The Dinner Without a Witness: Anthropic's Warm-Up and the Case for Verifiable AI Governance

For twenty-seven years in this industry, I have watched power consolidate — and learned to distrust the precise moment it starts calling that consolidation “partnership.” On a September evening in 2025, Dario Amodei, the chief executive of Anthropic, sat down for a private, one-on-one dinner with President Donald Trump. Two days later, he joined a larger table of technology leaders to meet the Speaker of the House. The reporting that surfaced afterward, drawn from a single anonymous source and only a handful of bare facts, described Amodei's place in this arrangement with one telling word: “warm-up.” That word is doing more work than any policy document published this year. It tells us that Anthropic — the laboratory founded on the promise of “safety first” — is not leading this relationship. It is warming up. It is seeking entry. And in any governance structure, a relationship defined by dependence is the first step toward capture. I spend my professional life designing systems in which no one needs to warm up to anyone. In the DAO frameworks I have architected over the past decade, legitimacy flows from transparent, verifiable participation rather than from proximity to power. That distinction is not an aesthetic preference; it is a structural safeguard. When the rules are legible to everyone, nobody has to whisper them across a dinner table. The Anthropic story sits at the opposite pole. The reporting describes a “long-standing tension” between the company and government officials — tension rooted, for anyone who has followed Amodei's public advocacy, in a philosophical collision over regulation. Anthropic built its reputation on the argument that frontier AI demands public guardrails. The administration it now courts has built its reputation on dismantling those guardrails in the name of national acceleration. This is not a private squabble. It is a conflict between two visions of how technology should be governed: one that trusts central authority to set the pace and shape of the future, and one that trusts distributed, verifiable consensus. For years, I have argued — in whitepapers, in workshops, in essays that cost me friends in the speculation economy — that the real question of any system is not who holds power, but whether those outside the room can verify what happens inside it. The dinner between Amodei and Trump is significant precisely because it is unverifiable. We know it happened. We do not know what was traded. Let me be precise about what “capture” actually means, because the word gets thrown around loosely. Governance capture occurs when the entity responsible for enforcing the rules becomes financially or structurally dependent on the entities it is meant to oversee. In the DAO systems I have audited, I have watched this happen in slow motion. A handful of whale wallets accumulate enough voting power that every proposal passing through the forum becomes theater — the discussion is public, but the outcome is predetermined. The participants are still voting. They have simply lost the ability to change anything. The government-lab relationship forming around Anthropic has the same silhouette. The government is both customer and regulator. It buys the AI, and it writes the rules that govern the AI. When those two roles collapse into one, the regulator loses its incentive to regulate strictly, because strictness would punish its own supplier. The customer gains an incentive to keep its supplier dependent, because dependency is leverage. Anthropic is not being welcomed into the room. It is being given a seat at a table it does not own. I know this dynamic from the inside. In 2026, I led the design of a decentralized governance framework for AI training-data ownership — a system in which every contributor of data receives a verifiable credential, and every use of that data leaves an auditable trace. We piloted the model with 10,000 data providers and negotiated its adoption with three major AI labs. The single hardest obstacle was not technical. It was convincing people that verifiable, distributed ownership could coexist with the commercial interests of the labs themselves. And the reason it was hard is that centralization is genuinely more efficient in the short term. Distributed governance costs time, coordination, and friction that no rushed product roadmap wants to pay. That is the trap. The efficiency argument is always plausible, and it is always the argument that precedes capture. Now look at the competitive landscape, because Anthropic is not playing this game alone. The reporting describes its position as a “warm-up,” and the word is revealing. OpenAI's leadership moved into the policy orbit earlier and more fluently; it appears, by every available signal, to be the incumbent of this new channel. Google brings cloud infrastructure, compute, and decades of Washington lobbying. Meta wields an open-source posture and a formidable political-spending machine. Even xAI, whose influence rises and falls with one man's personal relationship to power, occupies space that Anthropic does not. And the defense-native firms — the ones that have sold to the Pentagon for two decades — already own the contracts that these laboratories now