Harmonic Drives

Harmonic drives look comfortable on their announced curve, but once you account for the grinder pool they share, the wall arrives earlier and bites harder. Today the world builds roughly two million robot-grade strain-wave reducers a year⁴, and announced output is set to climb toward about six and a half million by 2030 as Japanese and Chinese lines expand⁴. On paper that keeps pace with the worldwide humanoid fleet’s harmonic demand³ until the mid-2030s (the figures here track the whole global humanoid fleet, not one country’s), but harmonic flexsplines share one finite precision-grinding pool with roller screws and cross-roller bearings⁷, and that pool throttles each part hardest through the late-2020s ramp, down to roughly sixty percent of its announced curve, then eases back toward eighty percent by 2040 as the curves self-limit at their own grinder ceiling. With the throttle applied, the effective crossover, the year cumulative output falls behind cumulative need, lands around 2031 globally for the combined Physical AI fleet. After that the arithmetic turns sharply against it: because every robot keeps its fourteen harmonic reducers¹ for life, the drives the fleet needs accumulate, and that growing stock outruns a pool-throttled production rate that reaches only about sixteen million units a year in 2040, against an announced curve near twenty million. By 2040 a fleet of roughly forty-seven million robots needs about 659 million harmonic drives, yet effective cumulative output has delivered only about 125 million, roughly eighteen percent of the combined requirement (humanoids need about 659 million harmonic drives, the cobot fleet about 22 million, industrial-arm wrists about 18 million and quadrupeds’ premium tier about 5 million more, roughly 704 million in all; see below). Building from a single bloc is far worse: China-only output falls behind around 2031, and US-only output, a thin aerospace-and-defence niche⁵, falls behind as early as 2025.

A harmonic, or strain-wave, drive is the precision gear that defines a humanoid’s rotary joints: a thin flexspline deforming inside a rigid circular spline gives very high single-stage reduction with near-zero backlash in a flat, light package, exactly what a robot arm-or-leg joint needs. Each humanoid uses fourteen of them¹ at its rotary joints, the shoulders, hips, knees and elbows, while the other fourteen body actuators are linear roller-screw units. Tesla’s Optimus is the canonical bill of materials for that rotary-versus-linear split¹. One honest caveat: fourteen is the premium Optimus-MAX configuration¹, budget designs substitute cheaper planetary or cycloidal reducers in some joints, so a fleet-blended harmonic count is lower than fourteen. The model carries fourteen for every robot, which therefore overstates harmonic demand to a degree we have not yet blended down.

Because a harmonic drive is built into a robot for life, the drives the fleet needs are a cumulative stock, the total number of robots ever built worldwide multiplied by fourteen¹, not a per-year flow. Integrating the consensus shipment ramp³, the global installed fleet reaches about 585,000 robots by 2030, ten million by 2035 and forty-seven million by 2040, requiring roughly 8.2 million, 141 million and 659 million harmonic drives respectively. In the chart below, the black line is that worldwide cumulative need, the grey band is cumulative global production, and the red wedge is the shortfall that opens once need outruns supply.

Production is the only thing that differs between the Global, China and US views. The worldwide humanoid fleet and the drives it needs stay identical in all three; only the supply line changes, asking whether one bloc’s factories alone could equip the whole global humanoid fleet. One disclosure up front: this curve credits the entire harmonic output to the humanoid fleet, even though industrial robots are the dominant consumer today at about eighty-four percent of harmonic volume¹⁰ and that existing demand is not subtracted, so the supply line plotted here is a generous upper bound on what humanoids could actually claim. Real output follows an S-curve rather than a smooth ramp: it sits near two million units in 2025⁴, climbs through Japanese and Chinese build-outs across the late 2020s (Harmonic Drive Systems’ Ariake plant alone runs about 220,000 units a month⁵, and Chinese makers led by Leaderdrive⁶ are localizing fast), and then plateaus against the limit of precision gear-grinding⁷. Global output runs from about two million units a year today⁴ to roughly six and a half million by 2030, fourteen million by 2035 and twenty million by 2040, but read these as announced output, the curve before the shared precision-grinding pool throttles it down (the effective curve, plotted in the gap chart, is lower; see “Why it binds”), and the post-2030 figures climbing toward twenty million are a modeled extrapolation, not a sourced forecast; China climbs from about 0.8 million to 4.5 million by 2030 and thirteen million by 2040 as domestic harmonic rises from under five percent to about thirty-five percent of Chinese-assembled-robot installs⁸; and the US holds a thin aerospace-and-defence niche, about fifty thousand units today rising toward half a million by 2040⁵, with no announced humanoid-grade lines. The market is concentrated (the top-ten makers hold about eighty-nine percent of revenue⁷), and the sector has been compounding at roughly thirteen percent a year⁹.

