Thesis
The bodies arrived early. Nobody owns the layer above them.
Hardware and intelligence are both being driven to commodity by people with more capital than us. That is the opportunity, not the threat — provided you are positioned to collect from it rather than compete with it.
Why now
Eighteen months ago this layer would have been worthless.
You cannot build Kubernetes before servers exist. The industry has only just produced enough competent, cheap, interchangeable hardware for an orchestration layer to be worth anything — and the intelligence gap above it is still wide open.
2,056 humanoids, 666 teams, 51 events at the World Humanoid Robot Games. 500 robots and 280 teams a year earlier.
SoThe body race is decided at volume, not at prototype.
Tiangong Ultra runs 100 m in 8.64 s — 21.50 s in 2025 — and crashes past the finish line.
SoMotor control is solved far faster than situational reasoning. The gap is the product.
Chinese manufacturers reported at >97% of global humanoid shipments. AgiBot ~8,400 units, Unitree ~5,900.
SoHardware is a commodity supply chain we should buy from, not a race to enter.
NVIDIA ships an openly licensed GR00T VLA, Isaac Sim/Lab and a reference humanoid assembled from third-party parts. Google ships Gemini Robotics 2.
SoEven the frontier is now assembling rather than integrating. Intelligence is arriving free.
Reuters finds Chinese humanoids still lose to conventional industrial arms on adaptable, precise factory work.
SoNobody has closed the loop between a capable body and a completed job.
Figures are as reported by the sources named and are not independently verified here. The thesis does not rest on their precision — it rests on a direction several independent sources agree on.
Discipline
Four things we refuse to build.
Each has a better-funded owner racing to give it away. Every one of them is a supplier to this layer, and every improvement they ship arrives here for free.
A humanoid body
China ships >97% of them and the price is falling toward $5,000. Every cut creates another endpoint for us.
A robot foundation model
NVIDIA gives one away and Google ships another. Competing costs billions to reach parity with free.
A simulator
Isaac exists, is better than ours would be, and is free.
Low-level control
Solved, commoditised, and owned by whoever made the actuator.
What compounds
The only asset in this stack that appreciates.
Robots depreciate. Models are superseded every quarter. The record of which combination actually completed paid work gets more valuable with every job — and it is generated as a side effect of serving the customer, on hardware we do not own.
Google has web data. Tesla has driving data.
We need physical cause and effect: attempt, failure, correction, success. Not footage of robots succeeding — the corrections are the rows nobody else has, because nobody else is on the hook for the outcome.
The metric we watch is not episodes ingested. It is distinct context fingerprints: ten million episodes of one robot doing one thing in one room is not a moat. Coverage is.
The asset
The table nobody with a frontier model can produce.
Which model, on which body, in which environment, actually finishes paid work. Google knows how its model scores on a benchmark. It does not know this, and structurally cannot — it is not holding the customer’s outcome.
Rows marked ? do not have enough evidence to trust — shown, but never ranked above a proven combination. Nobody holding a frontier model can compute this table, because nobody holding a frontier model is holding the customer’s outcome.
Defensibility
Why a customer cannot route around us once models commoditise.
The receipt
Settlement requires proof that physical work happened: verification, the routing trace, the episode. Nobody else holds it, because nobody else holds the customer’s outcome.
The priors
Which model, on which body, in which environment, with which recovery strategy, actually finishes paid work. A frontier lab knows its benchmark score. It does not know this, and cannot.
Portability
Skills declare morphology, never brands. A developer publishes once and runs on hardware invented after they shipped — which is why they publish here.
The switching cost is the fleet
Re-certifying safety envelopes across thousands of deployed machines is not a migration anyone undertakes for a few points of price.
The receipt, as a gate
Turn off any clause and watch settlement refuse. Each one blocks a specific way somebody could be paid for work that did not happen.
Integer cents, summing to gross exactly. Rounding remainder goes to us — a stray cent sent to a partner makes their statement unreconcilable, and that becomes somebody else’s accounting problem.
The split shown is illustrative. What is being demonstrated is the gate above it, not the percentages — those are set per agreement, and a published rate is a negotiating floor rather than a product feature.
Turn off any clause. This is why a customer cannot route around us once models commoditise: verification, the routing trace and the episode are the receipt, and nobody holding a frontier model is holding the customer’s outcome.
Note the order: none of these exist on day one. They are earned by the wedge year. The business model is not a plan to charge a take rate — it is a plan to become the party with the right to.
The arithmetic
What has to be true.
We price a percentage of physical labour performed by machines. Move the assumptions and see what the architecture must eventually support.
This is a specification, not a forecast. It says what the architecture and the market must eventually support, and therefore what we must not foreclose today. Two million verified jobs a day is roughly one job per machine per day across a fleet the size of the industrial robot base that already exists.
Now to December
The platform is built. What is left has lead times we do not control.
A robot ordered on 1 September arrives mid-October. A partner conversation opened on 1 September gets us on site in November. Both are longer than any remaining piece of software — which means December is decided by what gets ordered and who gets called, not by engineering velocity.