Goldman Sachs: AI Deleveraging Enters Phase Two as $106 Billion Debt Pressure Emerges
nashnova research
Goldman Sachs warns that AI-linked deleveraging has escalated from directional selling to full balance-sheet compression, with gross leverage still at 310%; meanwhile, AI infrastructure debt issuance has topped $106 billion, and credit-market stress is bleeding into equities.
What does "Phase One to Phase Two" actually mean?
Phase One was directional selling — funds trimmed their long positions, but overall book size barely moved.
Phase Two is gross-exposure compression — funds are unwinding longs and shorts simultaneously, shrinking the entire book.
This means → funds are no longer just "turning bearish on AI"; they are actively shrinking their balance sheets — a step up in panic.
Goldman trading head Rich Privorotsky noted that global systematic long-short strategies fell 1.4% in a single day — a move approaching negative four standard deviations.
What do the leverage numbers really tell us?
Goldman Prime data show net leverage dropped 4.3 percentage points this month to 73.3%, sitting at the 6th percentile over the past year — directional bets have been cut hard.
But gross leverage remains at 310.3%, the 94th percentile over five years, and has actually risen 0.6 pp since October.
In plain terms = funds pulled out the "directional bet" money, but the balance sheet itself hasn't shrunk — Thursday's sell-off marks the start of compressing that 310%.
How far has the AI basket fallen — and is the bounce enough?
Goldman's broad AI basket (GSTMTAIP) dropped 20% over five trading days. Goldman's Asia desk called this "on the lower end" of recent deleveraging episodes.
Asian markets bounced overnight — Korea's KOSPI rose 2–3% — but had already fallen nearly 20% over the prior seven sessions.
Privorotsky described the bounce as "slightly underwhelming relative to the scale of the sell-off." This reflects the absence of a clear capitulation signal.
Where does the $106 billion debt bill come from?
Goldman credit strategist Amanda Lynam's team disclosed that year-to-date dollar investment-grade and high-yield project-finance issuance has exceeded $106 billion.
This means → that figure is roughly three times the $37 billion issued in all of 2025 — the pace has far outrun expectations.
Goldman says upside risk to its $300 billion global forecast for 2027 is "significantly elevated." Put simply = the borrowing is running ahead of the model.
Where are the cracks in credit markets?
High-yield data-center bond prices have fallen from above $101 in June to roughly $95.
CoreWeave unsecured bond spreads — a gauge of default risk — hit 1,000 basis points for the first time; Oracle unsecured debt is priced as deep junk; Meta's "Project Beignet" bond spreads are approaching all-time wides.
This signals that credit stress appeared before equity stress — Goldman labels emerging cloud providers and issuers exposed to frontier AI labs as "the epicenter of the funding-cost debate."
What is the ultimate variable to watch?
Goldman's macro team argues that when capital is rationed, AI is likely the first line item cut.
This means → the question is not just how far stocks fall, but whether this $106 billion in outstanding debt can roll over smoothly as funding costs keep rising.
In plain terms = the sustainability of AI infrastructure expansion ultimately hinges on the cost of refinancing — if spreads keep widening, the most leveraged link in the chain breaks first.
市场有风险，内容仅供研究参考，不构成投资建议。