SoftBank is raising $40 billion through a bridge loan that routes to OpenAI via Vision Fund 2, and it is doing so weeks before OpenAI’s confidential S-1 filing. The banks syndicating that debt, including HSBC, BNP Paribas, and Intesa Sanpaolo, are effectively underwriting frontier AI capex with a 12-month maturity and roughly 6% all-in interest. When the company burning $50 billion a year on compute against $13.1 billion in revenue goes public, those lenders will have a seat at the pricing table.
The $40B Bridge: Terms, Syndicate, and What the Banks Are Betting On
On March 27, 2026, SoftBank signed a $40 billion unsecured bridge facility arranged by JPMorgan, Goldman Sachs, Mizuho, SMBC, and MUFG, with $30 billion earmarked for a follow-on OpenAI investment through Vision Fund 2. The facility matures March 25, 2027, giving SoftBank a one-year window to either refinance or convert the position.
By late April, at least eight additional banks, including HSBC, BNP Paribas, and Intesa Sanpaolo, had submitted commitments as sub-underwriters. Each was asked to commit a minimum of $5 billion. The deal pays approximately 250 basis points over SOFR, landing around a 6.14% all-in rate.
This is not a venture round. It is a syndicated bank loan with covenants, maturity pressure, and refinancing risk. The banks are not betting on OpenAI’s technology. They are betting that SoftBank can exit or refinance within 12 months without a credit event. S&P Global Ratings revised SoftBank’s outlook to negative on March 3, 2026 while affirming its BB+ issuer rating, specifically citing the risk that the additional OpenAI investment could weaken liquidity. SoftBank CDS spreads widened 74 basis points and touched a one-year high in April.
The $50B Compute Burn and the Revenue Gap
Greg Brockman testified that OpenAI’s 2026 compute spend alone will be $50 billion. Against 2025 revenue of $13.1 billion and a net loss of approximately $9 billion, that spend rate keeps OpenAI structurally dependent on external capital to fund infrastructure deployment.
OpenAI’s contractual compute footprint as of May 2026 is approximately $400 billion over three years across Stargate, Oracle (which committed $300 billion over five years, up to 4.5 GW), CoreWeave, SoftBank, and Microsoft. OpenAI shifted from owning to leasing compute mid-2025, converting what would have been balance-sheet capex into operating commitments. That swap improves the income statement in the short term but locks in fixed compute costs that must be covered by inference revenue, regardless of demand.
The $122 billion equity round closed March 31, 2026 at an $852 billion post-money valuation, with Amazon investing $50 billion, Nvidia $30 billion, SoftBank $30 billion, and Microsoft reinvesting at an undisclosed amount. That equity covers a portion of the compute commitments. It does not close the gap between $13.1 billion in annual revenue and a $50 billion annual compute bill.
Nvidia’s $30 billion investment warrants a specific note. OpenAI buys Nvidia GPUs. The investment is partially circular: OpenAI will spend a large share of that capital on Nvidia hardware. This is not net new capital entering the AI ecosystem; it is Nvidia pre-funding its own revenue. Characterizing it otherwise inflates the apparent capital available for inference-price subsidies.
Debt in the Capital Chain: How SoftBank’s Debt Load Constraints OpenAI’s Pricing
The mechanism is indirect but legible. SoftBank is separately seeking a $10 billion margin loan backed by its OpenAI shares at approximately 425 basis points over SOFR, roughly 7.89% all-in. It also raised $3.6 billion in a multi-tranche bond deal in April 2026, with a 10-year dollar tranche carrying an 8.5% coupon, the highest SoftBank has ever paid on such a security.
SoftBank is now triple-levered into its OpenAI position: the $40 billion bridge loan, the $10 billion margin loan, and the $3.6 billion bond issuance. All three instruments carry interest rates that demand the underlying asset appreciate or generate returns fast enough to service the debt.
Here is the constraint: SoftBank’s lenders need OpenAI’s equity value to hold or rise through the IPO. If OpenAI prices inference too cheaply to chase volume, revenue per query drops, margins compress, and the IPO valuation story weakens. If OpenAI prices inference too aggressively, user growth stalls, the same story weakens, and SoftBank’s $30 billion position loses value relative to the $852 billion round price.
The lenders are not setting API prices. But the cost of capital in the chain behind OpenAI’s largest investor creates a floor under how long OpenAI can run inference at a loss to buy market share.
