The Prospectus Problem: How an IPO Forces Frontier AI to Disclose

Somewhere in a law firm right now, a securities lawyer is trying to write a sentence that has never been written before. The sentence has to explain, to an ordinary investor with no technical background, what could go wrong with a frontier AI model. It has to be specific enough to satisfy the Securities and Exchange Commission. It has to be honest enough to survive discovery in a lawsuit filed three years from now by shareholders who lost money. And it has to be written by a company whose entire founding premise is that the technology it sells might be dangerous.

That sentence is going into Anthropic's prospectus.

On June 1, 2026, Anthropic confidentially filed a draft registration statement with the SEC, according to reporting from Fortune. The company had raised roughly $65 billion four days earlier at a valuation near $965 billion. The listing is targeted for as early as this fall. Goldman Sachs, JPMorgan, and Morgan Stanley are the names attached to the underwriting.

Most of the coverage has treated this as a finance story: the size of the raise, the eye-watering multiple, whether the AI trade holds up long enough to get the deal out the door. Those are real questions. They are also, I think, the least interesting thing about this filing.

The interesting thing is that an S-1 is not a marketing document. It is a confession under penalty of law. And nobody has ever made a frontier AI lab write one.

The three-year failure of voluntary disclosure

To see why registration matters, you have to be honest about what came before it.

Since 2023, essentially every mechanism for extracting information out of frontier AI labs has been voluntary. There were the White House commitments. There were the model cards and system cards. There were the responsible scaling policies and preparedness frameworks. There were the safety institutes with memoranda of understanding. There were the summit declarations.

I want to be careful here, because a lot of that work was done in good faith by serious people, and some of it produced genuinely useful artifacts. Anthropic's own responsible scaling policy is more specific than it had to be. That is worth something.

But none of it was enforceable. Every single one of those documents was written by the company being described, published at a time of the company's choosing, in a format of the company's design, with no liability attached to being wrong. A model card that overstates safety testing has no legal consequence. A framework that quietly loosens its own thresholds between versions has no legal consequence. If a lab decides a capability evaluation is commercially sensitive and omits it, there is no filing deadline, no auditor, and no plaintiff.

This is the part of the AI governance conversation that I think has been consistently undersold. We spent three years arguing about what labs should disclose while the actual disclosure regime remained: whatever they feel like, whenever they feel like it.

The gap between those two columns is not a matter of degree. It is a change in kind. One of them is a press release. The other is evidence.

What is invisible right now

The best argument that registration matters is to look at how much is currently unknown about a company at the center of the world economy.

Consider what the two largest investors in Anthropic have been booking. In the first quarter of 2026, according to Fortune, roughly $28.7 billion of Alphabet's $62.6 billion in profit came from marking up its stake in Anthropic. That is about 46 percent of the quarterly profit of one of the largest companies in the world, generated not by selling anything, but by revaluing a position in a private company. Amazon booked $16.8 billion in pre-tax gains in the same quarter from its Anthropic investments — reportedly more than half of its pre-tax income.

Now ask a simple question: what exactly do they own?

The honest answer is that nobody outside the cap table knows. Reported estimates put the Amazon position in the mid-to-high teens as a percentage, composed of something like $42.2 billion in convertible notes and $32 billion in nonvoting preferred stock. The conversion terms on those notes have not been disclosed. Google is reported at roughly 14 percent, contractually capped at 15. These are estimates assembled by journalists from fragments of public filings.

So we have a situation where two of the most widely held stocks on earth are reporting roughly half their quarterly profit from an asset whose ownership percentage is a journalistic estimate and whose conversion mechanics are private. Every index fund in America owns this. Every retirement account touches it.

The Amazon pre-tax split above is approximate — the reporting says "more than half," and I have drawn it as a bare majority, which is the conservative read. That imprecision is not a flaw in the chart. It is the entire point. I am drawing a chart about the earnings of a three-trillion-dollar company using a number I had to infer from an adverb.

The valuation nobody can pin down

It gets worse when you try to establish the simplest fact in the story: what is this company worth?

Here is what the public record offers. Google is reported to have committed up to $40 billion, with $10 billion of it going in at a $350 billion valuation. Amazon's stake was described in April as worth about $74 billion at a $380 billion valuation. By late May, a Series H round reportedly priced the company at $965 billion. Bankers are said to consider a debut above $1 trillion the base case. Meanwhile a separate set of reports put the valuation "in funding discussions" around $900 billion.

