Name: Nvidia 授權 Groq 技術，搶攻 AI 推理下一波浪潮！ (Nvidia licenses Groq to dominate the next wave of AI inference)
Uploaded: 2026-01-02T05:20:02.000Z
Duration: 8 min 24 s

程度副詞

Some big news in the acquisition and the Aqua hire front consolidation time, maybe, maybe they're trying to get this in before the end of the year, I don't know.

在收購和 Aqua 併購方面有一些大消息，也許，也許他們想在年底前完成，我不知道。

But Nvidia has agreed to a non exclusive deal to license Grok.

但 Nvidia 已同意一份非獨家協議，授權 Grok。

They're also hiring away some of its top leadership though GROK is going to continue to operate.

他們也挖走了部分頂尖領導層，不過 Grok 會繼續營運。

Um, and you know, like I've heard of Grok and again this is Groq, not X's Gro K.

嗯，你知道，我聽過 Grok，再次強調，這是 Groq，不是 X 的 Gro K。

But I'm not like intimately familiar with them other than kind of some of the research that I've done.

但除了我做的一些研究之外，我對他們並不熟悉。

They're known for super fast SRAM based AI inference chips.

他們以超快的 SRAM AI 推理晶片聞名。

So this is a, this deal is very largely around inference in AI, their chips and their systems.

所以這筆交易，這筆交易很大一部分是關於 AI 的推理，他們的晶片和系統。

That means low latency, high predictability for performance for chatbots and for LLMs, that sort of stuff.

這意味著低延遲、高可預測性的效能，適用於哈拉機器人和 LLMs 這類東西。

And like I said, it's all about inference.

就像我說的，這一切都關乎推理。

So it's targeting this really competitive inference market that's happening right now to kind of bring down cost, increase speed and take advantage of that.

所以它瞄準了這個目前競爭非常激烈的推理市場，目的是降低成本、提高速度並加以利用。

And I'm a bit of an aficionado of watching Nate B.

而且我有點算是 Nate B. 的粉絲。

Jones's and I've seen them, he has short versions on YouTube.

瓊斯，我看過他，他在 YouTube 上有短影片。

I subscribed and so I watched a 36 minute explanation of this deal.

我訂閱了，所以我看了一個 36 分鐘的關於這筆交易的解釋。

But Nate was very good and very clear about it and emphasizing exactly what you said.

不過 Nate 講得很好，非常清楚，而且強調了你說的重點。

This is about inference, that is to say the output of AI rather than the train of AI.

這是關於推理，也就是 AI 的輸出，而不是 AI 的訓練。

Nvidia owns the market for training chips and there's a huge shortage of what the, versus what the industry demands right now.

Nvidia 在訓練晶片市場上佔據主導地位，而且目前產業需求量很大，但供應卻嚴重短缺。

But this is on the other side and they're betting, Jensen Huang is betting that the models are trained, the models are commodified, the models are going to change.

但這是另一方面，他們正在賭，Jensen Huang 正在賭模型已經訓練好了，模型已經商品化了，模型將會改變。

是的，會有更多的訓練，是的。

But the real action is going to be in hoping that this becomes a major business with major usage by corporations and individuals.

但真正的重點將會是希望這能成為一個大型企業和個人都能廣泛使用的主要業務。

That's using the model to give you the output that you want.

那就是使用模型來給你想要的輸出。

And what Nate explained about this is that the GROK chips have a fairly small amount of memory, but

Nate 對此的解釋是，GROK 晶片擁有的記憶體相當少，但是

And so it's extremely fast access which makes it faster and cheaper to then be able to produce the inference output.

所以它能極快地存取，讓輸出推理變得更快、更便宜。

And so he was very complimentary of Jensen Wong saying this was highly Strategic that rather than saying, and

所以他對 Jensen Wong 讚譽有加，說這是極具策略性的，而不是說，

我看過這個，也有其他人說過。

Well, rather than Jensen Wong saying, well, we got to build this ourselves, he said, no, they're good, they're out there.

嗯，Jensen Wong 沒有說「好吧，我們必須自己蓋」，而是說「不，他們很棒，他們就在那裡」。

Again, as you said, the first reports were that he bought the company.

再說一次，就像你說的，最初的報導是他買下了這家公司。

And you think at $20 billion he would have bought the company, but

而你會認為，以 200 億美元的價格，他應該會買下公司，但

instead not exclusive license company stays out there.