covet. Anthropic, by contrast, arrives late, carrying a “safety-first” identity that was an asset in a regulatory-minded capital and is a liability in a deregulatory one. This is the cruelest irony of the moment: the very label that makes Anthropic trustworthy to defense and government buyers is the label that makes it politically inconvenient to the government now in power. Safety is simultaneously its procurement asset and its policy debt. A laboratory that trades that label for a seat at the table is simply wearing a newer, more expensive empty vest. There is a specific policy fight worth naming here. At the center of American AI policy in 2025 is a quiet war over whether federal law should preempt state-level AI regulation. The presence of the House Speaker at the Tuesday meeting points directly at this legislative question. If the federal government consolidates authority over AI rules, the laboratories closest to it gain enormous structural advantage — not because they are better, but because they helped write the standard they are measured against. This is what makes the warm-up so consequential. A company does not warm up to a government for a handshake. It warms up for access: procurement channels, compute allocations, regulatory shelter, and a voice in the standards that will define the next decade. None of that is disclosed in a press briefing, and none of it is auditable by the public whose interests those standards are supposed to protect. Here is the insight that the headlines miss. The competition between AI laboratories is migrating away from model capability and toward political-commercial capital. As the capability gap between frontier models narrows — and it is narrowing — the durable differentiation shifts to who gets the government contract, who gets the favorable compute policy, who gets the regulatory shelter. This is the same pattern I have watched in my own field. When technical differentiation compresses, the winners are decided not by the market but by who is closest to the rulebook. I have seen it happen in Layer 2. When data availability costs collapsed after Dencun, everyone celebrated the cheap fees — and almost nobody asked what happens when the blob space saturates. The economics that feel like abundance are often just an unpriced subsidy that resets the moment demand catches up. The AI version of that subsidy is government access, and it is already being priced into valuation narratives rather than engineering roadmaps. The labs that secure the channel will look more valuable right up until the terms of the channel are renegotiated from a position of dependence. I have also watched narrative become the decisive force in markets that pretend to be meritocratic. The inscription wave on Bitcoin is the cleanest example I know. For years, critics dismissed Ordinals as frivolous, but the inscription surge delivered something Bitcoin's security model structurally required: a new stream of transaction fees and a new reason to care. Without that narrative, the fee market would have looked far more precarious. The lesson generalizes. Narratives are not decoration on top of economics; they are the load-bearing beams. And the most powerful narrative available to an AI lab right now is not “safest model.” It is “the model the government trusts.” And yet I have to test my own bias, because the reflex to defend decentralization is not the same as the argument for it. It is entirely possible that frontier AI is one of the rare domains where centralization is the responsible choice. The failure modes are severe, the timelines are compressed, and the coordination required to prevent catastrophe may genuinely exceed what distributed governance can deliver in time. A pragmatist could argue that a warm dinner between a safety-conscious CEO and a deregulatory president is not capture — it is harm reduction. Better to be at the table shaping the outcome than shouting from outside it. I take that argument seriously. But it rests on one hidden assumption: that influence at the table is symmetric with dependence on it. It never is. The party who needs the meeting concedes more than the party who grants it. This is why I return to a principle I have carried through every governance system I have designed: don't govern the exit, govern the entrance. The integrity of a system is decided at the point of admission — who gets in, on what terms, with what verifiable conditions — not at the point of departure, when the leverage has already shifted. A private dinner is an entrance with no witness. And an entrance with no witness is a rule that cannot be audited. So the question worth asking is not whether Anthropic had a nice evening. The question is whether a civilization is comfortable letting a handful of private laboratories negotiate, over unrecorded dinners, the rules that will govern the most consequential technology it has ever built. Code is law, but people are the soul. And a soul cannot verify what it is not allowed to see. If the governance of artificial intelligence is to preserve human agency, it cannot be assembled at tables we are not invited to. The next decade of this battle will not be won by the loudest model. It will be won by whoever builds the system transparent enough that no one ever has to warm up to anyone again.