I keep returning to a single number, and it refuses to sit still.
According to the account now circulating, a researcher named Jacob Coxon joined Anthropic in July 2026 and resigned two months later. He walked away, the story goes, from unvested stock options — and before he left, he told the world that the two most valuable private companies on earth were quietly wagering human survival on a race toward self-improving superintelligence. David Sacks, seated in a government technology advisory role, then called for Anthropic's initial public offering to be paused until the allegations were investigated.
Except the calendar says none of it has happened yet.
That is the first thing any honest analyst must notice, and it is the thing almost no one wants to say out loud. We are not reading news. We are reading something that behaves like news — a document with dates, titles, quotations and institutional weight — that cannot be true on its own timeline. In my years of auditing contracts and reading market narratives from the inside, I learned that the anomaly is never noise. In the code, I found the ghost of the architect. Here, the ghost is the date.
The story is single-sourced, its central accusation is withheld, and its chronology runs past the present. And yet it moved through the channels — the timelines, the group chats, the pre-market whispers — with the fluency of something already true. That gap, between the verifiability of a claim and the speed of its pricing, is the actual subject of this brief. Not whether Anthropic is dangerous. Not whether Sacks is compromised. But how a market learns to value a warning it cannot verify, and what that tells us about the next decade of capital formation in frontier technology.
The Pipeline That Learned to Price Whispers
To understand why this story landed the way it did, you have to understand what an IPO has become in the age of narrative capital. It is no longer the moment a company meets the public. It is the last act of a long private performance, and the audience has already decided the ending long before the curtain.
Anthropic, per the account, has filed confidentially and is being discussed at a valuation approaching one trillion dollars. I want to sit with that number because it is the sort of number that should stop a room. For most of the company's short life, its public valuation lived in the tens of billions. A move toward a trillion is not growth. It is a re-rating — a claim that this entity should be priced less like a software company and more like a sovereign utility, or a currency, or a reserve asset.
I have watched this exact move before, in a different asset class, and I remember how it ended.
In 2017, at twenty-four, I was a junior researcher at a boutique security firm in Zurich, auditing smart contracts during the ICO boom. One of my assignments was a post-DAO successor project called Project Aether. I found a reentrancy vulnerability that put five hundred ether — about two point one million dollars at the time — at risk. I wrote it up with the care I had been trained to apply. The frontend team rejected the report for being, and I quote, too academic. They shipped anyway. The bug was never exploited, or was never noticed being exploited, which in this business is not the same thing.
What I learned there was not a lesson about Solidity. It was a lesson about how narratives price risk. Aether's valuation was not derived from its contract logic. It was derived from the story its community told itself about what the contract meant. My report threatened the story, so the story absorbed the report. This is the mechanism I want to map onto Anthropic, because the shape is identical: a valuation that floats on belief, and a technical warning that the belief cannot metabolize without cracking.
The difference between 2017 and now is only the sophistication of the instrument. A token sale let retail price a dream in real time. A confidential IPO filing lets institutions price a dream in private, through pre-IPO secondaries and allocation chatter, before anyone signs an S-1. But the underlying physics are the same. Narrative is the collateral. Cash is downstream of it.
The Anatomy of a Safety Brand, Priced
Here is where I have to be careful, because the temptation is to treat 'AI safety' as a marketing adjective. It is not. It is a capital structure.
Anthropic built its differentiation on a specific, defensible claim: that it takes alignment seriously in a way its competitors do not, that its Constitutional AI work and its research culture constitute a kind of institutional immune system. That claim is not decoration. It is the asset. It is what allows sovereign funds, pension allocators and conservative enterprise buyers to underwrite a technology they privately fear. Safety is the moat. The moat is the valuation.
So when a whistleblower says — as the account claims — that the people inside these companies genuinely believe the technology could end human life within a decade, and that they are racing anyway, he is not making a moral observation. He is issuing an impairment notice. He is walking into the room where the moat lives and saying the water is a painting.
