Who Elected Anthropic?

Nobody. And that’s not the interesting part.

The document that increasingly shapes how the United States, the United Kingdom, and the European Union think about the risk level of frontier AI models was not produced by a legislature, a treaty body, or an elected official.

It was written by Anthropic. For Anthropic. About Anthropic’s own models.

The Responsible Scaling Policy, now on version 3.3, established the capability thresholds that trigger safety requirements, the ASL levels that classify AI risk, the evaluation methodology that determines whether a model is dangerous, and the institutional vocabulary, safeguards, capability evaluations, deployment standards, safety cases, that governments now use when they speak about AI risk. [1] The UK AI Safety Institute, EU frontier-model policy discussions, and U.S. voluntary frontier safety commitments all became legible through overlapping language: capability thresholds, safeguards, evaluations, deployment standards, and safety cases. Anthropic’s own RSP v2.2 explicitly says the policy was designed in the spirit of emerging UK, EU, and U.S. proposals and “helps satisfy” the White House and Frontier AI Safety Commitments. [2]

Anthropic wrote the standard. Then the standard became the standard.

Nobody voted on this. Nobody had to.

That is the mechanism worth understanding.

The Oldest Power Move

There is a long history of private actors writing the rules that govern their own industries.

Rating agencies graded their own clients’ debt. Accounting firms audited the companies that paid them. Industry associations set the safety standards that determined whether competitors could enter the market. Pharmaceutical companies ran the trials that their own drugs had to pass.

None of this required conspiracy.

It only required that the institution with the most relevant expertise, the most resources, and the most immediate interest in having a standard be the institution that produced it.

The result, in every case, was a standard that happened to favor the institution that wrote it.

Anthropic is running the same move on a larger board.

The Responsible Scaling Policy is not a trick. It is a genuine safety document. The capability evaluations are real. The ASL thresholds represent actual attempts to define dangerous capability levels. The interpretability research that underlies them is substantive. Anthropic’s co-founder and Responsible Scaling Officer, Jared Kaplan, takes the work seriously. [3]

None of that is the point.

The point is that Anthropic determines whether Anthropic’s models cross Anthropic’s thresholds. The institution that wrote the standard assesses its own compliance with the standard and publishes the results. By version 3.3, the RSP had been updated six times, each time by Anthropic, each time in response to conditions Anthropic was best positioned to evaluate, each time producing a framework that Anthropic was best positioned to satisfy. [1]

A standard you wrote for yourself, that you assess yourself, and that you revise yourself is not a constraint. It is a credential.

The most consequential line in the RSP is not in the capability thresholds. It is the escape clause embedded in version 3.0: if Anthropic were to pause development unilaterally while others did not, “the developers with the weakest protections would set the pace, and responsible developers would lose their ability to do safety research.” [4]

Translation: Anthropic will not stop if others do not stop.

The commitment to safety is conditional on the competitive environment, an environment that Anthropic’s own regulatory advocacy, its vocabulary’s dominance in policy discourse, and its capital structure’s requirements all help shape.

The firm that defines the condition also influences whether the condition is met.

A constraint you can always explain your way out of is not a constraint.

It is a permission structure.

The Vocabulary That Governs

Power does not always announce itself as power.

Sometimes it arrives as vocabulary.

When the UK government stood up its AI Safety Institute, it needed a framework for thinking about frontier model risk. Anthropic’s was available. When European regulators needed categories for frontier-model risk, the conceptual architecture was already legible because Anthropic had published it. When the White House sought voluntary safety commitments from frontier labs in 2023, the commitments were drafted in a language Anthropic had already helped establish. [2]

None of this was imposed.

It was adopted.

That is what makes it durable.

Regulatory capture in the classical sense requires a regulator to capture: a specific agency, a specific rule, a specific official.

What Anthropic achieved was earlier and less reversible.

It produced the intellectual architecture before the regulators arrived.

When governments needed safety vocabulary, Anthropic’s was operational. When safety institutes needed evaluation methodology, Anthropic’s researchers were already running it.

The firm did not capture the regulatory environment.

It constituted it.

