SoftBank's OpenAI bet: Masayoshi Son's portfolio, leverage and the AI sell-off

Executive summary

SoftBank Group's shares fell 60.6% from their 2026 high close on 2 June to 9 November, the last close before its second-quarter results, while MSCI World net in yen slipped 1.3%. Arm, its largest holding, lost 53.7% in dollars over the same dates. The fall erased the shares' ten-year lead over the index: from October 2016 to 9 November 2026 they compounded at 15.86% a year in yen, against 17.57% for MSCI World.

SoftBank's own net asset value tells the same story: ¥72.3tn at 30 June, ¥59.8tn at 30 September and ¥44.7tn at 9 November prices, according to its second-quarter presentation, published on 10 November. Arm's part of that NAV fell from ¥43.4tn to ¥18.6tn.

Over the full period since the end of 2000, SoftBank still compounded faster than the index, 12.97% a year against 8.74%, but with three times the volatility, a lower Sharpe ratio and peak-to-trough losses of 88% and 87%.

Copio's assessment: SoftBank is a borrowed, concentrated bet on Arm and OpenAI, assets that depend on the same AI spending cycle. Its company-defined LTV of 15.9% at 30 September 2026 is below its 25% normal-times policy ceiling, but it measures debt against a holdings value that moves with Arm, and it rises to about 22.9% with margin loans and collar contracts added back (our calculation).

The qualification is real: the shares trade well below SoftBank's own NAV, about 57% below the 9 November pro forma, so a shareholder's result can diverge sharply from the holdings.

The same results show the October OpenAI payment made, the $40bn bridge loan repaid early with a term loan, Intel-backed derivative contracts and a larger Arm margin loan, and a six-month net loss of ¥768.0bn.

For a household, the practical check is to add SoftBank’s underlying AI exposure to the technology funds already held, then estimate the loss in euros or yen if those positions fall together. Keep money needed for spending or debt repayments separate from that scenario; a discount to NAV does not provide cash when it is needed.

What does SoftBank Group own?

SoftBank Group is a Tokyo-listed investment holding company whose value sits in a few large stakes. At 30 September 2026, its second-quarter presentation put the equity value of its holdings at ¥71.1tn after deducting margin loans and collar contracts, and its NAV at ¥59.8tn, which implies net debt of about ¥11.3tn.

NAV by holding, ¥tn, at 30 June, 30 September and 9 November pro forma: Arm 43.4, 32.1 and 18.6; Vision Funds including OpenAI 20.8, 20.6 and 19.7; SoftBank Corp. 2.4, 2.9 and 2.7; others 5.7, 4.3 and 3.7. Source: SoftBank Group Q2 FY2026 earnings presentation, slides 20 and 77; holdings data (CSV). SoftBank spreads net debt across holdings in proportion to their value, so each bar is a share of NAV; the pro forma reprices only Arm, SoftBank Corp. and Intel to 9 November closes.

At 30 September Arm was 54% of NAV and the Vision Funds 34%; OpenAI alone was about 21% of holdings, up from about 17% in June on SoftBank's method, so Arm and OpenAI together were about three quarters. At 9 November prices Arm's share had dropped to 42%, below the Vision Funds' 44%, because Arm's price fell while OpenAI stayed at SoftBank's own valuation.

At 30 June, Arm alone was worth ¥53.1tn before its margin loan: about 923 million American depositary shares at $354.57, according to the investor presentation. OpenAI is held through Vision Fund 2 and was carried at ¥14.5tn, or $89.6bn against $44.6bn invested, per the financial report. That was 75% of Vision Fund 2 and about 17.5% of all holdings (our calculation).

The February 2026 OpenAI investment commitment is now fully funded. SoftBank committed $30bn more on 27 February 2026 at a $730bn pre-money valuation and paid it in $10bn tranches on 1 April, 1 July and 1 October 2026, according to its financial report for the six months to 30 September. Cumulative investment reached $64.6bn, for about 13% of OpenAI. At 30 September, before the third tranche, the stake was carried at ¥15.7tn, or $99.6bn against $54.6bn invested, with the valuation unchanged since March.

