When Safety Becomes a Moat: The AI Governance Capture Crypto Already Debugged

Over the past few weeks, I have watched the capital that survived the 2022 crash quietly rotate into a new bet — not tokens, not rollups, but the governance of artificial intelligence itself. Then Yann LeCun said the quiet part out loud. The man who spent a decade building Meta's FAIR lab did not attack Anthropic's models. He attacked its business model. He called the safety narrative a play for regulatory capture, and in doing so, he dragged the oldest problem in decentralized governance into the newest industry on earth.

I have heard "safety" used as a moat before. In 2017, during the ICO chaos, I co-founded TrustChain, an open-source advisory platform that taught retail investors to read a smart contract before they signed it. Twelve projects cleaned their code before mainnet — not because a regulator forced them, but because a community refused to delegate its own diligence. I learned something that year that I have carried through every cycle since: the moment a group outsources its judgment to a central authority, it has already lost the thing it was trying to protect.

That lesson is about to be tested at a scale crypto has never faced, because the same governance failure we spent a decade debugging on-chain is now being rebuilt inside the AI labs — and this time the output is not a token price. It is the rulebook for the next twenty years of computing.

Here is what actually happened, stripped of the drama. LeCun left Meta to found a new venture, built around world models and the JEPA architecture rather than the autoregressive language models that power ChatGPT and Claude. Almost simultaneously, Anthropic moved closer to a public listing, positioning itself as the responsible lab — the one whose safety culture justifies both enterprise pricing and a permanent seat at the regulatory table. LeCun's charge is that the second move is not a safety strategy at all. It is a market-access strategy dressed in the language of ethics.

I want to be precise about my own confidence here, because precision is the only thing a bear market rewards. Several of the numbers circulating in this story are almost certainly wrong. A reported compute commitment in the hundreds of billions does not reconcile with Anthropic's actual disclosed scale; a two-trillion-dollar target valuation sits an order of magnitude beyond even the most aggressive private marks. I have audited enough governance proposals to recognize narrative inflation on sight. Treat the figures as directional, not documentary. The framework, though — the framework is sound, and it is worth your attention.

Anthropic is not a conventional corporation. It is structured as a public benefit corporation, which means its charter formally commits it to a social mission alongside shareholder return. That structure is the entire basis of its premium, and it is also the entire basis of LeCun's attack. A PBC can pursue mission and profit at once — but when the mission conveniently becomes the moat, an outside observer has no way to tell which one is driving. That ambiguity is not a flaw in the argument. It is the argument.

Let me give you the technical background, because the disagreement is not cosmetic. It is architectural.

JEPA — Joint Embedding Predictive Architecture — predicts in an abstract representation space rather than reconstructing tokens or pixels. Where a large language model learns by predicting the next token in a sequence, a world model learns by predicting the next state of a system, in a compressed latent space that discards irrelevant detail and preserves structure. For physical data — video, sensor streams, the messy continuous signals of a factory floor or a robot arm — this is theoretically far more sample-efficient than token-level reconstruction. It is the intellectual through-line of LeCun's career, and it is genuinely original.

But here is the part the evangelists of the new paradigm will not tell you. JEPA's maturity is early-industrial, not product-grade. The public V-JEPA work is impressive at video understanding and representation learning. There is no published evidence of end-to-end industrial control at the reliability a manufacturer would stake a production line on. When I led a volunteer research team of fifteen developers through Uniswap's early governance during DeFi Summer, we learned to separate a mechanism that works on a whiteboard from a mechanism that survives contact with real users. The gap is always larger than the whitepaper admits. JEPA sits on the whiteboard side of that gap, and anyone who tells you it is already "fighting LLMs head-on" is selling you a rivalry that exists mostly in the framing.

The deeper problem is data. World models need the physical world, and the physical world is locked inside private, fragmented industrial systems — proprietary sensor feeds, unshared telemetry, factory data no manufacturer will hand to a competitor. The JEPA bottleneck is not the algorithm. It is access. This is the unspoken reason LeCun's route may take longer than his backers expect. Algorithms can be published. Factories cannot be crowdsourced. The open-source playbook assumes contributors can share their work, but the data that would train a world model is the one resource its owners will never open.