This means every other frontier lab is now pressured to answer questions in a framework Anthropic helped design. A lab that does not conduct mechanistic interpretability research of the kind Anthropic pioneered lacks the safety infrastructure the standard implies is necessary. A lab that has not published ASL-style capability commitments has no visible restraint architecture. A lab that has not built the vocabulary cannot even speak the language in which its own safety practices are assessed.

The safety standard is simultaneously a bar all serious labs must clear and a bar Anthropic cleared first, at its own pace, on its own terms.

That is not a level playing field.

It is a field Anthropic surveyed and staked.

The Capital Structure of the Responsible Party

Anthropic’s safety mission requires staying at the frontier.

Staying at the frontier requires compute.

Compute at frontier scale comes from Amazon, Microsoft, and Google.

Amazon has committed up to $33 billion in equity, against Anthropic’s pledge to spend more than $100 billion on AWS over the next decade, securing five gigawatts of compute capacity. [5] Microsoft committed $5 billion alongside a $30 billion Azure spend commitment. Google and Broadcom entered a separate deal for multiple gigawatts. SpaceX agreed to provide 300 megawatts and more than 220,000 Nvidia GPUs. [6]

The firm most explicitly concerned about AI power concentration is infrastructure-dependent on all three major cloud incumbents and the world’s most prominent private launch company simultaneously.

This is not hypocrisy.

Hypocrisy implies the mission is false.

The mission is real. The dependency is also real. They are the same thing, not opposites.

If safety requires staying at the frontier, and the frontier requires hyperscale compute, then the safety mission and the infrastructure dependency are structurally identical. The more seriously Anthropic holds its mission, the more it must depend on the infrastructure controlled by the actors its mission claims to counterbalance.

Sincerity produces the dependency.

The dependency validates the mission.

The circle is load-bearing.

Revenue confirms the mechanism works. In 2022, Anthropic generated roughly $10 million. By April 2026, the run-rate had surpassed $30 billion, and by May 2026, Anthropic said run-rate revenue crossed $47 billion. [7] The fastest revenue ramp in the recorded history of the technology sector, built on a foundation of safety positioning. Enterprise customers want the responsible lab. Governments want the safety-certified vendor. Regulators want the actor whose vocabulary they already speak.

The safety architecture is not incidental to the commercial success.

It is the commercial success.

The Principled Act and the Strategic Act Were the Same Act

In 2025, Anthropic entered a Pentagon arrangement with restrictions on autonomous weapons and mass surveillance. In February 2026, Defense Secretary Pete Hegseth demanded removal of those restrictions and permission to use Claude “for all lawful purposes.” Anthropic refused. The administration cancelled the arrangement and declared Anthropic a supply chain risk. [8]

The refusal was principled.

It was also, structurally, the most consequential brand decision the firm could have made.

It produced moral differentiation from every other frontier lab. OpenAI became a Stargate partner in the Trump administration’s orbit. Google is a search monopoly defendant. Meta is a social harm defendant. Anthropic refused the weapons contract and lost the revenue.

Three months later, Anthropic co-founder Christopher Olah sat alongside Pope Leo XIV at the Vatican Synod Hall as co-presenter of the papal encyclical on AI ethics, Magnifica Humanitas, addressing the protection of human dignity in the age of artificial intelligence. [9] The Catholic Church has 1.4 billion members. Its moral authority extends across every jurisdiction on earth. No competitor can purchase that positioning. No lobbying budget produces it. The Pentagon refusal did.

Washington: policy testimony, White House voluntary commitments, NIST collaboration, UK AISI advisory relationships, EU engagement, $20 million donated to candidates backing AI regulation. [10] Rome: the Pope’s first encyclical on AI, co-presented by an Anthropic founder.

A firm that genuinely believes AI is dangerous and acts on that belief produces exactly the same institutional position as a firm that performs that belief strategically. The market, the regulators, the Vatican, and the public cannot distinguish between them. The only thing that matters is what the position does.

What the position does is accumulate authority over the question of what responsible AI development looks like.

The mechanism does not require bad faith. It only requires consistency.

Nobody Elected Them. Here Is What That Means.

Anthropic did not seize power.

It occupied a vacuum.

