, at the $300 million Series A round in July 2024.
Price the ramp, not the demo
Use the test to set the discount. If all three points are present, the company can be priced closer to a software deployment business. A strong model alone should be priced closer to a research bet. A growing customer base with a narrow control layer may be real, but the limit may be low.
Write the three answers into the term sheet before you negotiate the price:
- General-purpose control layer: can the same model move across robot types and sites without a rebuild?
- Paying-customer and robot-deployment growth: are contracts, renewals, and robot units expanding across industries?
- Post-raise capital deployed into deployments: is the new money funding field work, not just research?