美国要让AI算力扩张成功，不只受软体与晶片能力所限，也受电力基础设施的时间窗牵制。Sightline Climate估计，美国2026年预计上线的资料中心总计可耗电多达12吉瓦，但目前仅约三分之一进入施工。变压器交货期从2020年前的24至30个月，已延长到长达5年；这种延迟推高压力，像Crusoe透过大量预购及回收停运电厂旧变压器来缓解。进口数据显示，2025年10月前美国从中国进口高功率变压器超过8,000台，而2022年全年不足1,500台。企业投资正在反应，如GE Vernova并购Prolec耗资53亿美元、Siemens Energy在未来两年再投资10亿美元，但供给缺口在短期仍难解。
这种供应链失衡使美国在AI竞赛中出现「要赢必须先有电力」的悖论：特朗普行政团队同时推动扩大电力配套行政程序并提出关税与贸易壁垒，但过度切断进口可能扩大供应停滞。部分关键产品仍高度受中国影响，据称特定型号变压器与开关设备进口来自中国的占比仍约30%，而锂离子电池更长期维持超过40%。Crusoe与Equinix等公司开始自建或提前采购，以期缩短供应时间10%至15%，因为在没有稳定电力设备的情况下，资料中心项目无法按时开工、开张。
America’s AI data-center build-out is constrained less by software and chips than by electrical infrastructure timing. In Abilene, Texas, an OpenAI-backed Stargate facility built by over 6,000 workers expects 1.2 gigawatts of load, enough for nearly 1 million households, yet industry plans are stalling despite over $650 billion in spending commitments from Alphabet, Amazon, Meta, and Microsoft. Nearly half of US-planned data centers this year are delayed or cancelled because transformers, switchgear, and lithium-ion batteries cannot be sourced fast enough domestically, pushing builders toward imports.
The bottleneck now sits on component supply chains. Sightline Climate projects US data centers to need as much as 12 gigawatts in 2026, but only about one-third are currently under construction. Lead times for high-power transformers, once 24–30 months, can now stretch to five years. Wood Mackenzie’s Benjamin Boucher ties this directly to import dependence: US utilities and developers increasingly buy from China. By October 2025, more than 8,000 high-power transformers were imported from China, versus under 1,500 in all of 2022. Corporate responses include GE Vernova’s $5.3 billion Prolec acquisition and Siemens Energy’s $1 billion U.S. expansion plan, while Equinix invests $350 million to increase domestic switchgear capacity and seeks 10% to 15% faster lead times.
Even where policy pushes against Chinese dependence, immediate decoupling is risky. The United States remains exposed, with Chinese shares for some transformer and switchgear categories still around 30% and for batteries still above 40%. Trump’s framework to accelerate power-plant permitting helps only part of the problem: without transformers, switchgear, and storage, AI centers cannot run safely, and expensive servers are exposed to grid instability. Crucial voices, including Crusoe’s Andrew Likens and Professor Joshua Busby, warn that electrical lead times can make or break billion-dollar AI infrastructure despite tariff politics.