In the early 1990s, the obvious semiconductor bets were Intel, AMD, Motorola. The chip designers everyone knew. But the real wealth was in the component layer. ASML made the lithography machines. Applied Materials made the deposition tools. Lam Research made the etchers. Boring, obscure companies that sold to chipmakers, not consumers. ASML went from $20 to $1,000. Applied Materials from $5 to $200. The chip designers had brutal drawdowns. The tool makers just kept compounding, because every chip company needed their tools.

I keep thinking about that pattern when I look at the sensor market.

The Layer Cake

The sensor supply chain looks like a layer cake. Sensor market at the top: $168 million for tactile, $2.3 billion for force/torque. The next layer is the sensor integrators: ATI, XELA, Bluedot Touch. They buy components, assemble sensors, sell to robot OEMs. Below that, the component suppliers: strain gauges, signal conditioning chips, connector assemblies. At the bottom, raw materials: metal alloys, semiconductor substrates, specialty adhesives.

Each layer is more concentrated than the one above it. Each layer is less covered by analysts. And each layer has more pricing power than the one it sells to.

The sensor integrators get all the attention. The component suppliers get almost none. That’s where the alpha is.

Inside every force/torque sensor bolted into robot wrists, ankles, and joints, there’s a sensing element. A thin foil or semiconductor pattern bonded to a metal beam. When the beam flexes, the foil deforms. Its electrical resistance changes. That change gets measured, amplified, and converted into a number telling the robot how much force it’s applying.

That sensing element is a strain gauge. The companies that make them are more concentrated, less covered, and more critical than the sensor companies themselves.

Seventy Years, No Substitute

A strain gauge is a thin metallic foil, usually constantan alloy, patterned into a grid and bonded to a structural element. When that element bends under load, the foil stretches. Its electrical resistance changes proportionally. A Wheatstone bridge circuit measures the change, typically in microvolts. Signal conditioning electronics amplify and digitize it. The result is a precise force or torque reading.

The concept dates to 1856, when Lord Kelvin discovered that the electrical resistance of wires changed under stretch. The modern bonded foil strain gauge was independently developed in the late 1930s by Edward Simmons and Arthur Ruge during aircraft structural testing. The technology was classified for military use during World War II. After the war, it entered commercial manufacturing and became the foundation of the entire load cell and force measurement industry.

That was seventy years ago. Nothing has displaced it.

Every six-axis force/torque sensor uses strain gauges. Piezoelectric sensors exist for dynamic measurements, but they can’t hold a static reading. Capacitive sensors work for some applications but lack the precision and temperature stability for robotics-grade force measurement. Strain gauges are the only technology that does what robots need, at the precision robots require, at a cost that scales.

They’re inside every load cell, every force sensor, every structural test rig in aerospace, automotive, and manufacturing. The Boeing 787’s wing flex testing used strain gauges. Every crash test dummy has them. And they’re about to be inside every humanoid robot that ships.

$230 million. The entire component layer inside a multi-billion-dollar measurement industry. Growing at 3-4% annually. Boring, steady, invisible. Until humanoid robots start shipping at volume.

The global strain gauge market is roughly $230 million in 2026. The entire component layer inside a multi-billion-dollar measurement industry is smaller than a single mid-cap tech company’s quarterly revenue. Growing at 3-4% annually. Boring, steady, invisible. Until humanoid robots start shipping at volume.

If this analysis is useful, subscribe to see the four companies who make strain gauges and where the Alpha lives in sensors.

Subscribe now

Read more

Ouster’s software-attached bookings doubled in 2025. Software now ships with more than 15% of their sensors. When a customer buys an Ouster LiDAR unit, they’re increasingly also buying the software that processes the data. The sensor is becoming a distribution channel for perception software.

That pattern should look familiar. In every hardware market that scaled, the value migrated from the physical tool to the intelligence running on top of it. The same thing is happening in robotics perception. If you’re still thinking about the humanoid robot perception stack as a hardware problem, you’re looking at the wrong layer.

The core question: As sensor hardware commoditizes, where does the margin go? We think it goes to the fusion software layer. The companies building it today will tax every robot that ships when volumes arrive in 2028-2030.

What Sensor Fusion Actually Does

A humanoid robot carries between four and eight cameras, one or two LiDAR units, multiple depth sensors, and six or more inertial measurement units. Each sensor produces data in its own coordinate system, at its own clock rate, with its own noise profile. A camera delivers 30 frames per second of 2D pixel data. A LiDAR unit delivers 300,000 3D points per second. An IMU delivers acceleration and angular velocity at 1,000 Hz. None of them agree on where “here” is.

The robot needs one coherent model of the world, updated in real-time, accurate enough to grab a cup off a table without crushing it. Sensor fusion is the software that makes that happen. The pipeline has four stages.

1. Calibration aligns every sensor to a common coordinate frame. Get this wrong by two degrees and the robot’s hand lands three centimeters from where its eyes said the cup was.

2. Synchronization time-aligns data from sensors running at different rates. The IMU fires at 1,000 Hz. The camera fires at 30 Hz. The LiDAR spins at 20 Hz. The fusion algorithm interpolates, predicts, and merges these streams into a single timeline without introducing lag.

3. Filtering removes noise, handles occlusions, and manages conflicting data. When the camera says there’s an obstacle at 1.2 meters and the LiDAR says 1.4 meters, the software decides which sensor to trust based on noise models and confidence scores.

4. Registration combines 3D point clouds from different sensors into one unified model. Then SLAM (simultaneous localization and mapping) answers the two questions every mobile robot needs answered constantly: Where am I, and what’s around me?

Object recognition sits on top of the fused 3D model: “that’s a coffee mug, it’s full, and it’s on the edge of the table.”

The compute budget runs $500 to $1,500 per robot in dedicated neural processing units. That’s not a cloud workload. Every watt matters. Every gram matters. The fusion software has to run on the robot, in real-time, with a power budget that wouldn’t charge your phone.

Hardware Down, Software Up

The sensor itself is becoming the cheapest part of the stack.

Hesai has deployed more than one million LiDAR units. That volume turns a specialized optical instrument into a manufactured commodity. The engineering moat that justified premium pricing five years ago is eroding as manufacturing scales and alternative architectures (solid-state, flash, MEMS) compete on cost.

Camera modules already operate at 15-25% gross margins. LiDAR hardware sits at 25-40%. We saw this pattern in CMOS image sensors: Sony, OmniVision, and Samsung now compete in a market where the sensor itself accounts for less than 10% of the total camera system value. The value migrated to image signal processors, computational photography algorithms, and the AI models that turn raw pixels into usable data.

LiDAR is following the same trajectory. The hardware is commoditizing. The software that processes the point cloud is concentrating. If you’re investing in the humanoid robot perception stack based on who makes the best LiDAR unit, you’re looking at the commodity layer.

The specific cost trajectories and technology comparisons are covered in Part 1 of this series. The point here is directional: hardware margins will compress. The question is where the margin goes.

The Margin Stack

Gross margins tell the story in a single number.

Fusion software runs at 60-80% gross margins because every additional robot running your stack costs almost nothing in marginal compute. The fixed cost of developing the algorithms amortizes across millions of units. Hardware doesn’t work that way. Every sensor costs roughly the same to build whether you ship one or one million.

Ouster is betting on this margin migration. Sell sensors at hardware margins. Sell the software that processes sensor data at SaaS margins. The sensor becomes the customer acquisition cost for the software business.

The companies that still think they’re in the sensor business are fighting over margins that will be gone by 2028. The companies that know they’re in the spatial intelligence business are building the moat that will define the stack.

Subscribe to see where the value actually migrates, which tickers capture the fusion software layer, and why NVIDIA’s platform play is both the answer and the risk.

Subscribe to finish the article. See where the value actually migrates and which tickers capture the fusion software layer.

Subscribe now

Read more

ATI’s Gamma series force/torque sensor costs $5,000 to $10,000. Chinese alternatives deliver comparable performance for $1,500 to $3,000. At 10,000 robots per year, that’s $100-200 million in annual sensor spend. No CFO ignores that number. The cost gap is real. The question is why it exists, and whether it’s permanent.

In Part 1 of this series, we showed the $168 million market nobody’s tracking. In Part 2, we showed the touch gap: 12 sensing points on a robot fingertip versus 2,500 on yours. This is the part where it gets uncomfortable.

The Flywheel

Chinese sensors are cheap for a reason, and the reason compounds. China has 60-70 humanoid robot companies, more than 40% of the global total. That domestic scale creates feedback loops Western suppliers can’t access. Chinese sensor companies iterate on OEM feedback in weeks. They customize in days. Every cycle drives the cost advantage deeper, which pulls in more domestic customers, which drives more scale. The flywheel spins faster every quarter.

62% to 72.6% in one year.

Ten percentage points of consolidation. 70,000+ joint force sensor units shipped in H1 2025. 95% domestic market share in that category.

They're not alone. SRI Sensor (宇立仪器) holds 38.1% of the overall China 6-axis F/T sensor market. Kunwei Technology (坤维科技) drafted China's national standard for robot sensor testing. When your company writes the national testing standard, that's a regulatory moat. Kelio Sensing (柯力传感), publicly traded in Shanghai (603662.SS), has seen its market cap triple in 18 months.

China has 60-70 humanoid robot companies, more than 40% of the global total. GGII projects 12,300 six-axis F/T sensor shipments in 2025, with humanoid applications accounting for more than 30% of total demand. A year ago, humanoid was barely a line item. Now it's driving the market's growth trajectory. Chinese sensor companies get feedback loops from domestic OEMs that Western companies can't access. They iterate faster. They customize faster. They scale faster. That combination of production dominance and domestic demand creates a flywheel that's hard to compete with from the outside.

30-50% cost advantage. 28% of global demand.

China isn't just manufacturing sensors. It's consuming them.

A note on sourcing: the 72.6% and 38.1% figures trace to a single GGII report. We’ve cross-referenced against public filings and industry contacts where possible. Treat the directional trend as solid; the exact percentages as approximate. If anything, the real concentration may be higher. GGII’s methodology counts domestic sales, not design wins for future platforms where Bluedot is also winning.

The Double Chokepoint

The sensor flywheel is bad enough on its own. It’s worse when you remember the magnets.

In our Materials series, we showed how China controls 94% of NdFeB magnet production globally. When export controls tightened, prices surged 200-300%. Lead times stretched to 12 months. Some OEMs redesigned entire actuator assemblies to reduce magnet dependency. Companies that had sourced exclusively from Chinese suppliers found themselves with no fallback.

We showed you that. What we didn’t show you until now is that the same playbook is running in sensors. Different material. Same leverage. Same country.

94% of magnets. 72.6% of sensors.

Two chokepoints, one country. When China restricted rare earth exports in 2010, prices spiked 300% overnight.

^{Source: GGII, Machine Narratives analysis}

Every humanoid robot requires neodymium magnets in its actuators and force/torque sensors in its joints. China dominates both supply chains. The NdFeB chokepoint took a decade to form: a slow consolidation driven by state subsidies, environmental deregulation, and strategic acquisitions that Western competitors watched happen in real time and did nothing about. The sensor chokepoint is forming in two to three years. Same playbook. Faster execution.

Now you see why two chokepoints is fundamentally different from one. The first is a cost problem. The second is an existence problem.

Consider the scenario. A mid-tier Western OEM targeting commercial deployment in 2027 needs 12-16 force/torque sensors per unit and 20-30 neodymium magnets per actuator assembly. If China restricts magnet exports, the robot costs 20-30% more to build. It still ships. Now add sensor restrictions. There are fewer than five qualified Western F/T sensor suppliers globally. Their combined capacity covers maybe 30% of projected demand. Qualifying a new sensor takes 12-18 months. The robot doesn’t ship.

^{Source: Machine Narratives analysis}

Nobody is aggregating this risk. The sell-side covers magnets in one report and sensors in another. We've read the robotics coverage from every major bank. Not one models the compounding dependency. They treat magnets and sensors as independent variables. They're not. They're correlated exposures to the same country risk. If China restricts both, there is no humanoid robot industry outside China.

Read more

A camera-based perception stack costs $120 to $400. A LiDAR-supplemented stack costs $500 to $4,000. Same spot on the robot’s head. Same job description. The price gap is 5 to 10x, and the humanoid robotics industry is splitting into camps over which sensor technology earns the right to sit there. The winner determines which supply chain captures value for the next decade. We have seen this movie before.

If this analysis is useful, subscribe for free to follow the full Eyes of the Machine series.

The Format War Precedent

Toshiba surrendered on February 19, 2008. The Blu-ray versus HD DVD format war lasted two years, burned billions, and ended with a single winner. The lesson was not about which technology was “better.” HD DVD was cheaper. Blu-ray had more storage. The lesson was about adoption. Sony put Blu-ray in every PlayStation 3. Toshiba could not match that installed base. The supply chain locked in around Blu-ray. It still exists today, 18 years later.

The perception stack format war is happening right now. Once enough OEMs commit to a sensor architecture, the supply chain crystallizes. Component manufacturers optimize for the winning format. Software stacks mature around it. Switching costs compound. The technology that wins does not just win a spot on the robot. It wins a decade of supply chain lock-in.

The question is which camp has the PlayStation 3.

The Three Contenders

Every humanoid robot needs to answer one question: what is around me, and how far away is it? Each sensor answers that question differently, with a different cost structure and a different weakness.

1. Cameras are the cheapest option. A typical humanoid carries six to eight modules covering 360 degrees. Total cost: $90 to $400. Sony, Samsung, and OmniVision dominate the CMOS image sensor market. The weakness is compute: cameras need light, and extracting depth from 2D images requires GPU-intensive AI models. That compute cost shows up in the power draw and thermal envelope of the robot’s head.

2. LiDAR shoots laser pulses and measures return time to build a precise 3D point cloud. It works in complete darkness and delivers millimeter-level depth accuracy. The weakness is cost. A single LiDAR unit can cost more than the entire rest of a camera-based stack. For indoor humanoid applications, 200-meter range is overkill. The supply chain is thinner: Hesai, LIVOX, Luminar, and Ouster are the primary suppliers. The landscape is consolidating.

3. Depth sensors (structured light, time-of-flight, stereo) fill the gap. Better depth than a camera alone, cheaper than LiDAR, optimized for manipulation-range tasks at 0.5 to 5 meters. Intel’s RealSense is the most widely deployed. Orbbec and Stereolabs (now owned by Ouster) compete in the same space. The weakness is range. A depth sensor that works for picking up a coffee cup does not work for navigating a warehouse.

The cost spread is not subtle: a 5 to 10x gap that shows up directly in the bill of materials.

The camera camp argues the gap proves cameras win on economics. The LiDAR camp argues the gap is closing. Both are right. The question is timing.

The OEM Bets

The camps are forming. Five of the most visible humanoid robotics companies have made their sensor choices.

1. Tesla is the loudest voice in the camera-first camp. Optimus uses cameras exclusively. No LiDAR. No dedicated depth sensors. Tesla’s bet is that its AI vision stack, trained on billions of miles of driving data, can extract sufficient 3D understanding from 2D feeds. The cost argument is devastating: Tesla’s perception stack is the cheapest in the industry by a factor of five. The risk is that “sufficient” 3D understanding is not the same as “reliable” 3D understanding for manipulation tasks where millimeter precision matters.

2. Figure AI takes the pragmatic middle ground. Figure 02 pairs cameras with Intel RealSense depth sensors. No LiDAR. Better manipulation-range perception than cameras alone, at a fraction of LiDAR’s cost. Figure’s stack lands in the $200 to $500 range.

3. Unitree surprises. The G1 and H1 carry a LIVOX-MID360 LiDAR unit plus depth cameras at roughly $800 to $1,200. Unitree absorbed that cost because the robot’s navigation and SLAM capabilities required it. They are buying all three.

4. 1X Technologies uses camera pairs to derive depth through stereo vision, a technique that is computationally expensive but avoids dedicated depth hardware. Their bet is that compute gets cheaper faster than sensors get cheaper.

5. Boston Dynamics uses LiDAR plus cameras on Atlas. When a company that has spent 30 years building robots says LiDAR is necessary, the market listens.

Two things stand out. The camera-first camp (Tesla, Figure, 1X) has more adherents arriving through different engineering philosophies. The LiDAR camp (Unitree, Boston Dynamics) includes the companies with the most real-world deployment experience. The companies that have shipped robots into messy environments chose LiDAR. The companies building for scale chose cameras.

The Convergence

LiDAR costs are falling. Fast.

In 2025, a solid-state LiDAR unit suitable for indoor humanoid applications costs $500 to $2,000. By 2028, industry roadmaps target $200 to $500. Hesai, Luminar, and Ouster have all published product roadmaps with sub-$500 units for robotics applications by 2027 to 2028. The cost reduction is driven by silicon photonics integration (moving from discrete optics to chip-scale LiDAR) and volume scaling from the automotive market.

The camera stack is not standing still, but the absolute cost floor is already low. A camera module costs $15 to $50. There is not much room to compress. The total camera-based stack sits at $120 to $400. That number will not drop below $100 without a fundamental change in camera technology.

If LiDAR reaches $200 to $500 by 2028, the cost argument for camera-first perception evaporates. The 5 to 10x gap collapses to a 1.5 to 3x gap. At that differential, the capability advantages of LiDAR (works in darkness, native 3D, no compute-intensive depth estimation) become the dominant decision variable.

There is a second-order effect. When LiDAR costs fall below $500, the “LiDAR plus cameras plus depth sensors” stack that Unitree currently ships at $1,000 to $2,500 drops to $500 to $1,200. That is within striking distance of Figure’s camera-plus-depth stack ($200 to $500). A LiDAR-supplemented stack that costs 20% more than a camera-only stack but delivers 50% better depth accuracy becomes an easy engineering call.

The camera camp’s strongest argument is cost. That argument has a shelf life of 24 months.

If you want to understand where the supply chain value actually accrues, subscribe for free. Part 2 covers the fusion layer.

The Fusion Layer

Regardless of which sensor dominates the robot’s head, the data has to be combined into a single, coherent model of the world. That is the fusion layer. And the companies building it do not need to pick sides.

• Ouster acquired StereoLabs in February 2026. On the surface, this looks like a LiDAR company buying a camera company. Underneath, it is a fusion play. Ouster now owns both the LiDAR hardware and the camera-based depth perception software. They can sell a combined stack to any OEM, regardless of camp. The acquisition signals that the smart money is not betting on one sensor. It is betting on the ability to combine all of them.

