SoftBank seeks $10B margin loan backed by OpenAI shares at SOFR+425bps as leverage stack deepens


Summary: SoftBank is seeking a $10 billion margin loan backed by its OpenAI shares at SOFR + 425 basis points (~7.88%), a two-year term with one-year extension. The loan sits atop a $40 billion bridge loan from March and brings SoftBank’s total OpenAI commitment to ~$64.6 billion for a ~13% stake. At OpenAI’s $852 billion valuation, the stake is notionally worth ~$110 billion, but S&P has downgraded SoftBank’s outlook to negative (BB+) and the company faces a $32 billion funding gap over two years.

SoftBank is seeking a $10 billion margin loan backed by its OpenAI shares, Bloomberg reported on Wednesday, adding another layer of debt to the most leveraged bet in the history of artificial intelligence. The proposed loan carries a two-year term with an option to extend by one year, at an interest rate of approximately 425 basis points above the secured overnight financing rate, which translates to roughly 7.88% at current levels. The deal has not been finalised. The specific lending banks have not been named. What is known is that SoftBank is borrowing against paper wealth in a private company to fund more investment in the same private company, a recursive leverage structure that works brilliantly until it does not.

This is not SoftBank’s first margin loan, nor its most complex financing arrangement of the year. In March, the company secured a $40 billion bridge loan underwritten by JPMorgan Chase, Goldman Sachs, Mizuho Bank, Sumitomo Mitsui Banking Corporation, and MUFG Bank, earmarked for a $30 billion follow-on investment in OpenAI and general corporate purposes. The bridge loan entered a soft launch phase in mid-April, with additional lenders being invited to join at roughly $5 billion commitments each. The $10 billion margin loan sits on top of that. SoftBank’s total debt now stands at approximately 20.45 trillion yen, roughly $135 billion.

The OpenAI position

SoftBank’s cumulative investment in OpenAI will reach approximately $64.6 billion once its $30 billion follow-on closes, giving it roughly 13% of the company. The initial commitment, completed by late December 2025, totalled $40 to $41 billion: $7.5 billion in direct investment, $11 billion syndicated with co-investors, and a $22 to $22.5 billion final tranche. To fund the position, Masayoshi Son sold SoftBank’s entire Nvidia stake for $5.83 billion and $12.73 billion in T-Mobile stock between June and December 2025. He then borrowed $40 billion more. He is now borrowing $10 billion on top of that.

The collateral’s notional value has changed dramatically since the investment was made. SoftBank’s initial $40 billion went in at OpenAI’s March 2025 valuation of $300 billion pre-money. By March 2026, OpenAI’s record funding round closed at an $852 billion post-money valuation, anchored by Amazon at $50 billion, Nvidia at $30 billion, and SoftBank’s own $30 billion follow-on. At that valuation, SoftBank’s 13% stake is notionally worth roughly $110 billion, making a $10 billion margin loan seem modest, roughly 9% of the collateral’s paper value. The question is what that paper is worth if the music stops.

The price of illiquidity

The 💜 of EU tech

The latest rumblings from the EU tech scene, a story from our wise ol’ founder Boris, and some questionable AI art. It’s free, every week, in your inbox. Sign up now!

When SoftBank borrowed $8 billion against its Alibaba stake in 2018, the rate was LIBOR plus 150 basis points. Ten banks participated. Alibaba was publicly traded on the New York Stock Exchange, with daily trading volume in the billions of dollars. If SoftBank defaulted, the lenders could sell the shares on the open market the next morning. The rate reflected that liquidity.

The OpenAI margin loan is priced at SOFR plus 425 basis points, nearly triple the spread. The difference is not just the macro environment. It is the nature of the collateral. OpenAI is a private company. Its shares do not trade on any exchange. Secondary market transactions are infrequent, opaque, and subject to company approval. OpenAI’s $852 billion valuation has come under scrutiny from its own investors, with secondary market data showing a five-to-one ratio of sellers to buyers. If SoftBank were to default, the lenders would hold shares in a private company that they cannot easily sell, at a valuation that the secondary market is already questioning, in a sector where sentiment can shift in a quarter. The 275 basis point spread premium over the Alibaba loan is the banks’ price for that risk. Whether it is enough is another matter.

The leverage stack

SoftBank’s loan-to-value ratio stood at 20.6% as of December 2025, with management estimating current levels at roughly 19%. The company’s self-imposed ceiling is 25%, and it has acknowledged it could “temporarily” breach that limit. S&P Global Ratings downgraded SoftBank’s credit outlook from stable to negative in March 2026, affirming its BB+ rating and citing concerns about liquidity and asset credit quality. S&P noted that OpenAI is among the investments with the “weakest credit quality” in SoftBank’s portfolio and that the proportion of unlisted assets is expected to rise above 50%, up from 42% in December. Credit-default swaps on SoftBank debt widened roughly 10 basis points to approximately 360 basis points after the margin loan was reported, approaching a one-year high of 376.

