Lidar & Radar

Lidar is the rare link here that is both minor and elastic, on raw volume it does not bind, and its only real exposure is who makes it. It is minor because most humanoids do not carry one: Tesla’s Optimus and XPeng’s Iron are camera-only by design¹, while Unitree and AGIBot fit about one and a few designs carry two, so the build-weighted average is only about four-tenths of a lidar per robot. That makes the worldwide humanoid fleet’s² entire cumulative need just about nineteen million lidar units by 2040, a small new slice of an automotive-lidar industry that already shipped about three-and-seven-tenths million units in 2025, up sixty percent in a single year⁴, and is ramping toward tens of millions a year at roughly thirty-five percent annual growth⁵. Run the four-series gap and the volume crossover simply never arrives: cumulative global production, and China’s alone, both stay far ahead of the fleet across the whole horizon (the global and China crossover years are both null²), and by 2040 worldwide output has met the fleet’s need many times over. The bind is a different kind, and a softer one than magnets: concentration. Chinese suppliers hold roughly ninety to ninety-five percent of automotive-lidar units⁴⁸, Hesai alone about forty-three percent of the ADAS market, RoboSense about a quarter⁴³, so the supply base is overwhelmingly China-led. The one place the gap opens is the thin US/non-China bloc (Luminar, Ouster and a handful of others): even dedicating it entirely to the worldwide fleet, its cumulative output softly crosses the need only in the mid-2030s, around 2035¹⁰, and that bloc is fragile, with Luminar’s Volvo program terminated in late 2025 and its assets sold off¹⁰. Treat lidar as a concentration footnote, not a volume chokepoint: enough sensors will exist, but China makes almost all of them.

A lidar sends out laser pulses and times their return to build a direct, metric three-dimensional depth map of the scene, range, shape and motion measured rather than inferred, which is why mobile robots and self-driving cars use it for navigation and obstacle safety. But for humanoids it is optional and contested. The dominant designs are camera-first: Tesla’s Optimus carries eight autopilot-grade cameras and no lidar at all, on the explicit view that cameras are roughly a hundred times cheaper¹, and XPeng’s Iron follows the same vision-only philosophy; others such as Unitree and AGIBot fit about one lidar for navigation, and a few carry two. Because the volume leaders use zero, the model uses a build-weighted average of about four-tenths of a lidar per robot (a value of 0.0004 in thousand-unit terms)¹, deliberately a low fraction, and the single most uncertain assumption on this page, because if vision-only wins outright the per-robot count drifts toward zero and lidar all but disappears from the bill of materials. The hardware itself is a solid-state or mechanically-scanned assembly of a laser diode, a single-photon avalanche-diode (SPAD) detector and precision optics; automotive radar, the cheaper commodity cousin, is a different, fully elastic story.

Because a lidar is built into a robot for life, it does not wear out and there is no replacement tail, the units the fleet needs are a cumulative stock: the total number of robots ever built worldwide multiplied by about four-tenths each¹, not a per-year flow. Integrating the consensus shipment ramp², the worldwide installed fleet reaches about 585,000 robots by 2030, ten million by 2035 and forty-seven million by 2040, requiring only about two hundred and thirty thousand, four million, and nineteen million lidar units respectively. That is a strikingly small demand: nineteen million units cumulative across fifteen years is less than a few years of current automotive-lidar shipments. In the chart below the black line is that worldwide cumulative need, the grey band is cumulative global production, and the gap never opens, because lidar output dwarfs the small robot slice. That absence of a wedge is the finding, not an error.

