SoftBank’s massive bet on artificial intelligence infrastructure in the US is proving expensive. The Japanese investment giant reported billions of dollars in losses after revising its plans for equipment used to power AI data centres. It has also recorded losses linked to its agreement with OpenAI.
SoftBank had spent heavily to secure equipment needed to build power generation facilities for its planned AI data centres. Anticipating shortages, the company paid suppliers in advance and reserved manufacturing capacity to ensure it could get the equipment on time.
Between 2025 and the second quarter of 2026, these advance payments reached $9.5 billion.
However, as its data centre plans evolved, SoftBank changed its equipment requirements. Some of the equipment it had already paid for was no longer needed. The company has now written off $3.7 billion in losses related to unused equipment and changes to purchase agreements.
The investment firm clarified that the losses do not affect its PORTS-Pike power generation project, which is part of its broader US data centre expansion.
Salso reported a separate loss of ¥203.8 billion (around $1.3 billion) linked to warrants granted to OpenAI through its data centre business, Energy Global.
The agreement gave OpenAI the right to buy shares in Energy Global. As the value of the data centre business increased, the potential value of those shares also rose, increasing SoftBank’s cost.
Energy Global’s valuation has increased mainly because its data centre business has expanded. The company also reported ¥136 billion (around $900 million) in additional employee and director compensation expenses linked to this increase in value.
The developments highlight the enormous costs involved in building AI infrastructure. Beyond investing in data centres and computing equipment, companies must also secure reliable power supplies to keep these facilities running.
For SoftBank, the challenge now is to ensure its investments in power generation and data centre infrastructure translate into business returns.
