Last year, SoftBank Group Chairman and CEO Masayoshi Son declared his vision of turning the Japanese technology investment group into the biggest platform provider for “artificial super intelligence”. That required massive and aggressive investments into all spheres of the AI ecosystem, and it’s finally beginning to dent its bottomline.
The Japanese investment holding company reported a ¥1.12 trillion ($6.99 billion) net loss for July–September 2026, its first quarterly loss in about two years, after revising orders for power equipment intended for US data centres that resulted in a ¥597.2 billion ($3.80 billion) charge. But that doesn’t seem to be enough.
Elon Musk recently cried foul, posting on X that “SoftBank is defaulting on their turbine orders.” This comes as the $500 billion multi-year Stargate Project initiative to build AI data centres—backed by SoftBank, NVIDIA, Oracle, and OpenAI—hits roadblocks including partner squabbles, cancelled expansions, and local regulations.
Chief Financial Officer Yoshimitsu Goto told analysts on the earnings call there was “no such default.” SoftBank describes the charge as the cost of changing orders, not a missed payment.
SoftBank is ramping up AI infrastructure commitments, ramping up capital across chips, AI models and data centres, while relying on its existing assets and borrowing capacity to fund that expansion.
This brings Arm into focus. The chip designer is SoftBank's strongest AI asset, but it also serves as collateral for borrowing that supports its wider investment strategy, including its nearly $65 billion investment in OpenAI, where returns are slower and less certain. According to WSJ, the Sam Altman-led company’s operating losses widened to $10.9 billion in the third quarter.
When Arm's shares fall—such as when it dropped roughly 13% in early November as investors reacted to slower-than-expected growth during its second quarter—the value of SoftBank's holdings and the security behind its loans fall together.
That happened in the quarter. SoftBank's net asset value (NAV), the value of its holdings minus net debt, fell to ¥59.8 trillion ($380.04 billion) from ¥72.3 trillion ($445.22 billion) in June. Research firm Morningstar says Arm's 23% share-price drop in the relevant period was the main driver.
AI-related holdings accounted for over 60% of NAV by February 2026, up from about 18% three years earlier.
Arm is Both Bulwark and Collateral
SoftBank groups its AI investments into four pillars: chips (Arm), models (OpenAI), infrastructure (data centres and power), and physical AI (robotics).
Stephen Leung, Director, TMT Analyst, Investment Advisory at Astris Advisory Japan KK, tells AIM that chips offer the strongest and most proven returns. “Arm functions as the primary structural anchor of SoftBank’s balance sheet,” he says.
Arm accounted for 54% of SoftBank's equity value of holdings, after adjustments for asset-backed financing, at the end of September, its figures show.
Leung says Arm generates high-margin operating profits, while newer chip architectures command higher royalty rates per chip.
Arm said in its November 4 earnings call that data centre royalties doubled from a year earlier. Total royalty growth was more modest at 8%, as smartphone royalties stayed flat on higher memory prices. Licensing revenue rose 50%, helping lift overall revenue growth to 27%. Part of that licensing revenue came from SoftBank itself.
Despite its growth in data centres, Arm remains exposed to shifts in investor expectations around AI spending. Concerns about the pace of infrastructure investment and AI returns could weigh on its valuation. Competition from alternative chip architectures also presents a longer-term risk.
SoftBank also uses its Arm holdings to secure borrowing. Its financial report shows a ¥3.9 trillion ($24.79 billion) margin loan backed by Arm shares at the end of September, which Leung identifies as the main collateral supporting its wider AI expansion.
The arrangement includes cash-collateral and early-repayment provisions that can be triggered by certain events, including a significant decline in the value of pledged Arm shares. A fall towards the margin call threshold could require additional collateral or refinancing. Reaching the early repayment threshold could require the loan to be repaid in full, in cash or shares.
Some Bets Have Started Paying Off
SoftBank holds about a 13% stake in OpenAI. Leung says OpenAI offers substantial upside, but its valuation remains a private estimate that can fluctuate, and the eventual returns depend partly on when it lists.
In June, The New York Times reported that OpenAI was leaning towards delaying its listing to 2027. SoftBank's shares fell about 13% that day.
In August, Reuters reported that SoftBank's attempt to borrow against its OpenAI holding had been held up as lenders grew wary of credit backed by private companies. The eventual financing arrangement differed from the one backed by Arm shares.
SoftBank's $10 billion loan does not directly pledge OpenAI shares as collateral. Instead, it uses a cash-collateral account, carries a SoftBank Group guarantee, and includes margin call provisions linked to the value of OpenAI preferred shares held by SVF2.
A significant decline in their value could trigger additional cash-collateral requirements or mandatory prepayment, exposing SoftBank to changes in the value of its OpenAI-related assets despite the absence of a direct share pledge.
SoftBank has not yet reflected OpenAI's latest funding round in its numbers. An update could come in the October–December quarter.
The build-out continues, including plans for 5 GW of data centre capacity in France. Cumulative payments for power-generation equipment reached ¥1.5 trillion ($9.48 billion) by September 30.
Two Ways to Count Leverage
SoftBank also has a ¥1.2 trillion ($7.63 billion) margin loan backed by shares in its telecom unit, SoftBank Corp, its financial report shows.
Its policy is to keep its loan-to-value (LTV) ratio, calculated as adjusted net debt divided by adjusted equity value of holdings, below 25% in normal conditions, Leung says.
S&P's calculation, which uses a different methodology, puts SoftBank's LTV at between 20–25% in June—far more than SoftBank’s 13% figure.
Leung says the gap matters because SoftBank's holdings are concentrated in Arm, while a substantial part of its investment value is tied to OpenAI.
A sustained fall in Arm's share price would reduce the value of SoftBank's holdings while leaving the outstanding loan unchanged. This would push its LTV higher and put pressure on SoftBank's own 25% policy threshold.
Goto says SoftBank needs to balance the accelerator and the brake to keep LTV under 25%. He also says a steep fall in NAV would raise the risk to LTV.
SoftBank raised a record $56.5 billion through debt-market financing and asset monetisation in July–September. It also repaid its bridge loan ahead of schedule.
Leung says SoftBank has demonstrated access to capital markets to restructure its debt facilities, but sustaining this financing strategy requires continued market liquidity and stability in the value of its core collateral.
Market Swings and Structural Costs
Leung says part of the loss is accounting-related. As an investment holding company, SoftBank's profit moves with market prices and shifting listing timelines.
Intel is an example. Gains on the chipmaker's shares produced a ¥1.33 trillion ($8.30 billion) gain in April–June. In July–September, SoftBank booked a ¥321.9 billion ($2.01 billion) loss on financing contracts tied to Intel shares.
Another loss was linked to Energy Global, which granted warrants to OpenAI as incentives under data centre lease agreements. As the fair value of these warrants increased, SoftBank recorded a ¥203.8 billion ($1.30 billion) derivative loss for the six months ended September 30.
AI models and infrastructure carry structural costs, with large capital commitments preceding revenue.
Goto says an official OpenAI revaluation, which could be reflected in the October–December quarter, would raise NAV and improve LTV. Until then, much of SoftBank's borrowing rests on the price of Arm's shares.
- The Nightmare at NVIDIA
- India’s Accessibility Rules are Forcing Software and App Design Rethink
- Succession, Scrutiny, and the Future of NVIDIA
- Can You Trust Your ‘Instinct’ and Not ‘Muse’ Over the Risks?
- Is AI-Powered Reverse Engineering the Death of Proprietary Software?
- Smaller Language Models Could Reshape How Enterprises Deploy AI

