Anthropic at $850bn, or: the round before the round

The reporting, to be clear about what we have and don't have, is this: TechCrunch and Reuters say Anthropic has received unsolicited preemptive offers to raise $40–50bn at a valuation between $850bn and "$900bn+", with a board decision expected at a May meeting. Revenue run rate is said to have crossed $40bn, against roughly $9bn at the end of 2025. Some early investors are reportedly holding paper for an anticipated October IPO rather than selling into the preempt.¹

There is no filing. There is no term sheet on a wire. There is a company that has not confirmed the round, a number that is a range with a plus sign on the end of it, and a board meeting in the future. I want to be careful, because the gap between "preemptive offers received" and "round closed at this valuation" is exactly the gap where structured financings hide, and the press release version of these things tends to round in one direction.

Still, the shape is interesting enough to walk through.

The revenue line, first. A jump from ~$9bn ARR to >$40bn in roughly sixteen months is the kind of number where you want to read the disclosure rather than the headline, and there is no disclosure. ARR at frontier labs is a famously elastic concept: it can mean trailing-month annualised, it can mean last-week annualised, it can include committed but unbilled enterprise contracts, it can include compute resold through a hyperscaler partner at margins that would make a CFO wince. Anthropic's revenue is heavily intermediated by AWS and Google Cloud, which means the question of whose revenue it actually is, gross versus net, who books the inference margin, is non-trivial. I'd want to see the rev-rec policy before I'd build a model on the $40bn number. The directional claim that Anthropic's revenue has scaled hard through 2026 is consistent with everything else we can see (Claude code-gen share, enterprise seat penetration, API volume from agent frameworks); the specific multiple is a thing to hold lightly.

The valuation, second. $850–900bn against a $40bn run rate is roughly 21–22x revenue. That is, by the standards of the last cycle, not absurd for a company growing at this rate; by the standards of any prior software cycle, it is a number that requires inference economics to bend in Anthropic's favour for a sustained period. The implicit bet in the multiple is that gross margin on Claude inference is durable, that the $40bn isn't being bought with a $35bn compute bill it can't reprice. This is the inference economics frame doing its work: at frontier labs, the binding question isn't top-line growth, it's whether the unit economics on a token of output are stable, improving, or compressing. Anthropic doesn't disclose this. The hyperscaler partners do, obliquely, in their own segment reporting, and the picture there is that AI infrastructure revenue is growing faster than AI infrastructure margin, which is a polite way of saying somebody is eating cost.

The structure, third, and this is where I'd watch. "Preemptive offers" at this scale, in the months before a planned IPO, are not quite a primary round in the ordinary sense. They look more like a pre-IPO crossover, the kind of financing where late-stage funds, sovereigns, and crossover public investors get an allocation at a price that anchors the IPO range. The function of such a round is partly capital-raising and partly price discovery: you want a $850bn private mark on the cap table when you file the S-1, because it makes the IPO range defensible. The fact that some early investors are reportedly holding for October rather than selling into the preempt is the tell. If the preempt were the exit, they'd take it. They're holding because the IPO is expected to clear at a higher mark, and the preempt is the floor, not the ceiling.

This is a coherent strategy. It is also a strategy that depends on the IPO window staying open, on the revenue line continuing to compound through Q2 and Q3, and on no frontier-lab pricing event between now and October that re-rates the sector. Each of those is a real risk; the structure is built to convert paper gains into the IPO mark, which means the company and its existing holders are increasingly long the IPO date.

The frame that I think actually fits best is performativity. A round at $850bn, even if it closes at $700bn, even if revenue is half what's reported, is a capital event large enough to reshape the market it's measured against. It sets the comp for every other lab raising in 2026. It gives Anthropic a balance sheet that closes the compute-access gap with OpenAI's Microsoft relationship and with Google's in-house position. And it does this regardless of whether the underlying product economics support the multiple, because the spend the round enables, datacentre commitments, talent, long-dated compute contracts, is itself the thing that makes the multiple defensible ex post. The round funds the capex that justifies the round. This is not a criticism; it is the structural feature of frontier-lab financing in this cycle, and Anthropic is now executing it at the same scale OpenAI executed it eighteen months ago.

The safety-as-positioning frame is doing quiet work in the background here too. Anthropic's enterprise narrative, the RSP, the constitutional AI lineage, the defence-adjacent posture they've leaned into through 2025, is part of why the preempt exists at this price. It's the differentiator that lets them sell to risk-averse buyers who won't single-source on OpenAI, and it's the story that institutional investors find legible. "Safety-forward frontier lab" is a positioning, and at $850bn it is being priced as a durable one. Whether it survives contact with the revenue mix once the S-1 discloses customer concentration is a separate question.

What to watch.

• The May board outcome and, more importantly, whether any disclosure accompanies it. A confirmed round with a confirmed lead and a confirmed structure is one data point; a confirmed valuation range with no lead is another.
• The S-1, when and if it lands. Specifically: rev-rec policy, customer concentration, gross margin on inference, the structure of the AWS and Google Cloud agreements, and any committed-but-unbilled disclosure that would let a reader reconstruct the $40bn number.
• The compute commitments that follow. A round of this size will be partly recycled into multi-year datacentre and chip contracts; the counterparties and tenor of those deals will tell you what Anthropic thinks its inference cost curve looks like through 2028.
• Whether OpenAI re-marks. A $900bn Anthropic mark in May puts pressure on OpenAI's last private valuation. The response, a tender, a secondary, a fresh primary, is the next move in the same game.

The round itself is not yet a thing that has happened. The market it implies is.

Footnotes

¶