人工智能如何解决自身的能源危机 | 瓦伦·西瓦拉姆

选择瑞银，您将拥有一位真正全球化的合作伙伴，我们融合前沿科技、创新方法与意外机遇。以热忱之心、专业之姿，为您提供关键洞察，解答重要问题。因为我们深知，每日黎明即起，皆可更进一步。这就是我们理解的银行业——不仅是工作，更是一门精湛工艺。

With UBS, you have a truly global partner incorporating new technologies, innovative approaches, and unexpected opportunities. Leading you to insights that help answer the questions that matter, delivered with passion, care, and unmatched expertise. Because it's about rising with the dawn each day, knowing that we can do even better. That's what banking is to us. Not just work, but a craft.

瑞银。银行业是我们的精湛工艺。

UBS. Banking is our craft.

您正在收听TED每日演讲，我们每天为您带来激发好奇的新思想。我是主持人Elise Hu。AI算力需求正呈指数级增长。当众人为AI革命欢欣鼓舞时，我们却不得不面对一个严峻现实：全球资源正因AI技术飙升的能源需求而承受前所未有的压力。

You're listening to TED Talks Daily, where we bring you new ideas to spark your curiosity every day. I'm your host, Elise Hu. The need for AI computing power is growing at an exponential rate. And while many are excited about the AI revolution, we're also facing a terrible truth. The world's resources are being taxed at an unprecedented rate by the soaring energy needs of AI technologies.

本次演讲中，电网未来学家Varun Sivaram将分享其研发弹性AI数据中心的工作成果——这些中心如何反哺电网，以及如何以负责任的方式支持AI繁荣发展。

In this talk, grid futurist Varun Sivaram shares his work developing flexible AI data centers, how they could actually help power our energy grid, and how to support the AI boom but responsibly.

在亚利桑那州凤凰城酷热难当的某天，当百万台空调推高电网负荷时，一群高耗能AI服务器却逆势而行。它们实际提供了帮助：甲骨文数据中心的这些AI计算机在三小时内降低25%功耗，精准缓解了当日用电高峰压力。关键在于，搭载先进英伟达芯片的服务器在持续满足严苛性能要求的同时，仍能完成AI大语言模型的训练、微调与应用。我们Emerald AI团队成功策划了这场具有开创意义的弹性AI计算演示。

On a blistering hot day in Phoenix, Arizona, as a million air conditioners drove up demand on the power grid, a cluster of energy hungry artificial intelligence servers bucked the trend. They actually helped. For three hours, these AI computers at an Oracle data center dropped their power consumption by 25% to provide perfectly timed relief during that day's peak demand. And critically, the advanced NVIDIA chips continued to meet the stringent performance requirements of their tasks, training, fine tuning and using AI large language models. Our team at Emerald AI orchestrated this first of a kind demonstration of flexible AI computing.

我们并非孤军奋战。谷歌同样取得瞩目进展。若能将我们的技术推广至全国乃至全球，将有助于解决这个时代最严峻的挑战——在为AI革命供能的同时，推进建设更可靠、经济、清洁的电网。AI非但不会拖累电网，反而可能成为其救星。要理解其中缘由，我们需要重新构想AI供能这一挑战。为此，我彻底重塑了自己的职业生涯。

And we're not alone. Google's also made impressive strides, Scaling up our technologies across the country and around the world could help solve one of the biggest challenges of our time, powering the AI revolution, while also advancing a more reliable, affordable, and clean power grid. Far from undermining it, AI could actually help save the grid. To understand why, we need to reimagine the challenge of powering AI. And so I've reinvented my own career.

过去十五年，作为能源高管和美国首席清洁能源外交官，我专注于建设更多清洁能源。但能源供给仅是等式的一半。因此我创立Emerald AI聚焦另一半：需求侧。通过帮助AI智能用能、支持电网，释放现有的大量闲置电力容量。若不具备这种能力，我们将面临一场迫在眉睫的危机——两个数万亿美元级网络的历史性碰撞。

For fifteen years as an energy executive and as America's lead clean energy diplomat, I focused on building more clean energy. But energy supply is just half the equation. And so I founded Emerald AI to focus on the other half, demand. Helping AI intelligently use energy, support grids, and unlock massive stranded power capacity that already exists. Without this capability, we face an impending crisis, a historic collision between two multi trillion dollar networks.

AI数据中心网络正在快速增长，而老化的电网却完全无法应对这种新需求。各位，这有多方面的坏消息。首先，美国可能在AI领域落后。在全球数据中心之都弗吉尼亚州，新数据中心接入电网需要长达七年时间。其次，社区的电价正在飙升。

The network of AI data centers, it's rapidly growing, and an aging electricity grid utterly unprepared for all this new demand. That's bad news, folks, for multiple reasons. First, America risks falling behind in AI. In Virginia, the data center capital of the world, it takes up to seven years to connect new data centers to the grid. Second, power prices are soaring for communities.