The financing changed shape in the quarter. The $40bn bridge facility of March 2026 had $25.9bn outstanding when SoftBank repaid it on 15 September, six months before its 25 March 2027 maturity, and the undrawn $10bn was cancelled. The repayment came mainly from an $11.87bn term loan due by September 2028, $7.35bn of prepaid forward contracts backed by all of SoftBank's Intel shares and $5.0bn added to the Arm margin loan. The October OpenAI tranche was paid from foreign-currency bonds issued in September.

At 30 September the Arm margin loan stood at ¥3.9tn and the SoftBank Corp. margin loan at ¥1.2tn, both without recourse to SoftBank Group (financial report, PDF page 14). Gross interest-bearing debt at the parent, including its financing subsidiaries, SB Northstar and the OpenAI-linked loan, rose from ¥17.7tn in March to ¥23.9tn in six months.

In August a Vision Fund 2 subsidiary also borrowed $10bn, repayable in August 2028 and guaranteed by SoftBank Group, ¥1.5tn at 30 September. Its covenants require cash collateral or prepayment if the fair value of the OpenAI preferred shares falls significantly. On the 6 August 2026 conference call, chief financial officer Yoshimitsu Goto said: "This OpenAI-related financing has recourse to SBG."

SoftBank's LTV, its loan-to-value ratio, was 15.9% at 30 September 2026, after 13.0% at 30 June and 17.0% in March (presentation, slide 22). It deducts the margin loans and collar contracts, ¥6.5tn at 30 September, from both debt and holdings, but counts the OpenAI-linked loan as debt. Adding them back gives about 22.9%, against 17.6% in June (our calculation). We do not use an S&P leverage figure cited in press previews, because we did not obtain S&P's own report.

Why does Masayoshi Son use this structure?

Masayoshi Son wants SoftBank to lead the four businesses he believes artificial superintelligence will need; his finance chief's rules explain how the group borrows against what it already holds to pay for them. In his message in the SoftBank Group Report 2026, dated 24 June 2026, he writes: "I want us to become the undisputed No. 1 ASI Platform Provider in Japan, and in the world."

He names the four areas as AI models, where OpenAI is the "key partner", semiconductors, where "Arm is becoming even more important", AI infrastructure and robots. He compares them to the corners of an Othello board: "Once you secure the four corners in the middle of the game, the position can turn quickly in your favor." He sets a target of NAV of ¥1 quadrillion in 2042, against ¥74tn on 23 June 2026.

He rejects the bubble argument directly: "Some people say there is an AI bubble. I disagree. The AI revolution has only just begun." He also argues that the share price understates the holdings: the ¥74tn NAV on 23 June was "twice as much as SBG's market capitalization of ¥37 trillion as of the same date."

The financing rules come from Goto. In the CFO message, he restates the policy: keep LTV below 25% in normal circumstances, with an upper threshold of 35%, and hold cash covering at least two years of bond redemptions. Large deals are first funded with a bridge facility, then refinanced.

Goto separates what SoftBank controls from what it does not: NAV "is not something we can fully control. By contrast, LTV is something we can directly control."

On the August call he explained that the fall in LTV to 13% came from "the rapid increase in the value of our assets", "rather than a deliberate decision on our part to build up a buffer." He said the $40bn bridge facility "matures in March 2027" and that SoftBank intended to refinance it early, through bonds, loans, margin loans, derivatives or asset sales. It did so by 15 September, mainly with a term loan, an Intel-backed derivative contract and a larger Arm margin loan.

Historical performance versus MSCI World

A hypothetical ¥100 in SoftBank Group shares at 29 December 2000 was worth ¥2,340 on 9 November 2026, with dividends reinvested, against ¥874 in MSCI World net total return in yen. That is 12.97% a year against 8.74% over 25.9 years. Both are measured in yen, monthly, from Tokyo closing prices and MSCI's own index levels, before fees and taxes.

¥100 at end-2000 became ¥2,340 in SoftBank shares and ¥874 in MSCI World net by 9 November 2026; 2025: +92.4% and +20.8%; 2026 to 9 November: -22.5% and +11.5%. Source: 9984.T closes and dividends from Yahoo Finance and MSCI World net JPY month-end levels (index 990100); annual returns (CSV), monthly inputs (CSV).

The yearly bars show how the lead was earned: in a few enormous years. SoftBank gained 143% in 2003, 200% in 2005 and 195% in 2013, and lost 47% in 2001, 53% in 2006 and 32% in 2021. MSCI World had one year below -28%, 2008, when SoftBank lost less.