This is why the reaction was so fast and so total. A safety-branded company does not have the same damage function as an ordinary one. If a normal enterprise gets accused of reckless behavior, the market discounts the fine, the legal cost, the reputational drift. If a safety-branded enterprise gets accused of reckless behavior, the market does not discount a cost — it revalues the brand itself. The product was trust. The trust is now marked to a question mark.
The same researcher is said to have accused both Anthropic and OpenAI, which is important. A double accusation is not symmetrical in effect. It is worse for the company whose entire premium is safety, and comparatively less damaging for the company whose premium has always been capability and velocity. When you sell acceleration, a claim of recklessness is a rounding error. When you sell restraint, a claim of recklessness is a category error. The market is being asked to decide which story is now mispriced — and it is being asked with incomplete information, on a timeline that does not close.
The Cost of Walking Away
There is one detail in this account that I trust more than any other, and it has nothing to do with artificial intelligence.
The researcher, we are told, forfeited unvested stock options on the way out.
In mechanism design, this is the closest thing we have to a proof of sincerity. Unvested equity is a hostage. It is the firm's way of saying: your future is ours. When someone surrenders that hostage voluntarily, they are converting economic value into signal. They are saying the thing they believe is worth more than the thing they could have had.
I have spent enough time decompiling failed protocols to know how rare that is. In the 2022 collapse, I worked remotely from Auckland through the wreckage of Three Arrows Capital's related assets, and I read hundreds of vesting schedules, cliff dates, insider unlock cliffs. Everyone stayed for the options. Almost no one left on principle, because the options were the principle. When the pool empties, only the intent remains — and in a bear market, you find out very quickly whose intent was real.
So a forfeiture is a costly signal, and costly signals deserve weight. I want to give it that weight. But I also want to point out what the signal does and does not prove. It proves the person was willing to pay. It does not prove the person was right. The history of markets is littered with sincere believers who were sincerely wrong, and with sincere believers who were sincerely early, which in a mark-to-narrative system is indistinguishable from wrong until the mark finally clears.
This is the part that makes me uneasy about how the story was consumed. A forfeiture is being read as evidence of the allegation's truth. It is evidence of the allegation's cost to the accuser. Those are different coins, and the market is spending them as if they were interchangeable.
When Politics Enters the Order Book
The Sacks intervention is the piece of this account that should concern allocators most, and not for the reason the headlines suggest.
Consider what it actually is. It is not a regulator acting through process. It is not the SEC opening a file, issuing a comment letter, or delaying a registration statement under a published rule. It is a political figure, occupying an advisory position inside the government, publicly calling for a specific company's offering to be paused pending an investigation that, on the record we have, does not formally exist.
I have written extensively about this pattern in crypto, where projects preach decentralization while team wallets and foundation holdings remain perfectly traceable. In that world, the DAO is often not a governance structure at all — it is a compliance shield, a legal and rhetorical vehicle that lets a small group of insiders hold control while presenting the appearance of distributed legitimacy. The gap between the story and the cap table is the whole trick.
The same gap now appears in AI capital markets, except inverted. There, the shield is not decentralization. It is security. And the person pulling the lever is not a token holder but a state actor with a microphone.
What matters for pricing is not whether Sacks is right. What matters is that a single political statement can, in principle, move the timeline of the largest private listing in the history of technology. If that is true, then the correct valuation model for Anthropic is no longer a discounted cash flow. It is a political risk model. And political risk models, as every sovereign debt trader knows, price discontinuities — not growth rates.
I would want to know, before I allocated a dollar, what Sacks's relationships are. I would want the full disclosure of his investment positions, his advisory interests, and any entanglements with the competitive landscape. This is not cynicism. It is the same diligence I applied when I modeled yield farming incentives in 2020 and concluded that the governance was an illusion. You look for who benefits from the incentive structure, and you do not accept the stated reason as the operating reason.
The Valuation That Has No Denominator
Let me be direct about the trillion-dollar figure, because it is the softest number in the entire account and the one most people are treating as settled.
