A machine that can write its own code is, by construction, a machine that has outgrown the vocabulary of the people obliged to certify it. That asymmetry โ between what a system is capable of and what its operators can actually attest to โ is the quiet subject of Wang Xingxing's address at the World Physical AI Conference, and it is the reason a robotics keynote belongs in a publication about ledgers. The speech was not, in the end, about robots. It was about evidence, and about who is permitted to manufacture it.
The address was staged at a gathering co-hosted with Unitree's partner ecosystem summit, a structural detail that reframes everything that followed. What circulated afterwards as a technology keynote was, in its architecture, an investor briefing wearing a technical register: three claims about capability, one claim about timeline, and a single sentence about difficulty that has since been quoted everywhere and examined nowhere. I have spent seventeen years reading documents like this one โ first as a junior analyst auditing settlement messages in Geneva, later as a researcher tracking how capital narrates its own arrival โ and the tell is always the same. When the information density is low and the confidence is high, the document is doing work that has nothing to do with its stated subject.
By my count the address yielded seven discrete information points. Six were opinions. One was an identity claim. There were no product parameters, no benchmark results, no third-party measurements, no financial disclosure, and not a single named competitor. The materials describe the speaker as chairman, general manager and chief technology officer โ a title stack that does not reconcile with the founder-and-chief-executive formulation in every public record I can verify, and the English rendering of the company name itself appears to be a transliteration variant rather than a registered entity. None of this makes the speech unimportant. It makes it diagnostic.
The reflex in the token market was immediate and, in my assessment, aimed at the wrong layer. Anything with the words physical and artificial intelligence in its description caught a bid, as though a humanoid's ability to reconcile its joints with a language model's intent were a validation of decentralized compute, of tokenized data cooperatives, of machine-to-machine payment rails. It is not. What the market heard was resonance, not substance โ the hollow resonance of digital ownership applied to a domain that has not yet been tokenized and probably should not be.
Unitree's operational position is not in dispute. Its quadruped platforms ship at scale into research, education, power-line and utility-tunnel inspection, and security patrol โ unglamorous, high-frequency, reliability-tolerant deployments where the buyer cares about uptime rather than autonomy. Its humanoid line has been priced aggressively against overseas equivalents, a deliberate trade of margin for installed base. The company vertically integrates its own motors and joint modules, which is the least discussed and most consequential fact about it: vertical integration in precision actuation is what converts a robotics firm from an assembler into a cost-setter, and cost-setting is what wins a price war that nobody in the sector wishes to admit is already underway.
Unitree entered its listing guidance period in 2025 with an eye on the STAR Market, which places every public statement it makes inside a valuation conversation. The arithmetic of that conversation is unforgiving. A hardware manufacturer with a genuine installed base and genuine gross margin is ordinarily valued on price-to-sales multiples in the high single digits to low teens. An artificial intelligence platform company โ one that owns a model, a data flywheel, and an ecosystem of integrators โ can command multiples two to three times higher. The distance between those two numbers is not bridged by technology. It is bridged by narrative, and narrative is precisely what a partner ecosystem conference is designed to manufacture.
This is where the crypto reader should stop skimming. Physical AI is being marketed into our sector as the next frontier for decentralized infrastructure โ for compute markets, for data networks, for settlement between machines that will allegedly pay one another. If that thesis is correct, then the terms on which embodied data is produced, verified and priced will determine which of those networks have a business and which have only a token. The Unitree address is the clearest available statement of how the largest hardware player intends to define those terms, and it defines them in a way that quietly excludes everyone currently raising money against them.
I came to this question from the settlement side. In 2017 I ran a six-month audit of legacy messaging protocols against early on-chain settlement layers for a fintech in Geneva, and I spent part of it interviewing forty migrant workers in Zurich. Thirty-five percent of their transferred value was disappearing into intermediary fees they could not see, could not contest, and had never been shown. That experience taught me two things I have never been able to unlearn. Friction is rarely an accident of technology; it is usually a business model. And any system claiming to remove friction must be able to show its work.
