Everything About AI
NextFin News - SoftBank Group shares fell 10% in Tokyo on Wednesday after the Japanese conglomerate reported a net loss of 1.12 trillion yen ($7.1 billion) for the three months ended September 30, its first quarterly loss in almost two years and a stark reversal from the 2.50 trillion yen profit it posted a year earlier. The loss blew past the consensus estimate for a profit of 129 billion yen, and the driver was not SoftBank's famed Vision Fund portfolio so much as a $3.7 billion writedown on U.S. data-center projects — non-refundable advance payments for power-generation equipment tied to sites including an Ohio data-center campus. Masayoshi Son's all-in bet on artificial intelligence now carries a balance sheet with real operating assets, not just mark-to-market stakes, and the market is charging for it.
The quarter's investment results tell a more nuanced story than the headline loss. The flagship Vision Funds logged a 245.6 billion yen gain on investments, while SoftBank's own investment book recorded a 133.59 billion yen loss. Over the six months to September, the group booked a 597.25 billion yen net loss. The market's verdict was immediate: the 10% drop extended a decline that has already erased more than a quarter of SoftBank's value over the past month and cut its market capitalization roughly in half from the all-time high struck in June. The shares were up about 14% heading into the release, which makes the reversal a clean read of new information rather than a "sell the news" trade on old gains.
The Loss Was Not the Vision Fund — It Was the Data Center
The first question investors asked was where the loss came from, and the answer points away from the familiar story of Vision Fund volatility. SoftBank swallowed a $3.7 billion writedown tied to its U.S. data-center projects, and the company's finance chief framed it as a strategic rethink rather than a one-off accounting charge.
Yoshimitsu Goto, the chief financial officer, said the company had decided to "conduct a thorough reassessment" of its previous strategy to secure power-generation equipment for data centres before their construction.
That distinction matters. A Vision Fund swing is a mark-to-market event on a portfolio of private and public stakes — painful, but transient and largely outside management's direct control. A writedown on data-center assets is different: it is an admission that the economics of building and operating AI compute infrastructure are worse than the company modeled when it committed the capital. The mechanism is straightforward and unforgiving. Data centers are long-duration, capital-intensive assets whose value depends on contracted power, tenant demand, and the price of compute over a decade or more. When any of those inputs deteriorates — power costs rise, tenants delay, or the market price of GPU capacity falls as supply catches up with the 2025–2026 buildout — the impairment hits earnings directly and can recur. Unlike a listed equity stake, you cannot sell a half-built data center at yesterday's valuation.
This is why the market punished the stock harder than the headline loss alone would justify: investors are repricing SoftBank not as a venture-capital-style holding company, but as an owner of hard AI infrastructure with execution risk baked in. The numbers bear this out. The Vision Funds' 245.6 billion yen investment gain shows the listed portfolio still works as a trading book. The 133.59 billion yen investment loss on SoftBank's own book, combined with the $3.7 billion writedown, is where the new risk lives. In other words, the quarter separated SoftBank's two identities: the investor, which made money, and the builder, which did not.
The reassessment Goto described is itself a signal. The old playbook — buy the power equipment early, lock in capacity, and let the data-center tenant pay a premium for guaranteed electricity — assumed that AI demand would compound fast enough to make early commitments cheap in hindsight. The writedown says that assumption no longer holds at the margin. Power equipment purchased before a tenant is signed is not an option; it is a sinking asset that depreciates while it waits.
The OpenAI Bet That Cannot Be Marked to Market
At the center of that builder identity sits SoftBank's commitment to OpenAI. The group pledged $30 billion across three equal $10 billion tranches scheduled to close on April 1, July 1, and October 1, 2026 — the largest single investor commitment in a funding round that brought SoftBank's total investment in the ChatGPT maker to $64.6 billion. The final tranche closed on October 1, funded by an $11.1 billion high-yield bond sale that settled two days earlier, the largest of its kind on record.
