Roller screws, not chips, are the humanoid ramp's tightest mechanical chokepoint — demand runs to 770% of assessed 2030 capacity

The planetary roller screw — the part that turns a spinning motor into a robot's linear push — is the tightest mechanical node in the humanoid value chain. At 14 screws per robot and 7.7 million units of 2030 demand against roughly 1 million units of assessed capacity, the order book points to utilisation near 770%. Integrated rotary actuators sit at 385%, six-axis force-torque sensors at 440%, and four of ten modelled mechanical nodes carry a deficit-risk verdict — according to InfraMosaic's supply/demand order book.

Most of the public anxiety about scaling humanoid robots fixates on silicon and software — the onboard AI, the data, the GPUs to train it. That is the loud bottleneck. The quiet one is a machined steel cylinder the size of a thumb. InfraMosaic's supply/demand order book — robot ramp multiplied by a per-robot bill-of-quantities, set against assessed node capacity — flags the planetary roller screw (PRS) as the part most likely to bind first. It is not a software problem you can fix with a better model; it is a metallurgy-and-grinding problem measured in six-to-nine-month lead times.

The mechanism is straightforward. A humanoid like Tesla's Optimus uses roughly 14 linear actuators, each built around one roller screw, to drive the limbs that need high force in a compact envelope. Unlike a ball screw, a roller screw distributes load across multiple threaded rollers, surviving the shock and dead-stop loads of a walking machine. There is no easy substitute. So per-robot demand for screws scales one-for-one with the linear-actuator count, and the order book translates the ramp directly into a units number.

The node-by-node tightness map

Run the same arithmetic across every mechanical node and the picture is not uniformly tight — it is surgically tight. A handful of parts dominate the risk while the headline-grabbing components are, on InfraMosaic's numbers, comfortably oversupplied. LiDAR units sit at about 5.5% of 2030 capacity; battery packs and rare-earth magnet tonnage are effectively a capital glut. The constraint is concentrated in three precision-mechanical buckets where Western and Japanese incumbents — Rollvis, GSA, Ewellix, NSK, THK — still hold the high-precision share.

*Magnet capacity is a glut, but the rare-earth input (dysprosium, terbium) sits under Chinese export control — the real magnet constraint is upstream, not at the press. The order book separates the two so the tightness signal is not double-counted.

Why screws bind before everything else

Three properties make the roller screw the first domino. First, the count: at 14 per robot it is tied with the actuator for the most numerous precision-machined part, so it inherits the full force of the ramp. Second, the process: roller screws need precision-ground alloy-steel threads and matched roller sets, a low-throughput, high-skill operation with multi-month lead times that cannot be conjured by adding a shift. Third, the supplier base: high-precision PRS output has historically been an EU/Japan specialty (Rollvis, GSA, Ewellix, Rexroth), thin and export-controlled, only now being chased by Chinese entrants such as Xinjian Transmission, which broke ground on a million-unit line in 2025.

That is also why screws matter to the cost story, not just the supply story. On InfraMosaic's reference bill of materials — rebased on Morgan Stanley's own Exhibit 54 — a roughly $55,000 material BOM splits as sensors $20.4k (37%), motors $11.2k (20%) and screws $11.1k (20%), ahead of reducers at 13%. Screws are a fifth of the build cost and the tightest node. Where they go, the unit economics follow.

China holds the cost-down template

Here is the part that should reframe the whole debate. The roller screw is not only the tightest node — it is the node with a proven path down the cost curve. In the Chinese supply chain the per-unit roller-screw cost has been tracked falling from roughly $3,000 to $800 at Tesla-scale volumes: a near-four-fold reduction driven by localisation, line throughput and grinding-process maturity. That is the template the rest of the BOM is expected to follow.

The whole-robot curve carries the same fingerprint. InfraMosaic's cost track has the Western build falling from $200,000 in 2024 toward $40,000 by 2030, while the China build moves from $46,000 to roughly $18,000 over the same window — a discount that starts near 77% and only narrows to about 55% as Western localisation catches up. The gap is not a labour-arbitrage story; it is a precision-components story, and the roller screw is its sharpest edge. A reader betting on humanoid hardware deflation is, in large part, betting on who industrialises the screw line.

What to watch

The order book is a forward map, not a verdict — capacity is the variable, and the deficit signal is precisely the thing that pulls investment in. Three near-term tells will show whether the 770% closes or bites:

Greenfield PRS lines reaching grade. Announcing a million-unit roller-screw line is easy; shipping C3/C5-class precision at yield is the hard part. Watch whether Chinese entrants convert groundbreakings into qualified, in-spec output — that is what actually moves the capacity denominator.

Actuator-maker vertical integration. With screws at 770% and integrated actuators at 385%, expect module builders to pull screw-making in-house to secure allocation. Vertical integration is the rational response to a single-point chokepoint — and the early warning that the node is binding.

The cost-down spreading. The $3,000→$800 screw curve is the leading indicator. If force-torque sensors — the #1 cost line at 37% of the BOM and a 440% deficit — begin a comparable descent, the humanoid unit economics inflect years ahead of consensus. If they don't, sensors and screws together keep the floor under the price of a robot. Either way, the mechanical stack, not the compute stack, is where the 2030 ramp is won or lost.