A near-trillion valuation implies a story of extraordinary revenue growth, an unprecedented compute position, and a durable enterprise franchise — API revenue, enterprise subscriptions, cloud distribution through Amazon and Google. None of that is provided in the account. There is no revenue figure, no margin, no burn rate, no customer concentration, no compute commitment schedule. We are told the number and asked to feel its weight.
I have seen this before, and I mean this literally. In the summer of 2020, I was a mid-level analyst at a crypto-native fund in Singapore, modeling the yield mechanics of Compound and Uniswap across more than ten thousand on-chain transactions. The headline APYs were astronomical. They were also, structurally, marketing. The number was real in the sense that it existed on a dashboard, and fake in the sense that it could not survive the removal of the incentive that produced it. I wrote a paper called The Illusion of Decentralized Governance predicting that token incentives would concentrate power rather than distribute it. It got fifty thousand views and a citation from a major outlet. The market ignored it until the crash.
I did not enjoy being right. I went to a cabin in New Zealand for two weeks because the cognitive dissonance of being correct and unheard is its own kind of exhaustion.
What I took from that period is a habit I now apply to every valuation story, including this one: ask what the number would be if the narrative were removed. Strip the safety premium. Strip the sovereign-utility framing. Strip the existential-risk drama that makes the equity feel historically significant. What is left? A model company with real revenue, real customers, real compute costs, and a competitive landscape that is brutal. The residual is probably a very good business. It is almost certainly not a trillion-dollar one.
That does not mean the valuation is wrong. It means the valuation is a statement about belief, not a statement about cash flow. And when belief is the collateral, the collateral is only as good as the next headline.
The Whistleblower as an Asset Class
There is a structural observation buried in this account that I think will matter more over the next five years than anything specific about Anthropic.
Frontier AI companies are now in a position where their most credible internal critics are also their most dangerous counterparties. This is new. In traditional finance, a compliance officer who quits is a footnote. In frontier AI, a safety researcher who quits and speaks is a systemic event, because the company's entire public contract rests on the claim that such people are listened to.
I spent part of 2021 in London with a collective of women digital artists, minting a curated set of one hundred generative avatars on Ethereum and running the community Discord myself. The project sold out in fifteen minutes and raised three hundred thousand dollars. What I watched happen afterward taught me more about institutions than any balance sheet. The community had been built on a story about ownership and identity and mutual respect. The moment value arrived, the story became a liability. The people who cared most about the original meaning became the people the structure most needed to manage. To own a piece of art is to inherit its narrative — and inheriting a narrative means inheriting the fight over what it means.
The same dynamic is now embedded in AI labs, except the stakes are not a floor price. They are the terms of a governance pact with the public. A safety researcher is a living assurance. Their continued presence is a disclosure. Their departure is a restatement.
This is why I take the underlying worry seriously even while I distrust the wrapper it arrived in. The existential-risk narrative is not a technical claim, and I will not pretend otherwise. Self-improving superintelligence is today a research frame and a thought experiment, not a shipped capability. Current frontier models are static weights plus inference-time compute; they do not rewrite their own parameters in any reliable loop. There is no benchmark, no FLOP count, no internal safety memo in the account before us. So the technical confidence of E-grade material cannot support a strong claim in either direction.
But the governance claim is different. Whether or not the specific allegation is true, the structure it exposes is real: commercial pressure and safety assurance living inside the same legal entity, with the assurance being the thing that must not be contradicted. That tension is not fabricated. It is architectural.
The Blind Spot Everyone Is Standing In
Here is the contrarian move, and I want to make it precisely.
Everyone is arguing about whether the whistleblower is right. Almost no one is arguing about whether the whistleblower needs to be right for the story to work.
In a narrative-priced market, the truth of a claim and the price of a claim are separate variables, and the price moves first. This is not a defect of markets; it is their grammar. The moment the Sacks statement entered circulation, the IPO timeline acquired a probability distribution it did not previously have. That change in distribution is real, regardless of whether any of the underlying facts ever verify. Uncertainty itself is a cost. Optionality on a delay has value to someone. And in the interval between the claim and the confirmation, there is a tradeable window.
