大摩：市场低估了明年潜在的“AI重大利好”，但存在关键的不确定性

Core Viewpoint - A significant leap in AI capabilities driven by an exponential increase in computing power is anticipated by 2026, which the market may be underestimating [1][4]. Group 1: Computing Power Growth - Major developers of large language models (LLMs) plan to increase their training computing power by approximately 10 times by the end of 2025, which is expected to yield results in the first half of 2026 [1][4]. - A data center powered by Blackwell GPUs is projected to exceed 5000 exaFLOPs, significantly surpassing the computing power of the U.S. government's "Frontier" supercomputer, which is slightly above 1 exaFLOP [4]. Group 2: Scaling Wall Concerns - The report highlights a critical uncertainty regarding whether AI development will encounter a "Scaling Wall," where increased computing power leads to diminishing returns in model capabilities [2][6]. - Skeptics argue that simply increasing computing power may not sustain significant advancements in intelligence, creativity, and problem-solving abilities [6]. Group 3: Positive Signals and Risks - Recent research indicates that using synthetic data for large-scale training has not shown observable performance degradation, suggesting that there may still be room for continued improvement in model capabilities despite increased computing power [8]. - Other risks mentioned include challenges in financing AI infrastructure, regulatory pressures in regions like the EU, potential power bottlenecks for data centers, and the misuse or weaponization of LLMs [8]. Group 4: Asset Valuation Implications - If AI capabilities achieve a non-linear leap, investors should assess the multifaceted impacts on asset valuations, focusing on four core areas: 1. AI infrastructure stocks that can alleviate data center growth bottlenecks [10]. 2. The U.S.-China supply chain dynamics, which may accelerate decoupling in critical minerals [11]. 3. Stocks of AI adopters with pricing power, which could create an estimated $13 trillion to $16 trillion in market value for the S&P 500 [11]. 4. Long-term value appreciation for hard assets that cannot be easily replicated by AI, such as land, energy, and specific infrastructure [11].