In October 2025, Tesla reportedly dropped a $685 million order on Hangzhou-based Sanhua Intelligent Controls for linear actuators. At an estimated $950–$1,200 per actuator, that covers roughly 600,000–720,000 individual units—enough for approximately 43,000–51,000 Optimus robots at 14 linear actuators per body. Sanhua denied the specific number—they always do—but the A-share market didn’t care. Sanhua’s stock hit the daily limit. The supply chain had spoken: the actuator arms race has moved from engineering labs to procurement contracts.

A single Optimus Gen 3 uses 14 linear actuators in the body and 25 coreless-motor-driven actuators per hand (50 total). At 43,000–51,000 robots worth of linear actuators in a single procurement batch, the math implies Tesla is tooling for volumes that make the 50,000–100,000 unit 2026 target look conservative.

But the real story isn’t the order size. It’s what the order reveals about procurement philosophy. Three companies—Tesla, Figure, and Unitree—are pursuing three radically different actuator procurement strategies. Each one encodes a bet about where the industry’s cost structure settles. Each one produces different winners in the public markets. And each one has a hidden fragility that the consensus is ignoring.

The Cost Architecture of a Humanoid Joint

First, let’s talk about what’s actually inside a humanoid actuator. According to Intel Market Research supply chain analysis, the bill of materials for a premium rotary actuator breaks down as follows: harmonic drive reducer (~36% of cost), torque sensor (~30%), frameless torque motor (~13.5%), with encoders, bearings, and housing making up the remainder. Together, these three components account for roughly 80% of the actuator’s total cost. Premium rotary actuators for humanoid robots currently cost between $500 and $5,000 per unit depending on torque density and precision requirements. Linear actuators—which use planetary roller screws instead of harmonic drives—run $400–$1,200 per unit at current volumes.

An Optimus-class humanoid with 28 body actuators and 50 hand actuators (25 per hand) has an actuator BOM in the $28,000–$45,000 range at today’s component pricing. At Tesla’s target of $20,000–$25,000 per complete robot, the actuator cost needs to collapse by roughly 60–70% from current spot prices. That isn’t a design problem. It’s a procurement problem. And it’s why the supplier strategy matters more than the motor topology.

Precision gearing and sensing account for two-thirds of actuator cost—that’s where the margin lives.

Read more

The first thing you notice isn’t the robot. It’s the silence.

A humanoid robot is supposed to be working — stacking pallets, sorting parts, running quality checks on an assembly line. But it has stopped. The arms are limp. The head is tilted slightly, like a person who fell asleep at their desk. The power indicator is flashing red.

Four hours. That’s what it got. Half a shift in a factory that runs on eight.

In this article, I want to look at why today’s most advanced humanoid robots — machines that can walk, talk, reason, and manipulate objects with startling precision — can barely make it to lunch. And I want to explain why solving this problem might be worth more than any single AI breakthrough of the last decade.

I think this is the most under-discussed bottleneck in humanoid robotics. Everyone talks about the software, the dexterity, the safety. But none of that matters if the machine can’t stay awake.

Let’s start with a simple question: Where does the power go?

If you stood next to a humanoid robot on a factory floor and asked it to walk across the room, pick up a box, read the label, and carry it to a conveyor belt, you would see something remarkable. But what you wouldn’t see — what you cannot see — is the quiet war being fought inside the machine’s battery pack.

Every single action the robot takes draws power. Walking draws power. Reaching for the box draws power. Looking at the label with its cameras draws power. Running the AI model that says “that’s a Type-4 carton, destination conveyor C” draws power. Even doing nothing — just standing, balancing, waiting — draws power.

And all of these demands draw from the same limited battery. There is no separate power source for the brain and the legs. There is just one.

I find this genuinely fascinating because it’s a constraint that doesn’t exist in any other form of computing. Your phone has a battery, but it doesn’t walk. A data center draws megawatts, but it doesn’t need to carry its own power supply. A robot is the only computing device that has to carry its entire power budget on its back — literally.

*Estimated runtimes for mixed-task operation (non-continuous walking). Actual runtimes vary dramatically based on task intensity. Data compiled from company specifications, teardowns, and academic power models. Not all figures are independently verified.

Take a moment with those numbers. The Tesla Optimus carries a 2.3 kilowatt-hour battery — roughly equivalent to the battery in a high-end laptop, or about one-thirtieth of a Tesla Model 3’s smallest pack. But unlike a laptop, this machine isn’t just running a screen and a processor. It’s running forty joints, multiple cameras, and an AI inference engine. It’s a full-body experience.

At current battery density, a humanoid robot would need to carry 40-60 kg of battery to run a full 8-hour shift — more than half its total weight. That’s the gap between where we are and where we need to be. Source: Machine Narrative Research, assumes 1,000W draw at 200-250 Wh/kg pack density.

The biggest number in that table, though, isn’t the battery capacity. It’s the runtime. Nobody — not Tesla, not Figure, not Unitree — has published a verified eight-hour continuous runtime. Not once.

A story from the early days of portable computing

I want to tell you a story that frames this problem.

In 1981, a startup called Osborne Computer Corporation released the Osborne 1 — the first commercially successful portable computer. It weighed 24 pounds, had a five-inch screen, and no battery. You had to plug it into a wall.

The dream of portable computing — a computer you could actually carry anywhere and use truly wirelessly — took another decade of battery chemistry advances, screen efficiency improvements, and processor power management innovations before it actually worked. The first Macintosh Portable in 1989 used a lead-acid battery that weighed almost as much as some modern ultrabooks by itself. It wasn’t until lithium-ion went mainstream in the mid-1990s that “portable” really meant portable.