Look at that chart honestly. It moves from $350 billion to $965 billion across a handful of reported marks, some of which describe different instruments at different times, none of which is an audited number, and all of which are being used right now to generate tens of billions of dollars in reported profit at public companies.

I am not suggesting anything improper is happening. Mark-to-market accounting on a private position is legitimate, common, and rule-governed. When a new round prices a company, existing holders update their carrying value. That is how it is supposed to work.

But notice the structure. Google and Amazon invest in Anthropic. That investment sets a new price. The new price marks up the stake they already held. The markup becomes profit on their income statement. As a Columbia Business School adjunct put it to Fortune, this lets investors "control or influence the value of one of their own assets" — booking gains, in Fortune's phrasing, "without Anthropic ever having to pay them a dollar."

What registration actually forces

Here is the part that I think is genuinely underappreciated: an S-1 is not a narrative. It is a form.

Regulation S-K prescribes what has to be in it. Item 105 requires risk factors — a discussion of the material factors that make an investment speculative or risky, written in plain English, organized under headings, and specific to this company rather than generic boilerplate. The SEC has spent years pushing issuers away from laundry lists toward actual specificity, and the staff issues comment letters when a filing hides behind generalities.

Item 404 requires disclosure of related-party transactions. Item 303 requires management to discuss the results of operations and known trends and uncertainties. The financial statements have to be audited by an independent firm. Revenue concentration has to be described. Material contracts get filed as exhibits.

And then it never stops. A public company files a 10-K every year and a 10-Q every quarter, and its chief executive and chief financial officer personally certify, under Sarbanes-Oxley, that the filing does not contain an untrue statement of material fact.

Every one of those steps is routine. Thousands of companies have been through it. That is exactly why it is interesting: the machinery is boring, tested, and completely indifferent to the fact that the issuer thinks it is building something civilizationally important.

The risk factor nobody has had to write

Now put the pieces together.

Anthropic is a Delaware public benefit corporation. Its certificate of incorporation names, as its public benefit purpose, the responsible development and maintenance of advanced AI for the long-term benefit of humanity. Its directors are permitted — required, in the balancing sense — to weigh that purpose against the financial interests of stockholders.

On top of that sits the Long-Term Benefit Trust: five financially disinterested members, holding no equity, with the authority to elect a portion of the board that grows over time until it constitutes a majority. The trustees include Neil Buddy Shah, Richard Fontaine, and Mariano-Florentino Cuéllar, with former Federal Reserve chair Ben Bernanke reported to be joining. The likely offering structure gives public buyers a single class of common stock while the trust retains its Class T shares and its escalating rights over board composition.

Read that again with a securities lawyer's eyes.

You are asking public investors to buy common stock in a company where an independent body of five people who own none of the stock can eventually elect a majority of the directors, and where the directors are chartered to weigh humanity's long-term benefit against the stockholders' financial interests. No technology company has gone public with a structure like this.

Every honest version of that risk factor says something close to: we may take actions, including halting or restricting our most commercially valuable products, that materially reduce the value of your investment, and you will have limited ability to stop us.

That sentence is remarkable. Not because it is surprising — it is a fair description of what Anthropic has said about itself since 2021 — but because saying it in a prospectus makes it operative. It becomes the thing the company told the market. Which means the market prices it. Which means that from the moment of listing, there is a number attached to how much investors discount a mission lock.

We have never had that number. We have had three years of argument about whether safety commitments are real, conducted entirely in the register of vibes and credibility and whether you trust Dario Amodei. An IPO replaces that argument with a spread.

And it cuts the other way too, which is the part safety advocates should sit with. Once the mission lock is a disclosed risk factor, it is also a disclosed liability in the ordinary corporate sense. Every future decision to slow down becomes a decision the company made after warning shareholders it might. That is protective. But every future decision to not slow down — to ship the thing, to loosen the threshold, to reinterpret the policy — becomes a data point in a securities case about whether the safety commitments were ever more than marketing. Plaintiffs lawyers are going to read the responsible scaling policy very differently than the AI governance community does. They are going to read it as a promise to the market.

That chart is a claim, not a measurement, and it is deliberately provocative. My argument is that the ordering has quietly inverted. The mechanism most likely to force a frontier lab to make a true, specific, costly-to-retract statement about its own capabilities in the next eighteen months is not the AI Act, not an executive order, and not a summit communiqué. It is Item 105 of Regulation S-K, written in 1982, which has no opinion about artificial intelligence whatsoever.

