数据中心的平衡之道：驱动可持续的AI增长

The Data Center Balancing Act: Powering Sustainable AI Growth

生成式人工智能（GenAI）的崛起正在推动一场变革性的技术革命。数据中心价值链正处于现代史上**资本配置机遇之一的临界点上。为了满足计算力需求，到2030年，相关资本支出可能接近7万亿美元。

The rise of generative artificial intelligence (GenAI) is driving a transformative technological revolution. The data center value chain stands at the precipice of one of the largest capital allocation opportunities in modern history. This could amount to nearly $7 trillion in capital outlay by 2030 to meet compute power demand.

无论参考哪种方法来预测资本支出和计算力增长，这一大规模建设既带来了诱人的投资机会，也暴露出诸多瓶颈与约束。在这个时代，电力为王。随着生成式人工智能（GenAI）工作负载的增加，一些专家预测，到2030年，全球数据中心的电力需求可能翻倍至约945太瓦时（TWh），略高于日本当前的整体用电量，这对电网容量构成威胁。这也引发了投资者亟待思考的关键问题：超大规模数据中心将如何突破资源限制？哪些公司将从中受益，又有哪些可能面临颠覆？资本密集型投资将在多大程度上、何时才能获得回报？

Regardless of which methodology you reference to project capital expenditures and compute power growth, this massive buildout exposes both compelling investment opportunities and constraints. In this era, power is king. As GenAI workloads increase, some experts project that electricity demand from data centers worldwide could more than double by 2030 to around 945 terawatt-hours (TWh), slightly more than the entire electricity consumption of Japan today, threatening grid capacity. Critical questions arise for investors: How will hyperscale data centers overcome resource constraints? Which companies are poised to benefit and which will face disruption? To what extent, and when, will the capital-intensive investments pay off?

美国占全球数据中心能源消耗的45%。在经历了超过30年的电网用电需求下降或持平之后，电力公司预测，为了满足数据中心、电气化以及新兴产业的用电需求，到2028年，美国将需要新增约38吉瓦（GW）的发电能力，相当于约34座新的核电站。

The United States (U.S.) is home to 45% of global data center energy consumption. After more than 30 years of falling or flat demand for on-grid electricity, electric utilities forecast that the U.S. will need 38 gigawatts (GW) of new power production, equivalent to about 34 new nuclear plants, through 2028 to supply power for data centers, electrification, and new industry.5

历史上，尽管自2010年以来全球互联网流量增长了25倍，数据中心和数据传输网络的能耗增长却得到了一定的缓解，这主要得益于能源效率的快速提升以及离网电源供应（如家庭太阳能）的发展。具体来说，IT硬件和制冷系统的效率不断提升，同时，小型低效企业数据中心正逐渐被更高效的共置（Colocation）和超大规模（Hyperscale）数据中心所取代7。然而，面向未来的情景带来了新的、独特的挑战。专门用于人工智能的AI数据中心比传统数据中心需要更多电力，因为AI专用计算每机架的能耗远高于传统计算。

Historically, rapid improvements in energy efficiency and off-grid supply, such as domestic solar, have helped moderate growth in energy consumption from data centers and data transmission networks globally, despite global internet traffic expanding 25-fold since 2010. Specifically, there have been efficiency improvements in IT hardware and cooling, as well as a shift away from small, inefficient enterprise data centers towards more efficient colocation and hyperscale data centers.7 Yet, the forward-looking scenario presents new and unique challenges. AI-specialized data centers require more power than traditional data centers, as AI-specific compute consumes significantly more power per rack than traditional compute.

这引发了一系列环境影响，进而带来经济效应。除了全球对温室气体排放的关注之外，数据中心还在水资源保护、电子废弃物处理以及自然资源利用方面带来挑战，例如生产硬件组件所需的稀土开采。正如机构可持续投资负责人凯瑟琳·克罗尔（Katherine Kroll）在《能源扩张与独立》系列中所指出的：“发达经济体已切身体会到，能源基础设施的韧性与多样化对于国家安全和经济活力至关重要。”

This prompts a range of environmental implications, which in turn has economic effects. In addition to the global focus on greenhouse gas emissions, data centers present challenges for water conservation, e-waste, and concerns around natural resources, such as the mining of rare earth elements required to manufacture hardware components. As Katherine Kroll, Head of Institutional Sustainable Investing, depicts in the Energy Expansion and Independence series, “Advanced economies have learned firsthand that resilience and diversification in energy infrastructure are critical to national security and economic vitality”.

