‘FOMO has proven a stronger incentive than poor stock performance’: Goldman Sachs just issued a brutal verdict on the AI boom

Goldman Sachs projects about $7.6 trillion in cumulative AI capex from 2026 to 2031, but its analysts argue the spend is producing weak returns, with 95% of organizations seeing zero return on AI pilots and hyperscalers funding build-outs with more debt. The report flags rising data center costs, $182 billion of data center debt issuance in 2025, and only modest labor-market benefits, while Nvidia captures most of the economics. The piece is a bearish read-through for AI infrastructure and semis, though potentially supportive for hyperscalers if capex slows.

Analysis

The key second-order signal is not that AI spend is large; it is that the capex loop is becoming self-reinforcing while the monetization loop is weakening. That is the classic late-cycle setup where suppliers keep compounding because customers are locked in strategically, but the marginal ROI at the buyer level keeps deteriorating. The implication is that forward estimates for the ecosystem are likely still too high, because they assume efficient allocation of every incremental dollar, while the reality is more like duplicated capacity, stranded build-outs, and a rising cost of capital embedded in debt-funded expansion.

This argues for a near-term divergence between cash-rich platform companies and the chip stack. The hyperscalers can absorb a period of impaired returns because AI is strategically existential, but that same strategic urgency makes them willing to overbuild, which eventually becomes a margin headwind and balance-sheet story rather than a growth story. By contrast, NVDA’s economics are still excellent, but the stock is increasingly exposed to a regime shift where unit demand stays strong while the growth rate of aggregate capex slows; that is a dangerous setup for a premium multiple because the market is paying for acceleration, not just dominance.

The most interesting catalyst is not an AI adoption miss; it is a financing or governance inflection. If debt markets demand wider spreads for data-center and infrastructure issuance, or if one large hyperscaler signals a pause in expansion, the narrative can flip quickly from “arms race” to “capital discipline.” The timeline is months, not days: the market can ignore poor enterprise ROI for a while, but it will not ignore a visible slowdown in capex growth or a reset in free-cash-flow guidance. The contrarian view is that consensus may be underestimating how long strategic overspend can persist; that makes outright shorting hyperscalers early risky, but it also means the eventual downcycle could be abrupt rather than gradual.

Key Decisions for Investors

• Initiate a relative-value pair: long GOOGL / short NVDA over 3-6 months. Rationale: GOOGL has more optionality if AI monetization improves, while NVDA is priced for capex permanence; use the pair to isolate capex-slowdown risk with lower index beta.
• Sell 6-12 month out-of-the-money NVDA calls against a core long if already owned. Risk/reward favors harvesting elevated implied volatility while capping upside from a continuation of the current spending binge.
• Underweight IT versus a basket of large-cap platforms for 1-2 quarters. If enterprise AI ROI remains weak, infrastructure/software spend tends to be delayed first, creating a cleaner earnings downgrade path for services and implementation names.
• Go long MSFT or META on a 6-9 month horizon only on post-earnings drawdowns, not breakouts. These names are the best place to own the optionality of AI while limiting exposure to pure infrastructure multiple compression.
• Watch for a credit-led catalyst: if data-center and AI infrastructure issuance widens materially, add to short NVDA / long cash or short high-multiple semi exposure. That is the highest-conviction trigger for a regime change in the trade.