Featured essay
The robotics companies worth backing don't build robots
Why we started a fund for the constraints every robot depends on — and the test we run before we write a check.
Read the essay →
News & Blog · Field Notes
Field Notes is what we learn as we meet founders building the unglamorous middle of Physical AI. We argue with our own assumptions in public, and update when the evidence says to.
Why we started a fund for the constraints every robot depends on — and the test we run before we write a check.
Read the essay →The most valuable robotics company of the next decade might never ship a robot.
That sounds wrong, so let me explain. When you watch the field from the outside, the story is about machines: the humanoid that folds laundry, the arm that sorts packages, the rover that inspects a pipeline. Those get the headlines and most of the capital. But talk to the engineers actually putting machines into a warehouse or a hospital, and the conversation changes fast. They don't talk about the robot. They talk about what's holding it back.
The data is messy and expensive to collect. The simulation looks great and then reality disagrees. Evaluation is mostly vibes — the demo works, so ship it, and hope. The edge hardware runs out of headroom the moment the model gets interesting. None of these are the robot. All of them decide whether the robot works.
We started AWSM VC because that gap is where the durable companies are, and almost nobody is funding them on purpose.
Here's the belief underneath the fund. Value in a Physical AI system is not spread evenly across the stack. A small number of constraints govern most of the cost, the reliability, the deployment time, and the safety. Data quality can matter more than model size. Simulation fidelity can matter more than raw compute. Evaluation can matter more than the demo that raised the round. If you can name the constraint that governs the rest and back the company that controls it, you're not helping one robot. You're helping every system that hits the same wall.
So we built a test, because a belief without a filter is just a slogan. We call it the Pareto Constraint Method. Before we invest, we score a company on six questions — criticality, leverage, defensibility, productization, capital fit, and capture. A company we back should score well on all six. Most don't, and that's the point.
I should be honest about where we are, because the field has enough overconfidence already. AWSM VC is early. Our fund terms are still forming, and anything aspirational is marked as such. We're not claiming a track record we haven't earned. What's fixed is the thesis and the discipline: fund the constraint, unlock the system, and show our work along the way.
Latest
The cheapest path to a reliable robot isn't a bigger model — it's a data engine that turns every deployment into fresh training signal.
The evaluation layer is the least glamorous, most underfunded part of the stack — and where enterprise trust and budgets actually live.
Criticality, leverage, defensibility, productization, capital fit, capture — and why saying no is most of the job.
General-purpose simulators impress. The companies we back close the gap for a specific, performance-critical task first.
The moment a model gets interesting, on-device compute becomes the constraint. Whoever solves it owns the deployment.
Subscribe
New essays on the constraints Physical AI can't scale without. No hype, no cadence promises we can't keep.
How we write
Field Notes are working beliefs, not press releases. We state the thesis, show the evidence, and mark what's still aspirational.
When the evidence changes, we update the note and say what changed. No hype, no manufactured urgency, no results we haven't earned.
If you're working on the constraints Physical AI can't scale without, we want to hear about it.