TOKYO, November 10, 2026 — The SoftBank Q2 loss reached ¥1.115 trillion for the July-September quarter after a costly revision to equipment procurement for US artificial-intelligence data centres and a sharp increase in financing expenses. The result reversed a ¥2.502 trillion profit in the same quarter a year earlier, when large gains on investments had lifted the Japanese technology investor’s earnings. It places fresh scrutiny on founder Masayoshi Son’s strategy of committing enormous sums to AI software, chips and the infrastructure needed to run them.
The quarterly loss was disclosed with SoftBank’s results on November 10. For the full six months ended September 30, the group recorded a net loss attributable to owners of ¥768.0 billion, compared with a profit of ¥2.924 trillion a year earlier. The distinction matters: SoftBank had a profit in the first quarter, but the July-September loss was large enough to pull the half-year result into the red. Net sales for the six-month period rose 13.3% to ¥4.235 trillion.
The largest clearly identified project-related charge was ¥597.2 billion, equivalent to about $3.73 billion, arising from a realignment of procurement plans for power-generation equipment intended for AI data-centre projects in the United States. SoftBank said its subsidiary Silver Bands 7, in which the group and SoftBank Vision Fund II have invested, had reserved manufacturing capacity and entered equipment-purchase contracts to secure power infrastructure for existing and planned projects. The company said requirements had evolved and the procurement strategy had to be revised.
The group’s financial report shows that cumulative payments for the procurement programme from the previous financial year through September 30 reached ¥1.501 trillion, or about $9.48 billion. Of the recognised loss, ¥457.0 billion related to non-refundable advance payments used to reserve manufacturing capacity and purchase equipment. A further ¥140.2 billion came from contract modifications and settlements. The figures show how expensive it can become to secure power-generation equipment before the design and timing of a large data-centre project are settled.
SoftBank said the subsidiary became entitled to receive rebates of up to $2.1 billion over the next 10 years on future equipment purchases, subject to conditions. The company did not recognise those potential rebates as an asset because the amount could not be determined at September 30. The rebates therefore should not be treated as cash already recovered or as an immediate offset to the charge. Their value will depend on future purchases and whether the contractual conditions are met.
The procurement arrangements included equipment for the PORTS-Pike Technology Campus in Pike County, Ohio. SoftBank’s report also disclosed a separate impairment loss of ¥72.5 billion on property and buildings acquired in Lordstown, Ohio, for a manufacturing facility intended to support data-centre-related equipment. That charge followed a reassessment of the assets’ recoverable value in light of the facility’s intended use and future plans. It is separate from the ¥597.2 billion procurement loss and should not be combined with it as though it were one charge.
The company has not publicly set out every customer requirement or engineering change behind the procurement revision. Its explanation was that the needs of existing and planned projects had evolved. For large AI campuses, power supply is a central design constraint: developers must match generation equipment, grid connections, cooling, land, data-hall capacity and the timing of customer demand. If a project changes scale or configuration after suppliers have been booked, advance payments and cancellation terms can become costly.
SoftBank’s results are increasingly shaped by its concentration in artificial intelligence. The company’s report says cumulative investment in OpenAI stood at $54.6 billion at September 30, with a fair value of $99.6 billion and a cumulative investment gain of $45 billion at that date. SoftBank made a further $10 billion investment in October, completing the $30 billion of follow-on investments it had committed to in February. Reuters reported that the group’s total investment in OpenAI had reached $64.6 billion and represented about 13% of the company.
The distinction between investment cost and fair value is important. The $99.6 billion valuation in the September report is an accounting estimate at that date, not cash SoftBank can immediately realise. OpenAI remains privately held, and the value of the stake can change with future funding rounds, business performance and the terms on which shares can be sold. A public listing could eventually provide a route to liquidity, but timing and valuation are uncertain.
SoftBank’s exposure extends beyond a single model developer. The group owns Arm, has invested in semiconductor design businesses and has expanded its digital-infrastructure holdings. It completed the acquisition of DigitalBridge on September 30 for about $3.1 billion, adding a specialist investor in data centres, fibre networks, cell towers and other digital infrastructure. The company has also grouped Arm, Graphcore and Ampere-related operations within its AI Computing segment, linking chip design and computing capacity to the broader AI strategy.
That structure gives SoftBank potential participation in several layers of the AI economy: the models that generate demand, the chips that process workloads and the facilities that provide power and computing capacity. It also creates correlated risk. If AI adoption grows more slowly than expected, if customers delay construction or if the economics of large data centres deteriorate, several parts of the portfolio could face pressure at the same time.
