Down big, especially the tech-heavy Nasdaq.

跌幅較大，尤其是科技股密集的納斯達克指數。

CNN's Matt Egan is here to explain what we're seeing and why, Matt.

CNN的馬特-伊根（Matt Egan）在這裡解釋我們所看到的情況和原因，馬特。

Yeah, John, we've got a bit of a tech sell-off this morning, and it's being caused by earth-shattering developments in the AI space.

是的，約翰，今早我們看到了一些科技股的拋售，而這是由人工智能領域驚天動地的發展所引起的。

So let's take a look at this.

讓我們來看看這個。

Did you see the Dow?

你看到道指了嗎？

Down more than 140 points right now.

現在下跌了 140 多點。

The S&P solidly lower, but the real action is over here in the Nasdaq, 600 points lower.

標準普爾指數穩固走低，但納斯達克指數下跌 600 點，才是真正的動向。

Nearly 3% on track for one of its worst days in the past two years.

接近 3%，是過去兩年中最糟糕的一天。

And here's why.

原因就在這裡。

There's a Chinese startup that few people had ever heard of until the past few days, and it has emerged as a real player in the AI arms race.

在過去幾天之前，很少有人聽說過一家中國初創公司，但它已成為人工智能軍備競賽中的真正參與者。

It's called DeepSeek, and investors and I would imagine officials in Washington are stunned to learn that DeepSeek's AI model has developed technology that can actually be competitive with OpenAI and Google and XAI and all these more established players.

該公司名為DeepSeek，當投資者和我猜想華盛頓的官員們得知DeepSeek的人工智能模型所開發的技術實際上可以與OpenAI、谷歌、XAI以及所有這些更成熟的參與者競爭時，都驚呆了。

It's only been around for a bit, but DeepSeek has already vaulted to the top of the App Store on Apple as the most downloaded app, passing ChatGPT, which is pretty shocking.

DeepSeek 推出時間不長，但已躍居蘋果 App Store 下載量最高的應用程序，超過了 ChatGPT，令人震驚。

Look at that.

看看這個。

Number one veteran investor, billionaire Mark Andreessen, a legendary tech investor.

頭號資深投資者，億萬富翁馬克-安德森，一位傳奇的科技投資者。

He said that DeepSeek is, quote, one of the most amazing and impressive breakthroughs that he has ever seen.

他說，DeepSeek 是他所見過的最驚人、最令人印象深刻的突破之一。

Now, the most stunning thing here isn't necessarily that China has developed a pretty good AI app.

現在，最令人驚歎的不一定是中國開發出了一款相當不錯的人工智能應用程序。

It's how cheap it is.

這就是它的便宜之處。

DeepSeek says that their AI model only cost 5.6 million dollars.

DeepSeek 表示，他們的人工智能模型只花了 560 萬美元。

Now, we don't know that, but if that's true, that is pretty stunning.

現在，我們還不知道，但如果這是真的，那就太令人震驚了。

Anthropic, one of the leading AI companies, has said that it costs about 100 million to 1 billion dollars to develop an AI model.

Anthropic 是領先的人工智能公司之一，該公司表示，開發一個人工智能模型大約需要 1 億至 10 億美元。

We know that Mark Zuckerberg, the meta CEO, says that his company plans to spend 65 billion dollars on AI.

我們知道，Meta 首席執行官馬克-扎克伯格（Mark Zuckerberg）說，他的公司計劃在人工智能上投入 650 億美元。

So, John, look, this is all really questioning sort of the foundation here of the AI boom, which is that it requires a lot of spending, which is that the U.S. is running away with the AI arms race.

所以，約翰，你看，這都是在質疑人工智能熱潮的基礎，即人工智能需要大量的開支，美國正在進行人工智能軍備競賽。

And it's also questioning some of the big positives that had pushed markets to record highs.

同時，一些推動市場創下歷史新高的重大利好因素也受到了質疑。

A lot of people put a lot of money into AI.

很多人在人工智能上投入了大量資金。

And now they're wondering if that money is needed the way that some of these American companies have said it is.

現在他們想知道，這筆錢是否像某些美國公司所說的那樣需要。

Exactly.

沒錯。

We just don't know if the claims being made by these Chinese companies.

我們只是不知道這些中國公司的說法是否屬實。

Yeah, this is moving so fast.

是啊，進展太快了

It's stunning.

太震撼了

All right.

好的

We're going to stay on top of this, as we said, big drops, especially in the Nasdaq.

我們將繼續關注，正如我們所說的，大跌，尤其是納斯達克指數的大跌。

Speaking a lot more about DeepSeek.

進一步瞭解 DeepSeek。

It's one of these big stocks that has really changed the game, it seems, in just a couple of days.

它是這些大股票中的一隻，似乎在短短几天內就真正改變了遊戲規則。

Without warning, a big known, unknown Chinese firm rattling the U.S. market.

