
“我当然没预料到,这一切会在我新书出版之际集中爆发,”他笑着说。
“I certainly didn't predict that it was all going to kind of come to a head as my book was coming out,” he says with a laugh.
虽然数据中心正在影响美国大选——它们是可能决定不到两周后关键中期选举结果的议题之一——但世界各地围绕数据中心的斗争由来已久。马克思的新书带领读者走遍科技行业争夺权力及其遭遇的反抗:从越战时期针对美国计算机的反战抗议,到爱尔兰和智利近期的抵抗运动,再到新西兰为数据中心未来提出的另一种愿景。
While data centers are playing a role in shaping US elections—they’re one of the issues that could decide key midterm races in less than two weeks—there’s a history of struggle around the world. Marx’s book takes the reader on a tour of the tech industry’s quest for power and its opposition, from antiwar protests targeting American computers during the Vietnam War, to more recent resistance in Ireland and Chile, and an alternative vision for the future of data centers in New Zealand.
本周我采访了他,聊聊他所见所闻,以及如果 AI 泡沫破裂会发生什么。本次访谈为清晰起见经过编辑和压缩。莫莉·塔夫特:您何时决定写一本关于数据中心的书?
I caught up with him this week to talk about what he’s seen, and what could happen if the AI bubble pops. This interview has been edited and condensed for clarity. MOLLY TAFT: When did you decide to write a book about data centers?
巴黎·马克思:早在 2019 年,荷兰和新加坡就已实施暂停令。但到了 2022 年,人们越来越认识到数据中心在爱尔兰的需求,尤其是对电网的影响。数据中心在那里开始消耗大量电力。
PARIS MARX: In the Netherlands and Singapore, they had moratoriums going back to 2019. But in 2022, there was a growing recognition of the demands of data centers in Ireland, on the energy grid in particular. There were huge quantities of electricity that data centers were starting to consume over there.
另一个真正引起我关注的故事是谷歌在俄勒冈州达尔斯发生的事。谷歌极力隐瞒其用水数据,甚至促使市议会试图将其用水数据保密。这种做法持续了大半年,直到社区终于拿到实际数字,发现谷歌占用了当地 29% 的用水量。
The other story that really got on my radar was what happened with Google in The Dalles, Oregon. Google was really trying to hide its water use figures, to the degree that it got the city council to try to keep its water use figures secret. They kept that going for the better part of a year before the community finally got the actual numbers and found out that Google was using 29 percent of the water in the community.
ChatGPT 于 2022 年底推出。显而易见,生成式 AI 将需要更多数据中心。这些地方已深感大规模数据中心建设的冲击,感觉这种冲击很快就会波及更多社区。
ChatGPT came out at the end of 2022. It became apparent that there's going to need to be a lot more data centers for generative AI. These places were already feeling the effects of significant data center construction, and it felt like that was going to start coming to a lot more communities very soon.
我觉得很多美国人以为,数据中心是一两年间在美国突然冒出来的,恰好赶上了成为中期选举重大议题的时机。我喜欢你这本书的一点是,它让我们看到其他国家和社区如何充当了早期的警示信号。
I feel like a lot of Americans think data centers just came out of the blue in the US over the past year or two, just in time to be this huge midterms issue. One thing I like about your book is that it shows us how other countries and communities served as early warning signs.
爱尔兰就是一个真实的例子,说明了可能发生什么。都柏林早在2021年就通过了暂停令。实际上是电网运营商说的:“嘿,数据中心太多了。我们没有足够的电网容量再接入更多了。”作为回应,业界开始行动,以确保仍能在爱尔兰建设数据中心。他们在都柏林拿不到电网接入,于是开始考虑在爱尔兰其他地区建设。他们还考虑开始接入天然气网络,以便使用发电机。最终,他们开始建设专用的天然气发电厂,为自己的数据中心供电,而不是接入电网。
Ireland was a real example of what can happen. Dublin had a moratorium passed back in 2021. It was actually the grid operator that said, “Hey, there are too many data centers. We do not have the grid capacity to connect any more of them.”In response to that, you had the industry moving to make sure it can still build data centers in Ireland. They can't get grid connections in Dublin, so they start looking at building in other parts of Ireland. They also looked at starting to tap into the gas network so they can use generators. Eventually, they start building dedicated gas power plants to power their data centers instead of connecting to the grid.
