The Robot Industry Is Selling Factories, Not Robots

Between July 21 and July 23, four companies made major humanoid robot announcements, and what they did and did not say about production numbers is the whole story. A British startup called Humanoid raised $152 million at a $1.35 billion valuation. It also revealed a binding deal to place between 1,000 and 2,000 of its robots in the auto-parts giant Schaeffler’s plants by 2032. A startup literally named Foundation, with Eric Trump as chief strategy advisor, said it would open a factory in October capable of building 5,000 robots a year, then a second plant in early 2027 with capacity for 50,000. Samsung folded its robot efforts into a new division reporting straight to the Chief Executive Officer (CEO). And Tesla, heading into its second-quarter report, still had not published a single number for how many Optimus robots it has actually built.

Now the other number. In all of 2025, Tesla, Figure, and Agility Robotics each shipped roughly 150 humanoid robots. Not 150,000. One hundred and fifty, apiece.

Here is the arithmetic those announcements skate past. The entire planet shipped about 13,000 humanoid robots in 2025. Foundation’s second factory, by itself, is designed to build almost four times that many every single year. That gap, between the capacity being announced and the units actually shipped, is the entire story of the humanoid robot business in 2026. The factories are real. The robots are real, and some of them genuinely work. What almost nobody has is a customer with a purchase order, as opposed to a pilot, a subsidy, or a forecast. The industry has learned that it is far easier to sell the promise of a robot army than to sell the robots.

The week the capacity got announced

Start with what was actually promised, because the ambition is not subtle. Foundation, founded in 2024, will build Artificial Intelligence (AI) humanoids it calls Phantom, using Advanced Micro Devices (AMD) Ryzen AI Embedded X100 chips for the brain. Its 50,000-a-year plant is scheduled for early 2027, and it already holds a $24 million contract, not a grant, from the Pentagon for robots designed to breach enemy sites. Its industrial robots lease for about $100,000 a year; the defense versions sell to the government for $300,000 each. Eric Trump’s pitch is not about productivity. It is “We are America First. We have to win this race.”

Humanoid, the newly minted British unicorn, will start putting beta robots into customer sites in the fourth quarter of 2026, with Schaeffler committed to between 1,000 and 2,000 units by 2032 and also writing a check as an investor. Samsung, meanwhile, made itself the largest shareholder in the Korean humanoid developer Rainbow Robotics at the end of 2024 and has now built a whole division around commercializing the technology.

The money follows a clear logic, and it is worth naming who gets paid no matter how the robots perform. Foundation’s Phantom runs on AMD silicon. The more robot brains ship, the more AMD and Nvidia sell, which is why AMD’s push into robotics reads as a second front in its fight with Nvidia over who supplies the picks and shovels. A chipmaker gets paid whether or not any single robot maker survives, and so do the venture funds writing the checks. In Foundation’s case there is a beneficiary in the framing itself: “beat China” is a famously effective way to unlock a government purchase order. None of that requires the robots to be good yet. It requires the story to be good.

How many humanoids actually have jobs?

Here is the reality the announcements talk over. The single most impressive Western deployment on record is a pilot. Figure’s second-generation robots ran inside BMW’s plant in Spartanburg, South Carolina, and, in Figure’s own accounting, contributed to the production of more than 30,000 X3 vehicles over 1,250-plus hours of runtime, targeting better than 99 percent success per shift. That is a genuine achievement, and it is also the ceiling, not the floor. It is one robot line, at one plant, doing one narrow task, measured in hundreds of hours.

Scale out and the picture holds. The research firm Omdia counted more than 13,000 humanoid robots shipped worldwide in 2025, and China builds an estimated 85 percent of the world’s humanoids. China’s Unitree alone shipped over 5,500, the most of any company on earth, against roughly 150 apiece from the American leaders. But Unitree’s units are largely low-cost machines bound for research labs, universities, and entertainment, not factory shifts, which is why volume leadership and useful-work leadership are not the same crown.

And Tesla, the company that put humanoids on magazine covers, is the sharpest illustration of the gap. Going into its second-quarter report, Tesla had still published no Optimus production count at all, and volume production at Fremont had not begun. Independent reporting has put cumulative builds only in the low hundreds, all of them, by Elon Musk’s own description, primarily for learning rather than productive work. Against Musk’s promise of eventually a million a year, that is the capacity-versus-reality gap inside a single company, and the reality behind the BMW numbers is laid out in Figure’s eleven-month BMW exam.

Do the robots actually pay for themselves?

The reason order books stay thin while capacity explodes is not mysterious. It is arithmetic. Take Foundation’s own price: about $100,000 a year to lease an industrial humanoid. What that robot costs per hour of actual work depends entirely on how many hours it runs:

$\text{Cost per productive hour} = \frac{\$100{,}000 \text{ per year}}{H \text{ productive hours}}$

Run it for a single 2,000-hour human shift and the robot costs $50 an hour of work. Run it around the clock across two shifts, near 4,000 hours, and it falls to $25. Those are not absurd numbers against a human wage, but they only hold if the robot actually runs those hours at close to human reliability, with little supervision and almost no downtime. That is precisely the thing the current generation of machines cannot yet do. They handle narrow, repetitive tasks under close watch, which means in practice each robot still shadows a person rather than replacing one. A machine that needs a human babysitter is a second worker, not a substitute, and the payback math stays underwater until the AI gets reliable enough to cut the leash. The physical supply chain has to cooperate too: every one of these robots needs the same rare-earth magnets that are already a choke point for the whole robot industry.

The bet, and the warning already flashing in China

There is a serious case for building the plant before the customers arrive, and it deserves a real answer rather than a wave of the hand. It is a bet on option value. When a robot’s artificial intelligence finally becomes reliable enough to work an unsupervised shift, demand will be measured against every repetitive job on earth, and whoever already owns a 50,000-unit line wins the moment while everyone else is still pouring concrete. Schaeffler committing to as many as 2,000 units is a real buyer putting real money behind that bet. This is not a story about fraud, or about robots that do not work. It is a story about timing.

The trouble for the bet is that the one country already running the experiment at scale is not validating it. In June 2026, Caixin reported that China’s humanoid boom had founders and investors openly worried the sector would collapse into the same self-destructive competition that gutted its electric-vehicle industry. The word they used is “involution,” and it came from inside the industry: an executive at the robot maker AgiBot warned that prices were already falling fastest in the lower-end machines where barriers to entry are low.

The demand side shows the same crack. China builds an estimated 85 percent of the world’s humanoids, and the honest reporting is that finding buyers is the hardest part. A rental market has sprung up around the machines, and CNN found it is mostly exposing how little useful work they can reliably do. Overcapacity, falling prices, and thin real demand are the exact three ingredients that forced brutal consolidation on China’s carmakers. The humanoid sector is assembling them faster, and the West is now racing to pour the same concrete. Option value is real, but so is a glut, and only one of the two already has a dateline.

The number that matters is not on any factory blueprint. It is the count of humanoid robots that show up on a paid invoice for productive work, and in the West that still comes down to a handful of factory pilots, the most advanced of them at a BMW plant in South Carolina. Foundation opens its first plant in October. The question worth putting to its investors, and to everyone else racing to pour the same concrete, is not how many robots the line can build. It is who is waiting to buy them.