The Actuation Bottleneck

Section 01The cells of robotics

In 2025, somebody took a Unitree G1 apart.

The robot had been argued about for a year — its gait, its price, the videos. The teardown told a different story. Strip the shell and the value sat in one place: the joints came to roughly two-thirds of the bill of materials. Everything the industry had been discussing, the controller and the perception stack and the demo footage, was riding on top of a pile of gearboxes.

That is the shape of the whole category, and it is not a new shape. We have seen this film, with a different machine.

Ten years ago, the electric car was a software story

The pitch was the software-defined vehicle: over-the-air updates, autonomy, the car as a platform. Then the decade happened, and it was not decided by software. It was decided by who could buy cells.

Battery packs settled at roughly a third of vehicle cost. Energy density improved a few percent a year rather than doubling on a cadence. Supply concentrated in a handful of Asian manufacturers with multi-year expansion lead times. Every credible forecast of cheaper EVs turned out to be a forecast about cell manufacturing, not about cars. And the company that took the category built a battery factory before it built a mass-market vehicle — because it had worked out where the constraint lived.

The parallel holds line by line. Dominant share of the bill of materials. Improvement governed by materials science rather than lithography. Supply concentrated in a small number of specialists, increasingly Asian. Cost-down roadmaps that are really supply chain roadmaps. And vertical integration by exactly the players with the best information — Tesla designs its actuators in-house, and Schaeffler bought a roller screw manufacturer while taking an equity position in a humanoid company.

Where the analogy breaks — and why that is the bad news

Batteries, for all their difficulty, converged. The industry settled on a handful of cell formats. It agreed on what the numbers mean: watt-hours per kilogram, cycle life to 80%, C-rate. An engineer can compare two cells from two continents on a spreadsheet, because the specifications are defined against shared test conditions.

Actuation has not done this. There is no standard joint format, no manufacturer part number convention, and no shared rating basis. Continuous torque is quoted against undeclared thermal assumptions; load ratings against undeclared life definitions. Two suppliers can publish identical headline figures and mean materially different things.

So robotics has the battery problem without the battery industry's legibility. The constraint is mechanical, concentrated and slow to improve — and, unlike cells, you cannot yet put two of them side by side and tell which is better.

The rest of this report sets out what that costs, in money, in reliability, and in the design freedom it takes away.

Section 02The attention asymmetry

Robotics raised $55.8 billion in 2026, nearly double the previous record, on a narrative about foundation models and vision-language-action policies. Read the funding announcements and you would conclude the hard problem is intelligence.

Read the bills of materials and you conclude something else.

Compute, sensing and the model together are a minority of the hardware cost, and a smaller minority every year as silicon gets cheaper and steel does not.

This is not a humanoid curiosity. It is the same shape in a cobot arm, an exoskeleton and a quadruped. Actuation is where robots keep their cost, their mass, their failure modes and their physical limits.

Section 03There is no Moore's law for torque

The reason the industry's instincts are wrong here is that everyone's mental model of technology cost curves comes from semiconductors. Transistor density improved by orders of magnitude on a predictable cadence. Software inherits that curve for free.

Electromagnetic actuation does not. Torque in a rotary machine scales with rotor volume and the magnetic shear stress at the air gap. Shear stress is bounded by the flux density of available magnet material and the current density the winding can carry without cooking its insulation. Neither has improved by an order of magnitude in decades, and neither will.

The consequences are unfamiliar to anyone trained on silicon:

• Torque density is roughly constant for a given topology. You cannot make a joint twice as strong without making it substantially larger, heavier, or hotter.
• Cost falls with volume, not with physics. Every forecast of cheaper humanoids is a forecast about manufacturing scale and supplier competition, not about a technology curve.
• Mass compounds. A heavier joint requires a stronger joint above it, a bigger battery, and a stronger structure to carry both. Actuator mass is the term that drives the whole spiral.

Which means the cost-down trajectory the industry is forecasting — Bank of America's China-built BOM falling from roughly $35,000 in 2025 to below $17,000 by 2030, Bain's $10,000–20,000 by 2035 — is, in substance, an actuator supply chain forecast wearing a robotics label. It will be delivered or missed by reducer, screw and motor manufacturers, not by model labs.

