SoftBank Corp's AI Ambition Is a Leveraged Gamble

SOFTBANK CORP reported a record first quarter on 4th August 2026, with revenue of ¥1,815bn ($11.5bn), up 9.4% year on year, and operating income of ¥302bn, well above analyst expectations. The headline figures would satisfy investors in almost any mature telecom. Yet the details reveal a more disquieting story. The company that makes these results is not the same one that existed a year ago. It is becoming, by design and by necessity, an artificial-intelligence infrastructure company. The question is whether the arithmetic supports that transformation.

The structural pivot is already visible in the numbers. Cloud and AI revenue grew by 31% in the quarter. Management has raised its compound annual growth forecast for the segment to 30% and set long-term margin targets of 30% to 40%, a significant step up from earlier mid-term plans. The enterprise segment, which now houses cloud and AI services, saw its operating income surge 28% to ¥62.2bn. Mr Junichi Miyakawa, the chief executive, declared that the company is "entering the monetisation phase" of its heavy AI investments. He indicated that full-year results could finish "mid-hundreds of tens of billions of yen" above the start-of-year plan.

The trouble is on the other side of the balance-sheet. Net income rose by a mere 3.3% to ¥150bn, far behind the operating-profit gain. The gap was eaten by rising interest costs. Free cash flow turned negative, as strategic investment and working-capital timing drained the coffers. Net leverage increased, helped along by new euro-denominated bonds and upcoming data-centre spending. AI infrastructure depreciation is already weighing on enterprise profits, creating a timing mismatch between the costs of building and the revenue that will eventually flow from it.

This is the familiar pattern of a leveraged bet on infrastructure. The company borrows to build, depreciation and interest eat into early profits, and the thesis depends on future margins expanding enough to cover the cost of capital. It is not an unusual play — global hyperscalers collectively plan to spend roughly $740bn-$760bn on capital expenditure in 2026, most of it debt-financed. But SoftBank Corp is not a hyperscaler. It is a Japanese telecom whose core mobile business is losing customers: the smartphone subscriber base fell by roughly 180,000 in the quarter. Average revenue per user rose by ¥60, but that is an extraction exercise, not growth. The cash-flow cushion that once made a telecom predictable is being spent on a hope.

The ambition only accelerates from here. On 2nd July, SoftBank Corp and its parent, SoftBank Group, announced SB Neo, a Delaware-incorporated joint venture to operate a neocloud business in America. SoftBank Corp will hold 51%; the group will hold 49%. The venture is scheduled to launch in fiscal 2027, with a stated goal of deploying 10 gigawatts of AI compute capacity by 2030. It will use SoftBank's Infrinia AI Cloud OS, a software stack for GPU resource-sharing that has been in beta in Japan since April. Mr Miyakawa cites access to gas-fired energy as a competitive edge.

To be sure, power is the genuine bottleneck in the AI infrastructure game. Data centres are being delayed or cancelled because local grids cannot cope, and countries with dispatchable energy have an advantage. SoftBank Group's 10-gigawatt energy-and-infrastructure programme gives SB Neo a head start that pure-play neocloud operators lack. The company's domestic GPU cloud experience, though small, means it is not starting from zero.

Yet the neocloud business model is, by McKinsey & Co's own characterisation, "extremely fragile". The hardware it rents — GPUs — is increasingly commoditised, limiting operators' ability to differentiate. SB Neo will compete against CoreWeave, which already has scale; against Meta, which is launching its own neocloud offering; and against AWS, Azure and Google Cloud, which have deeper balance-sheets, established enterprise relationships and the pricing power to undercut newcomers. Ten gigawatts is an impressive target, but it is also a massive capital commitment that will land squarely on SoftBank Corp's balance-sheet as a consolidated subsidiary.

And then there is the parent. SoftBank Group's leverage has been a standing concern. The group has committed more than $60bn to OpenAI and related infrastructure, with $30bn of obligations due in the second half of 2026. A $40bn bridging loan matures in March 2027. S&P Global Ratings calculates the group's loan-to-value ratio at 33% as of March, including margin loans backed by its stake in Arm, though it expects improvement by June. S&P upgraded the credit outlook from negative to stable in July, driven by Arm's rising share price. That is the kind of collateral-dependent position that looks fine until the asset price turns.

For investors in SoftBank Corp, the relevant risk is not whether AI will grow. It is whether the margin expansion from cloud services can outpace the financial drag of the debt required to fund them, while the core telecom business continues to contract. The company's full-year guidance — ¥7.5tr in revenue, ¥1.1tr in operating income, ¥560bn in net income — remains unchanged. Those figures are slightly below average market expectations, which suggests management is hedging against the very capex and interest costs that are already showing up in the quarterly numbers.

The broader lesson is structural. SoftBank Corp's results illustrate the trap that confronts legacy operators everywhere: the old business is stable but shrinking, so management borrows against it to buy a new one. The incentive is clear. In a market that rewards growth multiples, being a telecom with a cloud segment is worth more than being a telecom full stop. The constraint is equally clear. Cloud infrastructure is capital-intensive, competitive and cyclical. It does not carry the predictable cash flows that once made telecoms attractive in the first place.

SoftBank Corp's first quarter is a reminder that operating profit is not the same thing as economic profit. The difference is the cost of the capital required to earn it. If cloud margins reach the 30%-40% range management now promises, the arithmetic may work. If they converge toward the low margins that characterise commoditised infrastructure, the leverage that made the bet possible will also magnify the loss. That is the gamble.

Better to start now than never, perhaps. But a gamble is still a gamble.