字号·· | 护眼
南华早报

美国最强大的数据中心排放量堪比一个小国America’s most powerful data centres emit as much as a small nation

点「原文对照」整页切到原文,或双击某段只看那段的原文。

周一发表的一项新研究发现,美国最强大的数据中心约占该国电力消耗的2%,其碳足迹相当于一个欧洲小国——而这个数字还在增长。

A new study published Monday found that the United States’ most powerful data centres accounted for around 2 per cent of the country’s electricity consumption and a carbon footprint comparable to that of a small European country – a figure that is only growing.

这项研究发表在《自然-可持续性》杂志上,基于2024年5月至2025年4月期间运营的403个“超大规模”数据中心,据通讯作者、哈佛大学生物统计学教授弗朗西斯卡·多米尼奇称,这仅代表一个“下限估计”。

The research, published in Nature Sustainability, was based on 403 “hyperscale” data centres that operated between May 2024 and April 2025, and represents only a “floor estimate”, according to corresponding author Francesca Dominici, a professor of biostatistics at Harvard University.

目前全美有1200个此类设施在运营,还有数百个正在规划中——这意味着真实数字很可能已经高出数倍。

There are now 1,200 such facilities operating across the US, and hundreds more are planned – meaning the true figure is already likely to be several times higher.

“我非常支持AI用于科学研究,”多米尼奇说。“但我们确实需要以客观、基于科学的方式来研究这个问题,我们需要对这次扩张进行规划。”这篇论文发表之际,联合国气候事务负责人西蒙·斯蒂尔在土耳其举行的第31届联合国气候变化大会(Cop31)上强调了人工智能在应对全球变暖方面的“巨大潜力和巨大风险”。

“I am extremely pro-AI for scientific research,” Dominici said. “But we really need to study this in an objective, science-based manner, and we need to plan this expansion.”The paper comes as the UN’s climate chief Simon Stiell highlighted the “great potential and great peril” of artificial intelligence in the battle against global warming during the Cop31 talks in Turkey.

这项自愿性的“安塔利亚人工智能承诺”以主办会谈的土耳其地中海城市命名,旨在挖掘该技术在寻找气候解决方案方面的潜力,并评估其日益增长的环境风险。数据中心是容纳大量计算机的设施,已经存在了几十年。

Named after the Turkish Mediterranean city hosting the talks, the voluntary “Antalya Artificial Intelligence Pledge” seeks to tap the technology’s potential in finding climate solutions and assess its growing environmental risks. Data centres, facilities that house large numbers of computers, have been around for decades.

但直到2022年和2023年AI热潮兴起、ChatGPT和其他大型语言模型崭露头角之后,“超大规模”数据中心——规模最大、最强大的一类——的数量才出现爆炸式增长。

But it wasn’t until the AI boom of 2022 and 2023, when ChatGPT and other large language models came to the fore, that “hyperscalers” – the largest and most powerful class – exploded in number.

内华达州亨德森的一座谷歌数据中心。图片来源:TNSA Google data centre in Henderson, Nevada. Photo: TNS

它们通常占地2000平方米(21500平方英尺),电力容量达数十兆瓦。研究团队确定了675个符合其定义的站点,但由于难以获得可靠数据,不得不排除数百个。

They typically span 2,000 square metres (21,500 square feet) and have power capacities in the tens of megawatts. The team identified 675 sites that met their definition, but had to exclude hundreds because of difficulty obtaining good data.

样本的电力中位数为40兆瓦。容量以兆瓦计量,因为计算和冷却所需的电力是主要制约因素。

The midpoint electrical power in the sample was 40 megawatts. Capacity is measured in megawatts because the electricity needed for computing and cooling is the main constraint.

多米尼奇表示,这项研究的动机是自2018年以来该主题缺乏同行评审研究,尽管人工智能和气候变化已成为主要讨论话题。团队确定了为每个数据中心供电的能源结构,从而能够量化排放。

Dominici said the work was motivated by the lack of peer-reviewed study on the topic since 2018, even as AI and climate change have become major talking points. The team identified the mix of energy sources powering each data centre allowing them to quantify emissions.

他们发现,这些数据中心的电力需求约为82太瓦时,占美国总用电量的1.8%,接近印第安纳州、密歇根州和田纳西州等个别州的水平。

They found the centres’ electricity demand was roughly 82 terawatt-hours, or 1.8 per cent of total US electricity consumption, approaching the levels of individual states including Indiana, Michigan and Tennessee.

堪萨斯州莱文沃思县一块反对建设数据中心的标牌。图片来源:路透社A sign opposing a data centre in Leavenworth County, Kansas. Photo: Reuters

二氧化碳排放的中位估计值为2570万吨——相当于五六百万辆汽油车的年排放量,或与丹麦的全部排放量相当。约54%的电力来自化石燃料,25%来自可再生能源,21%来自核能。

The midpoint estimate for carbon dioxide emissions was 25.7 million tonnes – equal to the annual emissions of five or six million petrol cars, or similar to Denmark’s entire output. Some 54 per cent of the electricity was attributed to fossil fuels, 25 per cent to renewables and 21 per cent to nuclear.

曾在乔·拜登总统任内担任白宫国家气候顾问的吉娜·麦卡锡在回应这项研究时表示:“遗憾的是,尽管对气候和健康造成了深远而持久的影响,大多数超大规模数据中心仍依赖化石燃料发电,这并不令我意外。”

Gina McCarthy, who served as White House national climate adviser under former president Joe Biden, said in response to the study: “Sadly, it doesn’t surprise me that the majority of hyperscale data centres rely on fossil fuels for their power generation despite the profound and lasting climate and health impacts”.

她补充说,这项研究是“重要一步”,为社区和政策制定者提供了做出决策所需的可靠、独立数据。

She added the study was an “important step” in giving communities and policymakers the reliable, independent data they need to inform decisions.

多米尼奇还呼吁对数据中心实施更严格的报告要求,以便研究人员得出更精确的结论。

Dominici also called for stricter reporting requirements on data centres so that researchers can draw more precise conclusions.

有迹象显示,在美国中期选举后可能出现放缓。中期选举中民主党夺取了国会两院的控制权,美国数十个城市的选民批准了对数据中心的限制或彻底禁令。

There are signs of a possible slowdown after the US midterm elections, where Democrats seized control of both chambers of Congress and voters in dozens of American cities approved restrictions or outright bans on data centres.

进步派议员伯尼·桑德斯和亚历山德里娅·奥卡西奥-科尔特斯提出了一项全国性暂停令,保守派共和党参议员乔什·霍利也支持这一想法。

Progressive lawmakers Bernie Sanders and Alexandria Ocasio-Cortez have proposed a national moratorium, with conservative Republican Senator Josh Hawley endorsing the idea.

多米尼奇说:“我确实认为人工智能已经扎根,而且我确实认为人工智能应用在解决重大社会问题方面具有巨大潜力。”

“I do think that AI is here to stay, and I do think there’s enormous potential for AI applications to solve important societal problems,” said Dominici.

“我还认为,我们需要找到一条前路,以负责任的方式建设这些数据中心。”

“I also think that we need to figure out the path forward to build these data centres in a responsible manner.”

阅读媒体原文 ↗