• Bosch is building multi-sensor perception modules that integrate cameras, radar, and LiDAR into a single package. Bosch is an automotive Tier 1 supplier bringing automotive-grade sensor fusion to robotics. The modules are not cheap, but they are certified, tested, and available at volume. For OEMs that do not want to build their own fusion stack, Bosch offers a turnkey solution.

• NVIDIA’s Isaac platform is positioning as the operating system for robot perception. Isaac does not care whether the input comes from cameras, LiDAR, or depth sensors. It processes all three. NVIDIA’s bet is that the value in robotics is not in the sensors (which are commoditizing) or in the actuators (which are commoditizing) but in the compute layer that makes sense of the sensor data.

The fusion layer is where the real supply chain value accrues. The sensor format war determines which hardware suppliers win. The fusion layer determines which platform suppliers win. And platform suppliers tend to capture more margin than hardware suppliers. In automotive, the Tier 1 suppliers (Bosch, Continental, Denso) capture more margin than the component manufacturers below them. In robotics, the same pattern will emerge.

Ouster’s software-attached bookings doubled in 2025. Software now ships with more than 15% of sensors. The fusion layer is already monetizing before the humanoid market even arrives.

The Bear Case

Two risks. Both real.

1. The numbers are small.

   Humanoid-specific sensor revenue is approximately $59 million today. That is a rounding error in a $12 to $13 billion total perception TAM. The humanoid-specific slice requires shipments to grow from roughly 13,000 units in 2025 to 1.4 million units in a decade. That is a 100x increase. The robotics market has burned forecasters before. Bull cases that assume hockey-stick adoption curves in hardware markets have a poor track record.

2. AI might make the debate irrelevant.

   If AI vision models advance to the point where a camera-only stack can extract reliable 3D depth from 2D images in real time, the entire LiDAR value proposition collapses. Not gradually. Suddenly. Tesla is already betting on this. Their FSD stack extracts depth from monocular cameras using neural networks trained on billions of labeled frames. If that approach transfers to manipulation tasks, cameras win by default. No format war. No convergence. Just cameras and compute. The counter-argument is that AI depth estimation is probabilistic, not deterministic. LiDAR measures distance with millimeter precision. AI estimates distance with confidence intervals. For navigation, estimation is good enough. For manipulation, where a robot hand needs to grasp a 10mm bolt, estimation might not be. The question is whether AI closes that precision gap before LiDAR closes the cost gap. We think the AI risk is the more dangerous one for the LiDAR supply chain. The convergence argument assumes LiDAR costs fall while camera capabilities stay flat. If AI vision models improve faster than LiDAR costs decline, the camera camp wins on both cost and capability. And the LiDAR supply chain becomes stranded.

Key Takeaways

• The format war is real. Once enough OEMs commit to a sensor architecture, the supply chain locks in. Switching costs compound. The winner captures a decade of value.

• The camera camp wins on cost today. Tesla, Figure, and 1X all arrive at camera-first through different engineering philosophies. The 5 to 10x cost gap is the camera camp’s strongest argument.

• The cost argument has a 24-month shelf life. LiDAR is targeting $200 to $500 by 2028. At that price, capability advantages (darkness, native 3D, no compute-intensive depth estimation) become the dominant decision variable.

• The fusion layer wins regardless. Ouster, Bosch, and NVIDIA are building platforms that process whatever sensors the OEM chooses. Platform suppliers capture more margin than hardware suppliers. This pattern held in automotive. It will hold in robotics.

• The real risk is AI, not cost. If AI depth estimation reaches sub-5mm precision at manipulation range, the LiDAR supply chain becomes stranded. That is the bear case that matters.

Format war timelines

VHS vs Betamax, 5 years. Blu-ray vs HD DVD, 22 months. The perception stack format war started in 2025. The clock is running.

We think the camera-first camp wins the next 24 months on cost. We think the LiDAR-supplemented camp wins the decade on capability. And we think the fusion layer captures the margin either way.

Part 2 examines the fusion layer in detail: who owns it, who is building it, and where the margin actually accrues. Subscribe for free to follow the series.

Frequently Asked Questions

Why can’t robots just use cameras like self-driving cars?

Self-driving cars operate outdoors in well-lit conditions with structured roads and known map data. Humanoid robots operate indoors, in variable lighting, manipulating objects at close range. The perception requirements, depth accuracy under 5mm, object recognition at arm’s length, operation in low-light rooms, are different in kind. What works for a car at 60 mph on a highway does not translate to a robot reaching for a door handle in a dim hallway.

Is LiDAR going to become as cheap as cameras?

LiDAR costs have fallen from $75,000 in 2015 to $500–2,000 today, with a $200–500 target by 2028. At that price, LiDAR would still cost 4–13x more than a camera module ($15–50). The gap narrows but does not close. The real question is whether LiDAR becomes cheap enough that the capability advantage, deterministic 3D point clouds that work in darkness, outweighs the remaining price premium. We think the answer is yes, at roughly the $300 unit price threshold.

What happens if AI makes cameras “good enough” for 3D perception?

This is the wildcard. Vision-language-action models are learning to extract depth from 2D images with improving accuracy. If that capability matures to sub-5mm precision at manipulation range, camera-first perception wins by default. No format war, just cameras and compute. But we are not there yet. Current AI depth estimation is probabilistic, not deterministic. For navigation, estimation is good enough. For grasping a 10mm bolt from a parts bin, it is not.

Which companies benefit regardless of which sensor wins?

The fusion layer, companies building software that combines data from multiple sensor types, wins either way. Ouster (via its StereoLabs acquisition), Bosch (via its NEURA Robotics partnership), and NVIDIA (via the Isaac platform) all sell to both camps. The fusion layer does not need to pick a side. It processes whatever sensors the OEM chooses and extracts margin at the platform level.

Eyes of the Machine: A 3-Part Series
1. Cameras, LiDAR, Depth Sensors: The $38B Format War
2. Software Eats the Sensors
3. Sony Won the Last Camera War. Who Wins This One?

The first time I watched a robot try to pick up a paper cup, I felt sorry for it.

It was a demo at a robotics lab. The robot’s hand closed around the cup with the same mechanical certainty it would use on a steel cylinder. Same grip. Same force. The cup collapsed instantly, coffee pooling on the table. The robot didn’t react. It opened its fingers, reset, and tried again with the same force. Same result.

The robot had cameras for eyes. It could see the cup. What it couldn’t do was feel the difference between a paper cup and a steel cylinder. It had no sense of touch worth speaking of.

The engineers weren’t embarrassed. They were used to it. “We’re working on it,” one said, and the conversation moved on to actuators and AI models. The things people actually want to talk about.

I went home and started digging. What I found was a big gap between what robots can sense and what humans can feel. Almost nobody outside the robotics world knows it exists.

In Part 1, we showed the $168M market that nobody is tracking. This article explains why that market exists and the engineering gap that makes these sensors so hard to build.

The Touch Gap in One Number: A robot fingertip has roughly 12 sensing points. Yours has 2,500 per square centimeter. That 200:1 ratio is the most expensive unsolved problem in robotics.

^{Sensing points per square centimeter: human fingertip vs. best commercial tactile sensor vs. best research prototype. Source: Machine Narratives Research, 2026}

What Tactile Sensing Actually Is

Force and torque sensors sit at the joints — wrists, elbows, shoulders, ankles. They measure mechanical load: how much force each joint is applying in each direction. These are mature. You can buy them from a dozen companies and have one shipped next week.

Tactile sensors are different. They sit at the contact surface — the fingertip, the palm, the forearm pressing against a doorframe. They measure what’s happening where the robot touches the world. Texture. Slip. Contact geometry. The local details that tell you whatyou’re holding, not just how hard you’re squeezing.

Your fingertip has about 2,500 mechanoreceptors per square centimeter. Each one reports something different: pressure, vibration, stretch, texture. They fire signals to your brain hundreds of times per second, telling you whether you’re holding a grape or a golf ball without looking. You do this unconsciously. You’ve done it your entire life.

The best commercial tactile sensor today has roughly a dozen sensing points. Research prototypes push into the hundreds — some impressive work from Stanford and Chinese universities gets close to a thousand. But those are hand-built, one-off devices costing thousands per patch. They’re not products.

Market Size:

• Force/torque sensors: ~$2.3 billion (2026).

• Tactile sensors: ~$168 million.

The tactile market is 14x smaller, roughly where force/torque was twenty years ago. Technically proven, commercially immature, waiting for an application that demands it at scale.

That application just showed up. If this analysis is useful, subscribe to follow the full sensor series.

Read more

In September 2010, a Chinese fishing trawler collided with Japanese coast guard vessels near the Senkaku Islands. China restricted rare earth exports to Japan. Neodymium oxide went from under $10/kg to nearly $239/kg. Molycorp went public at $14, hit $70 by early 2011. Lynas saw similar euphoria.

Then it collapsed. China backed off. The WTO ruled against Beijing in 2014. Prices cratered. Molycorp filed for bankruptcy in 2015. The entire rare earth thesis of the early 2010s turned out to be a geopolitical trade masquerading as a structural one.

The question before we talk about NdFeB magnet stocks today: is this time different?

I think it is. But the bear case deserves steelmanning first.

If this analysis is useful, subscribe for free to follow the full NdFeB series.

The 2010 Trade Failed. This One Might Not.

The 2010 analogy has a flaw most analysts miss: they think the lesson is “China weaponized rare earths and the world built alternatives.” The real lesson is that the world started to build alternatives, the crisis passed, and everyone lost interest. Lynas got its mine running. Molycorp got Mountain Pass back online. Neither became competitive with Chinese producers on cost. When prices collapsed, the economics evaporated.

The structural dependency never resolved. It stopped being urgent. Sixteen years later, China’s share of rare earth processing is higher than in 2010.

In April 2025, China imposed export controls on finished NdFeB magnets, not just raw ores. Automakers shut down production lines. The 2010 crisis was about export quotas on raw materials, resolved by alternative supply coming online and WTO pressure. This time, China restricts finished magnets, and the WTO can’t do much about controls framed as national security measures. The legal resolution from 2010 doesn’t apply.

The demand profile is also different. In 2010, the rare earth trade was a bet on a policy dispute. Today it’s a bet on physics. You cannot substitute away from NdFeB in humanoid robots. Ferrite magnets are 11.5 times weaker. Iron nitride magnets are at TRL 2 or 3, a decade from commercialization. Each humanoid robot needs 28 to 76 motors, each requiring NdFeB magnets, totaling 1 to 5 kilograms per unit.

The Demand Curve Is Real

The numbers stack fast across three vectors.

The EV transition alone would strain non-China supply. Adding robotics creates a demand profile the Western supply chain cannot serve.

The Timing Mismatch Is the Trade

Non-China NdFeB capacity today is roughly 12,000 to 18,000 tonnes per year. By 2028, it might reach 25,000 to 35,000. By 2030, optimistically 35,000 to 50,000.

Humanoid robot demand is small in absolute terms: 100 to 500 tonnes by 2028, maybe 1,000 to 2,500 by 2030. The squeeze isn’t about volume. It’s about priority.

Defense contractors and major EV OEMs will lock up most non-China capacity through long-term contracts. They’re already doing it. Humanoid robot companies are late to the procurement queue, startups competing with established industrial giants for every kilogram.

Chinese humanoid OEMs like Unitree, UBTECH, and Tesla’s Shanghai operations have direct access to Chinese magnet supply. American and European humanoid companies don’t. The supply gap isn’t just an investment opportunity; it’s a competitive moat for Chinese robotics.

The supply gap is real. The timing mismatch is real. The demand curve is locked in by physics, not policy. The question is not whether this trade works — it’s who captures the value when it does.

Below, I break down the only mine-to-magnet bet in the Western world, the boring alternative that might be smarter, why Japan’s position is weaker than it looks, the bear cases I find most convincing, and the specific trade setup I’d consider if I were allocating capital to this thesis today. Three companies. One clear winner. Two that might surprise you.

Read more

1X’s NEO hand has 25 degrees of freedom, ±0.2mm accuracy, 45N fingertip force, and is IP68 waterproof. It can fold laundry, crack eggs, and assemble furniture. On paper and in the videos, it’s the most advanced tendon-driven hand we’ve seen publicly demonstrated.

1X describes it as “an API to the physical world.” The metaphor is neat, but it misses the point.

APIs are only as good as the infrastructure behind them. And the infrastructure to manufacture this level of dexterity at scale doesn’t exist yet, not at the volumes 1X (and the rest of the industry) will eventually need.

The real question is no longer whether someone can build a highly dexterous hand. It’s whether anyone can build thousands of them at a cost that supports a viable robot business. That’s a supply chain and manufacturing problem, not an engineering one. And 1X’s own teardown quietly reveals how far the industry still is from solving it.

Read the full 1X announcement, then come back, because we’re about to unpack what they didn’t say.

New here? Start with the full supply chain research map

What 1X Actually Built

Let’s be clear: the engineering is real. Force-controlled quasi-direct-drive The engineering is real. Force-controlled quasi-direct-drive tendons mean losing mechanical advantage and gaining compliance. That trade-off only works with a fast control loop, strong tendons, and precise sensors. 1X appears to have solved all three.

The standout is shear-capable tactile sensing. Most hands detect normal force. Detecting shear force (sliding across a surface) requires denser sensor arrays and more complex signal processing. 1X integrated this into a hand that also hits IP68 waterproofing and food-safe compliance. That’s packaging problems solved that most labs haven’t attempted.

The demonstrations speak for themselves: LEGO assembly, individual screws, spinning light bulbs. 1X built something that works.

If this breakdown is useful, subscribe for free to follow the full humanoid supply chain series.

The Supply Chain Reality

1X TECHNOLOGIES’ CLAIM

“Every unit is built end-to-end in-house, from tendon materials and 1X Motors to the final soft polymers, skin, and tactile sensing stack.”

Source: 1X Technologies, NEO’s Hands

“End-to-end in-house” means assembly, not vertical integration. The raw materials, motor subcomponents, and sensor fabrication all come from external suppliers.

Three critical component layers stand between 1X and 10,000 hands. Each one has its own supply chain bottleneck. None of them are ready to scale.

Tendons: The Market That Doesn’t Exist Yet

25 DOF = 25-40 tendons per hand.

At 10,000 hands: 250,000+ tendons.

UHMWPE tendon material market: $113M (2025) → $1.23B (2032) at 40.6% CAGR (QY Research).

Dexterous hand tendon rope sub-market: $0.6M (2024) → $444M (2032) at 86.1% CAGR (IntelMarketResearch). Growing from essentially zero.

Carl Stahl Group leads the micro wire rope niche. Tesla Optimus Gen3, Shadow DEX-EE, and LinkerBot L30 all depend on the same supply chain.

Motors: Designed In-House, Built by Someone Else

Hollow-cup motors depend on precision magnets, micro windings, and encoders from a concentrated supplier base. The motor isn't the bottleneck. The motor components are. NdFeB magnets for micro-actuators come from a handful of suppliers. Precision wire winding has limited global capacity.

Sensors: Millions of Components, No Supply Chain

SENSOR SUPPLY CHAIN

Kunwei Technology (Beijing) raised ¥100M Series B+ (March 2026) to scale humanoid F/T sensor production. Each humanoid needs 2-6 wrist/ankle F/T sensors plus 28+ torque/pressure elements, requiring 10-30 signal chain ICs per unit. The signal chain is where cost and complexity live.

1X's quasi-direct-drive architecture reduces discrete joint torque sensors. Tendon tension can be inferred from motor current, eliminating the need for expensive per-joint torque sensing. But wrist-level F/T sensing and tactile skin are non-negotiable. Shear-capable tactile sensing requires dense arrays of micro-force sensors with their own signal processing. At 10,000 units, you need millions of sensor elements. That supply chain doesn't exist yet.

BOM Reality Check: $20K Robot, $5K Hands

To understand what 1X’s hand costs at volume, you need a baseline. Shadow Robotics (London) has spent 20 years building the gold-standard dexterous hand — the Shadow DEX-EE. 24 DOF, tendon-driven, used in research labs worldwide. It costs $100,000+ per unit and has never shipped at volume. It is the engineering benchmark.

On the other end, AGILINK (AGIBOT’s hand subsidiary) has shipped 8,000+ OmniHand units. 20 DOF, simpler design, no shear-capable tactile skin. But they are the manufacturing benchmark — the only dexterous hand company actually at volume.

1X sits between them: more complex than AGILINK, cheaper than Shadow, targeting 10,000 units by December. The BOM comparison shows why that’s hard.

THE COST REALITY

1X’s hand BOM: $2,400-5,000 at 10K volume. That’s 12-25% of a $20,000 robot’s total BOM in the hands alone. AGILINK comes in lower ($1,800-3,400). 20 DOF needs fewer tendons and simpler sensors. The 5-DOF gap compounds: each DOF adds a tendon, motor, sensor channel, and control loop.

1X’s $499/month subscription model (Forbes) makes more sense than the $20,000 purchase price. At $3-5K per hand BOM, the margin on a one-time purchase is razor-thin. The subscription front-loads hardware cost into recurring revenue: the Tesla FSD playbook.

What This Means

MANUFACTURING REALITY CHECK

1X: 10,000 target by Dec 2026. Most complex hand ever. Supply chain doesn’t exist at this scale.
AGILINK: 8,000+ shipped (per AGIBOT’s 2026 unicorn announcement). Simpler hand (20 DOF). Already at volume.
Shadow: 20 years of R&D. Zero volume. $100K+ per unit.Three players. Three very different manufacturing realities.

The hand race is now a manufacturing competition. The winners won’t have the best demo video. They’ll control the tendon supply chain, the F/T sensor manufacturing base, and the hollow-cup motor component ecosystem.

The next 12 months will tell us whether 1X can turn an engineering breakthrough into a manufacturing reality. Watch the component suppliers: Carl Stahl Group leads the micro wire rope niche. Kunwei Technology’s ¥100M Series B+ signals smart money flowing into F/T sensor capacity. The NdFeB magnet market for micro-actuators is another constraint nobody’s pricing into humanoid valuations.

Key Takeaways

• 1X built the most dexterous tendon-driven hand ever demonstrated. 25 DOF, shear-capable tactile sensing, IP68 waterproof. The engineering is genuine.

