«

Meta与核能公司签署合作协议

qimuai 发布于 阅读:7 一手编译


Meta与核能公司签署合作协议

内容来源:https://aibusiness.com/data-centers/meta-signs-deals-with-nuclear-companies

内容总结:

科技巨头押注核能:Meta签下千兆瓦级核电协议,为AI竞赛注入新动力

在生成式AI浪潮席卷全球的背景下,科技巨头对算力与能源的争夺已进入白热化阶段。近日,Meta公司宣布与三家核能企业达成一系列长期协议,预计到2035年将获得高达6.6吉瓦的核电供应,为其人工智能(包括“超级智能”系统)的宏伟目标提供动力基础。这一举措不仅使其在核电采购规模上超越亚马逊、微软和谷歌等竞争对手,也为激烈的AI竞赛增添了能源保障这一关键变量。

核能成为AI“新燃料”
随着生成式AI和智能体系统对训练与推理的能耗需求呈指数级增长,稳定且大规模的电力供应已成为AI超大规模企业的首要战略。Meta此次签约对象包括Vistra、泰拉能源以及由OpenAI首席执行官萨姆·奥尔特曼主要投资的Oklo公司。协议内容涵盖从现有核电站购电、资助新一代先进反应堆建设,并锁定未来多个核电机组的电力权益。

这些合作是Meta构建AI基础设施的重要一环,将为其正在俄亥俄州建设的“普罗米修斯”超级计算集群等项目供电。此前,Meta已与Constellation能源公司签署了购电协议。行业观察指出,由于现有电网难以满足激增的电力需求,转向核能等替代能源已成为Meta、谷歌和微软等AI厂商的共同选择。

机遇与争议并存
分析认为,Meta的核能布局一方面展示了其推动先进AI技术发展的影响力与野心,有助于改善其公众形象;另一方面,核能虽被视为比化石燃料更清洁的能源,但其放射性废物处理等潜在风险仍受关注。

然而,部分协议涉及的项目尚未建成,全部Natrium机组预计到2035年左右才能完工,有分析师提醒需审慎看待Meta的长期承诺是否存在过度扩张的风险。此外,大规模工业用电协议可能推高区域能源价格,普通消费者需警惕其对电费账单的潜在影响。

行业影响与未来展望
此举可能促使其他云服务巨头加速跟进类似的能源协议。对于关注AI市场的企业而言,专家建议保持谨慎态度——当前少数头部企业在基础设施上的“豪赌”,可能带来前所未有的回报,也可能面临巨大的失败风险。

随着AI发展对资源(电力、芯片、冷却用水等)的消耗日益加剧,科技公司在追求技术突破的同时,如何平衡创新需求与可持续发展,将成为整个行业必须面对的核心议题。

中文翻译:

由谷歌云赞助
如何选择首批生成式AI应用场景
要着手应用生成式AI,首先应关注那些能够优化人类信息交互体验的领域。

这些协议可能提升Meta在AI竞赛中的领导者形象,并彰显其保障AI数据中心能源供应的能力。
上周五,Meta宣布已与三家核能公司达成协议,为其数据中心供电。这家社交媒体巨头正积极推进其AI战略,包括对超级智能的追求。
Meta与核能公司Vistra、TerraPower以及Oklo(其最大投资者之一是OpenAI首席执行官萨姆·奥尔特曼)达成的协议,预计到2035年将为Facebook、Instagram和WhatsApp的母公司提供高达66亿瓦的电力,用于支撑其数据中心运营。
这些协议使Meta成为核电的重要采购方,其规模已超越亚马逊、微软和谷歌,为AI竞赛增添了新的维度。
近年来,数据中心的重要性日益凸显,尤其是生成式AI和智能体AI模型在训练和推理过程中需要消耗巨大能量。新协议可能推动Meta更接近其构建超级智能的目标——即在大多数认知任务中超越人类表现的系统。
Meta与Vistra签署了为期20年的购电协议,将从该公司与TerraPower在宾夕法尼亚州和俄亥俄州运营的核电站购买超过2600兆瓦电力。同时,Meta将支持Oklo在俄亥俄州派克县园区开发12亿瓦核能项目。Meta还将资助TerraPower在2032年前建设两座钠冷先进反应堆,并有权从未来六座反应堆中获得更多能源。
此前,Meta已于去年与Constellation能源公司达成协议,从明年开始购买伊利诺伊州克林顿清洁能源中心的1100兆瓦电力。这些协议也将为Meta的"普罗米修斯"超级计算集群等项目提供动力——该集群是正在俄亥俄州新奥尔巴尼建设的10亿瓦级AI数据中心。
西雅图华盛顿大学信息学院教授奇拉格·沙阿指出,获取持续增长的电力已成为Meta等AI超大规模企业的首要任务。
生成式AI和智能体AI驱动系统需要持续增加AI芯片、数据中心、电力、冷却剂和网络基础设施等资源。
"我们的能耗正在上升,各类芯片和数据中心如雨后春笋般涌现,"沙阿表示,"我们运行的所有芯片都消耗巨量电力,而通过常规电网获取电力的传统方式已难以为继。"
由于无法从现有电网获得充足电力,Meta与谷歌、微软等AI供应商正转向与核能公司合作。谷歌已与Kairos Power和NextEra能源公司达成核能协议,微软则与Constellation能源公司合作重启宾夕法尼亚州三哩岛核反应堆——该反应堆曾在1979年发生部分堆芯熔毁事故。
能源并非AI市场唯一的紧缺资源。
许多人还担忧数据中心冷却会耗尽水资源。
"所有这些都在让地球付出巨大代价,但目前尚无机制阻止它们通过签订协议获取资源、优先保障数据中心运行,"沙阿指出。
他补充说,这些协议能让Meta展示其在清洁能源方面的进展。尽管核电发展停滞数十年,但正缓慢复苏,部分原因在于许多人认为它比煤炭、石油和天然气发电更清洁。即便如此,许多观察家仍认为核电因放射性物质储存和处理问题而存在潜在危险。
Futurum集团分析师戴维·尼科尔森表示,Meta的核电计划彰显了其影响力、雄心壮志,以及推动先进AI技术发展、塑造公众对这家社交媒体巨头认知的能力。
"他们向资本市场传递出强势地位的信号,"尼科尔森说。Meta在能源领域的激进举措也可能促使谷歌或微软等其他超大规模企业达成更多电力协议。
但他同时指出,应以审慎眼光看待Meta的核电布局,因其部分协议涉及尚未落地的项目。
例如,部分提及的项目尚未开建,且时间表跨度极大——所有钠冷反应堆单元预计到2035年左右才能完工——这可能意味着Meta的规划过于冒进。
尼科尔森认为消费者需关注这些大规模协议会否影响能源账单。
"若以固定价格与Meta这样的企业锁定协议,其他用户将不得不支付更高费用,"他说,"我们必须密切关注此事,因为大型工业企业存在强烈的违规动机。"
对于关注AI市场的企业,他建议采取谨慎态度:
"那些进行巨额押注的参与者——包括OpenAI、英伟达和超大规模云服务商——其中部分将获得空前回报,另一些则将遭遇惨败。"

英文来源:

