Compute (AI SoC)

Compute is the rare component that scales ahead of the build ramp worldwide, there is no global deficit. A humanoid’s “bot brain” is a single fabless artificial-intelligence system-on-chip: Tesla’s Optimus runs on one Tesla SoC reused from its self-driving silicon¹, and the rest of the field runs on NVIDIA’s Jetson Thor, the robotics edge-AI standard². So the model counts one compute SoC per robot, and the numbers here track the robot-relevant slice, the high-end “bot-brain” autonomous-machine SoC tier (Jetson robotics modules, automotive autonomous-driving SoCs like NVIDIA’s Drive Orin, and China’s Horizon Journey), on the order of forty million high-end AI SoCs a year today⁴¹¹, not the roughly one-and-a-quarter-billion smartphone application processors⁶ or the billions-unit broad edge-AI chip base behind them⁵. At one SoC per robot, a worldwide fleet of about forty-seven million humanoids by 2040 needs only about forty-seven million SoCs in total¹², while cumulative global production over the same window delivers more than two billion, leaving worldwide output roughly forty-eight times ahead of need. The durable crossover, the year cumulative production drops below cumulative need, never arrives globally, and that no-deficit result is the robust finding: a semiconductor industry measured in hundreds of millions of capable chips a year simply does not bind at robot volumes. The real exposure is not raw silicon but concentration on advanced-node fabrication: more than nine in ten of the world’s leading-edge (seven-nanometre-and-below) logic chips are made in Taiwan, at TSMC⁷, and US export controls cap China’s access to advanced AI compute⁹. So building from one bloc alone is a different story. China is forced onto trailing domestic nodes and its own Ascend silicon¹⁰, so China-only output falls behind the worldwide fleet around 2036. The United States designs most of the world’s bot-brain SoCs but fabricates about nine-tenths of them in Taiwan⁷; with Intel and TSMC Arizona reshoring leading-edge capacity⁸, US-only output nearly keeps pace and crosses only at the very end of the window, around 2039. Treat compute as a concentration risk, not a volume chokepoint, the danger lives in where the advanced wafers are fabricated, not in whether enough exist.

A robot’s compute SoC is the central “bot brain”, the high-performance artificial-intelligence system-on-chip that runs perception, the world model, motion planning and the control loop for the whole machine. Tesla’s Optimus carries one Tesla SoC, downscaled from the dual-chip self-driving computer and reused from its vehicle silicon stack, exposing extra input/output to drive every sensor and actuator¹; across the rest of the industry the reference part is NVIDIA’s Jetson Thor, a two-thousand-and-seventy-trillion-operation module that has become the de facto robotics edge-AI standard², on a platform that has drawn more than two million developers and over seven thousand Jetson Orin customers since 2014³. The model therefore uses one compute SoC per robot (a value of 0.001 in thousand-unit terms). Crucially this is the high-end robot and autonomous-machine tier, the same class of silicon as automotive autonomous-driving SoCs such as NVIDIA’s Drive Orin at two hundred and fifty-four trillion operations⁴, not the roughly one-and-a-quarter-billion smartphone application processors made each year⁶, and not the billions-unit broad edge-AI chip market of small neural-processing units in cameras and appliances⁵, most of which is nowhere near powerful enough to run a humanoid. The chip is fabless: the robot maker designs or buys the SoC, and a foundry fabricates it.

Because a compute SoC is built into a robot for life, the SoCs the fleet needs are a cumulative stock, the total number of robots ever built worldwide multiplied by one¹, not a per-year flow. Integrating the consensus shipment ramp¹², the worldwide installed fleet reaches about 585,000 robots by 2030, ten million by 2035 and forty-seven million by 2040, requiring roughly the same count of compute SoCs: about six hundred thousand, ten million and forty-seven million respectively. That is a tiny demand by semiconductor standards, the entire cumulative 2040 requirement is less than a single year of high-end AI-SoC output and a rounding error against the broader chip industry. In the chart below, the black line is that worldwide cumulative need, the grey band is cumulative global production, and there is no red wedge: global output stays far above need for the whole horizon.

Production is the only thing that differs between the Global, China and US views. The worldwide humanoid fleet and the SoCs it needs stay identical in all three; only the supply line changes, asking whether one bloc’s fabs alone could equip the whole global fleet. This is the robot-relevant high-end AI-SoC slice, not the broad chip total, and real output follows an S-curve. Global production runs from about forty million high-end robot and autonomous-machine AI SoCs a year today⁴¹¹, NVIDIA’s Jetson and Drive lines, Qualcomm, Tesla’s self-driving silicon, plus China’s Horizon Journey, whose cumulative shipments passed ten million units by August 2025¹¹, toward roughly ninety million by 2030, one hundred and seventy million by 2035 and a modeled two hundred and eighty million by 2040, climbing on humanoid, vehicle-autonomy and edge-AI pull. The edge-AI hardware market alone grew to about twenty-four billion dollars in 2024 and is headed toward nearly eighty-eight billion by 2032⁵; the plateau is set by leading-edge wafer and advanced-packaging build cadence, not by design or raw demand. The post-2030 figures are a modeled extrapolation, a humanoid-and-autonomy step-change above the base rate, not a reported forecast. The supplier base is fabless-design-led but fabrication-concentrated: the chips are designed by NVIDIA, Qualcomm, Tesla, Horizon and others, but the advanced wafers come overwhelmingly from a single place, as of 2021 every one of the world’s sub-ten-nanometre logic chips was fabricated in just Taiwan (ninety-two percent) and South Korea (eight percent)⁷, and Taiwan still held about two-thirds of all advanced-process foundry capacity in 2024⁸. China climbs from a weak base, about one hundred and fifty thousand robot-grade advanced AI SoCs a year today, toward a modeled six million by 2040: it designs competitive parts (Horizon Journey 6 leads China’s assisted-driving market at about forty-six percent among domestic makers¹¹, plus Huawei Ascend and Black Sesame), but US export controls bar it from fabricating advanced AI chips at TSMC and force it onto domestic fabs⁹, where SMIC makes seven-nanometre silicon only through deep-ultraviolet multipatterning at roughly twenty-percent yield and forty-to-fifty-percent higher cost¹⁰. The United States holds far more design strength than fab: it designs most of the world’s bot-brain SoCs but fabricates about nine-tenths of them in Taiwan, so its domestic leading-edge output runs from about a quarter-million SoCs a year today toward a modeled six-and-a-half million by 2040 as Intel’s 18A and TSMC’s Arizona fabs reshore capacity, the US share of advanced-process fabrication is projected to climb from about ten to twenty-two percent by 2027⁸.