• “End-to-end in-house” means assembly, not vertical integration. 1X doesn’t mine UHMWPE, manufacture hollow-cup motor subcomponents, or fabricate tactile sensor arrays. The distinction matters when the supply chain is the constraint.

• Hand BOM: $2,400-5,000 at 10K volume. 12-25% of a $20,000 robot in the hands alone. The $499/month subscription is the rational path.

• AGILINK has shipped 8,000+ hands (per AGIBOT’s 2026 unicorn announcement). Simpler design, fewer DOF, but manufacturing volume is manufacturing volume.

• Component suppliers capture margin regardless of which OEM wins. Carl Stahl, Kunwei Technology, NdFeB magnet manufacturers are picks-and-shovels plays on humanoid scaling.

Manufacturing races are won by whoever owns the supply chain, not whoever owns the best design. The tendon material, motor components, and sensor ICs are where the real leverage lives.

Related Reading

• The Actuator Cartel — Who actually controls the motors, reducers, and drive systems every humanoid depends on

• The Magnet Nobody Talks About — Why NdFeB supply is the hidden bottleneck behind every robot motor

• Tesla, Figure, and Unitree’s Supplier Wars — The supplier-level competition for the same components 1X needs

• The Sensor Nobody Covers Series — The tactile sensor supply chain behind the BOM table in this article

  ---

Frequently Asked Questions

If 1X’s hand BOM is $3-5K at volume, can they hit a $20K robot price?

The hand alone consumes 12-25% of the $20K target. Add legs, torso, battery, compute, and software, and the total BOM likely exceeds $15K at 10,000-unit volume. That leaves thin margin for a consumer product. 1X’s $499/month subscription model (confirmed by Forbes) makes more sense than the $20,000 purchase price. It front-loads the hardware cost into recurring revenue.

Is AGILINK a serious competitor to 1X in dexterous hands?

AGILINK has shipped more dexterous hands than anyone. Their OmniHand is simpler (20 DOF vs 25 DOF), lower force, no shear-capable tactile skin. But manufacturing volume is manufacturing volume. In a market where supply chain control determines winners, the company that’s already shipping 8,000+ units has a structural advantage over the company targeting 10,000 by December.

Why does the tendon supply chain matter so much?

UHMWPE tendon material market is $113M today. It needs to hit $1.23B by 2032 to meet projected humanoid demand. That 10x growth in seven years requires massive capacity expansion. If capacity lags demand, tendon prices spike, margins compress, and production targets slip. Every humanoid maker depends on the same material. The tendon market is the bottleneck nobody’s pricing into humanoid stocks.

Should investors focus on humanoid OEMs or component suppliers?

Component suppliers. The tendon material, F/T sensor, and micro-motor subcomponent companies capture margin regardless of which humanoid OEM wins. Carl Stahl Group, Kunwei Technology, and NdFeB magnet manufacturers are picks-and-shovels plays on humanoid scaling. Hand assemblers compete on price. Component suppliers compete on capacity. And capacity is scarce.

What’s the most overhyped claim in 1X’s NEO hands announcement?

“End-to-end in-house.” 1X designs the hand. They don’t mine the UHMWPE, manufacture the hollow-cup motor subcomponents, or fabricate the tactile sensor arrays. “End-to-end” means they assemble the hand in-house using components from external suppliers. That’s integration, not vertical integration. The distinction matters when the supply chain is the constraint.

Where does Tesla Optimus fit into the hand race?

Tesla’s Gen 3 hand promises 22 degrees of freedom with a tendon-driven architecture — three thin tendons per finger, all 50 actuators relocated to the forearm. On paper, it’s the most direct competitor to 1X’s approach. In practice, Tesla has delayed the Gen 3 unveiling repeatedly and has not demonstrated the hand publicly. The patent filings are detailed. The working hardware is not. Tesla’s vertical integration gives them a structural advantage if they can execute — they design their own actuators, control systems, and manufacturing processes. But until we see a Gen 3 hand picking up screws on video, Tesla remains the elephant in the room: the competitor everyone assumes will arrive, but nobody can point to.

In 2008, the entire lithium-ion battery market for electric vehicles was worth about $400 million. Four hundred million dollars. Toyota was selling more Priuses in a quarter than the entire global EV battery industry was generating in revenue. If you went to a sell-side conference and pitched lithium-ion batteries as an investment theme, people looked at you like you were trying to sell them swampland.

Ten years later, that market was $44 billion. The companies that positioned early (Panasonic, CATL, LG Chem) became the backbone of the largest industrial transition since the internal combustion engine replaced the horse. Panasonic locked in its supply deal with Tesla. CATL rode the Chinese EV boom to become the world’s largest battery manufacturer. LG Chem spun off its battery division into a standalone company worth more than the parent. The early movers didn’t just participate in the growth. They defined it.

I think about that pattern a lot. The market is invisible, then it’s inevitable. And right now, I’ve found another one.

$168 million. That’s the entire global market for tactile sensors in humanoid robots. I had to read the number twice. I’ve been covering robotics and semiconductor supply chains for years, and I’d never seen this figure before. Not because it’s new. Because nobody thought it was worth tracking.

Let me explain why I think that’s about to change.

^{A tactile force/torque sensor, smaller than a fingertip, that enables robotic touch perception.}

The Number That Caught My Eye

$168 million. That’s the total addressable market for tactile sensors in humanoid robots, according to IntelMarketResearch’s 2026 report. To put that in perspective, Novanta paid $172 million to acquire ATI Industrial Automation (a single sensor company) in 2021. The entire global market for these sensors is worth less than one acquisition from five years ago.

$172 million. That’s what Novanta paid to acquire ATI Industrial Automation — a single sensor company — in 2021. The entire global market for tactile sensors in humanoid robots is worth $168 million. One acquisition cost more than the whole market.

It wouldn’t cover the quarterly R&D budget of a mid-tier automaker. It wouldn’t register on the radar of a thematic ETF manager. It’s the kind of number that gets buried on page 47 of an industry report that maybe three people read, and two of them work at the company that wrote it.

I didn’t stumble across this number through some brilliant research process. I found it because I was trying to build a bill-of-materials model for humanoid robots. Trying to understand what actually goes into these things, component by component. I kept getting stuck on the sensing layer.

The actuator numbers are well-documented. Motor specs, gear ratios, torque curves: all published. The compute costs are public. Nvidia’s Jetson pricing, Qualcomm’s robotics chips, the cost per TOPS: you can look it all up. But when I tried to find the sensor data, I ran into a wall. Nobody was aggregating it. Nobody was tracking the demand curve. The data existed, fragmented across a dozen research reports that don’t talk to each other, and nobody had pulled it together.

When I finally got the number, I thought it was wrong. A hundred and sixty-eight million dollars? For the sensors that give robots the sense of touch? For the technology that’s supposed to make humanoid manipulation viable? It seemed absurd.

Then I started to understand why it’s so small, and that’s when it got interesting.

The market is too niche for ETFs. It’s too small for sell-side coverage. It’s too early for earnings calls. The institutional world has a blind spot for markets under $1 billion: no model, no coverage, no analyst who wakes up thinking about tactile sensor procurement for humanoid robots. The companies making these sensors are small, mostly private, and operating at production volumes that would be dwarfed by a single large humanoid manufacturer’s annual needs.

That’s not a gap. That’s a blind spot. And blind spots are where alpha lives.

The Goldman Sachs Math

Then I found the projection that changed my mind.

Goldman Sachs projects 1.4 million humanoid robot shipments by 2035. That number alone is worth pausing on. 1.4 million humanoid robots, shipped and deployed, in just over a decade. But the number that made me sit up was what happens to the sensor market when you multiply it out.

$168 million becomes $3 billion.

I checked the math three times because I didn’t believe it. Let me walk through it so you can check it too.

The $3 billion figure isn’t a CAGR extrapolation. I want to be clear about that, because the two methods get confused constantly in market reports. A simple CAGR extrapolation of 11.7% over nine years would only get you to about $405 million. That’s meaningful growth, but it’s not the story.

The $3 billion comes from a bottom-up calculation: Goldman’s 1.4 million robot shipments by 2035, multiplied by the estimated sensor cost per robot. Each humanoid robot needs both force/torque sensors and tactile sensors. Force/torque sensors (the ones that measure load at wrists, ankles, and knees) are the mature market, roughly $2.3 billion globally and growing at about 6% annually. But tactile sensors (the ones that give a robot the ability to feel pressure distribution, texture, and slip) are where the exponential part kicks in.

When you scale from a few thousand prototype units to 1.4 million production units, the per-unit costs drop, the supply chains mature, and the total addressable market transforms. That’s the transition from lab curiosity to industrial commodity. I’ve seen this movie before. In lithium-ion batteries. In MEMS accelerometers. In CMOS image sensors. The pattern is always the same: the market looks too small to matter until it doesn’t.

And here’s the part that should make you pay attention: the Goldman baseline was set before Tesla committed $25 billion to AI and robotics for 2026. Before Unitree shipped 5,500 humanoid units in 2025 (the only verified mass-production data point we have). Before Figure AI started talking about production runs, not prototypes. The baseline was modeled when humanoids were still lab curiosities. The world has moved since then. Nobody revised the number.

Some forecasters use 39.2% CAGR for humanoid robot shipments, which would blow past the $3 billion figure well before 2035. I don’t know which number is right. But I think the direction is right, and in markets this early, direction matters more than precision.

Why Nobody’s Watching

I want to spend a minute on the institutional dynamics here, because I think they matter.

The reason nobody’s watching this market is structural. A $168 million market doesn’t clear the threshold for dedicated sell-side coverage. You can’t build a career on it. You can’t justify a conference panel. It doesn’t show up in anyone’s thematic ETF because the underlying companies are too small or too private. The sensor companies themselves are part of the invisibility problem: they’re small, they’re fragmented, and they’re not the kind of names that generate clicks.

Every institutional framework for evaluating robotics focuses on the same three things: actuators, AI, and compute. Those are the big-ticket items, the ones with billion-dollar markets and established supply chains. Sensors are a footnote. The bill of materials for a humanoid robot gets broken down into motors, batteries, controllers, and maybe “sensing systems” as a line item. Nobody drills into the sensing layer. Nobody aggregates the demand curve for tactile sensors across projected humanoid production.

Here’s what the career math looks like. A junior analyst at a major bank pitches their desk head on covering tactile sensors for humanoid robots. The desk head asks: “How big is the market?” The analyst says $168 million. The meeting lasts another thirty seconds. You can’t justify a headcount, a conference slot, or a morning note on a market that small. The analyst who writes that report is betting their career on a niche nobody at the firm has ever mentioned in an earnings preview. So nobody writes it. And the blind spot persists.

Think about what that means at scale. Every major bank, every asset manager, every research house runs the same calculus. The market is too small to cover, so nobody covers it. Because nobody covers it, the institutional world doesn’t know it exists. Because the institutional world doesn’t know it exists, nobody allocates to it. The feedback loop is self-reinforcing. The blind spot isn’t just structural. It’s stable.

The alpha isn’t in knowing something nobody else knows. It’s in paying attention to something nobody else thinks is worth paying attention to.

I’d argue this is the most important observation here. And the institutional incentive structure makes it almost impossible for anyone to be the first mover on this.

But the demand curve doesn’t care about institutional incentive structures. It just is.

The Timing

What makes this different from other small markets is the convergence of signals.

Tesla’s $25 billion commitment to AI and robotics in 2026 is not a research budget. It’s a production budget. That’s the kind of money you spend when you’re building factories, not prototypes. When the largest EV manufacturer in the world commits that kind of capital to humanoid robots, the supply chain implications ripple outward. Every component supplier in the chain gets pulled along.

Unitree shipped 5,500 humanoid units in 2025. Real units, shipped to real customers, with real sensors inside them. Not a press release about a prototype. That volume is small relative to the $3 billion projection, but it’s the first verified data point that humanoid production is real, not theoretical.

Figure AI is raising at a $39 billion valuation. Boston Dynamics switched from hydraulic to electric Atlas specifically to optimize for manufacturing. Agility Robotics signed a deal with Amazon for warehouse deployment. 1X Technologies is scaling its NEO line for home assistance. Sanctuary AI is targeting general-purpose labor. The list of companies making production bets, not science experiments, is growing every quarter.

The demand curve isn’t theoretical anymore. It’s happening. The question isn’t whether humanoid robots will ship in meaningful quantities. The question is how fast, and who supplies the components.

In the next article, we’ll show the engineering challenge — the touch gap between human and robot sensing. Your fingertip has 2,500 receptors per square centimeter. The best commercial tactile sensor has 12. That gap is real, and it’s expensive. But first, let’s talk about what that gap means for the market.

The sensor landscape is fragmented in a way that makes aggregation almost impossible. No single report pulls these names together. No ETF holds them. No sell-side model tracks the demand curve across projected humanoid production. Let me walk you through why.

The Incumbent

ATI Industrial Automation’s Nano17 is the de facto standard in every major robotics lab worldwide. Seventeen millimeters in diameter, 9 grams. If you’re doing force/torque research on a humanoid robot, you’re probably using an ATI sensor. Novanta (NOVT) acquired them for $172 million in 2021. At the time, that looked like a niche tuck-in. In hindsight, it might be the most prescient sensor acquisition of the decade.

Novanta’s market cap is around $5.5 billion today. ATI is a meaningful piece of their revenue mix, and if the tactile sensing wave drives demand for integrated force/torque plus tactile sensing across humanoid joints, Novanta benefits from the volume increase. It’s not pure-play, but it’s real revenue from real sensors on real robots. And the acquisition price tells you what Novanta’s leadership thought was coming. $172 million for a sensor company in 2021, before anyone was talking about humanoid production at scale. That’s conviction.

The Pure Plays

XELA Robotics, out of Tokyo and Taipei, makes 3D tactile sensors that detect forces as light as 0.1 gram-force. That’s roughly the weight of a mosquito landing on skin. Their customer list reads like a who’s who of robotics: Honda, Hitachi, Samsung, Denso, Sanctuary AI. Private, which means no direct public play. But they’re the name that comes up first in every conversation about tactile sensing.

Bluedot Touch controls 72.6% of the Chinese humanoid robot force/torque sensor market, per a 2025 GGII report. China is building its own humanoid supply chain deliberately, with government backing, and Bluedot is sitting at the center of it. Kunwei Technology drafted China’s national standard for robot sensor testing. When a company writes the standard, it tells you something about its position in the ecosystem. Both are private. Both are names that surface when you dig into the Chinese robotics supply chain. The fact that China has dedicated sensor companies writing national standards tells you how seriously Beijing takes the humanoid buildout.

The Enablers

GelSight makes high-resolution tactile imaging sensors: the kind that can map surface texture at micron-level resolution. SynTouch developed BioTac, a biomimetic tactile sensor that mimics the human fingertip. The company dissolved in 2024, but the technology set the standard for biomimetic sensing in research labs worldwide. Both names surface in every serious conversation about robotic touch.

And then there’s the connector and flexible circuit layer. Hirose Electric (6806.T) is developing miniature connectors and flexible printed circuits specifically for humanoid wiring harnesses. When a major connector company builds an entire product line around a single application, that tells you how serious the demand signal is. Hirose isn’t a sensor company. They’re an infrastructure company betting that humanoid robots will need a wiring architecture that doesn’t exist yet.

The problem is aggregation. These companies don’t show up together in anyone’s model. There’s no ETF, no index, no sell-side report that pulls together the demand curve for tactile sensors across projected humanoid production. The data exists, fragmented across a dozen research reports. But nobody’s built the composite picture.

I want to remind readers: the most cited number ($168 million) comes from a single report by IntelMarketResearch. No major research firm has published a comparable dedicated forecast for tactile sensors in humanoid robots. The methodology is somewhat opaque, and the definitional boundaries between tactile sensors, electronic skin, and haptic feedback are fuzzy: market researchers draw them differently. Hold the number loosely. But hold the direction tightly.

The Bear Case (Steelmanned)

Let’s steelman the counter arguments first, because that’s the right thing to do.

The bear case is straightforward. Humanoid robots are still mostly prototypes. Battery life is 2 to 4 hours. AI perception is immature. Dexterous manipulation is limited. Costs range from $50,000 to $250,000 per unit. If humanoid production stays niche (if these remain expensive toys for research labs and tech demos) the sensor market stays small. Not $3 billion. Maybe $400 million, which is what you’d get from a simple CAGR extrapolation of the current $168 million at 11.7% over nine years.

That’s still meaningful growth. But it’s not the kind that creates new market leaders or reshapes supply chains.

$400 million to $3 billion. Same market, two completely different stories. The bear case (CAGR extrapolation) says $400M. The bull case (bottom-up demand) says $3B. The direction is right either way.

For the bear case to hold, you need all of the following to be true simultaneously: Tesla’s $25 billion commitment doesn’t translate to production volume. Unitree’s 5,500 units in 2025 turns out to be a one-off. Boston Dynamics’ manufacturing pivot fails. Figure AI’s $39 billion valuation collapses. Every single production signal has to be wrong at the same time. Possible, sure. Probable? I don’t think so.

I think the bear case is plausible. I also think it’s wrong, and here’s why.

Tesla isn’t running a research lab. They’re building a factory. Unitree isn’t making prototypes. They’re shipping units. Boston Dynamics switched from hydraulic to electric Atlas specifically to optimize for manufacturing economics. These are production bets, not science experiments. The capital allocation tells you more than any market research report. When the largest companies in the world are committing billions to humanoid production, the demand curve for components follows.

The question isn’t whether demand for humanoid robot sensors will grow. It will. The question is whether it grows at 11.7% or 39% or something in between. And that range of outcomes is enormous for whoever positions early.

What I’m Watching

I’m not telling you to buy anything today. This market is too small, too fragmented, and too early for that kind of call. The companies closest to the tactile sensor opportunity (XELA, GelSight, SynTouch) are all private. The closest public proxy is Novanta, through ATI, and at a $5.5 billion market cap, it’s priced for steady growth, not explosive upside. The sensor demand curve could change that math, but it’s not a trade I’d make on a $168 million market.

What I am telling you is to watch this market. The companies that are positioned now (the ones with the sensor technology, the customer relationships, the production experience) will be the ones that benefit when the institutional world catches up.

