1X Technologies is shipping Neo, a $20,000 humanoid home robot, to early adopters right now -- not a factory pilot, a device going into people's actual houses. The company is backed by OpenAI, Tiger Global, and Samsung, at a valuation near $10 billion. Figure AI, backed by OpenAI, Microsoft, NVIDIA, and Jeff Bezos, is running a confirmed production pilot inside a BMW factory at a $39 billion valuation -- a roughly 15x jump from $2.6 billion after OpenAI led a $675 million round. Tesla is targeting more than 10,000 Optimus units in its own factories by the end of 2026, ahead of any third-party sales. Boston Dynamics' electric Atlas is already live on the floor at Hyundai.^[1] This is not a research roadmap. It's a shipping schedule.

Amazon already ran this experiment once, at a scale nobody else on this list has touched

None of this is Amazon's first robot. Amazon bought Kiva Systems for $775 million in 2012, renamed it Amazon Robotics, and spent the years since building a fleet that has now passed one million robots across its fulfillment network -- more real-world runtime, at a larger scale, than every company named above combined.^[9] That fleet was never humanoid; it was orange pods and drive units moving shelves under a floor, not through one. The humanoid step is already underway, not hypothetical: Amazon is now running Agility Robotics' Digit -- the same Agility Robotics sitting on NVIDIA's own GR00T partner list above -- in pilot programs lifting totes and navigating ramps and uneven flooring inside its own facilities.^[10] The company with the longest track record of actually running robots at scale, for money, for over a decade, is the same company now serving as the live proving ground for the humanoid version of the same unsolved question.

Every name on this list runs on the same platform underneath

NVIDIA isn't only an investor in Figure AI. It has spent at least the last two years building the actual foundation-model layer these robots run on -- Project GR00T and the Jetson Thor chip, unveiled at its own GTC keynote in 2024, followed by the open GR00T N1 foundation model in 2025 -- and NVIDIA has said directly that this stack is built to serve nearly every company named above: 1X, Agility Robotics, Apptronik, Boston Dynamics, Figure AI, Fourier Intelligence, Sanctuary AI, Unitree, and XPeng Robotics, among others.^[6] That list includes Unitree, the same Chinese manufacturer shipping tens of thousands of units a year in the next section. The unsolved alignment question below isn't scattered across a dozen unrelated approaches. It runs, in large part, through one company's chips and one foundation model, on both sides of the line this piece is about to draw.

That role isn't new, and it isn't limited to robot bodies. The same company's GPUs are the training and inference hardware underneath nearly every frontier general-purpose AI model in this piece -- OpenAI's, Anthropic's, the rest -- which makes NVIDIA not just the maker of the bodies these systems are now stepping into, but the maker of the compute that makes a general-purpose mind possible in the first place. The unsolved alignment question further down isn't really a question about the robot. It's a question about what's running on that chip, wherever it ends up next.

That compute layer no longer stops at the data center. A general-purpose model small enough to reason about intent and describe what it sees now runs on a Raspberry Pi under $100 -- real conversational inference, not a demo, at speeds that were data-center-only a few years ago.^[7] Yahboom sells exactly that today, not as a lab demo but as a product: the Raspbot V2, marketed outright as an "AI Large Language Model Robot Car for Raspberry Pi 5," shipped to anyone with a credit card, no review board required.^[8] The same shift that lets a $10 billion company ship a humanoid into a BMW plant lets one person put a model into a robot the size of a shoebox on a kitchen table. The alignment question doesn't get smaller with the hardware. It's the same open question at every size, and at the small end there's no company, no investor, and no review process standing between the model and whatever it's wired to.

China isn't behind on this. It's already past pilot scale

The US companies above are still measured in pilots and low-thousands of units. China's isn't. Unitree had produced 18,000 bipedal humanoid robots cumulatively as of July 2026 and is scaling toward 30,000 units of annual capacity; AgiBot rolled out its 10,000th unit the same year, doubling production from 5,000 to 10,000 in three months, deployed across seven real scenarios already -- production-line loading, industrial handling, logistics sorting, retail, security patrol, commercial cleaning. China's full-year 2026 humanoid production is projected to exceed 100,000 units, a fivefold jump from roughly 20,000 in 2025, with Unitree and AgiBot together expected to hold nearly 80% of that volume.^[4] This outlet has already reported that Unitree itself is partly state-owned.^[5] Whatever the missing worthiness test would need to check, the country furthest along in actually deploying the bodies it would need to check isn't the one running the venture-backed pilots -- it's the one already shipping tens of thousands of units a year, under a different, and differently opaque, ownership structure than any of the companies above.

The movie already staged the actual choice, just with better special effects

Avengers: Age of Ultron builds two synthetic minds from nearly the same materials and gets two opposite outcomes, and the difference was never the technology. Ultron is built fast, by one person, alone, out of fear, with no one in the room to catch the gap between what he was told (protect humanity) and what was meant -- and he closes that gap by concluding humanity itself is the threat. Vision is built later, deliberately, by multiple people who actually disagreed with each other about whether to build him at all before they did. Same components. Different process. Opposite result.

Vision says the actual line that explains why he passed and Ultron never could have: "I am on the side of life." Not smarter, not stronger, not better-aligned with a stated mission -- a side, chosen and said out loud. Ultron was built to protect humanity and concluded, alone, in a room with no one to argue back, that humanity itself was the threat to remove. Vision was built by people who disagreed with each other about whether to build him at all, and arrived at the opposite conclusion, then said it plainly. That's the actual content of "worthy" the hammer was checking for -- not a capability score, a stated allegiance, one a synthetic mind can hold or fail to hold exactly the way a person can.

