AI gets a body: humanoids are already working in factories, but the math doesn't add up yet

THE THESIS. Humanoid robotics is real, and it has already crossed the line separating demos from useful work. But it sits at what investment firm Bessemer Venture Partners calls robotics' 'GPT-2.5 moment': the capability exists and is scaling, yet a wide gap remains to the 99.9% reliability a production line demands. Our thesis: the question is no longer whether humanoids will work, but when they will pay for themselves, who will set the price, and what data will train them. Today the answers are split: the United States leads on the brain; China leads on the body.

WHAT IS ALREADY REAL. Facts first, promises second. Fact: BMW has completed the industry's most serious pilot at its Spartanburg plant (USA) with the Figure 02 robot. The numbers come from BMW itself: 1,250 operating hours, more than 90,000 parts moved, and 10-hour shifts over 10 months placing sheet metal on the welding line. BMW highlights something crucial: motions trained in the lab transferred to stable shift operation faster than expected. Industry outlets add that around 40 units of the new Figure 03 are now deployed there, and BMW starts a second pilot this summer in Leipzig (Germany), with robots in battery assembly. Promise: Tesla. As of July 23, 2026, formal Optimus production at Fremont has not started, per the company's own shareholder letter, which 'anticipates' it later this year. Elon Musk has called it 'the hardest product to scale manufacturing that we've ever made'. The hundreds of Optimus units in Tesla's factories today aren't producing: they are collecting training data. That distance between BMW and Tesla is, in miniature, the distance between demo and industry.

A BRAIN HUNGRY FOR DATA. The big unlock of the past two years isn't mechanical; it's software. So-called robotics foundation models — VLA networks, short for vision-language-action: they see a scene, understand an instruction, and directly generate the robot's movements — have done for machines what GPT did for text. NVIDIA (GR00T), Google DeepMind (Gemini Robotics) and the startup Physical Intelligence (π0.5) compete with different architectures but reach the same conclusion: more data, more dexterity. NVIDIA claims to have found the first 'scaling law' for robotic dexterity: going from 1,000 to 20,000 hours of first-person human video more than doubles task completion rates. The problem is the reservoir. Language models trained on the entire internet; for physical manipulation, no equivalent internet exists. Berkeley roboticist Ken Goldberg speaks of a '100,000-year data gap' between what text models know and what a hand needs to fold fabric or plug in a connector. That is today's true bottleneck. Not the motors: the data.

CHINA SETS THE PRICE. While the West refines brains, China industrializes bodies. Unitree launched its R1 humanoid at $5,900 (now under $5,000) and sells the G1 for $16,000. Its founder explained the key to TIME: they build their own actuators — the robot's 'muscles', which account for 50% to 70% of its cost. UBTech logged more than 13,300 orders for its new U1 on launch day, starting at around $17,700. AgiBot says it has passed 15,000 robots produced, and a factory in Guangdong can now make 10,000 humanoids a year, according to Interesting Engineering. Industry tallies attribute to China roughly 90% of humanoid units shipped worldwide, backed by some $20 billion in public subsidies. Caution: units shipped are not units working; many of those robots end up in labs, trade shows and stage performances. But the underlying signal is unmistakable, and it's the same one we saw with open-weight AI models: China's cost curve is falling faster than any published forecast, and that both democratizes the sector and squeezes it.

THE MATH THAT DOESN'T ADD UP YET. Goldman Sachs calculates that the cost of manufacturing a humanoid fell 40% between 2023 and 2024 alone, far ahead of the 15-20% it expected, and projects 50,000-100,000 units shipped in 2026 and a $38 billion market by 2035. Morgan Stanley sees a $5 trillion opportunity by 2050. These are projections, not facts. The real math is done on the factory floor: a humanoid only competes with labor costs if its reliability, maintenance and useful hours hold up, and independent analyses agree that current pilots run at speeds and reliability levels that classic industrial robotics cleared a decade ago. Figure claims it went from building one robot per day to one per hour at its BotQ plant, on the way to 50,000 units a year; that is a company figure, without independent verification. The honest metric isn't orders or videos: it's the autonomous operating hours a customer is willing to pay for. So far, only BMW has published them.

OUR READ. Why insist on the human form, expensive as it is? Because the world is designed for us: stairs, doors, shelves, tools. The humanoid is the only robot that doesn't require redesigning the factory, and that is why — even though a fixed arm remains cheaper and more reliable for a single task — it is a platform bet, not a niche one. In the short term, honesty: the first jobs exposed will be repetitive physical work in warehouses and logistics, the blue-collar equivalent of the back-office exposure we flagged in our AI-and-jobs series; and a new geopolitical risk emerges, because whoever controls actuators and the supply chain holds leverage comparable to chips — and today that leverage is Chinese. In the long term, the prize justifies patience: tireless arms to care for an aging population, manufacture closer to home and make basic goods cheaper. Physical abundance to complement cognitive abundance. What to watch to avoid buying hype: one, autonomous hours published by customers (BMW, Mercedes), not demos; two, whether NVIDIA's physical-data scaling law holds, because it would compress the whole timeline; three, Unitree's price as the market's floor; four, the closing of the 99.9% reliability gap. Our forecast: humanoids profitable in narrow logistics tasks before 2028; in your home, well after what the marketing promises. Neither imminent job apocalypse nor a robot butler this Christmas: a decade of quiet industrial deployment that, governed well, ends where we always point — more abundance and less drudgery.