When that first sell-side note drops, and it will, the analyst who writes it won’t be early. They’ll be late. The demand curve will already be established. The supply chain relationships will already be locked. The companies that got there first will have a significant advantage, and the coverage will read like a confirmation of what the market already priced in six months prior.

I’ve seen this pattern before. In lithium-ion batteries in 2008. In MEMS sensors in the early 2010s. In advanced packaging before CoWoS became a buzzword. The market is invisible, then it’s inevitable. The only question is how long the window stays open.

The first-mover advantage goes to whoever understands the demand curve before the sell-side catches up. And the sell-side isn’t even close.

Part 3 goes into the geopolitical risk: the supply chain chokepoints that could accelerate or constrain this market. If you think the sensor blind spot is interesting, wait until you see who controls the materials inside the sensors.

Anyways. That’s what I’m watching. Hope everyone has a great day, and I’ll talk to you soon.

The Sensor Nobody Covers: A 4-Part Series
1. The $168M Nobody Sees
2. 2,500 vs 12 - Why Robots Still Can’t Feel
3. Double Dependency - China’s Grip on Robots
4. The Real Alpha is One Layer Deeper

Frequently Asked Questions

What happens to the sensor market if humanoid robots don’t scale?

If production stays niche at 10,000–20,000 units per year, the sensor market probably grows to $300–400M by 2035 via CAGR alone. That’s still meaningful growth, but it doesn’t create new market leaders or reshape supply chains. The $3B number requires volume production. If the robots don’t ship, the sensors don’t sell. The bear case isn’t zero — it’s just slow.

Why should I trust a market projection from a single research report?

You shouldn’t — at least not the exact number. The $168M comes from IntelMarketResearch, and no major research firm has published a comparable forecast. The definitional boundaries between tactile sensors, electronic skin, and haptic feedback are fuzzy. Hold the number loosely. But hold the direction tightly. Multiple independent data points (Goldman’s robot projections, company filings, production data) all point the same way.

What about sensors for industrial robots and cobots, not just humanoids?

The broader force/torque sensor market is about $2.3 billion globally, growing at 6% annually. That market is real, established, and covered by analysts. The humanoid-specific opportunity is what’s invisible. If you’re interested in the broader market, ATI/Novanta and Kistler are the established players. If you’re interested in the alpha, you’re looking at the humanoid-specific demand curve that nobody’s modeling.

How do I actually track this market as it develops?

Watch three things. First, humanoid robot shipment numbers — Unitree, Tesla, Figure AI quarterly updates. Second, GGII and IntelMarketResearch reports on sensor market sizing. Third, earnings calls — when an analyst asks Tesla’s CFO about force/torque sensor procurement, that’s the signal that institutional coverage has arrived. We’re not there yet.

New here? Start with The Actuator Cartel

It’s the most-read piece on the site because it explains the single biggest hardware constraint in humanoid robotics and who actually controls it. Once you understand this, everything else (motors, magnets, costs, and timelines) makes much more sense.

Machine Narratives maps the supply chain that makes humanoid robots possible. Not the software, not the demos, not the pitch decks. The physical stack: the motors, the batteries, the magnets, the sensors, the materials. The parts that actually constrain when these machines walk into factories, warehouses, and homes.

Every article follows a specific chokepoint, traces who controls it, and works out what it means for the companies trying to build at scale.

New research is published weekly. Subscribe to stay ahead of the supply chain shifts that will determine which humanoid companies succeed and which ones struggle.

Note: All Part 1 articles and standalone analysis pieces are available in full for free. Paid subscribers get access to deeper parts of ongoing series, investment implications, and the complete supply chain map.

ACTUATION

The motors and gearboxes that move robots. Who controls them, what they cost, and why the supply chain is consolidating.

• Part 1: The Actuator Cartel

• Part 2: From Harmonic Drives to Torque Motors

• Part 3: The Chinese Are Coming

• Part 4: Tesla, Figure, and Unitree’s Supplier Wars

• Part 5: The Actuator Oligopoly’s Endgame

POWER

The batteries and thermal constraints that keep robots running. Why power is the silent bottleneck.

• Part 1: The 4-Hour Robot

• Part 2: The Leg Battery Problem

• Part 3: The Watt Tax

• Part 4: The Compute Mortgage

• Part 5: The $3 Worker

MATERIALS

The rare earths and specialty alloys underneath everything. China’s monopoly and what it means.

• Part 1: The Magnet Nobody Talks About

• Part 2: The Engineering Case for NdFeB

• Part 3: The China Chokepoint

• Part 4: Who Profits

SENSORS

The tactile and force sensors that give robots a sense of touch. The $168M market nobody’s covering.

• Part 1: The $168M Nobody Sees

• Part 2: The Touch Gap

• Part 3: The Second Chokepoint

• Part 4: The Real Alpha Layer

VISION

The cameras, LiDAR, and depth sensors that let robots see. From hardware to software to investment thesis.

• Part 1: Cameras, LiDAR, Depth Sensors: The $38B Format War

• Part 2: The Software Eats the Sensors

• Part 3: Betting on Robot Eyes (coming soon)

COMPUTE

Coming soon

STRUCTURE

Coming soon

END EFFECTORS

Coming soon

STANDALONE ANALYSIS

• SemiAnalysis Got Unitree’s Cost Advantage Right

• The Agility IPO: $2.5 Billion for 100 Robots

• 1X’s Hands are Real. 10,000 of Them Might Not Be

What Is Machine Narratives?

Machine Narratives is supply chain research for the humanoid robotics buildout. We trace the physical stack: actuators, power systems, materials, sensors, compute, and structure. We follow the chokepoints, map who controls them, and work out what it means for the companies trying to build at scale.

No hype. No pitch decks. Just the supply chain that decides whether these machines actually make it to production.

In 1780, a handloom weaver in Lancashire could earn a decent living. The work was skilled, the pay was adequate, and the demand for cotton cloth seemed insatiable. By 1820, that same weaver was competing against a mechanical loom that could produce cloth at a fraction of the cost, and required no skill at all. The handloom weaver didn’t lose to a better weaver. They lost to a machine that cost less per yard of cloth than a weaver cost in lunch.

The same thing is about to happen to the factory floor. Not in a decade. Not in some abstract “future of work” thinkpiece. In the next three to five years.

This is the final article in our series on the energy economics of humanoid robots. Over the last four pieces, we mapped the power budget of a walking machine: the legs that burn 650 watts, the gearboxes that bleed energy as heat, the AI brain that draws as much power as the walking itself, and the battery chemistry that cannot yet deliver an eight-hour shift. All of that was prologue. This is the payoff.

Every factory owner, every logistics executive, and every honest investor in humanoid robotics is asking the same question: when does a robot cost less per hour than a human, and what happens to the world on that day?

The math that keeps factory owners awake

Let’s start with the numbers, because the numbers are what matter here. In April 2026, the average production and nonsupervisory manufacturing worker in the United States earned $30.10 per hour. Total compensation (wages, benefits, payroll taxes, insurance) ran about $46 per hour, according to the Bureau of Labor Statistics’ Employer Costs for Employee Compensation survey. In high-cost regions like Germany or Japan, the fully loaded number is higher. In lower-cost manufacturing centers like Mexico or Vietnam, it is lower, but rising.

A human worker at $46 per hour, working three shifts of eight hours, 250 working days per year, costs a factory roughly $276,000 per year per workstation. A 1,000-worker factory spends $276 million annually on labor.

Now let’s price a humanoid robot. The assumptions matter enormously.

Source: Machine Narratives Research. Purchase price assumptions based on Tesla Optimus targets ($20-30K), Figure projections, and supply chain teardown estimates. Maintenance costs from traditional industrial robot benchmarks (5-12% annual maintenance), adjusted upward for humanoid complexity. Swap cost includes estimated labor and downtime.

Run those numbers through a per-hour calculation, and here is what you get:

Effective hourly cost includes amortized purchase price, maintenance, electricity, and battery swap costs. Does not include initial integration, programming, or safety fencing, which can add $5-15K per deployment station. Source: Machine Narratives Research Robot Economics Model.

That bottom-right cell, $2.80 per hour, is the number that should make every manufacturing executive sweat. At 17 times cheaper than the fully loaded cost of a US factory worker, the economic incentive to deploy humanoid robots would be overwhelming. The question is not whether factories will adopt them. The question is how fast they can deploy, and who benefits from the deployment.

The aggressive case is aggressive. I would not put $2.80/hr as my base case for 2028. But the conservative case is also too conservative, because it assumes battery technology and actuator efficiency stay static, and that is not how exponential manufacturing curves work. The price of humanoid robots is currently dropping at roughly 40% per year in real terms, driven by three things: manufacturing scale, component standardization, and the learning curve benefits that come from deploying real robots in real factories. Anyone who has watched the iPhone supply chain or the solar panel industry knows this curve. It is relentless, and it almost always moves faster than analysts project.

The math is already compelling even with today’s battery limitations. A $25,000 robot with 6-hour runtime and 5-year useful life costs about $5.50/hour to run, roughly 8x cheaper than a $46/hour US factory worker. The 8-hour breakthrough makes it 16x. Source: Machine Narratives Research Robot Economics Model.

The cost crossover. Even the conservative 4-hour scenario undercuts the human worker by 2029. The 8-hour breakthrough, the holy grail, produces economics that no factory can ignore. Source: Machine Narratives Research.

The Engels’ Pause, on fast-forward

A concept from an earlier piece is relevant here: Engels’ Pause. It is the period from roughly 1790 to 1840 when British working-class wages stagnated while per-capita GDP grew 46%. The handloom weaver, the iron forger, the cotton spinner, all of them watched machines do their work for less money than they could survive on. The gap between productivity and wages took two generations to close, and only after what economic historian Robert C. Allen called “the most consequential technology transition in human history.”

The difference this time is speed. The Industrial Revolution unfolded over decades because capital goods (steam engines, power looms, blast furnaces) were expensive to build, hard to transport, and required specialized infrastructure. A humanoid robot is software with legs. Once the hardware platform is production-ready, the deployment curve is limited by manufacturing capacity, not physics.

The Engels’ Pause for factory workers could unfold in under a decade rather than two generations. The machine that replaces a $46/hour worker today costs $25,000 to buy and $3-15 per hour to run depending on battery technology. When that machine can work a full shift without interruption, the incentive to deploy is not economic. It is fiduciary. A factory manager who doesn’t deploy robots at 10% of the labor cost will be replaced by one who does. This is not a moral argument. It is a competitive one.

The $3 robot crosses below minimum manufacturing wages everywhere except Vietnam and India, and even those markets are within striking distance as battery costs decline. The global labor arbitrage is not hypothetical. Source: ILO, national labor statistics, Machine Narratives Research.

Who is closest to the 8-hour shift?

The humanoid robot market is not a horizontal commodity market. It is a vertical integration market. The winners will not be the companies that assemble the best off-the-shelf parts. They will be the ones that design their own actuators, optimize their own compute, build their own battery management systems, and control the feedback loop between field data and hardware iteration.

Consider subscribing to see the conclusion of the battery series.

Read more

China produces 94% of the world’s sintered NdFeB magnets. Not 60%. Not 75%. Ninety-four percent. These are the magnets that make every servo motor in every humanoid robot on the planet spin. And on April 4, 2025, Beijing proved it was willing to weaponize that dominance, overnight, with zero warning, in retaliation for tariffs.

Export volumes plunged 74.3% year-over-year. European rare earth prices spiked to 6x Chinese domestic levels. Ford temporarily idled EV production lines. Elon Musk went on an earnings call and told the world that Tesla’s Optimus robot would need a Chinese export license to ship anywhere with magnets inside.

This is not a hypothetical supply chain risk. It already happened. The only question is whether it happens again, and what it means for the dozens of humanoid robotics companies that need these magnets to exist.

The Export Control Escalation — A Pressure Cooker Timeline

China didn’t flip the switch on April 4. They spent 18 months building the mechanism first. We need to understand this timeline because it reveals intent, and because every escalation was bigger than the last.

July 2023: gallium and germanium export controls. A test balloon. October 2023: graphite. August 2024: antimony. February 2025: tungsten, tellurium, bismuth, molybdenum, and indium. Tungsten prices tripled. Each wave was a message, we control the inputs your industry depends on, and we can cut them off whenever we want.

Then came April 4. MOFCOM Announcement No. 18. Seven heavy rare earths: samarium, gadolinium, terbium, dysprosium, lutetium, yttrium, scandium, restricted under non-automatic export licenses. Every compound, every metal, every magnet containing these elements. Case-by-case review. 60 to 120 day processing times. Military-adjacent applications? Almost never approved.

Effective immediately. No grace period. No stockpiling window. Companies with magnets already on order found their shipments frozen at customs.

But the real escalation came six months later. October 9, 2025: China introduced a Foreign Direct Product Rule for rare earths. Any magnet manufactured anywhere in the world containing even 0.1% Chinese-origin rare earth materials required Chinese government approval to export. This mirrors the US FDPR rules applied to Huawei, but applied to minerals. The nuclear option.

To put the FDP rule in perspective: there is essentially no NdFeB magnet on Earth that doesn’t contain at least trace amounts of Chinese-origin rare earths. China controls 91% of global rare earth refining and separation capacity. Even magnets manufactured in Japan, Germany, or the United States use feedstock that at some point passed through Chinese processing infrastructure. The 0.1% threshold was designed to capture everything. It mirrors the US FDPR on Huawei with surgical precision. Beijing studied Washington’s playbook and applied it to minerals.

Then, a reprieve, sort of. November 2025, Trump-Xi meeting in South Korea. China agreed to suspend export controls for one year, roughly through October 2026. Trump declared all rare earth issues settled. Markets exhaled.

But the suspension is fragile. China tightened controls on dual-use goods to Japan in January 2026. US imports of Chinese rare earth magnets fell 11% in November 2025 versus the prior year. EU imports surged 60% in the same period. Beijing using supply allocation as a geopolitical lever, rewarding allies and punishing adversaries. Yttrium exports to the United States collapsed to just 17 tons over eight months (April–December 2025), down from 333 tons in the same period a year earlier.

The suspension is fragile. The clock is ticking.

To put these numbers in perspective: the United States imported thousands of metric tons of rare earth magnets from China in 2024. Even with the suspension in effect, 2025 imports came in meaningfully below that baseline. The supply allocation game is already in motion. Beijing is choosing who gets magnets and who doesn’t, and Washington is on the losing end of that equation.

The 250% Price Premium — A Bifurcated Market

Here is the number that should keep every robotics CFO up at night: 250%. That’s the premium non-Chinese buyers paid for NdPr, the key rare earth input for NdFeB magnets, during peak restrictions in the second half of 2025.

Chinese domestic NdPr price: roughly $60 per kilogram. Non-Chinese NdPr price during the export control squeeze: $275 to $400 per kilogram. European rare earth prices hit 6x Chinese domestic levels at the peak.

This is not a tariff. Tariffs are predictable, measurable, and can be planned around. This is an asymmetric pricing structure created by export controls. Chinese manufacturers access materials at domestic prices, then export at whatever premium the licensing regime and arbitrage dictate. Chinese magnet producers operate at 20–30% gross margins. Western competitors? 10–15%, constrained by input costs that are structurally 2–4x higher.

The market is bifurcated by design. And it’s working exactly as intended.

Non-China Capacity — The Scale Problem Nobody Wants to Talk About

Every Western government has announced a rare earth independence strategy. Most of them are theater.

Total non-China NdFeB production capacity in 2025: roughly 8,000 to 12,000 metric tons per year. That’s 3 to 6% of global output. China’s current capacity: approximately 260,000 metric tons per year, expanding toward 500,000.

The companies actually building capacity:

Read more

So today I want to talk about whether the first humanoid robot IPO makes any sense. Not the hype. Not the TAM projections Goldman is putting out. The actual arithmetic of building a walking robot at scale, and what it means for a company trying to go from artisanal assembly to mass production in a handful of years.

What Agility Is Selling

Digit is a bipedal warehouse robot designed to move totes in distribution centers. Five feet tall, roughly 100 pounds, 35 pounds of cargo capacity. Agility doesn’t sell Digit outright. The go-to-market is RaaS, robotics as a service, at $8,000 to $10,000 per month. That’s $96,000 to $120,000 a year in operating expense for the customer, no upfront capital outlay.

The RaaS model is clever. It lowers the adoption barrier for mid-size warehouse operators who can’t write a $250,000 check for an unproven robot but can absorb $8,000 a month in opex during peak season. Digit has moved over 100,000 totes in live commercial deployment with GXO Logistics and Amazon, which is a meaningful proof point for operational maturity.

But 100 robots in nine years is not a production ramp. It’s a pilot program. The market is pricing this company as if 100 will become 50,000. Time to check the arithmetic.

The Three Things You Have to Believe

To justify $2.5 billion, you need to believe three things: Agility can produce 50,000 cumulative units by 2030, sell them at roughly $120,000 per unit in annual RaaS revenue, and do it at a BOM cost of $25,000 to $35,000 per robot.

Those are the three inputs. Quantity, price, cost. Everything else is derived.

Let me steelman each one before we stress-test them.

Quantity: 50,000 cumulative units by 2030. Not 50,000 in a single year. That’s 50,000 total, running at maximum factory output for four to five consecutive years. RoboFab, Agility’s purpose-built facility in Salem, Oregon, is designed to scale to 10,000 units per year. There are no public announcements of additional manufacturing sites or contract manufacturing partnerships. RoboFab is it.

Price: $120,000 per unit per year in RaaS revenue. That’s the midpoint of the $96,000 to $120,000 annual range at the $8,000 to $10,000 monthly rate. This assumes customers pay consistently and the robots stay operational enough to justify the fee. At scale, list prices for the hardware itself would need to hold above $100,000 to $150,000 if Agility ever shifts to direct sales.

BOM cost: $25,000 to $35,000 per unit at scale. Morgan Stanley pegs the current BOM at $50,000 to $60,000. Getting to $25,000 to $35,000 requires a 40 to 50 percent decline through volume manufacturing, actuator standardization, and supply chain optimization. Bank of America projected BOM could drop to $17,000 by 2030, but I think that’s optimistic. The realistic floor is $25,000 to $35,000.

Stack those three inputs and you get a picture of a company that needs to quintuple its production rate from a single factory, while simultaneously cutting the cost of building each robot in half. That’s the bet.