仅在2025年，当我们建设新电网和发电厂时，数据中心需求就使得俄亥俄州哥伦布市的家庭年均电费上涨了240美元。而这只是开始。随着数据中心用电量从目前占美国总需求的4%激增至2030年的12%，相当于给美国电网又增加了一个德国的用电量。第三，化石燃料将支撑AI数据中心的繁荣发展，因为当前它们需要可靠的电力供应。

Just in 2025, as we built new grids and new power plants, data center demand drove up the average annual household power price in Columbus, Ohio by $240. And this is just the beginning. As data centers surge from 4% of US power demand today to 12% by 2030. That's like adding another Germany to The US power grid. And third, fossil fuels are set to power the boom in AI data centers, which require reliable power supplies today.

在美国，天然气是推动AI发展的主要能源，而印度等国家将因煤炭使用增加导致全球碳排放上升。但事情不必如此。这个最大的新增用电群体实际上可以成为电网的最佳盟友。关键在于一个看似简单的概念：灵活性。

In The United States, natural gas is powering most AI growth, and countries like India will see rising coal use increasing global carbon emissions. But it doesn't have to be this way. The biggest new user of electricity could actually be our grid's greatest ally. The key lies in something deceptively simple. Flexibility.

这与提高效率或减少总体能耗不同。如果AI能在用电时间上稍作调整，就能消化当今电网大量闲置的电力。想象我们的电力系统就像一条高速公路，每月只有几小时面临高峰拥堵——就像亚利桑那州凤凰城夏季最热的那天，空调需求达到顶峰时。在这些时刻，电网可能被这些耗电量即将超过佛蒙特州整体用电量的巨型数据中心压垮。

That's distinct from efficiency or using less energy overall. Rather, if AI were just a little more flexible in when it uses energy, it could consume vast amounts of otherwise stranded power on today's grids. Think of our electric power system as a superhighway that faces peak rush hour just a few hours per month. Think of that hottest day of the summer in Phoenix, Arizona when air conditioning demand peaks. On those days, grids risk being overwhelmed by these massive new data centers that may soon consume more than a gigawatt or more energy than the state of Vermont consumes.

但大多数时候，发电厂的运行远低于其满负荷，输电线路的承载量也远未达上限，就像那条高速公路。全年平均来看，电力系统有一半的容量处于闲置状态。如果在电网真正吃紧的用电高峰时段，AI数据中心能动态降低耗电量，就能充分利用全年的闲置容量。这就像暂时让重型卡车离开公路，使剩余车流畅通行驶。事实证明，只要AI数据中心具备适度灵活性——每年不到2%的时间，每次仅需几小时削减25%的用电需求，美国现有电网就能额外承载100千兆瓦的新增数据中心。

But most of the time, power plants are running well below their full capacity and transmission lines are carrying less power than they could, just like that highway. On average, throughout the year, half of the power system's capacity goes unused. What if during those peak rush hour periods, when the grid is truly stressed, AI data centers could dynamically reduce their power consumption and take advantage otherwise of all that spare capacity throughout the year. It would be like briefly taking 18 wheelers off of that road to let the remaining traffic flow smoothly. Well, it turns out that if AI data centers were just modestly flexible, just less than 2% of the year, trimming demand by a quarter, just a couple hours at a time, America could fit up to a 100 gigawatts of new data centers on existing power grids across the country.

这意味着今天就能释放4万亿美元的AI投资潜力，而无需等待多年建设新基础设施。当然，随着数据中心、工厂等用电主体增加，美国经济持续增长确实需要更多能源。但通过赋予AI数据中心灵活性，我们既能审慎扩建电网，又能争取时间建设清洁的核能或地热电厂。更重要的是，具有灵活性的AI数据中心就像电网上的巨型减震器，可以整合间歇性但廉价的太阳能和风能，从而降低AI的用能成本。这就是我的工作方向。

That's $4,000,000,000,000 of AI investment unlocked today without waiting years for new infrastructure. Now, to be sure, America will need even more energy to power our growing economy as data centers, factories, and other users of electricity join. But by making AI data centers flexible, we can prudently expand our grid and buy ourselves time to build clean nuclear or geothermal power plants. And what's more, with flexible AI data centers acting as giant shock absorbers on the grid, we can integrate intermittent but cheap solar and wind power, driving down the cost of energy for AI. So that's what I do.

我和团队正在开发软件大脑，为AI数据中心提供这种关键灵活性。这是服务于AI的AI系统，我们称之为'翡翠指挥家'。其原理在于利用我们所说的'时空灵活性'——这个专业术语背后其实是个简单理念。

My team and I are building the software brain to give AI data centers this crucial flexibility. It's an AI for AI. We call it the Emerald Conductor. It works by harnessing something we call spatiotemporal flexibility. That's a fancy term for a simple idea.