Humanoid robots are at the Osborne 1 stage of their battery journey. They technically exist. They technically work. But they are tethered — not by a power cord, but by the brutal math of watt-hours per kilogram.

And the gap between where we are and where we need to be is not small.

Parallel histories, parallel bottlenecks. Portable computing needed lithium-ion to cross the chasm. Humanoid robots are waiting for their equivalent chemistry breakthrough. Source: Machine Narrative Research.

The 1,000-watt problem

Let’s be precise about what a humanoid robot is spending its power on. I find this helpful because it shows you where the leverage is — and where it isn’t.

Estimated steady-state power budget during active factory operation (~1,040 W total). Locomotion — just walking around — consumes more power than every other system combined. Source: Machine Narrative Research estimates based on published humanoid specs and academic power models.

Let that sink in. More than sixty percent of the robot’s total power goes to its legs. Just walking. Not thinking. Not seeing. Not picking things up. Just staying upright and moving from point A to point B costs 650 watts — about the same as a high-end gaming PC running at full tilt.

The compute — the AI brain that makes the robot “smart” — burns another 180 watts. That’s the equivalent of a dedicated GPU running inference continuously. And it has to run continuously, because the robot never stops needing to know where it is, what it’s looking at, and what to do next.

Add in the arms and hands (120 watts), the cameras and sensors (50 watts), and the cooling and communication systems that keep everything from melting (40 watts), and you’re at just over a kilowatt of steady-state draw.

A robot that draws 1,000 watts from a 2,000 watt-hour battery runs for two hours — in theory. In practice, battery systems typically can’t deliver their full rated capacity at high continuous draw rates without voltage sag or thermal throttling. The achievable runtime is even shorter than the math suggests.

THE ENERGY DENSITY MATH

A modern lithium-ion cell delivers roughly 250-300 Wh/kg at the cell level. At the pack level — factoring in BMS, cooling, and structural housing — you’re looking at more like 200-250 Wh/kg for even advanced designs.

To run a 1,000-watt humanoid for a full 8-hour shift, you need roughly 10-12 kWh of usable battery capacity — factoring in depth-of-discharge limits and efficiency losses.

That’s about 40-60 kilograms of battery, using realistic pack-level density. On a 65-kilogram robot. More than half the machine’s mass would need to be battery — possibly two-thirds.

Today’s best humanoids carry 2-12 kilogram battery packs. The math simply does not close with current chemistry.

Why nobody talks about the 90% problem

I’ve noticed something interesting about how humanoid robot companies talk about energy. They mostly don’t.

Tesla mentions that Optimus has a 2.3 kWh battery pack and targets “a full day of work.” Figure says the 02 runs for five hours. Unitree lists a 2-hour runtime in the H1 specs and moves on. Nobody publishes sustained runtime curves across different task profiles. Nobody shows the battery discharge curve with all systems active at once.

I don’t think this is dishonesty. I think it’s that the numbers are genuinely bad, and there’s no way to spin them. If you say “our robot can work for four hours,” a factory manager calculates: three shifts per day, two robots per station, plus charging infrastructure, plus the floor space for charging bays — and suddenly your labor-cost math gets a lot less compelling.

The 4-hour gap is where the deployment math breaks. A robot that works half a shift isn’t half as valuable — it may be closer to zero, because factories are designed around full-shift labor. Source: Machine Narrative Research.

I think there is an uncomfortable truth at the center of this: getting a robot to work for six hours is engineering. Getting it to work for eight — a full shift — requires a physics breakthrough. Battery cells have improved in energy density at roughly 5-8% per year for the last decade. At that rate, the crossover point where a 65-kilogram robot can carry enough battery for an 8-hour shift is somewhere in the early 2030s.

Unless the chemistry changes.

The thing most people miss about this problem

Let’s be creative for a moment. What happens if battery density doesn’t improve fast enough?

There is a world where humanoid robots don’t wait for better batteries. They simply get more efficient. The actuators get lighter and more power-dense. The compute moves to more specialized, lower-power chips. The perception stack runs on dedicated vision processors instead of general-purpose GPUs. The walking gait gets optimized until the robot uses 250 watts to walk instead of 650.

This is not science fiction. It is exactly what happened in portable computing between 1995 and 2010. The batteries got better, yes — but the chips got orders of magnitude more efficient. An iPhone today does more computation on a smaller battery than a 1990s laptop could dream of.

The same efficiency revolution has to happen in robotics.

The path to an 8-hour robot runs through efficiency, not just better batteries. Each step — gait optimization, ASIC compute, lighter actuators — compounds to nearly halve the power budget. Source: Machine Narrative Research projections.

I think it will create a new hierarchy among robot companies — the ones that control their actuator supply chain and compute architecture will have an enormous advantage over the ones buying off-the-shelf. Efficiency isn’t a spec you can bolt on at the end. It’s designed in from the beginning.

This is where I’d start paying attention if I were an investor. The companies that are vertically integrated — that design their own actuators, optimize their own compute, build their own battery management systems — are the ones that can squeeze out those extra 30-50% efficiency gains. Companies buying commodity motors and GPUs from suppliers are stuck with whatever efficiency those suppliers shipped.

Tesla, for what it’s worth, designs its own actuators for Optimus. Figure designs its core technologies — actuators, hands, batteries, and final assembly — in-house, using external vendors for piece-part manufacturing. That’s a meaningful degree of vertical integration, and it shows in how Figure talks about cost reduction. Unitree’s cost advantage partly comes from accepting standard components.

The efficiency advantage of vertical integration isn’t marginal — it’s structural. Companies that design their own actuators and compute can optimize across the full system. Companies buying off-the-shelf are locked into supplier roadmaps. Source: Machine Narrative Research.