Now the question becomes whether each input survives contact with reality.

Squeeze 1: The Factory

RoboFab is designed to scale to a maximum of 10,000 units per year. At maximum capacity, reaching 50,000 cumulative units requires five consecutive years of flat-out production starting from essentially zero today. The facility is still being built out. Even the most optimistic internal projections don’t claim 10,000 units per year before 2029.

I’d argue the closest historical parallel is Tesla’s Model S ramp, and it was brutal. In 2012, Tesla delivered 2,650 Model S sedans. By 2015, they were over 50,000 annually. That’s roughly a 19x increase in three years. It nearly bankrupted the company. Elon Musk has said publicly that Tesla was within weeks of running out of cash during the 2013 ramp. They needed multiple emergency funding rounds.

Tesla had something Agility doesn’t: the NUMMI factory in Fremont, a facility designed for mass automotive production since 1984. It had paint shops, stamping presses, welding robots, and assembly lines that could be retooled. Tesla bought it for $42 million in 2010. RoboFab is being purpose-built for humanoid robot production, which means Agility is simultaneously designing the product, the manufacturing process, and the factory itself. Three hard problems instead of one.

Here’s the timeline arithmetic. If RoboFab starts producing in 2027, which is optimistic, and ramps to 2,000 units in year one, 5,000 in year two, and 10,000 by year three, you’re looking at roughly 17,000 cumulative units by 2030. That’s a third of the 50,000 target the valuation requires. To hit 50,000 by 2030, Agility would need to start producing in 2026, hit 10,000 annually by 2028, and maintain that rate for three straight years. The ramp would need to be faster than Tesla’s, and Tesla had a factory designed for mass production.

The arithmetic is stark. You’re betting that a company that has shipped 100 robots can go from artisanal assembly to maximum factory output in a handful of years. Tesla did something similar. Tesla also almost died doing it, and they had infrastructure Agility can only dream of.

Squeeze 2: The BOM Floor

The BOM trajectory is the most important number in this entire thesis, and the part that gets the least attention.

At the $25,000 to $35,000 BOM range I laid out in the valuation section, the unit economics work. At $120,000 in annual RaaS revenue per robot, you’re generating enough gross profit to cover deployment, maintenance, and overhead while still contributing to growth. But that’s at scale. At the current $50,000 to $60,000 BOM, the math is much tighter.

The reason BOM has a floor comes down to physics and geopolitics, not engineering cleverness. I want to spend some time here because this is where most bull cases fall apart. People wave their hands about learning curves and scale economies without looking at what’s actually inside the robot.

Actuators alone represent 30 to 40 percent of the bill of materials. Each robot needs 12 joints, each with a harmonic reducer and servo motor, running $6,000 to $12,000 at current volumes. Four companies control the critical actuator components. That oligopoly doesn’t dissolve because you order more units. Volume manufacturing helps, but the supply chain concentration means pricing power stays with the suppliers, not the robot maker.

Structural materials, carbon fiber composites accounting for 10 to 20 percent of BOM, have their own concentrated supply chains. The compute stack, Nvidia inference hardware because there’s no alternative at this performance tier, adds another layer of cost that Nvidia has no incentive to compress. Every motor needs neodymium magnets from a supply chain where China controls 94 percent of NdFeB production.

None of this follows Moore’s Law. None of it compresses 10x every decade. The components inside a humanoid robot are governed by the same physics as the components inside a car, and car BOMs haven’t dropped 50 percent in five years either.

The hand alone is a microcosm of the problem. Forty-plus force and tactile sensors at $800 to $2,000 per robot, with no volume manufacturing path for most of them. The sensors are custom. The suppliers are small. The demand signal from 100 robots doesn’t move the needle on their production economics.

There’s another dimension that rarely gets discussed: the Chinese vs. non-Chinese BOM differential. Chinese competitors like Unitree are building humanoid robots at BOM costs of roughly $46,000. Agility’s estimated BOM sits at $50,000 to $60,000. That differential isn’t dramatic at first glance, but it compounds at scale. Agility’s 80 percent US-sourced supply chain reduces geopolitical risk but locks in a higher cost floor than Chinese competitors face.

As volume manufacturing kicks in and actuator designs standardize, those costs should decline meaningfully, maybe 30 to 40 percent from current levels. But that still leaves you with $4,000 to $8,000 per robot just in actuators. Add $3,000 to $5,000 for compute, $3,000 to $6,000 for structure, $1,500 to $4,000 for the battery, and you’re at $17,000 to $25,000 before you’ve wired a single connector.

The BOM has a floor. That floor sits somewhere between $25,000 and $35,000. Below that floor, you can’t build the robot. Above it, the unit economics start to work at the $120,000 per unit per year RaaS rate. But the floor exists, and it’s dictated by physics and geopolitics, not by software-style scaling laws.

To be clear: Bank of America projects $17,000 by 2030. I think that’s a fantasy. It assumes actuator costs compress by 60 percent, compute costs follow a Moore’s Law curve that Nvidia has no incentive to enable, and structural materials somehow get cheaper despite concentrated supply chains. The realistic range is $25,000 to $35,000, and even getting to $25,000 requires everything to go right.

Squeeze 3: The Competition

The bear case here isn’t about demand. It’s about whether Agility can win the race it’s in.

Let’s look at who else is building.

Tesla has announced ambitious Optimus production targets, 5,000 units in 2025, scaling to 50,000 to 100,000 in 2026, but has consistently missed them by enormous margins. As of mid-2026, Tesla has deployed only 300 to 500 Optimus units, almost entirely within its own factories. Musk declined to reaffirm the 2026 target on Tesla’s Q1 2026 earnings call. They’re still leveraging their existing automotive supply chain, established manufacturing infrastructure, and over $25 billion in cumulative AI and robotics investment. Tesla doesn’t need to build a factory from scratch. They have factories. But having factories and actually producing robots at scale are very different things, as the gap between target and reality demonstrates.

Figure AI has raised $745 million with backing from Microsoft, OpenAI, and NVIDIA. They haven’t shipped production units yet, but the capital stack and strategic partnerships give them a runway that most startups can’t match. Boston Dynamics has its electric Atlas program and a 1,000-unit agreement with DHL for its Stretch warehouse robot. Stretch isn’t a humanoid, but it’s competing for the same warehouse budget dollars. And then there’s Unitree, which shipped over 5,500 non-full-size humanoid units in 2025 (mostly to research institutions, with roughly 70 percent exported) and is targeting 20,000 units in 2026.

The competitive dynamics matter because they affect all three inputs in the valuation model. More competitors mean more pressure on the $120,000 per unit per year RaaS rate. More production volume across the industry means actuator suppliers have less incentive to cut prices for any single customer. And more capital flowing into the space means Agility’s first-mover advantage erodes faster.

At the $25,000 to $35,000 BOM floor I described earlier, the unit economics work at $120,000 per year in RaaS revenue. But Unitree is building at $46,000 BOM today, and they’re not standing still. If Chinese competitors get to $30,000 BOM while Agility is at $35,000, the pricing pressure on Agility’s RaaS rate becomes intense. You can’t charge $10,000 a month for a robot that costs the same to build as a competitor’s $6,000-a-month unit.

Agility’s advantages are real but narrow. They have first-mover experience in purpose-built humanoid manufacturing. They have operational deployment data that no one else has at this scale. And they have Amazon and GXO as validation partners. The question is whether that head start matters when Tesla decides to pour automotive-scale resources into the same market.

The window Agility had as the only serious player is closing fast. Two years ago, the humanoid robot market was essentially Boston Dynamics doing backflips on YouTube and Agility quietly building warehouse bots. Now you have Tesla, Figure AI backed by the biggest names in tech, Unitree shipping thousands of units from China, and a dozen other startups burning through venture capital. Agility’s moat is deployment experience, data on how robots fail in real warehouses, what breaks first, how operators actually use the machines. That data is genuinely valuable. But it’s a moat with a timer on it. Every month that Tesla, Figure, and Unitree deploy their own robots, that experiential advantage erodes.

Here’s the uncomfortable question: what happens to the $120,000 per unit per year RaaS rate when Unitree offers a comparable warehouse robot at $6,000 a month? Or when Tesla bundles Optimus with its existing warehouse automation stack? The pricing power assumption in the valuation model requires Agility to maintain premium pricing in an increasingly commoditized market. That’s possible if Digit is genuinely better than the alternatives. It’s not possible if the alternatives are 80 percent as good at 50 percent of the price.

The SPAC Structure

One more thing about the structure of this deal. The SPAC IPO was announced June 24, 2026, at a $2.5 billion valuation. Agility had previously raised a $400 million Series C in March 2025 at a $2.12 billion valuation, so the SPAC represents a modest markup from the last private round.

SPACs provide public market currency for future acquisitions and talent retention. Stock options in a publicly traded company are more liquid than private equity, which matters when you’re trying to recruit engineers away from Tesla and Google.

But the track record for SPAC’d hardware companies is, to put it politely, mixed. Nikola and Lordstown Motors are the cautionary tales everyone invokes, and for good reason. Both promised revolutionary hardware, both went public via SPAC, and both collapsed when the production reality failed to match the investor presentation. Digit actually exists and works, which is already more than Nikola ever had. But the structure carries baggage that investors should be aware of.

There’s also the timing question. Agility is going public at a moment when humanoid robotics hype is near peak levels. Goldman Sachs 6xing their TAM estimate will do that. The SPAC structure means Agility gets public market currency without the scrutiny of a traditional IPO roadshow. That’s good for Agility. It’s less good for investors who want to stress-test the production assumptions before buying in.

Where I Come Out

The bull case is real. The warehouse labor shortage is genuine. Goldman Sachs revised their humanoid TAM estimate upward by 6x, from $6 billion to $38 billion by 2035, and now projects 1.4 million humanoid robot shipments over that period. Digit works. The 100,000 totes moved in live deployment prove operational maturity. GXO and Amazon are validation partners, not science experiments. RaaS lowers adoption barriers. The demand is there.

But the bear case is arithmetic. Three squeezes, each independently problematic, each compounding the others.

Squeeze 1: Volume. The market prices 50,000 cumulative units by 2030. The single-factory constraint yields roughly 17,000. That’s a third of what the valuation requires.

Squeeze 2: BOM cost. The floor sits at $25,000 to $35,000 per robot. At $50,000 to $60,000 today, you need a 40 to 50 percent decline that physics and supply chain concentration make difficult. The BOM doesn’t compress like software.

Squeeze 3: Pricing power. Tesla has factories and $25 billion in investment. Unitree builds at $46,000 BOM. Figure AI has $745 million and the backing of Microsoft, OpenAI, and NVIDIA. The window is closing.

Stack the three squeezes together: a single factory that can produce a third of the required volume, a BOM floor that sits at $25,000 to $35,000 with no Moore’s Law to compress it, and competitors with deeper pockets and lower costs. The market is pricing Agility at roughly 2 to 3x what the supply chain can deliver in the next three years.

My number: if you give Agility every benefit of the doubt — RoboFab starts producing in 2026, hits maximum output by 2029, BOM drops to $25,000, RaaS rate holds at $10,000 a month, and no competitor undercuts them — you get to maybe $1.5 to $2 billion in enterprise value by 2030. That’s the bull case. And it requires everything to go right.

The more realistic scenario: RoboFab starts in 2027, hits 8,000 units per year by 2030, BOM lands at $30,000 to $35,000, and competition compresses the RaaS rate to $7,000 to $8,000 a month. That’s a $600 million to $1 billion company. Still a real business. Still worth building. Just not worth $2.5 billion today.

I think we’re watching the first chapter of the humanoid robot IPO story. Whether it’s a success story or a cautionary tale depends entirely on what happens in Salem, Oregon over the next 36 months. The demand is there. The technology works. The money is available. The only question is whether Agility can actually build the things. And if they can’t, someone else will. The market opportunity doesn’t disappear just because one company stumbles.

Anyways. The humanoid robot market is real. The supply chain constraints are real. And $2.5 billion is a lot of money for 100 robots. Hope everyone has a great day, and I’ll talk to you soon.

Related Reading from Machine Narratives Research:

• The Actuator Cartel: Who Controls the Robot Muscle Supply Chain — The hidden chokepoint behind every humanoid robot

• The Magnet Nobody Talks About: Why Humanoid Robots Can’t Scale Without NdFeB — Each robot needs 40 motors. Every motor needs rare earth magnets.

• The Leg Battery Problem — Why EV batteries fail inside walking machines

• The Compute Tax — Nvidia’s toll booth on the robotics revolution

• Tesla, Figure, and Unitree’s Supplier Wars — Three companies, three procurement strategies

• The Actuator Oligopoly’s Endgame — When Tesla builds its own motors and China prices at cost

There is a particular kind of frustration that engineers know well. It is the frustration of being told by the laws of physics that you have no choice.

I think about this a lot when I look at the current state of humanoid robotics. The companies building these machines (Tesla, Figure, Agility, Unitree, all of them) are running into the same wall. Not a software wall. Not a capital wall. A materials wall. Specifically, a magnet wall.

The problem is simple to state and maddening to solve. If you want to put 40 electric motors inside a machine shaped like a human body, in the ankles, the knees, the hips, the wrists, the fingers, and you want that machine to move smoothly and hold objects without crushing them, all without weighing 200 kilograms, then you need a very specific kind of permanent magnet. You need neodymium-iron-boron. NdFeB. There is no second option.

Not because nobody has tried to find one. People have been trying for decades. The physics simply will not cooperate.

In this piece, I want to walk you through why. Not as a lecture, but as a kind of detective story. We are going to follow the clues, one number at a time, until we arrive at the conclusion that the most ambitious robotics project in human history runs through a handful of mines and refineries in southern China.

The One Number That Rules Them All

Let us start with a concept that sounds more intimidating than it is: BHmax.

BHmax stands for maximum energy product. It measures how much magnetic energy a permanent magnet can store in a given volume. The units are MGOe (mega gauss-oersteds), or, if you prefer SI, kilojoules per cubic meter. The important thing is not the unit. The important thing is what the number means.

Think of BHmax as the magnet’s “energy density.” Just as a lithium-ion battery with higher energy density stores more power in the same package, a magnet with higher BHmax produces more torque in the same volume. If you are building a motor that needs to fit inside a human-shaped forearm, and that forearm needs to look like a forearm, not a fire hydrant, then BHmax is the number that determines whether your design works or does not work.

Here is where things get interesting. Look at the numbers for the permanent magnets currently available to engineers:

I want you to notice the gap. Not a small gap. A chasm.

NdFeB N52 stores 11.5 times more magnetic energy per unit volume than the best ferrite ceramic. It stores roughly twice as much as samarium-cobalt, the next-best permanent magnet. This is not a marginal difference. This is the kind of difference that determines whether a hand has five fingers or three. Whether a robot weighs 70 kilograms or 180. Whether it can walk across a room or has to be bolted to the floor.

Engineers did not choose NdFeB because they love rare earths. They chose it because the energy product left them no alternative.

Why Humanoids Are the Worst Possible Application for Substitution

When people talk about “reducing rare earth dependence,” they usually point to electric vehicles or wind turbines. Those are big, visible applications. But the truth is, those are the easy cases for substitution.

An EV motor sits in a well-defined space under the hood or between the axles. If you make that motor 30% larger to accommodate a weaker magnet, the car still drives. It might weigh a bit more. The packaging gets tighter. But it works. Engineers have room to negotiate with the physics.

A wind turbine nacelle is enormous. You have space. You can tolerate a heavier generator because the thing is already the size of a bus. Ferrite magnets have actually been making inroads in some wind turbine designs for exactly this reason.

Now consider a humanoid robot hand.

Each hand has somewhere between 12 and 22 degrees of freedom, meaning individual joints that need to move independently. Each joint needs its own actuator, or motor. These motors sit inside the fingers, the palm, the wrist. The available volume for each motor is measured in cubic centimeters. Sometimes single-digit cubic centimeters.

If you swap NdFeB for ferrite in a hand actuator, you need a motor roughly three to five times larger to produce the same torque. A finger that was 2 centimeters in diameter becomes 6 centimeters. That is not a finger anymore. That is a sausage.

I think this is the point that gets lost in the policy debates about rare earths. Yes, we should diversify supply. Yes, we should invest in alternatives. But the engineering constraints for humanoids are fundamentally different from EVs or wind. The robot’s body has to approximate a human form factor. The hands have to be able to pick up a coffee cup without crushing it and also open a jar lid. The legs have to support the robot’s weight and absorb the shock of each step. There is simply no room, zero room, for a motor that is three times too large.

A single humanoid robot in the Tesla Optimus class contains between 500 and 1,100 grams of NdFeB. That is about half a kilogram to a full kilogram of permanent magnets, spread across 40 or more motors. At current prices, roughly $65 to $80 per kilogram for finished magnets, that works out to $35 to $85 worth of magnet material per robot. A trivial cost, really. Two percent or so of the bill of materials.

The cost is not the problem. The availability is the problem. But we will get to that.

The Temperature Trap

Now here is where the story takes a turn. Because if BHmax were the only constraint, we would still have options. Limited options, but options. The real killer, the thing that makes the dependency even more acute, is temperature.

Read more

In 1982, a 38-year-old researcher named Masato Sagawa was told by his employer, Sumitomo Special Metals in Japan, to develop “a magnet that doesn’t break.”

Sagawa was, by most accounts, a failed academic. He had washed out of university research. He had lost confidence. He joined a corporate lab, the kind of place where ambition goes to file patents and retire. Everyone in the field knew that iron-based permanent magnets were impossible. The physics didn’t work. The crystal structures were wrong. The smart people had moved on.

Sagawa tried anyway. He spent years tinkering with different compositions, different ratios of rare earth elements and iron, different sintering temperatures. Most of his experiments failed. The magnets were brittle. The crystal structures wouldn’t hold. His colleagues at Sumitomo must have wondered what this quiet, stubborn man was doing with all that furnace time.

What followed was one of those quiet moments in history that nobody notices at the time but that reshapes everything downstream. In 1984, two teams on opposite sides of the world, Sagawa’s group at Sumitomo in Japan and a team at General Motors in the United States, independently discovered the same alloy: Nd₂Fe₁₄B. Neodymium, iron, and boron. A crystalline structure that, against all expectations, produced the strongest permanent magnet ever made.