Here lies the asymmetry that keeps SoftBank's holders awake. The OpenAI stake is not marked to market in SoftBank's quarterly results the way listed holdings are. Its carrying value is largely the cash put in, not the market's current verdict on OpenAI. That is a comfort only until it is not. If OpenAI's valuation resets lower — because revenue growth slows, because a cheaper-model competitor erodes pricing power, or because the capital intensity of training the next generation of models exceeds what the business can earn back — SoftBank's exposure would surface not as a quiet quarterly fluctuation but as a discrete, lumpy impairment. The group has already felt a version of this: OpenAI has reportedly trimmed the valuation target it is seeking from new investors, pulling it down from $1.4 trillion to $1.25 trillion after investor pushback, and closed an $11 billion round at the lower figure.
The warning signs around the broader AI infrastructure trade are multiplying. The Stargate venture — the $500 billion data-center initiative announced at the White House with OpenAI, Oracle, and MGX — has struggled to get off the ground, with no data-center deals completed six months after its unveiling. And SoftBank is not alone in feeling the pressure: shares of the Japanese group have lost more than a quarter of their value in the past month as AI-valuation anxiety spread from U.S. megacaps back to their financiers. On top of that, China blocked SoftBank's $5.4 billion acquisition of ABB's robotics unit earlier this month, closing off a diversification route just as the group needs optionality.
The Arm Collateral That Turns a Stock Drop Into a Balance-Sheet Question
Beneath the earnings headline sits a second-order risk that the 10% move only began to price. SoftBank has pledged 769 million Arm Holdings shares — 72% of the chip designer it controls — behind a margin loan that has reportedly been raised to $25 billion. SoftBank owns 86.4% of Arm, so a falling Arm stock price puts direct pressure on the parent's collateral position. That linkage turns an ordinary tech-stock decline into a potential funding spiral: the more SoftBank's shares fall, the more the market questions the cushion behind the loan, and the less room the group has to raise fresh capital without selling the very asset backing it.
The exposure runs both ways. A SoftBank affiliate supplied 14% of Arm's fiscal 2026 revenue, and as of June 30, the chip designer had booked $577.2 million of revenue from that affiliate it had not yet billed. Under a restructured work order, the affiliate has a fixed $300 million payment due during fiscal 2027, which ends in March. If that cash arrives and the balance keeps falling, Arm's SoftBank revenue is turning into real money. If the balance grows again while SoftBank is paying up to 9.75% to borrow, Arm has counted revenue it may wait a long time to collect. Arm trades at 72 times forward earnings — below its three-year average of 85 times and far from the 201 times it reached in June — but that is still a rich price for the shares that back a $25 billion loan.
The loss is a bigger threat to Arm's stock than to its business. Data-center royalties doubled again last quarter, and the operating business is sound. But until the affiliate's unbilled balance comes down, every SoftBank scare will hit Arm's shares too. The market is starting to treat Arm less as a standalone AI infrastructure play and more as the collateral that keeps SoftBank's funding engine running. This is the transmission channel that converts an earnings miss into a credit question: lower equity value compresses borrowing capacity, which raises the cost of the next commitment, which forces a choice between dilution and asset sales.
The Counter-Thesis: Son Has Been Wrong Before, and Right Bigger
The strongest case against this gloom is also the simplest: Masayoshi Son has built a career on being early, and the market has punished him for it before — only for his biggest bets to define the next cycle. The pattern is not ancient history. After the dot-com crash, SoftBank's net asset value fell more than 90% from its peak, and Son was written off as a relic of a bubble. He then made the Alibaba investment of roughly $20 million in 2000, which grew to more than $100 billion in value before SoftBank sold down most of it by 2023. The 2022 Vision Fund drawdown — one of the largest in venture history — was followed by the AI rally that pushed SoftBank's market value past Toyota's earlier this year. If AI infrastructure demand does compound at the rate Son assumes, today's writedowns will read in hindsight as the cost of admission to the defining asset class of the decade.
