Rules of relevance

Carlyle argues that the main market risk from the AI boom is not merely changing data, but a sudden change in the lens investors use to interpret it.

• Investors disagree over AI capex because engineers view the spending as a rational response to technical constraints, while statisticians see an unprecedented structural break with increasingly dispersed outcomes; credit spreads and required returns have not adjusted to the scale or duration of the buildout.
• Depreciating AI computing equipment over roughly six years means that each $100 of semiconductor revenue generates only a $17 depreciation charge in the year of sale, while about $150 billion of reported technology profits also reflects paper gains on stakes in frontier AI labs.
• More than 80% of anticipated AI capex through 2030 is expected to require external financing, but off-balance-sheet structures often protect lenders with residual-value guarantees and senior-tranche wrappers, leaving them exposed mainly to investment-grade counterparties rather than hardware economics.
• Top AI compute providers had disclosed more than $2.5 trillion of contracted backlog and remaining performance obligations by Q2 2026, but Carlyle estimates the net economic value after delivery costs at roughly $375 billion, or 15% of the nominal amount.
• AI borrowing is not the sole explanation for higher Treasury yields: US corporate net savings are down 86% from the comparable point in the previous cycle, while cumulative US borrowing from the rest of the world has reached 118% of GDP since 2000 and the foreign-demand subsidy is fading as globalization reverses.

Takeaway: Allocators should avoid assuming that direct AI exposure is scarce, favor credit risk over duration risk, expect fiscal strain to appear more through inflation than Treasury credit risk, and consider defense and energy as the clearest deglobalization beneficiaries; the conclusions remain exposed to uncertainty over how markets will reframe the same evidence.