Sagawa was awarded the Japan Prize in 2012. He has been nominated for the Nobel Prize repeatedly. He has never won it.

The material he discovered, neodymium-iron-boron, or NdFeB in the shorthand that engineers use, is the strongest commercially available permanent magnet in the world. Its energy product, the measure of how much magnetic energy a material can store, is over ten times that of the humble ferrite magnets you might find stuck to a refrigerator. In its highest grade, N52, NdFeB can produce a remanence of 1.44 Tesla, a magnetic field strong enough to snap your fingers together if you got them between two blocks of the stuff.

A material that didn’t exist 42 years ago is now the invisible foundation of every humanoid robot you’ve seen walk across a stage.

And this is not some niche curiosity. This is a material at the center of what might be the most consequential technology race of the next two decades. The humanoid robot industry, led by Tesla, Figure, Boston Dynamics, and a growing list of Chinese competitors, is projected to be worth $38 billion by 2035, and possibly ten times that by 2040. Every single one of those robots will need NdFeB. Every joint. Every finger. Every actuator.

What makes this story worth telling is not the technology itself. It is the supply chain behind it. Because the supply chain for NdFeB is one of the most geopolitically concentrated in the world, more concentrated than semiconductors, more concentrated than lithium, more concentrated than almost any critical material you can name.

The Secret in the Machine

Let me show you something that nobody talks about at robot launch events.

Every humanoid robot, every Atlas, every Optimus, every Figure, every Unitree, needs between 28 and 76 electric motors to move. Motors in the legs to walk. Motors in the arms to lift. Motors in the hands to grip. Motors in the neck to turn the head. Motors in the torso to balance. Tesla’s Optimus, the robot you’ve probably seen the most of, needs approximately 40 motors and roughly 4 kilograms of NdFeB magnets per unit. Over 95 percent of the motors used in humanoid robots contain rare earth permanent magnets.

picture that for a moment. About four kilograms. That is roughly the weight of a bag of sugar. Except this bag of sugar is distributed across 40 tiny, precision-engineered cylinders, each one containing a chunk of NdFeB magnet that makes the difference between a motor that works and a motor that doesn’t.

And here is the part that should make you sit up. You’ve watched these robots walk across stages. You’ve seen them wave, pick up objects, navigate rooms, fold laundry. Impressive, right? But nobody at those launch events told you that each robot is wearing several kilograms of a material that 94 percent of the world’s supply comes from one country.

Ninety-four percent.

China controls 94 percent of global sintered NdFeB magnet manufacturing. It controls 91 percent of rare earth processing and refining. And when it comes to the heavy rare earths, dysprosium and terbium, the elements that keep these magnets working at high temperatures, China controls roughly 99 percent of refining capacity.

One country. One supply chain. And every humanoid robot on Earth depends on it.

This is the story of the magnet nobody talks about.

The Material That Changed Everything

We should begin with what NdFeB actually does, because the physical intuition matters more than the chemistry.

Imagine you want to build a motor the size of a sugar cube, a motor powerful enough to curl a robot’s finger around an eggshell without cracking it. That motor needs a permanent magnet inside it. The magnet provides the static magnetic field that interacts with the electric current in the motor’s windings, creating the force that spins the rotor.

The stronger the magnet, the more force you get from a smaller motor. And NdFeB is, by a wide margin, the strongest permanent magnet material available. Its energy product, the number that engineers use to compare magnets, measured in mega-gauss-oersteds, or MGOe, reaches 52 in its best grades. Ferrite magnets, the cheap ceramic kind used in everything from speakers to refrigerator decorations, top out around 4.5. That is not a small difference. That is a tenfold advantage.

This matters enormously for humanoid robots, and here is why.

The motors in a robot’s hip and knee joints need to produce 50 to 100 Newton-meters of continuous torque. That is the force required to keep a 60-kilogram robot standing, walking, climbing stairs. Each of those motors requires 30 to 60 grams of sintered NdFeB. There is no way around it, not if you want the motor to fit inside a human-sized leg.

But the real problem is the hands.

Hand motors in humanoid robots are 8 to 15 millimeters in diameter. That is roughly the width of a pencil. Inside that tiny cylinder, you need enough magnetic force to give a robot finger the dexterity to pick up an egg, thread a needle, or turn a doorknob. NdFeB’s energy density advantage is magnified at small scales, the smaller the motor, the more the magnet matters. If you tried to use ferrite magnets instead, the motors would need to be three to five times larger. Try building a robot hand with fingers the size of bratwursts. The fingers won’t close. The hand won’t grip. The robot won’t work.

This is the detail that kills the casual objection of “just use a different magnet.” You can, in theory, substitute ferrite or samarium-cobalt into a large industrial motor where you have room to spare. You cannot substitute it into a robot hand. The physics simply will not allow it.

And samarium-cobalt, the other serious rare earth magnet? It is more expensive than NdFeB, less powerful, and its supply chain is, you guessed it, also dominated by China. Iron nitride magnets, which some researchers believe could eventually approach NdFeB performance, are at technology readiness level 2 or 3. That means laboratory demonstrations. Not production. Not for at least ten years, and probably longer.

So here we are. A material invented 42 years ago by a man who thought he was a failed scientist is now the only viable option for the motors that make humanoid robots possible. And there is no substitute on the horizon.

The Demand Wave

To put this in perspective, let us talk about numbers.

The global NdFeB magnet market was worth approximately $30.5 to $31.7 billion in 2025. Industry projections put it at $59.7 to $65.3 billion by 2034, a compound annual growth rate of roughly 7 percent. That sounds manageable. A market doubling over a decade. Supply chains have handled worse.

But those projections were made before the robots.

Goldman Sachs projects the humanoid robot market at $38 billion by 2035. Morgan Stanley is more aggressive, estimating $357 billion by 2040. And then there is Elon Musk, who has projected that Tesla could eventually produce 10 billion Optimus units, a number so large it almost doesn’t mean anything, except as a statement of intent.

Here is what that intent looks like in magnet terms. If the world were to produce 10 billion humanoid robots, Musk’s projection, not mine, it would require 186 times the current annual global NdFeB production. Not 186 percent. One hundred and eighty-six times.

Now, 10 billion robots is a fantasy number. Even 1 percent of it, 100 million robots, would require nearly double the world’s current NdFeB output. And that is just the robots.

Robots are not the only hungry mouth at the table. Every electric vehicle traction motor requires one to three kilograms of NdFeB. Global EV sales exceeded 17 million units in 2024, and that number is climbing fast. Direct-drive offshore wind turbines, the big ones, the ones that matter for the energy transition, use 500 to 700 kilograms of NdFeB per megawatt of capacity. A single large offshore installation can consume tonnes of the stuff. And defense applications, the F-35 fighter jet, submarine motors, missile guidance systems, have inelastic demand with absolutely no substitution path. A military planner does not have the option of saying “let’s use a weaker magnet.”

this like a dinner table. There are already three very large, very hungry guests sitting down, the EV industry, the wind industry, and the defense establishment, and now a fourth guest is walking through the door. A guest who, if the projections are even half right, will eventually eat more than the other three combined.

The humanoid robot industry wants to enter a market that is already stretched thin. The supply chain was not built for this. It was barely keeping up as it was.

The Chokepoint

I want to tell this part of the story carefully, because I think it reads better as a geopolitical thriller than as a table of statistics.

China's dominance in rare earth magnets did not happen by accident. It happened over three decades of deliberate industrial policy, subsidized mining, integrated supply chains, and a willingness to accept environmental costs that Western nations would not. The Bayan Obo mine in Inner Mongolia, the world's largest rare earth deposit, accounts for the majority of China's rare earth reserves. One mine. One province. One country.

The result is a supply chain so concentrated that it makes the semiconductor industry’s dependence on TSMC look diversified by comparison.

Between 2014 and 2024, China’s NdFeB production capacity grew from 30,000 tonnes per year to approximately 260,000 tonnes. That is an almost ninefold increase in a decade. During the same period, non-China capacity, everything built by the United States, Europe, Japan, and Australia combined, grew to roughly 8,000 to 15,000 tonnes. That is 3 to 6 percent of the global total.

The economics make this gap nearly impossible to close quickly. Chinese producers access neodymium-praseodymium, the key raw material, at around $110 per kilogram domestically. Non-Chinese producers pay $275 to $400 per kilogram. That is a 250 percent premium before you have even turned on the furnace. Before you have paid a single worker. Before you have shipped a single gram. Chinese magnets benefit from 30 to 50 percent lower production costs due to scale, integrated supply chains, lower labor costs, and government subsidies.

If you are a Western company trying to build a robot magnet, you are paying two and a half times what your Chinese competitor pays for the same raw material. That is not a competitive disadvantage. That is a wall.

And then, in April 2025, the wall got higher.

On April 4, 2025, the same week as President Trump’s “Liberation Day” tariffs, China added seven medium and heavy rare earths, along with dysprosium- and terbium-containing NdFeB magnets, to its export restricted list. The announcement, officially known as Announcement No. 18, did not ban exports outright. Beijing did something more elegant: it made you apply for an export license, then took its time processing the applications. License applications take 60 to 120 days, with no guarantee of approval and no clear criteria for decision-making.

The effect was immediate and devastating. China’s rare earth magnet exports plunged 74.3 percent year-over-year at their lowest point. But the restriction was not applied evenly. European imports surged 60 percent while US imports fell 11 percent. Beijing was using supply allocation as a geopolitical tool, rewarding allies, punishing adversaries, all without technically breaking any trade agreement.

The most striking number I found in my research was this: yttrium exports to the United States fell to 17 tons in the eight months following the restriction, compared to 333 tons in the preceding eight months. That is a 95 percent decline. Not a gradual slowdown. A spigot turned off.

Then there was the moment the abstract became personal.

On Tesla’s Q1 2025 earnings call, a call watched by millions of investors, analysts, and robot enthusiasts, Elon Musk admitted that Optimus production was “impacted” by China’s rare earth magnet export restrictions. His exact words:

“China wants some assurances that these are not used for military purposes, which obviously they’re not. They’re just going into a humanoid robot.”

The richest man in the world, on an earnings call watched by millions, admitted that his robot program was waiting on an export license from Beijing. Not a cutting-edge chip. Not a proprietary software library. Not an AI model. A magnet. A chunk of metal alloy that most people have never heard of, invented by a man most people have never heard of, in a country that has decided it gets to decide who receives them.

What Could Break This

Let us steelman the counter arguments first, because intellectual honesty demands it.

The West is building. MP Materials operates the only integrated US mine-to-magnet production facility. Its 10X manufacturing plant is targeting approximately 10,000 tonnes per year of NdFeB capacity by 2028, backed by a $200 million incentive package and $400 million from the Department of Defense. Announced expansions across the US, Europe, Japan, and Australia could bring non-China NdFeB capacity to 25,000 to 35,000 tonnes per year by 2028 to 2030.

That sounds promising until you remember that China currently produces 260,000 tonnes. Even the most optimistic Western buildout would represent roughly 10 to 13 percent of global capacity. The West is building. But it is building from 3 percent.

Then there is the diplomatic track. In November 2025, Presidents Trump and Xi reached an agreement to suspend export restrictions for approximately one year, until roughly October 2026. The truce has eased some of the immediate pressure. But “eased” is doing a lot of work in that sentence. Even during the truce, US companies report greater disruption than European ones. Export licensing remains slow, opaque, and highly discretionary. Military-affiliated companies are effectively barred from receiving licenses. They shook hands. The pipes are still clogged.

On the technology front, the one that optimists always point to, researchers are working on alternatives. Iron nitride magnets are the most exciting. Theoretical performance near NdFeB levels, made from abundant materials, no rare earths required. But they remain at technology readiness level 2 or 3, which in plain English means “we can make it work in a lab, sometimes, under controlled conditions.” A commercially viable, mass-produced iron nitride magnet is at least ten years away. Probably longer. Motor design innovations, like grain boundary diffusion, can reduce heavy rare earth content by 40 to 60 percent. That helps. It does not eliminate the dependency.

And recycling? The NdFeB recycling industry is growing, using processes like hydrogen decrepitation to recover magnets from old electronics and industrial motors. But the volumes are tiny relative to demand, and the feedstock from humanoid robots will not exist in meaningful quantities until the 2030s at the earliest. You cannot recycle robots that have not been built yet.

The Question That Lingers

So here we are.

A material invented by a failed academic in a Japanese corporate lab in 1984 is now the invisible foundation of the most ambitious technology project of the 21st century. Humanoid robots, the machines that Tesla, Figure, Boston Dynamics, and a dozen other companies are racing to build, cannot move without NdFeB magnets. There is no substitute. There is no workaround. Not today, and not for at least a decade.

Ninety-four percent of those magnets come from one country. That country has already demonstrated, in April 2025, that it is willing to use that leverage. The Trump-Xi truce expires in October 2026.

I keep coming back to something Musk said on that earnings call. “They’re just going into a humanoid robot.” As if that made it simpler. As if a robot is less geopolitically sensitive than a missile. But I wonder if he understands, if anyone in Silicon Valley truly understands, that the supply chain for the material inside every actuator, every joint, every finger of every robot they are building runs through a single mine in Inner Mongolia, a handful of smelters in Baotou, and a licensing office in Beijing that answers to people who have very different priorities than shipping magnets to Fremont.

In Part 2 of this series, we will go deeper into the engineering, why motors need this specific material, how the actuator supply chain actually works, and where the technical bottlenecks are. In Part 3, we will look at the geopolitics, the export controls, the diplomatic chess, and what happens when the truce expires. And in Part 4, we will look at who stands to benefit: the Western companies trying to build alternative supply chains, and whether any of them can actually compete.

But for now, I want to leave you with one image.

The next time you see a humanoid robot walk across a stage, smooth, confident, almost human, imagine roughly four kilograms of neodymium-iron-boron hidden inside its body. A material that didn’t exist 42 years ago. Discovered by a man the world overlooked. Controlled by a country that has already shown it will squeeze the supply.

The robots are coming. The magnets are finite. And nobody at the launch event is talking about it.

Next time, we dig into the engineering, and it is fascinating. Hope everyone has a good week, and I’ll talk to you soon.

Machine Narratives Research provides institutional-grade analysis of the humanoid robotics supply chain. This is Part 1 of a four-part series on NdFeB magnets and the future of robot manufacturing.

Sources: CSIS (Jun 2026); S&P Global (Jan 2026); Global Times / Securities Times (Apr 2025); Arnold Magnetics (2025); EurekAlert / Kobe University (Jan 2025); The Wire China (Jun 2025); IEA (2025); DOE (2024); BMI (H2 2025); Spherical Insights; Precedence Research; Goldman Sachs; Morgan Stanley; Adamas Intelligence (2025); IDTechEx (2026); Machine Narrative Research.

Read more

We’re not here to debunk SemiAnalysis. They got the big picture right. We’re here to go deeper on three things they didn’t fully unpack: what the margins actually look like when you use real selling prices instead of list prices, what’s inside the actuators that make up 66% of the BOM, and why the component suppliers, not Unitree, are the real winners of this cycle.

Thesis: Unitree's cost advantage is real and structural. But "dominance" is a timeline call, and the timeline depends on reliability data that doesn't exist yet, blended margins 12–17 points below the headline number, and 200+ competitors drawing from the same supply chain.

What SemiAnalysis Gets Right

Let’s give credit where it’s due. The BYD/DJI playbook framing is the cleanest lens anyone has applied to Unitree’s strategy. Own the most expensive component (actuators), bootstrap a willing market (researchers/hobbyists), ride the ecosystem, and let each hardware generation unlock the next market. That’s exactly what’s happening. The parallel to BYD’s battery-cell-first strategy and DJI’s flight-controller-first strategy is tight and well-argued.

The BOM analysis is also correct. We’ve run the same teardown through CNP Securities’ data, and we land on the same number: approximately $5,700 total BOM, with the actuator subsystem at roughly $3,780. The QDD architecture choice, cheaper planetary gearboxes with bigger motors instead of expensive strainwave drives, is a genuine structural advantage. Unitree can iterate a new actuator design in weeks, not quarters. That’s not a marginal benefit. That’s a compounding one.

The piece nails the ecosystem point too. China’s 200+ humanoid companies are all feeding and drawing from the same component supply chain. Every province now has multiple manufacturers of gearboxes, motors, and encoders sized for humanoids. Unitree didn’t just build a robot, they helped catalyze an entire manufacturing ecosystem. That’s the BYD parallel working in real time.

The Margin Nuance: List Price vs. Realized Price

Here’s where we diverge, not on the BOM, but on what price you attach to it.

SemiAnalysis cites 67% gross margins on the G1. That number is real at their cited export list price of $27,300. That’s the ceiling, not the average. It’s like quoting Apple’s gross margin on the iPhone Pro Max and pretending every iPhone sells at that price.

Here’s the actual pricing structure:

Source: CNP Securities teardown, Unitree First Round Inquiry Response (SSE), industry channel checks.

The majority of Unitree’s volume ships domestically at ¥85,000, that’s approximately $11,700 at current exchange rates. CNP Securities’ own analysis puts margins at “exceeding 40%” at that domestic price, which our math confirms at roughly 51%. Export deals in volume, the ones actually closing, not the sticker price, land between $16,000 and $22,000. Blended across channels, we estimate realized gross margins of 50–65%.

That’s still exceptional. A 50–65% gross margin on a $5,700 BOM is the kind of unit economics most hardware companies dream of. But there’s a meaningful difference between 50% and 67% when you’re projecting out to market dominance. At 50%, you have less room to absorb warranty costs, service contracts, and the inevitable reliability fixes that come with early deployments. At 67%, you’re swimming in margin. SA used the ceiling. The average is lower.

This matters because Unitree’s entire pitch to the market, and the narrative around their IPO, hinges on margin sustainability. If the blended margin is 50–55% instead of 67%, the path to profitability on deployment-grade robots gets longer, not shorter.

The Actuator Cost SA Didn’t Break Down

SemiAnalysis correctly identifies actuators as 66% of the BOM. They don’t break down what’s inside. We will.

The G1 uses PMSM motors with harmonic reducers, not pure QDD that eliminates the gearbox entirely. SA’s piece uses “QDD” loosely to describe the architecture, but the G1’s joints still use harmonic drives in key locations. This is an important distinction because harmonic reducers are the most expensive single component in the actuator stack.