Legislatures did not move fast enough. Regulators did not have the vocabulary. International bodies did not have the technical capacity. The institutions that normally produce legitimate standards for dangerous technologies, pharmaceutical regulators, nuclear safety bodies, financial supervisors, took years to develop and decades to acquire credibility.

AI governance needed a framework immediately.

The technology was moving. Governments were being asked questions they could not answer. The public wanted assurance. Investors wanted a signal. Other labs wanted a benchmark that distinguished serious from reckless.

Anthropic supplied the framework.

That is not corruption. That is what happens when private technical capacity moves faster than public institutional capacity. The actor with the relevant expertise, the resources, and the interest in having a standard produced the standard. The standard was adopted because it was the available one. The available one became the authoritative one. The authoritative one became the one everyone is measured against.

This is how governance works when democratic institutions are slow and technical change is fast.

The question “who elected Anthropic?” has the wrong form.

Elections are how democratic polities choose officials who exercise public authority. Anthropic does not exercise public authority. It exercises something more slippery: the authority to define what public authority should be asking for.

That is harder to contest, harder to remove, and harder to make legible as a form of power at all. You cannot vote it out. You cannot regulate it away without using its vocabulary to describe what you are regulating. You cannot replace it without building a competing standard, which requires the kind of long-horizon institutional investment that democratic governments are notoriously bad at making while a more immediate problem is running.

The power Anthropic holds is the power of the institution that arrived first with a framework that worked well enough to be adopted.

The IPO Is the Honest Moment

On June 1, 2026, Anthropic confidentially submitted a draft S-1 to the Securities and Exchange Commission. The filing came four days after closing a $65 billion Series H at a $965 billion post-money valuation, surpassing OpenAI for the first time. [11]

The IPO is not the end of the safety mission. It is the moment the safety mission becomes fully legible as an integrated institutional architecture.

An S-1 requires disclosing what the RSP actually commits Anthropic to in language that public investors can use to assess growth risk. It requires characterizing what happens to revenue if a model crosses an ASL threshold that triggers a pause. It requires disclosing the $100 billion AWS commitment in terms that allow assessment of dependency. It requires accounting for the fact that the firm’s largest compute partner is also one of its largest investors.

Securities law does not care about mission statements. It requires disclosure of material facts.

The RSP, which has always contained the escape clause permitting continued development when others are not restraining, will appear in a document alongside the revenue trajectory that depends on not restraining. The two grammars, mission and capital, have always been the same grammar. The IPO just requires them to appear in the same font.

What will not appear in the S-1 is the thing this essay is about.

The authority to define what safe means. The authority to assess whether its own models meet the standard it defined. The authority to revise the standard in response to conditions it is best positioned to evaluate. The authority, accrued through the Vatican, the White House, the AISI, and five years of policy testimony, to speak as the responsible party when governments need someone to speak.

None of that is a material fact in a securities filing.

All of it is what the near-trillion-dollar valuation is pricing.

The Answer

Nobody elected Anthropic. That is accurate and incomplete.

The more useful observation is this: the question of who should define safety standards for the most consequential technology being built has not been answered democratically. It has been answered by default, by the institution that moved first, built the vocabulary, occupied the regulatory vacuum, accumulated the legitimacy network, and turned a principled safety commitment into the architecture through which everyone else’s safety is assessed.

That institution is a private company that will shortly be publicly traded.

Its founders are sincere. Its research is real. Its safety commitments are more substantive than any competitor’s. None of that changes what the architecture does.

A genuine commitment to safety that also functions as the primary source of institutional authority over the definition of safety is still a concentration of unelected power over a question that affects everyone.

The RSP is not a leash. It is the document that made Anthropic the institution that writes the leash for everyone else.

Anthropic did not build safety to constrain its power.

It built power through safety.

The difference between those two sentences is the distance between a mission and a governance strategy.

At near $1 trillion, they are the same document.

