中泰证券：高频电力数据证实AI算力需求仍在加速

Core Viewpoints - The demand for AI computing power is accelerating, as evidenced by high-frequency monitoring of the PJM grid covering key data center clusters in Virginia and Ohio, showing significant increases in load and electricity prices [2][3] Group 1: Electricity Demand and Pricing Trends - In the Virginia DOM area, the average monthly load increment for 2025 is approximately 3 GW (excluding base load), an increase of 0.98 GW compared to 2024, with significant year-on-year growth in load increments of 73%, 53.2%, and 56.4% for the months of September to November [2] - The electricity price difference between ARCOLA, BOYDTNDP, and SHILOHDP nodes has significantly increased, with ARCOLA node, primarily powered by Google, showing a price increase of 197% year-on-year to $7.94/MWh in October 2025, and a staggering 680% increase to $13.11/MWh in November 2025 [2][3] Group 2: Price Volatility and Congestion Fees - The standard deviation of electricity prices in the ARCOLA, BOYDTNDP, and SHILOHDP nodes has increased, indicating that the grid's available capacity is nearing its limits, leading to noticeable short-term price fluctuations [3] - Congestion fees have also risen significantly, with ARCOLA showing the highest price volatility and congestion fee differences, which increased by 223% and 890% year-on-year in October and November, respectively [3] Group 3: AI Application Barriers and Investment Opportunities - The report identifies four major barriers to AI applications: weak scale effects for single-user costs, subscription limitations for user expansion, higher ROI and added value requirements, and the need for a closed data loop [4] - Google is highlighted as a key player in the AI ecosystem, leveraging its hardware (TPU, Tensor G5), smart devices, cloud infrastructure, and software to integrate AI across its platforms, significantly reducing computing costs [4] Group 4: Investment Recommendations - The report suggests that selecting companies capable of overcoming the identified barriers will significantly increase investment success rates, particularly in vertical scenarios where unique data can provide substantial added value [5]