Every major forecaster, including BofA¹, Goldman Sachs², Morgan Stanley³, Citi⁴, DBS⁵, UBS⁶, Macquarie⁷, and IDC⁸, models humanoid production reaching millions of units annually by the late 2030s. In addition, Citi projects a global fleet of 1.2 billion by 2050⁴. We ask the physical question: can the supply chain actually build it? By tracking production against demand across twenty-five critical components, we pinpoint the exact crossover year where manufacturing capacity fails to keep pace. The results reveal three tiers. The binding spine is restricted by precision processes; regional single-bloc chokepoints fail when supply chains decouple; and the commodity-elastic majority scales effortlessly.

For each component, the model tracks four metrics: (1) the active global robot fleet, (2) the total parts required to equip it, (3) regional factory production rates, and (4) the cumulative gap between supply and demand. This approach avoids complex economic assumptions, price elasticities, and substitution rates in favor of transparent, first-principles arithmetic.

We begin with the humanoid, which represents the deepest demand draw on the supply chain. In the diagram below, each body region is colored by its component supply tier. The binding precision spine (covering reducers, screws, bearings, hands, and force/torque sensors) runs through almost every joint. Hover over or tap any region to inspect its components and view the exact year each capacity limit is crossed.

While the model tracks the humanoid fleet in full depth, humanoids are not the only Physical AI systems drawing on these components. The model treats each form factor as a distinct embodiment with its own demand trajectory and bill of materials. Collaborative robots, or cobots, are the first addition. A six-axis cobot arm is built entirely around precision motion, carrying six harmonic drives, six crossed-roller bearings, twelve encoders, and six frameless motors¹¹¹², but it uses none of the humanoid-specific linear or hand parts. The world installed ~64,500 cobots in 2024⁹, and shipments are projected to grow at approximately 17% annually through 2030¹⁰. Because cobots ship in volume today, they front-load the demand for reducers, bearings, and encoders, pulling their crossover years earlier. By design, every supply curve credits 100% of world output to robots and never nets out the machine-tool or general-industry draw on the same parts, a deliberate upper bound: when a component still falls behind, it does so even under the most generous supply assumption. Cobots and industrial arms are already embedded in the “industrial robot” share of these precision-part markets (~20% of harmonic units)¹⁵, so they claim that existing pool rather than adding to it, which only pulls the wall in.

Outside of collaborative applications, the traditional industrial arm remains the manufacturing workhorse, representing major brands and accounting for about 88% of the ~540,000 annual industrial robot installations⁹. While its architecture resembles a cobot, it differs in two key ways. First, instead of being entirely harmonic-driven, it uses about three cycloidal or RV reducers for its high-torque base joints²⁴ and three harmonic drives for its wrist²⁵. Second, it relies on framed AC servo motors with separate gearboxes rather than the frameless torque motors found in cobots and humanoids. As a result, industrial arms draw from the shared rare-earth magnet pool but not from the frameless-motor manufacturing lines. These arms represent a mature base load of roughly 260,000 units per year with a modest 5% growth rate²⁶. This volume establishes cycloidal/RV reducers as a significant demand category and shifts the crossover point for shared harmonic drives slightly earlier. By 2040, the global fleet will include 6 million industrial arms alongside the humanoid, cobot, and quadruped fleets.

Moving from factory floors to unstructured environments, the quadruped stands as the closest architectural cousin to the humanoid. These four-legged systems, such as the Unitree Go2, Boston Dynamics Spot, or ANYbotics ANYmal, feature about twelve quasi-direct-drive (QDD) joints, with three per leg. Where other robots rely on harmonic or cycloidal drives, a quadruped is built around low-ratio planetary reducers paired with large-diameter frameless BLDC motors³⁰. Each unit requires about eleven planetary reducers, twelve frameless motors, twelve to twenty-four encoders, and one IMU for balance. It requires none of the humanoid-specific parts like roller screws, tactile skin, or dexterous hands. This makes quadrupeds the first embodiment to draw heavily on planetary reducers and the first non-humanoid user of IMUs. The quadruped fleet is a fast-growing, mid-volume segment. Production reached about 30,000 units in 2025, which is already above humanoids, and is projected to reach a few hundred thousand units annually by 2040³¹. While these robots consume shared NdFeB magnet feedstock and represent the largest non-humanoid encoder draw, their demand leaves the grinder-limited precision spine untouched.

