Masayoshi Son's bet was ambitious enough to deserve a case study in corporate strategy: deploy the cash of the SoftBank to secure, ahead of any competitor, the manufacturing capacity for natural gas turbines that would power the next generation of artificial intelligence data centers. Instead of a masterstroke, it resulted in a US$3.7 billion accounting loss.
In SoftBank's fiscal second quarter (ending September 30, 2026), the Japanese conglomerate announced a net loss of 1.12 trillion yen — roughly US$7 billion — its first quarterly loss in nearly two years. But the number that captured market attention wasn't the combined result of the investment division, but rather a specific line item within it: a US$3.7 billion write-down on power generation equipment for data centers that SoftBank had already ordered and prepaid, but now no longer intended to use.
The story begins in 2025, when the shortage of gas turbines reached a critical point. Manufacturers, led by GE Vernova, had order backlogs stretching into the 2030s. AI data centers needed power — far more power than any conventional electrical grid could deliver at the expected pace — and the industry's immediate response was to turn to on-site generation. SoftBank saw an opportunity and entered with unusual aggressiveness for an investor.
Through its subsidiary Silver Bands 7, linked to the second Vision Fund, SoftBank committed more than 13 gigawatts of gas turbine capacity between late 2025 and early 2026. The total value of advance payments reached US$9.48 billion. Of these, approximately 9.2 GW would be destined for a natural gas plant in Pike County, Ohio, designed to power the OpenAI data center — part of the Stargate project, which has already received US$64.6 billion in investment from SoftBank itself. The remaining 4 GW were allegedly reallocated to other projects.
But market reality shifted. As the first AI data centers came online, their operators began to understand better what was truly required to keep them running. The premise that on-site gas generation was the fastest and most reliable solution started to be questioned. Real-world consumption data revealed that the variable energy demand of AI training and inference workloads could be managed more efficiently with smaller grid connections combined with large-scale battery storage systems — an architecture that allows compensating for demand spikes and responding to utility requests to reduce consumption during restriction periods.
This shift in perspective was sufficient for SoftBank to reassess its project portfolio. Part of the 13 GW ordered was no longer needed. The company recorded a US$3.7 billion loss on non-refundable deposits and advance payments, and remains with US$6.6 billion in equipment contracts it still intends to use or transfer. But the mere existence of this loss — bigger than the entire deal Anthropic struck with OnePWR for 1.7 GW of generators — sent a clear signal to the sector.
The implication is not that demand for data center power has declined. On the contrary: BloombergNEF projected, in a report published the same day as the news, that global electricity demand from data centers should almost quadruple, reaching 2,354 terawatt-hours by 2035. The firm recorded 181 GW of natural gas generation announced as a supplement to data centers, more than 90% in the United States.
What SoftBank's loss reveals is that the path chosen to meet this demand — on-site generation with gas turbines — may not be the most efficient or cost-effective route in the long run. Brent Nelson, senior managing director of markets and strategy at Ascend Analytics, noted in an interview with TechCrunch's Equity podcast that gas turbines, despite being familiar, carry specific risks: supply can be curtailed during winter, when residential and commercial demand for natural gas peaks. Data centers cannot simply switch their power load on and off like a commercial building.
Furthermore, the gas turbine market remains extremely tight. GE Vernova registered 36.7 GW of new orders in the third quarter of 2026, compared to 19.6 GW in the same period the year before. The delivery queue, combined with the shortage of transformers and other critical components, means that the very solution data centers are seeking to bypass grid delays faces parallel bottlenecks.
Another factor that SoftBank likely didn't anticipate when making its bet was the impact of the 2026 US midterm elections. Several regions where data center projects were underway — including parts of Texas and Ohio — approved anti-data center referendums and new regulatory restrictions. While CFO Yoshimitsu Goto declined to comment on electoral impact during the earnings call, the political context may have directly influenced the decision to cancel part of the orders.
What follows from here? The most likely response, according to analysts, is that SoftBank will not be the last company to reassess its own equipment orders. The strategy of locking in long-lead equipment years in advance has become common among data center developers — and it is almost certain that other companies are going through the same review process.
The crucial question that remains is not whether AI will continue to require more power — the answer is already written in BloombergNEF's numbers — but which energy infrastructure will be able to meet this demand reliably, sustainably, and economically. SoftBank bet on the wrong answer, at least for now. The energy market of the AI era is still being built, and every US$3.7 billion error serves as a costly reminder that the physical fundamentals have yet to be solved.
Sources: TechCrunch, CNBC, Bisnow, BloombergNEF
✓ Independent sources cross-checked and verified before publishing
