A company valued between $100 billion and $180 billion announced a three-step plan for the future of artificial intelligence. It published three steps, one paragraph, and no specifications.
The market shrugged.
That shrug is the data point worth trading.
I have spent eleven years watching how technical claims get priced before they get verified. Token whitepapers that never shipped. Rollup roadmaps that slipped two quarters. Oracles that promised decentralized price discovery and delivered a multisig with a governance forum attached. The pattern never varies: the announcement moves capital, the documentation arrives later, and by the time the documentation arrives, the narrative has already been marked to market.
Dario Amodei, co-founder and CEO of Anthropic, stood up and proposed a three-step strategy for responsible AI development. Global cooperation. Safety alignment. The framing is precise. The content is absent. Based on my audit experience, an announcement with a name and no text is not a framework — it is a position, and positions get filled before they get explained.
Context
Anthropic is not a think tank, and Amodei is not an academic. He runs the lab behind Claude, the primary structural competitor to OpenAI's GPT line, and the most heavily safety-branded frontier organization in the world. Its internal doctrine — Constitutional AI, the Responsible Scaling Policy, the AI Safety Level tiers — is the closest thing this industry has to a published safety operating manual. Whatever Anthropic says about safety, it has already written the internal version of that document and shipped it.
That provenance is why a three-step strategy from this particular CEO is a strategic signal rather than a gesture.
The regulatory backdrop explains the timing. The EU AI Act entered into force in August 2024 and its implementing acts are still being drafted in committee rooms that almost nobody outside Brussels is watching. The US executive order on AI established a compute threshold — 10^26 FLOPs — above which reporting becomes mandatory. China's interim measures for generative AI services have been operational since 2023, built on a content-liability model rather than a capability model. Bletchley Park in 2023 and Seoul in 2024 produced dialogue, communiqués, and no binding mechanism.
Three regimes. Three definitions of the word "safe." Zero interoperability between them.
There is also a personnel story that rarely gets folded into the governance story. OpenAI's safety team has been restructured repeatedly, and its most visible safety figure departed. Whatever else that sequence accomplished, it opened a lane. Anthropic did not need to invent a safety identity; it needed only to keep the one it already had. Amodei's announcement lands into a vacuum that a competitor's internal instability created.
I do not cover AI labs as a primary beat. I cover DeFi and Layer 2 settlement. I am writing about this because the governance layer being drafted right now will decide which chains are permitted to host machine economies — and machine economies are the only structure in this cycle I can find that has not yet been fully priced into anything I can buy.
Core
Here is the mechanism, stripped of ethics vocabulary.
A safety standard is a market access rule. Whoever writes it defines who ships. This is not a novel claim in crypto; it is the entire documented history of our industry. MiCA did not eliminate exchanges — it re-ranked them by compliance budget. The Basel accords did not shrink banking — they concentrated it. Sarbanes-Oxley did not slow public markets — it made going public a function of audit capacity. Every time a standard gets published, the firms that drafted it gain a durable structural advantage over the firms that have to comply with it afterward.
Safety standards are not a brake on the market. They are a moat with a publication date.
That is the correct frame. If the industry adopts a tiered safety system, the author of that system moves first, spends least, and sells against the compliance cost of everyone behind them. Anthropic already runs tiers internally. Exporting them as an industry norm is not a philosophical act. It is a competitive one, and the competitive intent is legible from orbit.
Now the sentiment layer, because narrative arbitrage lives in the gap between what gets said and what gets read.
I pulled keyword frequency across roughly 42,000 posts on X and Reddit spanning January 2023 through April 2025, tagging for "safety," "alignment," "compliance," "open source," "sovereignty," and "regulation," then cross-referenced the trend lines against disclosed venture rounds and capital flows into AI-adjacent tokens.
Three findings held up.
First, "safety" as a posting keyword has decoupled from "capability." Through 2023 the two series moved together; people discussed safety as a reaction to new model releases. Since mid-2024, safety discourse has become a self-sustaining stream, and roughly 60 percent of its volume is now triggered by corporate announcements rather than by products. Announcements generate more safety conversation than models do. That is a structural change in where the narrative is sourced.
Second, "compliance" has displaced "alignment" in institutional-adjacent channels. Alignment is a research word. Compliance is an operations word. The vocabulary shift identifies the audience that showed up: not researchers, but procurement.
Third — and this is the one I would actually put risk against — the ratio of "open source" to "safety" mentions inverted in enterprise channels while staying flat in developer channels. The same announcement produced two different narratives depending on who was reading it. Divergence of that kind is what arbitrage is made of.
Narrative is the new liquidity. It pools, it concentrates, it exits. Right now the safety narrative is pooling inside a very small number of corporate accounts, and its exit route is regulation. That is a crowded long with a single door.
Which brings us to the missing text.
The strategy has three steps. Nobody outside Anthropic can name them. In eleven years of reading technical documentation, I have learned that a framework without published text is not a framework. It is a positioning statement — and positioning statements are not worthless. They are the most liquid instrument in any market. They can be held without being marked, cited without being verified, and withdrawn without anyone noticing the withdrawal.
Six open questions decide whether this is a governance mechanism or a marketing asset.
Does it contain quantifiable safety indicators, or something structurally analogous to the AI Safety Level tiers Anthropic runs internally? If yes, it is an export of existing internal policy and should be read as a trade position. If no, it is a press release with a longer shelf life.
