Over the past few weeks, a single portfolio move by Leopold Aschenbrenner has been dissected in AI circles. But I want to look at it from a different angle—one that sits at the intersection of trust, governance, and the uncomfortable truth that even the most brilliant minds can fall for the same centralization trap we crypto natives warn against.
Aschenbrenner, the author of the viral essay Situational Awareness, sold his entire stake in AI infrastructure stocks—Nvidia, Amazon, and others—and concentrated his fund into a single private company: Anthropic. The reported $45 billion figure (which I’ll flag as suspicious given the fund’s known size) is less important than the signal. This is a man who has argued that AGI could arrive by 2027, requiring trillion-dollar compute clusters. Yet he dumped the picks-and-shovels plays and went all-in on one model builder.
From a technical perspective, this is a bet on a specific route: the “capability + alignment” parallel track that Anthropic champions. Aschenbrenner’s background in AI safety makes him naturally drawn to their Constitutional AI, Responsible Scaling Policy, and interpretability research. But as someone who has spent years auditing smart contracts and watching DAO governance fail precisely because “code is law” is a myth—upgrade rights always sit with a few multi-sig admins—I see a parallel here. Anthropic’s alignment is ultimately controlled by a small team. The same centralized trust vulnerability that plagues crypto projects.
Democracy isn’t a transaction where every voice holds weight. In crypto, we’ve learned that governance isn’t about the code you write; it’s about the people who hold the keys. Aschenbrenner’s move reveals a belief that Anthropic’s team is uniquely trustworthy. But is that a defensible thesis, or just a charismatic founder bet? The core of my analysis comes from experience: I’ve seen too many projects where the whitepaper promised decentralization, but the multi-sig was held by three people. Anthropic’s “constitution” is a smart contract without a blockchain—beautiful on paper, but unverifiable without the ability to fork.
Where does that leave us? The contrarian angle is that Aschenbrenner’s concentration is actually rational from a game-theoretic standpoint. If AGI is a winner-take-all outcome, then betting on the single most likely winner makes sense. But it’s a bet on human integrity, not mathematical certainty. In crypto, we’ve built systems that don’t require trust. Aschenbrenner is trusting Anthropic’s leadership to stay aligned. That’s a fragile foundation.
Let me give you a specific, first-hand example. In 2017, I audited a DeFi project that had a brilliant economic model. The team was reputable. The code was clean. But the admin keys were held by the CEO. Six months later, the CEO’s wallet was compromised, and the entire pool was drained. The math was sound; the human layer was not. Anthropic’s alignment research is impressive, but it’s still a company with a board, investors, and a CEO. The moment profit incentives conflict with safety guarantees, whose constitution wins? Based on my audit experience, the answer is almost always the one with the signing key.
Aschenbrenner is arguably the most systematic thinker in the AGI debate. His decision to dump Nvidia suggests he believes the infrastructure bubble is overvalued, and that the real value accrues to the model that achieves AGI first. That’s a bold thesis. But for a crypto native, it’s also a warning: when you concentrate your trust in a single entity, you’re no longer participating in a decentralized system. You’re participating in a hierarchy.
The forward-looking thought is this: the next trillion-dollar asset will not be a token or a stock. It will be a trust protocol. Whether that’s a blockchain-based AI verification layer or a DAO-governed model, the market will punish those who rely on reputation alone. Aschenbrenner’s bet might pay off, but it’s a bet against the very principle of verifiable trust that we’ve spent a decade building.
Value flows from alignment, not just computation. The most dangerous code is the one you can’t read. Trust is a process, not a transaction.