August 6, 2025. SoftBank Group closed a $10 billion margin loan. The lenders: Goldman Sachs, JPMorgan Chase, Mizuho Securities, Apollo Global Funding, and Sumitomo Mitsui Banking Corporation. The term: two years. The drawdown: this month. The collateral: SoftBank’s shares in OpenAI.
I tried to find the smart contract. There isn’t one.
In DeFi, this position would exist as a public record. I could inspect the borrow function, the collateral balance, the health factor, the liquidation threshold, and the oracle address. I could fork it and simulate a 50% drawdown before breakfast. None of that is possible here. The terms are dense legal prose in a signed agreement that nobody outside a conference room has fully read. Code doesn’t lie. But this deal was never written in code — and that is the point. The largest collateralized AI bet in history is running on a ledger nobody can audit.
Let me be precise about the structure, because the headline hides the mechanism. A margin loan is a loan secured by securities. The borrower keeps ownership of the collateral but loses the practical freedom to sell it, and if the collateral falls, the lender can issue a margin call demanding more assets or cash. This is routine machinery in traditional finance. The novelty is the collateral class. OpenAI is not listed. It has no daily mark, no order book, no closing price, and no options market to hedge exposure. The most recent public tender offers valued the company at roughly half a trillion dollars, but those numbers are bilateral negotiations between one buyer and one seller, controlled by the company and its advisors. There is no continuous price discovery. In DeFi language, this loan’s oracle has enormous update latency, a sample size of one, and no consensus.
SoftBank’s path into the position began in October 2024, when it participated in OpenAI’s blockbuster funding round, and continued through early 2025 via tender-offer purchases from existing shareholders and employees at accelerating valuations. Alongside the direct stake, SoftBank is the financial engine behind Stargate, the AI-infrastructure joint venture with OpenAI and Oracle. The entity borrowing against OpenAI is not a diversified financial institution. It is an AI-concentration vehicle wearing a technology conglomerate’s balance sheet.
The banks, for their part, see a lending opportunity dressed as a secured trade. They underwrite the loan, collect fees, and book a position that appears low-risk because it is “collateralized.” In crypto, collateralized lending means the mechanism is visible, the collateral is priced by open markets, and liquidation is deterministic. Here, collateral means something softer: a negotiated valuation, a legal promise, and a shared belief that the AI trade will keep working.
Let me walk through the mechanical problems in this position, in the order they will surface.
The Collateral Has No Feed
When I audited the Uniswap V2 factory contract in 2020, I found an integer overflow in the liquidity token minting logic that automated scanners missed. The bug hid in plain sight: everyone was staring at swap math, nobody checked the mint path. I took a permanent lesson. The dangerous flaw is usually in the asset everyone treats as safe.
This loan has the same shape. The variable everyone is pricing is OpenAI’s valuation. The variable nobody can inspect is OpenAI’s liquidity. Private equity carries transfer restrictions, rights of first refusal, vesting conditions, and lockups. If OpenAI reprices down 40% tomorrow, the collateral cannot be sold to a willing buyer at any price, because there is no public bid. The banks own paper, and paper is only as good as the next tender offer another banker arranges. In a crisis, that tender offer does not arrive. It vanishes.
This is the difference between a liquidation-protected lender and a litigation-protected lender. In DeFi, the lender is protected by code: the moment a price feed crosses a threshold, a liquidator repays the bad debt, takes the collateral, and exits. The cycle takes under three minutes. Speed is the only shield in a flash loan. The banks have no equivalent. Their shield is a contractual clause in a memorandum signed by people whose paychecks depend on never exercising it.
The LTV Is a Banker’s Mark
The largest undisclosed number in this trade is the loan-to-value ratio. A $10 billion loan against SoftBank’s OpenAI stake implies an LTV somewhere between 25% and 40%, depending on the mark the lenders used. Neither number is public. Both are banker’s marks, not market prints.
In my experience reading raw Etherscan transactions, I distrust any metric that cannot be independently recomputed. I audit the logic, not the hope. The logic here is circular. The same banks that lend against OpenAI’s value are arranging the tender offers that establish the marks securing their own loan. The party that sets the price is the party that lends against the price, and the borrower is the party that buys at the price. Everyone has an incentive to mark high, and nobody has an incentive to mark low until a margin call is already underway.
The interest makes the circularity worse. A loan this size prices near SOFR plus a spread, landing between $700 million and $1 billion per year. Over two years, SoftBank is paying roughly $1.5 billion to rent $10 billion. OpenAI pays no dividends. OpenAI burns cash. The interest must come from future fundraising, future tender sales, or a rising mark that lets SoftBank borrow even more later. The loan is serviced by the narrative it relies on.
The Liquidation Engine Is a Lawsuit
I remember May 2022 exactly. When Terra collapsed, I didn’t panic sell. I moved my remaining stablecoins into overcollateralized DAI on MakerDAO and accepted the lower yield. The deeper lesson was about machinery. UST’s stability mechanism had a theoretical design: arbitrageurs would sell LUNA, buy UST, restore the peg. It worked until the moment it mattered. When the peg broke, inventory was insufficient and the arbitrageurs were gone. A mechanism without inventory and without willing counterparties is not a mechanism. It is a hope.
This loan has the same failure mode. The lenders’ liquidation engine is litigation. If SoftBank misses a margin call, the banks do not have a bot waiting to sell OpenAI shares into a liquid order book. They have a right to sue, a right to demand better collateral, and the privilege of waiting years for a court to sort it out. Algorithms don’t panic at a margin call. Bankers do.