本文关注的核心是数据中心——这些建筑承载着支撑数字世界的计算、存储和网络设备。数据中心是实现人工智能潜力的关键，但它们的快速增长对现有电网造成压力，并推动关键资源接近极限，同时也影响到当地社区。如果这些关键要素没有得到充分考虑，我们可能面临临界点。我们将分析数据中心的市场动态，以及超大规模数据中心运营商（Hyperscalers）正在采取的措施，以降低这些资源限制阻碍生成式人工智能（GenAI）前沿机遇的风险。同时，我们将探讨数据中心供应链各环节的投资机会——从节能芯片和可再生能源，到先进制冷技术——这些创新可能彻底改变电力平衡的方式，甚至推动电网的可持续转型。

Our focus in this piece is on data centers, the buildings that house the compute, storage, and networking equipment which underpins our digital world. Data centers are key to realizing the promise of artificial intelligence, but their growth puts pressure on existing power grids and pushes against limits on key inputs, which also impacts local communities. If these essential ingredients are not adequately accounted for, we could face a tipping point. We will examine data center market dynamics and steps hyperscalers are taking to increase the likelihood that these limits do not gate the exciting opportunities at the forefront of GenAI. We will highlight investment implications across the data center supply chain, from energy-efficient chips and renewable energy, to advanced cooling technologies, that could revolutionize the power balancing act and perhaps catalyze a sustainable grid transformation.

布朗咨询（Brown Advisory）基础设施董事维多利亚·施洛特巴克（Victoria Schlotterback）、伊丽莎白·希斯（Elisabeth Hiss）和蒂芙尼·欧内斯特（Tiffany Ernest）访问了QTS（Quality Technology Services，QTS）——这是一家黑石集团（Blackstone）投资组合公司，在北美和欧洲提供安全的数据中心解决方案，总电力容量超过2吉瓦。QTS在弗吉尼亚州劳登县拥有四个数据中心，分别是Ashburn-Broderick、Ashburn-Lockridge、Ashburn-Moran 和 Ashburn-Shellhorn。

Victoria Schlotterback, Elisabeth Hiss, and Tiffany Ernest, Brown Advisory’s Director of Infrastructure, visiting Quality Technology Services (QTS), a Blackstone portfolio company that delivers secure data center solutions with over two gigawatts of power capacity in North America and Europe. QTS owns four data centers in Loudoun County, Virginia, including Ashburn-Broderick, Ashburn-Lockridge, Ashburn-Moran, and Ashburn-Shellhorn.

自2017年以来，北美数据中心行业的年复合增长率（CAGR）达到20%。在此期间，大多数市场规模翻了一番甚至三倍。

The data center sector in North America has been growing at a 20% CAGR since 2017. Over that time, most markets have doubled or tripled in size.

北弗吉尼亚——全球数据中心最集中地区——比大多数地区更早面临资源约束。为该地区供电的Dominion Energy在其2024年年度报告中指出，数据中心在截至2024年12月31日的年度内占弗吉尼亚电力公司销售电量的26%，已成为电力需求显著增长的主要来源，并预计在未来十年持续增长。Dominion Energy表示，其已收到的客户订单可能在2028年前将弗吉尼亚的数据中心容量翻倍，预计到2035年市场规模将达到10吉瓦（GW）。

Northern Virginia, home to the greatest concentration of data centers in the world, is experiencing constraints sooner than most. Dominion Energy, a company that supplies electricity in the region, stated in its 2024 Annual Report that data centers — representing 26% of Virginia Power’s electricity sales for the year ended December 31, 2024 — have been a source of significant increase in demand, which is expected to continue over the next decade. Dominion Energy says it has received customer orders that could double the amount of data center capacity in Virginia by 2028, with a projected market size of 10GW by 2035.

数据中心主要集中在弗吉尼亚州劳登县，这带来了复杂的挑战。当前的数据中心电力约束往往源于输电和配电的限制，而非发电能力不足。换句话说，该地区短期内面临的难题并不是电力不足，而是无法有效分配现有电力。这导致了对电力输配设施的大量投资，以满足不断增长的需求。

The concentration of data centers primarily in Loudon County, Virginia, represents a complex challenge. Current data center power constraints often arise due to transmission and distribution limitations rather than a lack of power generation capabilities. In other words, the near-term hurdle in the region isn’t insufficient power generation but rather the inability to distribute it effectively. This is leading to significant investments in electric transmission facilities to meet the growing demand.

