ARKHOSReports

AI demand is not contract flexibility

What SoftBank’s procurement loss changes about the next board approval

ARKHOS | Investment intelligence | Analysis: 14 November 2026 | Reviewed and revised: 16 November 2026

The next AI investment paper should not begin and end with a demand forecast. It should show what the company will owe if that forecast arrives late, in a different location, or with different technical requirements.

This is not an argument against building. It is an argument for pricing the right to change the build.

The facts, without the wrong inference

SoftBank Group’s 10 November report records a $3.73 billion procurement-realignment loss for the six months ended 30 September 2026: $2.85 billion in non-refundable advances and $0.88 billion in contract changes and settlements. This is not a disclosed return on the entire AI portfolio. SoftBank separately states that the loss is unrelated to PORTS-Pike. The remaining-advance recovery discussion below is a different measure, not an explanation of a loss at that project. [1, 4]

A 12 November update adds essential balance. Of a $6.6 billion remaining advance-payment balance, management expects $5.6 billion to be recouped from the PORTS-Pike power-generation project and $1.0 billion over time. Within the $5.6 billion, $3.4 billion is expected near term and $2.2 billion by end-2027. These are recovery expectations, not cash already recovered. [2]

The loss, the remaining balance and the timing of recovery answer different questions. Treating them as one number would obscure the decision.

Exhibit 1. Two measures, two different questions

Independent composition panels. Each bar totals 100% of its own stated amount. Equal full widths do not mean equal dollar totals; no flow or causal link is implied.

A / Recognized procurement-realignment loss

$3.73bn · Six months ended 30 September 2026

76.4%
23.6%
Loss componentUSD bnShare
Non-refundable advances2.8576.4%
Contract changes and settlements0.8823.6%

Not portfolio ROI or current-period cash paid. SoftBank states that the loss is unrelated to PORTS-Pike. Sources [1] and [4].

B / Remaining advances and expected recovery

$6.6bn · Remaining balance at 30 September 2026

51.5%
33.3%
15.2%
Management’s expected recoveryUSD bnShare
Near term · PORTS-Pike3.451.5%
By end-2027 · PORTS-Pike2.233.3%
Over time · timing unspecified1.015.2%

Expected recovery, not cash already received. Widths encode amounts, not time intervals. The first two categories total $5.6bn expected from PORTS-Pike. Source [2].

Reported USD equivalents; shares calculated by ARKHOS from disclosed rounded amounts. [1] Financial Report, 10 Nov 2026, p.5; [2] Equipment update, 12 Nov 2026; [4] Investor Briefing v1.1, revised 11 Nov 2026, Accounting p.8. Full source links appear in “Sources and boundaries”. Data labels and tables provide the text equivalent.

The real exposure sits between four clocks

An AI infrastructure project runs on at least four clocks: the supplier’s manufacturing slot, the site and power-delivery schedule, the technology specification, and the customer’s actual workload ramp. A project can have attractive long-run demand and still experience a costly mismatch between these clocks.

The IEA’s October 2026 analysis describes constrained grids, lengthy connection queues and uncertainty about eventual connection demand. It also warns that connection pipelines can include speculative or duplicate projects. A queue is not contracted consumption; eventual demand is not a commissioning date. [3]

Our interpretation is that procurement should be evaluated as a sequence of commitments, not a single capital-expenditure total. The relevant question is not merely “How large is the AI opportunity?” It is “Which uncertainty are we paying to remove, and which uncertainty remains with us?”

Three choices, none automatically superior

Commit early. A binding reservation may protect a scarce delivery slot. It can be rational when credible customer demand and site readiness are sufficiently strong, and the cost of losing the slot exceeds the exposure being accepted. The board should see evidence of scarcity, not just a supplier deadline.

Buy flexibility. A smaller initial tranche, transferable allocation, specification-change window or deferred release may preserve choices. It has value only if the right is enforceable and usable when needed. A nominal transfer right without an eligible buyer is weak protection.

Wait. Deferral preserves cash but can sacrifice delivery priority, revenue or customer commitments. It is not free. Management should quantify the incremental economic cost of delay rather than assume that avoiding a deposit maximizes value.

The right comparison is between executable alternatives. An unavailable “perfectly flexible” contract is not a benchmark.

