The physical AI supply chain, three levels down (Part 1 of 5)

Hey {{first_name|Investor}} -

Robots are having their moment. Unitree humanoids dance on Chinese national television, Figure bills BMW by the robot-hour, and Tesla is converting its old Model S line in Fremont for Optimus. TrendForce forecasts global humanoid shipments could pass 50,000 units in 2026, roughly a 700% jump over 2025; treat that as a scenario, not a schedule. The robot makers and their visible component suppliers have already re-rated hard on the story (their stock prices rose sharply on the theme).

That re-rating is exactly why I want to look somewhere else. When a theme gets crowded at the surface, the asymmetry moves down the chain. This report builds the framework for going three levels deeper than the coverage you see on social media, maps the full supply chain from the robot down to the raw material, and flags where I see the strongest combination of bottleneck exposure and pricing power. Each future report in this series takes one link and goes all the way down.

As of early July 2026, China dominates physical volume. Industry estimates put close to 90% of the humanoids sold in 2025 as Chinese, led by Unitree, which shipped over 5,500 units and is targeting somewhere between 10,000 and 20,000 in 2026. The Western programs are earlier but accelerating: Figure reports a robot-per-hour production rate at its BotQ factory and has around 40 units working at BMW's Spartanburg plant, while Tesla's Optimus V3 is expected to enter mass production in the second half of 2026, with first external sales targeted for late 2026.

Asia also front-ran the component trade. Chinese reducer, screw, and motor names re-rated through 2025 on the visible supply-chain story, and Chinese suppliers hold an estimated 63-70% of the global humanoid component chain with a 30-40% cost advantage, figures from domestic Chinese industry reports rather than global consensus. That first layer is where attention already lives.

The deployment numbers stay small relative to the projections. Fifty thousand robots is a rounding error against the multi-hundred-million-unit forecasts that banks publish for the 2040s, and I treat any projection past 2030 as scenario-building rather than forecasting. What matters for us is the direction: unit volumes are inflecting now, every incremental robot pulls a bill of materials behind it, and the capacity to produce those materials is being ordered today.

That last point is the whole thesis of this report. Robots are uncertain. The tooling to build robots is being bought right now.

Picture the supply chain as strata. Each layer down gets less attention, and attention is what prices things.

Level 0 is the robot itself. Tesla, Figure, Unitree, AgiBot. This is where the narrative lives and where the valuations carry the most expectation.

Level 1 is the component. Actuators, reducers, roller screws, cameras, LiDAR, torque sensors. This is where most "picks and shovels" coverage stops, and it is crowded: China's reducer and screw names were retail favorites through 2025, and the accessible Western names get mapped in every bank's value-chain report.

Level 2 is the machine that makes the component. A roller screw is only as precise as the thread grinder that cut it, and an image sensor is only as good as the equipment that deposited its coatings. This layer is thinner, more specialized, and much less discussed.

Level 3 is what the machine consumes and what the component is made from. Grinding wheels, cutting tools, bearing-grade steel, optical glass, heat treatment, and the metrology gear that certifies all of it. This layer is nearly invisible in the humanoid conversation, and it has a property I love: it gets paid on production volume regardless of which component design wins. Call it the toll booth layer.

The asymmetry logic is simple. At Level 0 you are betting on which robot wins. At Level 1 you are betting on which component architecture wins, and there are live architecture battles at that layer: rotary versus linear actuation, vision-only versus LiDAR. By Level 3, most of those bets cancel out. Whichever screw maker wins, the threads get ground, the grinder burns through wheels, the screw starts life as a bar of special steel, and someone has to measure the result to sub-micron accuracy. You are no longer betting on outcomes. You are betting on activity. For a stock picker, that is the entire attraction: the deep layers get paid on robot activity regardless of which robot brand wins.

Consumables are the second property. A thread grinder is a one-time sale, but the superabrasive wheels it wears out are a subscription tied to every meter of thread ground, the way a printer sells you ink for as long as you print. Level 3 is full of these quiet toll collectors.

The full chain for both systems, robot to raw material. Names in bold are ones I discuss below.

A humanoid's big load-bearing joints, the hips, knees, and ankles, mostly run on planetary roller screws: a motor spins a threaded shaft, and the spinning becomes straight-line force, the way twisting a car jack's handle becomes lifting force. These screws run roughly $1,000 to $3,000 each and a single robot can need ten or more of them. Morgan Stanley's value-chain work expects roller screws to become the dominant screw type in humanoids over time. The delicate joints, wrists and arms, run on harmonic reducers, gearboxes whose flexible ring eliminates slack between gear teeth.