行业专家指出，输电限制迫使一些数据中心等待长达七年甚至更久才能接入电网。尽管如此，弗吉尼亚的数据中心扩张仍在继续。例如，Alphabet计划在2026年前在弗吉尼亚投资额外90亿美元，用于云计算和人工智能基础设施建设。

We have heard from industry experts that electric transmission constraints are forcing some data centers to wait up to seven years or more to secure grid connections. Nonetheless, Virginia expansion continues. Alphabet, as one example, is investing an additional $9 billion in Virginia through 2026 in cloud and AI infrastructure.

尽管弗吉尼亚持续受到投资，但由于该地区的资源限制，数据中心开发正逐步向北弗吉尼亚以外的二级和三级市场扩展。高电力容量与电网可用性、用电成本、可靠性、网络连接性以及法规要求是选址的关键因素。例如，Meta正在路易斯安那州建设一座价值100亿美元、以人工智能为核心的数据中心——这是西半球规模**的项目，占地2,250英亩。该项目预计将重塑企业对电网可靠性的考量，通过向电网新增至少1,500兆瓦（MW）的可再生能源，并推进下一代核能技术的发展。

Despite the continued investment in Virginia, due to the region’s constraints, data center development is expanding beyond Northern Virginia into secondary and tertiary markets. High power and grid availability, electricity cost, reliability, connectivity, and regulations are key factors in site selection. Meta is developing a $10 billion AI-focused data center — the largest in the Western Hemisphere — on a 2,250-acre site in Louisiana, a project expected to reshape how companies approach grid reliability by adding at least 1,500 MW of new renewable energy to the grid and bringing next-generation nuclear technologies on track.

数据中心扩张的监管框架在地方、州、联邦及国际层面差异巨大。行业必须适应不断变化的分区规则、需求环境和资源影响评估，并对水资源使用制定指导方针。美国在联邦层面日益重视此问题，这可以从近期发布的《美国人工智能行动计划》及随后颁布的三项行政命令中体现，其中包括《加快联邦数据中心基础设施审批的行政命令》。与此同时，弗吉尼亚州《清洁经济法案》（Clean Economy Act）要求Dominion Energy和American Electric Power分别在2045年和2050年前实现100%可再生能源转型。欧洲则规定，对超过500千瓦的设施必须进行排放报告。一些地区为了满足数据中心需求，甚至延长燃煤电厂的运行时间。这导致了开发禁令，例如都柏林至2028年的禁令，以及北弗吉尼亚的类似限制。如前所述，当地社区可能会受到扩张的负面影响。征收权——地方和州政府为公共用途征用私人财产并给予补偿的权力——用于完成输电线路项目，是居民团体关注的潜在问题。

Regulatory frameworks governing data center expansion vary widely across local, state, federal and international jurisdictions. The industry must adapt to the evolving zoning rules, demand environment and resource impact assessments, and impose guidelines on water usage. The U.S. is increasing focus at the Federal level, evidenced by the recent America’s AI Action Plan and subsequent publication of three Executive Orders, including the Executive Order “Accelerating Federal Permitting of Data Center Infrastructure.” Concurrently, Virginia’s Clean Economy Act mandates that Dominion Energy and American Electric Power transition to 100% renewable energy sources by 2045 and 2050, respectively. Europe mandates emissions reporting for facilities larger than 500 kilowatts.Some regions are extending coal plant operations to meet data center demand. This has prompted development moratoriums such as Dublin through 2028 and similar restrictions across Northern Virginia. As we mentioned previously, local communities can be negatively impacted by this expansion. The potential use of eminent domain – the power of local and state governments to take private property for public use, in exchange for compensation – to complete transmission line projects is a looming concern for citizen groups.