The cash benchmark is the Bank of Japan's overnight call rate, which compounded at 0.11% a year over the period. SoftBank’s higher Sortino ratio reflects more excess return per unit of measured downside deviation; it does not establish that losses were less frequent. Its lower Sharpe ratio indicates less excess return per unit of total volatility. At 30 September 2026, the ten-year comparison still favoured SoftBank, 23.7% a year against 17.9%; the fall since then reversed it.

SoftBank sits far to the right of the MSCI World and cash line: 52.5% annualised excess-return risk for a 24.9% mean annual return, against 17.8% and 10.0% for the index. Source: the same inputs plus the Bank of Japan call rate series; frontier data (CSV). The curve covers only long-only mixes of MSCI World and yen cash, rebalanced each January, without leverage or costs; it is not an efficient frontier for SoftBank's own opportunities.

The gap between the 24.9% mean and the 12.97% compound rate is the cost of volatility on a single balance: a 50% loss needs a 100% gain to recover.

SoftBank's own NAV gives a second, shorter comparison. From June 2021 to June 2026, at its 21 reported quarter-ends, NAV per share grew 26.35% a year, the share price with dividends 25.89% and MSCI World 20.31%, all in yen. NAV per share fell in 8 of the 20 quarters and the index in 7.

Source: SoftBank earnings datasheets, NAV and LTV sheets, Q2 FY2021 to Q1 FY2026; figures before 2026 divided by four for the January 2026 share split; quarterly NAV data (CSV). The discount ranged from 18.9% at December 2025 to 58.2% at March 2025. Dividing SoftBank's 30 September NAV by its June share count gives about ¥10,490 a share against a ¥6,430 price, a 39% discount; at the 9 November pro forma it is about ¥7,840 against ¥3,400, or 57% (our calculations).

NAV is SoftBank’s pre-tax figure and includes private stakes at its own valuations. Quarterly marks can smooth changes in those private stakes, and the share price can also move as the discount changes; the two measures describe different risks.

The crash record belongs to the shares. Measured on daily closes, SoftBank has had six falls of more than 55% since 2000, while MSCI World in yen had one, in 2007-2009.

Source: daily inputs (CSV); our calculation on split-adjusted closes, without dividends. The price provider carries some closes across exchange holidays; equal highs use their first observation, so the 2005-2008 episode starts on 30 December 2005, not the carried-forward 3 January price.

Most of these losses were SoftBank’s own. In the 29 months when SoftBank fell more than 15%, MSCI World averaged -3.7%; in the 43 months when the index fell more than 5%, SoftBank averaged -7.6%.

Copio's reading

What does the record say about compounded growth?

In this historical sample, combining SoftBank with MSCI World compounded faster than holding SoftBank alone. That does not establish which allocation will grow fastest in future. We take the objective as the expected logarithm of a long-term shareholder's wealth in yen, with no withdrawals. In the usual approximation, compound growth is the mean return minus half the variance; at 52.5% volatility that penalty is about 13.8 points a year.

The comparison selects weights with hindsight from a restricted set of alternatives. Holding the full 2001-2026 path fixed, we mixed SoftBank shares with MSCI World and rebalanced each January. Compound growth peaked at a 70% SoftBank weight, 13.76% a year, above the 12.97% of SoftBank alone.

Over the ten years to 9 November 2026 the peak was at 40% SoftBank, 18.38% a year, against 15.86% for SoftBank alone and 17.57% for the index. Even with perfect knowledge of the outcome, holding SoftBank alone compounded less than holding it with the index. These are in-sample results on one stock's realised path, not an optimum for the future; our blends data (CSV) shows each weight.

The forward question is whether Arm and OpenAI will earn enough to compensate for being most of the portfolio. No market price exists for OpenAI; its value rests on a $730bn pre-money round and SoftBank's own fair value.

Financing adds a cost: finance cost was ¥790.7bn in the six months to 30 September, against ¥340.0bn a year earlier, according to the financial report. The report attributes the rise mainly to interest on the Arm margin loan, the bridge loans and more bonds, plus the bridge's remaining unamortised financing costs, recognised in full when it was repaid early. The group reported a six-month net loss of ¥768.0bn, after a ¥597.2bn loss on revised power-equipment orders for its US data-centre projects.

Is it crash aware and protected?

