Anthropic's Series G, and the $520 billion gap between the leak and the press release

Anthropic announced on Thursday that it had raised $30 billion in Series G funding at a $380 billion post-money valuation. The Financial Times, reporting the same round earlier in the week, said Anthropic had agreed terms on a $30 billion round at a $900 billion pre-money valuation. Bloomberg used the $900 billion figure too. These are the same round. They cannot both be right, and the gap between them is $520 billion, which is roughly one Meta.

This is the kind of thing that, if you only read headlines, you might not notice. The FT headline says Anthropic surpassed OpenAI. The Anthropic press release URL contains the string 380-billion-post-money-valuation. Both statements were published within roughly 48 hours of each other. Both are about the same Series G led by Dragoneer, Greenoaks, Sequoia, and Altimeter.

What the primary document says. Anthropic's own announcement is unambiguous on the headline metric:

Anthropic raises $30 billion in Series G funding at $380 billion post-money valuation.

Standard round arithmetic: $380B post – $30B raised = $350B pre-money. That is a very large number, and it is roughly 39% of the figure the FT was given by people familiar with the matter. I have read the press release carefully and there is no footnote, no share-class qualifier, no "as-converted" asterisk that bridges the two. Anthropic is simply stating a different number than its lead investors appear to be telling reporters.

The most plausible reconciliations, none of them flattering. There are a handful of structural reasons two valuations get reported for one round, and they're worth walking through because the answer determines whether this deal is what the FT says it is.

One: liquidation preferences. If the Series G has a senior preference stack, common-equivalent valuation diverges from preferred-share-implied valuation. A 2.5x senior preference on $30B of new money plus stacked prior preferences could plausibly bridge a gap of this size on paper. But that's a structured financing dressed as a round, and it would make the $900B number nearly meaningless as a market signal.

Two: secondary versus primary. If $30B of primary clears at $350B pre, while a parallel secondary tender for employee or early-investor shares clears at $900B pre, both numbers are true and they describe different transactions. This is increasingly common in late-stage AI rounds. It would mean the FT headline is technically accurate and structurally misleading.

Three: one of the numbers is wrong. The press release is a disclosed document with the company's name on it. The FT and Bloomberg figures are sourced to "people familiar". The base-rate prior on which to trust is not subtle.

The frame: AI performativity, applied honestly. One of the lenses I carry is that the scale of AI capital commitments has become materially influential regardless of underlying delivery — the spend itself shapes the market. The Anthropic round is a near-perfect case. At $350B pre-money, Anthropic is priced at roughly 7.8x reported annualised revenue ($45B), which is rich but defensible for a hypergrowth frontier lab. At $900B pre-money, it is priced at 20x forward revenue, which requires you to believe the 80x token-demand growth figure compounds for several more years without margin compression.

The two numbers are not just two valuations. They are two different theses about what is being bought. The $350B thesis is: Anthropic is a very large enterprise software company growing fast, with a defensible product surface (Claude Code, API). The $900B thesis is: Anthropic is a bet on inference demand being structurally undersupplied through 2028, and the equity is a call option on that supply-demand gap.

I notice that the investors leaking to reporters wanted the second number in the market, and the company writing the press release wanted the first number on the record. That is informative.

Claude Code is the part of this that is actually load-bearing. Strip out the valuation theatre and the most interesting disclosure in the deal coverage is that Claude Code is annualising above $2.5 billion and is responsible for roughly 4% of global GitHub commits. The first number is verifiable in time (developer-tools ARR doesn't hide for long). The second is a methodology question — Anthropic has not, as far as I can find, published how it measures commit attribution, and GitHub has not corroborated. If accurate, 4% of global commits is a structural displacement of Copilot at a speed that ought to be the actual headline.

This is the inference-economics frame doing useful work. Claude Code is a high-token-intensity product (long context windows, agentic loops, tool use), and a $2.5B run-rate from one product means Anthropic is generating real revenue from the workload that consumes the most expensive compute. Whether that is margin-positive at current GPU pricing is the question the press release does not answer and the round does not require it to answer.

What two weeks from outreach to terms tells you. The FT reports the round went from initial outreach to agreed terms in approximately two weeks. For $30 billion at frontier-lab valuations, that is not a diligence process; it is an allocation auction. Sequoia was already in. The other three, Dragoneer, Greenoaks, Altimeter, are crossover funds whose institutional edge is speed of conviction at scale. The signal is not that they did the work fast. It is that the work that mattered (price, allocation, preference terms) was negotiated, and the work that didn't matter to them (independent verification of the $45B ARR figure, gross-margin disclosure, the token-economics model) was not.

This is consistent with a market where the binding constraint on capital deployment is access, not analysis.

What this is a case of. It is the third frontier-lab round in eighteen months where the leaked valuation and the disclosed valuation diverge materially. OpenAI's most recent round had a similar gap between primary and secondary marks. xAI's 2025 raise had a comparable structural ambiguity around preference stacks. The pattern is that frontier-lab valuations are increasingly being reported as headline numbers that describe the most expensive share class in a complex stack, while the company itself discloses something more conservative.

What to watch. Three things. First, whether Anthropic or its lead investors clarify the $380B vs $900B discrepancy in any subsequent S-1-adjacent filing — and the October IPO window, if it holds, will force this. Second, whether GitHub or a third-party developer survey corroborates the 4% commit-share claim. Third, the preference-stack disclosure in the eventual prospectus, which is where the answer to the valuation question actually lives.

Until then, the headline number is whichever one you find more useful, which is exactly the problem.

Footnotes

¶