虽然“在哪里建设数据中心”的问题尚在探讨，但这些资源限制也引发了更广泛的讨论。基于太空的数据中心可能在未来十年成为可行的解决方案。据称，轨道数据中心能够通过利用太空的低温真空进行被动制冷，并利用太阳能，其效率比地面系统高出多达40%，从而显著提升能源利用效率。其运行成本低至每千瓦时0.1美分（相比地面约5美分），排放量最多可降低10倍，为可持续、高性能计算提供了**吸引力的替代方案。Starcloud是少数敢于突破地面限制的公司之一。其首席执行官设定了远大目标，预计Starcloud的轨道数据中心能够将推理计算的能耗成本降低10倍，而**的制约因素仍是当前的发射成本。

While the question on where to build data centers is fluent, these constraints have also called for a bigger debate. Space-based datacenters could be a viable solution in this next decade. Orbital data centers are said to dramatically improve efficiency by leveraging the cold vacuum of space for passive cooling and harnessing solar energy with up to 40% greater efficiency than Earth-based systems. With operating costs as low as 0.1 cents per kilowatt-hour—compared to 5 cents on Earth—and emissions up to 10 times lower, they offer a compelling alternative for sustainable, high-performance computing.19 Starcloud is one of a handful of companies daring to think beyond terrestrial constraints. Starcloud’s CEO has lofty goals, expecting Starcloud’s orbital data centers to be able to slash the energy cost of inference by 10X, with the biggest inhibitor being current cost to launch.

美国的超大规模数据中心运营商是全球**的计算能力买家和数据中心建设者之一。超大规模运营商为需要大规模数据处理和存储基础设施的组织提供云计算和数据管理服务。亚马逊、微软和Alphabet（谷歌）合计占全球超大规模数据中心容量的60%以上，Meta紧随其后。这些公司在获取有限资源方面具有显著竞争优势，我们认为，它们在推动电力系统的可持续转型方面也具备独特的地位——这一点在更广泛的市场中可能尚未得到充分认识。

The U.S. hyperscalers are among the largest buyers of compute and installers of data center capacity. Hyperscalers provide cloud computing and data management services to organizations that require vast infrastructure for large-scale data processing and storage. Amazon, Microsoft, and Alphabet (Google), in aggregate, account for over 60% of global hyperscale data center capacity, with Meta closely following in ranking.22 These companies have a distinct competitive advantage in accessing limited resources, and we believe, are uniquely positioned to catalyze a sustainable transformation of our power systems — which may still be underappreciated by the broader market.

众所周知，超大规模运营商正在通过大规模资本投入争夺人工智能市场份额。仅Alphabet、亚马逊、微软和Meta今年在数据中心上的投资就将超过3,500亿美元，预计到2026年将达到4,000亿美元（见图3）。

It’s well known that hyperscalers are racing to capture share in the AI market with a sizable wave of capital investment. Alphabet, Amazon, Microsoft and Meta alone are set to spend more than $350 billion this year on data centers and $400 billion in 2026. See Figure 3.

超大规模运营商通过提升效率努力降低成本和资源消耗，其中半导体行业是关键杠杆。NVIDIA的Blackwell平台使组织能够在万亿参数的大型语言模型上构建和运行实时生成式人工智能（GenAI），其成本和能耗相比前一代产品最多可降低25倍。然而，尽管半导体效率将继续发挥关键作用，但随着晶体管尺寸接近单个原子，行业从摩尔定律中获益的空间正变得越来越有限。

Hyperscalers have worked to reduce costs and resource consumption through increasing efficiency, with a strong lever being the semiconductor industry. NVIDIA’s Blackwell platform enables organizations to build and run real-time GenAI on trillion-parameter large language models at up to 25x less cost and energy consumption than its predecessor. However, while semiconductor efficiency will continue to play a critical role, as transistors approach the size of individual atoms, it’s becoming more difficult for the industry to benefit from Moore’s Law.

近期发展凸显了效率提升带来的机遇与风险。中国公司DeepSeek于2025年1月发布的R1模型对市场造成了冲击。据称，DeepSeek AI的模型在能效方面比ChatGPT高出最多40%，通过优化算法和硬件，实现了能耗和碳排放的降低。这类创新有望显著降低人工智能训练和部署的计算成本、财务成本以及能源成本。然而，这一效率突破也揭示了一个关键挑战，即所谓的 “反弹效应”（rebound effect）或杰文斯悖论（Jevons paradox）：当AI的运行成本变低、效率提升时，需求通常会按比例增长，甚至可能超过效率提升带来的节约。虽然在提高模型效率方面已取得重大进展，而且这种趋势必将持续，但效率问题同时涉及软件和硬件两个层面。