It is aware of leverage limits but not protected against a fall in its main assets, which is what happened from June. The financial policy is real and specific: an LTV ceiling, a cash floor of two years' bond redemptions and non-recourse margin loans on Arm and SoftBank Corp. shares. At 30 September, the ¥5.0tn cash position exceeded the ¥1.6tn of bonds due over the next two years (presentation, slide 23).

What the policy does not do is pay out in a fall. SoftBank discloses no hedge on Arm or OpenAI.

Since August its Intel shares, ¥1.6tn at 30 September, have sat under prepaid forward collar contracts that raised $7.35bn; collars limit both the gain and the loss on the shares, and these cost a ¥321.9bn derivative loss as Intel rose before the quarter-end (financial report, PDF page 9). The only other collar contracts are on T-Mobile shares, a ¥0.26tn liability at 30 September. The LTV improved because Arm rose, as Goto said, so it worsens when Arm falls.

On 30 September figures, holdings would have to fall 36% for LTV to reach 25%, with net debt unchanged (our calculation). SoftBank's own pro forma at 9 November prices puts NAV at ¥44.7tn, 25% below 30 September; with net debt unchanged, that implies an LTV of about 20% (our approximation, not a SoftBank figure).

Early refinancing removed one risk: the bridge, due in March 2027, was repaid in September, before the steepest part of the fall; Arm lost a further 35.7% in dollars from 30 September to 9 November. It moved debt onto collateral instead: the Arm margin loan rose to ¥3.9tn, and the ¥1.5tn OpenAI-linked loan carries margining that follows OpenAI's valuation.

A borrower whose collateral falls may have to post cash, sell assets or borrow on worse terms, the pressure that forced Aschenbrenner's Situational Awareness fund to sell its public book in July 2026. The two structures use different leverage measures, so the comparison identifies a financing mechanism rather than a probability of forced sales.

Is it diversified without sacrificing growth?

The measured listed holdings share substantial AI exposure. SoftBank Corp., the less-correlated telecoms holding in our sample, was only about 5% of holdings at 30 September 2026. Over 37 months from October 2023 to October 2026, SoftBank's monthly yen returns correlated 0.68 with Arm, 0.57 with Intel and 0.56 with MSCI World. SoftBank Corp., the Japanese telecoms operator, correlated 0.04 with SoftBank Group and 0.06 with the index.

October 2023-October 2026, 37 months: SoftBank Group with Arm 0.68, Intel 0.57, MSCI World 0.56 and SoftBank Corp. 0.04. Source: Yahoo Finance closes and dollar-yen rates, MSCI 990100; correlation matrix (CSV), inputs (CSV). Thirty-seven observations give wide confidence intervals, and these are ordinary months, not crash estimates.

The fall since June is the stress test. From 30 June to 9 November 2026, Arm lost 47.5% and Intel 57.5% in dollars, while SoftBank Corp. gained 14.6% in yen. In SoftBank's own NAV, Arm's part fell from ¥43.4tn at 30 June to ¥18.6tn at 9 November prices, while SoftBank Corp.'s rose from ¥2.4tn to ¥2.7tn; the Intel stake has been under collar contracts since August.

OpenAI has no market price, so its moves cannot be measured, but its business and Arm's depend on the same AI spending. Our study of investments that fall together shows how correlations rise in bad months across asset classes.

Copio does not grade this portfolio in its checker, because single stocks, private stakes and holding-company borrowing cannot be represented by an ETF replica.

For a household, the useful checks are specific. If you hold SoftBank shares to own OpenAI, at 30 September more than half of what you bought was Arm. Count that look-through Arm exposure alongside any semiconductor or Nasdaq fund you hold, and remember that a holding company's borrowing adds to your own. The roughly 57% discount at 9 November prices can widen when you need to sell; our recovery maths shows that a 60% loss needs a 150% gain to recover.

To make the household check concrete, a hypothetical 10% SoftBank position losing 60% would subtract 6% from the starting portfolio if everything else stayed flat. Other AI holdings could fall too. In copio’s private beta, use the combined allocation and fund look-through views to inspect overlapping exposures, then check SoftBank’s private stakes and borrowing against its own reports. The app view does not turn this holding company into a fully replicated portfolio.

For the same AI theme without borrowing, our ARK study examines a concentrated technology fund, and our Berkshire Q3 2026 review a holding company that keeps its cash.