The dollar index, which measures the U.S. dollar against six major currencies, fell 0.49% to 102.055 at the New York close. EURUSD rose to 1.1212 from 1.1153; GBPUSD to 1.3272 from 1.3181. USDJPY was 158.14 (prior 158.20); USDCHF 0.8271 (prior 0.8345); USDCAD 1.4190 (prior 1.4247); USDSEK 9.9597 (prior 10.0321).

Jensen Huang told a Stockholm event that data-center operator CoreWeave’s financial condition is improving and Nvidia, as its compute supplier, does not need to provide a financial guarantee. He said CoreWeave’s strength is increasing daily and that he is optimistic, adding its cost of capital is falling. The remarks aim to restore market confidence after credit markets signaled rising default risk for CoreWeave. Market attention remains on CoreWeave as a test case for the sustainability of debt-fueled AI infrastructure models and on concerns that heavy borrowing supported by suppliers could create circular financing and artificial demand.

《紐約時報》10月15日發佈互動報道，以歷代GPT模型爲例，展示AI能力提升背後急劇擴張的計算需求。報道估算，2026年GPT-6 Astra的訓練計算量接近10²⁷次，達到初代GPT的5000多萬倍。文章通過可視化演示指出，大模型訓練依靠海量數據和計算，不斷調整數以萬億計的參數，以提高預測和生成能力。隨着模型規模擴大，算力和電力需求也持續增長。報道援引研究預測，到2030年，訓練最大規模AI模型所需的電力功率可能接近當前紐約市的用電負荷；同期，美國數據中心用電量可能佔全國總用電量的約12%。這一趨勢凸顯了AI技術進步與能源基礎設施擴張之間日益緊密的聯繫。