Net loss, Q2 FY2026
¥1.12 trillion
July–September 2026; first quarterly loss in almost two years
US power-equipment charge
¥597.2 billion
$3.7 billion booked as US AI data center procurement plans changed
Prepayments written off
$2.9 billion
Non-refundable advance payments already paid; $6.6 billion of $9.5 billion still carried
Five-year credit default swap
466 basis points
Nov 6, 2026; implies a 30.6% five-year default probability
Net asset value per share
¥10,499
As of Sept 30, 2026, against a Nov 10 share price near ¥3,879
Quarterly financing raised
$56.5 billion
Record quarter; the $40 billion OpenAI bridge was repaid ahead of schedule
The print
One Line Did Almost All of the Damage
SoftBank Group swung to a net loss of ¥1.12 trillion ($7 billion) for the July–September quarter from a ¥2.5 trillion profit a year earlier. Almost all of that swing came from a single decision: it changed the procurement plan for power-generation equipment serving US AI data centers and wrote off $2.9 billion of advance payments it had already made. Everything else in the quarter — a shrinking investment gain, the Intel mark, a doubling of financing costs — was a consequence of the same buildout being financed with borrowed money.
| Line | Q2 FY2026 | Comparison |
|---|---|---|
| Net income (loss) | ¥(1.12) trillion | ¥+2.5 trillion a year earlier |
| Six-month net income (loss) | ¥(0.8) trillion | Down ¥3.7 trillion year on year |
| Investment gain, six months | ¥1.7 trillion | Down ¥2.2 trillion year on year |
| US power-generation equipment | ¥(597.2) billion | Changed procurement plans, US AI data centers |
| Net asset value | ¥59.8 trillion | Up roughly ¥20 trillion from Mar 31, 2026 |
- The charge is mechanical: of $9.5 billion advanced for US power-generation assets, $2.9 billion already paid was expensed and $6.6 billion stayed on the balance sheet at Sept 30, 2026.
- SoftBank can recover part of it as rebates of up to $2.1 billion on future equipment purchases rather than as refunds on what was written off.
- The hedge that monetized Intel did the opposite of cushioning the quarter: $7.4 billion raised through prepaid forward contracts backed by Intel shares produced a ¥321.9 billion derivative loss, because the share price rose after the contracts were struck.
- The balance sheet took the strain where it usually does — in the loan book. Financing costs more than doubled to ¥462 billion ($2.9 billion) in the quarter, and the margin loan backed by Arm shares was upsized by $5 billion to roughly $25 billion.
Cause
A Procurement Decision, Not a Demand Verdict
SoftBank's US AI buildout was never a spot purchase. It pre-committed cash to scarce generation and interconnection slots, then expected to recoup that cash through project financing once the campuses were built and leased. When the political and permitting environment shifted — the Democratic midterm sweep sharpened local opposition to data centers, and the 10-gigawatt Ohio flagship went back for review — the procurement plan changed first and the accounting caught up. The group pre-paid for the power because the buildout itself was the point.
| Item | Disclosed amount |
|---|---|
| Advance payments authorized for US power generation | $9.5 billion |
| Expensed in Q2 FY2026 | $2.9 billion |
| Still carried at Sept 30, 2026 | $6.6 billion |
| Rebates available on future equipment purchases | Up to $2.1 billion |
| US campus plan under review | 10 gigawatts in Ohio, per the group's own materials |
| Demand check from the operating arm | Cloud & AI revenue up 37% year on year in H1 FY2026 |
The mark
Intel Shows What the Buildout Cost Before the Buildout Arrived
Intel cut both ways in the quarter, which is why the loss looked like a mark-down story when it was mostly a financing story. SoftBank had monetized its Intel stake at 3.7 times the $23 per share acquisition cost and raised $7.4 billion through prepaid forward contracts backed by Intel shares — then recorded a ¥321.9 billion derivative loss because the shares rose after the contracts were executed, alongside a ¥270.1 billion Intel valuation loss in the fiscal 2026 investment-gain table. The hard-currency verdict came from Intel's own accounts: a quarter that looked healthy inside a trailing year that still does not cover its own depreciation.
