SoftBank Is Refinancing a $40 Billion Bridge With Margin Loans on a Paper Mark. The OpenAI Stake Is a Leverage Stack Now.

Bloomberg had it on Friday afternoon. SoftBank is talking to lenders about a second $10 billion loan against its OpenAI stake, priced roughly 275 basis points over SOFR on a two year term, with Mizuho as the mandated lead arranger. That is three weeks after SoftBank closed the first $10 billion margin loan against the same stake, and two days after a separate Bloomberg piece put a potential $10 to $20 billion bond sale on the table for September. All of it is layered on top of a $40 billion unsecured bridge SoftBank signed in March, which comes due in March 2027.

The headline is the second loan. The story is the shape of the stack.

The Ledger

For scale on the other side of the ledger: SoftBank has publicly committed $64.6 billion in equity to OpenAI across the initial 2024 round, the $22.5 billion follow-on, and the conversion tranches announced in February 2026, for a fully diluted position of roughly 13 percent. At the $852 billion secondary mark set in March, that paper stake is worth about $110 billion. So the borrowing stack is not larger than the mark. It is a large fraction of it, and the fraction grows every time OpenAI closes at a lower number than the last one.

What Is Actually Being Pledged

The subtle line in every SoftBank OpenAI financing headline this year has been the collateral. Pledging equity in a private company is not the same as pledging listed shares. When SoftBank first went to lenders in April for $10 billion secured on OpenAI shares, the process broke on valuation: nobody could agree what a private secondary mark was worth as collateral. The target was cut to $6 billion in May, and only revived at $10 billion in July after SoftBank added a corporate guarantee. That guarantee is the fact under the fact. The lenders on the August 6 loan have recourse to SoftBank Group the parent if the pledged OpenAI shares fall short of the collateral schedule. It is a margin loan in name and a partly recourse loan in substance.

The corporate guarantee is what let the second loan get priced this fast. It is also what turns an OpenAI mark-down into a SoftBank balance-sheet event. Under a pure non-recourse structure, a stake that lost value would just get repossessed. Under a guarantee, a lender that took a mark hit could go looking for cash from the parent, or more precisely from whichever piece of the parent still trades on a public market. That is Arm, mostly. There is a reason the earlier reporting on a $5 billion loan proposed Arm shares as separate collateral.

Why the Bridge Is the Real Deadline

Every piece of this stack points at one date: March 2027. That is when the $40 billion unsecured bridge comes due. The margin loans and the September bond are the tools SoftBank plans to use to get out from under that bridge before the deadline. Read that way, the August 6 loan and the August 28 loan are not additional leverage. They are the first two tranches of the takeout. The bond is meant to be the third.

That framing is friendlier than the leverage headline suggests, and it is the version SoftBank has been quietly encouraging with the banks. It is also the version that only works if two things hold. One, the OpenAI mark does not slip between now and when the bond prices. Bond investors will look at the same collateral schedule the margin lenders did, and a private-company mark that got a haircut in the intervening quarter would reprice the coupon. Two, the OpenAI IPO clears inside the window. The prospectus was filed confidentially in June. The read across the confidential-filing to first-print timeline for other frontier IPOs is nine to fourteen months, which puts the earliest plausible print somewhere between March and August 2027. That is exactly the same window as the bridge deadline. The whole stack is timed against the same event.

What Changes at a Mark Cut

Assume the OpenAI IPO prices at $600 billion instead of the $850 billion the March secondary implied. That is not a bear scenario. That is a modestly conservative one, the kind that shows up in a diligence report when the underwriter compares AI revenue multiples to the mature software cohort. At $600 billion, the 13 percent stake is worth roughly $78 billion, and the collateral coverage on the August 6 loan falls from about eleven times to about eight times. That is still comfortable, on paper. It is not comfortable if the corporate guarantee triggers a margin call at a lower ratio, and most margin loans do.

Assume $400 billion. Now the stake is worth about $52 billion, the two $10 billion margin loans plus a $20 billion bond stack to $40 billion against $52 billion, and the coverage ratio flirts with breach. The bridge is not paid off in that scenario. It gets rolled again, at a higher coupon, or it eats into Arm proceeds. Neither is a solvency event for SoftBank. Both are the kind of thing that shows up as a Vision Fund line write-down in the November quarterly, and shows up in the OpenAI cap table two quarters later as a distressed sale of a preferred position.