Representing the capstone of this analysis, the drone is by far the largest Physical AI category by volume. The world currently produces between 15 and 18 million drones annually across consumer, commercial, and military segments⁴⁸. This output is projected to grow to approximately 50 million units per year by 2040, which is five times the forecasted annual humanoid build. Despite this massive volume, a drone is a direct-drive propeller machine. Its outrunner motors couple straight to the propellers, meaning there is no gearbox anywhere in the system. Consequently, drones draw nothing from the precision-motion spine, requiring no reducers, roller screws, crossed-roller bearings, absolute encoders, or force-torque sensors. Their outrunner motors represent a distinct commodity class that does not share a supplier base with robot-grade frameless motors⁴⁹. Drones consume shared rare-earth magnet feedstock and draw heavily on commodity-elastic components like batteries, cameras, IMUs, and power semiconductors. As a caveat, many military drones are single-use, so this model tracks cumulative units built rather than a standing fleet. The central lesson is clear: adding the highest-volume embodiment in the world does not shift a single crossover year because raw volume is not what binds. The true bottleneck is the precision manufacturing required for legged and armed robots.

When mapped as a grid, the contrast is clear. The humanoid column is dense with red binders while the drone column touches none of them. In this matrix, cell density indicates how many units of a component a single robot carries, and the colored dot represents the part's supply chain tier.

The baseline demand curve is constructed from the geometric mean of eight global shipment forecasts, including BofA¹, Goldman Sachs², Morgan Stanley³, Citi⁴, DBS⁵, UBS⁶, Macquarie⁷, and IDC⁸. This consensus model anchors at approximately 18,000 robots in 2025, reaching 250,000 annually by 2030^12345678, 4.4 million by 2035, and 10 million by 2040. These figures are forward-looking projections based on the stated methodology and the cited third-party forecasts, and actual outcomes may differ. These projections span a wide range. Goldman Sachs represents the conservative edge with 1.38 million units in 2035², while BofA models a rapid acceleration to 10 million units by that same year¹.

To understand the physical bottlenecks, we convert these annual shipment numbers into the cumulative installed base. This allows us to focus on the cumulative count of robots in operation requiring parts rather than the annual flow. Under this approach, the active global fleet reaches approximately 585,000 in 2030, 10 million by 2035, and 47 million by 2040, which is the horizon of this analysis; values are not extrapolated beyond 2040, where reliable forecaster coverage ends.

When broken down by embodiment, the mature base of traditional industrial robots and early-scaling quadrupeds dominates the fleet through the early 2030s⁹. Drones exceed all other categories by raw units, with 16 million built in 2025 alone. However, because drones consume almost none of the precision components, they are best evaluated independently. The combined Physical AI fleet ramp is detailed below.

A component's cumulative demand is determined by multiplying the fleet size by its per-robot requirement. These parts are split into durables, which last for the lifetime of the robot, and consumables, which wear out and must be replaced on a regular schedule. For example, a planetary roller screw is a durable part purchased only once per robot. In contrast, a battery is a consumable part that requires periodic replacement to keep the robot operational.

This model tracks four consumable components: hand tactile sensors and hand tendon drives in the binding spine, foot contact sensors, and batteries. All other components are durables. This distinction is critical. Because consumables like tendon drives must be replaced periodically, their ongoing replacement demand eventually exceeds the initial installation demand, which pulls their capacity crossovers years earlier. Under this cumulative approach, the 47-million-robot fleet of 2040 requires approximately 133 million roller screws in total, a fleet average near three per robot rather than the fourteen a pure work-humanoid would carry.