Notes

[1] Anthropic RSP version history: v1.0 September 2023, v2.0 October 2024, v2.1 March 2025, v2.2, v3.0 February 2026, v3.1 April 2026, v3.3. Each version produced and published by Anthropic.
https://www.anthropic.com/responsible-scaling-policy

[2] RSP v2.2 states the policy was designed “in the spirit of” the UK Frontier AI Taskforce responsible capability scaling proposal, the EU AI Act obligations for systemic-risk models, U.S. government proposals, and the METR RSP framework. It also states the policy “helps satisfy” the Voluntary White House Commitments (2023) and Frontier AI Safety Commitments (2024).
https://www-cdn.anthropic.com/872c653b2d0501d6ab44cf87f43e1dc4853e4d37.pdf

[3] Anthropic announced Jared Kaplan as Responsible Scaling Officer in October 2024. Kaplan is a co-founder of Anthropic and co-author of the 2020 scaling laws paper.
https://www.anthropic.com/news/announcing-our-updated-responsible-scaling-policy

[4] Anthropic RSP v3.0, February 24, 2026: “If one AI developer paused development to implement safety measures while others moved forward training and deploying AI systems without strong mitigations, that could result in a world that is less safe — the developers with the weakest protections would set the pace, and responsible developers would lose their ability to do safety research and advance the public benefit.”
https://www.anthropic.com/news/responsible-scaling-policy-v3

[5] Amazon committed up to $33 billion total in equity and Anthropic committed to spend more than $100 billion on AWS over the next decade, securing up to 5GW of compute capacity.
https://www.anthropic.com/news/anthropic-amazon-compute

[6] Microsoft, NVIDIA, and Anthropic announced a strategic partnership in November 2025 under which Anthropic committed to purchase $30 billion of Azure compute capacity and contract additional compute capacity up to one gigawatt. Reports also described Microsoft as investing up to $5 billion in Anthropic. Anthropic separately announced a Google/Broadcom agreement for multiple gigawatts of TPU capacity and a SpaceX agreement for more than 300MW of capacity and over 220,000 NVIDIA GPUs.
https://blogs.microsoft.com/blog/2025/11/18/microsoft-nvidia-and-anthropic-announce-strategic-partnerships/
https://www.anthropic.com/news/google-broadcom-partnership-compute
https://www.anthropic.com/news/higher-limits-spacex

[7] Anthropic’s AWS announcement stated that run-rate revenue had surpassed $30 billion, up from approximately $9 billion at the end of 2025. Anthropic’s Series H announcement stated that run-rate revenue crossed $47 billion earlier in May 2026.
https://www.anthropic.com/news/anthropic-amazon-compute
https://www.anthropic.com/news/series-h

[8] TechPolicy.Press timeline of the Anthropic-Pentagon dispute, citing Reuters and WSJ reporting, describes the dispute over military-use restrictions, Anthropic’s concerns about autonomous weapons and mass surveillance, and Hegseth’s February 2026 demand that Anthropic allow use “for all legal purposes.” AP later reported that Anthropic lost a $200 million contract and was labeled a supply chain risk after refusing compliance.
https://www.techpolicy.press/a-timeline-of-the-anthropic-pentagon-dispute/
https://apnews.com/article/d5fbaee17ee0bdb9738dbb808ea2d047

[9] Anthropic stated that Chris Olah was invited to speak at the May 25, 2026 Vatican presentation of Pope Leo XIV’s AI encyclical, Magnifica humanitas. National Catholic Reporter reported that Pope Leo XIV and Olah appeared at the Vatican Synod Hall during the encyclical’s presentation.
https://www.anthropic.com/news/chris-olah-pope-leo-encyclical
https://www.ncronline.org/vatican/vatican-news/pope-leo-anthropic-co-founder-call-church-tech-ethics-partnership-magnifica

[10] Religion News Service, May 22, 2026: “Anthropic, by contrast, has donated $20 million ahead of the 2026 midterm elections to a U.S. political group backing candidates who endorse AI regulation.”
https://religionnews.com/2026/05/22/why-anthropic-is-helping-unveil-the-popes-new-encyclical-on-ai/

[11] Anthropic announced on June 1, 2026 that it confidentially submitted a draft S-1 registration statement to the SEC. Anthropic announced on May 28, 2026 that it had raised $65 billion in Series H funding at a $965 billion post-money valuation.
https://www.anthropic.com/news/confidential-draft-s1-sec
https://www.anthropic.com/news/series-h