Own the Body, Rent the Brain?

A deployable humanoid robot now costs about $15,400. Japan Airlines put two of them to work at Haneda Airport this spring on a three-year commitment; the units, built on Unitree hardware, cost less than a single year of the wage they stand in for.Unitree / JAL^[1] That number is striking, because it crosses a line that has held since the first power loom. For two centuries, automation meant capital substituting for labor — but the labor that survived stayed an expense. A wage is a variable cost: you scale it up in a boom and shed it in a bust. When the machine that does the work costs less than a year of the worker it replaces, something subtler happens. Labor stops being a line on the income statement and becomes a line on the balance sheet. You no longer employ the worker. You own it.

This is the inverse of the move I traced in “The Landlord in the Loop.” Cognitive AI made labor more rented, not less: you summon intelligence by the token and owe nothing the moment you stop. Embodied AI swings the pendulum the other way. The same technological wave now travels in opposite directions on the capital-labor axis — renting cognition while re-capitalizing physical work. And that split surfaces a question every founder who touches the physical economy will face within a few years. Not whether to use embodied AI, but what, exactly, they will own when they buy it.

The Slogan and Its Appeal

The intuitive answer is a slogan: own the body, rent the brain. Buy the chassis — it depreciates like a forklift — and subscribe to the intelligence that drives it, the way you rent cloud compute. The appeal is genuine. A robot you own is collateral; it sits on your books; it can be financed, resold, and run as hard as you like. In a world where cognitive work has become a metered subscription to a handful of frontier labs, owning the means of physical production feels like getting leverage back — capital returning to the small firm after a decade of platforms collecting rent on everything that mattered.

The slogan is clean. It is also wrong about where the dependence lives, and seeing why requires a forty-year-old observation about machine intelligence.

What Moravec Knew

Hans Moravec noticed in the 1980s that the things humans find hard — symbolic reasoning, chess, calculus — are easy for computers, while the things humans find effortless — walking, grasping, reading a face in a crowd — are computationally brutal.^[2] Moravec's paradox has aged into the central fact of robotics. The high-level reasoning we associate with intelligence is the cheap part, and open-weight models are making it cheaper by the token. The expensive frontier is sensorimotor control: holding balance on a wet floor, closing a hand around a deformable object, recovering from a stumble. That competence is exactly the kind of skill Michael Polanyi had in mind when he wrote that we know more than we can tell.^[3] And it carries a property the planner does not: it is latency-bound. Routing a balance-correction loop through a data center two hundred milliseconds away is, for many tasks, simply too slow, so the controller still has to live on the body.

So the brain is not one thing. There are three layers, not two. The body — chassis and actuators, the atoms, increasingly cheap and increasingly Chinese. The controller — the on-board sensorimotor policy, latency-critical, integrated, and the actual Moravec frontier. And the planner — high-level task reasoning, cloud-capable, modular, and the layer open weights are commoditizing. The slogan folds the controller and the planner into a single “brain,” and that fold hides the whole game.

Here is what follows. The rent does not disappear when the planner goes open; it migrates to the layer that stays hard. A founder buys the body, downloads a capable open-weight planner for nothing, and finds the robot still will not reliably unload a truck without the manufacturer's proprietary control stack and its over-the-air updates. You own the body. The planner is nearly free. And you are locked in precisely at the seam between them — the controller — which Moravec's paradox guarantees will stay the hard, defensible, rentable layer for years.

Own the body, rent the brain understates the problem. You own the body, you are handed the easy half of the brain, and you rent the half that matters.

The Architecture Fork

This relocation sets up the question that will organize the industry: does the controller stay welded to a single vendor's stack, or does it become a standard anyone can build against?

One path is integrated — the Apple model. Because the controller must fuse to the body, the natural firm boundary wraps body and controller into one product, and you lease the bundle. The landlord here is the manufacturer, not the AI lab; Tesla and Figure are building exactly this stack. The other path is modular — the Android model. Open interfaces, commodity bodies, planners and controllers you mix and match. The force pushing hardest in that direction is geopolitical. China owns the body: Unitree's price point, the actuators, the rare-earth supply chain Beijing still controls. Chinese open-weight models, in parallel, commoditize the planner. Push cheap commodity hardware and open cognition together, and the pull is toward an open stack in which “own the body, rent the brain” becomes nearly literal and the rent collapses at two of the three layers.

If I am forced to guess, integration still wins in the near term. Moravec is on its side, and so is a quieter mechanism I will discuss below. But that judgment turns on three tensions that are arguably more useful to discuss than the forecast. The first is whether the controller standardizes: whether an “Android for robots,” a shared control layer good enough that bodies and planners can be built against it, actually emerges. The second is the Chinese open-plus-hardware vector, the one combination with both the manufacturing scale and the open-weight commitment to pry the stack open from below. The third is the latency frontier itself: if on-device compute and better simulation-to-reality transfer make a strong controller cheap to train and run locally, the layer that anchors integration loses its grip. Move any one of the three and the near-term call inverts. I am not predicting that integration wins; I am saying that is where the weight currently sits, and pointing at the levers that would move it.