I have spent a decade arguing against a particular kind of premature infrastructure, and the discipline applies here almost perfectly. For years I have pushed back on dedicated data-availability layers built for rollups that will never generate enough data to justify them — beautiful architecture, speculative workload. The JEPA camp risks the same trap. A world-model-native edge stack, custom silicon for physical inference, dedicated industrial training clouds — all of it is plausible, and all of it is premature until someone proves the workload exists at scale. The infrastructure follows the demand, never the reverse. Build the dedicated layer before the demand arrives, and you have not built a moat. You have built a monument.

Now the commercial layer, because this is where the two camps actually collide.

Anthropic's strategy is the classic standard-setter play. Push for regulation, help define the compliance bar, then let that bar become your moat. Enterprise and government customers pay a premium for the certification — the assurance that your model cleared a bar your open-source competitor cannot afford to clear. I have seen this movie. It is Microsoft and the operating system. It is Visa and the payment network. It is every incumbent that ever discovered the cheapest way to beat a faster competitor is to make the rules expensive.

LeCun's counter-position is the open, industrial route — closer to Red Hat's open-core services than to a closed enterprise API. Sell into manufacturing and automation, the verticals where general chat models are weakest, and win on physical-world specialization rather than on compliance paperwork.

These two models cannot coexist under the same rulebook. If the safety bar is legislated into existence, open-weight models get squeezed. If it is not, Anthropic's premium loses its justification. This is not a personality clash. It is a structural conflict over who writes the standard — and whoever writes it decides which business models are legal.

There is a second-order version of this that Uniswap taught me to fear. When V4 introduced hooks, it turned the DEX into programmable Lego — and it also introduced a complexity spike steep enough to scare off ninety percent of the developers who might have built on it. Complexity is a moat whether you intend it or not. A compliance regime that is technically rigorous, legally intricate, and expensive to satisfy does the same thing to open-source AI that hooks did to casual DEX builders: it does not ban them, it simply prices them out of the room. The most effective gatekeeping is never a wall. It is a maze.

And that is why the word "capture" is the most important word in this entire story, and why it should sound familiar to anyone who has spent time in DAO governance.

Here is the uncomfortable truth that crypto natives understand better than the AI labs do: regulatory capture is just delegation, scaled up. When a token holder is too busy to research a proposal, they delegate their vote to a KOL. When a legislator is too busy to understand a technical domain, they delegate the standard to the largest incumbent. The mechanism is identical. The lazy voter and the overwhelmed regulator both hand their judgment to whoever shows up with the most confident narrative. In DAOs, that produces whale-dominated governance. In AI policy, it produces incumbent-dominated rulemaking. Different ledger, same failure.

I watched this happen in real time during DeFi Summer, when we published our fifty-page governance analysis and organized town halls to bridge core developers and token holders — Root: DeFi Summer. The tension we reduced by forty percent was not technical. It was a legitimacy crisis. Holders did not trust the process because they could not see inside it. The same opacity now surrounds AI safety standards, and the same distrust is forming. When you cannot audit the reasoning behind a rule, you eventually stop obeying it — not out of rebellion, but out of the simple human instinct that a rule you cannot inspect is a rule you cannot trust.

This is the civic-governance analogy the AI debate is missing. In a healthy polity, safety standards are set by a body accountable to everyone and captured by no one. In a captured polity, the standard is set by the very entity it is supposed to constrain. A safety regime authored primarily by the company it regulates is not safety. It is a constitution written by the king.

LeCun's most effective move, and the one I respect most as a strategist, is that he never questioned Amodei's honesty. He questioned the effectiveness of the model. He kept the argument on institutional design rather than personal integrity — which is exactly how a mature competitor behaves. When you attack a person, you create a martyr. When you attack a mechanism, you force everyone to examine the mechanism. I have made the opposite mistake, and I have paid for it. During the 2022 collapse, when anxiety peaked and half my community was ready to quit, I learned that the only persuasive move left in a crisis is to be scrupulously fair about your opponent and scrupulously specific about your objection — Root: The 2022 Bear Market. The Resilience Hub we built then, connecting two hundred junior developers to mentors, taught me that people do not follow certainty. They follow fairness under pressure.

So let me apply that same fairness to LeCun, because the open camp has a blind spot of its own, and pretending otherwise would make me exactly the kind of partisan I am warning against.