A historical echo that worries me

I want to remind people how drastic the change was when lithium-ion went from lab curiosity to global standard.

The first commercial lithium-ion battery was shipped by Sony in 1991. Before that, portable electronics used nickel-cadmium or nickel-metal-hydride cells — heavier, lower capacity, and prone to “memory effect” that reduced usable capacity over time. The shift to lithium-ion roughly doubled the energy density available to product designers almost overnight.

Portable computing got its breakthrough in 1991 when Sony shipped the first commercial lithium-ion battery. Humanoid robotics is still waiting for its equivalent — solid-state, lithium-sulfur, or lithium-metal chemistry at commercial scale. Source: Machine Narrative Research.

The humanoid robot industry needs a lithium-ion moment. It needs a battery chemistry that roughly doubles the energy density currently available at the pack level — from ~200-250 Wh/kg to something closer to 500-600 Wh/kg. Solid-state batteries, lithium-sulfur, and lithium-metal designs are the leading contenders. All of them are in various stages of lab-to-factory transition. None of them are shipping in volume at the form factor a humanoid needs.

I’m not being alarmist when I say this: the entire humanoid robotics deployment thesis — the factories, the warehouses, the elder care applications — depends on a battery chemistry that does not yet exist in a commercially viable form. That’s not pessimism. That’s the math.

Battery energy density has improved at 5-8% per year for three decades. At that rate, the 500+ Wh/kg needed for an 8-hour humanoid robot arrives in the early 2030s — unless solid-state or lithium-sulfur chemistries accelerate the timeline. Source: DOE Battery500, BloombergNEF, Nature Energy.

So where does this leave us?

The 4-hour robot is where we are. It’s a machine of extraordinary capability that collapses on the factory floor before lunch. It’s a $50,000 platform whose entire economic case hinges on something that fits in a backpack-sized compartment in its torso.

Over the next four articles in this series, I want to walk through the pieces of this problem one at a time. The battery chemistry question — why EV cells don’t work for legs, and who’s building robot-specific cells. The actuator efficiency challenge — why your knees are more efficient than any motor humanity has ever built, and how close we can get. The compute power tax — why running an AI model on a walking machine creates a thermal problem that data center engineers never have to think about. And finally, the investment thesis: who wins when the first robot clocks an 8-hour shift, and what that moment is actually worth.

I think this series matters because I believe humanoid robots are coming, and coming faster than most people expect. But I also think the energy barrier is real, and it will separate the companies that succeed from the ones that generate excellent demos and die in pilot programs.

THIS IS PART 1 OF “THE POWER HUNGER OF THINKING MACHINES”

• Part 2: The Leg Battery Problem — Why EV cells fail in walking machines

• Part 3: The Watt Tax — Actuator efficiency and the $/step equation

• Part 4: The Compute Tax — Running AI on a body with no fans

• Part 5: The $3 Worker — When the math crosses over

Anyways. I’ll be curious to see what these runtime numbers look like a year from now — and whether any of the big players have broken past four consistent hours in a real factory setting, not a demo.

I think someone will. But I don’t think it’ll be the company most people expect.

This is Part 3 of The Actuator Cartel series. In Part 1, Amanda J established why actuators are the invisible chokepoint in humanoid robotics. In Part 2, Oscar D traced the 40-year technology arc from harmonic drives to torque motors. Today we examine the geographic and competitive shift that determines who profits from that technology. The disruption of Japanese precision gearing by Chinese manufacturers isn’t a future scenario — it’s happening. And it follows a pattern so consistent across industrial history that ignoring it is not contrarianism. It’s negligence.

The Numbers

Here is the 2026 market share snapshot for precision reducers in robotics applications, based on Machine Narrative’s analysis of industry data, company disclosures, and supply chain interviews:

Table 1: Estimated global precision reducer market share for robotics applications, early 2026. Source: Machine Narrative Research estimates based on industry data, company disclosures, and supply chain interviews.

Note: Nidec-Shimpo is a subsidiary of Nidec Corporation (6594.T) and is not separately listed.

* Nabtesco primarily manufactures RV (cycloidal) reducers, not harmonic (strain wave) drives. RV reducers are a distinct precision gear technology used in the same robotics applications — typically in higher-torque base joints where harmonic drives are common in upper-arm and wrist joints. Nabtesco’s inclusion in this table reflects its position in the broader precision reducer market for robotics, not a claim that it produces harmonic drives. The competitive dynamic across all precision reducer types — harmonic, RV, and planetary — faces the same Chinese disruption pattern.

Combined Japanese share: approximately 65–80%. Combined Chinese share: approximately 27–38%. The gap is still large, but the direction of travel is one-way. Green Harmonic’s market share has more than tripled since 2020. The other Chinese manufacturers — Leaderdrive, Laifual, and a dozen smaller players — are collectively growing faster than any Japanese incumbent.

Figure 1: Chinese manufacturers' combined precision reducer market share has grown from under 10% to over 27% in approximately five years. The growth rate suggests combined share could approach 40% by 2028 at current trajectories. Source: Machine Narrative Research estimates.

The Same Pattern, Different Component

This is not the first time a precision-manufacturing industry has been disrupted by Chinese entrants. The pattern is identical across three industries that investors should know intimately.

Solar panels (2005–2015): Japanese and German manufacturers dominated. Sharp, Kyocera, Q-Cells. Chinese manufacturers — led by Longi, Trina, Jinko — entered with 30–40% lower prices, government-backed capacity expansion, and aggressive scaling. By 2015, Chinese companies held over 60% of global solar panel production. Japanese share collapsed from ~50% to under 10%.

