Goldman’s top strategist just added hard numbers to his earnings-bubble warning

Goldman Sachs warns that AI-linked technology stocks may face an "earnings bubble" as AI infrastructure capex competes with government borrowing for capital, raising long-term funding costs. AA-rated technology issuer capex rose 65% year over year in Q2, AI-related borrowers represented 44% of $135B in U.S. convertible issuance, and Goldman raised its 2026 U.S. investment-grade issuance forecast by $200B to a record $2.3T. While tech balance sheets and AI-compute demand remain strong, Goldman says any profit-growth slowdown amid higher capital costs could pressure valuations; Nvidia fell more than 3% and the Philadelphia Semiconductor Index nearly 6% after calls to slow frontier-AI development.

Analysis

The key asymmetry is not AI demand but capex reversibility. MSFT, GOOG and META own distribution, proprietary data and end-user monetization channels; they can lengthen deployment schedules and preserve cash flow if incremental compute returns weaken. NVDA and the broader AI hardware chain instead face a sharper operating-leverage reset: a modest reduction in customer build plans can create inventory digestion, lower utilization and simultaneous multiple compression across semis.

The market should distinguish between funding stress and a credit event. Strong interest coverage makes a near-term balance-sheet break unlikely, but a higher long-end discount rate raises the hurdle rate for projects whose cash returns are several years out; this pressures terminal-value-heavy AI suppliers first, even if reported revenue remains strong over the next one to two quarters. The more important 6-18 month risk is that depreciation, power commitments and financing costs arrive before AI revenue is sufficiently visible in hyperscaler margins.

A contrarian point: the relative strength of hyperscalers is not unequivocally bullish. It may reflect investors rotating toward firms with the ability to cut orders, which is negative for the entire physical AI stack; however, hyperscalers are not fully insulated because contracted power, land and data-center commitments limit capex flexibility. The thesis is falsified if hyperscalers sustain elevated capex while simultaneously demonstrate accelerating AI-linked revenue or margin expansion, or if long-dated Treasury yields and hyperscaler credit spreads retrace despite heavy issuance.

Key Decisions for Investors

• Initiate a 1-3 month pair trade: long MSFT and GOOG equally / short NVDA, sized beta-neutral. The target is further relative outperformance if AI deployment timelines slow; exit if NVDA guides materially above expectations while MSFT or GOOG raise capex without corresponding monetization evidence.
• Avoid adding broad semiconductor beta through SOXX/SMH until the next hyperscaler capex disclosures clarify 2027 deployment plans. Use any sharp hardware-led rally as an opportunity to reduce exposure rather than chase, because order revisions would transmit to suppliers faster than to platform owners.
• For existing AI exposure, buy 3-6 month downside protection on SMH or NVDA rather than hedge with broad SPX puts. The relevant downside catalyst is a customer order or utilization reset, which is more concentrated in AI infrastructure than in the index.
• Maintain an alert on long-end rates and credit spreads: a persistent rise in the 10-30 year Treasury yield or widening in long-dated MSFT/GOOG/META bonds would validate a capital-cost shock and warrants increasing the hardware short. A material spread tightening would weaken the near-term funding-stress thesis.
• APO is a watch, not a directional recommendation: private-credit fundraising and financing demand can benefit from the capital shortage narrative, but underwriting losses rise if data-center assets are financed against aggressive utilization assumptions. Require disclosure of data-center lending exposure and asset-level covenants before adding.