| Metric | Quarter ended Sept 26, 2026 | Trailing twelve months |
|---|---|---|
| Revenue | $16.51 billion | $59.89 billion |
| Operating income | $1.92 billion | $1.92 billion |
| Net income (loss) | $3.32 billion | $(12.03) billion |
| EBITDA | $7.54 billion | $3.36 billion |
| Depreciation and amortisation | $3.23 billion | $12.62 billion |
| Interest expense | $0.32 billion | $1.19 billion |
- The trailing numbers are the warning: Intel's twelve-month EBITDA of $3.36 billion covers barely a quarter of its $12.62 billion of depreciation, and it still paid $1.19 billion of interest.
- For SoftBank the arithmetic was harsher — the quarter's Intel derivative loss of ¥321.9 billion landed on top of a ¥270.1 billion valuation loss, so the hedge that raised cash also surrendered the upside it was meant to protect.
- The credit read-through is direct: the same chips, sold forward at a fixed price, are collateral for the money funding the campuses, so a flat or falling chip price quietly raises the cost of the AI buildout itself.
Transmission
Who Absorbs the Write-Down Now That the Equity Has a Creditor
AI buildout exposure that sits outside the buyer's income statement
Contracted and contingent exposure tied to US AI data centers, in $ billions
Unit: $ billion
Oracle additional lease commitments
Off balance sheet, 15-19 year terms, commencing Q1 FY2027
260
CoreWeave net debt
Neocloud funded on debt against contracted compute
45.2
SoftBank US prepayments still carried
Exposed to another procurement revision
6.6
Vertiv customer prepayments
Customers funding equipment before delivery
4.1
Oracle lessor borrowing guarantee
Guaranteed lessor debt, maturing Sept 2026
3.3
| Company | Role in the buildout | What the numbers say |
|---|---|---|
| SoftBank Group | Pre-pays for power equipment, owns the equity that funds AI tenants | ¥(597.2) billion equipment charge; Arm alone is ¥38.2 trillion of ¥59.8 trillion NAV; 466 basis point credit default swap |
| SoftBank Corp | 51% owner of the US neocloud SB Neo | H1 revenue ¥3,798.7 billion, up 12%; Cloud & AI up 37%; full-year guidance raised |
| Oracle | Lessee of large campuses and guarantor of lessor debt | $260 billion of additional lease commitments; $3.3 billion lessor borrowing guaranteed |
| CoreWeave | Neocloud landlord financed almost entirely on debt | $45.2 billion net debt; $1.87 billion trailing interest expense against minus $231 million trailing operating income |
| Vertiv | Sells power and cooling into the buildout | Q3 net sales up 61%; customers prepaid $4.1 billion before delivery |
| NVIDIA | Supplies chips and credit support for the Ohio campus | $1.5 billion invested in SB Energy; up to $105 billion of financing support |
The impairment does not vanish; it changes hands. Oracle has $260 billion of additional lease commitments, mostly data center, running 15 to 19 years and commencing from its fiscal 2027 — commitments that sit off its balance sheet but that its lessors and lenders have already funded. CoreWeave carries $45.2 billion of net debt against a $5.0 billion equity base, and its trailing interest bill of $1.87 billion already exceeds its revenue base of $7.59 billion. Meanwhile Barclays estimates $682 billion of US data center construction-debt funding needs for 2026 to 2030 and JLL projects $722 billion of North American debt financing for 2026 to 2028.
What makes this a credit story rather than an accounting story is the structure of the cushions. Most standalone data center construction projects carry a debt service reserve account sized to absorb setbacks lasting three to six months, and named structures reflect that: one2026 issuance requires an interest reserve equal to the greater of three months of Class A interest and fees or twelve months of priority expenses and maintenance capex, another holds a six-month senior interest reserve, and a third adds cash-retention thresholds with debt-service coverage triggers. A delay of a quarter does not trigger a default. A delay of two quarters does — and the Financial Times put the point plainly: a data centre with no power yet is just a construction loan accruing interest.