While fleet size and component demand are uniform worldwide, regional production capacity varies by location. We project factory capacity using S-curves instead of static compound growth rates. This matches the physical reality of manufacturing, where output increases in step-functions as new facilities open and plateaus as lines reach capacity. S-curves model events like the sudden expansion of Chinese roller screw capacity between 2026 and 2028⁶³⁶⁴ on top of today's 2.25 million annual base⁶².

We categorize supply into low-volume, robot-grade precision components and high-volume commodity components. This structural split is the reason commodity items never limit growth. The global supply chain already produces commodity parts by the billion for other markets, meaning robotic demand is easily absorbed. The resulting supply curves represent an optimistic view of manufacturing limits.

We define the crossover year as the point when cumulative production fails to meet the needs of the growing fleet. In the early stages, factories produce parts faster than robots are built, creating an inventory buffer. As deployment accelerates, cumulative demand outruns factory capacity and depletes this inventory, leading to a permanent supply deficit.

Each component's year-by-year table makes this explicit. The Gap column is cumulative production minus cumulative fleet need in that year: a positive value means production is keeping pace and a negative value marks a deficit, and the Status column reads "ok" until the first year the gap turns negative, which is the crossover year. Because production and need both accumulate from 2025, a component can run a large surplus for years and still cross once the fleet's durable, compounding demand overtakes a slower-growing supply curve.

Where the report cites a "decoupled" single-bloc crossover, for example a United States-only or China-only supply chain, it re-runs this identical calculation using only that bloc's production measured against the same unchanged worldwide fleet need. That single-bloc series is distinct from the global table shown alongside the text: because one bloc's output is a fraction of global output, its decoupled crossover generally lands earlier than the global one, while for volume-ample components it can land later or not occur within the horizon at all.

For each component, we document regional production forecasts alongside year-by-year calculations of the supply gap. The Master Table centralizes these metrics, summarizing every crossover and showing how much of the 2040 demand each component satisfies.

Before examining the individual components, we must address a critical constraint. Three of the most severely restricted components, which are harmonic-drive flexsplines, roller screw threads, and cross-roller bearing raceways, all rely on a single type of scarce machinery: the ultra-precision grinder. This machinery is controlled by a narrow oligopoly of manufacturers, including Reishauer, Studer, Kapp, and Matrix⁶⁵. The global market for these specialized grinders is extremely small, with a total value of only 251 million dollars in 2024⁶⁶. This capacity translates to only a few hundred new machines each year, while Chinese manufacturers remain less than 50% localized on these high-end tooling lines⁶⁴. Grinder throughput, rather than raw materials, dictates the production ceiling for all three components.

Because these grinding machines are scarce and slow to build, we apply a throttle factor to our capacity forecasts. This factor caps effective output at roughly a third below announced factory capacity by 2030, reflecting the physical unavailability of new machinery. Actuator design changes cannot bypass this constraint. They only shift the allocation of grinding capacity. Swapping a linear joint for a rotary one shifts grinding requirements from roller-screw threads to harmonic flexsplines and bearing raceways, but the overall machinery bottleneck remains. The pool capacity, combined draw, and throttle factor by year:

Evaluating all twenty-five components side-by-side reveals the exact timelines where production capacity falls short of cumulative demand. These crossover years are calculated using a combined Physical AI scope that aggregates the demand of humanoids, cobots, industrial arms, quadrupeds, and drones. When these separate fleets are layered together, their combined draw pulls the shared-reducer crossover years one to two years earlier than a humanoid-only model would project. For instance, the global capacity deficit for force/torque sensors shifts from 2031 to 2029, harmonic drives move from 2033 to 2031, and absolute encoders advance from 2039 to 2038.