爱尔兰的电网是公共电网——只有公共事业公司才能建设电力线路。而数据中心企业实际上游说推动了立法,实质上允许私营公司建设电力电缆。这显然让很多人感到担忧,也引发了关于系统公平性的问题。
The Irish energy grid is a public grid—only the public utility can build electricity lines. And data center companies actually lobbied for legislation that essentially allows private companies to build electricity cables. That obviously concerns a lot of people and presents questions about equity in the system.
天然气方面的做法,对照美国现在发生的情况,显得极具先见之明——那里的公司正竞相为数据中心建设这些大型私营天然气发电厂。
The gas stuff looks so prescient with what’s happening in the US now, where companies are racing to build these massive private gas power plants for data centers.
(数据中心运营商)从爱尔兰环境保护局获得这些排放许可证,以便能够燃烧所有这些化石燃料。该机构当时警告说,现在要实现气候目标将会困难得多。
[Data center operators] were getting these emission licenses from the Irish Environmental Protection Agency to be able to burn all these fossil fuels. The agency was warning that climate targets are going to be a lot harder to achieve now.
当我开始在美国看到这种情况发生时,我并不感到意外。我已经看过这个故事上演过一次了。但其规模令人震惊,而且升级的速度之快也令人咋舌。
When I started to see that happen in the US, I wasn’t surprised. I’d already seen the story play out before. But the scale of it is staggering, and how quickly it’s escalated.
如果这个行业崩溃,正在建设的实体基础设施——从燃气电站到数据中心本身——会怎样?有没有历史先例?
What happens to the physical infrastructure—from the gas plants to the data centers themselves—that's getting built out if this industry falls apart? Is there historic precedent for that?
我认为我们正处于人工智能泡沫中。但泡沫破裂会是什么样子?他们计划新增所有这些数据中心以及为数据中心供电的燃气电站。但在获取涡轮机、芯片以及数据中心所需的各种部件方面都存在问题。这些设施能否在重新评估人工智能需求之前完成建设?
I think we're in an AI bubble. But what does the collapse of that bubble look like? They're planning to add all these data centers and all of these gas power plants in order to power the data centers. But there are issues with getting turbines, with getting chips, with getting different kinds of pieces of the puzzle for the data center itself. Are these even going to get built out before there’s a reassessment of the demand for AI?
我们常常谈到点网泡沫期间光纤容量的过度建设。但当时也出现了数据中心的建设——显然比我们现在所处的阶段要早得多。许多数据中心公司倒闭。像Equinix这样幸存的公司收购了大量数据中心,以备容量回升时使用。
We often talk about the excess build-out of fiber capacity back in the dotcom boom and bust. But there was also a build-out of data centers—obviously a much earlier stage than what we're at now. A lot of data center companies went under. Companies like Equinix, which did survive, bought up a ton of data centers to be ready for when that capacity returned.
但我们现在面临的问题是,如果数据中心是为人工智能用途而规划的,而我们假设芯片将在几年内被淘汰。即使数据中心已经建成,其最有价值的部分可能在需求回升时已不再有用。
But the issue we have now is that if data centers are being planned for AI use, and we’re operating under the assumption that the chips are going to have to be cycled out within a few years. Even if the data center is there, what makes it most valuable is potentially not going to be useful by the time that demand returns.
不可避免地,即使数据中心没有建成,那些燃气电站仍会存在。它们将寻找新客户来证明自己的存在价值。我确信人们建造这些电站并不是打算随后关闭它们。
Inevitably, even if the data centers don't come, then those gas plants will exist. They will be looking for new customers to justify their existence. I'm sure that people are not building them with the intention to just shut them down.
你在书的结尾举了新西兰的一个具体例子:一个社区主导的项目,利用本地云服务商托管数据,创建更准确的毛利语语音识别模型。我非常喜欢这种非掠夺性利用人工智能力量的替代愿景。你是否考虑过我们可以做些什么来推动改变,以及这种建设进程将如何进行?
You ended the book with a specific example in New Zealand of a community-led effort to create a more accurate speech-recognition model of the Māori language, using local cloud providers to host the data. I really liked this alternative vision of a non-extractive way to harness the powers of AI. Have you thought about what we could do to make some changes and how this build-out is going?