The signal
The Equity Shrugged While the Credit Market Did Not
The two halves of the same balance sheet produced opposite verdicts in the same week. SoftBank Group shares closed up 13.9% at roughly ¥3,879 on Nov 10, about 37% of the ¥10,499 net asset value per share the company itself publishes — the market is pricing the write-down as a timing event that a balance sheet with 15.9% loan-to-value can absorb. The credit market was not reassured: the five-year credit default swap spread hit as much as 466 basis points on Nov 6, the widest since October 2022 and equivalent to a 30.6% five-year default probability on Markit's data, and the shares fell 10.8% that day. The spread was already 432 basis points, the widest since November 2022, on Oct 20.
- The bridge is gone: SoftBank repaid the $40 billion facility raised to fund its OpenAI follow-on ahead of schedule in September 2026 and cancelled $10 billion of undrawn capacity, so the bond market is no longer bridging the group's own AI bet.
- What replaced it is not equity. Q2 financing of a record $56.5 billion included $21.9 billion of new borrowing, $18.0 billion of bond issuance, $16.6 billion of asset monetisation, $7.4 billion of prepaid forwards backed by Intel shares and $10.0 billion of loans using OpenAI shares.
- Leverage against the winners rose alongside: the margin loan secured on Arm shares was upsized by $5 billion to roughly $25 billion, which the group's own net asset value disclosure reconciles as a ¥3.9 trillion asset-backed finance adjustment against its Arm holding.
- The exit got further away. Reuters reported that OpenAI delayed its public offering to 2027, removing the liquidity event the whole financing chain had been underwritten around.
Horizons
What the Next Two Quarters Test, and What2030 Already Implies
| Window | Catalyst | What it tests |
|---|---|---|
| Days to quarters | SoftBank's next quarterly results, covering October to December 2026 | Whether the remaining $6.6 billion of US prepayments converts into operating assets or is written down again |
| Days to quarters | SB Energy's postponed US listing | Whether private capital will fund AI data centers at prices that make the parent's write-down look small |
| Days to quarters | Tencent replacing the Oracle capacity it contracted in September | Whether AI chip leases are durable assets or revocable options |
| Days to quarters | Oracle lease commitments beginning to commence in its fiscal 2027 | Whether $260 billion of off-balance-sheet commitments arrive before the demand they were built for |
| 1 to 3 years | Gartner's 2030 arithmetic: $8.5 trillion of revenue needed for a 12% return against $3.6 trillion of attributable AI revenue | Whether any of this capacity earns its cost of capital at all |
| 1 to 3 years | Expected consolidation, with more than 80% of new cloud service providers ceasing to operate independently by 2030 | Who inherits the assets when the operators fail |
Gartner's October 2026 analysis, reported alongside the results, put more than $3.5 trillion of AI infrastructure investment at risk of impairment by the end of 2030, and the arithmetic behind that number is the same arithmetic now visible in one quarter of SoftBank's accounts: data center owners would need $8.5 trillion of cumulative revenue to cover capital costs, financing expenses and a 12% return on invested capital, against enterprise and consumer AI revenue attributable to those data centers that tops out near $3.6 trillion. If most of the remaining $6.6 billion of prepayments and the next cohorts of campus capital are financed on the same terms, the impairment does not stay with today's buyer — it consolidates onto the largest lease commitments and the most leveraged lenders, whose reserve accounts only cover three to six months of delay.
Falsifiers
What Would Break This Reading
- If the remaining $6.6 billion of US prepayments converts into operating assets on schedule, the write-down was a timing event and the 466 basis point spread was pricing a procurement stumble as a credit event.
- If vendor financing keeps absorbing the strain — the $105 billion of support alongside the $1.5 billion NVIDIA put into SB Energy, and $638 billion of Oracle performance obligations — then the buildout can outrun the funding constraint and the credits are being paid for an asset that works.
- If Oracle's cancellation of the $7 billion Tencent lease proves idiosyncratic rather than systemic, lease commitments remain as durable as the assets behind them and the holder of the risk is still the equity.
Where the AI capex risk actually sits
- The ¥597.2 billion equipment charge equals about 1% of the ¥59.8 trillion net asset value, so the write-down alone does not impair the franchise.
