02 / Physical AI
Picks and shovels for autonomy.
The layer every autonomous machine runs on: simulation, perception and data, validation, and fleet operations. We do not pick the winning robot. We own the ground all of them stand on.
What it is
The software and data beneath the machines.
Simulation environments and synthetic data. Perception stacks, sensor calibration and the pipelines that turn raw sensor logs into training data. Evaluation and validation tooling, safety cases, and the record that proves a system is safe enough to ship. Fleet operations, remote assistance and the software that runs a thousand machines once they leave the lab. World models and the data infrastructure that trains them.
The machines themselves are fragmenting: road vehicles, warehouse and factory robots, drones, agricultural and mining equipment, humanoids. The layer beneath them is doing the opposite. The same simulation, the same data discipline and the same validation standard serve all of them, and the companies that own those layers grow with the number of programs, not with the fortunes of any one machine.
We do not need to know which robot wins. We need to own what every one of them has to run on.
Why now
More programs, cheaper hardware, general models.
The number of autonomous programs in flight is a multiple of what it was three years ago, and each one needs the same infrastructure. Sensors and actuators are commoditizing, which pushes the differentiation into software and data. And general robot foundation models, far from replacing the layer, make it more valuable: a general model still has to be adapted, evaluated and proven on the specific machine in the specific environment, and the data and validation infrastructure is what does that.
The same logic that governs the rest of our thesis applies here. The model is becoming the commodity. The data with physics in it, the validation record and the toolchain everyone integrates are what the commodity has to pass through.
Where value accrues
Edge cases, proof, and the toolchain.
Data with physics in it
Multi-sensor, multi-environment, edge-case-rich. The long tail of what goes wrong on a road, a floor or a field is expensive to collect, and it is what separates a demo from a deployment.
The validation record
Regulators, insurers and customers will require proof that a system is safe. The company whose evaluation record is the accepted proof holds a standing no model release removes.
Integration into the program's toolchain
When the simulation, the data pipeline and the fleet tooling are how the engineering team works every day, the company is the system of record for the program. Switching means re-validating everything.
Breadth across programs
Every additional program on the platform improves the data and the tooling for all of them. That compounding is the distribution moat in this layer.
The fork
Stronger, or erased.
Owns edge-case data and the validation standard, and is embedded in the toolchain of many programs. A better general model raises the value of its data and evaluation, because someone has to prove the model works on this machine.
A single-task perception model a foundation model absorbs. A robotics application without a data flywheel. Hardware that competes on cost. A growth plan that depends on a defense contract, which is outside our thesis regardless of the technology.
What we ask
The questions the memo has to answer.
What does the company own when the customer's own team catches up?
Large programs build in-house. The answer has to be data, standing or breadth the customer cannot replicate alone.
How many programs run on it, and would one switch?
Breadth is the moat. Concentration in one customer is a risk we price.
Whose validation counts?
If the company's record is what a regulator or insurer accepts, that is a standing. If it is an internal metric, it is a feature.
Does a better robot foundation model make the data more or less valuable?
The right answer is more. If the honest answer is less, the company is a model, not a layer.
What is the real gross margin?
Compute, data operations and support are cost of goods here. We underwrite the margin with all of it counted.
How does it price?
Per vehicle, per scenario, per mile, per program. Pricing that scales with the customer's fleet is what makes the layer compound.
The bar by stage
What good looks like from first cheque to Series A.
Pre-seed
A team from inside an autonomy program that has felt the missing layer. A wedge one program will pay for now. A clear view of which data or validation asset the wedge starts to accumulate.
Seed
Several programs on the platform, at least one in deployment rather than research. Data or evaluation results the customers rely on and cannot produce alone. Unit economics that survive the cost of data operations.
Series A
A repeatable sale across program types, annual recurring revenue in the low millions and growing, net revenue retention above roughly 110 to 120 percent, and evidence the platform got more valuable when the last general model shipped.