AI实验室如何解决电力危机：现场天然气发电深度解析

AI实验室如何解决电⼒危机：现场天然⽓发电深度解析 BringYourOwnGeneration,SayonaraElectricGrid,Turbinesvs.Recips.vs.FuelCells,WhyNotBuildMoreCCGTs?,OnsitePowerTCO ⾃带发电，告别电⽹，燃⽓轮机vs.往复式发动机vs.燃料电池，为什么不多建联合循环燃⽓轮机（CCGT）？，现场发电的总拥有成本（TCO） TheGridisOldandTired电⽹⽼旧⽽疲惫 Nearly two years ago, we were thefirst to predict a looming power crunch. In our reportAIDatacenter Energy Dilemma - Race for AI Datacenter Space, we forecasted AI PowerDemand in the US to grow from ~3GW in 2023 to over 28GW by 2026 – a pressure thatwould overwhelm America’s supply chains. Our prediction proved very accurate. 近两年前，我们率先预测了⼀场迫在眉睫的电⼒危机。在我们的报告《AI DatacenterEnergy Dilemma - Race for AI Datacenter Space》中，我们预测美国的 AI⽤电需求将从2023 年约 3GW 增长到 2026 年超过 28GW——这种压⼒将令美国的供应链不堪重负。事实证明，我们的预测⾮常准确。 Readfullstory阅读完整报道 The chart below tells the story: in Texas alone,tens of gigawatts of datacenter loadrequestspour ineach month. Yet in the past 12 months, barely more than a gigawatt has beenapproved. The grid is sold out. 下⽅的图表揭⽰了这⼀现状：仅在得克萨斯州，每个⽉就有数⼗吉⽡的数据中⼼负载接⼊申请涌⼊。然⽽在过去 12 个⽉⾥，获批的容量 barely 超过⼀吉⽡。电⽹已经售罄。 However, AI infrastructure cannot wait for the grid’s multiyear transmission upgrades. An AIcloud can generate revenue of $10-12 billion dollars per gigawatt,annually. Getting a 400MW datacenter online even six months earlier is worth billions. Economic need dwarfsproblems like an overloaded electric grid. The industry is already searching for new solutions. 然⽽，AI 基础设施⽆法等待电⽹多年才能完成的输电升级。⼀座 AI 云每年每吉⽡可产⽣100⾄120 亿美元的收⼊。⼀个 400 兆⽡的数据中⼼即便只提前六个⽉上线，其价值也⾼达数⼗亿美元。与这种经济需求相⽐，电⽹过载之类的问题显得微不⾜道。整个⾏业已经在寻找新的解决⽅案。 Eighteen months ago, Elon Musk shocked the datacenter industry by building a 100,000-GPUcluster in four months. Multiple innovations enabled this incredible achievement, but theenergy strategy was the most impressive. xAI entirely bypassed the grid and generated poweronsite, using truck-mounted gas turbines and engines. As shown below, xAI has alreadydeployed over 500MW of turbines near its datacenters. In a world whereAI Labs are racing tobefirst with a Gigawatt datacenter,speed is the moat. ⼗⼋个⽉前，埃隆·马斯克在四个⽉内建成了⼀个拥有 10 万块 GPU 的集群，震惊了数据中⼼⾏业。多项创新促成了这⼀惊⼈成就，但其中最令⼈印象深刻的是能源策略。xAI 完全绕过了电⽹，采⽤车载燃⽓涡轮和发动机在现场发电。如下⾯所⽰，xAI已在其数据中⼼附近部署了超过 500 兆⽡的涡轮机。在 AI 实验室竞相率先建成千兆⽡级数据中⼼的世界⾥，速度就是护城河。 One by one, hyperscalers and AI Labs are following suit and temporarily abandoning the gridto build their own onsite power plant. As we discussed months ago in theDatacenter Model,in October 2025, OpenAI and Oracle placed the largest order ever for onsite gas generation,with a 2.3GW plant in Texas. The market for onsite gas generation is entering an era oftriple-digit growth annual growth. ⼀家接⼀家，超⼤规模云⼚商和 AI 实验室正在效仿，暂时放弃接⼊电⽹，转⽽⾃建现场发电⼚。