For most components this section names the crossover year; for compute there isn’t one to name globally. Cumulative global production runs about forty-eight times ahead of the fleet’s need at 2040, more than two billion SoCs produced against a forty-seven-million requirement¹², so the durable crossover never arrives, the robust finding. One fabless SoC per robot against a semiconductor base of hundreds of millions of capable high-end chips a year is simply too shallow a draw to drain it⁵⁶. Where the picture changes is building from one bloc alone, because the advanced wafers are made in one place. China-only output crosses below the worldwide fleet’s need around 2036: it designs capable AI SoCs but is barred from TSMC for advanced AI chips and forced onto trailing domestic nodes and its own Ascend silicon⁹¹⁰, so even a fast self-sufficiency ramp cannot equip the whole global fleet from China’s constrained advanced-node fabs alone. The United States, the design leader, is the stronger bloc: it designs nearly all the world’s bot-brain SoCs and is reshoring leading-edge fabrication through Intel and TSMC Arizona⁸, so US-only output very nearly keeps pace and crosses only at the very end of the window, around 2039, and only just, among the latest single-bloc crossovers in the whole component set. The contrast with the grinder-bound parts is the whole point: where roller screws or cross-roller bearings cross in the mid-2030s and meet well under half the 2040 need, compute tracks the ramp globally the entire way, and even a single strong bloc nearly does. One honest caveat on those two bloc years: unlike the global no-deficit result, they are the lower-confidence finding, because they depend on the robot-grade carve-out, treating only the high-end advanced-node AI-SoC slice, fabbed domestically under export controls, as each bloc’s binding base⁹, and on modeled per-bloc fab shares. Read the broad chip universe instead and even the single-bloc views stay elastic; it is the advanced-fabrication interpretation that makes a bloc cross. The global-never-binds verdict holds either way.

What makes the precision-mechanical parts tight does not apply here. There is no scarce grinder, no thin skilled-labor base, no multi-year qualification on a single European machine tool, and the design talent is broadly distributed across NVIDIA, Qualcomm, Tesla, Horizon and a dozen others. The compute SoC is fabless: design is abundant and the part is, at robot volumes, trivially small, which is exactly why the global verdict is elastic. The one real constraint is geographic concentration of advanced-node fabrication, and it is about where the wafers are made, not how many. More than nine in ten of the world’s leading-edge logic chips, the seven-nanometre-and-below silicon a humanoid bot brain needs, are fabricated in Taiwan, at TSMC; as recently as 2021 the entire sub-ten-nanometre logic capacity of the planet sat in Taiwan and South Korea alone⁷, and Taiwan still held about two-thirds of all advanced-process foundry capacity in 2024⁸. Layered on top is policy: US export controls prohibit Chinese firms from using TSMC to fabricate advanced AI chips and treat Huawei’s Ascend accelerators as off-limits worldwide⁹, so China is pushed onto domestic fabs where SMIC reaches seven nanometres only through deep-ultraviolet multipatterning at roughly twenty-percent yield and a forty-to-fifty-percent cost penalty¹⁰. So the binding step, in the one scenario where it binds, is not a shortage of chips but the concentration and reshoring of leading-edge fabrication outside Taiwan, a vulnerability the global market masks (TSMC supplies the world fine) but a single decoupled bloc cannot escape: China is fenced out by controls, and even the United States, which designs the silicon, still fabricates most of it in Taiwan and is only now reshoring through Intel’s 18A and TSMC’s Arizona fabs⁸. The verdict, globally elastic, no global crossover, is robust across a wide range of assumptions; it rests on the world keeping access to Taiwan’s leading-edge fabs, which is precisely the concentration the bloc views expose.

Each cobot and each industrial arm ships with one industrial-PC controller, roughly one compute unit per arm, the same order as a humanoid, so both fleets add modestly to the robot-side compute draw. That draw is dwarfed by AI-datacenter demand for the leading-edge silicon and never binds on the robot side. Quadrupeds carry one Jetson Orin-class SoC apiece, the same single-brain draw as a humanoid, so even the fast-growing quadruped fleet of about five million by 2040 adds only a few million more SoCs to a market measured in hundreds of millions. Drones barely register despite being the largest fleet by units: only the autonomous roughly fifteen percent carry an AI system-on-chip companion computer, consumer and racing craft run a commodity flight microcontroller not counted here, so the blended draw is only about fifteen-hundredths of an SoC per unit, and even a cumulative drone fleet of roughly four hundred and seventy-eight million by 2040 (units built, since about half are single-use military first-person-view craft) adds only some seventy million SoCs. The robot-side draw never binds, all five embodiments included.