Test report DSG-6461 · Rev D · tested November 17, 2026

AI Datacenter InfrastructureDevice under test

SoftBank’s $7B Loss Exposes AI Data Centre Power Risk

SoftBank’s ¥1.12 trillion quarterly loss links AI data-centre equipment changes to higher financing costs, showing the cash and contract risks in the power build-out.

Read
5 min
Words
944
Node
65nm
Operator
Elena Vasquez

Spec summary

  1. SoftBank lost ¥1.12 trillion, about US$7 billion, in July-September 2026.
  2. The same quarter in 2025 produced a ¥2.5 trillion net profit.
  3. Quarterly financing costs reached ¥462 billion, about US$2.9 billion, more than doubling year over year.
  4. The six-month loss through September 2026 reached ¥597.2 billion, about US$3.7 billion.
  5. Future-purchase rebates run up to US$2.1 billion; their yen value remains unconfirmed.
SoftBank Books $7B Loss on US Data Centres [2026]
Fig. ASoftBank Books $7B Loss on US Data Centres [2026] — evgenia_axsenova / pixabay

By Elena Vasquez

SoftBank Group posted a ¥1.12 trillion net loss, about US$7 billion, for July-September 2026, reversing a ¥2.5 trillion profit in the same quarter of 2025.

By our arithmetic, the year-on-year swing was roughly ¥3.62 trillion. SoftBank also recorded a ¥597.2 billion loss, about US$3.7 billion, for the six months through September. The reporting does not reconcile that total with the quarterly figure. By subtraction, the first fiscal quarter appears stronger, but that is an arithmetic inference, not a SoftBank statement.

The result landed as investors worried about the cash burn required by the AI build-out.

What changed in SoftBank’s quarter?

SoftBank revised procurement plans for power-generation equipment serving US artificial-intelligence data centres. Costs connected with advance payments for equipment and manufacturing capacity formed part of the loss.

The company had paid suppliers early to reserve turbines, generators and factory slots. When procurement changed, some prepayments no longer matched planned orders and could require accounting charges. Reporting does not show how much came from written-off deposits or other adjustments.

Quarterly financing costs rose to ¥462 billion, about US$2.9 billion, more than double the year-earlier level. SoftBank did not disclose whether higher borrowing needs, larger balances or pricier terms drove the increase.

Why did power equipment become the constraint?

AI data centres need more electricity than local grids can often supply on developers’ timelines, so operators plan on-site generation. Turbines, generators and switchgear come from few manufacturers with long production queues, forcing buyers to commit years before servers arrive.

Reserving manufacturing capacity resembles a take-or-pay arrangement: funding secures priority but exposes the buyer if a project slips or shrinks. Pike County, Ohio, was among the projects covered by the revision; its size, schedule and status were not disclosed.

Each project also depends on land, transmission lines and local approvals. Related reporting put used turbines at $1.3 million per megawatt. Early commitments therefore carry a high price, while late orders leave campuses waiting for power.

What is the rebate worth?

SoftBank said it is entitled to rebates of up to US$2.1 billion on future purchases. The yen value, timing and conditions remain unconfirmed, and the ceiling means SoftBank may receive less if it reduces later orders.

The rebate equals a little more than half the US$3.7 billion six-month loss, but it cannot simply cancel that loss. It applies to orders not yet placed, rather than cash returned today. One plausible structure is a supplier credit against part of an earlier prepayment, but no named executive or supplier confirmed that explanation.

What does the financing bill mean?

Financing expenses equal roughly two-fifths of the quarterly loss’s dollar value. The figure does not show whether SoftBank borrowed more, carried larger balances or paid higher rates; its filings must resolve that gap.

Investors and lenders will watch interest coverage, further procurement revisions, expected rebates and the pace of new commitments. They will also test whether suppliers can replace SoftBank as a customer for released production slots; the reviewed reports did not establish whether those slots filled.

This resembles capital-heavy technology cycles. The late-1990s telecom build-out placed capacity ahead of demand, although AI services already generate revenue. SoftBank’s concentrated investments make reported profit volatile; infrastructure spending and interest add cash costs that valuation marks can reverse.

How does SoftBank compare with rival builders?

Recent plans show a market split between large power purchases and project pullbacks:

  • Anthropic reportedly agreed to pay $3.4 billion for 1.7 GW of gas power.
  • SpaceX plans a $24 billion, 748 MW gas plant at Starbase for compute.
  • Google took over Crusoe’s Wyoming site under a $17 billion plan.
  • NextDC paused uncontracted projects and cut A$140 million.
  • Norway halted data-centre grid deals covering 4,200 MW.

These figures describe reported plans or claims, not completed outcomes. A separate Harrisburg proposal covers $10.5 billion and five buildings. The mix shows no uniform appetite for new capacity.

Power procurement now sets the schedule: a campus without equipment slots cannot energize servers, regardless of GPU orders. A US Senate data-centre permitting bill carries an estimated 31% to 90% cost-cut range, while reported local opposition reached 76% and 41 measures have put the AI boom on notice.

What remains unconfirmed?

No named SoftBank executive or supplier provided an attributable quote in the reviewed coverage. The rebate’s yen value, reconciliation between quarterly and six-month losses, Pike County’s status and the financing-cost breakdown also remain unknown. Share-price reaction was not covered.

The available evidence does not show falling AI demand. The reported drivers concern equipment commitments, prepayment accounting and financing, not customer orders, and nothing in the reports points to service disruption.

What may happen over the next two quarters?

These are editorial forecasts, not SoftBank or analyst guidance:

  • Financing costs remain well above the year-earlier level as long-lived projects continue drawing capital.
  • More power-equipment adjustments emerge across the industry.
  • Disclosures clarify the rebate’s yen value, timing and conditions.
  • Developers secure generation or grid commitments before finalizing server orders.
  • Local opposition continues to shape site selection alongside power and tax terms.

What should cloud buyers do?

If infrastructure backers become more cautious, expansion could slow in some regions and GPU-heavy pricing could stay firm, although the available reports quantify neither outcome. Teams dependent on one provider or region should map a fallback. They should also review capacity-reservation contracts for refund terms, credit rights and protections if expansion slips.

via shattered.io (Source)

Filed under

  • softbank
  • ai-data-centres
  • power-generation-equipment
  • data-centre-financing
  • ai-infrastructure
Share this article:

More from Elena Vasquez

Elena Vasquez

Show full bio

Senior reporter covering people, launches and day-to-day practice in hardware layer of artificial intelligence at Die Signal.

675 articles

Same lot · LOT-362D

« Previous articleNext article »