在毫無徵兆的情況下，一家名不見經傳的中國大公司在美國市場劍拔弩張。

Shelly Palmer is CEO of the Palmer Group and a professor of advanced media.

Shelly Palmer 是帕爾默集團的首席執行官和高級媒體學教授。

He joins me now from New York.

他現在從紐約與我連線。

Shelly, great to have you with us.

謝利，很高興你能和我們在一起。

It seems like there's a bit of a coup happening in the AI space.

人工智能領域似乎正在發生一場政變。

Tell me about DeepSeek, what you know about it.

告訴我 DeepSeek 的情況，以及你對它的瞭解。

I mean, it's a startup.

我的意思是，這是一家初創公司。

It's one year old company.

這是一家成立一年的公司。

A lot of people are downloading it.

很多人都在下載它。

And the question is, is it taking on the big guys in the United States?

問題是，它能與美國的大公司抗衡嗎？

So DeepSeek is a company out of China.

DeepSeek 是一家中國公司。

It is a startup.

這是一家初創公司。

As you said, it's about a year old.

正如你所說，它已經有大約一年的歷史了。

They have a model called DeepSeek V3.

他們有一種型號叫 DeepSeek V3。

And in the last little bit, they released R1, which is a reasoning engine.

最後，他們發佈了 R1，這是一個推理引擎。

And it's open source on an MIT license.

它採用 MIT 許可的開放源代碼。

You can download it.

您可以下載它。

The magic here is that instead of taking months to train and hundreds of millions of dollars, it took under two months to train and cost under, according to the company, under $6 million to train.

這裡的神奇之處在於，培訓不需要花費數月時間和數億美元，而只需要不到兩個月的時間，而且據公司稱，培訓費用不到 600 萬美元。

When you download it, you can post-train it.

下載後，可以進行後期培訓。

And pre-trained transformer models need to be post-trained in order to be valuable.

預訓練的變壓器模型需要經過後訓練才能發揮價值。

And the compute cost is a fraction.

而計算成本只是一小部分。

We're talking a few percentage points fraction of what it costs to use the large language models from the big hyperscalers and foundational model builders like OpenAI or Anthropic or any of the other large foundational model builders.

我們所說的只是使用大型超級計算機和基礎模型構建商（如 OpenAI 或 Anthropic 或其他大型基礎模型構建商）的大型語言模型所需的幾個百分點。

So this cost savings along with the basically lightweight of the model, meaning you can download it and run it on a laptop, has sent kind of shivers through the spines because what this really means is there's a possible future where algorithmic efficiency beats brute force computation.

是以，這種成本的節省以及模型的輕量化，意味著你可以下載它並在筆記本電腦上運行，都讓人不寒而慄，因為這真正意味著未來可能會出現算法效率戰勝蠻力計算的情況。

And the entire AI industry in the United States and around the world has been thinking about high-level computation and big data centers and a lot of energy use and chips.

美國和全世界的整個人工智能產業都在考慮高級計算和大數據中心，以及大量的能源使用和芯片。

And now they're thinking, well, maybe not.

現在他們在想，也許不會了。

So it's a really good one to look at in terms of the numbers, which you're saying and DeepSeek saying that they spent just $5.6 million on this new AI model.

是以，從你所說的數字和 DeepSeek 說他們在這個新的人工智能模型上只花了 560 萬美元來看，這確實是一個很好的數字。

Meta last week said it would spend upward of $65 billion.

Meta 上週表示，它將花費高達 650 億美元。

How come DeepSeek was able to crack this code and do it so much more efficiently and the big guys in the U.S. unable to do so?

為什麼 DeepSeek 能夠更高效地破解這個密碼，而美國的大公司卻做不到呢？

First of all, it's hard to speculate, but let's just say the following.

首先，這很難推測，但我們只能說以下幾點。

There is a cliche that I believe is a cliche for a reason.

有一句老生常談，我認為老生常談是有原因的。

There are more honors students in China than there are students in the United States.

中國的優等生人數超過了美國。

You're talking about a very deep pool of very smart people.

你說的是一個非常聰明的人才庫。

That's thing one.

這是第一件事。

Two, they've been restricted from getting all of the tools they need from the West, including specifically the kind of chips that would do the computation they need to do the advanced computation.

其次，他們一直受到限制，無法從西方國家獲得所需的全部工具，包括專門用於進行高級計算所需的芯片。

This is a mathematical solution to a hardware problem.

這是一個硬件問題的數學解決方案。

It's called algorithmic efficiency.

這就是所謂的算法效率。

They've actually written code that's more efficient than the code that's being used generally in this space throughout the world.

實際上，他們編寫的代碼比全世界在這一領域普遍使用的代碼效率更高。

So how did they do it?

他們是怎麼做到的？

It's open source.

它是開源的。

You can see their research.