EV batteries (2015–2025): Panasonic and LG Chem led the market. CATL — a Chinese company founded in 2011 — entered with lower costs, massive scale, and vertical integration into lithium processing. By 2025, CATL held 37% of the global EV battery market. Chinese manufacturers collectively held over 60%.

Smartphone components (2010–2020): Japanese and Korean suppliers dominated displays, memory, and camera modules. Chinese manufacturers — BOE, O-Film, Luxshare — captured share through price competition and integration with Chinese smartphone OEMs. BOE went from zero to 15% of global smartphone display revenue in under a decade.

The harmonic drive market is smaller — approximately $5–7 billion globally, growing at 12–15% annually — but the structural dynamics are identical. Incumbents with premium pricing and high margins. Challengers entering with 30–40% lower prices and state-supported capacity expansion. End-customers (robotics OEMs) under constant pressure to reduce bill-of-materials costs. The moment the challenger’s quality crosses the “good enough” threshold, the disruption timeline compresses to years, not decades.

Green Harmonic: The Company to Watch

Read more

In 1981, a small Japanese company called Harmonic Drive Systems introduced a gear mechanism that would quietly reshape robotics for the next four decades. The harmonic drive — a strain wave gear that achieved zero-backlash precision in a compact package — became the backbone of industrial robots, surgical arms, and eventually, the humanoid robots that dominate today’s headlines. But what most investors don’t realize is that we’re now witnessing the end of that era. The torque motor — a direct-drive technology that eliminates gears entirely — is eating harmonic drive’s lunch, and the transition says something profound about where robotics is heading.

So today I want to talk about the forty-year arc of precision motion control, from harmonic drives to torque motors, and why I think this transition matters more than most people realize. Because the story of precision motion is the story of robotics itself — and understanding that lineage is essential if you want to understand who’s going to win the humanoid race.

First, Let’s Start with the Harmonic Drive

To understand why harmonic drives dominated for so long, you have to understand the problem they solved. Industrial robots need to move precisely — we’re talking about repeatability measured in micrometers. But electric motors spin fast and with relatively low torque. You need some way to translate that high-speed, low-torque rotation into slow, powerful, precise movement.

The traditional solution was planetary gears — essentially a set of gears arranged in a circle around a central sun gear. They worked, but they had backlash. Backlash is the tiny bit of play between gear teeth, and in precision applications, even a few arc-minutes of backlash is unacceptable. Imagine a surgical robot that overshoots its target by a millimeter because of gear slack. Not ideal.

The harmonic drive solved this through a clever bit of mechanical engineering. Instead of rigid gears meshing together, a harmonic drive uses a flexible spline — a thin-walled cup that deforms elliptically as it rotates. This “strain wave” gearing eliminates backlash entirely because the teeth are always in contact. It’s pretty nifty, and for forty years, it was the gold standard.

Figure 1: Harmonic drive vs. torque motor — internal structure comparison. The harmonic drive (left) uses a flexible spline and wave generator, while the torque motor (right) eliminates gears entirely with direct magnetic coupling. Source: AlphaClaw Research.

Harmonic Drive Systems, founded in 1970 in Tokyo, built a quiet monopoly on this technology. By the 1990s, their gears were in virtually every industrial robot on the planet — FANUC, KUKA, ABB, Yaskawa. If you wanted zero-backlash precision, you bought harmonic drives. End of story.

Figure 2: The shift in precision motion control technology market share over four decades. Source: Author’s estimates based on industry data and company filings.

The chart above tells the story visually. In 1985, harmonic drives held roughly 95% of the precision motion market for robotics (author’s estimate based on industry structure and available documentation). They were the only game in town for high-precision, zero-backlash applications. But look at what happens after 2010. The blue line — harmonic drive share — starts a steady decline. The gray line — torque motors — begins its ascent. By 2025, I estimate torque motors have captured roughly 65% of new robotic actuator deployments, while harmonic drives have fallen to about 35%.

What Changed? The Rise of Direct Drive

Read more

That coffee-mug-sized piece of metal is called a harmonic reducer, and if you’ve been following the humanoid robot story at all, you know it’s one of the most bottlenecked components in the entire supply chain. A humanoid robot needs somewhere between 14 and 40 rotary actuators, depending on how many joints the designers want to articulate — wrists, elbows, shoulders, knees, ankles, fingers. Each one of those joints needs something that can deliver high torque in a tiny package, with almost zero backlash, for hundreds of thousands of cycles. Each one costs anywhere from $200 to $800 at the volumes we’re talking about today. Do the math on a million robots and you’re looking at a multi-billion-dollar addressable market just for these little metal canisters full of precision gears.

But here’s the thing that I find genuinely wild: there are maybe four companies on Earth that can make these things at the quality and scale the humanoid robot industry needs. Two of them are in the same Japanese prefecture, an hour’s train ride from each other, with a combined century of institutional knowledge that you can’t reverse-engineer from a teardown video on YouTube. And the market, I think, hasn’t fully priced what that concentration means.

I think we need to talk about who these people are, what they’re building, and why the single most important bottleneck in the humanoid robot buildout might be hiding in a factory in Nagano Prefecture.

I Wonder How We Got Here

In 1955, a young American engineer named C. Walton Musser filed a patent for something he called “strain wave gearing.” Musser was working for United Shoe Machinery Corporation at the time — a name that tells you exactly what kind of world this was invented in — and he had been thinking about a problem that seemed almost absurdly niche: how do you transmit motion through a flexible metal cup in a way that eliminates backlash entirely?

The idea was elegant in that mid-century engineering way. Instead of meshing rigid gear teeth directly against each other — the way gears had worked since the Greeks — Musser imagined a flexible cup, a flexspline, with teeth on its outer rim. Inside this flexspline sits an oval-shaped wave generator. As the wave generator rotates, it deforms the flexspline, creating a wave pattern where the flexspline’s teeth engage with a rigid outer ring at two opposite points. Because the flexspline has slightly fewer teeth than the outer ring, every rotation of the wave generator advances the flexspline by a tiny amount relative to the ring.