Where This Sits in the Broader Financing Stack

The reason to stack this piece against the ones we ran on the Ohio guarantee and the circular equity loop is that it is the third face of the same architecture. Nvidia agreed to backstop up to $105 billion of the Piketon build via its Ohio guarantee. Google recycled its $40 billion Anthropic equity into $200 billion of TPU commitments on the compute side. Now SoftBank is borrowing against its own OpenAI mark to keep funding an OpenAI position it already paid for in equity. The buildout is being paid for by a rotating cast of guarantors, and each one is layering financial commitments on top of paper marks. The bubble-debate scoreboard gets an extra column this week, and the column is labeled recourse.

Counterreads

The strongest bear read on this piece is that we are pattern-matching to a leverage crisis that has not happened. SoftBank has run large balance-sheet positions for a decade. Masayoshi Son ran WeWork exposure and Alibaba exposure through worse volatility than an OpenAI mark cut in diligence. The $40 billion bridge got 21 additional lenders in July, which is the opposite signal a stressed borrower would generate. Mizuho would not lead-arrange a second $10 billion facility at SOFR + 275 if the collateral picture were shaky. Every one of those points is true, and none of them changes the fact that the structural risk has moved. What used to be equity risk on a paper mark is now equity risk plus a schedule.

The second counterread is that private markets have moved on. Bloomberg reported this spring that the secondary OpenAI mark was clearing above $850 billion with real volume, and Coatue and Sequoia and Fidelity are still writing checks at those numbers. A single desk cutting a bond diligence mark is not the same as the market clearing at a new price. That is fair. It also underlines exactly why the March 2027 wall matters: the private mark holds up as long as private buyers keep clearing at it, and private buyers only clear at it as long as the IPO is visible on the horizon. Delay the print by six months and half the diligence spreadsheets swap the secondary price for a public comp.

The third counterread is the one worth sitting with. The whole stack is not really about collateral coverage. It is about SoftBank ensuring that its OpenAI position is fully funded before the equity gets diluted again in a pre-IPO round. Cheap dollars secured on the current mark are the cheapest way to defend the 13 percent, and defending the 13 percent is what keeps SoftBank in the Founders Fund tier of the eventual public cap table. On that read, the leverage stack is a rational insurance policy, and the headline risk is a rounding error against the alternative of getting diluted below the board seat threshold. That framing is coherent. It also is a bet that the equity upside on the marginal share exceeds the interest cost across the stack, and interest costs across the stack are already north of $2 billion a year on the drawn portion.

Our Take

Priced against the way frontier-lab exposures were funded eighteen months ago (mostly equity from a handful of names, held on unlevered balance sheets), the SoftBank stack is a step-change in how the AI capex conversation reaches the credit markets. Priced against the way LBO sponsors have been funding private-company positions for two decades, it is aggressive but not exotic. Both readings are true. What matters this quarter is that the borrower is a public company with a very visible balance sheet, the collateral is a private position that trades on infrequent secondary marks, and the repayment schedule is stacked against a single IPO window nobody controls.

Practical implication for anyone modeling the compute buildout. The next time an Anthropic or an OpenAI headlines a multi-year hyperscaler commitment, ask a follow-up question the reporter probably did not: which piece of the stack is being funded by fresh cash flow, which piece is being funded by equity, and which piece is being funded by debt secured against another party's equity. Two years ago the answer was almost entirely equity. This week the answer is mostly debt-on-paper, and the paper is not liquid.

Three signposts to watch. Whether the September bond prints at the marketed size, or gets cut to the $10 billion floor, which tells you whether the credit market is repricing SoftBank OpenAI paper. Whether the OpenAI S-1 moves from confidential to public inside 120 days, which is the earliest visible checkpoint on the IPO clearing before March 2027. And whether any of the eight banks on the August loans quietly refuses to backstop the bond, which is the way this kind of situation surfaces first, one syndicate desk at a time.

We are tracking the OpenAI cap table and the March 2027 refinance clock on our OpenAI provider page and the SoftBank Piketon build via the Nvidia guarantee analysis. Next data point: the SoftBank Q2 FY27 print later this quarter, which is the first time the second margin loan and any priced bond show up on the same page.