That window is where the interesting money lives, and it is also where the ethics get murky. Because a story like this — single-sourced, timeline-inconsistent, unverifiable — is perfectly shaped for the very thing the crypto market industrialized: the pre-truth. The rumor before the news. The position before the correction.
The deeper blind spot is this. If safety is Anthropic's premium, then the company is permanently short its own candor. Every safety researcher who leaves is a partial put option on the valuation. Every internal disagreement that leaks is a repricing event. The brand that justifies the moat is also the brand that makes the moat fragile. You cannot sell restraint and then be accused of abandoning it without paying a nonlinear price. The more seriously you market your conscience, the more violently the market reprices when the conscience speaks.
And here is the part that should bother the industry most: none of this requires bad faith from anyone. The researcher may be entirely sincere. Sacks may be entirely sincere. The company may be entirely sincere. And the outcome can still be that the most safety-conscious lab in the world is structurally more exposed to a single resignation than the least safety-conscious one. That is a perverse incentive, and it is now priced in. Capital markets just learned that safety branding carries a hidden liability — which means the next cohort of founders will think twice before building it. That is the real injury, and no investigation will repair it.
The Disclosure That Comes Next
Let me say what I think actually happens, and what I will be watching.
First, if this pattern holds — political figures commenting pre-listing on private AI companies — then AI safety disclosure migrates from the marketing page to the risk factors. Not a paragraph about responsible scaling. A formal, auditable section in registration documents, with named internal escalation channels and independent verification. The moment that becomes standard, the entire safety narrative becomes a compliance cost, and the premium attached to it compresses. This is the same path crypto governance walked. Slogans became legal engineering. Foundations became vehicles. Decentralization became a filing.
Second, the double indictment is structurally unstable. When a whistleblower accuses two competitors simultaneously, the market will inevitably decide which accusation is load-bearing, and it will make that decision based on brand, not evidence. That asymmetry is not a legal process. It is a narrative settlement, and it will be wrong in at least one direction.
Third, and this is the part I find genuinely new: we are entering the era of the pre-event. A document that describes something that has not yet happened, circulated as though it has, arriving in a bull market where everyone is searching for the next catalyst. I have watched identical consumer NFT collections mint out on identical promises, watched identical protocols fork and re-fork with identical incentives, and now I am watching a trillion-dollar private company get repriced by a story dated in a month that has not arrived. The pattern is not AI-specific. It is how narrative capital has always worked. It has simply found a larger vessel.
The audit is not a check; it is a confession. Every institutional statement about its own safety is a claim about what it would do under pressure, and the only way to test such a claim is for the pressure to arrive. This account is that pressure, arriving early, unverified, and possibly imaginary. Which may be precisely why it is instructive.
What I Am Watching, and What I Am Not
I am not watching the valuation. The valuation is a story, and stories are cheap to tell.
I am watching three things. The first is whether any formal regulatory process ever attaches to these allegations — because a statement from an advisor is weather, and a filing is climate. The second is whether Anthropic's safety organization visibly changes shape in the coming months, and how the change is described; identity is a protocol, and soul is the private key, and you learn a great deal about an institution by which of the two it protects. The third is whether the phrase 'AI safety' survives its own disclosure regime, or whether it becomes, like liquidity mining before it, a number that looked real until the incentive was removed.
There is a version of this story in which everything is verified, the IPO is delayed, the sector reprices, and safety becomes a line item. There is another version in which nothing is verified, the story evaporates, and the episode is remembered as noise. Neither of those versions matters as much as the third: the version in which the industry quietly learns that candor is expensive, that a conscience is a liability, and that the safest thing a frontier lab can do is say less.
That is the outcome I fear most, and it requires no villain. It requires only a market that prices a warning faster than it can verify it. When the pool empties, only the intent remains — and I am no longer sure we will like the intent that this pricing mechanism leaves behind. So let me leave you with the question I keep circling: if a single unverifiable warning can reset the timeline of the largest listing in history, what exactly is it that we are buying when we buy the future?