Recent developments highlight both the promise and peril of efficiency improvements. Chinese company DeepSeek’s release of its R1 model in January 2025 disrupted the market. DeepSeek AI’s models are said to operate up to 40% more efficiently than ChatGPT, achieved through optimized algorithms and hardware that reduce energy consumption and carbon emissions.Such innovations could dramatically reduce computational, financial and energy costs of AI training and deployment. Yet this efficiency breakthrough illustrates a critical challenge known as the “rebound effect” or Jevons paradox. When AI becomes cheaper and more efficient to operate, demand typically increases proportionally — or even exceeds the efficiency gains. Substantial progress has been made in making models more efficient, and this is certain to continue. Efficiency is both a software and a hardware issue.

数据中心的电力挑战不仅仅是化石能源与可再生能源的对立。实际上，电网运营商和电力公司需要采取“全方位”能源扩展策略——从天然气和核能，到太阳能、风能，以及地热和小型模块化反应堆（SMR）等新兴技术，充分利用所有可用的能源来源。人工智能基础设施加速带来的电力需求规模巨大，超出了任何单一发电类型在规定时间内的供给能力。这就需要引入更多能源、优化能源使用方式，并推动创新能源解决方案的发展。

The data center power challenge is not merely a fossil fuels versus renewables debate. The practical reality is that grid operators and utilities need to adopt an “all-of-the-above” energy expansion strategy - deploying every available power source from natural gas and nuclear to solar, wind and emerging technologies such as geothermal and small modular reactors (SMRs). The sheer scale of electricity demand accelerated by AI infrastructure exceeds what any single generation type can provide within required timeframes. This requires bringing more energy online, optimizing how we use it, and advancing innovative energy solutions.

随着可再生能源的持续发展，电网互操作性的重要性日益凸显。亚马逊、微软、Meta和谷歌是企业可再生能源购电协议（PPA）的四大买家，总计签约容量超过50吉瓦，相当于瑞典的发电能力。

As renewables continue to develop, grid interoperability becomes increasingly important. Amazon, Microsoft, Meta, and Google are the four largest purchasers of corporate renewable energy power purchase agreements (PPAs), having contracted over 50GW, equal to the generation capacity of Sweden.

利用热能电池及其他储能方法的公司，已经在让多余的可再生能源储存变得更简单、更经济。关键在于将数据中心潜在地转变为巨型电池，吸收多余的太阳能和风能。行业正在试点一些支持电网的早期案例，例如Alphabet的需求响应（demand response）方法，通过将当前紧急的计算任务调整到其他时间和地点，减少电网压力时期的数据中心用电需求。需求响应可以成为降低电网新建电力投资需求的有效工具，从而在电网扩展中最小化环境风险，同时减轻对社区的负面影响。

Companies utilizing heat batteries and other storage methods are already making it easier and cheaper to store excess renewable generation. The key lies in potentially transforming data centers into giant batteries that absorb excess solar and wind energy. The industry is piloting early examples of supporting grids such as Alphabet’s demand response method which allows reduced data center power demand during periods of grid stress by shifting now-urgent computing tasks to an alternative time and location. Demand response can be a useful tool to reduce the need for new power investment on the grid. This minimizes environmental risk in grid expansion as well as detrimental community impact.

核能作为可靠、低碳、可全天候供电的能源，正重新受到关注，有望帮助平衡间歇性可再生能源的波动。不过，监管、成本、建设周期及安全等方面仍存在挑战。数据中心运营商正在探索在现有核电站附近选址的可能性。随着核技术的持续进步，公用事业监管机构对批准新核能容量的态度也可能更加宽容。

Nuclear power is gaining renewed attention as a reliable, low-carbon source of always-on electricity that could help balance intermittent renewables. Although, hurdles exist around regulatory, cost, timing, and safety concerns. Data center operators are exploring the potential to site facilities near existing nuclear plants. As nuclear technology continues to advance, utility regulators could become more comfortable approving new nuclear capacity.