The screw and reducer makers are the Level 1 story, and China has been busy: domestic reports put Chinese localization of harmonic reducers, roller screws, and six-axis force sensors above 90%, with component costs down 30-40% as local capacity scales. Margin compression at Level 1 is already visible in that sentence. So step down a level.

A roller screw's threads have to be ground to micron-level tolerances, and the machines that do this grinding are made by a small club of mostly European and Japanese firms. Chinese industry commentary consistently flags high-precision thread grinders as an import dependency, and the Jamestown Foundation estimates China still imports around 90% of its high-end machine tools, mostly from Japan. When Chinese screw makers announce billion-yuan capacity expansions, as Beite did with a base designed for up to 2.6 million screw sets a year, those factories get filled with grinding machines first. The machine order comes before the screw revenue, so this layer moves first.

The listed ways in:

Qinchuan Machine Tool (SZSE: 000837) sells the thread grinders that make roller screws and the gear grinders that make reducers, and its Hanjiang subsidiary can build the screws themselves. Vertically integrated across Levels 1 and 2.

Hiecise Precision (SZSE: 300809) is China's roll-grinder specialist pivoting into thread grinders, and it booked a 100-machine order in March 2025 specifically for roller screw machining, reported as the first bulk domestic order of its kind.

Chongqing Machinery & Electric (SEHK: 2722) owns UK-based PTG Holroyd, a world leader in thread and rotor grinding machines. A Chinese conglomerate holding one of the West's crown-jewel thread-grinding franchises is its own geopolitical story.

Tsugami (TSE: 6101) makes precision thread and form rolling machines in Japan, with a separately listed China subsidiary.

Nidec (TSE: 6594 / OTC: $NJDCY) is known as a motor company, but its machine tool arm, built from the former Mitsubishi Heavy machine tool business and OKK, sells gear grinders and cutting machines into exactly this capacity build-out. The motor side already supplies humanoid joints, so it touches the theme at two levels.

Klingelnberg (SWX: KLIN) is the Swiss-listed specialist in gear grinding and, importantly, gear measurement. The private Swiss and German names, Reishauer and Kapp Niles, dominate continuous generating grinding but you cannot buy them, which concentrates the listed exposure in the few names above.

Precision grinding of hardened steel threads and gear teeth consumes superabrasive wheels, typically CBN, cubic boron nitride, a synthetic material nearly as hard as diamond. These wheels wear out and get replaced on a schedule tied directly to production volume, making them the most overlooked layer in the whole map. The machine is the razor and the wheels are the blades.

Noritake (TSE: 5331) and Asahi Diamond Industrial (TSE: 6140) are Japan's listed superabrasive specialists, and Asahi explicitly markets vitrified CBN wheels for super-finished bearing raceway grinding, the same precision class the screw and bearing chain requires. Neither shows up in humanoid value-chain reports, and that quiet is exactly what I am screening for. Every screw factory in China and every reducer line in Japan feeds this duopoly-like niche.

The same logic extends to cutting tools. OSG (TSE: 6136) is the world's benchmark in taps and thread-cutting tools, and Sandvik (STO: SAND) owns the broadest carbide tooling franchise on earth. Tooling is a spread bet on machining activity everywhere, so humanoids are one tailwind among many rather than a pure play, but that diversity is also what keeps the downside boring.

Roller screws, bearings, and gears start life as high-cleanliness special steel, the kind where a single microscopic inclusion can become a fatigue crack under millions of load cycles. This is a concentrated industry: CITIC Pacific Special Steel (SZSE: 000708) leads globally, with Japan's Sanyo Special Steel (TSE: 5481) and Daido Steel (TSE: 5471) as the quality benchmarks, and Carpenter Technology (NYSE: $CRS) playing that role in the US. Daido just consolidated further, acquiring Nippon Koshuha Steel in February 2026. Steel is the least humanoid-levered layer in this report, since robots will be a small slice of demand for years, but it is also where the deepest moats sit. Nobody qualifies a new bearing-steel supplier quickly.

Two service layers certify everything above. Heat treatment hardens the screw and gear surfaces, and Bodycote (LSE: BOY) is the world's largest outsourced heat-treatment network. Metrology proves the parts meet spec, and precision component production is metrology-bound: six-axis torque sensors, for example, are calibration-intensive to the point that McKinsey flags measurement infrastructure as a scaling constraint. Renishaw (LSE: RSW) sells the probes and calibration systems that live inside precision machine shops, and also happens to supply magnetic encoders into humanoid joints through its RLS stake, another two-level touch. Hexagon (STO: HEXA-B) owns the coordinate-measurement franchise. The private names here, Zeiss, Mitutoyo, and Heidenhain, again concentrate listed exposure into few hands.