In the same city last year I facilitated a roundtable between European regulators and developers building AI-adjacent crypto infrastructure, working through how decentralized compute markets might satisfy the transparency obligations of the EU AI Act. The finding that stayed with me was not the one I expected. Roughly seventy percent of the training data behind the systems under discussion had no verifiable provenance โ no attributable source, no chain of custody, no way for a supervisor to reconstruct how a given capability had been acquired. Zero-knowledge attestations were the only mechanism on the table capable of closing that gap without forcing disclosure of the underlying data itself, and even that mechanism was being discussed as an afterthought by people who had built their entire thesis around compute.
Hold that number beside the Unitree speech and the shape of the real problem becomes visible. The company's central claim โ that the core challenge of embodied intelligence is the alignment of model inputs and outputs with real machines โ is directionally defensible and analytically evasive. It is a true sentence arranged to do a false job. What follows is an attempt to separate the two.
Begin with the word. In the language-model literature, alignment denotes the conformity of a system's behavior with human values and intentions, pursued through preference optimization and reinforcement. In the robotics literature the same English word is used for something entirely different: embodiment alignment, the mapping between abstract representation and physical actuation. The two are not cousins. They are strangers sharing a coat. Embodiment alignment is itself not one problem but at least five โ sensor calibration and noise characterization; the representation gap between vision, language and action spaces, including the tokenization of continuous motor commands; temporal alignment between inference and control; the simulation-to-reality gap; and the semantic translation from an abstract instruction into a continuous action sequence.
The temporal problem alone should end the conversation about imminence. A large model's inference latency sits somewhere between one hundred milliseconds and a full second. A real-time control loop for a legged or humanoid platform runs at one to ten milliseconds. That is a gap of two to three orders of magnitude, and it is not a gap that a better model closes. It is a gap that architecture closes, by decomposing the system into a slow deliberative layer and a fast reflexive one. Anyone who describes alignment as the single remaining obstacle is describing an architecture they have already chosen, not a problem they have already measured.
Alignment is not a bottleneck; it is a bundle of five bottlenecks wearing one name, and the name was chosen because it is the only one of the five that a hardware company does not have to solve alone. Sensor calibration is a manufacturing problem. Temporal decomposition is a control-engineering problem. Simulation fidelity is a tooling problem. Each can be distributed to partners, to suppliers, to an ecosystem of integrators who will absorb the cost of integration and call it innovation. Data scale cannot be distributed. Data scale is a bill.
Which brings us to the claim that models have achieved self-evolution, that they can write their own programs and collect their own data. The known mechanisms for model self-improvement โ self-play, self-rewarding generation, evolutionary program search, active learning with automated labeling โ have all produced credible demonstrations inside bounded task domains. Extending them to a physical agent changes the risk profile rather than the principle. Autonomous data collection in the physical world means a model exploring environments it was not trained on, which demands a safety envelope, an exploration policy, and a sample-efficiency regime that no one has demonstrated at production scale. Real-machine sampling costs five to seven orders of magnitude more per sample than text. The claim, as delivered, was a second-hand summary from a speaker who is not an insider at any frontier model laboratory, offered without a single accompanying metric.
If the constraints were ranked by severity rather than by rhetorical convenience, the order would read differently. Public robotic manipulation datasets contain on the order of ten to the fifth, perhaps ten to the sixth, trajectories; frontier language corpora contain on the order of ten to the thirteenth tokens. Seven orders of magnitude is not a gap that alignment work closes. Long-tail generalization in open environments is combinatorial and unresolved โ the object, illumination and deformation combinations multiply faster than any data collection campaign can chase them. Industrial reliability demands three nines, while contemporary vision-language-action policies in unfamiliar scenes frequently fall below seventy percent success.
The binding constraint on embodied intelligence is not intelligence. It is the cost and the verifiability of experience โ how expensive it is to obtain a physically grounded sample, and how confidently anyone outside the originating laboratory can attest to what that sample contains.