所有这些发展都指向一个未来：数据中心不仅是电力的消费者，更将成为构建更可持续、更具韧性的电网的积极合作伙伴。超大规模运营商凭借其巨大的能源需求，既有动力，也具影响力推动这一转型。布朗咨询（Brown Advisory）投资组合经理、私人客户及捐赠基金联合首席投资官Erika Pagel形象地将此称为技术与能源之间的“非同寻常的关系或结合” 。

All of this points to a future where data centers are not just power consumers, but active partners in laying the foundation for a more sustainable, resilient grid. Hyperscalers, with their substantial energy demand, have both the incentive and the influence to drive this transformation. Erika Pagel, Brown Advisory Portfolio Manager, Co-CIO of Private Client, Endowments & Foundations, aptly describes this as an “unlikely relationship, or marriage” of technology and energy.

在用水方面，制冷带来了多层次的挑战。尽管相关研究仍在进行，且不同数据中心的用水强度各异，但我们通常听到的数据是平均每个数据中心每天消耗约300万至500万加仑水。大部分为当地淡水，这对饮用水供应造成压力，尤其是在干旱多发地区。世界资源研究所（World Resources Institute）预测，全球约三分之一的数据中心位于水资源高度或极度紧张的地区。因此，领先的数据中心运营商正在将节水优先化，并探索替代制冷方法，例如浸没式冷却（immersion cooling）和直接芯片冷却（direct-to-chip cooling）。

As it relates to water usage, cooling presents a multifaceted challenge. While research continues and each data centers’ water intensity varies, the most common range we hear is the average data center consumes 3-5 million gallons of water per day. Most is local freshwater, putting pressure on drinking water supplies, especially in drought-prone areas. The World Resources Institute forecasts that about one-third of data centers globally are now located in areas with high or extremely high levels of water stress.Accordingly, leading data center operators are prioritizing water conservation and exploring alternative cooling methods, such as immersion cooling and direct-to-chip cooling.

一个常被忽视的事实是，人工智能本身正被用于应对能源和电网约束。全球知名能源公司迅速采用AI，认识到其在优化规划、设施管理、环境影响评估、能源储存与分配等方面的潜力。利用AI提升效率在科技行业并非新鲜事。早在2016年，谷歌就宣布，通过将DeepMind的机器学习应用于自家数据中心，其制冷能耗最多降低了40%。

What is often overlooked is how AI itself is being harnessed to solve the energy and power grid constraints. Prominent energy companies worldwide have swiftly embraced AI, recognizing its potential to optimize planning, facility management, environmental impact, energy storage, and distribution. Leveraging AI to drive efficiencies is not a new phenomenon in the technology sector. In 2016, Google announced that by applying DeepMind’s machine learning to their own Google data centers, the company managed to reduce the amount of energy used for cooling by up to 40%.

微软最新的数据中心采用了芯片级液冷设计。在建设期间注满液体后，该系统可持续循环水流于服务器与冷却机之间，实现散热，而无需额外用水。亚马逊也在新建及现有数据中心推出了一种创新的直接芯片液冷（direct-to-chip liquid cooling）解决方案，用于高密度AI计算芯片。据称，这些组件在制冷高峰期可将机械能耗降低高达46%，且不增加用水量。

Microsoft’s newest datacenters are designed with chip-level liquid cooling. Once filled during construction, the system continuously circulates water between the servers and chillers, dissipating heat without requiring additional water.Amazon also introduced a novel, direct-to-chip liquid cooling solution for high-density AI compute chips in new and existing data centers. These components are said to reduce mechanical energy consumption by up to 46% during peak cooling — without increasing water usage.

尽管生成式人工智能（GenAI）的扩展带来挑战，超大规模运营商仍致力于实现其雄心勃勃的可持续发展目标。这不仅仅是“锦上添花”，而是商业必需。数据中心是超大规模运营商的收入核心，因此成本效益高且资源（能源与水）使用高效是经济增长的生命线。微软在2024财年的间接排放（Scope 3）比2020年的基线增加了26%，“主要来自于更多数据中心的建设”。尽管如此，微软仍计划在2030年前成为碳负排放、水资源正贡献、零废弃物的企业，并保护生态系统。

Hyperscalers remain committed to their ambitious sustainability goals despite the challenges deterring progress as GenAI scales. This is not a “nice to have”, it’s a business imperative. Data centers are the heart of hyperscalers revenue generation, so the incentive is clear – cost effective and efficient resources (energy and water) are lifelines to economic growth. Microsoft’s indirect emissions (Scope 3) increased in FY24 by 26% from their 2020 baseline, which “primarily comes from the construction of more datacenters”. Still, Microsoft aims to be a carbon negative, water positive, zero waste company that protects ecosystems — all by 2030.