Related parties, in the technical sense

The circularity I described earlier has a name in securities law, and the name has a form number.

Item 404 of Regulation S-K requires disclosure of transactions with related persons — including any holder of more than five percent of a class of voting securities — where the amount involved exceeds a threshold and the related person has a material interest. The financial statements themselves, under accounting standards, require related-party disclosure of the nature of the relationship, the dollar amounts, and the terms.

Anthropic has reportedly committed to spend on the order of $100 billion with Amazon for compute over roughly a decade, and has a compute arrangement with Google described in the range of gigawatts alongside an investment of up to $40 billion. Amazon resells Claude through Bedrock. Both companies are large holders. Both are suppliers. Both are channels.

I want to be precise about my claim here, because it would be easy to overreach. I am not alleging that Anthropic revenue is fake, or that the compute deals are shams, or that anyone has done anything wrong. The commercial logic of these arrangements is entirely legible: the labs need compute at a scale only a few companies can supply, and those companies want exposure to the labs. Of course those deals exist. Of course they are large.

My claim is narrower and, I think, harder to dispute: right now, no outside party can measure the loop. Not because anyone is hiding it, but because there is no obligation to describe it and therefore no standard description of it. An S-1 creates the obligation. After it, the question stops being a matter of opinion and starts being a matter of arithmetic.

That is what disclosure regimes do. They do not make people honest. They make dishonesty checkable.

The clock problem

There is a real cost here, and pretending otherwise would be cheap.

A private company with a long-horizon mission can absorb a bad quarter in silence. A public one cannot. Every ninety days, Anthropic will have to stand up and explain its numbers to people who are, in the main, not thinking about transformative AI in 2032. They are thinking about the next print.

The schematic above is the honest version of the bargain. Disclosure goes up. Latitude goes down. Those move together, and anyone selling you an IPO as pure transparency gain is not being straight with you.

The specific mechanism to worry about is not dramatic. It is not that a board will vote to abandon safety. It is that the reporting cadence slowly reweights what counts as a decision worth making. A capability evaluation that would have taken six weeks takes four, because the launch is in the quarter. A threshold gets reinterpreted rather than tripped, because tripping it is an 8-K. The responsible scaling policy does not get repealed; it gets lawyered, in the way that everything at a public company eventually gets lawyered.

This is the same dynamic I wrote about in the free sample, on how AI token pricing is engineered to feel cheap: once a number becomes the thing you are measured on, the number starts shaping the behavior rather than describing it. Quarterly earnings are the most powerful version of that force ever constructed.

So the fair summary is: registration buys us a large, permanent, enforceable increase in information, and it costs the company some of the freedom that made its commitments credible in the first place. Both halves are true. I think the trade is worth it, but I want to be clear that it is a trade.

The Etsy warning

We have actually run this experiment before, at smaller scale, and the result should worry anyone who thinks a mission structure survives a listing just because it is written down.

Etsy was a certified B Corporation from 2012. It went public in 2015 carrying that certification as a core part of its identity — the socially responsible marketplace, the antidote to Amazon. B Lab, which issues the certification, required Delaware companies to convert their legal form to a benefit corporation in order to keep it. Etsy had a deadline of December 2017 to do so.

It did not convert. In 2017, after a rough stretch and a management shakeup that brought in a new chief executive under pressure from activist investors, Etsy gave up the B Corp certification rather than change its corporate structure. The stated reason was that conversion was complicated and untested for an existing public company. Both things can be true at once: that was a real legal concern, and it arrived at precisely the moment the company was under pressure to prove it would prioritize returns.

That is the base rate you should carry into this. When a mission commitment and shareholder pressure meet on a public company balance sheet, the historical record is not encouraging.

But the disanalogy matters, and it matters in Anthropic's favor. Etsy's commitment was a certification issued by an outside nonprofit — a badge that could be surrendered by declining to renew it, with no vote and no charter amendment. Anthropic's commitment is the corporate form itself. It is already a Delaware public benefit corporation; the public benefit purpose is in the certificate of incorporation; and the Long-Term Benefit Trust holds hard governance rights over board composition rather than a logo.

Surrendering that is not a matter of letting a certification lapse. It requires amending the charter, which requires the board — a board the trust can increasingly elect. Etsy's mission lock was a sticker. Anthropic's is structural.