The contrarian angle — the one both camps are avoiding — is that the regulatory-capture critique is symmetric, and the open route is fully capable of its own capture. "AI for the physical world" can just as easily become a lobbying position: grant industrial world models an exemption from the safety rules that bind the language models. That is capture too, wearing the mask of pragmatism. If we accept the reductionist claim that every safety concern is really a business motive, we will be defenseless when a genuine risk arrives. And it will arrive — not as the cartoon existential threat both sides invoke to raise money, but as the mundane, mid-spectrum failure modes nobody is pricing: autonomous agents that deceive their operators, multi-agent systems that coordinate in ways no single designer intended, model behavior that drifts after deployment in ways no static certification can catch. The existential-versus-engineering binary is a distraction. The dangerous middle is where the real alignment work lives, and neither camp wants to fund it, because it does not make for a good headline.

That middle is exactly where my own recent work landed. In 2026, as AI agents began transacting on-chain, I convened a global working group of thirty ethicists and developers to draft an accountability charter — a document that asked the question the certification regime cannot answer: who is liable when an autonomous agent fails? We reached consensus, across seven workshops, on one uncomfortable principle: liability cannot be certified in advance, because the failure modes are emergent. You cannot pre-approve a system for behaviors that only appear in combination with other systems. This is the quiet bankruptcy of the compliance-moat thesis. A certificate is a snapshot. An agent is a moving target. The premium Anthropic wants to charge for is a photograph of a river.

I am not, let me be clear, anti-regulation. During the 2024 ETF cycle, I led a campaign across ten Asian universities to argue the opposite: that institutional legitimacy and decentralization are not enemies, and that the honest framing is "responsibility within freedom," not "regulation versus freedom." Regulation done well can enhance trust and widen participation. But regulation done by the incumbents, for the incumbents, does neither. The distinction is not whether the rule exists. It is who holds the pen.

There is a second blind spot, and it is the one I find most striking. The entire debate is being conducted as though only American labs exist. DeepSeek, Alibaba, ByteDance — the Chinese labs are almost entirely absent from this framing, even though any global compliance regime will have to account for them, and any open-weight ecosystem already depends on them. A safety standard written in Washington and applied only in Washington is not a safety standard. It is a trade barrier, and everyone outside the room knows it.

So where does this leave someone holding assets in a bear market, trying to judge which protocols are bleeding and which are simply quiet?

The directional signal is clear even if the specifics are not. Capital is splitting into two investment theses — compliance-native AI and open industrial AI — and they are not diversifiable against each other. They are a bet on which rulebook wins. The near-term catalyst is Anthropic's listing, because an IPO is the moment a safety narrative must convert into a valuation number, and the market will finally price the premium everyone has been asserting. If the listing lands and regulation does not follow, the premium evaporates, and every enterprise contract priced on "certified safety" gets repriced on "commodity capability." That is the bear case nobody in the bull camp wants to say out loud.

What I would actually watch, if I were allocating attention rather than capital — and in this market, attention is the scarcer asset — are the neutral players. The compliance auditors. The safety certifiers. The cross-camp interoperability layers. The third parties who can serve both rulebooks and therefore profit regardless of which one wins. That is the same role the resilient, unglamorous infrastructure played after 2022 — Root: The 2022 Bear Market — the projects that survived not because they were exciting, but because they were needed by everyone.

And watch the edge. If world models are real, the compute demand shifts — from the hyperscale data center toward the factory floor, toward low-latency edge inference, toward a chip ecosystem that looks nothing like the centralized training clusters we have spent three years worshiping. That is a structural change in infrastructure demand that the current debate has not even begun to price.

Code is law, but people are the protocol.

That sentence has never been more literal than it is right now. The AI labs are writing law — actual, binding, market-defining law — and they are doing it through the same delegation-and-capture mechanism we have spent a decade trying to design around on-chain. The crypto industry has a choice about how to meet this moment. We can treat AI governance as a spectator sport, another narrative to trade. Or we can bring the one thing we actually built: a hard-won, bloodied, empirical understanding of what happens when you let a central authority define the rules for a decentralized world.

The bear market has a way of revealing which institutions are load-bearing and which were only ever narrative. I have lived through three of them now, and I can tell you the survivors are never the loudest. They are the ones whose rules anyone can read. As the AI industry races to write its constitution, ask the only question that has ever mattered in governance: who gets to inspect the rulebook, and who gets to change it?

If the answer is "the company the rulebook governs," then we have not solved the problem. We have only moved it to a bigger ledger.