Five percent of Anthropic is worth $7.2 billion. It is also worth $48.25 billion. It is also worth $75 billion. One sentence of equity, three anchors, a 10.4x spread.
New Constructs' October note implied a $144 billion valuation. Investor channels were circulating a $1.5 trillion number. The announcement itself anchors on $965 billion. Multiply each by 0.05: $7.2 billion, $48.25 billion, $75 billion. Public coverage collapsed all three into "$48 billion."
A pledge whose headline value swings by an order of magnitude depending on the reference price is not a number. It is a timestamp. That is the finding. What follows is the arithmetic behind it, and the three clauses that were never written.
The structure as announced: Anthropic commits 5% of its equity to a philanthropic vehicle, disclosed weeks before a planned IPO and explicitly subject to shareholder approval. A separate July commitment โ $200 million in cash for labor-displacement research โ sits alongside it. Coverage immediately set the figure against OpenAI's 26% nonprofit foundation stake, framing a governance arms race between the two labs.
The framing obscures that these are not comparable instruments. OpenAI's 26% is residual equity retained by a nonprofit after a 2025 restructuring โ passive, historically accretive, constrained by charter. Anthropic's 5% is a newly designed, still-expanding commitment layered on top of a Delaware public benefit corporation and a Long-Term Benefit Trust that holds a special share class electing a portion of the board. Aggregate the mission clauses, the trust, the cash pledge and this new equity and you approach 20% โ but you are adding four legally distinct things and calling the sum a percentage.
Translate it into token terms, which is the only vocabulary I trust for this class of claim. A PBC is a token with a mission clause in the contract. The LTBT is a governance wallet with a near-veto role. The new vehicle is an emissions schedule whose recipient address has not been published. And the $48 billion is a price quote from an illiquid order book.
My first pass here was the pass I ran on 45 whitepapers in Bangalore in 2017: strip the adjectives, reconcile the numbers, find the clause that isn't there.
Start with the vehicle, because the vehicle is the disclosure. The fund is an LLC, wholly owned by the company. Not a 501(c)(3). That single legal choice removes five constraints at once: no tax deductibility for the donor, no IRS Form 990, no 5% minimum annual payout, no self-dealing prohibition, no excess-business-holdings cap. Each omission was a mechanism by which an external party could have forced performance. None of them exist here.
A commitment with no independent trustee, no mandatory distribution, and no public filing is not a commitment. It is an intention with a legal wrapper. In six years of reconstructing failed protocols I have learned that the exploit is almost never in the code that was written; it is in the constraint that was never imposed. The 2020 yield aggregator that drained $30 million had no bug in its swap logic โ it had an oracle feed nobody had audited, and an architecture that assumed the feed would behave. Same pattern, different asset class.
Now the three missing clauses.
First: share source. If the 5% is newly issued, the cost is borne by all post-IPO holders, including the buyers who fund the offering. If it is secondary โ existing holders transferring โ the cost sits entirely with pre-IPO shareholders. That is the difference between a dilution event and a redistribution event, and the announcement does not say which. A buyer cannot price a security whose dilution incidence is undefined.
Second: the baseline. The escalation clause reads "add 2% for every 50% increase above the baseline." Set the baseline at $965 billion and the schedule is: $1.45 trillion triggers 7%, $2.17 trillion triggers 9%, $3.26 trillion triggers 11%. That is a reflexive curve โ a commitment that accelerates with success and has no stated ceiling. In token design I would call it an unbounded emissions function with a positive feedback loop into the treasury. Elegant if you believe the mission scales with market cap. A permanent dilution tax on long holders if you do not.
Third: monetization and voting. Who votes the 5% to 11%? If the company votes it, the company is voting its own shares โ a governance loop with no external check. If an independent trustee votes it, a philanthropic LLC becomes a material voting bloc in a company with a concentrated cap table. Lockup terms and liquidation rules are equally absent. A stake that cannot be sold has no market value; a stake that can be sold in size is a permanent overhang.
Then the accounting, where the headline number dissolves. The LLC is wholly owned, which means it is almost certainly consolidated. Shares move inside the group. No cash leaves. No expense is recognized. On a DCF the effect is not value destruction โ it is float reduction. The $48 billion is a notional figure, not a cash commitment, and the distance from the July pledge is instructive: $200 million of deployable cash against an equity notional 240 times larger that cannot be spent, cannot be deducted, and has no named beneficiaries. That is not generosity. That is a ratio chosen for its optics.
The timing completes the picture. A major governance announcement in the weeks before a filing, requiring shareholder approval, does three things at once: it lowers regulatory friction, it courts long-horizon ESG capital, and it supplies a responsible-AI narrative at exactly the moment price discovery happens. Gas fees are the price of truth, and the cheapest transaction in this entire structure is the announcement.
One layer the coverage missed entirely: index methodology. PBC charter plus LTBT special shares plus a new philanthropic holder is a multi-class structure. FTSE Russell and S&P both apply governance screens to multi-class arrangements. A structure designed to look responsible can quietly disqualify a company from passive inclusion โ a cost measured in basis points of demand, not in press releases.
Six questions would close most of the gap: pre- or post-IPO share count, down-round adjustment, voting control, lockup and liquidation, the approval threshold, and whether the pledge sits in the charter or only in a press release. None are answered. In modeling terms, "unknown" is not a neutral setting.
Here is what the skeptics, myself included, are getting wrong.
The exclusions are precise. The fund is barred from lobbying, political donations, and executive compensation. Those are not random carve-outs; they are the two loudest criticisms aimed at AI labs โ political influence accumulation and insider monetization โ answered in a single sentence. That is competent defensive drafting, not hand-waving.
The LLC choice is rational, not sinister. A private foundation would have subjected Anthropic to the 5% payout rule, self-dealing bans, and public disclosure of every grant. No founder optimizing for mission flexibility would choose that. The question is not whether the structure is unusual. It is whether the operating agreement contains an irrevocable clause. That document is not public.
And the bear anchor deserves its own scrutiny. New Constructs runs a DCF-plus-ROIC model and has been structurally short high-multiple tech for years โ Tesla, Palantir, the whole cohort. Its $144 billion is an extreme scenario, not a median. Citing it as consensus overstates the downside by roughly the amount the $1.5 trillion bull case overstates the upside.
The honest read: this pledge is cheap to make and expensive to verify. That is not fraud. It is an unpriced option written by the company and held by the public, and nobody has marked it.
Watch three documents, not three headlines: the S-1 dilution base, the operating agreement's amendment clause, and the trustee list. If the 5% is new issuance and the operating agreement is amendable by the sole member, the pledge is a marketing line with a vesting schedule that never starts. The rug is not pulled; it was never tied. For a company asking public markets to underwrite a decade of compute spend, that distinction is worth more than $48 billion.