As artificial-intelligence capital expenditures rise, so do the risks for AI stocks, Goldman Sachs tells investors

Goldman Sachs warns that the market’s 2027 AI capex consensus of $920 billion is likely too conservative, implying hyperscaler spending may stay elevated longer than expected. The note also flags that AI-stock valuations have returned to post-ChatGPT-launch highs, increasing the risk of volatility across the sector. The message is cautionary for AI-related equities, but it is commentary rather than a direct company-specific catalyst.

Analysis

The market is treating AI infrastructure spend as a linear growth story, but the more important setup is a classic capex reflexivity loop: when hyperscalers keep raising spend, earnings visibility for the picks-and-shovels names improves near term, while the system-level free-cash-flow burden eventually compresses equity multiples across the whole basket. That means the first derivatives still look bullish for semis, networking, power, and data-center equipment, but the second derivative is turning hostile for the highest-duration AI software and platform names that are priced off a perpetually falling cost of capital.

The bigger risk is not the absolute level of spend, but the market’s tolerance for funding it. As long as hyperscalers can self-fund via operating cash flow, investors will ignore the burn; once incremental capex starts crowding out buybacks or forcing balance-sheet expansion, the penalty will show up quickly in the group’s factor behavior. That transition tends to happen over months, not days, and it often begins with leadership loss in the most crowded mega-cap AI winners before broadening into a volatility spike across the whole complex.

Consensus appears to be underestimating how much of this capex becomes a demand-pull story for the supply chain rather than a pure margin story for the builders. Power generation, grid equipment, liquid cooling, and advanced packaging can stay tight for longer than the market expects because the bottleneck shifts away from chips toward energy and deployment infrastructure. The contrarian miss is that the “too conservative” capex estimate is not automatically bullish for AI equities overall; it can be bullish for infrastructure vendors while negative for return on capital across the hyperscalers themselves.

Catalyst-wise, watch any commentary on capex growth rates, depreciation assumptions, and buyback pacing in upcoming quarterly prints. If management teams start narrowing guidance or emphasizing phase timing rather than absolute spend, that is usually the first signal that the market has moved from rewarding scale to punishing overbuild. Volatility should be highest around earnings and guidance windows, with the downside skew strongest in names trading at the richest sales multiples and weakest cash-flow conversion.

Key Decisions for Investors

• Long SMH / short QQQ AI-elite basket for 1-3 months: express a view that infrastructure spend supports semis while multiple-sensitive mega-cap AI leaders de-rate if capex crowds out buybacks; target 8-12% relative outperformance, stop if hyperscaler commentary turns explicitly accelerationary.
• Buy call spreads in VRT or ETN out 3-6 months: these are cleaner beneficiaries of data-center power and thermal bottlenecks, where order backlogs can re-rate faster than the broader AI trade; prefer defined-risk structures because the market is already crowded into the theme.
• Short a high-multiple AI software basket over 2-4 months: use an ETF or paired short against an infrastructure winner, since the funding burden from capex is more likely to hit duration-sensitive software multiples first than the physical buildout names.
• Own downside volatility in a top hyperscaler via put spreads into earnings: the catalyst is not revenue miss but any hint that capex intensity is outpacing incremental cash flow or buyback capacity; risk/reward improves if implied vol stays below realized into the print.
• Pair long CEG/utility-grid exposure against short a crowded AI platform name for a 6-12 month thematic hedge: if AI buildout becomes an electricity and interconnectivity story, capital should migrate toward the enabling infrastructure rather than the most expensive end-user names.