Goldman Sachs Trading Desk: Momentum Trading May Take Weeks to Bottom Out, AI Capital Expenditure Narrative Is Shaking

AI models are achieving cutting-edge performance with computing power allocations far below market expectations, once again challenging the market narrative that “only continuous expansion of capital expenditures can win the AI race.” Goldman Sachs One-Delta trading desk believes that momentum strategy adjustments are not yet complete, but the U.S. stock market is not showing systemic risk, and the market structure remains resilient.

Rich Privorotsky, head of the Goldman Sachs One-Delta trading desk, said their momentum models indicate that the current correction in momentum trading still has several weeks before it truly bottoms out, and the final drop is likely to approach the historical median level. However, considering the previous upward slope was well above the historical average, there is also a possibility that this correction will exceed historical averages.

He also pointed out that the emergence of a new generation of highly efficient AI models is once again sparking market reassessment of the AI infrastructure investment logic. As model training efficiency continues to improve, the core narrative of "continually investing massive capital to build ever-larger compute clusters" is being increasingly questioned, while AI capital expenditure remains the most important pricing driver for global tech stocks.

Momentum Trading Not Yet Fully Unwound, Internal Market Rotation Persists

Privorotsky stated that his tracked momentum indicators show relative volatility remains high, and there are still no signals strong enough to lift the alarm.

However, obvious divergences have already emerged within the market. On one hand, some AI hardware-related stocks have entered oversold territory; on the other hand, certain sectors that previously underperformed have rebounded despite no clear fundamental improvement, with sector rotation being quite apparent.

At the index level, the U.S. stock market as a whole is still demonstrating notable resilience. Correlations between sectors remain at low levels, and capital is largely shifting between industries rather than triggering a broad-based sell-off. Even though implied volatility rose last Friday, this market structure has not fundamentally changed.

Higher AI Efficiency Again Challenges Capex Logic

Privorotsky said he was impressed by the engineering capabilities of the Kimi K3 model after actual testing. This model has 2.8 trillion parameters and still requires enterprise-level GPU clusters for self-hosting, meaning it cannot run on ordinary local devices.

He believes that the real focus should not be on the inference phase but on training efficiency.

Compared to simply relying on greater computing power, the new generation of models are improving training efficiency mainly through algorithm optimization, architectural innovation, and more efficient Mixture of Experts (MoE) routing mechanisms. For example, Kimi K3 has 896 expert modules, but only activates 16 per inference, drastically reducing computational resource consumption.

This is prompting markets to reconsider: if state-of-the-art models can significantly improve training efficiency through algorithmic innovation, does the AI industry still require continuous construction of ever-larger, capital-intensive data centers and training clusters?

However, Privorotsky believes this mainly challenges the investment logic on the training side, while demand for inference-side computing power remains strongly supported, and the long-term demand for AI infrastructure has not fundamentally reversed.

Earnings Season Will Determine if the AI Narrative Persists

As the Federal Reserve enters a blackout period ahead of its policy meeting, the market’s short-term focus will shift to macro events such as the European Central Bank meeting, UK CPI, and preliminary PMI data from major global economies.

Nonetheless, Privorotsky believes that the real market direction will be determined by the upcoming earnings season.

Beyond Alphabet, the performances of tech companies such as Tesla, Texas Instruments, and Intel, as well as AMD’s upcoming "Advancing AI" event, will all be crucial windows for market participants to observe the AI investment cycle, and will further test whether the multi-trillion-dollar AI capital expenditure thesis can continue to gain market recognition.