On August 6, SoftBank Group signed a $10 billion margin loan against its shares in OpenAI. The two-year facility is underwritten by Goldman Sachs, JPMorgan Chase, Mizuho Securities, Apollo Global Funding, and Sumitomo Mitsui Banking Corporation. Drawdown is planned for this month. I have audited this trade before. Not on Wall Street. In DeFi. Borrow against volatile collateral. Mark it to a narrative. Wait for the leverage to exceed what the market will accept.
The difference is not the leverage. It is the visibility. In DeFi, a liquidation is executed by open-source code, published on a public ledger, and available for audit by anyone who cares. In this deal, the liquidation parameters live in a confidentiality clause. The collateral is a private company's shares with no continuous market. The “market price” is a committee's conviction. This is not a SoftBank story. It is a systemic one.
SoftBank's OpenAI position is substantial. Through Vision Fund vehicles and direct investment entities, the group has accumulated a stake reportedly worth tens of billions at OpenAI's latest marks. That mark has been one of the most aggressive upward revisions in venture history: roughly $29 billion in early 2023, $157 billion in October 2024, and reports of a $300 billion conversation in recent syndicate discussions. Each of these values was set in a private round. In several of those rounds, SoftBank was a leading buyer.
Why borrow rather than sell? Selling a block of private shares into a secondary market would force price discovery. It might trigger a markdown. It would cap the dream. A margin loan lets SoftBank extract liquidity from paper appreciation without letting the shares trade. The lenders are not fringe actors. Goldman Sachs and JPMorgan do not underwrite $10 billion facilities without serious internal risk approval. Mizuho, Apollo, and SMBC add balance sheet breadth. The structure is a margin loan, not a plain secured term loan. That means the collateral is subject to mark-to-market and maintenance calls.
And that is where the logic starts to break. A margin loan requires a market. OpenAI shares have a restricted secondary market, negotiated between accredited parties, with no order book, no closing price, and no public tape. So when the banks mark the collateral, what exactly are they marking? They are projecting. They are estimating what a desperate seller might receive if the position needed to be monetized during a crisis. That estimate is not a price. It is an opinion.
The first red flag is provenance. OpenAI's valuation is not independent of SoftBank's participation. SoftBank has helped finance OpenAI's rounds, and now it borrows against the asset those same rounds inflated. This is a closed loop of confidence. In 2022, I spent six months modeling the death spiral of UST. The core flaw was circular dependency: UST's value relied on Luna's, Luna's on UST's, and the entire architecture on new buyers arriving to break the cycle. The mechanics were algorithmic. This loan's circularity is narrative, but the structure is the same. OpenAI's value is supported by strategic capital inflows, not by distributed cash flows. Revenue is real and growing, reportedly crossing a $10 billion run-rate in 2025. But the cash cost of inference compute, model training, and talent scales with that revenue. Profits are an expectation, not a premise.
I do not mean to dismiss OpenAI's engineering. The technology is remarkable. What I am flagging is the capital structure. A $10 billion loan is a concentrated bet on a profitability path that has not yet booked sustained net income. That is not collateral. It is a swap of one conviction for another. My MakerDAO work taught me the same lesson in 2020: the riskiest collateral is the asset that looks the safest during a bull run. KNC looked liquid, supported by a respected team and strong on-chain volume. The manipulation vector was the single Chainlink oracle feed. The asset's own market was fine until it was not. Here, the fragile pointer is not an oracle contract on Ethereum. It is a marks committee inside five banks.
Standard margin-lending math suggests this facility was sized conservatively. At a 30 to 40 percent loan-to-value ratio, the pledged OpenAI stake would need to be worth $25 to $35 billion to support a $10 billion draw. That ratio looks safe on a spreadsheet. But loan-to-value protection fails when the collateral cannot be sold at the marked price. Illiquid assets do not decline in smooth decrements. They gap. In March 2020, even US Treasury bonds traded through every model. A private AI stock in a forced sale would fall to a buyer's number, not a banker's.
Imagine the cascade. OpenAI's next internal mark drops by a third. The banks issue a margin call. SoftBank can post more collateral, pay down principal, or default. But SoftBank's balance sheet is already stretched across Arm, its Vision Fund obligations, and its substantial debt load. The shares get surrendered. The five banks now hold restricted stock in a private AI company. They cannot dump it into a public market. They can only sell in private blocks, at an undisclosed discount, in a process that produces no public price. The entire risk lifecycle moves in the shadows.