The data-centre charge was not the only reason for the weaker result. SoftBank’s total investment gains for the six months ended September 30 fell 56% to ¥1.726 trillion from ¥3.927 trillion a year earlier. Gains at the SoftBank Vision Funds segment dropped to ¥705.6 billion from ¥3.415 trillion. The comparison is unusually demanding because the previous-year period included substantial valuation gains, particularly from investments in private technology companies.
The investment business of holding companies recorded a gain of ¥1.078 trillion in the first half. SoftBank said that included a ¥1.063 trillion gain on Intel shares. That contribution helped offset weaker results elsewhere, but gains on listed holdings can move in either direction as share prices change. The Vision Fund result, which depends heavily on private-company valuations and exits, was not strong enough to reproduce the prior-year performance.
Financing costs also rose sharply. SoftBank recorded ¥790.7 billion in finance costs for the six-month period, an increase of 132.5% from a year earlier. Reuters reported that financing costs in the July-September quarter alone more than doubled to about ¥461.9 billion. The company attributed the increase to borrowings, higher funding rates and the accounting recognition of remaining financing costs after it repaid a large bridge facility before maturity in September.
The report says SoftBank raised ¥17.16 trillion and repaid ¥9.70 trillion on a consolidated basis during the first half. It fully repaid and closed a bridge facility through longer-term funding before its scheduled maturity. The company’s cash and cash equivalents rose to ¥7.43 trillion by September 30, but the larger financing bill shows that the group’s investment programme is not costless even when it can refinance or repay individual facilities.
SoftBank’s balance sheet has expanded as it has added investments and digital infrastructure assets. Total assets reached ¥69.39 trillion at September 30, while total equity was ¥19.40 trillion. Those figures provide scale, but they do not remove the need to manage borrowing, collateral and the volatility of the assets used to support financing. A sustained decline in the value of major holdings could make lenders and investors more sensitive to the group’s leverage.
SoftBank’s equipment charge highlights a practical problem for the AI industry. Data-centre developers are racing to secure electricity, generators, transformers, cooling systems and grid access, often before every customer requirement is final. Suppliers may demand advance payments to reserve scarce manufacturing capacity. Those commitments can protect a project from shortages, but they also leave developers exposed if the project changes or the required equipment no longer fits the revised design.
Power has become a strategic constraint because high-performance AI systems require large, continuous electricity supplies and substantial cooling. A campus can have land and computing hardware available but still be unable to operate at the planned scale if power infrastructure is delayed. That has encouraged technology groups and infrastructure investors to move earlier in the procurement process. SoftBank’s loss shows the other side of that strategy: securing capacity too early can create expensive obligations when project plans change.
The group’s exposure also reflects the uncertain economics of AI infrastructure. Data-centre investment requires large upfront capital, while the revenue depends on customers committing to use capacity over time. Developers must estimate demand, utilisation, power prices and the pace at which new generations of chips make existing equipment less competitive. If customer demand shifts or a project is redesigned, the cost of equipment reservations and construction can remain even when the original business case changes.
SoftBank’s investment in OpenAI may offer a route to large returns if the company expands and achieves a public listing at a strong valuation. But the same exposure means the group is highly sensitive to the future economics of one of the world’s largest AI developers. The company’s October investment was completed even as its latest results showed how much capital is required to build the supporting infrastructure. Investors will assess the model investment and the data-centre spending together rather than as unrelated bets.
SoftBank shares had lost nearly half their value since June, according to Reuters, as investors questioned the size of the group’s AI commitments and the financing needed to support them. The stock rose sharply before the results were announced on Tuesday, but the quarterly loss gives shareholders a fresh set of numbers to weigh. A market rally ahead of an announcement does not establish how investors will value the company after they have assessed the details of the report.
For customers and businesses planning AI capacity, the result is a reminder that data-centre construction is not simply a matter of ordering servers. Power procurement, site preparation, equipment design and long-term financing can determine when a facility is ready and how much it costs. For suppliers, the case highlights the importance of contract terms governing advance payments, modifications and cancellations. For investors, it raises questions about how much risk should sit on the balance sheet of a technology investor that is also financing physical infrastructure.
The next test for SoftBank will be whether its data-centre projects can be brought back into line with customer requirements without further large write-downs. Investors will also watch whether the conditional equipment rebates become usable, how quickly the company can convert AI demand into cash flows and whether financing costs begin to ease. The group’s Vision Fund returns, the valuation of its OpenAI stake and the performance of Arm will remain important indicators of how much support its wider portfolio can provide.
SoftBank’s quarterly loss does not by itself prove that its AI strategy has failed. The group continues to hold valuable technology and infrastructure assets, and the procurement charge is tied to a specific set of project changes rather than a write-down of every AI investment. But the result shows that scale and ambition do not eliminate execution risk. The ability to match capital commitments to actual customer demand, power availability and project schedules will be as important as the growth of AI models themselves.
Comments (0)
Be the first to comment. Comments are published after moderation.