margin loan
SoftBank raised another twelve billion against OpenAI last week. The number is not the story. What the loan is secured by is.
On 15 September SoftBank paid down $25.9 billion, the remaining balance on a $40 billion bridge it took earlier this year to fund its OpenAI position. In the same stretch it closed a new two-year facility of $11.87 billion from about twenty banks — upsized from a $10 billion target, which means the book was oversubscribed. Bloomberg and the Japan Times carried the terms. Sitting alongside those is a $10 billion margin loan secured by the OpenAI stake itself, and reporting on a possible overseas bond sale of as much as $20 billion.
Every write-up led with the size. The size is the least interesting thing here.
The structure is migrating. A bridge loan is corporate debt: the lender is underwriting SoftBank, a listed company with a balance sheet you can read. A margin loan against a specific stake is asset-backed: the lender is underwriting OpenAI, through a pledge. Those are different bets wearing the same word.
Here is the arithmetic that makes the second kind of bet strange.
By October, SoftBank’s cumulative investment in OpenAI is expected to reach $64.6 billion for roughly 13% of the company. Divide it out and you get an implied valuation near half a trillion dollars.
Now ask the ordinary question a lender asks: who set that price?
OpenAI is private. There is no tape, no continuous quote, no short interest, nothing that marks the position between rounds. The mark is the last primary round. And the last primary rounds are rounds in which SoftBank was among the largest checks — which is to say, among the prints that cleared the price.
So the collateral on the margin loan is valued at a number that the borrower’s own buying helped establish. Not exclusively. Not improperly. But materially.
I want to be careful here, because the cynical version of this sentence is wrong and it is the version that will get repeated. This is not fraud, and it is not unusual. Every private mark in the world works roughly this way: the last round is the price until there is another round. Venture capital has always been an asset class where the buyer helps set the valuation of the thing they already own. Lenders know it. They haircut for it.
The difference is not kind. It is scale, concentration, and what the mark is being asked to do.
Hans Vaihinger gave us the right word for this in 1911, in Die Philosophie des Als Ob, and it has never been more useful than it is in private markets.
Vaihinger’s subject was the fiction: a construct we know to be false, or at least unproven, that we adopt because reasoning with it produces results. Mathematicians use the infinitesimal. Jurists use the legal person. Physicists used the frictionless plane. None of these are lies, because a lie is told by someone who wants you to mistake it for the truth. A fiction is held knowingly, by both parties, as an instrument.
His condition — the whole of his ethics on the matter — is that you must not forget. The fiction works as long as everyone remembers it is one. The failure mode Vaihinger names is when the as if hardens into an is, when a useful construct is promoted to a belief and people start deducing from it as though it were a fact about the world.
A private valuation is a textbook Vaihingerian fiction. Everyone in a board meeting knows the mark is a convention. It is useful for accounting, for option strikes, for narrative, for the next round. Nobody mistakes it for a price at which half a trillion dollars of stock could actually change hands, because no such transaction exists or could.
A margin loan is where the fiction stops being held as a fiction. Collateral is not a convention. It is a number attached to a covenant, and a covenant is a mechanism that does something when the number moves. The moment you pledge a privately marked asset, you have promoted an as if to an is, and you have done it in a document that will be enforced literally.
Which leaves the question of what marks it down.
This is the part of the structure that I find genuinely novel, and I have not seen it laid out. In a normal margin arrangement, the collateral trades. It gets repriced every second by people with opposing opinions, and the margin call is the mechanical consequence of a price that a market produced. The discipline is external and continuous.
Here, the collateral does not trade. Between rounds it cannot fall. Structurally, the only ways this mark moves are a down round, an impairment written by an auditor, or a lender’s own internal revision — and all three are discretionary acts by people who would rather not perform them, performed slowly, and performed by parties with a position.
So the collateral is stable in the way that a thing with no thermometer is always room temperature.
There is exactly one continuously priced security in the entire structure, and it is SoftBank’s own equity. It fell as much as 13% on Monday, the sharpest drop since late June, on nothing more than frontier-lab executives publicly worrying about AI in the same week. Bloomberg tied the move to that.
Read that against the private mark and the picture is clean. The listed instrument, the one thing in the stack that has to answer to a market every day, is the one thing that moved. The unlisted asset it is levered against did not move, because it has no mechanism for moving.
That is not evidence the private mark is wrong. It is evidence that we would not learn it was wrong from here.
Otto Neurath’s image is the one I keep coming back to for the refinancing itself. We are sailors who must rebuild our ship on the open sea, never able to dismantle it in dry dock and construct it from the best materials. You replace a plank and you stand on the rest.
That is what 15 September was. Retire $25.9 billion, raise $11.87 billion, keep the mast up. There is nothing wrong with rebuilding at sea; it is the only way anything gets rebuilt. The condition is that you replace one plank at a time and the rest holds.
The stack now has a plank that is held up by the plank it is holding up.
The bet may well be right. OpenAI’s revenue is real and growing at a rate that makes half a trillion arguable rather than absurd, and Masayoshi Son has been correct about exactly this sort of thing before, expensively and at length. This is not a call that the position is bad.
It is a call that nobody in the structure is in a position to find out.
Sources: Bloomberg: SoftBank gets upsized $11.9bn loan ↗ · The Japan Times ↗ · Tech Startups: loan detail and 13% stock drop ↗ · Business Standard ↗ · Vaihinger, The Philosophy of ‘As If’ (1911) ↗ · Neurath’s boat ↗