The counter-thesis has real teeth. Arm remains the other crown jewel in the AI stack; a recovery in chip demand or a new AI-driven upgrade cycle would flow directly to SoftBank's book. The Vision Funds' 245.6 billion yen gain this quarter shows the investment engine still generates alpha when markets cooperate. And if Stargate does get moving under a project-finance structure, SoftBank could capture the economics of the AI buildout with far less balance-sheet risk than the headline $500 billion figure implies. In that world, today's 10% drop is a panic over a temporary impairment, and the $3.7 billion writedown is a one-time cleanup of overoptimistic assumptions rather than a structural flaw.
But the burden of proof has shifted, and the two prior cycles are not clean analogs. The Alibaba bet was a tiny fraction of SoftBank's balance sheet when made — an asymmetric option, not a $64.6 billion concentrated position. The 2022 drawdown was a broad market event that lifted every private-tech valuation on the way back up; there is no guarantee of a symmetric reflation when the impairment is specific to the economics of one industry's capital spending. The cyclical view — that this is a mark-to-market dip that will revert when AI sentiment recovers — requires three things to hold: AI capex growth must stay intact, OpenAI's valuation must not reset materially lower, and SoftBank's funding cost must not crowd out its next commitments. The structural view — that the AI infrastructure buildout is entering a period of lower returns and higher financing costs — requires only one: that the marginal dollar of data-center capital earns less than the market assumed in 2025. The writedown, the reassessment of the power-equipment strategy, and the collateralized Arm position suggest the market is starting to believe the latter.
Who Benefits and Who Is Exposed
The repricing does not land evenly. The direct beneficiaries of a softer SoftBank are the counterparties and competitors that gain pricing power as the group's urgency to monetize rises: other AI-infrastructure financiers can demand better terms, and rival chip designers can press Arm on licensing as SoftBank pushes its affiliate to convert unbilled revenue into cash. Japanese index funds are also exposed — SoftBank's weight in the Topix means a 10% move in the stock mechanically drags the broader index, which is why the group's swings have repeatedly accounted for a disproportionate share of daily index volatility this year.
On the liability side sit three groups. First, SoftBank's bondholders, who now face a balance sheet that is more infrastructure-heavy and less liquid than the one that priced the $11.1 billion deal. Second, Arm holders, whose shares carry a SoftBank-funding premium that can vanish even if Arm's own fundamentals hold. Third, the Vision Fund's limited partners, who are exposed to the group's need to monetize assets — if SoftBank sells stakes to raise cash, it may do so into a weaker market, crystallizing losses that would otherwise remain unrealized.
What Comes Next
In the short term, expect volatility to stay elevated. The next catalyst is how management frames the data-center writedown on the earnings call and whether it signals more impairments to come. A second consecutive quarter of large writedowns would confirm the structural read; a clean bill of health with no further impairments would support the cyclical-dip view.
Over the medium term, three signals matter. First, OpenAI's next funding round or IPO filing — any valuation below the level at which SoftBank's stake sits would force a reassessment. Second, Stargate's ability to sign actual data-center deals; more delays would validate the execution-risk concern. Third, SoftBank's funding pipeline — whether it can continue to raise debt at manageable cost without selling core assets, and whether the $300 million Arm payment due in fiscal 2027 arrives on time.
The falsifying signal for the bearish read is specific: if SoftBank reports no further data-center impairments over the next two quarters, OpenAI's next funding round clears above $1 trillion, and Arm's unbilled balance falls while the $300 million payment lands, the impairment was a one-off cleanup and the structural-downgrade thesis fails. Conversely, if the group books another large writedown while AI capex growth slows and Arm's stock falls toward the level that strains the margin-loan collateral, today's 10% drop will look like the first step of a longer repricing.
SoftBank's quarter is a clean illustration of a broader truth about the AI boom: the companies that own the picks and shovels can write down their tools, but the investors who financed the diggers cannot mark their stakes to comfort. The market is no longer paying for the dream of AI infrastructure — it is charging rent on it.
Explore more exclusive insights at nextfin.ai.
