Tesla, Figure and Apptronik All Shipped Humanoid Robots This Year. The Same $200 Part Is Slowing Down All Three.
Three rival humanoid-robot programs began shipping to real customers within the same quarter of 2026 β a first for the industry. All three lean on the same scarce component: a thumb-sized precision gearbox, made almost entirely by a handful of Japanese manufacturers, that lets a small motor move a heavy limb without slipping. By mid-2026, the wait for one had stretched past six months.
Race to fill one robot's 40 joints: today's line vs. the 2027 fix
A simplified model, not a literal production schedule. Blue tiles are joints filled by today's precision-machined gearbox line (Line A). Unlock Line B to simulate Schaeffler's 2027 formed-gearbox process β dramatized here at roughly 8Γ Line A's pace, a stand-in for its "seconds instead of minutes" claim rather than an audited throughput figure. Hit Run and watch the backlog behave differently.
Press Run to simulate production, one week at a time, toward outfitting one 40-joint humanoid robot.
The plain version
Picture what it takes for a robot to lift its arm smoothly, hold a box steady, and stop exactly where it means to β without a motor the size of a car engine. Engineers solve this with a strange little gearbox called a "harmonic drive": instead of ordinary gear teeth, it uses a thin, flexible steel cup that bends slightly as it spins, meshing with a rigid ring around it. That flex is the whole trick β it turns a small, cheap, fast-spinning motor into a slow, extremely precise, very strong twist, packed into a space no bigger than a fist. It's the same basic idea as a bicycle's gears trading speed for power, just shrunk down and made nearly frictionless.
A humanoid robot needs 30 to 40 or more of these joints to move like a person β shoulders, wrists, hips, a flexible spine. And almost none of the companies racing to build humanoid robots make their own gearboxes. A tiny handful of manufacturers, mostly in Japan, make nearly all of them. This year, as Tesla's Optimus, Figure's robot and Apptronik's Apollo all started shipping to real customers for the first time, orders overwhelmed that supply. Wait times for a single gearbox stretched past six months. Elon Musk himself said the robot ramp-up would be "long and flat," in part because, unlike a car, a robot can't lean on a supply chain that already exists.
One company is trying to fix this the way factories usually break a bottleneck: change how the part is made. In August, the German manufacturer Schaeffler showed off a way to stamp these gearboxes into shape with high-pressure forming instead of slowly machining them β cheaper, faster, using a fraction of the raw metal. It won't reach mass production until 2027.
That's the twist: humanoid robots look like a software and AI story. In practice, whoever can get their hands on enough of one obscure, thumb-sized metal cup gets to actually ship robots this year β everyone else waits.
The expert version
The component is a strain-wave (harmonic) gearing system: a flexspline β a thin-walled, elastically deformable cup with external teeth β is driven by an elliptical wave generator that forces it into contact with a rigid, internally toothed circular spline at two opposing points. Because the flexspline carries slightly fewer teeth than the circular spline, each full rotation of the wave generator advances the flexspline by only that tooth-count difference, yielding single-stage reduction ratios roughly in the 30:1-320:1 range with near-zero backlash in a housing a fraction the size and weight of an equivalent planetary or worm-gear assembly. That combination of high reduction, high torque density and a compact envelope is why harmonic drives β rather than planetary gearboxes, which run up to 80% cheaper but at 95-98% efficiency versus 85-90% and in a bulkier package β dominate humanoid robots' rotary joints: shoulder, wrist roll, torso yaw.
Manufacturing one is unusually precision-intensive: flexspline wall thickness and tooth geometry are held to micron-level tolerances, since the part must survive millions of elastic-deformation cycles without fatigue-cracking. That's kept the segment concentrated in a small number of specialists. Japan's Harmonic Drive Systems holds an estimated 85% share of the strain-wave gearbox segment specifically; Nabtesco holds roughly 35% of the separate market for RV (cycloidal) reducers used in larger robot joints. China's Leaderdrive, Suzhou Green Harmonic and Zhejiang Shuanghuan (adding a 500,000-unit-capacity plant in Suzhou this year) are contesting the lower-precision tiers. By mid-2026, industry lead times on these gearboxes had stretched to an estimated 26 weeks or more.
Actuators built around these gearboxes account for an estimated half of a humanoid robot's total bill of materials, at an audited per-joint price near $200 β meaning a 30-40 DoF robot's gearbox-and-motor stack alone can run several thousand dollars before sensors, batteries or compute. Nabtesco is targeting doubled reducer capacity by year-end 2026 to meet demand from Tesla, Figure and Agility-class programs. Schaeffler introduced a stamped, high-pressure-formed strain-wave gearbox in August 2026 that it says cuts unit manufacturing cost over 25% and raw-material input over 75% by replacing subtractive machining with net-shape forming β producing key components in seconds rather than minutes on a conventional line. Mass production isn't scheduled until 2027, leaving 2026's supply constraint largely intact through the ramp Optimus, Figure 02 and Apollo are all attempting at once.
Why it matters for tech + supply chain: a robot's intelligence gets the headlines, but the machines actually shipping this year are bottlenecked by a decades-old mechanical part almost nobody outside factory-automation circles had heard of.
Why it matters for tech + supply chain: humanoid robotics is running the AI-chip and EV-battery playbook again β a fast-scaling demand curve colliding with a slow-scaling, geographically concentrated precision-manufacturing base with multi-year capacity lead times.