Goldman Sachs flags $100B CTA selling risk as AI CapEx narrative shifts

Goldman Sachs warns the momentum trade correction may take weeks to bottom as efficient AI models challenge the trillion-dollar CapEx narrative that has driven the rally.

The emergence of efficient AI models like Moonshot's Kimi K3 — a 2.8 trillion-parameter system — is shaking the entrenched market narrative that winning artificial intelligence requires ever-expanding capital expenditure on data centers and training clusters.

"The momentum factor has not yet completed its adjustment process, and relative volatility remains elevated," said Rich Privorotsky, head of the One-Delta trading desk at Goldman Sachs. "But the more structural question is whether the market is correctly pricing the trajectory of AI infrastructure investment."

Kimi K3, developed by Chinese AI startup Moonshot, uses 896 expert modules in a Mixture-of-Experts architecture but activates only 16 per inference, dramatically reducing computational load. The open-weight model surpassed systems from OpenAI and Anthropic on several benchmarks and became the first Chinese model to top the Frontend Code Arena leaderboard — a live evaluation platform for AI-generated web applications. Its release comes roughly 18 months after DeepSeek first rattled U.S. AI leaders with a similarly efficient approach, suggesting the efficiency trend is accelerating rather than fading.

The challenge to the CapEx consensus arrives as Goldman Sachs' momentum model signals the current correction in momentum-driven stocks may take weeks to fully resolve. Bank of America estimates a further 3% decline in the S&P 500 could trigger as much as $100 billion in forced programmatic selling from Commodity Trading Advisors, whose short-term thresholds have already been breached in the Nasdaq. The upcoming earnings season for major tech firms — including Alphabet, Tesla, Texas Instruments, Intel and AMD — will serve as a critical test of whether the trillion-dollar AI CapEx thesis can withstand scrutiny from both investors and technological reality.

Training Efficiency vs. Inference Demand

Privorotsky said he was impressed by Kimi K3's engineering after testing the model, but cautioned that the efficiency challenge applies primarily to the training side of the AI investment equation. "Inference-side compute demand still has strong support," he said, suggesting that AI infrastructure demand has not fundamentally reversed even if training clusters face more scrutiny.

The distinction matters for investors. Nvidia, whose data center revenue reached $47.5 billion in its most recent fiscal year, benefits from both training and inference demand. A shift toward more efficient training architectures could compress the total addressable market for training-specific hardware, even as inference workloads continue to scale with user adoption. AMD, which holds its "Advancing AI" event later this month, and Intel, which reports earnings alongside Texas Instruments this quarter, are among the semiconductor names most exposed to the evolving CapEx narrative.

HSBC equity strategists independently warned that the momentum trade has become vulnerable to "a sharper and more sustained reversal," noting that the global long-versus-short momentum factor has already fallen 15%. The bank favors U.S. consumer discretionary stocks excluding Amazon and Tesla, U.S. banks, and European cyclicals including airlines, hotels and defense — a rotation away from the AI-concentrated momentum trade that has dominated performance.

Rotation, Not Systemic Risk

Despite the pressure on momentum names, Goldman's analysis shows the broader market structure remains resilient. Sector correlation is holding at low levels, indicating that money is rotating between industries rather than exiting equities entirely. Some AI hardware-related stocks have entered oversold territory, while previously lagging sectors have bounced without fundamental improvement — a pattern consistent with rotation rather than systemic liquidation.

The VIX has ticked higher but remains below levels typically associated with panic selling. Privorotsky said the firm does not see systemic risk in U.S. equities, though he cautioned that the momentum adjustment process could take weeks to fully play out. CTAs added roughly $23 billion in U.S. equities in a single week earlier this year and approximately $53 billion over a month-long window, according to Goldman's data — positioning that now faces potential unwinding as price thresholds are breached.

The Nasdaq's short-term CTA threshold of approximately 19,608 has already been broken for the first time since April, Goldman's data shows. The S&P 500's threshold sits at roughly 5,472, with the index hovering about 3% above a more consequential medium-term level. If that level breaks with conviction, the forced selling cascade could accelerate, testing the appetite of even conviction-driven buyers across equity markets.

For investors, the key question is whether the AI CapEx narrative — which has underpinned the valuation of semiconductor and data center stocks trading at elevated multiples — can withstand a period of both momentum-driven selling and fundamental re-evaluation. The next two weeks of earnings reports will provide the first major data point.

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