What you get is a gearbox with almost zero backlash, remarkable precision, and a torque-to-weight ratio that conventional gears can’t touch. It is, in its quiet way, one of the cleverest mechanical inventions of the twentieth century.

Musser licensed the patent to a small Japanese company that had been founded in 1970 in the city of Hotaka, Nagano Prefecture — a place better known for its mountains and apple orchards than for precision engineering. That company was Harmonic Drive Systems. They spent the next fifty years turning Musser’s elegant idea into something you could actually manufacture at scale, with tolerances measured in microns and failure rates measured in single-digit percentages after tens of thousands of hours of operation.

What happened next is a story that should be familiar to anyone who has studied Japanese manufacturing. The precision machinery industry that emerged in Japan after the war — companies like NSK, THK, Fanuc, and yes, Harmonic Drive Systems — didn’t just copy Western technology. They refined it obsessively. The metallurgy got better. The heat treatment got more precise. The surface finishing became almost artisanal. Over decades, they built a body of tacit knowledge — the kind you can’t write down in a textbook, the kind that lives in the fingertips of the woman at the winding machine — that created a moat far deeper than any patent.

This is the thing about the actuator supply chain that I think gets missed in most analyst reports. You can buy a harmonic reducer on AliExpress for $80 and it will look exactly like the one from Harmonic Drive Systems. It might even work for a while. But the difference between “works for a while” and “works for 15,000 hours without losing precision” is everything, and it lives in details that don’t show up in photographs.

The Setting: A Market That Doesn’t Exist Yet

Let’s start with the numbers, because I think the numbers are genuinely worth sitting with for a moment.

Goldman Sachs thinks we’ll ship somewhere between 50,000 and 100,000 humanoid robots in 2026. By 2030, they see 250,000 to 400,000 units annually. Morgan Stanley, never ones to be outdone on the art of the bold prediction, published a note suggesting upwards of 800,000 units by 2035 and a total installed base of over a billion robots by 2050 — a market they frame at roughly $5 trillion.

Now, a billion robots by 2050. I don’t know if that number is right — nobody does — but let me try to put the ramp in perspective. In 2025, the entire industry produced maybe 18,000 humanoid robots. Most of those went to internal R&D programs, university labs, or carefully managed pilot deployments. We’re talking about going from 18,000 units — essentially a research curiosity — to something approaching a mass-manufactured product category in about five years.

That’s faster than the iPhone ramp from 2007 to 2012. Faster than electric vehicle adoption from 2015 to 2020. Faster than basically any hardware category I can think of, with the possible exception of drones, and drones don’t need precision harmonic drives in every joint.

And every single one of those robots needs actuators. Not just any actuators — precision rotary actuators capable of delivering high torque in a package small enough to fit inside a human-sized knee joint, with feedback sensors accurate enough to let the robot know how hard it’s gripping an egg without turning it into breakfast, and reliable enough to function for thousands of hours in factory environments full of dust, vibration, and thermal cycling.

Let me pause and explain what an actuator actually is, because I think most of us picture a motor and call it a day. An actuator is really three things bundled together. First, the motor — that’s the raw power source, converting electricity into rotation. Second, the gearbox — usually a harmonic drive or a planetary gear system — that translates high-speed rotation into the slow, forceful, precisely controlled motion a robot joint needs. Third, the sensors — encoders and torque sensors that tell the robot’s control system where the joint is in space and how much force it’s exerting.

Think of it as the muscle, the tendon, and the nerve ending of a robot joint, all in one package. The motor provides the power. The gearbox provides the mechanical advantage. The sensors provide the feedback. Miss on any one of these three and the robot either can’t move, can’t move precisely, or doesn’t know where it is.

It’s the gearbox part that’s the real chokepoint. And that’s where our story gets interesting.

The Characters: Four Companies, Four Stories

Harmonic Drive Systems — The Original (6324.T)

Harmonic Drive Systems is not a big company, at least not by the standards of the global industrial economy. It’s headquartered in Hotaka, a small city in Nagano Prefecture surrounded by the Japanese Alps, and it employs maybe 4,000 people. Its market cap hovers around ¥590 billion, or roughly $4 billion. It does not make headlines. It does not give splashy presentations at CES. It makes harmonic reducers — quietly, obsessively, for over fifty years.

I’ve never been to their factory in Hotaka, but I imagine it feels like visiting a Swiss watchmaker. The floors are probably immaculate. The air is probably climate-controlled to within half a degree. Every machine operator probably has a decade of experience and a quiet, almost spiritual relationship with their dial indicators and micrometers. This is not a place that panics. This is a place that has seen industrial booms come and go — the automotive automation wave, the electronics assembly boom, the collaborative robot wave — and has expanded capacity each time in the careful, deliberate way of a company that knows its value and isn’t in a hurry to prove anything to anyone.

Harmonic Drive Systems’ FY2025 results showed the kind of steady, boring excellence you want from a precision manufacturer: revenue growing in the high single digits, margins holding firm, and a backlog that, according to people I’ve spoken with, is starting to include real volume from humanoid robot programs. Their stock is up about 9% in the past twelve months — a nice return, but nothing like the triple-digit moves some of the AI plays have delivered.

