Who owns the robots, the power, and the compute

The industry has moved past whether a machine can walk. The question inside a buying committee is whether anyone can afford to build it at volume and keep it working.

Humanoids get the photograph, but the work—and the cost—is in the stack. Actuators alone account for roughly 56 percent of a humanoid’s bill of materials: the harmonic drives, planetary roller screws, strain gauges, and brushless motors where heat and failure actually live. Structure and battery packs take another 26 percent; sensors and edge compute take the rest. A precise joint does not get cheap because a demo looked fluent.

On the factory floor, the sticker price is secondary to utilization. An industrial lease at $8,500 a month is $51 an hour on a single eight-hour shift—well above the warehouse wage the robot is meant to stand beside. Only when a machine runs three shifts—roughly 6,000 hours a year, without waiting on a charger or a human reset—does the math drop toward $17 an hour. Western labs currently lead on world models. The metal is a different contest, run down Tier-1 automotive supply lines. We report on who makes the joints, who placed the firm order, and what is actually on a customer floor versus what is still a memorandum of understanding.

The models behind physical AI are being bought at a scale that no longer looks like ordinary corporate spending, with hyperscaler capital expenditure tracking toward $1 trillion and single frontier campuses pricing near $41 billion per gigawatt.

But compute spent on internet text and video has a ceiling. A language model can describe a doorway; a robot has to pass through it without taking the frame along. A model trained on video can look fluent and still put a foot through wet snow, because a video frame does not contain friction, glare, or contact force. Closing that gap requires tactile failure data recorded in the physical world, and edge silicon small enough to run vision-language-action models without waiting on a distant data centre.

A data centre cluster or an automated fleet is only as real as the megawatts behind it. Compute schedules are now governed by gas turbine bookings, transformer lead times, grid interconnection queues, and cooling water.

Alongside physical power, we track machine infrastructure and decentralized physical networks (DePIN)—strictly when they change what a machine can do. Our filter is simple: if you remove the token and the price chart, does a robot, a fleet operator, or a data buyer still have a reason to care? A machine identity tied to a secure element that survives a repair or an ownership transfer is infrastructure. A simulated door in a press release is not.