让我们拆解一下。首先是时间灵活性。并非所有AI任务都相同。有些工作负载，比如训练或微调AI模型、进行深度研究或运行大型科学模拟，我们称之为可批处理任务。它们极其重要，但不必立即完成。

Let's break it down. First, temporal flexibility. Not all AI jobs are created equal. Some workloads, like training or fine tuning an AI model, conducting deep research or running a massive scientific simulation, are what we call batchable. They're incredibly important, but they don't have to be completed right this second.

软件可以在电网压力大时智能地短暂暂停或减缓这些工作负载，待电力充足时再加速运行。其次是空间灵活性。以生成式AI聊天机器人的查询为例，虽然无法暂停响应查询的任务，但可以以光速将其转移到全国各地。因此即便我们在电力传输建设上遇到困难，仍可利用虚拟传输或遍布全国乃至全球的光纤网络，将AI工作负载从电网紧张的城市数据中心（比如炎热天气下的凤凰城）转移到当前电力充沛的区域（比如风大的大平原地区）。

Software can intelligently pause or slow these workloads briefly when the grid is stressed and then speed them back up when there's plenty of power available. Then there's spatial flexibility. Think of your query to a generative AI chatbot. You can't pause the job of responding to that query, but you can move it across the country at the speed of light. So even as we struggle to build electric power transmission, we can take advantage of virtual transmission or the network of fiber optic cables that crisscrosses the country and the planet to move AI workloads from a data center in a city where the grid's currently strained, let's say Phoenix on a hot day, to a data center in a region where there's presently abundant power, say the windswept Great Plains.

AI工作负载得以完成，而电网在最需要时获得喘息。用户甚至毫无察觉，因为幕后有AI在协调AI。数据中心成为电网的智能协作伙伴。我们知道这行之有效。还记得我提到的演示吗？

The AI workloads get done, but the grid gets a break right when it needs it most. And the user never even notices because behind the scenes, there's an AI orchestrating AI. Data centers become smart, cooperative partners to the power grid. And we know it works. Remember that demo I told you about?

它实现了。2025年5月在亚利桑那州凤凰城，我们采用256个GPU服务器集群运行了混合AI工作负载——部分高度灵活，部分完全刚性，更多介于两者之间。某个炎热午后，软件收到当地公用事业公司将达到需求峰值的信号，于是Emerald Conductor精准地在电网要求的三小时内优雅地将AI计算负载降低了25%。

It happened. In May 2025 in Phoenix, Arizona, we took a cluster of two fifty six GPU servers and we ran a mix of AI workloads. Some highly flexible, others entirely inflexible and many in between. One hot afternoon, our software received a signal that the local utility was going to reach its peak demand. And so Emerald Conductor gracefully reduced the AI computational power load by 25% for the exact three hours requested by the grid.

我们证明了AI数据中心能在电网紧张时灵活调节，在用户需要时全速运转。但技术验证只是第一步，最困难的是说服庞大的能源与AI行业改变运营方式合作共赢。一个多世纪以来，电力公司始终认为用户无法在用电高峰时主动降耗。确实，

We proved that AI data centers can flex when the grid is tight and sprint when users need them to. But proving the technology was just the first step. The hardest part will be to convince the enormous energy and AI industries to cooperate and to change the way that they operate. For over a century, electric power utilities have assumed that their users can't simply reduce their power consumption when the grid faces peak rush hour. Sure.

在极少数情况下，电力公司会要求家庭调节恒温器或大型工业负载降低消耗，但这些干预通常微乎其微。而AI数据中心具有根本性的不同，拥有变革性的灵活潜力——相比需要聚合的分散家庭用电，它们是巨型能源用户；比大型制造设施响应更快更优雅；还能以光速在全国转移工作负载，这是其他能源用户无法做到的。

In limited situations, a utility may request homes to adjust their thermostats or large industrial loads to dial down consumption, but these interventions are typically tiny and marginal. But AI data centers are fundamentally different with a transformative potential to be flexible. They're massive energy users compared with tiny household loads that need to be aggregated. They respond faster and more gracefully than large manufacturing facilities. And they can move their workloads around the country at the speed of light, which no other energy user can do.