The Flywheel That May Decide It

There is a mechanism beneath the architecture fight, and it is why the near-term trends continue to favor integration: the controller improves with use — every deployed robot's sensorimotor experience becomes training data for the next policy. Whoever runs the largest fleet trains the best controller, sells more robots, and grows the fleet again. The controller is not merely hard; it is a data flywheel, and flywheels reward scale and the vertical integration that captures the data cleanly.

Autonomous driving is the same argument run a decade ahead, and the two leaders illuminate different halves of it. Tesla made the purest flywheel bet: millions of consumer cars running its software, a commodity-camera stack, and one end-to-end controller trained on the largest data fleet on the road. Waymo made the opposite bet — a smaller fleet of sensor-heavy, lidar-equipped cars confined to mapped, geofenced cities. By mid-2026 the bounded bet leads where it counts. Waymo runs roughly thirty-eight hundred vehicles and half a million paid driverless trips a week across eleven citiesWaymo^[4]; Tesla, despite the larger data fleet, fields only a few dozen driverless robotaxis in a single Texas market and still stumbles on the long tail — school buses, emergency vehicles, pedestrians.^[5] Data scale alone has not closed the Moravec gap. Restricting the domain has. The bounded bet is not flawless — Waymo paused service in several cities this spring after its cars drove into standing water, despite a recall meant to prevent exactly that — but bounding the problem is still what has produced reliable driverless operation at all. The lesson for embodied labor is direct: the first reliable robots will work in bounded, mapped, structured spaces — the warehouse, the airport gate where Japan Airlines put its humanoids — long before they work anywhere at all. The flywheel still turns; it just turns fastest inside a fence.

That turns the ownership question inside out. Tesla already runs the play in miniature: owners pay for the software, and their everyday driving trains the controller they depend on — the asset improving fastest is the manufacturer's, not the driver's. So when you “own” a humanoid, ask who owns the sensorimotor data it generates on your floor. If that telemetry flows back to the manufacturer to train the controller you depend on, you have bought a machine that makes your landlord stronger every shift you run it. My colleagues and I have named the broader pattern digi-serfdom — value accruing to whoever controls the rails rather than the operator who does the work.^[6] Embodied AI gives it a literal form by pointing to an asset that appreciates for the platform and depreciates for you.

When you buy a robot whose data flows back to its maker, you have bought an asset that appreciates for the platform and depreciates for you.

What This Changes for the Founder

The practical consequence is that “make or buy” is no longer one decision. It is a decision per layer. You will almost certainly own the body, increasingly rent or freely run the planner, and the live question — the one that sets your margins and your bargaining power — is the controller. Own it, which is rare and hard, and you have a moat. Rent it on good terms and you have a business. Rent it on bad terms and you are a franchisee of the firm that sold you the robot.

Step back, and the wave splits the entrepreneur in two. In bits, cognitive AI lowers the barrier to entry: a solo founder rents intelligence cheaply and does the work of five. In atoms, embodied AI raises it: you must own and finance the capital up front rather than scaling a workforce as revenue arrives. The same technology democratizes the idea business and re-concentrates the physical one. And the tax code is quietly pressing its thumb on the scale — wages are expensed and taxed while capital is depreciated and shielded, so owning the machine is advantaged over employing the person before the sticker price even lands. I don't have much of an appetite for wading into policy debates, but that is the one policy fact worth scrutinizing in more detail given the larger questions it raises about the balance between capital and labor in the tax code.

Despite these advantages tilting the playing field toward capital, one important caveat that cost curves tend to hide is that humans are still underrated for many physical tasks. Humans not only act; they notice anomalies. They exercise judgment and handle the exception. Replace them with owned capital and you can buy the doing while losing the noticing. Whether embodied AI becomes a replacement or a partner rides on the same architecture fork: an integrated robot that closes the human judgment loop is substitution; a modular one that keeps a person in the controller's oversight loop is an ensemble — a cyborg arrangement in which the human keeps the sensing and the machine takes the strain. The architecture you buy is also a decision about how much of your own organization's intelligence you are willing to stop using.

What to Watch

The cost crossing is real and it is not reversing; Unitree alone is aiming from a few thousand units last year to the tens of thousands this one.^[7] The question is no longer whether physical labor gets re-capitalized but on whose terms. Watch the controller. Watch whether an open standard forms there or a few integrated vendors hold it closed. Watch whether China's pairing of cheap bodies and open models sets the world's default architecture the way Android set the phone's. And watch the contract language about data, because that is where ownership is quietly decided.

Own the body, rent the brain was always too simple. Reality is much more interesting. You will own the body. You will be handed the easy half of the mind. And the terms on which you get the hard half will decide whether you have become a capitalist again — or a tenant with a heavier balance sheet.