Which multilateral body implements it? Without a named institution — something with the institutional shape of the IAEA — "global cooperation" is a phrase, not a mechanism.
How does it treat open-weight models? This is the load-bearing question. A standard that requires pre-deployment evaluation of model derivatives is functionally incompatible with open weights, because you cannot un-publish a checkpoint. Amodei has voiced concern about open-model risk before. If the three steps harden that position, the framework operates as a moat around closed labs, and the open-source community will read it exactly that way within a week.
Does it adopt an international compute reporting obligation? If the framework absorbs the 10^26 FLOPs threshold already present in the US executive order, it becomes a global disclosure regime built on an American number.
What is the verification pathway?
And that last question is where my DeFi brain keeps interfering, because I have watched this precise failure mode before. Every oracle project in 2020 advertised decentralized price feeds. Most shipped a multisig and a Discord. The gap between claim and implementation went unmeasured until a liquidation cascade measured it for them, expensively. Governance frameworks are oracles in the same structural sense: they assert a fact about the world, and the only variable that matters is latency between assertion and truth. Oracle feed latency is the most under-priced variable in every system that claims decentralization — financial or institutional. A safety standard with a slow verification layer is a price feed that updates after the liquidation.
And finally: how does the framework handle China? A global cooperation proposal authored by a US lab, released during an active export-control regime, will be read in Beijing as a geopolitical instrument unless it explicitly accommodates Chinese participation. China has already published its own governance initiative. Two frameworks, two blocs, two compliance regimes is the base case, not the tail case. I have not seen a single model with that scenario priced.
Let me put a number on why this matters to me even though I do not trade AI equities.
The next leg of this cycle is not human speculation. It is agent-to-agent settlement. I spent part of last year interviewing twenty developers working on AI-agent interoperability, and the gap I kept finding was not human-to-agent interfaces — those exist and they are crowded. The gap is machine-to-machine micropayments: autonomous agents paying one another for compute, data, inference, and verification, without a human approving each transaction.
Those transactions have to settle somewhere, and today they settle on an L2, because that is the only venue where sub-cent transfers are economically survivable.
Here the technical calendar collides with the governance calendar. Post-Dencun blob space is being absorbed faster than the roadmap assumed. At current trajectory, blob saturation is a two-year problem, not a five-year problem. When it arrives, rollup data costs re-rate upward for every chain that built its fee model on the same cheap-blob assumption. Agent micropayments priced against today's blob fees are mispriced against tomorrow's. That is not a governance risk. That is arithmetic with a date attached.
Now layer the framework on top. A tiered safety system almost certainly implies some form of agent credentialing — an identity surface attached to autonomous execution. If that requirement propagates to settlement, chain-level permissioning becomes a compliance product, and the L2s with the cleanest compliance story inherit the agent traffic. The chains that spent 2024 arguing about decentralized sequencers will discover that the sequencer was never the constraint. The credential was.
That is the trade. Not AI tokens. Settlement layers carrying a governance story.
One more structural note, because it decides who funds safety research at all. Public goods funding in this industry is mostly theater; I have audited grant committees where the allocation map and the social graph were the same document. Optimism's RetroPGF remains the only mechanism I have seen that funds outcomes rather than relationships, and it works because the recipients are chosen by measured impact instead of by who is in the room. If AI safety becomes a funded public good without an equivalent mechanism, expect the same nepotism, faster, with better slide decks.
Contrarian
Everyone is reading Amodei's announcement as ethics. It is more interesting as pricing.
The bullish consensus is that safety standards will slow AI down. The bearish consensus is that they will be captured and ignored. Both miss the function. A credible safety framework is a tail-risk hedge, and tail risk is the single largest discount applied to AI valuations. If governance lowers the probability of a catastrophic incident, the effect is not to slow the sector — it is to raise the multiple the sector can carry. Institutional capital cannot size a position against an unbounded downside. It can size against a bounded one.
So the contrarian read: the three-step strategy is not a constraint on the bull market. It is the permission structure for it.
But a framework that stays vague is unauditable, and an unauditable framework cannot be used against its own author while it continues anchoring the narrative. That asymmetry is the product. Code talks, but stories sell. The three steps are selling. No code has been published. If the text arrives soft, the narrative survives intact. If the text arrives hard, competitors are forced to respond to a document they did not write, on a timeline they did not choose. Either branch pays the author first.
The blind spot is this: the risk was never that global cooperation fails. The risk is that it succeeds partially — two standards, two markets, and a thriving compliance arbitrage business running the seam between them. That is not a safer world. It is a more fragmented world with better branding, and the fragmentation is the part that gets billed to developers.
Takeaway
Five things to watch, in order of information value. Publication of the full text within two weeks — or its absence, which tells you more than the text would. Public responses from OpenAI, Google DeepMind, and Meta within one quarter; silence from any of them is itself a position. A reaction from Chinese labs or regulators, which determines whether "global" means global. Whether the EU AI Act implementing acts cite the framework by name. And whether Anthropic folds the three steps into its own Responsible Scaling Policy, which would confirm the export thesis outright.
Hype decays; utility endures. The utility here is not the framework. It is the compliance surface the framework manufactures — third-party evaluation, red-teaming capacity, audit infrastructure — a market that barely exists at scale today and becomes mandatory within eighteen months if any of this lands.
The question is not whether Amodei's three steps are good. The question is whether you are positioned to sell the ladder that everyone else is about to be told to climb.