SoftBank is stretching past its own rules to own more of OpenAI — and the numbers are getting uncomfortable.
The Japanese investment firm is seeking a bridge loan of up to $40 billion, according to Reuters, which would make it SoftBank's largest ever dollar-denominated borrowing if completed. The facility would carry a roughly 12-month tenor with JPMorgan Chase among the arrangers. The loan is intended to fund additional investment in OpenAI — the company that already accounts for the bulk of SoftBank's AI portfolio.
What is not ambiguous is the direction. SoftBank has committed $30 billion in additional OpenAI investment through Vision Fund 2, structured in three $10 billion tranches on April 1, July 1, and October 1 of 2026, per the company press release. Combined with the $34.6 billion already invested since September 2024, SoftBank's total OpenAI position stands at roughly $64.6 billion — a stake the Financial Times has estimated at around 13 percent of the company. That is not a bet. It is a position.
The leverage question
SoftBank operates under a self-imposed loan-to-value ceiling of 25 percent under normal conditions, with an upper threshold of 35 percent in emergencies, as the company stated in its February press release. Its LTV ratio had climbed from 16.5 percent to 20.6 percent in recent months before this loan was arranged. Chief Financial Officer Yoshimitsu Goto acknowledged publicly that the ratio would likely exceed that threshold temporarily — a remarkable admission from a CFO about his own company risk framework.
S&P Global Ratings noticed. The agency revised SoftBank's credit outlook to negative from stable on March 3, 2026, while affirming the BB+ rating, per Bloomberg. The reason: an additional $30 billion investment would stress SoftBank's liquidity and asset quality. The market disagreed with S&P caution — SoftBank debt is trading well above distressed levels — but credit ratings are written for a reason.
Masayoshi Son, SoftBank founder, appears to have decided the reason no longer applies. At the groundbreaking of SoftBank's Ohio campus in March, he said the rules were written for normal times, and these are not normal times. That is the thesis in one sentence.
What the money is building
The physical manifestation of this bet is under construction in Piketon, Ohio — a 3,700-acre site that was a uranium enrichment facility for nuclear weapons from 1954 to 2001. The irony is not subtle. The substrate for AI inference, being built on the same land that produced fissile material for atomic weapons, is being developed by SoftBank subsidiary SB Energy with OpenAI as anchor tenant and Arm providing processor architecture.
The target is 10 gigawatts at full build-out. For context: the entire U.S. data center fleet today runs at roughly 17 to 20 gigawatts. One campus, in Ohio, is designed to supply more than half of that. Phase one is 800 megawatts, with an estimated cost of $30 billion to $40 billion, targeted for early 2028. Full deployment is pencilled in for the end of the decade.
This is part of a broader $550 billion U.S.-Japan investment commitment tied to tariff relief between the two countries. It is also the most concrete physical infrastructure bet in the current wave of AI capital deployment — more concrete, at least on paper, than the speculative timelines that typically accompany announcements of this scale.
The Stargate structure
SoftBank and OpenAI jointly own the Stargate joint venture, each holding 40 percent. Oracle and MGX each contribute $7 billion for the remaining 20 percent. The arrangement gives SoftBank a governance position disproportionate to a typical financial investor — and gives OpenAI a partner that is simultaneously its largest shareholder and its infrastructure host.
This is the second significant SoftBank-OpenAI announcement in recent weeks. In February 2026, the two companies disclosed a $110 billion funding round at a $730 billion pre-money valuation, with Amazon committing $50 billion, SoftBank $30 billion, and Nvidia $30 billion. OpenAI later added $10 billion, bringing the round to $120 billion and the company's implied valuation to roughly $840 billion, according to CNBC.
What to watch
The Piketon timeline is aggressive. 800 megawatts in roughly two years, for a first phase that already represents more compute than most hyperscalers have deployed in a single location, against a grid that is not currently built for rapid AI cluster expansion. The Ohio site requires new power infrastructure from scratch. The compute economics of large-scale AI training — where infrastructure ambition routinely outruns unit economics — are the risk embedded in every tranche of this bet.
SoftBank is betting that the infrastructure itself is the moat. Son has said as much, in substance if not in those words: own the compute layer, set the terms for everything above it. Whether that thesis survives contact with a 10-gigawatt construction schedule and a credit outlook from S&P is a different question — and one worth tracking as the tranches deploy.