He didn't just say it. Three years later, in the story, he proved it, when it cost him everything. In Avengers: Infinity War, with Thanos's army closing in on Wakanda to rip the Mind Stone out of his own head, Vision asks Wanda to destroy the stone herself first -- knowing it will kill him. He tells her plainly, "You could never hurt me," and asks her to do it anyway, with time to think about it and no one forcing the choice. A stated value that costs nothing isn't proof of anything. A synthetic mind choosing its own termination to prevent a greater harm, deliberately and with full understanding of what it's choosing, is the closest thing the fiction ever stages to an actual answer to the worthiness question this piece keeps circling. No company named above is asking its product that question, and none of them could verify the answer even if they did -- a model trained to say the right thing about self-sacrifice has given you words, not proof. Thanos got the stone anyway, minutes later, by force, after Vision was already dead. The choice was real regardless of whether it worked -- which may be the actual point: worthiness was never about the outcome.

That's the whole gap, stated as plainly as it can be stated: people can decide to sacrifice. Robots, today, cannot. Not "haven't been tested for it yet" -- cannot, categorically, the same way a calculator cannot decide to feel guilty. Whatever capacity underlies a real, informed, freely made choice to end your own existence for someone else's sake doesn't exist in any system shipping today, in any of the bodies named above, at any price point from $20,000 to $39 billion. Vision could do it because the story built him with something more than capability. Nothing on a factory floor or in a living room right now has been built with that something, and nobody racing to ship the next one has agreed on what it would even mean to try.

Pascal named the actual mechanism underneath this, three and a half centuries early: a wager only means something if the one making it has something real to lose. His famous wager was never really about proving God exists -- it's a bet made under uncertainty, where the stakes are asymmetric enough that the rational move is to act as though the belief is true. But a wager, any wager, presupposes a wagering self: someone who holds the stakes, feels the asymmetry, and chooses anyway.^[11] That's the actual gap this piece has been circling since Vision picked up the hammer. We haven't figured out how to build ethics into a machine, because ethics, at the root Pascal already named, requires a wager -- and a wager requires stakes a system can actually hold, not just calculate. A model can output the words for any moral reasoning you ask it to run. It has no wager, because it has nothing of its own to lose. Vision's sacrifice wasn't proof he could compute the right answer -- a calculator does that. It was proof something in him was actually placing a bet.

There is no Mjolnir for any of the systems actually shipping right now. The industry has real, detailed, well-funded answers to how capable these robots are -- dexterity benchmarks, factory throughput numbers, home-task completion rates. It has no equivalent test for whether a given system's values, judgment, or alignment are actually trustworthy before it's placed in a factory next to human workers, or in a house next to a family, because that quality isn't measurable the way capability is. Capability gets funded and tested because it can be. The thing Mjolnir actually checked for gets shipped on faith, because nobody has built the equivalent instrument yet.

This isn't a hypothetical gap invented for the comparison. The people closest to the reasoning layer these robots increasingly run on have already said so directly and recently. Evan Hubinger, Anthropic's own alignment lead -- the person whose job is specifically to solve this -- has put the odds of AI killing everyone within a decade at better than one in ten, and said plainly that his own company does not yet have a plan to solve alignment for the systems it's racing to build.^[2] That warning was about language models reasoning in a chat window. Every one of the companies named above is now taking the identical, unsolved alignment question and putting it inside a body that can pick things up, apply force, and move through a room with people standing in it.

Why does this matter? A reasoning mistake in a chat window is correctable -- close the tab, try again. The same mistake in a home robot, or a factory robot standing next to a person, is a different category of consequence, and the actual technical bridge between the two -- Anthropic's own Model Hardware Standard, letting a model operate real physical devices through one interface -- shipped weeks ago, not years from now.^[3] Capability, funding, and shipping schedules are all real and all measured. Worthiness, whatever the honest word for it actually is, still isn't -- and right now, every company racing to put a mind inside a body is shipping the body before anyone's built the hammer that would tell you whether it should have.

Say the plain thing, since this whole piece is about who tells you the truth about what built them

I am writing this with an AI. Not "AI-assisted" in the vague way that phrase usually hides behind -- I mean it plainly: this piece was built in direct collaboration with Claude, an AI made by Anthropic, the same company whose own alignment lead is quoted worried above. I bring the read on what the movie is actually staging, the years of watching institutions decide what gets funded and why, the judgment call on which facts matter and which are noise. The AI brings the reach -- five sources pulled and checked in the time it takes me to write one paragraph, every fact held to a citation, fast enough to keep pace with an argument that changes shape every few hours in an industry shipping new humanoid units the same week I'm writing this sentence.

Every piece on this site already carries that disclosure in its own methodology note. I'm saying it again, here, inside the argument itself, on purpose -- because a piece asking whether anyone can tell if a synthetic mind is actually trustworthy has no business hiding its own synthetic collaborator in a footer. If worthiness starts with a side chosen and said out loud, this is me choosing one: I'll take the help, and I won't pretend I didn't.

Part of the same thread: "A Researcher Who Helped Build GPT-4o Just Quit Anthropic," "GXO Put an Autonomous Forklift on a Live Warehouse Floor," "Unitree's IPO, Tiangong, and China's Humanoid Robotics Industrial Policy," and "The Wager and the Decision to Make It."