If robotics is the next intelligence explosion, then actuators are the detonator.
You can have the smartest foundation model in history, but if it can’t move a finger with enough speed, precision, and force, it’s just a very expensive paperweight. The dirty secret of 2025 is that the software is already leaping ahead of the hardware. The gap is closing, but only for the handful of players who can secure enough of the one component nobody outside the industry has ever heard of: the strain-wave gear, also known as the harmonic drive.
Let’s start with the brutal numbers.
A single high-end harmonic drive reducer for a humanoid-sized joint (say, a shoulder or knee that needs to handle 50–80 Nm of torque at 1:100 reduction ratio with zero backlash) costs $1,200–$2,500 today in volume. A full humanoid has roughly 28–40 actuated degrees of freedom. Do the math: $40,000–$80,000 of the BOM is just gears. That is more than the batteries, more than the motors, more than the computers in most designs. Tesla’s Optimus Gen 2 hand alone has 22 actuators and 11 of them are driven by tiny harmonic cups the size of a shot glass that still cost $300–$500 each.
Now multiply by millions of robots.
Goldman Sachs and Morgan Stanley both published reports in late 2025 estimating that 10 million humanoids shipped by 2035 would require roughly 300–400 million harmonic drive units. Global production capacity in 2025 is approximately 1.8 million units per year, almost entirely controlled by one Japanese company: Harmonic Drive SE (market cap ~$4B, 70 % global share in precision strain-wave gears).
Yes, the entire humanoid revolution is currently bottlenecked by a 68-year-old invention from 1957 that looks like a metal donut with teeth on the inside.
How a quick physics primer
A strain-wave gear works by flexing a thin metal cup (the “flexspline”) with an elliptical wave generator. This creates a moving zone of tooth engagement that gives you enormous reduction ratios in a very small, backlash-free package. It is mechanically elegant, incredibly reliable (some units from the Apollo era are still running on the moon), and infuriatingly hard to manufacture. Tolerances are measured in single-digit microns across the entire circumference. A speck of dust in the wrong place ruins the batch.
For decades this was a niche product sold to aerospace and semiconductor equipment makers at insane margins. Then the Chinese entered the market. In 2020, Leader Harmonious Drive Systems (Shenzhen) and Laifual Drive (Zhejiang) started shipping clones at 30–50 % lower cost. Quality was spotty at first, but by 2025 the best Chinese units are within 5–10 % of HDSE performance while costing half as much. Capacity is exploding: Leader alone is targeting 2 million units in 2026 and 5 million by 2028.
The race is now fully geopolitical. Japan is trying to keep the crown jewels (the highest-precision, zero-backlash variants needed for robot hands), while China is happy to flood the market with “good enough” 5–10 arc-second units for legs and arms. Meanwhile, everyone else is scrambling.
The actuator land-grab map – December 2025
- Harmonic Drive SE (Japan): still the gold standard, but capacity capped at ~2M/year. Stockpiling by Boston Dynamics, Agility, and Figure is rumored to have taken most of 2026 production already.
- Leader Harmonious + Laifual (China): combined ~1.5M units/year now, scaling aggressively. Tesla and Chinese EV giants are signing 10-year offtakes.
- SPINEA (Slovakia): high-quality European alternative, tiny volume, mostly going to KUKA and ABB.
- Cone Drive (USA): bought by Timken in 2018, trying to restart domestic production. Current output: <50k units/year.
- New entrants:
– Genesis Robotics (Canada) – revolutionary “LiveDrive” direct-drive actuator, no gearbox at all. Raised $200M, but still pre-volume.
– Kubo Robotics (Japan) – magnetic geared motors, promising but stuck in prototypes.
– Tesla in-house – Elon tweeted in November 2025 that “we will make our own harmonic drives if we have to.” Vertical integration threat is real; Tesla already casts its own gigacastings and winds its own motors.
Beyond harmonic drives: the other chokepoints
- Frameless torque motors – maxon (Switzerland) and Kollmorgen (USA) still dominate the high-performance segment. China’s Topband and Inovance are catching up fast.
- Ballscrews and linear guides – THK, NSK, HIWIN. Again, Japan/Taiwan oligopoly.
- Encoders – almost all high-resolution optical encoders still come from just four companies (Heidenhain, Renishaw, Canon, Mitutoyo).
Every single one of these is seeing 3–5× demand growth and lead times stretching to 18–24 months. Startups are literally flying executives to Japan with briefcases full of cash trying to lock in supply.
The Saudi Arabia play
The country or company that breaks the harmonic-drive cartel wins the 2030s. The prize isn’t just selling $2,000 gears; it’s capturing the economic rent on every hour of robotic labor forever. If you control the actuators, you control the physical leverage of AI.
China seems to understand this best. Beijing designated “high-precision reducers” as one of the 35 “choke point” technologies in the Made in China 2025 plan and has been pouring subsidies in ever since. The result: domestic market share in harmonic drives went from <5 % in 2018 to >40 % in 2025, and rising.
What happens when the bottleneck breaks?
Assume (conservatively) that global harmonic drive capacity reaches 20 million units/year by 2028 and average price falls to $400–$600 in volume. Suddenly the actuator cost for a full humanoid drops from $60k to $15–20k. That is the moment the hockey stick goes parabolic. Every marginal $10k drop in robot BOM translates to millions of additional units sold, which creates more data, better software, lower servicing costs, and another $10k drop. Classic Wright’s Law on steroids.
We are one or two breakthrough years away from actuators becoming boring commodities instead of exotic artifacts. When that day comes, the limiting factor shifts from “can we build the joints?” to “how fast can we mine copper and rare earths?”
Next post: “The Data Flywheel That Actually Matters – Why Embodiment Is Eating Vision-and-Language Models for Breakfast.”
We’re still just scratching the surface.
Stay strapped in.