The effective crossover, the year cumulative production drops below cumulative need³, with the shared precision-grinding pool throttling output to its realistic slice, lands around 2031 globally for the combined Physical AI fleet (about 2033 for humanoids alone), a year or more earlier than harmonic’s standalone curve alone would suggest. Before then, the standing Japanese base plus the Chinese build-out banks enough output to cover the still-small fleet; after it, the durable need explodes, because the fleet is adding close to ten million robots a year by 2040 and each one demands fourteen harmonic drives¹, while pool-throttled grinding cannot follow, effective output reaches only about sixteen million units a year by 2040 against an announced twenty million, so cumulative supply meets just eighteen percent of the requirement. The gap then widens without bound. Building from a single bloc brings the wall forward sharply. China-only output crosses around 2031, it expands quickly but starts from a smaller base and is throttled by the same pool toward roughly thirteen million units a year of announced capacity⁸, and US-only output crosses as early as 2025, since a thin niche that reaches only about half a million units a year⁵ never keeps up with even the early fleet.

Humanoids are not the only claimant on the grinder pool. By 2040 the collaborative-arm fleet reaches about 3.7 million units, each cobot joint a strain-wave reducer, six per arm versus a humanoid’s fourteen, adding about 22 million harmonic drives; and the ~6-million-strong industrial-arm fleet carries about three harmonic drives on each wrist, another ~18 million, and quadrupeds add a small premium-tier tail, most four-legged robots are quasi-direct-drive on planetary gears, but the high-end harmonic units (ANYbotics ANYmal, the harmonic side of Boston Dynamics Spot) put roughly five million strain-wave reducers on the ~5-million quadruped fleet by 2040. Together they lift the ~659 million humanoids need to about 704 million. Cobots and industrial arms are, in fact, harmonic’s existing baseline, the traditional industrial-robot market already consumes on the order of a fifth of harmonic output today, and because cobots ship in volume now (~65,000 built in 2024 against ~18,000 humanoids) they pull the effective global crossover forward from about 2033 (humanoids alone) to about 2031, coverage about eighteen percent met. The model credits total harmonic output to the robot fleet, so adding them makes that supply line less of an upper bound, a partial reconciliation, not a full one. And drones, the largest fleet in the whole report by units, add nothing here at all: a direct-drive multirotor has no gearbox of any kind.

The plateau is not set by steel but by flexspline gear-grinding, heat-treat and skilled precision assembly. Each flexspline is a thin steel cup whose teeth must be continuous-generating-ground to micron-class tolerances on scarce Reishauer/Kapp/Klingelnberg-class machines⁷, then heat-treated for fatigue life and assembled to zero-backlash tolerances by a thin bench of skilled hands. Crucially, that same precision-grinding pool is shared with roller screws and cross-roller bearings, the three precision-reducer links compete for the very same grinders and the very same machinists⁷. Japan’s Harmonic Drive Systems built that capability over decades and still dominates the premium tier⁵; China is localizing fast⁸, but every entrant is gated by the same finite grinder base. It is that grinding throughput, not the drive itself, that caps how fast the world can build harmonic reducers, and it is why production plateaus so far below the fleet’s need. This shared-pool constraint is now operationalized in the model, not just flagged. The roughly twenty-million-units-a-year plateau is harmonic’s announced standalone curve, what it could build if the whole precision-grinding pool were available to flexsplines. But that pool is the same flexspline-, thread- and raceway-grinding labour and metrology base that roller screws and cross-roller bearings draw on too⁷, and it grows only about nine percent a year, gated by a grinder-OEM oligopoly, a thin bench of skilled operators, and scarce metrology, against roughly fourteen and a half percent a year of combined demand from the three parts. Because the three over-subscribe that single finite pool, the model throttles each to one shared factor, but that factor is not flat across time. It bites hardest through the late-2020s ramp, when the three announced lines together call on the shared grinder base by roughly one and a half times what it can supply and the model throttles each to about sixty percent of its announced curve; it then eases back toward eighty percent by 2040, because by then each part’s announced curve has already plateaued against its own standalone grinder ceiling, so the collective pool is no longer the binding constraint and the curves self-limit. Three caveats keep this honest. First, the grinder ceiling is applied only once: it already sits inside each part’s standalone plateau, and the shared pool adds the collective ramp-window coupling between the three parts, not a second ceiling stacked on top. Second, the throttle is a band, not a point estimate, it is sensitive to the modeled grinder-draw weights, so read sixty-to-eighty percent as a range. Third, it is an upper-bound approximation that bundles three different machine types, flexspline thread-grinding, gear-grinding and raceway-grinding, into one shared budget. Within that budget, harmonic’s flexspline grinding draws roughly three times the pool of a plain bearing per part, heavier than a cross-roller raceway, but far lighter than a roller screw’s multi-start thread grind at about ten times, so it is not the worst-throttled of the three, yet it is throttled all the same. With the throttle applied, harmonic’s effective crossover moves to about 2031 and only about eighteen percent of the 2040 need is met, both worse than its standalone curve implies, and the red wedge in the gap chart reflects this pool-throttled effective output, not the announced ceiling.