Goldman Sachs warns that AI capex narrative is under pressure; momentum trading bottom could be weeks away as fundamentals falter.

New-generation artificial intelligence models are achieving cutting-edge performance with significantly less computing power than the market anticipated, challenging the prevailing narrative that "only continuously expanding capital expenditure can win the AI race." According to Rich Privorotsky, head of Goldman Sachs' One-Delta trading desk, the current momentum trade drawdown remains weeks away from a true bottom. While the preceding rally's steeper-than-average slope suggests the correction's magnitude could exceed historical norms, U.S. equities overall exhibit no systemic risk. Market structure remains resilient, cross-sector correlations stay low, and capital is rotating between sectors rather than triggering a broad-based sell-off.

Privorotsky notes that momentum indicators still show elevated relative volatility and no signal sufficient to sound the all-clear. However, even as implied volatility ticked up recently, market structure has not fundamentally changed. Clear divergence has emerged: some AI hardware-related stocks have entered oversold territory, while certain previously lagging sectors are rebounding without obvious fundamental improvements—classic sector rotation characteristics.

After testing the Kimi K3 model, Privorotsky was impressed by its engineering. The model features 2.8 trillion parameters and still requires enterprise-grade GPU clusters for self-hosting. Critically, however, its training efficiency deserves greater attention. New-generation models increasingly rely on algorithmic optimization, model architecture innovation, and efficient Mixture of Experts (MoE) routing to improve training efficiency. Kimi K3 exemplifies this: it contains 896 expert modules but activates only 16 per inference, dramatically reducing computational resource consumption. This has prompted market reconsideration: if frontier models can substantially improve training efficiency through algorithmic innovation, does the industry still need continuously expanding, capital-intensive data centers and training clusters?

Privorotsky contends this primarily challenges the investment logic for training infrastructure, while inference-side computing demand remains relatively strong. Long-term AI infrastructure demand has not fundamentally reversed. However, as model training efficiency improves, the narrative of "continuously investing massive capital in larger computing clusters" faces mounting skepticism—even as AI capital expenditure remains the most important pricing anchor for global technology stocks.

With the Federal Reserve in its blackout period ahead of its policy meeting, near-term market focus will shift to the European Central Bank policy decision, UK Consumer Price Index data, and preliminary Purchasing Managers' Index readings from major economies. Yet Privorotsky believes the upcoming earnings season will truly determine market direction. Beyond Alphabet, earnings from Tesla, Texas Instruments, Intel, and AMD's "Advancing AI" event will serve as critical windows for the market to evaluate whether the logic behind trillions of dollars in AI capital expenditure can maintain market acceptance.

At the index level, U.S. equities continue demonstrating strong resilience. Despite dual pressure from an ongoing momentum trade correction and a wavering AI narrative, capital has rotated toward sectors with reasonable valuations or previously lagging performance rather than fleeing in panic. This suggests the market is pricing the AI investment theme with greater granularity rather than rejecting it outright. Privorotsky's momentum model indicates neither the duration nor magnitude of this correction has met the threshold for historical warning signals, making upcoming tech earnings and AI-related events critical for determining whether the AI capital expenditure cycle is approaching a turning point.