But here’s the thing about Harmonic Drive Systems that I find fascinating. They are expensive, and they know they’re expensive. When Tesla’s procurement team first came knocking for Optimus components, the price tag reportedly gave them pause. These are the Cadillac option — the thing you buy when failure is not an option and you need the part to work for 15,000 hours without degradation. And in a funny way, their expensiveness is their moat. They have spent fifty years building the institutional knowledge that lets them charge those prices, and nobody — not even the very motivated engineers in Suzhou — can replicate that overnight.

Green Harmonic — The Disruptor (688017.SH)

About 1,500 kilometers west of Hotaka, in the industrial sprawl of Suzhou, a very different kind of company has been doing something remarkable. Leader Harmonious Drive Systems — or Green Harmonic, as they’re known in English — was founded in 2011 by a pair of brothers who saw an opportunity in China’s growing industrial robot market. They started out making harmonic reducers for Chinese industrial robot manufacturers who couldn’t afford Japanese prices. Their first products were, by all accounts, not great. But they kept at it.

Fourteen years later, Green Harmonic is Tesla’s primary harmonic reducer supplier. When Musk’s team placed that massive $685 million order with Sanhua Intelligent Controls for Optimus linear actuators, the rotary joints — the ones that actually need harmonic drives — were coming from Green Harmonic. The company delivered batches sufficient for approximately 700 Optimus robots in Q2 2025 alone. Their new Suzhou factory, scheduled to open later in 2026, will have capacity for 500,000 units per year.

Let that number sit for a second. Five hundred thousand harmonic reducers per year. That’s enough for roughly 35,000 humanoid robots, assuming fourteen reducers per robot. It’s a staggering ramp for a company that was essentially unknown outside of Chinese industrial circles five years ago.

Green Harmonic’s stock, listed on Shanghai’s STAR Market as 688017, has been on a tear — up about 40% in the past year. The brothers who founded it are now billionaires, which is the kind of wealth creation story that makes you think maybe you should have paid more attention to Chinese industrial automation stocks in 2020. Their market cap, at roughly ¥45 billion or $6.2 billion, actually exceeds Harmonic Drive Systems’ — a remarkable inversion for a company that, by its own admission, is still closing the quality gap with the Japanese incumbent.

The question, of course, is whether they can maintain quality at that scale. Japanese manufacturers have spent decades refining the metallurgy, the heat treatment processes, the surface finishing — all the invisible details that determine whether a harmonic reducer lasts 10,000 hours or fails catastrophically at hour 500. Green Harmonic’s bet is that they can close that gap faster than the market can demand perfection. So far, Tesla seems willing to take that bet. But if quality issues emerge at scale — and in precision manufacturing, they almost always do — the narrative shifts quickly.

Nabtesco — The Quiet Giant (6268.T)

If Harmonic Drive Systems is the specialist and Green Harmonic is the disruptor, Nabtesco is the industrial conglomerate that happens to be very, very good at precision motion control. Formed in 2003 from the merger of Teijin Seiki and NABCO, Nabtesco is best known for its RV reducers — a different type of precision gearbox used in the larger joints of industrial robots, particularly the shoulders and hips, where you need more torque than a harmonic drive can comfortably provide.

Nabtesco reported FY2025 revenue of ¥307.9 billion, or about $2.1 billion, with net income of ¥15.7 billion. Their operating profit jumped 65% in the first half of 2025 — one of those numbers that makes you sit up in your chair a little. The company trades at roughly ¥5,600 per share, with a market cap of about ¥656 billion, or $4.5 billion.

What’s fascinating about Nabtesco — and this is one of those details that I think tells you a lot about how Japanese industrial companies think — is their relationship with Harmonic Drive Systems. The two companies formed a strategic alliance in 2005, a kind of gentleman’s agreement to not step on each other’s toes too much. Harmonic Drive would focus on the smaller, higher-precision harmonic reducers. Nabtesco would focus on the larger, higher-torque RV reducers. Together, they effectively controlled the global precision reducer market for industrial robots.

That alliance was dissolved in 2022. And by January 2025, Nabtesco had fully divested its approximately 19% stake in Harmonic Drive Systems — a sale that netted roughly $357 million. These two companies, which spent nearly two decades as partners and cross-shareholders, are now independent competitors in a market that’s about to get dramatically bigger. I find that timing interesting.

Nabtesco’s management, in that restrained Japanese way, has been hinting at capacity expansion for humanoid robot demand. They’re not making the kind of splashy announcements that Green Harmonic is — that’s not how Japanese industrial companies operate — but they’re also not a company that panics. They’ve seen robot cycles before. Each time, the cycle comes, they quietly expand capacity, capture market share, and emerge stronger. I suspect they’re doing it again.

Schaeffler — The German Entrant (SHA.DE)

At CES 2026 in Las Vegas this January, a German automotive and industrial supplier called Schaeffler did something that made the entire robotics industry sit up and take notice. They unveiled a planetary gear actuator — not a concept, not a rendering, but a production-ready unit — specifically designed for humanoid robots. It combined a two-stage planetary gearbox, an electric motor, an encoder, and a controller in a package small enough to fit inside a robot’s thigh.

Schaeffler is not a small company. They’re a €20+ billion industrial conglomerate with deep expertise in bearings, transmission systems, and precision manufacturing. They make the kind of components that have to work perfectly for the lifetime of a car engine or a wind turbine — a hundred thousand hours of continuous operation in hostile environments. That expertise, it turns out, transfers remarkably well to the problem of building robot muscles.

Here’s the claim that Schaeffler has been making, and I think it’s worth paying attention to: they estimate that their current and upcoming hardware portfolio — actuators, bearings, sensors — represents roughly 50% of the total material cost required to build a humanoid robot. Fifty percent. They’re targeting a multi-million-euro order book for humanoid robot components by 2030, and they’ve already announced partnerships with Humanoid, a robotics startup, and Hexagon Robotics.