The contrast with my own trading is instructive. In 2021, I ran a flash-loan arbitrage script between SushiSwap and Uniswap and extracted about $14,500 over three weeks. The edge came from the mechanism: one atomic transaction, no legal friction, no counterparty delay, no negotiation window. This margin loan runs in the opposite direction. The gap between the marked value and the recoverable value is permanent by design. The only party who can make the position whole is a borrower whose incentive, at the moment of distress, is to delay every step. That is not a liquidation engine. It is a negotiation.
Correlation Is the Hidden Collateral
Name the variable no term sheet can fix: correlation. SoftBank is not simply a shareholder in OpenAI. It is an entity whose AI thesis concentrates in OpenAI, whose infrastructure commitments flow through Stargate, and whose public portfolio includes Arm, a chip supplier that trades on the same AI assumptions. The margin loan is not independent of SoftBank’s balance sheet. It is the same bet, triple-levered.
In 2022 I learned what correlation risk does. I thought I was diversified because my stablecoin yield sat in several protocols. They fell together anyway, because the underlying risk was a shared assumption that liquidity would stay cheap. This loan has the same structure. OpenAI equity, SoftBank capital, AI infrastructure, and the banks’ exposure all sit inside one assumption: the AI asset class continues to appreciate.
If that assumption cracks, the banks do not simply write down OpenAI. They write down SoftBank’s creditworthiness, Arm’s public equity, and the Stargate commitments all at once. The margin call arrives at the exact moment the borrower is most impaired. That is not a bug. It is the design of leveraged finance. It is also why “guaranteed returns” is the first phrase I filter out of any pitch.
Why the Banks Take the Trade
The answer is not conviction. It is fee structure. Underwriting a $10 billion loan to a name like SoftBank generates tens of millions in fees, plus annual interest, plus the positional prize: Goldman and JPMorgan become the banks that handle OpenAI’s eventual IPO and SoftBank’s next financing. A margin loan is a relationship product dressed in secured-trade clothing.
In late 2023 I allocated capital into EigenLayer restaking, targeting EigenDA, and spent weeks manually verifying slashing conditions. What I found was that complexity had outrun the advertised security model. Incentives became unclear, and I exited half the position. The pattern repeats here. The banks are running models on an asset whose security model — the ability to sell, reprice, and reclaim in default — has never been tested at this scale. Complexity outpaces security models every time.
OpenAI is a remarkable company. None of this is a bet against the product. It is an examination of the container. The wrapper here is a two-year loan with a margin call and no public oracle. The banks do not need to be right. They need to collect fees while the outcome is undecided. If the collateral doubles, they collect. If it halves, they collect — then they negotiate.
The Two-Year Roll Is the Real Event
The last structural detail is maturity. Two years from August 2025, this loan must be repaid or refinanced. That is the true liquidation date. Everything before it is re-pricing theater.
Run the scenarios. If OpenAI’s valuation has declined, the banks renew at a lower LTV and demand capital repayment SoftBank must source elsewhere. If OpenAI has grown and generated real revenue, the roll is trivial. If OpenAI has listed, the exit is a market trade — the only outcome where the collateral becomes genuinely liquid.
Notice what the two-year term does to bank risk. It is short enough to preserve re-pricing power. The banks are not locking in a twenty-year view. They are renting exposure, and at maturity they reserve the right to reprice the entire relationship. This is the TradFi equivalent of a health factor that updates once at maturity, instead of every block.
Rollover failure, not outright default, is what kills leveraged institutions. I am not predicting SoftBank’s insolvency. I am pointing out that the loan’s risk does not live in the interest payment, or even in the margin call. It lives in the quiet moment in 2027 when someone has to state aloud what OpenAI shares are worth, with no feed, no order book, and no bid — only a number negotiated in a boardroom. Arbitrage is just patience wearing a speed suit. The real arbitrage in this trade will belong to whichever lender correctly prices the exit before the market notices it is unpriceable.
The Signal the Market Will Misread
The obvious read: the largest banks in the world are validating OpenAI, and SoftBank’s credit is strong enough to borrow $10 billion. That is the retail read. The alternative is not bearish on OpenAI; it is a read on the signal.
When a sophisticated borrower chooses a margin loan over an equity raise or a conventional bond, it is choosing the structure with the least price discovery. An equity raise forces SoftBank to sell discounted shares and dilute Vision Fund returns. A bond forces public markets to price SoftBank’s credit. A margin loan forces only a private negotiation. The structure is the message: the borrower values opacity. The banks’ participation is not bullishness; it is fee neutrality. They get paid in every outcome, so their signature proves nothing about conviction.
The sharper observation is for crypto natives. This is what treasury teams did in 2021 when they borrowed against their own illiquid protocol tokens and called it financial engineering. The framing is different. The mechanics are identical: borrow against a token you cannot sell, mark it to a thesis, and cover the carry with further appreciation. Crypto has a name for that pattern: farming your own collateral. The banks have reinvented the self-referential loan, upgraded it to ten figures, and put it in a law firm’s filing cabinet instead of on a chain.
The useful question is not whether OpenAI succeeds. It is what the exit looks like if the mark stops rising. I audit the logic, not the hope. The logic of any leveraged position is only as strong as its exit path, and this one has no price, no venue, and no date certain. Trust the stack, verify the exit.
Watch two data points. First, the next OpenAI tender mark, and whether it prints above or below the banks’ internal valuation. Second, SoftBank’s credit default swap spread. If both move in the wrong direction before the 2027 maturity, this loan becomes the canary — not for AI, but for the entire leveraged architecture of the AI trade.
When $10 billion comes due and the only bid for the collateral is another loan, someone will finally have to answer the question SoftBank never asked: who buys the position if the thesis stalls? The banks already know. That is why they insisted on the margin call clause. And that is why the margin call, when it comes, will arrive at the worst possible moment.