Analysts estimate SoftBank faces a $32 billion funding gap over the next two years for bond redemptions and committed acquisitions. The company held 3.8 trillion yen in cash as of December and has issued 1.12 trillion yen in domestic bonds alongside $2.2 billion and 1.7 billion euros in foreign currency bonds. Undrawn credit lines totalled 945.2 billion yen. The financing is not yet stretched to breaking, but the margin for error is narrowing with each new facility.

The Son thesis

Masayoshi Son has described SoftBank’s current posture as “total offence mode” and is repositioning the company as an “AI-era industrial holding company.” He believes artificial superintelligence, AI 10,000 times smarter than humans, will arrive within 10 years. He has stopped investing in China entirely. Sixty percent of SoftBank’s assets are now classified as “ASI-oriented investments.” Son chairs the Stargate joint venture with OpenAI, Oracle, and MGX, the Abu Dhabi investment firm, which has committed an initial $100 billion with plans to invest up to $500 billion by 2029 in AI data centres across the United States. Ten facilities are under construction in Abilene, Texas, with expansion planned to the UK, Norway, Japan, and the UAE.

The thesis requires believing several things simultaneously: that OpenAI will maintain its position as the leading AI company; that its $852 billion valuation will hold or increase; that the Stargate infrastructure buildout will generate returns commensurate with its cost; and that SoftBank can service its debt stack while the returns materialise. Son has been right about this kind of bet before. His $20 million investment in Alibaba in 2000 returned more than $100 billion. He has also been wrong. The Vision Fund 1 returned 7% IRR on $100 billion in committed capital. Vision Fund 2 returned 0.2%, essentially flat, on $72 billion. The combined funds generated roughly $5 billion in gains from $172 billion in commitments. The question is which pattern the OpenAI bet follows.

The AI financing environment

SoftBank’s margin loan sits within a venture capital environment deploying record capital into AI, with $297 billion in venture investment in the first quarter of 2026 alone, 2.5 times the previous quarter. The concentration is extreme. OpenAI’s $122 billion round, the largest private funding round in history, drew $50 billion from Amazon and $30 billion each from Nvidia and SoftBank. The capital is flowing to a handful of companies at valuations that assume the entire global economy will be reorganised around their technology within a decade.

The risk is not that AI fails to matter. It is that the returns are competed away, that open-source models erode pricing power, that the infrastructure buildout overshoots demand, or that a single company at an $852 billion private valuation is mispriced even if the sector as a whole delivers. OpenAI has already paused its Stargate UK data centre project over energy costs and regulatory obstacles, a reminder that not every infrastructure plan proceeds on schedule. SoftBank is borrowing against a future that it is simultaneously helping to build and betting will arrive on time. The margin loan is a financial instrument. The margin for error is a strategic one. At $135 billion in total debt and $64.6 billion committed to a single private company, the distance between visionary and reckless is one down round.



Source link

Leave a Reply

Subscribe to Our Newsletter

Get our latest articles delivered straight to your inbox. No spam, we promise.

Recent Reviews


As I’m writing this, NVIDIA is the largest company in the world, with a market cap exceeding $4 trillion. Team Green is now the leader among the Magnificent Seven of the tech world, having surpassed them all in just a few short years.

The company has managed to reach these incredible heights with smart planning and by making the right moves for decades, the latest being the decision to sell shovels during the AI gold rush. Considering the current hardware landscape, there’s simply no reason for NVIDIA to rush a new gaming GPU generation for at least a few years. Here’s why.

Scarcity has become the new normal

Not even Nvidia is powerful enough to overcome market constraints

Global memory shortages have been a reality since late 2025, and they aren’t just affecting RAM and storage manufacturers. Rather, this impacts every company making any product that contains memory or storage—including graphics cards.

Since NVIDIA sells GPU and memory bundles to its partners, which they then solder onto PCBs and add cooling to create full-blown graphics cards, this means that NVIDIA doesn’t just have to battle other tech giants to secure a chunk of TSMC’s limited production capacity to produce its GPU chips. It also has to procure massive amounts of GPU memory, which has never been harder or more expensive to obtain.

While a company as large as NVIDIA certainly has long-term contracts that guarantee stable memory prices, those contracts aren’t going to last forever. The company has likely had to sign new ones, considering the GPU price surge that began at the beginning of 2026, with gaming graphics cards still being overpriced.

With GPU memory costing more than ever, NVIDIA has little reason to rush a new gaming GPU generation, because its gaming earnings are just a drop in the bucket compared to its total earnings.