Sponsored by Google Cloud
Choosing Your First Generative AI Use Cases
To get started with generative AI, first focus on areas that can improve human experiences with information.
The agreements could enhance Meta's public image regarding its leadership in the AI race and its ability to secure energy sources to power its AI data centers.
Meta on Friday revealed it has struck deals with three nuclear energy companies to power its data centers, as the social media giant advances its ambitions in AI, including the pursuit of superintelligence.
The agreements with nuclear energy companies Vistra, TerraPower, and Oklo, one of whose biggest investors is OpenAI CEO Sam Altman, are expected to provide the parent company of Facebook, Instagram and WhatsApp with up to 6.6 gigawatts of electricity by 2035 to power its data centers.
The deals position Meta as a significant buyer of nuclear power, surpassing Amazon, Microsoft, and Google, and adding another element to the AI race.
Data centers have become increasingly important in recent years, particularly as generative and agentic AI models require massive amounts of energy for training and inference. The new deals could bring Meta closer to its goal of building superintelligence, systems that surpass human performance in most cognitive tasks.
With Vistra, Meta signed a 20-year power purchase agreement to buy more than 2,600 megawatts from nuclear stations in Pennsylvania and Ohio run by Vistra and TerraPower. Meanwhile, Meta will help support Oklo's development of 1.2 gigawatts of nuclear energy from Oklo's campus in Pike County, Ohio. Meta will also fund the development of two of TerraPower's Natrium advanced reactor units by 2032, with rights to more energy from six future units.
The deals follow Meta's agreement with Constellation Energy last year to purchase 1,100 megawatts from the Clinton Clean Energy Center in Illinois starting next year. The deals will also help power projects such as Meta's Prometheus supercluster, a 1-gigawatt AI data center being built in New Albany, Ohio.
The need to acquire ever-increasing amounts of power has become a top priority for AI hyperscalers, including Meta, according to Chirag Shah, a professor in the Information School at the University of Washington in Seattle.
Generative and agentic AI-driven systems require continuously more resources such as AI chips, data centers, power, coolant and networking infrastructure.
"Our consumption is going up, and all these chips and data centers are just coming out left and right," Shah said. "All these chips that we are running … they consume a ton of power, and we are running out of normal means to get that power from the usual grid."
Because they can't acquire enough electricity from existing power grids, Meta and other AI vendors, such as Google and Microsoft, are partnering with nuclear companies as an alternative source of energy. Google has struck nuclear deals with nuclear energy firms Kairos Power and NextEra Energy. Microsoft has an agreement with Constellation Energy to restart the Three Mile Island nuclear reactor in Pennsylvania, which partially melted down in 1979.
Energy and power are not the only limited resources in the AI market.
Many are also concerned about depleting water sources to cool the data centers.
"All of these things are costing planet Earth a lot, but right now, nothing is preventing them from making these deals and securing things, prioritizing things for running their data centers," Shah said.
These deals allow Meta to demonstrate its progress toward a cleaner form of energy, Shah added. While nuclear power has stalled for decades, it has been slowly making a comeback, partly because many consider it cleaner than coal, oil, and natural gas plants. Even so, many observers still consider nuclear power to be potentially dangerous because of the problem of storing and disposing of radioactive materials.
However, Meta's nuclear power initiatives demonstrate the scope of its influence, ambition, and ability to drive the development of advanced AI technology and shape public opinion about the social media giant, said David Nicholson, an analyst at Futurum Group.
"They're signaling to the equity markets that they are in a position of strength here," Nicholson said. Meta's aggressiveness on the power front could also trigger other hyperscalers, such as Google or Microsoft, to make more power deals of their own.
However, Meta's move toward nuclear power should be examined with a critical eye, as it has made agreements for projects that have yet to be actualized, Nicholson said.
For example, some of the projects mentioned have yet to be built, and the timetable appears to be so long -- with a completion date of around 2035 for all the Natrium units -- that it suggests that Meta could be overextending itself.
Nicholson said consumers will have to monitor whether these large-scale deals affect their energy bills.
"If you can lock up an agreement with somebody like Meta at a certain price, everybody else is going to have to pay more," he said. “We're going to have to closely follow that because there are huge incentives for big industrial players to behave very badly."
For enterprises paying attention to the AI market, it might be best to take a cautious approach, Nicholson continued.
"The people who are making the huge bets, and that is the Open AIs, the Nvidias, the hyperscale cloud providers -- some of those bets are going to pay off like never before, and others are going to fail spectacularly," he said.

商业视角看AI

文章目录


    扫描二维码,在手机上阅读