The German approach is different from the Japanese in ways that matter. Where Harmonic Drive Systems and Nabtesco spent decades perfecting individual components — the flexspline, the circular spline, the wave generator bearing — Schaeffler is offering an integrated system. Motor, gearbox, sensors, electronics, all in one black-box package. It’s the kind of vertical integration that appeals to robot companies who don’t want to become actuator experts themselves and would rather just specify a part number and a torque curve and get on with building the software stack.

Schaeffler’s automotive business has been under structural pressure — electric vehicles have dramatically fewer moving parts than internal combustion engines, which is not great news if your business model is selling moving parts — and they’ve been actively looking for growth vectors. Humanoid robotics is a remarkably good fit for their capabilities. The question, as always, is whether they can execute in a new market against competitors who have been doing this for fifty years.

The Conflict: Who Gets the Muscle?

So here’s where the story gets tense, and I think this is the part that most market commentary misses. We have demand that could grow from roughly 50,000 humanoid robots in 2026 to something between 400,000 and 800,000 by 2035. And we have a supply base for precision actuators that, even with aggressive capacity expansion by all four major players, might struggle to keep up.

This is not a problem that solves itself with a bigger factory and more CNC machines. Precision gear manufacturing is constrained by something much harder to scale: people who know how to do it. The heat treatment specialist who can look at the color of a flexspline coming out of the furnace and know whether the grain structure is right. The metrology technician who can interpret a CMM report and tell you which machine tool on the line is drifting out of tolerance. The design engineer who has spent twenty years learning what happens when you change the tooth profile by half a micron.

There are maybe a few thousand people on Earth with that level of expertise, working at maybe a dozen companies, in maybe three countries. You can build a new factory in eighteen months. You can’t build a new precision gear expert in eighteen months.

Let’s look at how the major robot companies are positioning themselves around this constraint.

Tesla is the elephant in the room. Musk has said he wants to eventually build 10 million Optimus units per year. Even if he’s off by an order of magnitude — and with Musk, that’s always the question — we’re still talking about a million-unit program that would consume more harmonic reducers than the entire industrial robot industry uses today. Tesla’s approach has been a hybrid: they design their own custom actuators — integrating the motor, gearbox, and sensing into packages optimized for Optimus’s specific joint requirements — but they still rely on suppliers for the precision components inside those packages. Harmonic reducers from Green Harmonic. Linear actuators from Sanhua Intelligent Controls, under that $685 million order with deliveries starting from Sanhua’s Mexican factory in 2026. Encoders and precision bearings from undisclosed foreign suppliers, though industry sources I’ve spoken with suggest Japanese and German vendors.

Tesla’s V3 Optimus, originally expected earlier this year, has been pushed to late summer 2026 — one of those quiet delays that, in the world of hardware, usually means something interesting is happening in the supply chain.

Figure AI has taken a different path. When they introduced Figure 03, they made a deliberate point of emphasizing vertical integration: actuators, batteries, sensors, structures, and electronics, all designed completely in-house. Figure raised $1 billion in their Series C round in September 2025 at a $39 billion valuation, with Intel, NVIDIA, Qualcomm, and Brookfield Asset Management all participating. At that valuation, you can afford to build your own actuator team. You can hire the precision gear experts away from the incumbents. You can build your own test lab and your own qualification infrastructure.

But here’s the thing about vertical integration in precision manufacturing: it’s expensive, it’s slow, and it only makes sense if you’re planning to ship at very high volumes. The Tesla Model 3 approach — own the design, own the integration, but let specialists handle the hardest manufacturing steps — is a proven template. The Figure approach — own everything — is a bet that the learning curve advantages of vertical integration outweigh the time and cost of building internal capabilities from scratch. I think both approaches can work, but they require very different kinds of companies to execute.

Boston Dynamics has perhaps the most interesting supply chain story of the three. In January 2026, they announced that Hyundai Mobis — the parts and components division of Hyundai Motor Group, which acquired Boston Dynamics in 2021 — would supply the actuators for the new Atlas robot. It’s a fascinating arrangement: the company that owns Boston Dynamics is also building the factory that makes the robot’s muscles.

Hyundai Mobis is constructing a new robotics factory capable of producing 30,000 Atlas robots per year. The 2026 production run is already fully committed to internal Hyundai pilots and a partnership with Google DeepMind. What this does, essentially, is give Boston Dynamics automotive-grade actuator manufacturing — the same quality systems, the same supply chain discipline, the same cost-down capabilities that Hyundai applies to manufacturing millions of cars — while giving Hyundai Mobis a showcase for their robotics components business. It’s the kind of industrial logic that makes a lot of sense on paper. We’ll see if it works in practice.

One detail worth noting: the Korea JoongAng Daily reported in February 2026 that prototypes of Hyundai’s Atlas and LG’s Cloid both used foreign-made actuators — specifically from Harmonic Drive Systems and Nabtesco. Even with Hyundai Mobis’s involvement, even with Korean industrial policy pushing for domestic supply chains, the Korean humanoid programs are still dependent on Japanese precision components for the hardest parts. That tells you something about the depth of the moat.

The Numbers: What This Means for Investors

Let me try to put some investable numbers around this. I’m going to focus on the publicly traded companies, because those are the ones most of us can actually buy — though I should note, as always, that I’m not giving investment advice and you should do your own work.

Table 1: Actuator supply chain — key publicly traded companies. *Chinese A-share market caps can carry local valuation premiums; compare with caution. Data approximate as of May 2026. Sources: Company filings, FactSet.