Here the ledger finally earns its place in the argument, though not in the form the token market imagines. If experience is the scarce input, then the infrastructure that matters is the one that makes experience attributable: a signed record linking a trajectory to the platform that generated it, the environment it was generated in, the software version that controlled it, the operator who supervised it, and the consent under which it was captured. That is a provenance problem, and provenance is the one thing an append-only distributed ledger does better than a database โ not because it is decentralized, but because it is independently verifiable and cheap to audit after the fact, when the only question that matters is what happened and in what order. The European transparency regime does not ask whether your model is decentralized. It asks whether you can reconstruct, following an incident, what the system was doing and why.
The scarce asset in physical AI is not compute and not the body. It is a signed, attributable, replayable record of what the machine experienced and what it did in response โ and almost no one in this sector is building one, because attestation infrastructure is unglamorous, sells to compliance departments, and cannot be pumped on a chart.
The places where crypto does not fit should be named just as plainly, because the misallocation is already visible. Compute is not the binding constraint on embodied models: vision-language-action systems are two to three orders of magnitude smaller than frontier language models, and a cluster in the low thousands of accelerators is sufficient for training. Decentralized GPU markets are solving a problem physical AI does not yet have. The edge is where the real constraint lives โ the single-robot inference platform, in the class of embedded modules that determine real-time capability, sitting squarely inside export-control sensitivity. And beneath that, the true ceiling is not silicon at all. It is harmonic reducers, planetary roller screws, six-axis force sensors, dexterous hands. Capacity and yield, not artificial intelligence.
The strategic consequence is arithmetic. When the body becomes manufacturable at scale, the body converges toward manufacturing margins โ the same trajectory that turned handset assembly into a low-margin contract business while the operating system and the application store captured the profit. Value migrates to the layer that owns the model, the data, and the operational relationship with the customer. Unitree's speech is a defense of the integration layer because the body layer is already, in competitive terms, lost โ not lost to a rival, but lost to commoditization, which is worse, because there is no one to blame and no one to acquire.
The second-order industrial effect is the one least discussed and most likely to matter. Embodied systems require spatiotemporally aligned multimodal data โ synchronized video, joint angles, torque readings, tactile response โ and the annotation complexity is not a linear extension of text labeling. It requires physical capture, supervised teleoperation, and verification against ground truth. Unit cost per usable sample may run two to three orders of magnitude above text. A new occupational category is forming around this: the teleoperator as data worker, in facilities that resemble neither a factory nor an office, and whose labor conditions no regulator has yet thought to examine.
There is a coherent counter-case, and it deserves stating rather than dismissing. If deployment scenarios remain radically fragmented โ if every warehouse, substation and hospital demands bespoke integration โ then the generalist model advantage erodes and supply-chain depth plus channel reach become decisive. Unitree's vertical integration and cost-setting power would compound rather than decay, and the commoditization thesis would invert. Which future arrives depends on a single empirical question no keynote can settle: whether cross-scenario generalist policies emerge before scenario-specific integration economics saturate the market.
Now the contrarian claim, which I make with some reluctance because it implicates a great deal of my own prior work. The sector's proposed solution to the data problem โ token-incentivized collection, decentralized networks paying contributors for teleoperated demonstrations or household robot footage โ will reproduce, with unusual fidelity, the failure mode of liquidity mining. In 2020 I analyzed more than five thousand transactions across stablecoin pools to understand peg stability, and what I found was not a market in liquidity but a market in subsidy. Yield was the product; deposits were the byproduct; and when the emission schedule ended, the deposits left within days, because they had never been users. They had been arbitrageurs of a promise.
Token-subsidized embodied data has the same structure and a worse outcome. Subsidies buy volume, volume buys distribution shift, and distribution shift buys a dataset that is expensive, abundant and useless โ because a demonstration captured in a subsidized teleoperation facility, under one lighting condition, with one gripper, on one object set, tells a policy almost nothing about the warehouse where it will actually be deployed. Liquidity, at least, is fungible. Experience is not. You cannot arbitrage one robot's afternoon into another robot's competence, and no emission schedule can fill the hollow resonance where the evidence should be.