I have written critically about DeFi oracles for years. In 2020, I flagged the risk in Maker's KNC feed: one point of manipulation from a cartel. But a decentralized oracle is transparent. You can inspect its parameters. You can model its failure. This loan has no oracle. It has a loan agreement. The banks' internal valuation groups will argue quarterly about what OpenAI is worth, referencing comparable private rounds that are themselves confidential.
The borrower will argue back. SoftBank has its own marks, its own LP reporting requirements, and its own conviction about AGI's future. The banks are not independent price machines. They are counterparties protecting a relationship that spans decades and multiple facilities. This was the same issue I raised in my 2024 critique of the spot Ethereum ETF filings: custody and verification are treated as legal questions when they are actually operational ones. The banks hold a security interest, but the shares sit on a cap table controlled by an issuer that never consented to the pledge. OpenAI can restrict transfers. It can litigate. It can simply refuse to cooperate on a future round that would set the next mark. Complexity hides risk in the legal layer as much as in the pricing layer.
None of this is a knock on margin loans per se. Leverage against private equity has existed since the first leveraged buyout. The novelty here is the concentration and the narrative. Ten billion dollars of debt against a single private AI position, pledged as if it were a treasury, is the kind of structure that ends badly in every historical cycle. The 2008 lesson was not about collateralized debt generally. It was about collateral that looked diversified but was, in fact, correlated at the moment of stress. AI equities, AI tokens, and AI compute companies will all trade down together if the rate regime or the narrative shifts.
The tokenized RWA crowd will point to this deal as proof that private assets can collateralize institutional credit. I read it differently. This facility works precisely because it is off-chain, opaque, and relationship-based. The participants can pretend the marks are real. On a public ledger, that pretense faces the scrutiny of market participants who can short the asset and arbitrage the gap. If you take this exact structure and place it under a smart contract, it breaks. Not because of the code. Because of the underlying illiquidity. Sharding is easy; consensus is hard. Moving this loan on-chain would not make it safer. It would only make the risk visible.
Now the uncomfortable part. What have the trade's supporters gotten right? First, SoftBank's OpenAI stake is not vaporware. OpenAI holds a defensible technical position, real revenue, and the deepest institutional relationships in AI. Converting an illiquid asset into liquidity without triggering a taxable sale is rational treasury management for a conglomerate that has been forced to sell assets into bad markets before. This might be the smartest financing SoftBank has negotiated in a decade.
Second, the lenders are not passive. Five banks underwriting $10 billion means the underwriting was adversarial, the same way I reverse-engineer a protocol's tokenomics before I publish a teardown. They have access to OpenAI's financials through the loan terms. They will demand covenants. They can ask for additional collateral. Their downside protection may be genuine. Third, and this is the point I carry with me, this deal creates a lane for tokenized private credit. If a private equity position can credibly secure an institutional facility, then RWA protocols have a real reference point. The problem was never collateral. It was price discovery. The banks have the staff to pretend price discovery exists for a two-year window.
So I will apply my own rule: audit the code, not the pitch. The “code” in this deal is the loan agreement, the covenants, the transfer restrictions, the events of default, and the final marks mechanism. If that document holds, this can be a template for intelligent private-asset leverage. If it does not, the failure will be quiet, correlated, and felt across every fund, protocol, and regulator pretending AI valuations are a market instead of a meeting.
The next major shock will not be a smart contract exploit. It will be a margin call on an asset that cannot be sold at the number on the spreadsheet. I have spent twenty-seven years watching leverage rebuild itself in new packaging. This SoftBank facility is a new package. The collateral is real. The underwriting is sophisticated. The risk is the same one that brought down Terra, that nearly broke Maker, and that every bull market sells as optional: the belief that paper wealth converts to cash whenever someone rings the bell. Bell-ringers are not listed on term sheets. When the AI cycle turns, the test will not be the technology. It will be whether a private company's shares, marked by a committee and held by five of the world's largest banks, can survive the cold, factual request for repayment. Trust no one. Verify the liquidation path.