The sensing chapter of this theme usually gets told as a chip story, and the chips matter: a humanoid carries seven to eight cameras, Optimus reportedly carries eight, and robot cameras need global-shutter sensors, where every pixel captures the same instant, rather than the rolling-shutter type in your phone, which scans the scene line by line like a photocopier and smears anything that moves. Sony holds roughly half the CIS market and is deepening its robotics commitment through a joint venture with TSMC announced in May 2026 for next-generation AI image sensors. onsemi leads in global shutter for machine vision and is absorbing Synaptics.

But the chip story is Level 1, and it is covered. The same asymmetry that runs through the joints runs through the eyes: the finishing layers get paid per camera whichever sensor brand wins, and the deepest layer shares its supply chain with the datacenter optics build-out. Go down.

A CMOS image sensor leaves the fab unfinished. Someone has to deposit the microscopic color filters and micro-lenses on top of every pixel, and someone has to package the die into a module that survives in the real world. VisEra (TWSE: 6789), a TSMC affiliate, is the specialist foundry for exactly that color-filter and micro-lens step, supporting pixel work from 0.61 microns up to the large pixels machine vision uses. Tong Hsing (TWSE: 6271) is the independent CIS packaging house. Every global-shutter sensor that goes into a robot eye, whoever designed it, is a candidate to pass through this Taiwanese finishing layer. These two are as close as the sensing chain gets to the "every design wins" property of the grinding consumables.

Every depth camera and LiDAR works by firing infrared light and catching it again, and catching faint infrared in a bright room requires a narrow bandpass filter, a coating stack that admits the sensor's exact wavelength and rejects everything else, like noise-cancelling headphones for light. Viavi Solutions (NASDAQ: $VIAV) built this niche, from the Kinect through smartphone face unlock, and markets its filters into robotics and LiDAR directly. Its main challenger is China's Crystal Optech (SZSE: 002273), and the two are in active patent litigation, which tells you the niche is worth fighting over. If Chinese robots keep shipping with LiDAR and structured-light depth while Western robots go vision-only, Crystal Optech rides the higher-filter-count architecture and Viavi rides the Western security-of-supply angle.

Those filter stacks and lens coatings come out of optical coating machines, and Optorun (TSE: 6235) is Japan's listed specialist with a leading share in optical thin-film deposition equipment across Asia. It is the Level 2 machine seller behind every Level 1 lens and filter maker, the same structural position the thread-grinder names occupy in the actuation chain.

Below that sits the glass itself. Robot optics lean toward glass and glass-hybrid lenses over the molded plastic in phones, because robots need optics that hold focus across temperature swings and vibration. Ohara (TSE: 5218) is the listed pure-ish play on optical glass melting, Hoya (TSE: 7741) and AGC (TSE: 5201) carry optical materials inside much larger portfolios, and Corning (NYSE: $GLW) anchors the Western supply. Schott, the other global optical-glass power, sits privately inside the Zeiss foundation, which once again narrows the listed field.

LiDAR emitters and 3D-sensing dot projectors are VCSELs, tiny lasers grown as arrays on compound semiconductor wafers, and the value chain runs down through epitaxy to the substrate. IQE (LSE: IQE) grows epitaxial wafers for the VCSEL makers, Taiwan's VPEC (TWSE: 2455) does the same for a broad laser customer base, and AXT (NASDAQ: $AXTI) supplies the gallium arsenide and indium phosphide substrates underneath. Readers of the photonics series will recognize this layer immediately: it is the same substrate chain, with the same China concentration risk in raw indium phosphide, now with a second demand story attached. The overlap between the photonics build-out and the robot-perception build-out at the substrate layer is one of the quieter compounding stories I am watching.

Three tests, applied to every name in the map.

First, bottleneck severity: does production physically stop without this input, and how few qualified suppliers exist? Thread grinders and superabrasive wheels score highest here, since the machine club is small and Chinese screw capacity depends on it. Optical glass and CIS finishing score well. Commodity bearings and camera modules score lowest.