This holistic view also highlights the unequal demands of different robot form factors. Quadrupeds represent the first meaningful shared draw on planetary reducers, although the vast global gear-cutting base prevents this link from experiencing a global crossover. Meanwhile, the drone fleet consumes only commodity-elastic components, which explains why the highest-volume robot category has no impact on any of the bottleneck timelines.

The following data organizes these components by the severity of their global crossovers. The earliest and deepest deficits appear at the top, followed by the components that never experience a global capacity constraint. The subsequent sections of the report trace each of these components in detail across three distinct tiers: the binding precision spine, the regional chokepoints, and the commodity-elastic majority.

Geopolitical shifts are driving concrete legislative actions to restrict foreign-made robotics. In June 2026, the bipartisan GUARD Act (H.R. 9129) was introduced in the United States Congress, directing national security agencies to review Chinese-manufactured humanoid and quadruped robots and their control software for inclusion on the FCC covered list¹³⁹. That campaign has since hardened into action: on 28 July 2026 the FCC moved to bar Chinese-made humanoid and quadruped robots, along with power inverters, from the US market, turning the covered-list threat into an enacted import restriction¹⁴⁸. Although it is a proposed bill, not law, and as written targets finished robots rather than the parts inside them, and does not reach allies, it follows a broader regulatory trajectory. The Commerce Department's 2025 connected-vehicle rule already bans Chinese and Russian hardware and software within finished automotive products, signaling that similar component-level restrictions in robotics are a plausible next step rather than a certain one¹⁴⁰.

Implementing these trade restrictions would immediately accelerate supply chain deficits. Removing Chinese manufacturing does not create new chokepoints, but it pulls existing global bottlenecks forward. Relying entirely on domestic United States production would drag eighteen of the twenty-five key components into immediate supply deficits, compared to only ten under a globalized model. If that step comes, the achievable build rate is exactly what the Global ex-China and US-built columns show.

Ramping up domestic and allied manufacturing capacity would need to begin well before fleet demand peaks. Because the precision-motion spine, which comprises specialized reducers, screws, bearings, and force-torque sensors, falls into deficit years before fleet demand peaks, the lead time to add capacity is long. Standing up precision gear-grinding machinery and multi-axis sensor calibration facilities requires years of preparation, so localized component production would need to expand years ahead of demand to keep pace.

Quantifying the timeline required to establish manufacturing capacity reveals the steep hurdle facing the humanoid supply chain. Sourced production projections assume highly aggressive growth trajectories over the next fifteen years, with roller-screw output rising five-fold, harmonic drives eight-fold, encoders fifteen-fold, and force-torque sensors scaling seventy-fold. These targets require sustained annual growth rates between 7% and 45%. Eight of the ten components stay within their historical maximum growth rate, but for six of the ten the modeled output runs past their total-expansion ceilings, meaning a stricter limit would only make the shortfalls deeper. By 2040, harmonic drives and force-torque sensors are the deepest deficits, each falling roughly six-fold short of cumulative fleet needs, with cross-roller bearings about three-fold and tendon and ball screws about twofold; roller screws, once the deepest deficit, fall only about 1.2-fold short after correcting for the rotary-heavy shipped fleet.

Two specific constraints make these supply shortfalls an immediate concern rather than a distant threat. First, force-torque sensors present a rigid expansion limit. Their 20% annual growth rate is locked by the multi-axis metrology and calibration bottleneck, a process that requires physical rig time and cannot be easily accelerated by capital injection. Second, the machine-tool base that supports the gear-grinding spine scales far slower than the components relying on it. The global precision grinding machinery market expands only about 4.4 percent a year¹⁴¹, whereas roller screws, harmonic drives, and crossed-roller bearings require sustained growth rates of 8% to 10%.

Lead times for expanding manufacturing facilities further delay supply response. Expanding an existing, qualified precision production line requires one to one and a half years of lead time. Building and qualifying a new facility from the ground up takes three to four years. Because these timelines are fixed by physical tool installation and qualification processes, developers cannot rely on rapid capacity spikes to resolve deficits.