超大规模运营商正在加速并推动创新解决方案的规模化应用。微软2025年《环境可持续发展报告》强调，其首批使用大规模木材（mass timber）建造的数据中心已投入运营。大规模木材是一种坚固、超轻的木材，采用混合结构建造模式，预计相比典型预制混凝土，新数据中心的隐含碳足迹可降低高达65%。微软还宣布了一项重大100亿美元可再生能源交易，与布鲁克菲尔德资产管理公司（Brookfield Asset Management）合作。从2026年开始，布鲁克菲尔德将主导部署超过10.5吉瓦的可再生能源容量，相当于10座核电站的发电量。此外，微软与能源开发商Helion Energy签订了全球首个聚变能源购电协议，预计其首座聚变电厂将于2028年投入运行。微软还开始对小型模块化反应堆（SMRs）进行早期投资，预计本十年将建造首批SMR，并在2030年代加速全球部署。类似地，亚马逊以6.5亿美元收购了宾夕法尼亚州的一座核电数据中心园区，以补充其可再生能源布局。Alphabet签约新增8GW清洁能源发电，并在2024年投运2.5GW，同时签署首创性合作协议，通过Kairos Power开发的一系列SMRs解锁新的清洁能源。2025年7月，Alphabet宣布投资基于CO₂的电池储能技术，可调度清洁能源长达24小时，而传统锂电池仅为约4小时。

There are ongoing examples of hyperscalers accelerating and enabling scale of innovative solutions. Microsoft’s 2025 Environmental Sustainability Report highlights the launch of their first datacenters constructed with mass timber, a strong, ultralightweight wood in a hybrid construction model which is projected to reduce the embodied carbon footprint of new datacenters by up to 65% compared to typical precast concrete.Microsoft also announced a monumental $10 billion renewable energy deal with Brookfield Asset Management. Commencing in 2026, Brookfield will spearhead the deployment of more than 10.5GW of renewable energy capacity, equivalent to the output of 10 nuclear power plants.Microsoft entered into the world’s first fusion energy purchase agreement with energy developer Helion Energy for the provision of electricity from its first fusion power plant expected to be online by 2028.We’re also seeing companies such as Microsoft showing early signs of investment in SMRs, with the first SMRs expected to be built this decade and accelerated global deployment in the 2030s.Similarly, Amazon purchased a nuclear-powered data center campus in Pennsylvania for $650 million to supplement its renewables. Alphabet added new clean energy generation by signing contracts for 8GW and bringing 2.5GW online in 2024 alone and signed a first-of-a-kind partnership to unlock new, clean power from a series of SMRs, which will be developed by Kairos Power.In July, Alphabet said it invested in CO₂-based battery storage that can dispatch clean energy for up to 24 hours compared with the typical four hours of lithium-ion batteries.

生成式人工智能（GenAI）引发了一个拐点，它在考验我们电力基础设施承载能力的同时，也推动了创新，使数据中心——我们的数字基础设施——更加高效。我们认为，超大规模运营商具备规模、资本和工程能力，能够在长期内实现计算能力增长与资源可持续性之间的平衡。这需要多管齐下的策略，包括：大规模投资可再生能源、储能和电网平衡能力；关注水资源节约及替代制冷技术；在可行情况下部署现场发电和专用可再生能源；通过政策激励推动电网及市场结构现代化；以及在地球或太空中创新AI基础设施以提高效率。作为投资者，我们会关注数据中心价值链上的各类公司，包括运营商、半导体、电力电子、制冷及软件供应商等。

GenAI sparks an inflection point to stress-test our power infrastructure while unleashing innovation to make data centers, our digital infrastructure, more efficient. We believe that hyperscalers have the scale, capital, and engineering prowess to balance compute growth with resource sustainability in the long run. This will require a multi-pronged approach including substantial investments in renewables, storage, and grid balancing capabilities; focus on water conservation and alternative cooling technologies; deployment of onsite generation and dedicated renewable energy where feasible; policy incentives to modernize power grids and market structures; and innovation in AI infrastructure (potentially on Earth and in space) to be more efficient. As investors, we monitor companies across the data center value chain, including operators, semiconductors, power electronics, cooling, and software providers, to name a few.