What the S-1 Will Have to Disclose
Bloomberg reported in May 2026 that OpenAI is preparing to confidentially file its S-1, seeking a public debut in Fall 2026. Reuters has separately reported that management is considering a retail allocation.
The S-1’s risk-factors section will have to address several items that most venture-backed IPOs can gloss over:
Related-party compute deals. Microsoft is both an investor and OpenAI’s primary cloud provider. Nvidia is both an investor and OpenAI’s primary hardware supplier. Oracle, CoreWeave, and SoftBank hold similar dual roles. The SEC requires disclosure of material related-party transactions, and in OpenAI’s case, these transactions represent the majority of the cost structure.
Covenant exposure through investors. SoftBank’s bridge loan and margin loan mean that OpenAI’s second-largest equity holder is itself carrying debt that could force liquidation of its OpenAI position. A forced sale by a major shareholder during the post-IPO lockup period would be material. The S-1 will need to disclose this risk, or the underwriters will.
The circular capital loop. Nvidia invested $30 billion. OpenAI’s compute spend flows back to Nvidia. This is not opaque, but the specific dollar amounts flowing through that loop will need to appear in the related-party disclosures.
Revenue concentration and target misses. If the WSJ report on missed internal targets is accurate, the S-1’s discussion of revenue trends will either confirm or contradict those targets. Either way, the filing becomes the first audited public look at whether $13.1 billion in 2025 revenue is growing fast enough to service the implied capex burden.
The $852 billion post-money valuation from March 2026 may or may not hold through the IPO pricing window. What holds regardless of the specific number is the structural shift it represents: bank lenders are now in the capital chain behind the largest AI lab, and their covenants will be in the S-1. The frontier model companies are no longer funded solely by equity and venture debt. They are funded by the same syndicated loan market that finances leveraged buyouts, and that market expects its money back.
Frequently Asked Questions
How does SoftBank’s triple-leveraged OpenAI position compare to its WeWork exposure?
SoftBank used a similar margin-loan structure against its WeWork stake in 2019 and was forced to write down over $7 billion when that IPO collapsed. The OpenAI position is roughly ten times larger by notional value, but the structural pattern is identical: debt collateralized by private-market valuations that have not been tested in public trading. One difference is that OpenAI generated $13.1 billion in 2025 revenue, whereas WeWork had no viable path to covering its lease obligations at the time of its attempted IPO.
What happens if S&P downgrades SoftBank from BB+?
BB+ is the highest speculative-grade rating. A one-notch cut to BB pushes SoftBank deeper into high-yield territory, raising refinancing costs for the $40 billion bridge when it matures in March 2027. If replacement financing is unavailable at comparable terms, SoftBank may need to liquidate assets, including a portion of its OpenAI stake, potentially during OpenAI’s post-IPO lockup period. That scenario would put selling pressure on a newly public stock from its own largest shareholder.
Do other frontier labs carry comparable debt exposure in their cap tables?
No. Google DeepMind funds compute through Alphabet’s investment-grade balance sheet. Anthropic’s capital comes from equity raises (Amazon’s $4 billion, Google’s contributions) without a major investor carrying margin loans against the position. No other frontier lab has an investor triple-levered on its shares. If the SoftBank structure survives the IPO cycle without a credit event, compute providers may replicate it for other labs. If it doesn’t, this becomes the boundary condition that limits debt-financed AI infrastructure.
What does Oracle’s 4.5 GW commitment mean for OpenAI’s cost flexibility?
Oracle’s $300 billion, five-year deal includes up to 4.5 gigawatts of dedicated power capacity. Power procurement and data-center buildout at that scale involve physical infrastructure with lead times measured in years, not months. Unlike cloud compute contracts that can scale with demand, this is a fixed-capacity commitment: OpenAI pays for the megawatts whether or not inference volume fills them. That rigidity compounds the floor under per-query costs and limits how quickly OpenAI can reduce compute spend if revenue growth disappoints.
If OpenAI allocates shares to retail investors, who bears the covenant risk?
Reuters reports management is considering a retail allocation for the IPO. Retail investors would buy into a company whose second-largest shareholder (SoftBank) carries $50+ billion in debt partially collateralized by its OpenAI shares. Institutional lenders priced that risk into the bridge loan’s 250 bps spread and the margin loan’s 425 bps spread. Retail shareholders have no such protection, no covenants, and no senior claim in a distressed scenario. The S-1 will disclose SoftBank’s leveraged position, but the informational asymmetry between syndicate banks and retail buyers will be significant.