Goldman Sachs warns AI capex bubble as 3 risks converge

Goldman Sachs warns the AI capital expenditure market has stretched like a rubber band, with the first major spending cut set to trigger a full-sector repricing.

Goldman Sachs warned the AI capital expenditure market has become dangerously overstretched, with the gap between spending commitments and investor returns reaching a critical point that could trigger a full-sector repricing. The bank's strategists said markets are extrapolating near-term trends too far into the future, making the sector vulnerable to any challenge to that optimism.

"The market is extrapolating near-term trends — including profits fueled by the investment boom itself — too far into the future," Rich Privorotsky, a strategist at Goldman Sachs, said in a note Tuesday. "The possibility of 'slightly less' spending has not been priced into anyone's expectations."

The warning landed as South Korea's Kospi index plunged 10% after hitting a record high, with Samsung Electronics and SK Hynix each falling more than 12%. Nasdaq futures dropped about 2.5%, while Micron Technology slid more than 7% in pre-market trading and Intel fell 6.5%. The VanEck Semiconductor ETF showed signs of what Stifel equity strategist Thomas Carroll called "mania," with price and volatility rising simultaneously.

The sell-off reflects a growing recognition that the $300 billion-plus annual AI capex cycle may be approaching an inflection point. Amazon, Alphabet and Meta have all increased spending commitments in recent quarters, yet their shares have underperformed the broader market — a divergence Privorotsky flagged as a pricing distortion signal. If any one of these hyperscalers cuts spending, the entire AI valuation chain from chipmakers to cloud platforms would face a comprehensive repricing.

The Asian Cost Disruption

A structural risk compounding the valuation concern comes from Asia, where Chinese and Japanese technology companies are developing AI models at a fraction of Western costs. China's GLM-5.2 large language model was trained entirely on 100,000 Huawei processors, using no Nvidia products whatsoever. The case directly challenges the assumption that Western hyperscalers must maintain ever-increasing GPU procurement from Nvidia and AMD.

"The biggest capital allocators are also the ones most exposed to overinvestment risk," Privorotsky said. If frontier intelligence can be developed in the East at a small fraction of Western costs, the current spending trajectory for companies like Microsoft, Amazon and Alphabet may represent significant overinvestment.

Multiple Pressures on Tech Valuations

JPMorgan's international market intelligence team, led by Federico Manicardi and Victoria Campos, identified additional headwinds: excessive market expectations, euphoric sentiment, free cash flow scarcity, chief technology officers pushing back against rising token costs, a shift toward lower-cost models, regulatory constraints from Washington, and increasing debt and equity supply.

The strategists also flagged advances in orchestration, model fusion and quantization as technologies that point to continued efficiency gains and potential pricing headwinds for incumbent AI providers.

The First Cutter

In Goldman's framework, the breaking point for the current AI capex cycle will come from a core spending company recognizing that returning capital to shareholders is a better allocation than continuing to pour money into AI infrastructure. The first company to pull back would become the trigger for a market-wide reassessment.

Goldman's Dominic Wilson and Vickie Chang, who previously compared the AI boom to the 1990s tech bubble, said the AI capex boom now matches the scale of that era even if it is less broad-based. They do not recommend exiting tech stocks entirely, but suggest using options strategies such as puts or call replacements to limit downside until the investment cycle peak draws closer.

Nvidia trades at about 35 times forward earnings, while AMD and Broadcom trade at elevated multiples that assume uninterrupted growth in AI-related revenue. A single spending cut from a major hyperscaler would test whether those valuations can hold.

This article is for informational purposes only and does not constitute investment advice.