半导体通常是计算周期中的早期受益者，随后是数据中心基础设施的其他组件。时间将验证数据中心在推动电网储能和灵活性扩展、创造新技术与软件市场中的作用。一个有趣的软件投资案例是Cadence Design Systems的“Cadence Reality DC”，这是全球首个覆盖整个园区的数据中心数字孪生模拟器。客户正在使用Cadence Reality DC对液冷进行建模，以提升高密度AI设备的制冷效率。供应链中最具创新性的公司将推动可持续数据中心的发展。一种潜在方案是模块化数据中心（MDCs）。MDC是一种紧凑、可移动的集合体，包含传统数据中心的所有关键组件，包括服务器、网络设备、制冷和存储。例如，戴尔拥有一系列MDC解决方案，可显著降低现场部署所需的资源与时间。

Semiconductors are the early beneficiaries in a compute cycle, with the other components of data center infrastructure to follow. Time will tell the extent to which data centers catalyze scaling grid storage and flexibility, creating markets for new technologies and software. An interesting software investment example is Cadence Design Systems’ “Cadence Reality DC,” the first campus-wide digital twin simulator for data centers.Customers are modeling liquid cooling within Cadence Reality DC to facilitate more efficient cooling for high-density AI equipment.The most innovative companies in the supply chain will enable sustainable data center development. One potential solution is Modular Data Centers (MDCs). An MDC is a compact and portable collection of all the key components of a traditional data center, including servers, networking equipment, cooling, and storage.Dell, for example, has a portfolio MDC solutions which significantly reduces deployment and resources needed onsite.

在平衡短期挑战与长期潜力的过程中，我们发现技术赋能型企业提供了吸引人的投资机会，这些企业帮助数据中心克服限制，提高效率并增强电网灵活性。投资组合经理Eric Gordon, CFA指出：“这一主题几乎单枪匹马推动了市场在近期复苏阶段的回暖。”

In balancing near-term challenges with long-term upside, we find attractive investment opportunities in technology enablers across the broader ecosystem that are helping data centers overcome constraints and become more efficient and grid flexible. As Portfolio Manager Eric Gordon, CFA, explains, "This One Theme [GenAI] has almost single-handedly driven the market recovery post-Liberation Day."

从电力管理、储能、制冷到能源解决方案，机会广泛且多样化。Pure Storage作为全闪存数据存储解决方案的领先提供商，通过每TB减少高达54%的能耗、并大幅提升存储密度，提升了数据中心的效率 。Vertiv通过其CoolPhase Flex混合冷却系统及基于液体的CoolChip技术，在AI数据中心推动制冷效率提升，可减少高达50%的能耗和部署时间。技术发展迅速，GenAI的故事正以前所未有的速度展开。绝大多数GenAI应用与实现仍处于早期阶段，但可以明确的是，随着GenAI应用和推理规模扩大，电力需求将相应增长，使效率问题愈发关键。未来充满未知，我们无法预见一切，这也正是投资分析师角色的激动人心之处。数据中心展示了环境可持续性同时是商业必需和竞争优势的关键交汇点。如果超大规模运营商在推动计算效率跨越式提升的同时，促进更强大、韧性更高、可持续的电网建设，我们相信，将释放巨大的价值。

From power management, storage, cooling and energy solutions, the opportunity set is large and diversifying. Pure Storage, a leading provider of all-flash data storage solutions, enhances data center efficiency by delivering up to 54% lower energy consumption per terabyte and dramatically higher storage density than traditional systems.Vertiv is another example driving efficiency in AI data centers through advanced cooling systems such as its hybrid CoolPhase Flex and liquid-based CoolChip technologies — reducing energy consumption and minimizing deployment time by up to 50%.Technology evolves quickly, and the tale of GenAI is unfolding at an unprecedented pace. The vast majority of GenAI use cases and implementations remain in the early stages. Yet, what is quite clear is that as GenAI applications and inferencing scale, power demand is set to increase commensurately — making efficiency increasingly critical. There are endless unknowns of what the future holds, and we do not hold a crystal ball. This is what makes the role of an investment analyst thrilling. Data centers demonstrate the critical intersection of environmental sustainability being both a business imperative and a competitive advantage. If hyperscalers catalyze a more robust, resilient and sustainable power grid while continuing to achieve step-changes in compute efficiency, we believe immense value will be unlocked.