Anthropic's IPO Asks You to Fund Someone Else's Exit

Anthropic has filed for an initial public offering, and the most interesting detail is the one buried in the plumbing: the company is considering letting existing shareholders sell stock in the deal. That means some — perhaps a lot — of the money raised in this IPO wouldn't go to Anthropic at all. It would go to the people who already own Anthropic, as a cash-out.

It's a distinction that sounds small until you follow the money.

Primary. Secondary. Same ticker. Different destination.

An IPO has two parts, and they do completely different things. The primary portion sells newly created shares — the company issues more stock, the money goes to the company's bank account, and that cash funds growth, compute, hiring, whatever. The secondary portion sells shares that already exist — the company creates nothing, the money goes to the seller, and the buyer simply becomes the new owner of a slice Anthropic already issued.

Both types of shares trade at the same price under the same ticker once the IPO is done. But in the offering itself, one side builds the company and the other side enriches its shareholders.

Anthropic is reportedly weighing a secondary component, which would break from the SpaceX IPO earlier this year — a primary-only deal that let no existing shareholders cash out; instead following CoreWeave and Figma, both of which included seller exits alongside new capital.

No one has confirmed the size of the secondary tranche, or whether sellers would include early investors, executives, employees, or some mix. But the mere possibility changes what the IPO actually is for the person on the buying side. If the secondary tranche is large, you're not just buying into Anthropic's future. You're also underwriting someone else's exit.

And you're doing it at a price that assumes a staggering valuation.

The number that makes no sense until it does.

Anthropic's last private valuation was $965 billion, set in May during a $65 billion Series H round led by Altimeter, Dragoneer, Sequoia, and others. Bankers are reportedly discussing an IPO target of $2 trillion or more. Goldman Sachs and JPMorgan, by contrast, are modeling a valuation in the $400 to $500 billion range.

That is not a tight spread. It is a gap between what investors want the story to be and what the numbers currently support.

The revenue is real and moving fast. Anthropic hit $10.9 billion in Q2 2026, up from $4.8 billion in Q1 — more than doubling in a single quarter. The run rate was about $47 billion by mid-2026, up from roughly $9 billion at the end of 2025. The company posted its first operating profit in Q2, which covers salaries, compute, and research costs.

But operating profit is not net profit. Anthropic is not yet making money on a bottom-line basis. Cumulative losses are substantial, and the company remains in the heavy capital-expenditure phase of running a frontier AI lab. To justify a $2 trillion valuation at even generous public-market multiples, Anthropic would need to post annual net profits somewhere in the $70 to $80 billion range. That is a number only a handful of companies on earth generate. Even Amazon, which Fortune used as the comparison, took more than two decades to get there.

The revenue trajectory is impressive. The gap between $10.9 billion a quarter and $70 billion in annual profit is just about the entire future of the business compressed into a single number.

What you buy vs. what you control.

Here's where the governance structure adds another layer of strangeness — or rather, clarifies it.

Anthropic's co-founders, including CEO Dario are preparing to receive a new class of stock with enhanced voting rights. Amodei holds roughly 2% of the company by economic stake. That 2% would carry disproportionate voting power through a dual-class share structure. The purpose, reported by The Information, is to let founders maintain control and avoid being ousted by outside shareholders — the same architecture used by Google, Meta, and Snap.

But Anthropic also has something neither of those companies has: the Long-Term Benefit Trust (LTBT). The LTBT is a three-member body — now including former Federal Reserve Chair Ben Bernanke — that holds power over board appointments and is meant to ensure the company stays on its stated mission of "responsible development of advanced AI."

Public shareholders would buy economic exposure to Anthropic's cash flows but have almost no say over how the company is run. The founders control the vote. The trust controls the board. The public controls neither.

This is not inherently fraudulent — it's just what dual-class structures do. They separate economic ownership from control. But it's worth being honest about what that means for a buyer. You're not buying a vote. You're not buying a voice. You're buying a financial claim on a company run by people who don't need your permission to do anything.

The LTBT adds a veneer of mission-aligned governance on top of that. A trust of three respected people, watching over the AI safety mission. In practice, the trust has appointed only one of its planned board seats so far, and its actual enforcement power — what happens if management diverges from the mission — is uncertain. It's a promise, not a mechanism, at this point.

The compute commitment no one gets to vote on.

While the IPO is being prepared, Anthropic just signed a six-year, $45 billion deal with Nscale to rent roughly 460 megawatts of compute capacity in West Virginia. The deal uses Nvidia's Vera Rubin chips and is separate from Anthropic's existing commitments with Google and Amazon.

$45 billion in committed infrastructure spend. Locked in before the company has even shown it can sustain profitability at that scale. The board appoints through the trust. The founders control the vote. The public would have no say in the decision that just committed more cash than most countries' annual defense budgets.

This is what a fully separated ownership-and-control structure looks like in action.

What happens to the margin.

One of the more encouraging details is on the economics side. SemiAnalysis reported that Anthropic's inference margins — the portion of each API dollar left after compute — jumped from roughly 38% a year ago to about 70%. That's a genuine structural improvement: as models get more efficient and volume scales, the cost per inference drops faster than the price.

But Anthropic also lowered its 2025 gross margin projection to 40%, citing higher inference costs even as revenue skyrocketed. The margin story is real but still being written, and it may not stabilize at 70% if compute costs scale faster than efficiency gains. The $45 billion Nscale deal suggests compute spending is accelerating, not decelerating.

The investor question.

So what does a retail investor actually face here?

You'd be buying into a company with explosive revenue growth that has not yet proven sustained profitability, at a valuation that ranges from $400 billion on the conservative end to $2 trillion on the aspirational end — a five-fold difference that reflects belief, not calculation. A portion of the IPO proceeds might go directly to existing shareholders exiting at a massive gain (early money went in at valuations below $20 billion, two years ago). The people who control the company will have super-voting shares and a mission trust. You'll have a ticker.

That's not a recommendation one way or the other. It's the machine, as it stands. The revenue growth is extraordinary. The valuation uncertainty is extraordinary. The separation between what you buy and what you control is extraordinary.

The Anthropic IPO is interesting because it concentrates every question about how to price AI into one offering. But for the person buying shares on day one, the first useful question isn't whether Claude is the best model or whether Anthropic will win the AI race. It's whether you're comfortable funding someone else's exit while buying a voiceless claim on an unproven profit story at the highest end of every valuation debate in the market.

That's not a moral question. It's a plumbing question. The answer tells you what you actually bought.