NVIDIA is an AI company now

Gaming GPUs are taking a back seat

A graph showing NVIDIA revenue breakdown in the last few years. Credit: appeconomyinsights.com

NVIDIA’s gaming division had been its golden goose for decades, but come 2022, the company’s data center and AI division’s revenue started to balloon dramatically. By the beginning of fiscal year 2023, data center and AI revenue had surpassed that of the gaming division.

In fiscal year 2026 (which began on July 1, 2025, and ends on June 30, 2026), NVIDIA’s gaming revenue has contributed less than 8% of the company’s total earnings so far. On the other hand, the data center division has made almost 90% of NVIDIA’s total revenue in fiscal year 2026. What I’m trying to say is that NVIDIA is no longer a gaming company—it’s all about AI now.

Considering that we’re in the middle of the biggest memory shortage in history, and that its AI GPUs rake in almost ten times the revenue of gaming GPUs, there’s little reason for NVIDIA to funnel exorbitantly priced memory toward gaming GPUs. It’s much more profitable to put every memory chip they can get their hands on into AI GPU racks and continue receiving mountains of cash by selling them to AI behemoths.

The RTX 50 Super GPUs might never get released

A sign of times to come

NVIDIA’s RTX 50 Super series was supposed to increase memory capacity of its most popular gaming GPUs. The 16GB RTX 5080 was to be superseded by a 24GB RTX 5080 Super; the same fate would await the 16GB RTX 5070 Ti, while the 18GB RTX 5070 Super was to replace its 12GB non-Super sibling. But according to recent reports, NVIDIA has put it on ice.

The RTX 50 Super launch had been slated for this year’s CES in January, but after missing the show, it now looks like NVIDIA has delayed the lineup indefinitely. According to a recent report, NVIDIA doesn’t plan to launch a single new gaming GPU in 2026. Worse still, the RTX 60 series, which had been expected to debut sometime in 2027, has also been delayed.

A report by The Information (via Tom’s Hardware) states that NVIDIA had finalized the design and specs of its RTX 50 Super refresh, but the RAM-pocalypse threw a wrench into the works, forcing the company to “deprioritize RTX 50 Super production.” In other words, it’s exactly what I said a few paragraphs ago: selling enterprise GPU racks to AI companies is far more lucrative than selling comparatively cheaper GPUs to gamers, especially now that memory prices have been skyrocketing.

Before putting the RTX 50 series on ice, NVIDIA had already slashed its gaming GPU supply by about a fifth and started prioritizing models with less VRAM, like the 8GB versions of the RTX 5060 and RTX 5060 Ti, so this news isn’t that surprising.

So when can we expect RTX 60 GPUs?

Late 2028-ish?

A GPU with a pile of money around it. Credit: Lucas Gouveia / How-To Geek

The good news is that the RTX 60 series is definitely in the pipeline, and we will see it sooner or later. The bad news is that its release date is up in the air, and it’s best not to even think about pricing. The word on the street around CES 2026 was that NVIDIA would release the RTX 60 series in mid-2027, give or take a few months. But as of this writing, it’s increasingly likely we won’t see RTX 60 GPUs until 2028.

If you’ve been following the discussion around memory shortages, this won’t be surprising. In late 2025, the prognosis was that we wouldn’t see the end of the RAM-pocalypse until 2027, maybe 2028. But a recent statement by SK Hynix chairman (the company is one of the world’s three largest memory manufacturers) warns that the global memory shortage may last well into 2030.

If that turns out to be true, and if the global AI data center boom doesn’t slow down in the next few years, I wouldn’t be surprised if NVIDIA delays the RTX 60 GPUs as long as possible. There’s a good chance we won’t see them until the second half of 2028, and I wouldn’t be surprised if they miss that window as well if memory supply doesn’t recover by then. Data center GPUs are simply too profitable for NVIDIA to reserve a meaningful portion of memory for gaming graphics cards as long as shortages persist.


At least current-gen gaming GPUs are still a great option for any PC gamer

If there is a silver lining here, it is that current-gen gaming GPUs (NVIDIA RTX 50 and AMD Radeon RX 90) are still more than powerful enough for any current AAA title. Considering that Sony is reportedly delaying the PlayStation 6 and that global PC shipments are projected to see a sharp, double-digit decline in 2026, game developers have little incentive to push requirements beyond what current hardware can handle.

DLSS 5, on the other hand, may be the future of gaming, but no one likes it, and it will take a few years (and likely the arrival of the RTX 60 lineup) for it to mature and become usable on anything that’s not a heckin’ RTX 5090.

If you’re open to buying used GPUs, even last-gen gaming graphics cards offer tons of performance and are able to rein in any AAA game you throw at them. While we likely won’t get a new gaming GPU from NVIDIA for at least a few years, at least the ones we’ve got are great today and will continue to chew through any game for the foreseeable future.



Source link