You want a US-listed, pure humanoid, multi-bagger name — like buying NVIDIA early. Plainly: it's basically an empty set. The hardest chokepoints are almost all private, China/Japan-listed, or big-company internals. The US-buyable pure-plays are pricey and cash-burning: TSLA diluted (Optimus ~14%), Aurora burns most, Pony eats an ADR discount. Manage them as disciplined options, not a core bet.

You may have seen "AgiBot tops 2025 global humanoid shipments, ~39% share" AND "Unitree ships 5,500+, world's largest." Both cite sources. Who's #1? It depends how you count: AgiBot ~5,168 delivered (Omdia #1 by share), Unitree ~5,500 (more by raw units). Change the yardstick, change the champion. Two facts matter more: China holds ~85-90% of global humanoid shipments; and never let a single data source draw your conclusion — always ask the definition.

Want to "buy the leader at the chokepoint" in embodied AI? See one map: the four hardest bottleneck links — dexterous hands, tactile sensing, the robot brain (VLA), harmonic reducers — who leads each, and whether you can buy them. A brutal pattern: the harder the chokepoint, the less buyable. Leaders are almost all private, China/Japan-listed, or big-lab internals. The few US-buyable "real chokepoints" you can count on one hand (MP, NOVT). The hardest things are systematically un-buyable.

The humanoid bull story assumes cost falls all the way down — Musk targets $20K for Optimus. But Morgan Stanley calc'd that "de-China"-ing its supply chain jumps the BOM from ~$46K to ~$131K — nearly 3x, because the actuators, reducers, screws, and magnets that dominate cost are where China leads. The body maker's tradeoff: cheap (lean on China) or secure (swallow 3x BOM). A body can pick only one — which is why it's thinnest-margin.

If part of your reason for owning Tesla is "Optimus," look at Morgan Stanley's $425 SOTP: FSD/network $145 (34%), Robotaxi ~$165 (39%), Optimus $60 (14%, already halved for risk), Auto $55 (13%). The Optimus you wanted to bet on is ~14% — you're actually 73% buying Robotaxi + FSD. You think you're betting A; your exposure is mostly B. That's the most hidden error in a portfolio: thesis and exposure don't match.

Two numbers: Figure closed its Series C at a $39B valuation on a few million dollars of revenue; the same season, John Deere's weeding robots ran across 5M+ acres, charging by the acre and cutting herbicide 59%. One is worth $39B on near-zero revenue; the other counts cash in a field and nobody calls it a robot stock. The most expensive lesson: fame isn't fortune. The real cash flows sit downstream, in names that don't carry "robot."

The entire humanoid bull case rests on one belief: the VLA scaling law holds, like LLMs. That belief drives every valuation. But it has three layers most people collapse into one: L1 baseline (holds; most demos), L2 real-world generalization (emerging), L3 deployable autonomy (blank today). What supports sky-high valuations is L3 — and it's not just missing, its timeline keeps sliding right. The discipline: only pay for L3 money, not L1 demos.

Planning to buy the "robot supply chain" to ride the humanoid wave? The two ends of the smile curve you can buy; the richest middle links you can't. High-end harmonic reducers and planetary roller screws — the highest-value parts in a humanoid's actuators — are carved up by Japan, Europe, and China. US-market exposure is basically zero. The US can build the gears, but not the 40-year moat. Value sits at the two ends: allocate to the ends, avoid the middle.

The LLM playbook gives people the wrong intuition: that crossing from digital to physical is just "scale once more." It isn't. Embodied AI hits four walls of different natures — data (scale can grind it down), safety (liability), reliability (physics), economics (accounting). Only the first yields to scale. Underneath sits Moravec's paradox: what's easy for humans is brutal for machines.

In Sept 2025 Figure closed a $1B Series C at a $39B valuation — on revenue still in the low millions, a price-to-sales in the thousands. That's the embodied-AI paradox: capital pricing runs far ahead of financial delivery. The market isn't paying for cash flow; it's paying for an option on a paradigm. The disciplined fix isn't picking a winning body — it's a barbell at the border of iron and silicon.