Sensors are the most expensive thing in a humanoid robot. At an assessed $20,350 — 37% of a ~$55,000 material bill of materials — the sensing stack outweighs actuator motors and screws, each ~20%, and dwarfs compute at 4%. The six-axis force-torque sensor anchoring that line is also the supply chain's tightest knot: 2030 demand runs to 440% of assessed capacity. That corrects the "joint modules are 60–70% of cost" framing in Morgan Stanley's own Humanoid 100 — contradicted by its Exhibit 54 — according to InfraMosaic's humanoid reference BOM and supply/demand order book.
The conventional shorthand for humanoid economics is the actuator. Morgan Stanley's influential Humanoid 100 (February 2025) put it bluntly in prose: "joint drive modules are 60–70% of a humanoid's cost." That framing has propagated through sell-side notes and supplier pitches ever since, training investor attention on reducers, roller screws and frameless motors as the parts that matter.
It is the wrong cut of the data — and the corrective sits in the same report. Morgan Stanley's Exhibit 54, a component-level teardown of Tesla's Optimus Gen-2, ranks the bill of materials as: Sensor 37%, Motor 20.3%, Screw 20.2%, Reducer 12.6%, Encoder 3.9%, compute 3.8%, bearing 0.8%, battery 0.5%. Sensing leads. The "60–70%" figure is real, but only at the sub-system cut — once you bundle motors, screws, reducers, encoders and bearings into a "joint module," that envelope dominates. Slice the BOM the way a procurement team actually buys it — by component — and the single largest line you negotiate is the sensor.
The bill of materials, line by line
InfraMosaic re-bases its humanoid reference BOM directly on Exhibit 54, mapped onto a ~$55,000 material build. The breakdown, in dollars:
| Subsystem | Assessed cost | Share of BOM |
|---|---|---|
| Sensors — 6-axis force-torque + IMU + encoders + vision + tactile #1 | $20,350 | 37% |
| Motors — frameless + BLDC actuator motors | $11,165 | 20% |
| Screws — planetary roller + ball screws | $11,110 | 20% |
| Reducers — harmonic + cycloidal | $6,930 | 13% |
| Encoders — joint position feedback | $2,145 | 4% |
| Compute — SoC + chips + camera ISP | $2,090 | 4% |
| Bearings — cross-roller + ball | $440 | 1% |
| Battery pack — 2.3 kWh | $275 | 1% |
| Others — frame, structure, wiring, power electronics | $495 | 1% |
| Material BOM total | ~$55,000 | 100% |
Source: InfraMosaic reference BOM, re-based on MS Humanoid 100 Exhibit 54 (Optimus Gen-2). Total resolves to ledger figure kpi.bom_material_total_usd. Shares rounded; columns may not sum to 100.
Three things stand out. First, sensing is not one part — it is a stack: the six-axis force-torque sensors in the wrists and ankles, MEMS IMUs, joint encoders, vision and an emerging tactile layer. Second, the cost density sits in force-torque: InfraMosaic assesses a six-axis F-T sensor at ~$6,000 a unit (range $4,000–$10,000), with four per robot in the hands and feet — the single most expensive discrete component class in the body. Third, the cheap stuff is genuinely cheap: a 2.3 kWh battery pack is $275, an automotive-grade MEMS IMU is a few dollars. The robot's intelligence about the physical world, not its battery, is what runs up the bill.
Slice the bill of materials the way a buyer actually negotiates it — by component — and the single largest line is the sensor, not the actuator.
The same part is also the tightest bottleneck
Cost concentration would be a curiosity if supply were elastic. It is not. InfraMosaic's supply/demand order book — robot production ramp multiplied by per-robot bill-of-quantities, set against assessed manufacturing capacity — flags the six-axis force-torque sensor as one of the three deficit-risk nodes for 2030.
At 440%, demand outruns assessed capacity more than fourfold. The supplier base is thin — ATI, Kunwei and Hypersen sit at the top of the order book — and the technology does not scale like a commodity sensor. A six-axis cell must resolve forces and torques across three axes simultaneously, with the temperature stability and crosstalk rejection that lets a robot grade its grip on a fragile object. That is a precision-machining and calibration problem, not a lithography one. The cost is falling — InfraMosaic tracks the F-T sensor on a ~44% cost-down trajectory — but the capacity to make them is the gating variable.
It is not alone. The order book's other 2030 deficit-risk verdicts run alongside it:
- Planetary roller screws — 770% of assessed 2030 capacity. The single tightest node in the body.
- Integrated rotary actuators — 385%. The "joint module" the old narrative obsessed over, also short.
- Six-axis force-torque sensors — 440%. Top cost line and a top-three bottleneck at once.
The pattern is the point. The parts that cost the most are, disproportionately, the parts the world cannot yet build at scale. The glut sits elsewhere — LiDAR at 5.5% of capacity, battery packs near zero — exactly the components a generic "EV supply chain" instinct would worry about. For humanoids, the binding constraints are the precision-mechatronic nodes, and force-torque sits at the intersection of dearest and scarcest.
Why the framing matters
The cost curve is moving fast underneath all of this. InfraMosaic's whole-robot BOM curve has the Western build falling from ~$200,000 in 2024 to ~$40,000 by 2030, and the Chinese build from ~$46,000 to ~$18,000. As the actuator drivetrain commoditises — the planetary roller screw alone is on a path from roughly $3,000 to $800 at Tesla scale — the sensor share of the remaining bill only grows. Getting the largest cost line wrong in 2026 means mis-pricing the bottleneck that will still be binding in 2030.
For investors and procurement teams, the re-frame changes where to look. The "joint module 60–70%" story points you at motors, reducers and screws — a crowded, increasingly Chinese, increasingly commoditised field. The component story points you at the force-torque sensor: dearest line, thinnest supplier base, hardest to scale, and the node where a 440% deficit gives incumbents real pricing power through the decade. Morgan Stanley's report had the right number in its own exhibit; the prose buried it.
Rystad asks you to trust the trace. InfraMosaic lets you check it.
Every figure in this article resolves to a hash-chained record in the InfraMosaic Publication Ledger — input file, SHA-256, methodology version, test suite and named approver. The headline BOM figure is logged as:
content_hash abb1125afeddf999…7872c014e
ledger v1 · 363 records · chain_head 15a725e0…ee41c29d · 134/134 tests passed
What to watch
- Force-torque capacity announcements. Any meaningful greenfield from ATI, Kunwei, Hypersen or a new entrant is the single most important signal for whether the 440% deficit eases before the 2030 ramp.
- The sensor cost-down rate. If the F-T unit price falls faster than ~44%/yr, the sensor share of BOM compresses and the bottleneck loosens together. Slower, and sensing's cost lead over motors widens further.
- Tactile-skin integration. A full tactile layer would push the sensing share above 37% — the one move that could extend, rather than erode, sensing's #1 position.
- Whose narrative the market adopts. Watch whether sell-side BOM models migrate from the "joint module 60–70%" sub-system cut to the component cut. The re-rating of the force-torque supply base depends on it.