It signed a deal to build seven small nuclear reactors to power its AI data centers. The first is scheduled to go online this decade, with the rest following through 2035. To talk more about this, I'm joined by an expert in everything AI. She's a senior fellow for emerging technologies at the Alliance for Securing Democracy at the German Marshall Fund in Washington DC.

它簽署了一項協議，將建造七個小型核反應堆，為其人工智能數據中心提供電力。第一個反應堆計劃在本十年內投入使用，其餘反應堆將在 2035 年之前陸續投入使用。為了進一步討論這個問題，我請來了一位人工智能方面的專家。她是位於華盛頓特區的德國馬歇爾基金會保障民主聯盟（Alliance for Securing Democracy）新興技術高級研究員。

Lindsay Gorman, welcome to the day. Thanks for having me. Is there any alternative to going nuclear if we want to fully realize AI's potential?

林賽-戈爾曼，歡迎來到今天的節目。感謝您邀請我。如果我們想充分發揮人工智能的潛力，除了核能，還有其他選擇嗎？

Well, one thing that's become increasingly clear as this AI revolution takes hold is that the power consumption needs are just astronomical. One study recently found that only AI uses could make up half a percent of the entire worldwide electricity use. So this is an enormous amount of power required. And just to give a benchmark, that's the rough equivalent that Argentina uses for its entire electricity supply. And the push towards nuclear, I think, is twofold. First, to make up for this exorbitant demand as these tech giants are building out new data centers, that's pushing the needs up to support AI. At the same time, though, there is a push towards cleaner sources of energy. And as the Biden administration and the U.S. government has put forward, nuclear is one of a suite of technologies, along with wind and solar, that the U.S. government is trying to push as part of its climate goals. And it's not just the United States, but other countries around the world as well are trying to build out this clean energy for the future. So it's not that nuclear is the only tool in the toolkit. But as we're trying to move away from fossil fuels at the same time as we have this explosion in need, that's why you're seeing this turn towards new and advanced nuclear technologies. And Google wants to have its seven reactors up and running by 2035. Won't that be too late at the current rate of development?

隨著人工智能革命的興起，有一點變得越來越清楚，那就是耗電量需求簡直就是天文數字。最近的一項研究發現，僅人工智能的用電量就能佔到全球總用電量的一半。是以，這是一個巨大的電力需求。作為一個基準，這大致相當於阿根廷的全部電力供應。我認為，推動核能發展有兩方面的原因。首先，這些科技巨頭正在建設新的數據中心，這將推高支持人工智能的需求，從而彌補這種過高的需求。與此同時，人們也在推動使用更清潔的能源。正如拜登政府和美國政府所提出的，核能是美國政府正在努力推動的一系列技術之一，此外還有風能和太陽能，作為其氣候目標的一部

Well, I think what Google is banking on is that the current rate is only going to explode, that the needs that we anticipate today are only going to compound if AI is as important and as revolutionary as it's expected to be. I don't think this is going to be an energy need for the next five years, but perhaps the next 50. Can you explain why AI sucks up so much power?

我認為，谷歌所依賴的是，目前的速度只會爆炸式增長，如果人工智能像預期的那樣重要和具有革命性，那麼我們今天預計的需求只會更加複雜。我認為這不是未來五年的能源需求，也許是未來五十年的需求。你能解釋一下為什麼人工智能會消耗這麼多電力嗎？

Yeah, it's a great question. And the long and short of it is that AI runs on data. And the way these artificial intelligence models work is that in order to present you with human-like reasoning and the responses that we're accustomed to on chat GPT, first, the model needs to ingest massive, massive quantities of data. For example, in the case of open AI and chat GPT, it's scraping much of the internet, the whole reported history of human output and written output to mimic those responses and create an AI-generated human-like reasoning.

是啊，這是個好問題。長話短說，人工智能是基於數據運行的。而這些人工智能模型的工作方式是，為了向你展示類似人類的推理和我們在哈拉 GPT 上習慣的回覆，首先，模型需要攝取大量、海量的數據。例如，就開放式人工智能和哈拉 GPT 而言，它需要從互聯網、整個人類輸出和書面輸出的報告歷史中獲取大量數據，以模仿這些迴應，並創建人工智能生成的類人推理。

Can I just interrupt? Are you saying that this happens every time we ask it a question? It goes whoosh across the whole of the internet?

我能打斷一下嗎？你是說每次我們問它問題，它都會這樣？它就會 "嗖 "地一下傳遍整個互聯網？

Not every time we ask it a question. Now, the models that you've seen released, first it was chat GPT, then there's GPT-4, and then the next generation models that come out.

我們不是每次都問這個問題。現在，你們看到的已發佈的機型，首先是哈拉式 GPT，然後是 GPT-4，再然後是下一代機型。

Those are the breakpoints of we train it on a certain dataset, and then we release the model.

這些是我們在特定數據集上進行訓練的斷點，然後我們就會發布模型。

So it's not each query that is scraping the entire internet in real time. But in order to build that model to give you those queries, it is scraping and training on all of that information. And that's processing all of that information, bringing it all together and training the model on all of it is what sucks up just so much power and so much electricity uses as these data centers that store the processors on which these models are trained need to suck up all this power.

是以，並不是每個查詢都會實時搜索整個互聯網。但為了建立模型，為你提供這些查詢，它需要對所有資訊進行搜索和訓練。處理所有這些資訊，將它們彙集在一起，並對所有這些資訊進行模型訓練，這需要消耗大量的電力，因為這些數據中心存儲了用於訓練這些模型的處理器，需要消耗所有這些電力。

I just want to, okay, we've got less than a minute. Can you see Google needing all seven nuclear reactors it's ordering? Because I wonder about them opening a secondary market as either a power supplier to other computer companies or indeed selling power to domestic consumers.

我只想說，好吧，我們還有不到一分鐘的時間。你認為谷歌需要訂購全部七個核反應堆嗎？因為我不知道他們是否會打開一個二級市場，成為其他計算機公司的電力供應商，或者向國內消費者出售電力。

Yeah, I mean, they say that AI is an electricity, AI is the new oil, and data is the new oil. And I think there's something to it now. And Google's not alone in this. OpenAI, Sam Altman has tried to move into these markets as well. So I think we're going to see a fusion of both energy for AI and then eventually also AI for energy. Fascinating. Thank you for guiding us through that. Lindsay Gorman from the Alliance for Securing Democracy. Thank you so much. Thanks.

是的，我的意思是，他們說人工智能是一種電力，人工智能是新的石油，而數據是新的石油。我覺得現在確實有道理。在這一點上，谷歌並非孤軍奮戰。OpenAI、薩姆-奧特曼（Sam Altman）也試圖進軍這些市場。是以，我認為我們將會看到人工智能與能源的融合，最終也會看到人工智能與能源的融合。令人著迷。謝謝你的指導。林賽-戈爾曼（Lindsay Gorman），來自民主安全聯盟。非常感謝。謝謝。