SoftBank's $6.6 Billion Equipment Bet Has a $5.6 Billion Repayment Test
The procurement loss is booked. The next wager is whether project lenders reimburse the advances left behind.
By 30 September 2026, SoftBank's SB7 had paid $9.48 billion for power equipment, and SoftBank recognised a $3.73 billion loss after SB7 changed its procurement plans. Chief accounting officer Kazuko Kimiwada described gas turbines, transmission equipment, substations and manufacturing reservations in a 10 November 2026 Q&A. These were commitments made while planning American AI data centres. As project requirements evolved, the commitments became expensive to revise. No turbine had to fail for a supplier contract to hurt.
The loss was the cost of inflexible procurement in a scarce market. The live risk sits in SB7's remaining $6.6 billion of advances as of 30 September 2026, especially the $5.6 billion SoftBank expects the PORTS-Pike power project to repay. That expectation is a wager on project finance. I would value the prospective repayment below face until lenders' money arrives. When you own SoftBank, you own a financing exposure alongside equipment access.
The Bridge Trap is the gap between cash spent on scarce equipment and cash a project's financing returns to the spender. SoftBank calls its spending a bridge investment. Its 12 November 2026 slide schedules $3.4 billion for recovery from PORTS-Pike "in the near term". The slide also schedules $2.2 billion by the end of 2027. It expects another $1.0 billion back over time. Those amounts add up to the $6.6 billion held at 30 September 2026: $3.4 billion + $2.2 billion + $1.0 billion = $6.6 billion. The timetable comes from management. Its entries describe intended repayment rather than cash already returned.
Look at the cash, since the arithmetic refuses to be talked away. Through 30 September 2026, SB7 had paid $9.48 billion. For the six months ended then, $2.85 billion of non-refundable advances plus $0.88 billion of contract modifications and settlements equals the $3.73 billion loss. That $3.73 billion divided by $9.48 billion is 39.3%, though only $2.85 billion divided by $9.48 billion, or 30.1%, is the non-refundable advance component. SoftBank's 10 November 2026 presentation said cash settlement payments were expected from Q3 onward.
Now take the generous scenario for that $3.73 billion loss. SoftBank's accounts put conditional rebates at up to $2.1 billion over the next ten years as of 30 September 2026, with no rebate asset recognised. In the scenario where every dollar arrives, $3.73 billion of loss minus $2.1 billion of rebates leaves $1.63 billion of nominal cost. With zero rebates, the cost remains $3.73 billion. Neither scenario puts a cent of project-financing proceeds into SB7. A rebate on future purchases rewards the next purchase. It cannot turn a prior reservation into collateral.
Timing explains the attraction. SoftBank's November 2026 earnings presentation assigns $3.9 billion of payments to FY2025 and $5.6 billion to the first half of FY2026. The latter $5.6 billion divided by the slide's rounded $9.5 billion total is 58.9%, spent in one half-year. A quick purchase order secures manufacturing capacity while exposing the buyer to charges when project specifications change. You see why the first decision felt urgent. That urgency explains the mistake more persuasively than a story about collapsing electricity demand.
SoftBank has a sensible answer. In a 10 November 2026 Q&A, CFO Yoshimitsu Goto said "we determined that this was a risk we could absorb in the course of operating the business." He described timing and shortages, including cases where equipment is secured before due diligence is complete. At 30 September 2026, SoftBank listed ¥59.8 trillion in NAV and 15.9% loan-to-value. Its ¥5.0 trillion cash position also included investments and undrawn borrowing capacity. That context matters: ¥597.248 billion of procurement loss divided by ¥59.8 trillion of NAV is about 1.0%. A bruising purchase can coexist with a formidable balance sheet.
GE Vernova supplies the physical argument. In its 28 October 2026 filing, gas-equipment backlog plus reservations rose from 116 to 122 GW in the third quarter, while 13 GW of reservations became orders. CEO Scott Strazik expected "to exit 2026 with at least 130 GW of gas equipment under contract." NVIDIA's commentary for the quarter ended 26 July 2026 describes guarantees entered in August. They give credit support on the land, power and shell buildout for about 4.25 GW at SB Energy's PORTS-Pike campus, which will host NVIDIA infrastructure under 20-year leases to OpenAI. The guarantees are capped at $105 billion. They become effective in phases, with first service expected in fiscal 2029. With a 20-year OpenAI lease signed on 17 August 2026, you see why early access mattered.
The strongest counterexample to a stranded bridge is Cheniere's Sabine Pass financing. On 31 July 2012 it closed a $3.6 billion credit facility with eleven joint lead arrangers and ten other financial institutions, supported by completed commercial agreements for 7.7 mtpa and necessary authorisations. Cheniere Partners also reported receiving the remaining $333 million of funding from the purchase of Class B units by its parent. Infrastructure lenders do write large cheques when contracts and permits line up. Those banks financed construction, whereas SB7's exposure is an earlier equipment advance made by a separate entity. The analogy stops at the destination of the money.
The distinction is where the claim sits. SoftBank says the booked procurement loss is "completely unrelated to PORTS-Pike." SB7 made the advances. SB Energy's filing of 21 September 2026 says its planned gas-fired generation facilities are "not owned by the Company." PORTS-Pike was "Preparing for construction" in SoftBank's November 2026 materials. The 9.2 GW of utility service agreements were subject to final regulatory approval in the September filing. NVIDIA's capped guarantees back the PORTS-Pike buildout, rather than SB7's deposits. The disclosures identify neither a closed takeout of the $5.6 billion nor a lender commitment to reimburse old reservations at book value. A project can get financed while the reservation owner absorbs a haircut. That's the Bridge Trap.
There is a smaller historical version. Yingli Green Energy's 2017 filing records an RMB450.8 million provision in 2015 against supplier prepayments, followed by a full RMB225.4 million impairment of a supplier prepayment in 2017. Yingli incurred successive prepayment losses over several years. It was confronting polysilicon contracts after prices fell. SoftBank's disclosures have no comparable equipment price collapse; GE Vernova's reservation conversions point elsewhere. You can have a booming market for the supplier's product and an unfortunate contract for the buyer. Supplier orders are a poor guide to the amount SB7 will recover.
I expect some of the $5.6 billion earmarked for PORTS-Pike to remain unrecovered at the end of December 2027, even with genuine demand for power. I would discount that allocation against book value today and pay zero for conditional rebates. The supplier has its contract, the tenant has a lease and the shareholder owns the wait. My judgment would be wrong if a SoftBank financial filing showed that the full $5.6 billion of PORTS-Pike advances reached SB7 at carrying value from closed project financing by December 2027. Until then, par is a generous offer.