News

SoftBank just swung to a $7 billion loss on its AI power gamble

SoftBank swung to a $7 billion loss as a power equipment pivot cost $3.7 billion. Our read: locking in AI power early hits earnings before the upside shows.

SoftBank just reported a fiscal second-quarter net loss of 1.12 trillion yen ($7 billion), weighed down by a costly power equipment procurement loss and a slide in Intel shares—flipping a 2.5 trillion yen profit from the same quarter a year earlier.

The core wound is straightforward: SoftBank pre-bought power generation equipment to lock in capacity for U.S. AI data center projects, then changed those procurement plans. That pivot crystallised a 597.2 billion yen ($3.7 billion) loss across the six months ending September 30. Cumulative payments for the equipment—tied to projects including the PORTS-Pike Technology Campus in Pike County, Ohio—totalled $9.48 billion by quarter-end. Intel share declines added pressure on top. The Vision Fund segment swung to a quarterly loss of 81.58 billion yen from a profit of 2.38 trillion yen a year earlier; at Vision Fund 1, Coupang and Grab share declines offset a $1.3 billion gain on ByteDance. Revenue held: 2.22 trillion yen in the quarter against LSEG estimates of 2.02 trillion yen, up 13.3% across the half-year on telecom and AI computing growth.

The OpenAI position is the counterweight. SoftBank's cumulative investment stands at $54.6 billion, fair value at $99.6 billion, cumulative gain at $45 billion—but a potential valuation increase from a new OpenAI funding round hasn't yet been reflected in results and could be reflected in the fiscal third quarter. SoftBank also announced plans with Grab and Malaysia's PETROS to commit a combined $1.85 billion in initial capital toward AI infrastructure in Malaysia, targeting at least five gigawatts of AI computing capacity long-term.

The commitment-versus-payoff gap

SoftBank's loss illustrates what it can cost to bet on AI infrastructure ahead of need. The company made a rational call—lock in power generation equipment amid concerns over potential shortages and long procurement lead times—and booked a $3.7 billion loss when the plan changed. It rhymes with the logic pushing hyperscalers into long-term power purchase agreements and data center operators to acquire generation assets outright. The same pressure is visible in SoftBank's own 20-year lease with OpenAI for eight gigawatts of IT capacity at a data center project in Ohio: the obligation is long and capital-heavy; the return is still forming.

Power is increasingly a key constraint on AI scale. Whoever controls generation capacity stands to set the pace of the buildout—and as SoftBank just showed, controlling it early carries a price that shows up in earnings before the upside does. For anyone running or funding AI infrastructure, the question is no longer whether to commit capital ahead of demand; it's how much of that commitment you can absorb when the plan shifts.

Get the Daily Frontier

Stories like this one land in your inbox every weekday morning. Subscribe free at factua.com/blog.

Know someone who should be reading this? Send them to factua.com/blog to subscribe.