SoftBank takes a $3.7 billion hit on AI power equipment it ordered and then cancelled

SoftBank booked a $3.7 billion charge in its latest quarter for power equipment it had paid for in advance and then did not need. The SoftBank power equipment writedown is small against the hundreds of billions pledged to AI infrastructure this year, but it is one of the first clean, dated examples of a cost that comes from ordering the buildout faster than the plans behind it could settle.

What SoftBank reported

SoftBank released fiscal second quarter results on Tuesday, November 10, 2026, after the Tokyo close. According to investingLive's summary, the group posted a net loss of about 1.12 trillion yen, roughly $7 billion, compared with a 2.5 trillion yen profit a year earlier. The main drag was a writedown of about 600 billion yen, around $3.7 billion, on "nonrefundable advance payments for power-generation equipment intended for US data centres." Smaller drags were losses tied to Intel's share price decline and higher financing costs.

The rest of the results were not weak. Revenue came in at 2.22 trillion yen against a 2.02 trillion yen consensus from LSEG, and net asset value hit a record 59.8 trillion yen at the end of September, up 20 trillion yen from the prior fiscal year end. Loan to value stood at 15.9% with about 5 trillion yen of liquidity. SoftBank also completed a further 3 trillion yen investment in OpenAI, taking its stake to roughly 13%.

What the charge actually is

Semafor's Chloé Farand describes the charge as following the cancellation of a surplus of orders for gas producing equipment meant to power AI data centers that SoftBank ended up not needing. The wording, as summarized by investingLive, is: nonrefundable advance payments. That is the core of the story. When developers pay deposits to secure a place in a turbine maker's production line, the money is gone if they cancel. SoftBank is eating that cost rather than keep paying for capacity it does not plan to use.

The company has not said which projects the cancelled orders served, and neither Semafor nor investingLive gives the value of the cancelled orders or the specific equipment. Until SoftBank names them, that remains unresolved.

Why turbines, and why now

Gas turbines are the scarcest part of the power supply chain for AI campuses. Customers face multiyear waits, and developers are increasingly looking for second hand equipment, per Semafor. Siemens Energy's chief executive told Semafor: "Some of the growth rates for data centers were unrealistic, it's just that nobody wanted to hear it." Siemens Energy itself expects to triple its dividend in 2027 as US data center demand lifts its gas turbine and grid equipment sales, so the maker is not describing a collapse in demand. Our explainer on the turbine bottleneck, published alongside this report, covers the backlog numbers.

The pattern is a familiar one in a shortage: buyers reserve more capacity than they will use because they cannot know which sites will clear permits and financing, then release what they cannot place. That is rational for the first mover and costly for the last one holding the deposit.

How to read a loss like this

A writedown is an accounting event, and it can be larger than the cash that left. Here, the company says the payments were advance deposits that cannot be refunded, so the loss reflects money already spent. That is different from a fall in the value of an asset SoftBank still owns. It also sits inside a quarter in which SoftBank's net asset value reached a record 59.8 trillion yen, per investingLive, so the group is not short of balance sheet. The point for the AI sector is narrower: a developer with ample capital still chose to pay to walk away from capacity rather than build around it.

That tells you something about the market for turbines. Reservations are sold as scarce, and holding one is treated as a hedge. SoftBank's quarter shows the hedge has a price when the plan changes, and that the price is paid in full and up front. Other developers with large reservations will face the same decision when their own sites change, which is why we flag the next few earnings seasons as the place to look for similar charges.

The Nvidia link

The loss arrived in a week when investors were already asking who underwrites the buildout. Semafor reports they worry Nvidia could scale back its practice of financing AI hardware now that Jensen Huang has resigned, and quotes a fund manager: "I think the fear is that the music might stop." We cover that angle in our report on Nvidia's CEO resignation and the earlier Nasdaq selloff. A Federal Reserve governor has also warned about AI related debt, as we reported in our coverage of Governor Bowman's speech.

What this does not show

It does not show that AI power demand is shrinking. Siemens Energy, the turbine maker, is still expecting higher dividends on US data center demand. It also does not show a pattern: one developer's cancelled reservations are not an industry trend.

Our take

We would treat this as a data point on the cost of speed. The AI buildout is being financed with deposits, backstops and reservations that are cheap to make and expensive to unwind. SoftBank's charge puts a number on one unwind. The next signals worth watching are whether other large developers disclose similar cancellations, whether turbine makers report slot resales, and whether the November 17 Nvidia earnings call says anything about its own customer commitments. One writedown is a footnote; three would be a trend.