Here’s the cost structure of a typical linear actuator in the G1:

Source: CNP Securities teardown, component supplier pricing, industry contacts.

The planetary roller screw alone accounts for nearly two-thirds of the linear actuator cost. And a basic ball-screw runs $100 to $120 per unit. A planetary roller screw, higher precision, better load capacity, longer life, runs $200 to $400+. That delta is where Unitree’s cost advantage lives. They’re not using the cheapest possible components. They’re using mid-tier components at scale pricing, which is a different and more defensible position.

The force/torque sensors are another line item SA glossed over. At $250–$400 per unit, these are not commodity items. They’re precision instruments that measure contact forces in real time, and they’re critical for any robot that needs to interact with the physical world safely. Unitree sources these domestically, which keeps costs down, but the sensor market is consolidating. If a single supplier gains pricing power, as HarmonicDrive did in strainwave gearboxes for decades, Unitree’s BOM advantage narrows fast.

No Custom Silicon: Manufacturing Advantage, Not Technology Moat

“Every critical semiconductor in the G1 is commercially available. Any competitor with a purchasing department can buy the same compute, the same sensors, the same motor drivers.”

This is the point nobody in the Unitree hype cycle is talking about.

Every chip in the G1 is off-the-shelf. The compute stack runs Intel i5/i7 processors paired with NVIDIA Jetson Orin NX modules. The LiDAR is a Livox MID-360. The depth camera is an Intel RealSense D435i. There is no custom silicon. No Unitree-designed ASIC. No proprietary inference chip. Nothing.

Compare that to Tesla Optimus, which is designing custom Dojo-derived inference silicon for its humanoid. Or to Apple, which has shipped custom silicon in every product category it enters. The companies that dominate hardware categories long-term almost always design their own chips. Unitree hasn’t done this.

Now, does it matter today? Not really. The commodity compute stack is cheaper, available at volume, and good enough for teleoperated and semi-autonomous tasks. But it means Unitree’s advantage is manufacturing and cost, not technology. Anyone with access to the same Chinese supply chain can build a robot with the same Intel + Jetson stack, the same Livox LiDAR, the same RealSense camera. The hardware is reproducible. The manufacturing process, the iteration speed, the supplier relationships, the vertical integration, is the moat. And manufacturing moats erode faster than technology moats.

There’s also a reliability implication. Consumer-grade compute components aren’t designed for continuous industrial duty cycles. That’s acceptable in a research lab. It’s a warranty nightmare in a warehouse running two shifts. SA mentions viability but doesn’t quantify what happens when a $5,700 robot needs a compute board swap after extended deployment.

Source: Teardown estimates, industry contacts, CNP Securities.

The Reliability Gap: 250 Units Is a Pilot, Not a Threshold

SemiAnalysis frames approximately 250 deployed G1 units as crossing a “viability threshold.” We’d frame it differently. Two hundred fifty units is a pilot program.

Forward risks (18 months):

1. Reliability data will surface failure modes that don’t show up in lab testing, narrative headwind for the “dominance” thesis.

2. Tesla Optimus may break $15,000 by 2027 with custom silicon and the same Chinese supply chain, compressing Unitree’s cost advantage from structural to marginal.

3. The first major service incident (property damage or injury) will trigger a regulatory and perception reckoning across the industry.

There is no published MTBF data for the G1. There are no duty cycle specifications. There are no long-term reliability studies. The robot carries 5 kg for 10–15 minutes before needing a cooldown. That's up from 2–3 minutes a year ago, impressive iteration, but it's not a factory floor number. A human warehouse worker handles 10–20 kg payloads for eight hours. The G1 handles 5 kg for 15 minutes with teleoperation.

The economic calculation SA presents, Unitree below $30/hour labor cost, assumes full teleoperation, a two-year useful life, zero residual value, and only two shifts. Those are conservative assumptions, and we respect the honesty. But “conservative assumptions” on a 250-unit deployment base is still a projection, not a track record. The first time a G1 drops a tote on a warehouse floor and damages $50,000 of product, or the first time a unit fails mid-shift and halts a conveyor line, those economics change overnight.

We’re not saying it won’t work. We’re saying the data doesn’t exist yet to prove it does. Two hundred fifty deployed units across multiple customers with no published uptime data is a proof of concept. BYD shipped 189,000 EVs in 2020 before the market took them seriously. Unitree has shipped 5,500+ G1 humanoids through 2024–2025. The scale difference is an order of magnitude.

The Real Winner: China’s Actuator Supply Chain

The picks-and-shovels play: Harmonic drive manufacturers, frameless motor specialists, and sensor fabs in Shenzhen and Suzhou are the real winners. They sell to every humanoid OEM, Chinese or Western. The component supply chain captures value regardless of which humanoid wins.

Here’s the part of the story SA touched on but didn’t fully develop, and it’s the most important investment thesis in the piece.

There are over 200 humanoid robot companies in China right now. Every one of them needs actuators. Every one of them needs gearboxes, motors, encoders, and sensors. And the supply chain that Unitree helped build is now available to all of them.

Source: Industry estimates, company filings, CNP Securities.

This is the real BYD parallel. BYD’s battery cell strategy created an ecosystem of suppliers, Hunan Yuneng, Shenzhen Dynanonic, Inovance, Sanhua, that now supply the entire Chinese EV industry. BYD won, but so did the component suppliers. The same dynamic is playing out in humanoids.

Companies like Leaderdrive (harmonic reducers), Inovance (motors), and the dozens of planetary gearbox manufacturers now shipping at volume, these are the picks and shovels of the humanoid gold rush. They win whether Unitree dominates or whether AGIBot, UBTECH, Fourier, or any of the other 200+ companies takes the crown.

Unitree’s verticalization is impressive. They self-develop BLDC motors, planetary gearboxes, LiDARs, and depth cameras. Their S-1 filing explicitly states that scaling production gave them upstream bargaining power, which translated to quadruped margins improving from 42.36% to 55.49% while costs dropped nearly in half. But even Unitree sources from the broader ecosystem. And that ecosystem is commoditizing fast.

Source: Industry contacts, CNP Securities, company filings.

If you’re looking for the investment thesis buried in SA’s piece, it’s not “buy Unitree.” It’s “watch the component suppliers.” The actuator supply chain is the infrastructure layer, and infrastructure layers capture value regardless of which application-layer company wins.

What Comes Next

Unitree’s upcoming IPO is a milestone, not a turning point. BYD’s IPO didn’t make it dominant, a decade of iteration did. DJI never needed to go public at all. An IPO gives Unitree capital to scale manufacturing and fund R&D. It doesn’t solve the reliability question or the margin question.

We expect three things in the next 18 months:

First, reliability data will disappoint. Not because the robots are bad, but because the expectations are too high. When you’re selling a $5,700 robot into industrial environments with consumer-grade compute and no published MTBF, early deployments will surface failure modes that don’t show up in lab testing. This is normal. Every hardware company goes through it. But it will create a narrative headwind for the “dominance” thesis.

Second, Tesla Optimus will break the $15,000 price barrier by 2027. Tesla has the same Chinese supply chain access, a larger engineering team, and custom silicon in the pipeline. If Optimus hits $15,000 with better reliability and Tesla’s brand in industrial sales, Unitree’s cost advantage compresses from structural to marginal.

Third, the first major service incident will reset expectations. When a humanoid robot causes property damage or injury in a commercial deployment, and it’s a matter of when, not if, the entire industry will face a regulatory and perception reckoning. Unitree, as the highest-volume player, will be first in the crosshairs.

Bottom Line

SemiAnalysis wrote the definitive overview of Unitree’s strategy, cost structure, and market position. The BYD/DJI analogy is tight. The BOM teardown is accurate. The ecosystem analysis is thorough. It’s a strong piece.

But “will dominate global robotics” is a timeline call, and the timeline is longer than the piece suggests. Unitree’s 50–65% blended margins are real but lower than the 67% headline. Their 250 deployed units are a pilot, not a threshold. Their lack of custom silicon means the moat is manufacturing, not technology. And 200+ Chinese competitors are drawing from the same supply chain that Unitree helped build.

The cost advantage is real. The strategy is sound. The BYD/DJI playbook is the right one. But dominance requires reliability data that doesn’t exist yet, and the deeper play, the one that doesn’t depend on which humanoid wins, is the component supply chain.

Watch the actuators. Watch the sensor suppliers. Watch the gearbox manufacturers. They’re building the infrastructure layer of humanoid robotics, and they win regardless of whether Unitree’s G1 becomes the iPhone or the Galaxy of this market. That’s the thesis worth betting on.

Every humanoid robot carries an invisible mortgage. It’s not paid in dollars. It’s paid in watts, drawn from the same battery, competing with the same actuators, generating the same heat. The more intelligence you put onboard, the less machine you have left to move.

We’ve been tracking this problem across the humanoid robotics supply chain for 18 months. Our Humanoid Robot Power Model, first detailed in Part 1 of this series, identified compute as the single largest non-locomotion power draw in every major platform. Not sensors. Not comms. The brain. 180 watts, on average, across the current generation of humanoids performing full-stack autonomous tasks. That’s roughly equal to what the legs consume during brisk walking.

The compute tax is the most underappreciated bottleneck in humanoid robotics. Everyone talks about actuator torque density, battery energy density, finger dexterity. Almost nobody talks about how the robot’s brain is burning a lighting circuit’s worth of power just to figure out where to put its foot next, and that every additional watt of intelligence comes directly out of the locomotion budget.

The Zero-Sum Power Budget

A humanoid robot is not a data center on legs. In a data center, if your GPU draws 700W, you bolt on a liquid cooling loop, pull another 30kW from the grid, and call it Tuesday. Nobody cares about the power budget, the hyperscalers sign PPAs for 500MW campuses without blinking.

A humanoid has a battery. A 2.5 kWh pack, typically. Maybe 3 kWh in the next generation. At 500W total system draw, our model’s baseline for full-stack autonomous operation, that gives you five hours of runtime. Cut it in half for battery degradation, safety margins, and peak load headroom, and you’re looking at 2–2.5 hours of real useful work.

Now look at where the power goes. Our breakdown across six representative task loads makes the trade-off brutally clear:

Compute power scales with task complexity, but the locomotion budget is non-negotiable. Source: Machine Narratives Research. Humanoid Robot Power Model.

At idle, compute sips ~40W and actuators draw ~20W for balance. Walking ramps locomotion to ~180W while compute stays modest at ~60W, basic gait control doesn’t require heavy inference. But the moment you add real-time perception. SLAM, object detection, visual language models parsing human instructions, compute jumps to 150–170W while locomotion is still running. Full-stack autonomous operation pushes compute past 200W and locomotion past 300W. The total exceeds 500W.

the robot’s ability to think and its ability to move compete for the same electrons. There is no budget separation. No separate battery for the brain. Every watt burned on inference is a watt the knee actuators don’t get, and the knee actuators don’t negotiate.

Why Reasoning Burns as Much as Walking

This surprises people who haven’t worked with edge inference. The intuition is that walking should be the expensive part, all those motors, all that torque, all that mechanical work. And it is expensive. But inference is worse than you think, for reasons that are obvious once you look at the silicon.

Walking is biomechanically efficient. Human gait evolved over millions of years to minimize energy expenditure. A human walking at 5 km/h consumes about 280W of metabolic power for the whole body, brain included. A robot’s actuators are less efficient than muscle, but the gap is closing: modern electric actuators with harmonic drives and regenerative braking can achieve 60–70% electromechanical efficiency. The locomotion subsystem in a well-designed humanoid can run at 180–250W for steady-state walking.

Inference, meanwhile, has no evolutionary optimization. A vision transformer processing six RGB camera feeds at 30 FPS, plus a language model parsing instructions, plus a world model predicting outcomes, plus a planning stack generating trajectories, this is not a lightweight compute workload. It’s a data center inference server compressed into a package the size of a lunchbox, running on a battery.

Our supply chain contacts report that the perception + planning stack on current-generation humanoids commonly requires:

• 50–80 TOPS for real-time SLAM and depth estimation across 6–8 cameras

• 30–50 TOPS for the VLM running instruction parsing and scene understanding

• 20–30 TOPS for the world model and trajectory planning

• 10–20 TOPS for lower-level motor control and safety monitoring

Total: 110–180 TOPS sustained. At the power efficiency of current edge silicon, roughly 1–2 TOPS per watt, that’s 55–180W for the compute subsystem alone, before you add sensor power, comms, and overhead. Our model says 180W is the right number for a full-stack autonomous workload. That matches independent teardown estimates from three separate integrators we’ve spoken with.

Walking is 180W. Thinking is 180W. The robot spends as much energy deciding what to do as it does doing it.

The perception-to-action pipeline runs at 100% duty cycle, every frame, every step, every reach. Unlike a phone or a data center GPU that spikes and idles, the robot's inference engine never gets a coffee break. Source: Machine Narratives Research based on integrator teardowns.

The Thermal Problem Nobody’s Solved

A 180W chip in a data center is trivial. A 180W chip inside a human-shaped enclosure is a thermal nightmare.

Data centers have forced air at 2 meters per second, cold aisles at 18°C, hot aisle containment, rear-door heat exchangers, and direct-to-chip liquid cooling for anything above 500W. A humanoid has none of this. The compute module sits inside the torso, encased in structural polymers and, increasingly, covered in bionic skin that has thermal conductivity roughly 1,000 times worse than aluminum. There is no airflow path. No heat pipes to an external radiator. No chilled water loop.

We’ve reviewed the thermal management approaches across the major platforms:

• Tesla Optimus uses a passive chimney effect, channel linings in the torso garment that pull cool air in at the waist and exhaust warm air at the collar. Clever, but limited to roughly 80–100W of sustained dissipation before the chip throttles.

• Figure 02 runs dual NVIDIA GPUs in the torso and relies on conductive heat spreading through the chassis frame. Our contacts suggest thermal throttling kicks in after ~20 minutes of sustained full-stack operation.

• Boston Dynamics Atlas takes the defense-grade approach: custom conduction-cooled compute modules mated to the chassis structure, with the entire torso acting as a heat sink. Effective but heavy, adds roughly 8 kg to the upper body mass.

• Unitree H1 uses an Intel Core processor with a small active fan. The fan is a reliability risk, moving parts in a dusty factory environment fail, and fan noise interferes with the robot’s own microphone array.

The academic literature is converging on microfluidic cooling, flexible microchannels embedded in the torso structure that circulate coolant through the robot’s frame. A paper published in Device in late 2024 demonstrated 120W heat rejection through a flexible microfluidic network without adding external radiators. But lab-scale is not production-scale. We’re years away from a humanoid walking off the assembly line with liquid cooling loops in its chest.

The thermal ceiling, right now, is roughly 120–150W of sustained compute heat dissipation in the human form factor. Above that, you throttle. Below that, you leave inference performance on the table. The compute tax has a thermal component: you pay for every watt of intelligence with a watt of heat that has nowhere to go.

A sealed humanoid torso is a Faraday cage for heat. No airflow. No radiator. No place for 180W of compute to go except into the battery, the motors, and the plastic housing. The thermal ceiling is the hard constraint on onboard intelligence. Source: Machine Narratives Research.

The edge AI chip landscape. NVIDIA's Jetson Thor pushes 100W, right at the thermal ceiling for an unassisted humanoid torso. Source: Machine Narratives Research. Vendor specs, May 2026.

The Silicon Landscape: Who’s Solving It?

There are four distinct approaches to the humanoid compute problem, each with different assumptions about what a robot brain should cost, in watts, dollars, and supply chain dependency.

Consider subscribing to see the rest of the article below

Read more

In 1999, two Japanese companies. Harmonic Drive Systems and Nabtesco, made a quiet agreement. Harmonic Drive would focus on smaller, high-precision harmonic reducers for the wrist joints and elbows of industrial robots. Nabtesco would focus on larger, high-torque RV reducers for the shoulders and hips. Together, they divided the global precision reducer market between them, a gentleman’s arrangement that held for nearly two decades. If you wanted to build a robot that moved precisely, you bought from one of these two companies. End of story.

That agreement was dissolved in 2022. Nabtesco sold its approximately 19% stake in Harmonic Drive Systems by early 2025, netting roughly $357 million in the process. The two companies, which had been cross-shareholders and strategic partners since 2005, are now independent competitors in a market that’s about to get dramatically bigger, and dramatically more competitive. that timing tells you everything you need to know about where this industry is heading.

the endgame. Over the past week, we’ve traced the actuator supply chain from its origins in a 1955 patent by C. Walton Musser, through forty years of Japanese manufacturing dominance, to the Chinese disruption that’s now remaking the competitive landscape. Amanda J mapped the bottleneck in Part 1. I traced the technology arc from harmonic drives to torque motors in Part 2. Phil D showed us the Chinese market share surge in Part 3 and the procurement wars between Tesla, Figure, and Unitree in Part 4. Now I want to synthesize all of that into a single question: who wins, who loses, and what do you do about it?

the answer has three parts. But before I get to the investment thesis, let’s be precise about what we’re actually looking at, because the numbers matter, and in my experience, most of the market commentary on the actuator supply chain is about five years behind the data.

This matters because the precision reducer oligopoly isn’t one company controlling 85% of the market, as some earlier analyses suggested. It’s more nuanced, and in some ways, more interesting. What we actually have is a market dominated by roughly five players, with the top two (HDSI and Nabtesco) holding a combined ~60% share of the high-end precision reducer market for robotics applications, and a third Japanese player (Nidec-Shimpo) adding another ~12%. The Chinese challengers. Green Harmonic, Leaderdrive, Laifual, and a dozen smaller players, have collectively grown from under 10% in 2020 to somewhere around 30% today.

Figure 1: Estimated precision reducer market share in robotics applications, early 2026. Japanese incumbents still hold roughly 70% combined, but the direction of travel is clear. Source: Machine Narrative Research estimates.

HDSI is the clear leader in harmonic drives specifically, the smaller, higher-precision reducers used in robot wrists, elbows, and increasingly, humanoid hand joints. I'd estimate their share of the high-end harmonic drive market at roughly 60%, though exact numbers are hard to come by since the market isn't tracked by any single authoritative source. Nabtesco dominates RV reducers, the larger, higher-torque gearboxes for shoulder and hip joints. The distinction matters because the humanoid robot buildout is likely to require both types: harmonic drives for the dozens of small joints in the hands and wrists, RV reducers for the larger load-bearing joints in the arms and legs.

