Chokepoint deep-dives

Precision Reducers: the joint problem

DUOPOLY · CRACKING · CHALLENGER ASCENDANT

THE THESIS

A humanoid moves at its joints, and inside each joint sits a precision reducer — ~20–40 per robot, among the most expensive single components in the build. Two Japanese firms historically made essentially all the good ones. The crack that isn't fully in the price: the same humanoid demand wave that re-rates the incumbents is funding the Chinese challengers taking the harmonic segment first.

THE NUMBERS THAT MATTER · MID-2026

• —Harmonic Drive Systems FY2025: sales +7% YoY, March-quarter orders at a four-year high; FY2026 guide ~14% growth.
• —Leaderdrive 2025: net profit more than doubled to ¥124.4M on revenue +47%; JPM puts its China harmonic share at 30–40%.
• —HDS + Leaderdrive together hold >63% of global harmonic revenue — the duopoly changed nationality at the #2 seat, it didn't break.
• —Shuanghuan's Suzhou plant targets ~500k reducer units/yr in 2026; Nabtesco is doubling RV capacity in defense.

THE EDGE

• 01Harmonic vs RV is the fork: humanoids lean harmonic at the distal joints — the segment with the most humanoid upside is also the one where the Chinese challenge lands first.
• 02Reducer purity matters: one incumbent is a near-pure reducer play, the other buries reducers inside a diversified conglomerate. 'The reducer duopoly' is not one trade.
• 03Challenger valuations carry their own risk: at least one Chinese name prices in near-flawless humanoid execution, on top of A-share access constraints.

TOUCHES

Planetary Roller Screws: the linear-actuator chokepoint

CONCENTRATED · LOCALIZING FAST · CHALLENGER ASCENDANT

THE THESIS

Reducers are the rotary chokepoint; the linear one is the planetary roller screw — arguably the most expensive single part in the robot. Tesla chose roller screws for Optimus's high-load joints because they survive the shock loads of walking, putting a low-volume, Western/Japanese-dominated component on the critical path of a million-unit production target. The whole story is a cost-and-localization race.

TOUCHES

NdFeB Magnets: the China lever

CONCENTRATED → CONTROLLED · CHINA-CONTROLLED UPSTREAM

THE THESIS

Every torque motor in every actuator needs an NdFeB permanent magnet. There is no humanoid without them, and no NdFeB supply chain that routes around China at scale. This is the single point where the robotics buildout touches a deliberate geopolitical lever — China pulled it in 2025. The trade is the small set of Western escape-valve producers, priced on policy as much as on tonnes.

TOUCHES

LiDAR: the China cost-curve chokepoint

CONCENTRATED · COST-CURVE MOAT · CHINA-CONTROLLED COST CURVE

THE THESIS

LiDAR is the one sensing node where China doesn't just compete — it sets the price floor. Hesai and RoboSense drove automotive LiDAR from five-figure exotica to sub-$1,000 parts, and the same cost curve is flooding into robotics. In Q1 2026 RoboSense's robotics LiDAR shipments surpassed its automotive shipments for the first time — the demand mix is tilting toward physical AI in real time.

TOUCHES

Frameless Torque Motors & the Inovance servo story

CONCENTRATED · CHALLENGER ASCENDANT

THE THESIS

Humanoids use two motor kinds: frameless torque motors for the big body joints and coreless motors for the dexterous hands. Frameless is the harder chokepoint — and it sits at the center of the most important architecture decision in the robot: precision joints (motor + harmonic reducer + torque sensor) versus quasi-direct-drive (beefier motor + cheap planetary, no sensor). That fork decides which other chokepoints even appear in the BOM.

TOUCHES

Six-Axis Force/Torque Sensors: the contested chokepoint

CONCENTRATED · ERODING · CHALLENGER — ARCHITECTURE-DEPENDENT

THE THESIS

A six-axis F/T sensor is the robot's sense of touch at the wrist and ankle. ATI and Bota dominate, China is entering hard on price — and the whole node carries a structural risk the others don't: quasi-direct-drive joints infer torque from the motor and skip the sensor entirely. The one Layer-2 chokepoint whose TAM is hostage to a design decision one layer up.

TOUCHES

Edge-AI Compute: the inverted chokepoint

CONCENTRATED · NVIDIA-LED · THE US HOLDS THE LEVER

THE THESIS

Every other chokepoint is a story of Japanese or Chinese dominance. Compute is the inversion: the US holds the choke — NVIDIA specifically — and export controls point outward at China. For the China-edge thesis this layer matters because sanctions create a forced-localization tailwind for Chinese SoC makers (Horizon, Black Sesame, Huawei Ascend) — the mirror image of the rare-earth lever.

TOUCHES

Tactile / E-Skin: the successor chokepoint

COMPETITIVE · IMMATURE · CHINA-LED ON SHIPMENTS, NO MOAT LOCKED

THE THESIS

The chokepoint that doesn't fully exist yet — which is why it's worth pre-mapping. E-skin gives the robot touch across fingertips and body, the missing piece for contact-rich manipulation. No incumbent owns it, no standard has set, and the current shipment leader is a Chinese full-stack player most Western investors have never heard of. The play is to know the names before they're consensus.

TOUCHES

Power: cells, C-rate, and the CATL double-play

CONTROLLED CELLS / COMPETITIVE ELECTRONICS · CHINA-CONTROLLED CELLS

THE THESIS

The battery chokepoint isn't energy density — it's discharge rate. A walking, lifting robot demands high C-rate from a light pack, a different spec from an EV's range chemistry. China owns this twice: it dominates cell manufacturing (CATL #1, BYD #2) and the high-rate know-how. CATL is simultaneously the supplier, a humanoid investor (Spirit AI), and a humanoid customer on its own lines. Of all six layers, power has the thinnest non-China escape route.

TOUCHES