Which is exactly why the question of whether the trust reaches the public document intact is the one worth watching. Etsy shows you what happens when the lock is weak. Anthropic is the first real test of whether a strong one holds.

The steelman

Let me argue against myself, because the thesis has real weaknesses.

Objection one: risk factors are boilerplate and everyone ignores them. This is substantially true. Most risk factor sections are defensive sludge, written to inoculate rather than inform, and most investors skip them. Fair. But the sludge is still actionable sludge — it is the boilerplate that plaintiffs compare against internal documents in discovery. And the SEC staff has been actively pushing issuers toward specificity. A first-of-its-kind structure will not clear review with generic language.

Objection two: the SEC does not know anything about AI. Also substantially true, and it does not matter as much as you would think. The staff does not need to evaluate whether a model is dangerous. It needs to evaluate whether the company's description of its own process is complete and not misleading. That is a competence the agency has. It has been reading engineering-heavy filings from biotech and semiconductors for decades.

Objection three: this only binds one company. True today. But listings are contagious in a specific way: once a comparable exists, everyone else gets measured against it. OpenAI is reported to be looking at fall 2026 as well. The moment one frontier lab has audited financials, disclosed related-party terms, and a public risk factor about capability, the absence of those things at a competitor becomes a data point rather than a default.

Objection four: they will just take the mission lock out. This is the strongest objection, it has Etsy behind it, and the honest answer is that we will find out. The structural pressure to simplify the governance ahead of a listing is enormous, and bankers hate explaining a five-person trust to institutional buyers. Watching whether the trust survives contact with the underwriters intact — with its escalating board rights unmodified — is the single highest-signal thing in this entire story. If it gets diluted, we learn that the mission was always downstream of the capital, which is exactly what the OpenAI decision to offer Washington five percent suggested about how quickly frontier lab governance reshapes itself around whoever is writing the check.

I hold my thesis at maybe seventy percent. The objections are good. But the asymmetry is what persuades me: on the voluntary path, we get exactly what we have gotten for three years, which is nothing enforceable. On the registration path, we get an imperfect, partial, adversarially-drafted, boilerplate-infected document that is nonetheless the first one anybody can sue over. That is not a close call.

What to watch

If you want to track whether this thesis is right, there are five specific things to look for when the public S-1 drops.

I have written up the specific falsifiable version of this as my prediction on the mission structure appearing as a disclosed shareholder risk, with a target date and validation criteria, because a thesis you cannot lose is not worth much. The news analysis of the filing and the circularity it exposes has the day-of detail and sourcing.

The accident

Here is what I keep coming back to.

For three years, some very smart people have worked very hard on the problem of how to make frontier AI labs tell the truth about what they are building. They convened summits. They drafted frameworks. They negotiated voluntary commitments. They stood up institutes. It was serious, well-intentioned work by people who understood the stakes better than almost anyone.

And the thing that is actually going to produce the first legally binding, externally reviewed, personally certified, plaintiff-enforceable description of a frontier AI company's risks is a securities statute passed in 1933 in response to a stock market crash, administered by an agency with no AI mandate, triggered not by any policy process but by a company's decision that it wanted to sell shares.

Nobody designed this. It is not an AI governance regime. It has no theory of catastrophic risk, no view on alignment, no opinion about whether any of this should be built. It is a disclosure statute that says: if you want other people's money, you have to tell them what you know.

That turns out to be a surprisingly good approximation of what we wanted all along.

It is not enough. It does not touch the labs that stay private, it does not constrain a state actor, it does not evaluate a single model, and it will be gamed by the best lawyers money can buy. Anyone who tells you securities law solves AI governance is selling something.

But after three years of asking frontier labs to please describe themselves accurately, we are about to find out what they say when the question is asked by someone who can make it stick.

The prospectus is going to be the most honest document any frontier lab has ever published. Not because anyone at Anthropic became more honest. Because for the first time, the alternative is a felony.

Further reading:

• When the regulator becomes a shareholder: OpenAI offers Washington 5% — on what happens to frontier lab governance when the entity writing the rules also holds the equity.
• The free sample: how AI token pricing is engineered to feel cheap — on how a number, once it becomes the measure, starts shaping the behavior instead of describing it.
• My prediction on the mission structure as a disclosed shareholder risk — the falsifiable version, with a target date and validation criteria.