The least reliable number in any AI story
Two reputable trackers currently report Anthropic at roughly $965 billion and at $380 billion. Both are cited as current. The disagreement is not a data-quality glitch — it is what private-market valuation actually is.
One tracker reports roughly $965 billion post-money after a Series H; another reports $380 billion at Series G with $14 billion in annualised revenue. At least one is wrong. More usefully: the question “which one” is less interesting than the question of why a number this large can be this uncertain.
What a private valuation is
A post-money valuation is the price of the most recent share class multiplied by all shares outstanding, as though every share carried the same rights. They do not. Preferred shares in a late-stage round routinely carry liquidation preferences, ratchets and participation terms that make them worth substantially more per share than common stock. Multiplying the preferred price across the whole cap table is a convention, not a measurement, and in a round with aggressive terms it can overstate the company by a wide margin.
Add the ordinary mechanics — rounds that close in tranches, valuations reported at announcement versus at close, secondary sales at prices unrelated to the primary — and two trackers watching the same company in good faith can land far apart.
Why the spread is worse in AI specifically
Rounds are closing faster than tracker update cycles, which means a stale figure and a current figure coexist in circulation for weeks. And AI rounds are disproportionately strategic — investors who are also customers, suppliers or partners — where the headline price may be doing work that has nothing to do with the investor’s view of fair value.
The same distortion shows up further down the market. Shield AI raised $1.5 billion within a $2.25 billion package at $12.7 billion, up 140% in a year, against projected 2026 revenue above $540 million. That is roughly a software multiple applied to a company that manufactures hardware — and hardware businesses have inventory, unit economics and gross margins that software businesses do not.
How to use these numbers anyway
Treat a private valuation as a directional indicator with a wide error bar, not a quantity. It is reasonably good evidence that a company raised, that investors were willing, and roughly what tier the company sits in. It is poor evidence of anything requiring arithmetic — market share, comparative worth, or whether a sector is “overvalued” by some percentage.
And when two trackers disagree, report the disagreement. Picking the more dramatic figure and presenting it as fact is how a convention becomes a number, and a number becomes a thesis.
Cheatsheets — 8 Frontier AI Labs Compared (2026): OpenAI vs Anthropic → · Second Talent — Top 100 AI Startup Funding & Investment Statistics (Aug 2026) → · Crescendo — Latest AI Startup Funding News and VC Investment Deals — 2026 → · The Business Perspective — AI Startups That Raised Funding in August 2026: Big Rounds →