我希望看到各国政府切实采取强硬态度,对数据中心的开发实施更严格的监管。但这些监管也应与针对人工智能产品本身的监管相结合。我们曾听到过许多关于生成式人工智能将带来什么成果的承诺,但这些承诺尚未兑现。这些技术还带来了一长串我们被告知不会随之而来的危害。我们的政府似乎并不太愿意真正承担起责任。他们正在接受这种说法,即生成式人工智能仍然是萨姆·奥尔特曼在2023年初所兜售的那种样子,而不是它今天的实际模样。
I would like to see governments go really hard on actually putting firmer regulations on the development of data centers. But those should also be paired with regulations on the AI products themselves. We had a lot of promises about what generative AI was going to deliver that haven't been realized. There has also been a long list of harms that have arisen from these technologies that we were not told would come along with them. Our governments do not seem to be so willing to actually take that on. They're buying the narrative that generative AI is still what Sam Altman sold it to be in early 2023 rather than what it has become today.
我认为,我们确实需要思考如何以不同的方式发展技术,特别是在我们正投入如此多的资源来建设支持生成式人工智能的基础设施的这一时刻。我们应该追问:我们究竟需要多少计算能力?投入到这个项目中的资源到底有多少是合理的?我们究竟从中得到了什么?
I think that we do need to be thinking of how we do technology differently, especially in this moment where we're dedicating so many resources to the infrastructure to enable generative AI. We should be asking, how much computation do we really need? How many resources are actually justifiable to put into this project? What are we actually getting out of it, right?
硅谷创造的东西所带给我们的一切,真的能证明我们为实现它而共同牺牲的资源是合情合理的吗?如果我们真的进行了评估,我认为我们会非常清楚地发现,无论从财务角度来看,还是从社会和环境角度来看,我们现在所做的事情都是根本无法合理解释的。
Does what we're getting from what Silicon Valley is creating really justify the resources that we are collectively sacrificing in order to realize it? If we actually did the assessment, I think that we would very clearly find that there's no way to justify what we are doing right now—from a financial perspective, but also from a social and environmental one.
我们正在撰写的内容 说到那些气候目标,不仅爱尔兰政府在考虑它们。我的同事乔尔·哈利勒(Joel Khalil)报道了欧盟对人工智能的渴望如何与欧洲各国的气候目标发生冲突。乔尔告诉我,随着各国明年举行大选,他不会对数据中心成为欧洲的一个政治热点感到惊讶——就像它们在美国一样。他说:“欧洲数据中心建设引人入胜之处在于各种利益冲突的碰撞:国家级政策与地方优先事项相碰撞,经济增长的雄心与气候目标相冲突,而规划官僚主义又阻碍了人工智能公司迅速扩大其机群的尝试。很难看出欧洲怎么可能鱼与熊掌兼得。”
What We’re Writing Speaking of those climate targets, it’s not just the Irish government that’s thinking about them. My colleague Joel Khalil covers how the European Union’s thirst for AI is running into European countries’ climate goals. Joel tells me he wouldn’t be surprised if data centers become a political flash point in Europe, like they are in the US, as countries go to the polls next year. “The fascinating thing about the European data center build-out is the collision of competing interests: national-level policy butts up against local priorities, ambitions for economic growth stand in conflict with climate objectives, and planning bureaucracy foils attempts by AI companies to rapidly expand their fleets,” he says.
我们正在阅读的内容 《纽约时报》报道了人工智能和数据中心的繁荣如何推动了一种令人意想不到的大宗商品:灰色油漆。
“It’s hard to see how Europe could possibly have it all.”What We’re Reading The New York Times covers how the AI and data center boom is boosting a surprising commodity: gray paint.
得克萨斯州大规模的数据中心建设威胁到了该州多年来在削减空气污染方面取得的进展。《得克萨斯论坛报》有更多报道。
The massive build-out of data centers in Texas threatens years of progress in the state cutting air pollution. The Texas Tribune has more.
谷歌为在芬兰未获得必要环境许可的情况下清理了超过800英亩(约合324公顷)的森林以建设一项耗资1300万欧元的数据中心项目的一部分而道歉。《卫报》报道称,这家科技巨头在该国正面临强烈的反对,这可能会波及明年春季的选举。
Google has apologized for clearing more than 800 acres of forest in Finland without getting the necessary environmental permits to build part of a €13 million data center project. The Guardian reports that the tech giant is facing backlash in the country that could bleed into elections next spring.