Production is the only thing that differs between the Global, China and US views. The worldwide humanoid fleet and the lidar units it needs stay identical in all three; only the supply line changes, asking whether one bloc’s factories alone could equip the whole global fleet. Global automotive-lidar output runs from about three-and-seven-tenths million units a year in 2025⁴ toward a modeled twelve million by 2030, twenty-eight million by 2035 and forty-two million by 2040, a steep S-curve as advanced-driver-assistance adoption and robotics pull draw investment, the market growing at roughly thirty-five percent a year and crossing a billion dollars for the first time in 2025⁵. The post-2030 figures are a modeled extrapolation of that ramp, not a reported forecast. China is almost all of it, Chinese suppliers hold about ninety to ninety-five percent of automotive-lidar units⁴⁸, led by Hesai (about forty-three percent of the ADAS market, and wildly elastic: it shipped about half a million units in 2024, one-and-six-tenths million in 2025, a two-hundred-and-twenty-three-percent jump, and is guiding three to three-and-a-half million for 2026⁶), RoboSense (the first to a million cumulative deliveries, up from a quarter-million in 2023 to over half a million in 2024⁷) and Huawei³, so the China curve tracks the global one from about three-and-three-tenths million units today. The US/non-China bloc is the thin line: about thirty thousand units a year today, led by Ouster (roughly seventeen thousand sensors in 2024, a record seven thousand two hundred in a single quarter of 2025) and Luminar, with Aeva and Cepton smaller and Innoviz in Israel¹⁰. It is not only thin but fragile, Volvo terminated its Luminar program in late 2025 and MicroVision bought Luminar’s assets for thirty-three million dollars¹⁰, and most US-headquartered brands contract-manufacture in Asia anyway, leaving only Ouster’s small San Francisco “Buy American” line genuinely US-made. Modeled to recover and ramp slowly, the US bloc still reaches only about one-and-one-tenth million units a year by 2040.

For most components this section names the crossover year; for lidar there isn’t one to name globally. Cumulative global production runs far ahead of the fleet’s need across the whole horizon, hundreds of millions of units produced against a cumulative requirement of only about nineteen million by 2040², so the durable crossover never arrives, and the same is true of China alone: its output is so large and so elastic that it never falls behind the worldwide fleet⁴. Both crossover years are legitimately null, and that is the point, lidar is a minor, optional component whose demand is a rounding error against an auto-lidar industry shipping millions of units a year. The wall appears only in the US/non-China bloc view, and even then it is a soft one: if the West had to supply the entire worldwide fleet from that thin slice alone, cumulative US output keeps pace only into the mid-2030s, crossing around 2035, when the fleet’s slowly-accumulating need finally outruns a bloc that tops out near one million units a year¹⁰. But read this honestly, it is a concentration finding, not a volume wall. The reason the US bloc falls behind is not that lidar is hard to make; it is that almost all lidar is made in China⁸, and the small Western alternative is fragile and contracting¹⁰. The crossover year here is the lowest-confidence number on the page: it rests on the contested four-tenths-per-robot assumption, and if vision-only designs prevail it never crosses at all. The robust verdict is the one above, globally elastic, no volume bind, with concentration, not capacity, as the real exposure.

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. Lidar is a fast-moving semiconductor-and-optics product: a solid-state assembly of a laser diode, a SPAD detector and precision optics whose cost is falling steeply, Chinese average selling prices have dropped to four-to-five hundred dollars with next-generation units targeting under two hundred⁵, and capacity is wildly elastic, as Hesai’s leap from half a million to over one-and-a-half million units in one year shows⁶. So the global verdict is robustly elastic. The exposure that remains is not volume but concentration, and it is geographic: Chinese suppliers hold roughly ninety to ninety-five percent of automotive-lidar units⁴⁸, Hesai, RoboSense and Huawei between them, with the Western bloc reduced to a thin, fragile tail of Luminar, Ouster and a few others¹⁰. That concentration becomes a hard constraint only where decoupling forces it to: the United States’ FY2025 National Defense Authorization Act, section 164, bars the Department of Defense from procuring or operating China-made lidar and names Hesai expressly⁹, which pushes US and defense robots toward either vision-only designs or the scarce, expensive non-China supply⁹. That is the whole bind in one line: there is no shortage of lidar in the world, only a shortage of non-Chinese lidar, a concentration and policy exposure of exactly the kind that shapes the magnets story, not a precision-process volume ceiling. For a humanoid program that can go vision-only, lidar is barely a constraint at all; for one that cannot, the constraint is which flag is on the sensor, not whether enough exist. Quadrupeds tell the same story: only the EDU and premium tiers carry a lidar, about three-tenths of a unit per robot on a build-weighted average, with the cheap consumer units none, so even a quadruped fleet of about 5.15 million by 2040 adds under two million units, a commodity draw dwarfed by the same China-led supply base. Drones draw less again: only high-end survey and mapping craft carry a lidar, about one-hundredth of a unit per drone on a build-weighted average, so even a cumulative drone fleet of roughly four hundred and seventy-eight million by 2040 (units built, since about half are single-use military craft) adds only a few million units, a fractional draw on that same base.