A one-page commitment test for the next investment committee

Before approving the next non-refundable payment, request five fields for each material contract:

1. Next irreversible date. When does the deposit become non-refundable, the specification freeze, or the cancellation charge increase? Name the owner who must act before that date.

2. Evidence due before that date. Identify the customer commitment, permitted site, deliverable power milestone and validated equipment configuration required for the next release. Separate signed evidence from management assumptions.

3. Cash still at risk. Show unpaid binding commitments, cancellation and redesign costs, carrying costs and timing of recoveries. Show sunk payments separately. Do not justify new spending merely to defend old spending.

4. Three synchronized cases. Test: delivery on plan; power or customer ramp delayed; and a material specification change. For each, show the cash timeline, the contractual remedy and who bears the mismatch. A project-level upside/downside percentage does not reveal this mechanism.

5. A decision that changes. Specify the condition for releasing the next tranche, renegotiating, transferring equipment, or stopping. If no plausible scenario or evidence threshold could change the decision, make that assumption explicit and test whether the cases are genuinely discriminating. All credible cases can support continuation; that conclusion still needs evidence.

This is an ARKHOS proposed decision tool, not a claim about SoftBank’s internal controls or an audited prediction of outcomes.

Two measures that expose a hidden trade-off

Alongside expected project value, report the unavoidable future cash commitment before the next evidence gate and the cost of changing course at that gate. Calculate both from actual contracts and finance-reviewed assumptions, with ranges where outcomes are contingent.

Then test a negotiable alternative: is the premium for a changeable reservation smaller than the scenario-weighted avoidable mismatch cost? Include the alternative’s probability of execution, delivery penalty and financing effects. Where probabilities cannot be defended, show the break-even probability instead of inventing precision.

This makes flexibility visible without pretending every delay can be forecast.

Why this matters in Saudi Arabia and the Gulf

For a regional sponsor, local power availability does not eliminate procurement, commissioning or customer-ramp risk. A practical investment paper should reconcile the equipment delivery schedule with the local connection agreement, sector-specific data requirements and the customer’s binding capacity commitment. Export-oriented capacity also needs a credible customer-acquisition case beyond domestic strategic ambition.

These are due-diligence questions, not claims that any particular regional project lacks these protections.

The implication for enterprise buyers is equally practical: a multi-year AI platform or minimum-consumption commitment can become difficult to reverse before adoption is proven. Separate a pilot’s technical success from the commercial evidence needed to authorize a larger locked-in commitment.

The board question

A strong AI thesis is necessary. It does not make every contract written around that thesis equally valuable.

Before your next commitment becomes irreversible, what evidence will improve, what choices will disappear, and what is that loss of choice worth?

Sources and boundaries

[1] SoftBank Group, Consolidated Financial Report, 10 November 2026, printed p.5 and Note 12. https://group.softbank/media/Project/sbg/sbg/pdf/ir/financials/financial_reports/financial-report_q2fy2026_01_en.pdf

[2] SoftBank Group, Equipment for U.S. AI Data Center (3), updated 12 November 2026. https://group.softbank/media/Project/sbg/sbg/pdf/ir/presentations/2026/us-ai-data-center_q2fy2026_01_en.pdf

[3] IEA, Powering Artificial Intelligence, executive summary, 14 October 2026. https://www.iea.org/reports/powering-artificial-intelligence/executive-summary

[4] SoftBank Group, Investor Briefing, 10 November 2026, Ver. 1.1 (revised 11 November), Accounting p.8 / PDF p.12. https://group.softbank/media/Project/sbg/sbg/pdf/ir/presentations/2026/investor-presentation_q2fy2026_01_en.pdf

The loss is not described here as a current-quarter cash payment or as proof of negligence, failed AI demand or realized project ROI. Conditional future-purchase rebates are not cash refunds. The financial report was not subject to interim auditor review. Management’s recovery expectations remain forward-looking. Figures are reported US-dollar equivalents, not independently reconstructed currency conversions. Analysis originated on 14 November 2026 and was reviewed and revised on 16 November 2026. Evidence cut-off: 16 November 2026. Recheck the cited disclosures and recovery expectations before publication or reuse after a material update.

© 2026 ARKHOS. All rights reserved. Third-party sources retain their respective rights.