SoftBank’s ¥1.1 Trillion AI-Era Loss: When the Data Center Power Bill Comes Due

When AI meets the bill for electricity

SoftBank Group has become one of the defining financial backers of artificial intelligence. Its November 10 earnings put a price on that ambition: a ¥1.1153 trillion net loss attributable to the parent for July through September 2026, against a ¥2.5022 trillion profit in the equivalent quarter a year earlier.

The most revealing element is not a single failed bet on a technology stock. Equipment purchased or reserved for American AI data centers needed a new procurement plan. Debt and equity-linked financing created additional expenses. In an industry that promises increasingly powerful intelligence, the hard constraints of electricity, construction schedules and capital markets have become impossible to ignore.

Two earnings clocks, two different losses

SoftBank's financial report covers the six months ending September 30. For April through September it recorded ¥4.235 trillion in revenue, up 13.3%, and a ¥768.0 billion loss attributable to parent shareholders. The previous six-month period produced a ¥2.9241 trillion attributable profit.

The larger ¥1.1153 trillion figure is the second quarter alone, July through September. The first fiscal quarter had generated about ¥347.3 billion in attributable profit, explaining why the half-year deficit is smaller than the second-quarter deficit. Readers should never mix these time periods or confuse net income with pretax results.

The consolidated group's net loss including non-controlling interests was ¥541.6 billion over six months. The parent-attributable loss was ¥768.0 billion; the accounting distinction matters to shareholders of SoftBank Group Corp.

The $9.5 billion procurement gamble

The company tried to secure scarce power-generation equipment early, anticipating long lead times and intense demand as data-center projects expanded. By the end of September, its cumulative payments associated with these purchases and capacity reservations had reached approximately $9.5 billion, according to its accounting presentation.

Changing requirements for existing and planned US projects, including the PORTS-Pike power-generation project in Ohio, forced a realignment. SoftBank recognized ¥597.2 billion in losses in the first half, including about ¥457.0 billion tied to nonrefundable advances on equipment no longer expected to be purchased and ¥140.2 billion linked to contract modifications or settlement.

The company says a portion of the commitments remains directed toward equipment purchases and that it negotiated the right to as much as $2.1 billion in future discounts under specified conditions. A conditional future rebate is not cash already recovered; nor is the entire $9.5 billion equivalent to the recognized loss. Those distinctions are essential to a fair reading.

OpenAI paper gains are not the same as cash receipts

SoftBank's disclosure places cumulative investment in OpenAI at about $54.6 billion as of September 30, valued at $99.6 billion, with a cumulative valuation gain of approximately $45.0 billion. Another $10 billion was invested in October, taking the amount invested to roughly $64.6 billion.

Those valuation figures do not establish that the company could sell its entire position at that valuation, on that date, without restrictions or market effects. Nor do they tell readers OpenAI's operating cash flow or when an investor might ultimately receive distributions.

The contrast is instructive: private-company equity valuations can move very differently from the immediate contractual costs of building a data center. Model capability, customer revenue, project execution, construction requirements and investment valuations all move on different schedules.

How debt changes the economics of a technology bet

SoftBank's first-half finance costs rose to approximately ¥790.7 billion from about ¥340.1 billion a year earlier. In July–September alone, press reporting placed those costs at about ¥461.9 billion. This is one reason a discussion confined to investment gains cannot explain the net result.

Borrowing can accelerate investment without selling a prized asset, but creates interest expenses, refinancing requirements and exposure to collateral and market conditions. For an investment group, those constraints can shape decisions as strongly as the commercial performance of portfolio companies.

An accounting loss is not automatically a matching cash outflow during that same quarter. Conversely, the absence of an immediate cash outflow does not make a recognized obligation or impairment financially irrelevant.

Intel shares: an asset can gain while its financing loses

An August 2026 prepaid forward transaction backed by Intel shares raised approximately $7.35 billion. SoftBank's earnings materials describe a roughly ¥321.9 billion derivative loss on the arrangement following a rise in Intel's share price.

