Wall Street Journal, citing Rich Privorotsky, head of the Goldman Sachs One-Delta trading desk, said that recent global stock market pullbacks have revealed signs of fragility in the popular trades surrounding AI capital expenditures. The core assessment is that the market has nearly ignored all negative signals over the past several weeks, with capital increasingly concentrating in a narrow group of AI beneficiary stocks.
Trading is concentrated on a few AI assets.
Following a sharp decline in South Korea’s stock market on Tuesday, pressure quickly spread to global markets. The Nikkei Index fell nearly 3.5% that day, South Korea’s KOSPI Index dropped about 10%, and SK Hynix slid approximately 13% in a single day. According to Goldman Sachs, this is not merely a one-day fluctuation but reflects that trading related to memory chips has approached a structural bottleneck.
Privorotsky noted that the current market's pricing trajectory is highly concentrated. Returns from AI spending are being funneled into sectors such as semiconductors, memory, power, networking, and infrastructure, ultimately converging further into a small number of leading stocks. Goldman Sachs' institutional brokerage data also shows that global markets have evolved into a highly concentrated bet.
- Nikkei Index falls nearly 3.5%
- The South Korean KOSPI index fell by approximately 10%.
- SK Hynix fell approximately 13% in a single day.
Leveraged products amplify volatility risk.
The article notes that behind SK Hynix's sharp decline is the rapid accumulation of leveraged capital. According to the report, the assets under management of CSOP’s 2x daily leveraged product for SK Hynix, 7709.HK, have increased to $16.7 billion, placing it among the world’s largest single-stock leveraged ETFs.
In Goldman Sachs' view, the rapid expansion of such products is itself a signal of rising concentration risk. Should the underlying assets decline, leveraged capital could amplify price volatility and intensify market repricing of similar AI trading activities.
Cost reduction signals are not fully priced in
Privorotsky also noted that while the market is chasing the narrative around AI capital expenditures, it has overlooked how technological advancements are reducing model development costs. Iterations such as GLM-5.2, progress in small models, and the emergence of new architectures all point toward lower costs and higher efficiency.
He specifically noted that GLM-5.2 is reportedly trained on 100,000 Ascend 910B processors, without using NVIDIA chips. If cutting-edge AI capabilities can be achieved at a lower cost, the largest cloud service providers currently making the most aggressive capital investments may face questions about overinvestment.
The key variable is cloud service provider spending.
Goldman Sachs’ trading desk believes the key focus should not be on individual chip stocks, but rather on the stock performance of hyperscale cloud providers. These companies are the primary source of AI capital spending; if any one of them decides that spending less money better serves shareholder returns, the entire industry’s valuation foundation could be significantly impacted.
From a structural perspective, the Nasdaq Index has failed to form a decisive new high, and the support following options expiration has weakened. Combined with month-end and quarter-end rebalancing pressures, the market may face a shift in capital flows from selling stocks to buying bonds. Goldman Sachs believes that current AI trading is in an unstable equilibrium; if expectations for cloud service providers' spending loosen, the market could experience a more pronounced repricing.