In Today's Issue:

🤖 Why the most-hyped humanoids are the least finished

🏭 The maturity test: shipped and deployed versus demoed and announced

🔩 The real bottleneck: models, actuators, or data

📊 Three counts of the funding boom, and the numbers the discourse gets wrong

🧠 Who actually controls the stack, and where it gets decided

Dear Readers,

Watch the promotional videos and you would think the robots have already arrived. A humanoid folds laundry, another sorts packages through the night, a third walks a runway with an unsettlingly human gait. In November 2025, XPENG's IRON moved so smoothly on stage that the audience openly wondered whether a person was hidden inside the suit, and the company's CEO had to open the machine's back and show the internal structure to prove otherwise. That anecdote is the whole field in miniature: the performance is extraordinary, and the performance is the point of confusion. The distance between what a robot does on a stage and what it does on a paying customer's floor has never been wider, or more carefully blurred.

The gap is widening now because the ground under the AI industry is shifting. The previous chapter of this story argued that as frontier intelligence turns into a commodity, value slides downward, toward chips, compute, and the infrastructure the models sit on. If the model is no longer the scarce thing, then the scarce thing becomes whatever stands between a capable model and useful physical work: hands that manipulate, bodies that move, the supply chains that build them, and the data that teaches them. Embodied AI is the name for that frontier, and mid-2026 is the moment it became the loudest arena in the whole AI economy.

The loudness is the problem. Almost every headline number in this field sits on one side of a line that reporting usually smudges: the line between shipped or deployed, meaning real units in production or on customer floors, and demoed or announced, meaning a stage, a video, or a roadmap. Hold that line and the market reorganizes itself; drop it and you end up quoting a valuation as if it were revenue, a state guidance fund as if it were a robotics budget, or a CEO's target as if it were an installed base. So the question this piece holds is not the easy one of who has the flashiest robot. It is sharper: if the contest has moved into the physical world, who is actually winning it, and is it decided on the stage, in the model, or on the factory floor?

All the best,

Kim Isenberg

Start with the roster, sorted by the only axis that matters here, and read it with one rule in mind: the higher a system sits, the more its placement should rest on something an outsider could actually check. The strongest temptation in this field is to rank the systems nobody has independently verified above the ones that have been scrutinized, precisely because scrutiny surfaces the caveats. A clean-looking number is often just an unchecked one.

At the top, in the only tier that needs no one's word, sits Unitree's G1. You can order it. The product page lists a price from 13,500 USD, with a common base configuration around 16,000 USD, roughly 35 kg, and between 23 and 43 joint motors depending on the build (Unitree product page).

This is the single hardware fact in the roster that does not depend on trust: a buyer can place an order today. A first commercial deployment, baggage and cargo handling at Tokyo Haneda with Japan Airlines, and a stated target of 20,000 units in 2026 against 5,500 in 2025 sit one notch lower, because those are manufacturer and partner claims rather than something you can buy off a page.

Below that come the systems reported on industrial floors, but resting entirely on unaudited manufacturer self-report. UBTech says its Walker S2 entered mass production in November 2025, a first batch of several hundred, with orders above 800 million yuan for the Walker series and a named customer list running from BYD and Geely to Foxconn and SF Express (UBTech, 11/17/2025).

UBTech's Walker S2 showing its autonomous battery swap, the kind of clean product video that drives the coverage. (Video: UBTech Robotics)

AgiBot reports building its 10,000th humanoid by March 2026, with a striking acceleration curve: about two years to the first thousand, a year to five thousand, and only three months from five to ten thousand (AgiBot, 03/30/2026). Both stories look like real factory use, and both are, in every particular, the company's own account, unverified by any outside party. The widely repeated claim that AgiBot holds a 39 percent global market share has no cleanly traceable origin at all, and should be treated as an unsourced relay, not a fact.

Then the two Western cases, which look more hedged precisely because someone independent looked. Figure's F.03 is in Hall 52 at BMW's Spartanburg plant on a sequencing task, running the company's Helix vision-language-action model, and Figure's own post calls it the first demonstration of Figure 03 performing a logistics workflow at BMW (Figure AI). Set that against The Information's reporting, and a real but gradual progression appears: in early 2025, two robots were shipped to Spartanburg and practiced on weekends while the line was idle, with a BMW spokesperson calling the picture sobering against the summer-2024 videos, which had been filmed during a plant shutdown (The Information, 03/04/2025). That is genuine progress, and it is not the same thing as a robot doing a paid daily shift unsupervised, and we know that only because an independent outlet checked. The Chinese floor claims have no equivalent check, which is exactly why they should not be read as more advanced.

We’re planning next year’s coverage and building it around you.
Three taps, no typing.