正如我们⼏个⽉前在《数据中⼼模型》中所讨论的那样，2025 年 10⽉，OpenAI 与甲⾻⽂下达了史上最⼤的⼀笔现场燃⽓发电订单——在德克萨斯州建设⼀座2.3GW 的电站。现场燃⽓发电市场正在进⼊⼀个年增长率达到三位数的新时代。 The beneficiaries extend far beyond the usual suspects. Yes, GE Vernova and Siemens Energyhave seen their stocks surge. But we’re witnessing an unprecedented wave of new entrants,such as: 受益者远不⽌传统的那些巨头。没错，GE Vernova 和西门⼦能源的股价已经⼤幅上涨。但我们正在见证⼀波前所未有的新进⼊者浪潮，例如： Doosan Enerbility, the Korean industrial giant, timing its H-class turbine launchperfectly. It alreadybooked a 1.9GW order to serve Elon’s xAI - as we exclusivelyunpacked to our Datacenter Industry Model subscribers several weeks ago. ⽃⼭能源（Doosan Enerbility），这家韩国⼯业巨头精准把握时机推出 H 级燃⽓轮机。正如我们⼏周前向《数据中⼼⾏业模型》的订阅者独家解析的那样，它已经拿下了⼀笔 1.9GW 的订单，⽤于服务埃隆的 xAI。 Wärtsilä, historically a ship engine manufacturer, realized the same engines that powercruise ships can power large AI clusters. It has already signed 800MW of US datacenter contracts. ⽡锡兰（Wärtsilä）历来是⼀家船⽤发动机制造商，但它意识到，为邮轮提供动⼒的同款发动机同样可以为⼤型 AI 集群供电。它已在美国签署了总计 800MW 的数据中⼼合同。 Boom Supersonic—yes, the supersonic jet company—announced a1.2 GW turbinecontract with Crusoe, treating the margin from datacenter power generation as anotherround of funding for their Mach 2 passenger jets. Boom Supersonic——没错，就是那家超⾳速喷⽓式飞机公司——宣布与 Crusoe 签署了⼀份 1.2 GW 的涡轮机合同，并将数据中⼼发电带来的利润空间视为其 Mach2 客运喷⽓式飞机的又⼀轮融资。 To understand growth and market share by supplier, we built a building-by-building trackerof sites deploying onsite gas in ourDatacenter Model. The results surprised us:12 differentsuppliers have now secured >400 MW of datacenter orders each in the US alone, for onsitegas generation. 为了了解各供应商的增长情况和市场份额，我们在《数据中⼼模型》中构建了⼀个按建筑逐⼀追踪部署现场燃⽓发电站点的系统。结果令我们惊讶：仅在美国，就已有 12家不同的供应商分别拿下了超过 400 MW 的数据中⼼现场燃⽓发电订单。 However, onsite power generation brings its own set of challenges. Power costs are often(much) more expensive than via the grid, as detailed below. Permitting can be a lengthy andcomplicated process. And it’s already causing some datacenter delays - most notably one ofthe Oracle/Stargate GW-scale facilities, which ourDatacenter Industry Modelpredicted three weeks prior to the Bloomberg headlines by analyzing the whole permitting process. 然⽽，现场发电也带来了其⾃⾝的⼀系列挑战。如下⾯所述，其电⼒成本往往（要）⽐通过电⽹获取的电⼒⾼得多。审批流程可能既漫长又复杂，⽽且这已经导致了⼀些数据中⼼项⽬的延误——其中最引⼈注⽬的是 Oracle/Stargate 的⼀座吉⽡级设施。我们通过分析整个审批流程的⽅式，在彭博社相关头条出现前三周，就已在我们的数据中⼼⾏业模型