华尔街见闻·2025-09-19 11:51Core Viewpoint - The current AI infrastructure investment by major tech companies is significantly underestimated, with potential implications for future depreciation costs and a looming supply-demand imbalance that could lead to a price war by 2027 [1][3]. Group 1: Capital Expenditure Trends - Major players like Amazon, Google, Meta, Microsoft, and Oracle are projected to have capital expenditures as a percentage of sales reach 26% by 2027, nearing the peak of 32% seen during the internet bubble [2]. - The actual scale of investment is likely underestimated due to the increasing use of off-balance-sheet financing tools like leasing, which accelerates data center expansion without fully reflecting in traditional capital expenditure figures [2][5]. - Microsoft and Oracle are expected to see their capital expenditure to sales ratios rise significantly, with Microsoft projected to increase from 28% to 38% and Oracle from 41% to 58% by fiscal year 2026 [8]. Group 2: Depreciation Costs and Future Implications - Analysts at Bank of America highlight that the market is underestimating future depreciation expenses, with a projected shortfall of nearly $16.4 billion in depreciation costs for Google, Amazon, and Meta by 2027 [16][18]. - The trend of increasing capital expenditures will lead to accelerated depreciation and amortization (D&A) expenses starting in 2026, as these companies ramp up their investments [16][18]. - The lifespan of AI-related assets, such as GPUs, is shorter than traditional servers, with effective lifespans potentially only three to five years, which could further increase depreciation costs [20][21]. Group 3: Supply-Demand Dynamics and Pricing Strategies - There is a risk of overcapacity in the AI infrastructure market, with supply potentially exceeding demand by 2027, leading to aggressive pricing strategies among major tech firms to maintain utilization rates [25][30]. - The increasing similarity in performance among large language models may further commoditize infrastructure services, exacerbating pricing pressures [26]. - Major companies like Meta are investing heavily in new data centers, with significant projects expected to come online between 2026 and 2029, indicating a continued push for capacity expansion [28].