The Rare Earth Trap

Let me pause here and talk about something that I think gets lost in most supply chain analyses, because it sits one layer deeper than the gearboxes and the torque motors. Every precision actuator, whether it uses a harmonic drive, a planetary gear, or a direct-drive torque motor, depends on permanent magnets. Specifically, neodymium-iron-boron (NdFeB) magnets, the most powerful commercially available permanent magnets in the world. And China controls roughly 90% of the global supply chain for NdFeB magnet manufacturing.

This is not a mining story. The rare earth elements themselves, neodymium, praseodymium, dysprosium, are mined in multiple countries, including the United States (Mountain Pass in California), Australia (Lynas’s Mount Weld), and Myanmar. The bottleneck is in the midstream processing: separating the oxides, producing the alloys, and sintering the finished magnets. That processing infrastructure took China roughly fifteen years to build, and it cannot be replicated quickly. The US, Australia, and Canada are all investing in domestic processing capacity, but even under the most optimistic timelines, meaningful non-Chinese NdFeB magnet production is five to seven years away.

In April 2025, China imposed new export controls on selected rare earth magnets and separation technologies, citing national security concerns. The rules require export licenses for a range of rare earth products, including NdFeB magnet materials. Starting December 1, 2025, the controls were extended extraterritorially, meaning overseas entities using Chinese rare earth processing technologies also need licenses. This is not a hypothetical risk. It’s an active policy framework that can be tightened at any time.

Figure 2: Humanoid robot NdFeB magnet demand as a percentage of global production, 2025–2030F. Today it’s a rounding error (0.03%). By 2030, at 1.2 million robots, it becomes a meaningful demand driver. Source: Adamas Intelligence, Tianfeng Securities, Machine Narrative Research estimates.

Each humanoid robot requires approximately 3.5 kilograms of NdFeB magnets across its actuators. At 50,000 robots, roughly the midpoint of consensus 2026 estimates, that’s about 175 metric tons, or less than 0.1% of global NdFeB production. Nobody in the rare earth industry loses sleep over 175 tons. But the math changes quickly. At one million robots, you need 3,500 metric tons. At ten million. Musk’s stated ambition for Optimus, you’d need 35,000 metric tons, which would represent roughly 14% of current global production. That’s not a rounding error anymore. That’s a market-moving quantity of a material where one country controls the processing infrastructure.

this is the underappreciated risk in the actuator supply chain thesis. Everyone focuses on the gearboxes, who makes the best harmonic drive, who’s got the Tesla contract, whether torque motors will displace gears entirely. But all of those debates assume the raw materials are available. The rare earth trap means that even if you solve the gearbox problem, even if you build the factory and qualify the supplier and lock in the price, you’re still dependent on a single country for the magnets that make every actuator work. That dependency is not diversifiable in under five years. It might not be diversifiable in under ten.

The Chinese Disruption, In Context

Phil D laid out the numbers in Part 3 with admirable precision. Chinese harmonic drive manufacturers have grown from under 10% market share to roughly 30% in five years, with Green Harmonic leading the charge at 12–18% and growing. I want to add some context to those numbers, because I think the pattern matters as much as the level.

Green Harmonic (688017.SH) was founded in 2011 by a pair of brothers in Suzhou. Their first products were, by most accounts, not competitive with Japanese harmonic drives in terms of precision or durability. But they were 40–60% cheaper, and for a Chinese industrial robot manufacturer trying to build a cost-competitive product, that price difference was the difference between viable and non-viable. Over the next decade, Green Harmonic improved its quality steadily, not through any single breakthrough, but through the same grinding, iterative refinement that Japanese manufacturers had used in the 1980s and 1990s. By 2023, their products were good enough for Tesla to qualify them as the primary harmonic reducer supplier for Optimus.

The economics are instructive, and they invert the conventional narrative. HDSI’s gross margins have compressed to approximately 27% as of FY2024. Green Harmonic’s gross margins are approximately 37.5%, a full ten percentage points higher. The Chinese challenger has better margins than the Japanese market leader, and it’s selling at a 40–60% discount. That is not a healthy competitive dynamic for the incumbent. A company with 37.5% gross margins and a large discount to the market leader has room to keep cutting prices. A company with 27% margins and declining revenue does not.

“The challenger has better margins than the market leader and charges 40–60% less. That’s not a competitive dynamic. That’s a countdown.”

The other Chinese players. Leaderdrive, Laifual, and a growing number of smaller manufacturers, are collectively growing faster than any Japanese incumbent. Their products are still primarily sold into the Chinese domestic market, where industrial policy actively favors local suppliers. But the moment a Western humanoid OEM qualifies a second-source Chinese harmonic reducer supplier, the pricing umbrella that supports Japanese margins collapses entirely. that moment is coming within 2–3 years, not 5–10.

I should be fair here. The Japanese incumbents are not passive. HDSI has launched integrated motor-gearbox modules, essentially a harmonic drive with a built-in torque motor, that they’re marketing as a drop-in solution for humanoid robot joints. Nabtesco is expanding RV reducer capacity with humanoid-specific product lines. Nidec-Shimpo is leveraging Nidec’s broader motor expertise to offer bundled actuator packages. These are smart moves, and they reflect the kind of thoughtful, long-term strategic thinking that Japanese industrial companies are famous for. But I think the historical parallel is instructive here, and it’s not a comforting one for the incumbents.

The Solar Panel Parallel

In 2005, Japanese and German manufacturers. Sharp, Kyocera, Q-Cells, dominated the global solar panel market with a combined share of roughly 70%. Their products were premium. Their manufacturing was world-class. Their institutional knowledge was measured in decades. Over the next ten years, Chinese manufacturers. Longi, Trina, Jinko, entered with 30–40% lower prices, government-backed capacity expansion, and an aggressive scaling strategy that prioritized market share over near-term margins.

The incumbents didn’t stand still. They improved their technology continuously. Their products remained, by many objective measures, superior. But by 2015, Chinese companies held over 60% of global solar panel production. Japanese share collapsed from roughly 50% to under 10%. The premium product didn’t lose on quality, it lost on cost, in a market where “good enough” was, in fact, good enough.

The harmonic drive market is smaller, roughly $5–7 billion globally compared to solar’s $200+ billion, but the structural dynamics are identical. A premium incumbent with high margins. A challenger entering at 40–60% lower prices with state-supported capacity expansion. An end-customer base under constant pressure to reduce bill-of-materials costs. And a “good enough” quality threshold that, for the fastest-growing applications, is lower than the incumbents would like it to be.

Not to be alarmist, but I think the precision reducer market is about five years into a ten-year disruption that ends the same way the solar disruption ended: with Chinese manufacturers controlling the majority of global production, Japanese incumbents retreating to the highest-precision, highest-certification niches, and a fundamentally restructured competitive landscape that nobody in Nagano Prefecture wanted but that the economics made inevitable.

Three Strategies, Three Bets

Read more

In 1991, Sony shipped the first commercial lithium-ion battery. It was a cylindrical cell, 18 millimeters across, 65 millimeters tall. You have almost certainly held one. They call it the 18650, and for the next three decades, this little metal tube would become the default shape of stored energy. It powers laptops, power tools, electric cars, and, reluctantly, walking robots.

The 18650 is a beautiful piece of engineering. It is cost-optimized, supply-chain-proven, and produced at a scale of billions of units per year. When Tesla decided to build the Model S, they didn’t design a custom battery. They bought 6,831 little metal tubes and wired them together. The design was elegant in its own way: a flat chassis floor, a rectangular battery compartment, and thousands of identical cylinders packed like sardines.

But a humanoid robot does not have a flat chassis floor.

The shape that started it all

Let us start with the 18650 itself. The name tells you everything you need to know: 18 millimeters in diameter, 65.0 millimeters in height. The naming convention is literally just the dimensions, which I find charming in its industrial honesty. A 21700 cell is 21 by 70 millimeters. A 4680 cell. Tesla’s latest, is 46 by 80 millimeters.

These cylinders are not an accident of physics. They are an accident of manufacturing economics. When you wind a battery, rolling up layers of cathode, anode, and separator into a spiral, the natural shape is a cylinder. The “jelly roll” construction that gives the 18650 its cylindrical form is also what makes it cheap: the winding machine just spins, the can slides on, the electrolyte gets pumped in, and the whole thing seals shut. It is a stunningly efficient manufacturing process that produces a stunningly awkward shape for a thigh.

Here are the numbers that matter for what we are discussing:

Source: Battery Design, HighStar Lithium. Energy density ranges reflect NMC cathode formulations. Pack-level density is typically 34% lower due to BMS, cooling, and structural housing.

I want you to look at the rightmost column. A single 18650 holds about 12.5 watt-hours. The Tesla Optimus carries a 2,300 watt-hour battery pack. To build that pack from 18650s, you would need roughly 184 individual metal tubes, plus the wiring, the battery management system, the cooling hardware, the structural housing that holds them all in place. And all of that has to fit inside a torso that is roughly the size of a human ribcage, shaped like an inverted teardrop, and already occupied by motors, wiring harnesses, sensor arrays, and compute modules.

The elementary arithmetic of robot batteries. At current energy densities, energy storage consumes a third of the robot's mass budget, before you account for motors, structure, or compute. Source: Machine Narrative Research.

This is the leg battery problem. It is not that we lack the energy. It is that the energy comes in the wrong shape.

Why an EV battery pack is a terrible leg

If you have ever looked at an EV battery teardown, and I recommend it, it is fascinating, you will notice something immediately. The pack is a flat rectangle. The cells sit in rows, standing upright or laid on their sides, packed into modules that are packed into a flat floor pan. The entire structure is optimized for something that is essentially a skateboard. A 65-kilogram humanoid robot does not have a skateboard. It has two legs, a pelvis, a torso, and a head. The available volume for energy storage is, to put it mildly, architecturally hostile.

There are three distinct problems that cylindrical EV cells create when you try to put them in a walking machine. it is worth walking through each one.

First, the packing fraction problem. When you pack cylinders into a rectangular box, at best you fill about 91% of the volume, the theoretical maximum for circles in a plane. In practice, with the spacing needed for thermal management and structural integrity, you are looking at more like 70-75% volume utilization. The remaining 25-30% of the battery compartment is wasted air. In an EV, this is acceptable because the floor pan is large enough that the penalty is absorbed. In a humanoid torso that can hold maybe 10-15 liters of battery at the absolute maximum, losing a quarter of that volume to empty space is catastrophic.

Second, the center of mass problem. A humanoid robot is bipedal. It balances on two feet. The placement of every gram matters for stability, gait efficiency, and fall recovery. Cylindrical cells want to be arranged in grids, which creates rectangular blocks of mass that cannot easily follow the organic contours of a chest cavity or a thigh. Put the battery in the torso, and the robot is top-heavy. Distribute it through the legs, and you have lumpy cylinders interfering with joint range of motion. There is no good answer with a cylindrical cell, only less-bad compromises.

Third, the thermal problem. Walking draws bursts of high current, a humanoid robot going from standing still to lifting a 20-kilogram box can spike its power draw from 100 watts to over 1,000 watts in under a second. Cylindrical cells, with their wound internal structure, have relatively high internal resistance compared to pouch or prismatic designs. Under burst loads, they heat up unevenly, with the center of the “jelly roll” running hotter than the outer layers. This thermal gradient accelerates degradation and creates hot spots that trigger the battery management system to throttle power, precisely when the robot needs it most. An EV accelerates gradually. A robot takes a step.

THE BURST VS. STEADY PROBLEM

An EV drawing 20 kW to accelerate from 0-60 mph is pulling about 5C from a 4,000-cell pack, around 5 amps per cell, well within the comfort zone.

A humanoid robot spiking from 100 W idle to 1,000 W during a lift from a 100-cell pack is pulling 10C, or roughly 30 amps from a single 18650 cell.

At 10C, internal heating becomes significant, voltage sag is pronounced, and effective usable capacity drops by 15-25% compared to the rated 0.2C capacity.

The same battery chemistry under these two different loading profiles produces two different machines.

EV cells are engineered for steady discharge at moderate C-rates. Robot cells must survive violent pulse loads — walking, lifting, catching a stumble — that create thermal and mechanical stress no automotive cell was designed for. Source: Machine Narrative Research.

A humanoid robot’s power draw is anything but steady. Walking creates 3-5x spikes in instantaneous demand — the battery must deliver peak power at a moment’s notice without voltage sag triggering a BMS shutdown. Source: Machine Narrative Research based on published bipedal gait models.

The shape of batteries to come

Look at what happens when engineers are freed from the cylinder.

Pouch cells, flat, flexible, laminated packages of electrode layers, are perhaps the closest we have to a robot-native battery format today. Instead of a metal can, a pouch cell wraps the active materials in a thin aluminum-polymer laminate. The result is something that looks more like a high-tech envelope than a battery. You can make them thin, as little as a few millimeters, and shape them to follow curved surfaces. You can stack them, bend them, and fit them into spaces that would make a cylindrical cell engineer weep.

The trade-off is real, though. Pouch cells lack the rigid mechanical protection of a metal can. They swell slightly with age and cycling, which means you need to leave expansion gaps in the pack design, gaps that cost volume. Their energy density at the cell level tends to be lower than the best cylindrical cells, more like 180-240 Wh/kg, because the flexible packaging is less space-efficient than a thin steel can. But at the pack level, the math reverses. When you can fill 95% of the available volume instead of 70%, the system-level Wh/kg can actually exceed what you get from cylindricals, even if each individual cell stores less.

This is the counterintuitive thing about robot battery design. The best cell on a spec sheet is often not the best cell when it is sitting inside a thigh.

No single format wins on all dimensions. Cylindricals are proven but waste space. Pouches conform beautifully but need armor. Prismatics split the difference. The robot industry is still experimenting. Source: Machine Narrative Research.

Prismatic cells represent a middle path. They use a rigid rectangular aluminum housing, more protective than a pouch, more space-efficient than a cylinder. The internal structure can be either wound (like a flattened jelly roll) or stacked (like a pouch). Prismatics have become the dominant format in Chinese EVs, supplied by companies like CATL and BYD, and they share many of the same advantages for humanoid robots: high packing efficiency, good thermal contact surfaces, and the ability to serve as partial structural elements.

But the format that I think matters most, the one that could change the shape of this problem, is something called a structural battery.

Three paths to the robot battery: cylindrical cells in a traditional pack, flexible pouch cells conforming to the body, and structural batteries that serve as load-bearing chassis elements. Each trades off density, manufacturability, and mechanical integration. Source: Machine Narrative Research.

The structural battery idea

Let me tell you about something remarkable that happened with electric aircraft.

Read more

The humanoid robot industry is not fighting over AI models. It’s fighting over supply chains for electric motors the size of a fist. The reality is that whoever controls the actuator supply chain controls the manufacturing cost curve, and whoever controls manufacturing cost controls the TAM. The market is mispricing who wins this fight.

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.

Tesla: The Hybrid Empire

Tesla’s approach is best described as a hybrid: design in-house, manufacture through captive Chinese suppliers, assemble in Austin. It is neither pure vertical integration nor pure outsourcing. It’s the automotive supply chain playbook applied to robotics, and it’s working.

The key partnerships tell the story. Ningbo Tuopu Group (601689) is the exclusive supplier of rotary actuators to Tesla Optimus. Tuopu wasn’t a robotics company, it was an automotive chassis and suspension supplier that had been in Tesla’s EV supply chain since 2016. When Tesla needed someone who could design frameless torque motors and integrate harmonic reducers at automotive scale and automotive cost, Tuopu was the obvious candidate. Mass production of Tuopu’s rotary actuators began in 2024, with a single-vehicle value exceeding ¥10,000 (~$1,400).

Sanhua Intelligent Control (002050) got the linear actuator side. Sanhua controls the #1 global market share in electronic expansion valves and over 35% of the broader thermal management components market, a completely different product category, but its precision manufacturing of small electromechanical assemblies transferred directly to planetary roller screw integration. The $685 million order in October 2025 wasn’t a gamble on an unproven supplier. It was capacity reservation on a partner that had already proven it could hit automotive quality at automotive cycle times.

Behind both of these sits Green Harmonic (688017), the domestic Chinese harmonic reducer manufacturer that has broken the monopoly of Japan’s Harmonic Drive Systems. Green Harmonic now controls over 60% of China’s harmonic reducer market and is the primary harmonic reducer supplier for Tesla Optimus. Supply chain analysis shows Tesla is Green Harmonic’s largest customer by volume, though exact revenue percentages are not publicly disclosed. As we detailed in our actuator cartel supply chain mapping, when we talk about supply chain concentration risk in humanoid robotics, we’re really talking about one factory in Suzhou.

Actuators account for approximately 56% of the total value of major components in Tesla Optimus, per supply chain analysis from Yicai Global and Futu Securities.

HCFA Technology (688320) supplies frameless torque motors to both the Tesla chain and the Huawei chain, as one of the leading domestic frameless torque motor manufacturers with significant market share. Keli Sensing (603662), the only Chinese manufacturer of six-dimensional torque sensors, benchmarked against ATI in the United States, has been sampled to Huawei and is seeking Tesla qualification. The pattern is clear: Tesla is building a China-dependent actuator supply chain that looks increasingly like its China-dependent battery supply chain circa 2019.

The fragility here is geographic concentration. China controls roughly 90% of global NdFeB rare earth magnet supply, and export licenses have been required since April 2025. Each Optimus requires approximately 3.5 kg of NdFeB magnets. At 100,000 units, that’s 350 metric tons, a rounding error in the global rare earth market. At 10 million units, it’s a strategic vulnerability. The same supply chain that gives Tesla cost advantage today is the one that creates single-point-of-failure risk at scale. Our supply chain model shows that a rare earth export disruption would hit Tesla’s actuator cost curve harder than any competitor’s.

Three procurement strategies, three supply chain architectures, and no consensus on which approach produces structural cost advantage at scale.

Figure: The Full Stack Gambit

Figure AI is doing something fundamentally different, and more expensive in the near term. The company designs everything in-house: motors, actuators, sensors, battery systems, embedded software, firmware, controls. Brett Adcock has been explicit: “Figure is vertically integrated and designs everything in house.” This is not a procurement strategy. It’s a declaration of war on the supplier concept itself.

Read more