Its holding-company investment segment also reported a roughly ¥1.0628 trillion investment gain on Intel shares during the six-month period. Those figures come from different accounting lines and must not be presented as if one mechanically cancels the other in a single transaction.

The episode captures a central paradox of asset-backed finance: stock appreciation can be beneficial to the investor's holdings while creating a negative valuation movement on an associated derivative agreement. Capital access has a structure, and that structure has a cost.

Refinancing a $30 billion bridge loan

In March 2026, SoftBank arranged a $40 billion bridge facility, principally to support further OpenAI investment. It had drawn $30 billion. By September 15 the drawn amount was repaid early, and the $10 billion unused commitment was canceled at the end of September.

SoftBank listed a September term loan of $11.87 billion, the Intel-backed prepaid forward, and a $5 billion enlargement of a margin loan secured on Arm shares among its sources of replacement financing. The group also issued foreign-currency and yen-denominated bonds and entered other loans; their maturities and collateral terms matter as much as the sums raised.

Across the group, financing inflows recorded as borrowing amounted to about ¥17.16 trillion in the half-year, and repayments approximately ¥9.70 trillion. These are gross flows, not a measure of cash freely available for new investments.

A company born in 1981 that kept changing industries

Masayoshi Son founded Nihon SoftBank in September 1981 as a distributor of personal-computer software. The name reflected a plan to become a bank-like infrastructure provider for the information industry, not a bank in the conventional deposit-taking sense.

Its transformation continued through stakes in Yahoo in the 1990s and Alibaba in 2000; Japanese broadband in 2001; the purchase of Vodafone's Japanese operations in 2006; Sprint in 2013; Arm in 2016; and the first Vision Fund in 2017. OpenAI investment beginning in 2024 marked another shift in emphasis.

SoftBank Group Corp., the listed investment holding company behind these results, must also be distinguished from SoftBank Corp., the listed telecommunications operator. They have different business structures and results even where consolidated ownership links them.

A vertical map of the AI economy

The increasingly interconnected portfolio spans model development, chip architecture, server processors, physical data centers and power-generation plans. Arm, Ampere, the consolidated DigitalBridge business and US infrastructure projects illustrate how the group is placing capital across layers of the computing economy.

SoftBank's 2026 annual report describes an Ohio initiative envisioning 10 gigawatts of generation capacity and 10 gigawatts of AI data-center capacity. These are project ambitions and stated plans, not verified operating capacity already delivered. Land, permits, transmission connections, customer commitments, equipment delivery and construction execution must be established separately.

This is why the procurement loss is significant. It arose not from a software experiment but from the sequence in which physical infrastructure must be designed, reserved and purchased.

What investors should measure instead of one quarterly headline

A more useful performance framework separates realized gains from fair-value gains, calculates net asset value against debt, tracks debt maturities and collateral, measures recurring financing cost, and checks the status of infrastructure construction and revenue contracts.

In the first-half accounts, total assets reached roughly ¥69.39 trillion, while equity attributable to parent owners as a percentage of assets fell to 23.6% from 29.0% at March 31. This ratio does not by itself establish a liquidity emergency, but it is evidence that the financing architecture deserves scrutiny.

Undisclosed contract economics for individual data-center projects, returns on future OpenAI funding and the enforceability of every future equipment discount cannot be independently derived from the headline figures. Treating them as settled would go beyond the evidence.

The test ahead: converting ambition into durable returns

The AI investment cycle creates legitimate opportunities: advanced computing can have enormous long-term value, and strategic control of infrastructure can be valuable. But the sequence of supply contracts, debt issuance, equity valuations and delayed facility revenues can expose an investor to substantial losses before those opportunities become cash-generating businesses.

SoftBank's latest quarter is neither proof that artificial intelligence has failed nor evidence that its economics are already assured. It is a vivid demonstration that financial engineering and industrial execution are now inseparable from the technology story. The coming years will test not how large the commitments look on paper, but what reliably operates, what customers pay for, and what eventually returns to investors.