The governance layer compounds the problem rather than resolving it. The data cooperatives being designed around embodied capture inherit the legal condition of nearly every decentralized autonomous organization I have examined over the past decade: no recognized legal personality, no limited liability, and no clear allocation of responsibility when something goes wrong. Consider a household robot whose continuous visual capture is contributed to a tokenized dataset, and then breached. Who is liable โ the manufacturer, the model developer, the integrator who installed it, the cooperative that pooled the data, or the individual holders who voted on the pooling policy? In the current structure the honest answer is that a court would decide, and the natural persons nearest the entity would discover that their exposure had never been limited at all.
The regulatory mispricing is subtler and, for anyone allocating capital, considerably more expensive. The market has broadly priced artificial intelligence plus crypto as a transparency trade โ the intuition being that cryptographic attestation will satisfy the emerging European regime, and that compliance-adjacent tokens therefore carry a structural premium. My reading of the Act's operative provisions points elsewhere. The binding obligations are not declarations of principle made at deployment; they are post-market monitoring, incident reporting, and the maintenance of a technical documentation trail that survives contact with a supervisor after harm has occurred. A token is not an audit trail. A token is a claim on a database that somebody else operates. The premium belongs to whoever supplies the evidentiary substrate, and that supplier is most likely to be a boring company with a boring contract and a very boring renewal cycle.
There is a conceptual trap embedded in the vocabulary, and it will produce bad regulation if it is not named. If supervisors read alignment in the language-model sense, they will legislate about values, intent and behavioral conformity. The physical risk is not located there. It sits in the safety envelope โ torque limits, collision detection, emergency stop, redundant sensing โ and in the allocation of liability when a machine of several dozen kilograms with meaningful force output operates in a space shared with people. Regulating the wrong alignment would be worse than regulating none, because it would consume the political capital without reducing the hazard.
The silence on safety in the address is, in commercial terms, entirely rational and, in systemic terms, indefensible. Every participant in this sector is racing to accumulate deployment before the regulatory perimeter closes, which is the correct strategy for any individual firm and a catastrophic strategy for the category. The first serious physical injury in an open environment will not be adjudicated as an engineering failure. It will be adjudicated as an industry failure, and it will arrive with the same suddenness as the collapse of a major lending platform in 2022 โ the moment a sector discovers that the trust it spent years accumulating was rented, not owned. Here is the hollow resonance of digital ownership in art, transposed to actuation: tokens that purport to convey rights over trajectories nobody can verify, generated by machines nobody can audit, governed by entities nobody can sue.
I spent the 2022 contraction watching forty billion dollars of stablecoin liquidity leave cross-border payment corridors in a matter of weeks, and I learned to distrust any growth metric that could not survive the withdrawal of its own incentive. The discipline that replaced it is simpler and duller: measure what remains when the subsidy stops. Applied to physical AI, that means asking which participants hold a paying customer, a verifiable dataset, and a liability structure a lawyer would recognize on a first read.
Three observable thresholds will settle the question within twelve months, and none requires access to a laboratory. Whether Unitree discloses, in any quantitative form, the scale and provenance of its training data and the architecture of its policy stack โ the absence of such disclosure after a listing would itself be informative. Whether a provenance schema for embodied data achieves adoption beyond its own authors, in the way that attestation formats either become infrastructure or become footnotes. And whether credible third-party safety certification arrives before the first incident rather than in the grim aftermath of it.
In a market where survival has replaced appreciation as the organizing principle, the assets that endure inside this narrative will be the unglamorous ones: the attestation layer, the verification market, the settlement rail that a compliance officer can defend under oath. Everything else is a claim on a story about a robot that has not yet been built, trained on data that has not yet been collected, under a liability regime that has not yet been written.
The body, in the end, was never the business. The question worth carrying into the next cycle is not whether the machine can learn to move through the world, but whether anyone will be able to prove what it learned there โ and who will be standing in the room when the proof is demanded.