你可以看看他們的研究。

They've been very good and transparent about publishing their research.

他們在公佈研究成果方面一直非常出色和透明。

They haven't exactly shown the secret sauce, but it is something that we'll be able to replicate.

他們並沒有完全展示祕訣，但這是我們能夠複製的東西。

And by the way, the markets are reacting in a knee-jerk and most people don't understand any of what I'm about to say to you.

順便說一句，市場的反應是膝跳式的，大多數人根本聽不懂我接下來要跟你們說的話。

No one in the United States, no American, no company anywhere on Earth, to be fair, is going to lay down and die because a couple of engineers figured out a better algorithm.

公平地說，美國人、美國人、地球上任何地方的任何公司，都不會因為幾個工程師想出了更好的算法就躺下等死。

Math is math and engineering is engineering and these algorithmic efficiency approaches will be figured out worldwide.

數學就是數學，工程就是工程，這些算法效率方法將在全球範圍內得到驗證。

Now, what does that mean for NVIDIA's chip prices?

現在，這對英偉達的芯片價格意味著什麼？

What does that mean for nuclear power plants for data centers?

這對數據中心核電站意味著什麼？

What does that mean for future data centers?

這對未來的數據中心意味著什麼？

The jury's still out.

陪審團還沒有定論。

By the way, no one's replicated this.

順便說一句，沒有人複製過這個。

For all we know, this is the Sputnik moment or this is the cyber Pearl Harbor where they faked this whole thing and it really wasn't done for $5.5 million in two months.

我們都知道，這是人造衛星的時刻，或者這是網絡珍珠港，他們偽造了整個事件，而這真的不是在兩個月內用 550 萬美元完成的。

We don't know.

我們不知道。

My guess is it was done for $6 million and was done in two months.

我猜是用 600 萬美元在兩個月內完成的。

That's my guess.

這是我的猜測。

And we're all going to have to rise to the challenge.

我們都必須迎接挑戰。

And by the way, I am confident that everybody will.

順便說一句，我相信每個人都會這樣做的。

Yeah, this is the thing.

是的，就是這樣。

We were talking about the race, the AI race, and I think this is pretty much indicative of the race that has truly just begun.

我們一直在談論人工智能競賽，我認為這恰恰說明這場競賽才真正開始。

As you say, you can crack this code.

正如你所說，你可以破解這個密碼。

If it's a mathematical solution, then you've got people that could work on it.

如果是數學上的解決方案，那麼你已經有了可以研究它的人。

But you touched on something in terms of the hardware and the GPUs and NVIDIA and so forth and even what it means for electricity consumption for data centers.

但你提到了硬件、GPU 和英偉達（NVIDIA）等方面的問題，甚至還提到了這對數據中心耗電量的影響。

If what we're hearing from DeepSeek is, in fact, all factual, what does it actually mean?

如果我們從 DeepSeek 所聽到的都是事實，那麼它究竟意味著什麼呢？

Does this change the calculation completely for AI?

這是否完全改變了人工智能的計算方法？

So at the moment, it is much, much less expensive to run DeepSeek R1 locally than it is to, or even in the cloud, than it is to run any of the other models.

是以，目前在在地運行 DeepSeek R1 的成本遠遠低於在雲端運行，甚至低於運行任何其他模型。

So the question is not, is this better or worse?

是以，問題不在於這是好還是壞？

The question is, going forward, is algorithmic efficiency, this ability to run smaller models more efficiently, going to be a better path forward than the brute force, massive compute models that we've used so far at OpenAI and Anthropic and at Google Gemini and at Metas Llama?

問題是，與我們迄今為止在 OpenAI、Anthropic、谷歌雙子座和 Metas Llama 所使用的蠻力、大規模計算模型相比，算法效率，即更高效地運行較小模型的能力，是否會成為一條更好的前進道路？

We don't know the answer to that.

我們不知道答案。

Right now, it points towards two futures.

現在，它指向兩種未來。

One future is algorithmic efficiency.

未來之一是算法效率。

The other is brute force compute.

另一種是蠻力計算。

There's no one in the world who can answer this question right now.

現在世界上還沒有人能回答這個問題。

It's all speculation.

這都是猜測。

Time will tell.

時間會證明一切。

And as you've just said, it's very early days, which is like the beginning of the race.

正如你剛才所說，現在還很早，就像是比賽的開始。

We're nowhere near the end of this race.

我們離比賽結束還早著呢。

We've never seen anything come this quickly.

我們從未見過這麼快的事情。

We've never seen exponential improvement that has really hit us at this speed.

我們從未見過指數級的改進能以如此快的速度給我們帶來如此大的衝擊。

So it's accelerating, and we're going to all learn together.

是以，它正在加速發展，我們將共同學習。

That's just the way it's going to go.

事情就是這樣發展的。

And the markets are going to follow along, because markets do that.

市場也會跟著走，因為市場就是這樣。