‘At some point you’ve got to make money’: Goldman’s top AI skeptic warns the clock is running out ahead of OpenAI and Anthropic IPOs

Goldman Sachs’ Jim Covello argues the AI investment cycle is still far from proving enterprise ROI, warning that companies are spending more to implement AI while returns remain elusive. He cites estimates of $7 trillion to $8 trillion in eventual AI infrastructure spend, with MIT research showing 95% of organizations reporting zero return on AI pilots and an EY survey finding 99% of companies reporting AI-related financial losses averaging $4.4 million each. The article highlights rising capex, FOMO-driven spending by hyperscalers, and growing political and operational risks, but it is primarily a thematic/analyst-driven warning rather than a direct market catalyst.

Analysis

The market is still treating AI as a growth perpetuity story, but the more important inflection is margin compression for the entire buyer stack. If enterprise ROI stays weak, the economic rent will continue migrating upstream to the infrastructure layer for longer than consensus expects, but that is also the setup for eventual capital discipline: hyperscalers are already carrying the capex burden, so any slowdown will show up first as lower growth and multiple compression in the builders, not in the chip vendor immediately. The key second-order effect is that the longer adoption lags, the more exposed the cloud platforms become to scrutiny over depreciating assets, power costs, and return on invested capital.

The underappreciated vulnerability is not model quality but implementation friction. Legacy-data plumbing, workflow redesign, and internal trust decay mean the near-term productivity benefit is likely to accrue to a smaller set of AI-native firms while incumbents keep spending to avoid competitive irrelevance. That creates a bifurcation: public megacaps can still win on relative scale, but many downstream enterprise software vendors and integrators may see a tougher sales cycle as customers demand proof, not pilots.

Politically, the thesis is moving from economic skepticism to cost-of-living backlash, which raises the odds of localized permitting, power-pricing, and regulatory friction over the next 6-18 months. That matters because data-center buildout is increasingly constrained by electricity and community pushback, so the bottleneck may shift from semiconductor supply to grid access and political license. The consensus is underpricing how quickly sentiment can flip from "strategic necessity" to "subsidized utility burden," especially if job displacement remains visible while wage gains do not.

The contrarian takeaway is that the current winner may be the least durable winner: semis are monetizing the spend now, but hyperscalers own the customer relationship and can eventually ration capex, renegotiate procurement, or internalize more compute through custom silicon. If enterprise payback remains elusive for another few quarters, the market may re-rate AI not as a secular compounding story but as a capex cycle with delayed payback and lower terminal margins than implied today.

Key Decisions for Investors

• Long AMZN/MSFT/GOOGL basket vs short NVDA on a 3-6 month horizon: thesis is that hyperscalers can absorb and eventually discipline spend, while semis face multiple risk if capex growth decelerates; target a 1.5-2.0x payoff if AI capex growth slows even modestly.
• Buy 6-12 month put spreads on NVDA into strength: use any post-earnings or AI-event rally to position for a valuation reset if investors start questioning incremental ROI; defined risk with favorable convexity if sentiment shifts.
• Long GS vs short AI-exposed software/IT services basket over 6 months: if enterprise AI ROI remains unproven, capital markets activity around monetization and restructuring should hold up better than implementation-heavy vendors.
• Reduce exposure to names levered to aggressive data-center buildouts and grid-spend assumptions; the asymmetry is worsening because political and utility constraints can bite faster than revenue monetization ramps.
• Watch for a relative-value long hyperscaler / short semiconductor pair entry after the next capex guide-up cycle: the trade improves if management teams keep raising spend while evidence of enterprise payback stays weak.