- Loan-to-value of 15.9% and a ¥10,499 net asset value per share against a near ¥3,879 share price mean the equity is priced for a lot of disappointment already.
- The five-year credit default swap spread at 466 basis points implies a 30.6% default probability, which is where the market disagrees with the equity price.
- Over the next two quarters, the $6.6 billion of US prepayments still carried is the number that decides whether this was timing or impairment.
- Record first-half revenue of ¥3,798.7 billion, up 12%, with operating income of ¥642.1 billion and Cloud & AI revenue up 37% year on year.
- Full-year guidance was raised to ¥7,900 billion of revenue, ¥1,160 billion of operating income and ¥610 billion of net income, helped by a gain on selling preferred equity interests in SB Energy.
- Owning 51% of the US neocloud SB Neo gives it the operating upside from AI capacity without the parent's prepayments exposure.
- Over 1-3 years, the risk is that neocloud ramp costs land in the operating company while the revenue is still contracted years out.
- Trailing twelve-month EBITDA of $3.36 billion covers barely a quarter of $12.62 billion of depreciation and amortisation, so the fab plan depends on external funding.
- Trailing net loss of $12.03 billion and $1.19 billion of interest expense leave no cushion if AI demand plateaus.
- SoftBank's ¥321.9 billion derivative loss on prepaid forwards and ¥270.1 billion Intel valuation loss show chip prices now drive the financing cost of AI capacity itself.
- Over the next two quarters, watch trailing depreciation coverage rather than the quarterly profit, which was $3.32 billion on $16.51 billion of revenue.
- $1.5 billion invested in SB Energy plus credit support of up to $105 billion for the Ohio campus turns chip demand into secured, financed capacity.
- Trailing revenue of $302.97 billion and operating income of $197.58 billion make the vendor guarantee a rounding error next to the balance sheet behind it.
- Vendor credit is what keeps campuses funded after a buyer's equity takes a write-down, which protects the order book even as equity holders absorb losses.
- Over 1-3 years, the risk is that guarantee exposure grows faster than the chip revenue it secures.
- $260 billion of additional lease commitments, substantially all data center, run 15 to 19 years from a fiscal 2027 start and sit off balance sheet.
- Trailing capital expenditure runs at roughly 105% of revenue while net debt to EBITDA is 3.17 times, so the lease book is being built against a stretched balance sheet.
- Cancelling the $7 billion, five-year Tencent chip lease shows contracted demand can be withdrawn, which is the risk a 15-to-19-year commitment carries.
- Over the next two quarters, the first lease commencements and any further cancellations will show which side of that trade the market is on.
- $45.2 billion of net debt against $5.0 billion of equity leaves the balance sheet thin enough that a delayed ramp becomes a credit event.
- Trailing interest expense of $1.87 billion against minus $231 million of trailing operating income means debt service is paid before any return.
- Trailing capital expenditure of $20.57 billion was funded with $16.80 billion of net debt issuance and $5.84 billion of stock issuance, so growth is bought, not earned.
- Over 1-3 years, credit spreads on neocloud debt become the price of the AI buildout, not the equity's earnings.
- The 20-year, 890 megawatt nuclear agreement with Google, alongside a 20-year Amazon agreement, monetises the one input the buildout cannot substitute quickly.
- Constellation committed to invest more than $4.3 billion upgrading its fleet to deliver it, so the scarcity is being converted into contracted cash flow.
- Every campus that is financed but delayed still needs firm power, which keeps the option value with the generator rather than the landlord.
- Over 1-3 years, incremental nuclear and gas PPAs are the cleanest way to own AI capex without owning AI equity risk.
- A $7 billion, five-year lease for about 100,000 advanced chips signed with Oracle in late September was scrapped within weeks under US pressure.
- The episode shows AI compute leases from US suppliers are revocable political instruments, not long-dated contracted assets.
- Over the next two quarters, watch whether replacement capacity is secured and on what terms, and whether further US-linked leases are pulled.
- Over 1-3 years, the strategic cost is access to advanced non-domestic compute, which is the scarce input for its own model roadmap.