Section 04The multiplication problem

A six-axis industrial arm has six joints. A NEURA 4NE-1 has 42 degrees of freedom. A 1X NEO carries 22 in each hand.

Joint count is a multiplier on everything: cost, mass, wiring, thermal load, assembly labour, and — most punishingly — failure probability. If joints fail independently, robot-level reliability is per-joint reliability raised to the power of joint count.

This is the quiet reason humanoid fleets are hard in a way that pilots are not. A demonstration needs one robot to work for an afternoon. A fleet needs thousands of joints to work for years, and the arithmetic is unforgiving: joints that would be entirely acceptable in an industrial arm produce an unserviceable humanoid.

It also explains a commercial behaviour that otherwise looks irrational — why robot builders qualify actuation components so slowly and switch suppliers so reluctantly. At 42 joints, a supplier change is not a procurement decision. It is a bet on the reliability of the whole product.

Section 05The thermal wall

Continuous torque — how hard a joint can push indefinitely, rather than for a second — is not a magnetic limit. It is a thermal one. The motor can produce more torque than it can survive producing.

Because losses in the winding rise with the square of current, and torque rises roughly linearly with it, doubling continuous torque requires roughly four times the heat removal. In a test fixture that is a larger heatsink. In a robot joint the heatsink is a structural link that also has to be light, and which sits next to the encoder and the electronics that degrade when hot.

Two consequences follow, and they explain a great deal of robot design.

Published continuous torque figures are not comparable between manufacturers unless the heatsink, ambient and permitted winding temperature rise are stated, and they usually are not. This is the central finding of our first report and it originates here.

Robots are designed around duty cycle, not peak capability. A humanoid that can lift 100 kg momentarily may not be able to hold 15 kg for ten minutes. Specification sheets quote the first number; deployments are limited by the second.

Section 06Where the cost actually concentrates

Within actuation, the money is not evenly spread. Locomotion hardware — thigh, calf and feet — is roughly $21,000 of a $55,000 Optimus-class BOM, the single largest cluster. Legs account for around 39% of the BOM and hands around 17%.

The components underneath are precisely the ones our supply map identifies as most concentrated. Morgan Stanley reports planetary roller screws at $1,350–$2,700 each, with a full-size humanoid carrying around a dozen. That is $16,000–32,000 of screws in the legs of a single robot — from a component family made by a small number of specialists worldwide, and for which we found only 29 manufacturers globally.

Section 07What the incumbents are doing about it

The clearest evidence for the thesis is behavioural. Firms with the most information are integrating vertically into actuation.

Tesla designs its actuators in-house — a decision that makes sense only if actuation is the dominant cost and the dominant risk. Schaeffler acquired Ewellix, a roller screw manufacturer, in 2023, and holds an equity position in NEURA Robotics; it now sits on both sides of the transaction. Chinese entrants are attacking the reducer and screw categories directly, with at least one publishing a list of seven European and American brands it intends to displace, and a roller screw maker filing for IPO guidance on a line rated at a million units a year.

Nobody is vertically integrating into vision models. They are integrating into gearboxes.

Section 08What follows

For robot builders. Your cost curve, your reliability and your ship date are set by actuation decisions made early and changed slowly. Treat actuation supply as a strategic function rather than a purchasing one, and establish second sources before the ramp rather than during it. At 42 joints, the qualification calendar is your product calendar.

For investors. Diligence on a robotics company that does not examine the actuation BOM, supplier concentration and qualification status is diligence on the minority of the hardware cost. The question is not whether the model works in a demonstration. It is whether 42 joints can be bought, at a price, reliably, for years.

For component manufacturers. Demand in several of these families is about to be set by an application that did not exist three years ago, on volumes an order of magnitude above historic norms. The firms that publish comparable specifications and honest lead times will be designed in. The ones that quote on application will be designed out, because an engineer who cannot evaluate you cannot specify you.