Anthropic writes the rules it already clears

Anthropic published two policy documents and a Dario Amodei essay on Tuesday, proposing that governments be empowered to block deployment of frontier AI models above a specific compute threshold from companies above a specific revenue threshold. The thresholds, conveniently, bracket Anthropic and roughly three other firms. The same week, Anthropic and OpenAI both filed confidential S-1s.

The package is called the Policy on the AI Exponential, and contains the Advanced AI Framework (AAF), the Economic Policy Framework (EPF), and Amodei's accompanying essay. I read all three. The structural story is in the threshold numbers.

What was actually filed. The AAF proposes that governments have authority to "block deployment of models above 10^25 FLOPs from firms over $500M AI revenue pending safety review."¹ FLOPs (floating-point operations, a measure of training compute) and revenue are the two gates. The EPF pledges $200M toward an Economic Futures Research Fund and names basic income, sovereign wealth funds, and equity-sharing as appropriate policy tools in a scenario of 10% AI-driven unemployment.² Amodei's essay reframes public concern about AI as "a democratic accountability problem," not a PR problem.³

The headline reading is that Anthropic is volunteering itself for regulation. The structural reading is that Anthropic is proposing a regulatory perimeter drawn tightly around itself, OpenAI, Google DeepMind, and (depending on revenue timing) xAI, and excluding everyone else.

The threshold geometry. Two gates have to be cleared simultaneously for the AAF to bite. A frontier lab training a 10^25 FLOP model but selling under $500M of AI revenue is outside the perimeter. A profitable AI vendor running smaller models is also outside. Only firms that have both the compute scale and the commercial scale fall in. As of this month, that is a four-firm list, and Anthropic put itself on it.

This is the standard incumbent-regulator move. The compliance infrastructure Anthropic has already built, Responsible Scaling Policy, red-teaming, model cards, third-party evaluations, becomes a sunk-cost moat. A firm at $400M of AI revenue training a 10^25 FLOP model has a sharp incentive to stay at $400M, or to structure revenue so the AI line stays under the threshold. Digital Applied's readout describes the result as a two-tier market: large regulated labs above the line, a largely unregulated long tail below it.⁴

The framework also leaves the technical definitions soft. "Training run" is not pinned down precisely enough to prevent workload-scheduling games — distributed training across legal entities, sequenced runs that each clear the threshold separately, fine-tuning on top of a sub-threshold base. The compute gate is gameable in a way that the revenue gate is not, which is itself revealing about which firms the framework is actually designed to bind.

The IPO context is hard to ignore. Both Anthropic and OpenAI reportedly filed confidential S-1s with the SEC this week.⁵ Anthropic's last disclosed valuation is $61.5 billion from the March 2025 round led by Google and Spark, against approximately $14.7 billion in total disclosed funding.⁵ A comprehensive, regulator-friendly policy package landing in the same window as confidential filings is a specific kind of document. It is the document you want quoted on the roadshow.

The narrative writes itself: Anthropic is the lab that proposed binding safety standards, that named labour-displacement scenarios concretely, that pledged research capital toward distributional effects. Whether or not any legislature adopts the AAF, the policy document does work for the listing. Enterprise procurement teams reading the S-1 see a counterparty positioned as the responsible incumbent. Institutional investors see one less category of regulatory tail risk. The narrative cost of writing the framework is near zero. The optionality value of having written it, in an IPO window, is not.

The EPF is more interesting than the headline number. Naming basic income, sovereign wealth funds, and equity-sharing as serious policy tools is not standard US corporate orthodoxy. For a company filing an S-1 into a political environment that is, at best, ambivalent about redistributive policy, endorsing these mechanisms in writing is a non-trivial commitment. It is the most surprising piece of paper in the package.

It is also $200M against a $61.5B valuation. That is 0.32%. The mismatch between the problem described, a scenario of 10% unemployment requiring sovereign-scale interventions, and the resource committed is the kind of thing the EPF itself would, in a different document, call out. A research fund is not a hedge. It is a research fund.

The capital chose research, not redistribution.

I do not think this is cynical exactly. I think it is what a pre-IPO company can credibly pledge without spooking the book. A sovereign-wealth-scale commitment from a private firm would be a different kind of document, and would not survive contact with a roadshow. The EPF is calibrated to be publishable.

Amodei's "democratic accountability" framing. The essay's sharpest line is the rejection of the "PR problem" framing. It is also the line that contains the most tension. The Advanced AI Framework was not produced by an elected legislature, a standards body, or a multi-stakeholder process. It was produced by one private company, which placed itself inside the regulatory perimeter it drew. Asking who appointed Anthropic to define the accountability framework is not a gotcha; it is the question the framework itself raises and does not answer.

What this is a case of. Incumbent firms proposing the rules that bind their tier, calibrated to leave their competitive position intact while raising costs for adjacent entrants. The closest analogues are Basel capital rules for tier-one banks and the EU's GDPR enforcement geometry. In both cases, the largest firms ended up advantaged by the regime they had influence in writing. The AAF is on that pattern.

What to watch.

• Whether any legislature picks up the 10^25 FLOPs / $500M thresholds verbatim or modifies them. Verbatim adoption would be the strongest signal of Anthropic's regulatory standing; modification toward broader perimeters would be the signal that legislators see the framing.
• Whether sub-threshold competitors (Mistral, Cohere, the open-weights camp) respond with their own framework proposals, or argue against the threshold geometry directly.
• The S-1s. If the Anthropic and OpenAI filings cite policy posture as a risk-mitigation factor, the IPO-prep reading is confirmed.
• Whether the $200M EPF fund actually deploys capital in 2026, and to whom. Research grants to think tanks aligned with Anthropic's policy positions would be one signal; grants to independent labour economists would be another.

Footnotes

¶