Capability
Robotics is at a “GPT-2.5 moment”
Capability is real, but reliable deployment remains incomplete. Robot data is structurally scarce, and data pipelines, simulation and world models are central to competitive advantage.
Field Notes
We argue with our own assumptions in public.
The evidence we are underwriting against, cited so it can be checked, followed by the register of ways this thesis could turn out to be wrong.
Evidence
What the public record actually says.
Six readings that shape the position. Each links to its source; where we have drawn a conclusion the source does not state, that is our inference and should be treated as such.
Framing
Focus on capability, not embodiment
Mature perception separates cleanly from the harder problems of manipulation, planning and reasoning — and software-defined configuration can address a meaningful share of setup and reconfiguration cost.
Capital
Record capital, rising scrutiny
A reported $16.3B across 492 robotics and Physical AI deals in Q1 2026, with investors increasing scrutiny of manufacturing scale, paying customers and integration cost. Capital is abundant; proof is not.
Concentration
Concentration in large platform rounds
$18.8B invested in robotics startups during 2026 by the article date, exceeding the full-year 2025 total, with capital concentrating in large platform and system rounds — precisely what leaves the enabling layer underfunded.
Infrastructure
World models as strategic infrastructure
Cosmos world models, Isaac GR00T, Isaac Lab and Isaac Sim validate simulation, synthetic data and accelerated compute as strategic infrastructure — and create the platform the enabling layer must interoperate with.
The gap
Capability advances, evaluation lags
Gemini Robotics shows continuing progress in vision-language-action models, embodied reasoning and on-device robotics — while its model cards and usage constraints reinforce the unresolved need for evaluation and safety tooling.
Risk register
How this thesis could be wrong.
Published because a thesis that cannot be falsified is marketing. Each line is a way the construction fails, and we update this page when one of them starts looking likely rather than after it has happened.
- Platform vendors absorb the enabling layer into the base stack, for free
- Constraints resolve faster than a seed company can build a durable position
- Leverage does not transfer — every robot needs a bespoke solution after all
- The buyers stay research budgets rather than becoming operating budgets
- Open-source alternatives reach good-enough quality before capture is established
- Capital concentration at the platform layer starves the enabling layer of Series A
- We are simply early, and the constraint window opens after Fund I is deployed
Think we have read the evidence wrong?
That is the most useful message we can receive. Tell us which reading is mistaken and why, and we will publish the correction.