这就是为什么我对EPRI的DC Flex等能联姻能源与科技行业的计划如此兴奋。在美国即将进行的演示中，以及与英国国家电网的合作中，Emerald将展示AI工作负载如何跨区域灵活迁移，并证明Conductor等软件能像指挥交响乐般协调AI工作负载，与电池等现场能源设备配合，为电网提供更强灵活性。我们正与合作伙伴英伟达共同构建下一代数据中心（或称AI工厂）的参考设计，使其具备电力灵活性——这样看到认证的电力公司就能更快接入对电网友好的AI工厂。那么这一切将把我们带向何方？

That's why I'm so excited about initiatives that bring together the energy and technology industries like EPRI's DC Flex. In upcoming demonstrations in The United States and with National Grid in The United Kingdom, Emerald will showcase how AI workloads can flex and move across regions. And will prove that software like Conductor can orchestrate a symphony of AI workloads in concert with on-site energy equipment like batteries to deliver even more flexibility to power grids. And with our partner NVIDIA, we're building a reference design for next generation data centers or AI factories to be power flexible so that utilities that see the certification can more swiftly connect a grid friendly AI factory. So where does this all leave us?

这意味着我们无需等待电网升级多年，现在就能建设所有所需的人工智能基础设施来增强竞争优势。灵活的AI数据中心不仅不会压垮电网，还能在电网达到临界点前提供缓解，避免轮流停电。它们不仅不会推高电价，反而可能降低电价，因为灵活的AI数据中心能更高效地利用现有能源基础设施，推迟昂贵的升级。AI激增的能源需求不仅不会只刺激化石燃料消耗，还可能推动国内外更多清洁能源接入电网。如今太阳能已是全球最便宜、增长最快的能源。

Well, means rather than wait years for grid upgrades, we can build all the AI infrastructure we need right now to sharpen our competitive edge. And far from crashing the grid, flexible AI data centers can provide relief before the grid hits a breaking point, avoiding rolling blackouts. Rather than increasing power prices, they could actually go down as flexible AI data centers more effectively utilize the existing energy infrastructure, deferring expensive upgrades. And rather than goose demand only for fossil fuels, AI's soaring energy needs could encourage more clean energy onto the grid at home and abroad. Solar today is the cheapest, fastest growing power source on the planet.

想象一下灵活的AI数据中心能够调整能耗以适应日间太阳能峰值，或转移负载以更好地将清洁能源整合入电网。AI革命已经到来。我相信我们可以兼顾：迅猛的创新、对AI的大规模投资，以及为所有人提供充足、可负担、可靠且清洁的能源。而支持灵活AI基础设施的人工智能，可能成为未来能源系统的关键。

Imagine flexible AI data centers capable of ramping their energy consumption to match daytime solar peaks or shifting their loads so that they better integrate clean energy onto the grid. The AI revolution is here. And I believe we can have it all. Breakneck innovation, massive investments in AI, and abundant, affordable, reliable, and clean energy for all. And AI for flexible AI infrastructure could be a linchpin for our future energy system.

谢谢。

Thank you.

这是瓦伦·西瓦拉姆在2025年纽约TED Countdown活动上的演讲，该活动与贝索斯地球基金合作举办。若您对TED的内容策划感兴趣，可访问ted.com/curationguidelines了解更多。今天的节目就到这里。《TED每日谈》是TED音频合集的一部分。本演讲经TED研究团队事实核查，由玛莎·埃斯特瓦诺斯、奥利弗·弗里德曼、布莱恩·格林、露西·利特尔和坦西卡·桑玛·尼冯制作编辑。

That was Varun Sivaram at a TED Countdown event in New York in partnership with the Bezos Earth Fund in 2025. If you're curious about TED's curation, find out more at ted.com/curationguidelines. And that's it for today. Ted Talks Daily is part of the Ted audio collective. This talk was fact checked by the Ted research team and produced and edited by our team, Martha Estevanos, Oliver Friedman, Brian Greene, Lucy Little, and Tansika Sungmar Nivong.

本期节目由克里斯托弗·法伊齐·博根混音，艾玛·陶伯纳和丹妮拉·巴拉雷佐提供额外支持。我是艾莉丝·胡。明天我将为您带来新的思想火花。感谢收听。

This episode was mixed by Christopher Faizy Bogan. Additional support from Emma Taubner and Daniella Balarezo. I'm Elise Hu. I'll be back tomorrow with a fresh idea for your feed. Thanks for listening.

我们总在谈论政府失灵，但成功案例呢？我必须说，我第一次在线更新护照的经历简直棒极了。

We talk about how the government isn't working but what about when it does? I have to say I renewed my passport for the first time online and it was awesome.

听到这个消息我太高兴了。

I am so glad to hear that.

关于通过改善系统运作来重建民主信任的理念。这些内容将在下一期的NPR TED广播时间播客中呈现。无论您通过何种平台获取播客，都欢迎订阅或收听TED广播时间。

Ideas about restoring trust in democracy by making systems just work better. That's next time on the TED Radio Hour podcast from NPR. Subscribe or listen to the TED Radio Hour wherever you get your podcasts.