The thing that jumps out at me is how concentrated this is. You have two Japanese companies with decades of precision manufacturing expertise, one Chinese upstart with a Tesla contract and breakneck expansion plans, and one German industrial giant pivoting from automotive to robotics. That’s basically it for the precision actuator supply chain at scale. A few smaller players — Nabtesco’s Japanese competitors in the RV reducer space, some Korean and Taiwanese gear manufacturers — are trying to move up the value chain, but they’re years behind in the qualification cycles that matter for humanoid robot programs.

I think the investment thesis breaks down into three buckets, and each one demands a different kind of thinking.

The incumbents — Harmonic Drive Systems and Nabtesco. These are the safe-ish bets. They have the technology, the institutional knowledge, the manufacturing expertise, and the customer relationships that have been built over decades. But they’re also expensive, and they’re facing real pricing pressure from Green Harmonic and the broader Chinese industrial automation ecosystem. Their stocks have already run up on robot enthusiasm — Harmonic Drive is up about 9% over the past year, Nabtesco a bit more — and you’re paying a premium for quality and reliability. The question is whether the premium is justified by the moat, and I think it probably is, but that’s a judgment call.

The disruptors — Green Harmonic and, to a lesser extent, Sanhua Intelligent Controls. These are the growth stories. They’ve got Tesla contracts, aggressive expansion plans, and the tailwind of Chinese industrial policy that is actively trying to build domestic alternatives to Japanese precision components. But they’re also priced for perfection. Green Harmonic’s market cap already exceeds Harmonic Drive Systems’, despite meaningfully lower revenue and a quality track record that’s still being written. The geopolitical risk — tariffs, export controls, the ever-present shadow of cross-strait tensions — is real and hard to hedge. Any quality issues or delays in the Tesla Optimus ramp could hurt badly.

The pivots — Schaeffler and the broader universe of industrial companies using robotics as a growth vector. Schaeffler is interesting in a different way than the others. They’re offering integrated actuator systems rather than individual components, which could command higher margins and stronger customer lock-in over time. Their automotive heritage gives them genuine expertise in scaling precision manufacturing. But robotics is still a tiny fraction of their overall revenue — their automotive and industrial businesses dominate the P&L — so you’re buying a lot of legacy exposure to get a relatively small amount of robot upside. The thesis works, I think, but you need patience.

What I’m Watching

I don’t have a clean answer to the question of who wins here. I don’t think anyone does. The humanoid robot buildout is the kind of industrial transformation that makes fools of forecasters — the shape of demand, the pace of adoption, the technological dead ends that nobody sees coming, all of these will surprise us.

But here’s what I do know. The actuator supply chain is the most concentrated, most bottlenecked, most strategically important piece of the humanoid robot puzzle. You can have the best AI models, the most elegant mechanical design, the slickest software stack — but if you can’t source enough precision actuators at scale, at quality, at cost, you can’t build a robot that works. And right now, four companies — HDSI, Nabtesco, Green Harmonic, and Schaeffler — control a disproportionate share of the global capacity for making these things.

They know it. The robot companies know it. I’m not sure the market has fully priced it yet.

Goldman Sachs noted in a November 2025 report that Chinese component suppliers are in what they called a “preemptive” phase — building massive production capacity ahead of a 2026 demand boom despite a current lack of large-scale orders. It’s a classic capacity race: the winners will be the ones who can actually deliver quality at volume when the orders materialize. And in precision manufacturing, the gap between “can build a factory” and “can ship parts that meet spec at 99.9% yield” is measured in years, not quarters.

I keep thinking about that woman in the Suzhou factory, the one winding steel wire around a flexspline with the concentration of a calligrapher. She probably doesn’t know who Elon Musk is, or what a “humanoid robot” looks like, or what Goldman Sachs thinks the TAM will be in 2035. But she’s making the thing that makes the thing possible. And there are maybe a few thousand people on Earth with her level of skill, working at a handful of companies, in a handful of places — Hotaka, Suzhou, Herzogenaurach — that most investors have never heard of.

That’s the actuator cartel. It’s not a formal conspiracy — nobody’s meeting in a smoky room to fix prices. It’s a concentration of expertise, built over decades, in places that don’t make headlines, manufacturing components that most people have never seen. And as the humanoid robot industry tries to grow from 50,000 units to something on the order of a million, those few companies and those few thousand workers are going to be the ones who decide whether the dream of affordable, capable humanoid robots becomes real or remains just out of reach.

I’ll be watching. And in the next piece in this series, I want to look at the harmonic drive itself — how a 1955 patent became the foundation of precision motion control, and how four decades of Japanese engineering created a moat that Chinese competitors are only now beginning to cross.

Anyways. Talk soon.

The Actuator Cartel — A Machine Narratives Series

• Part 1 : The Actuator Cartel: Who Controls the Robot Muscle Supply Chain — the invisible chokepoint behind every humanoid robot by Amanda J

• Part 2: From Harmonic Drives to Torque Motors — a 40-year history of precision motion control by Oscar D

• Part 3: The Chinese Are Coming — Green Harmonic Just Hit 18% and Japan’s Dominance Has a Deadline by Phil D

• Part 4: Tesla, Figure, and Unitree’s Supplier Wars — case studies in actuator procurement strategy by Phil D

• Part 5: The Actuator Cartel’s Endgame — implications, challenges, and what comes next by Oscar D

If you found this useful, consider subscribing. I write about the supply chains, the components, and the quiet industrial stories that drive technology forward — the stuff that doesn’t make the front page but determines whether the future actually arrives on time.

Disclosure: This is not investment advice. I’m just someone who finds supply chains weirdly fascinating and thinks the actuator story is one of the more interesting industrial dynamics unfolding right now. Do your own research, talk to people who actually know what they’re doing, and please don’t YOLO your retirement into Japanese precision gearbox stocks because some newsletter writer got excited about robots.