Second, pricing power and revenue quality: does the name get paid per machine (lumpy), per part (volume-tied), or per consumable (recurring)? The consumable and service layers, wheels, tooling, heat treatment, metrology, carry the revenue quality I want. Machine sellers get the earliest orders but the cyclical exposure that comes with them.

Third, attention: is the humanoid connection already in the story? The Level 1 names that appear in every bank's value-chain report fail this test by definition. Names like Noritake, VisEra, Bodycote, and Ohara pass it, since none of them are framed as robot plays anywhere I can find, which leaves the humanoid connection as pure optionality on top of their existing businesses.

Scored that way, the areas I am prioritizing for the next reports, in order: the grinding chain (machines plus consumables together, since they are one economic system), the CIS finishing and filter layer, the special steel and heat treatment layer, and the optics equipment and glass layer.

Four risks to this framework.

Timeline risk is the big one. Humanoids remain years from true scale, and at 50,000 units a year the Level 3 revenue contribution rounds to nearly zero for most of these companies. What we are really positioning for is the capacity build-out and the eventual re-rating, and both can stall if robot deployment disappoints. The saving grace of going deep is that most of these names run real businesses in autos, aerospace, semis, and general industry. At today's volumes you are paying mostly for those existing businesses, with robot upside layered on top, so you are waiting inside real companies rather than inside stories.

Localization risk cuts at the machine layer specifically. China is working hard to domesticate precision grinding, and there is progress: Shuanglin acquired a domestic grinding-machine maker to cut its equipment costs, and Hiecise's bulk order shows domestic machines winning real business. If Chinese thread grinders reach import parity faster than expected, the Japanese and European machine franchise erodes, though the consumables and metrology layers are stickier.

Cyclicality is structural. Machine tools, special steel, and abrasives are industrial cyclicals, and a broad manufacturing downturn would hit them regardless of what robots do. These names trade with the industrial cycle first and the robot story second, for now.

Access and liquidity are practical constraints. Many of these names trade in Tokyo, Taipei, Shanghai, or Stockholm, and some OTC listings are thin. For readers without international brokerage access, broad vehicles like the KraneShares KOID ETF hold slices of the visible chain, though almost nothing at Level 3, which is rather the point of this series. As always, vet any fund yourself before using it.

This report is the map. The next ones are the digs, in this order:

Report #2 covers the grinding economy: thread and gear grinding machines, superabrasive wheels, and cutting tools as one system, with full company-level work on the four or five names that matter most.

Report #3 covers the sensor finishing chain: CIS post-processing, packaging, filters, and coating equipment, including how the vision-only versus LiDAR split changes each name's math.

Report #4 covers steel, heat, and measurement: the materials and verification layer where the moats are deepest and the waiting happens inside businesses that already work.

Report #5 connects physical AI to the security layer: every robot is an edge device with a machine identity, and the perception stack this report mapped is also an attack surface. This is where the physical AI theme meets the cybersecurity work we have already done.

If humanoids follow anything like the trajectory the industry is ordering capacity for, the companies in this report get paid on the way there, whichever robot wins.

Planetary roller screw — A threaded shaft with rollers that converts a motor's spin into straight-line force; the "muscle" behind a humanoid's hips, knees, and ankles.

Harmonic reducer — A gearbox using a thin flexible ring whose teeth stay pressed against the outer ring, removing the slack between gear teeth; used in wrists and arms where precision matters most.

CIS (CMOS image sensor) — The chip inside every digital camera that converts light into electrical signals.

Global shutter — A sensor design where every pixel captures the same instant, avoiding the motion smear of phone-style rolling shutters; the standard for robot vision.

CBN (cubic boron nitride) — A synthetic crystal nearly as hard as diamond, used in the grinding wheels that finish hardened steel parts.

Superabrasive — The family of diamond and CBN grinding materials consumed by precision manufacturing; wears out with use and gets replaced on a production schedule.

Metrology — The science and equipment of precision measurement; how a parts maker proves its components meet specification.

Bandpass filter — An optical coating stack that lets one narrow slice of light wavelengths through and blocks the rest; essential for depth sensors and LiDAR.

VCSEL — A tiny laser built in arrays on a chip; the light source inside LiDAR and 3D sensing systems.

Epitaxy — Growing ultra-thin crystal layers on a semiconductor wafer; the step where laser chips get their light-emitting structure.

Stay disciplined - Koh

Disclaimer: Nothing in this newsletter constitutes investment advice or a recommendation to buy or sell any security. Numbers and observations are as of publication. I may hold positions in companies discussed above. Always do your own research and consult a licensed financial advisor before making investment decisions.