来源：环球网 报告分析指出，美国AI数据中心面临的并非系统性电力短缺，而是电网交付节奏与算力扩张速度的致 命错配。一个AI数据中心的建设周期可压缩至12-24个月，而电网扩容、输电建设及并网审批的典型周 期仍长达3-5年。以德州电力可靠性委员会（ERCOT）为例，2024至2025年间，数据中心提交的新增负 荷申请高达数十吉瓦，但同期成功接入的仅约1吉瓦。 "电网并非没有电，而是慢到无法匹配AI的节奏。"SemiAnalysis在报告中总结道。这种时间差，在AI算 力堪比印钞机的商业逻辑下，变得无法承受。 为何AI公司愿意承担更高成本的现场发电？核心在于算力的"时间价值"重塑了决策逻辑。报告测算，一 个1吉瓦规模的AI数据中心，年化潜在收入可达百亿美元量级。即便是中等规模集群，提前数月上线所 带来的商业价值，也足以覆盖自建电厂带来的更高电力成本。电力，由此变成了决定AI项目能否存在 的"准入门票"，而非后续的运营开销。 报告将这种模式称为"BYOG"（自带发电）。其目标并非永久脱离电网，而是"抢时间"：前期以离网方 式快速投产，后期再逐步接入电网，现场电厂转为备用。其中，xAI的实践被视为行业标杆。 【环球网 ...

Core Insights - The article discusses the urgent need for AI companies to bypass the aging public power grid by building their own gas power plants to meet the exponential demand for computing power, which has become a critical constraint for timely deployment [1][3][4]. Group 1: Power Crisis and AI Demand - The real bottleneck for AI data centers is not the lack of electricity but the slow delivery of power that cannot keep pace with the rapid expansion of computing needs [4][8]. - AI data centers are now being constructed in 12-24 months, while the typical cycle for power grid expansion and approval is still 3-5 years, making waiting for the grid a significant risk [5][6]. Group 2: Economic Implications of Power Supply - The time value of computing power is reshaping decision-making, with a 1GW AI data center potentially generating annual revenues of up to $10 billion, making it economically viable to incur higher electricity costs for faster deployment [9][10]. - Power is no longer just an operational cost but a prerequisite for the existence of AI projects, emphasizing the need for immediate power solutions [10][22]. Group 3: Onsite Power Generation Solutions - The BYOG (Bring Your Own Generation) model has emerged as a practical solution, allowing data centers to quickly start operations without waiting for grid connections [11][48]. - Major AI companies, including xAI, OpenAI, and Oracle, are leading the trend of onsite power generation, with significant projects underway, such as a 2.3GW gas power plant in Texas [16][29]. Group 4: Gas as the Preferred Energy Source - Natural gas has become the dominant choice for onsite power generation due to its scalability, stability, and rapid deployment capabilities, unlike nuclear or renewable sources [20][21]. - The competition in AI is increasingly defined by speed rather than cost, with companies prioritizing quick power access over traditional cost considerations [22]. Group 5: Market Dynamics and New Entrants - The onsite gas power generation market is experiencing unprecedented growth, with over a dozen suppliers securing contracts for AI data centers, indicating a shift in how power is viewed within AI infrastructure [17][30]. - New entrants, such as Doosan Energy and Wärtsilä, are capitalizing on this trend, with significant orders for gas turbines to support AI data centers [30][31]. Group 6: Challenges and Considerations - While onsite power generation offers speed, it also presents challenges, including higher long-term costs compared to grid power and complex permitting processes [34][36]. - The deployment of onsite power systems requires careful planning to ensure redundancy and reliability, as the complexity of managing power independently from the grid increases [94][100].

为了不被时间淘汰，越来越多美国AI数据中心正在做一件过去几乎不可想象的事：不等电网，直接在园区内自建电厂。燃气轮机、燃气发动机、 燃料电池被快速部署到数据中心旁边，只为一个目标——尽快把电接上，让算力跑起来。 AI的战争中，指数级增长的算力需求，正狠狠撞向美国老化且缓慢的公共电网。结论残酷而清晰——谁等电网，谁就出局。 2025年最后一天，知名半导体与算力研究机构SemiAnalysis发布了一份长达60多页的付费深度报告——《How AI Labs Are Solving the Power Crisis: The Onsite Gas Deep Dive》（AI实验室如何破解电力危机：现场燃气发电深度解析）。报告系统梳理了这一变化的底层逻辑：当AI进入超大规模 部署阶段，电力问题已经从"成本问题"，升级为决定算力能否按期上线的第一性约束。 电力危机的本质：不是不够，而是太慢 在传统认知中，美国并不存在系统性"缺电"。但SemiAnalysis指出，AI数据中心遭遇的真正瓶颈，并不在于电力资源是否存在，而在于电力交付节 奏与算力扩张速度的严重错配。 AI数据中心的建设周期，已被压缩至12—24个月；而电网扩 ...