而是非獨家的授權，公司還在外面。

I don't know what its ownership stake will be in Grok.

我不知道它在 Grok 的持股比例會是多少。

But then he got the top two executives or thereabouts from Grok over to Ross and Madra, right over to John Nvidia.

但後來他把 Grok 的兩位高層，或者差不多的人，帶到了 Ross 和 Madra 那裡，直接帶到了 John Nvidia 那裡。

And that's Jonathan Ross who's the founder of Grok, and then Sunny Madra who's the president.

那位是 Jonathan Ross，Grok 的創辦人，還有 Sunny Madra，他是總裁。

That's so interesting to me from a deal making standpoint, just this whole, this whole point we find ourselves in right now because we keep hearing about the Aqua Hire without the acquisition strategy.

從交易的角度來看，這點對我來說非常有趣，就是我們現在發現自己處於的這個情況，因為我們一直聽到關於 Aqua Hire 的收購策略。

And you know, you've got Nvidia doing it.

This isn't the first time Nvidia has done it.

You've got Microsoft, Meta, Amazon, all of them employing this strategy.

你看微軟、Meta、亞馬遜，他們都在用這個策略。

If their top founder, like their founder and their president, the people who, you know, arguably led the, the company to this major point, suddenly they're hollowed out.

如果他們的創辦人、總裁，這些帶領公司走到這個重要階段的人，突然被掏空了。

Well, but again, Nvidia didn't buy it, simply licensed technology.

嗯，不過話說回來，Nvidia 並沒有收購，只是授權了技術。

所以這是一個，這是一個 Nvidia 的水晶。

It's an if Nvidia has a stake and again that's the part I don't know if they have a stake in the equity in Grok, which I assume they do for that amount of money, it's in their interest for Grok to keep succeeding.

這是一個，如果 Nvidia 在 Grok 裡有股份，再次強調，這是我不知道的部分，他們是不是在 Grok 裡有股權，我猜想他們有，因為花了這麼多錢，他們當然希望 Grok 繼續成功。

they have interlocking corporate structures where everybody owns a piece of everybody else and they all kind of benefit if it happens.

他們有互相交織的公司結構，每個人都持有對方的一部分股份，如果事情發生，大家都能受益。

I think it starts to build a new kind of structure here where they can do that and he gets the technology, he gets the view of it, he expands his business.

我認為這開始在這裡建立一種新的結構，他們可以這樣做，他得到了技術，他得到了它的視野，他擴展了他的業務。

The other thing I've read about this is that Google with the TPUs is somewhat, I'm going to get this completely wrong and I apologize to everybody out there because I just don't know enough.

我讀到的另一件事是，Google 的 TPU 有點像，我可能會說錯，先向大家道歉，因為我真的不太瞭解。

Google Cloud and DeepMind models and Gemini were a bit more like what Grok brings.

Google Cloud 和 DeepMind 模型以及 Gemini 有點像 Grok 帶來的東西。

But Google is not selling their chips, at least not at scale to the world.

但 Google 並沒有賣他們的晶片，至少沒有大規模地賣給全世界。

Yeah, I think they will, but they're not yet.

嗯，我想他們會，但目前還沒有。

It rounds out Nvidia's offerings in its hosting in a way that becomes competitive with Google, sets them up strategically.

這讓 Nvidia 的產品線更完整，在託管方面，讓他們能與 Google 競爭，並在策略上佔據優勢。

That's it makes them more integrated and in, in that sense kind of covering all of the different aspects now covering on a deeper level the inference aspect of their business, you know, so that they can compete with, you know, AMD and I guess Cerebras is another company that, you know, competes heavily in this space.

這讓他們更整合，從某個層面來說，現在涵蓋了他們業務中所有不同的面向，特別是推理的部分，這樣他們就能與 AMD，還有 Cerebras 這類在這個領域激烈競爭的公司抗衡。

Now Nvidia is really moving to kind of control, maybe not own because they didn't buy GROK outright, but

現在 Nvidia 真的在朝著控制，或許不是擁有，因為他們沒有直接收購 Grok，而是

And as we've talked about often, everybody thinks of Nvidia as only a chip making company, but

而我們常說的，大家想到 Nvidia 就只覺得他們是做晶片的公司，但

CUDA is critical and to keep people on the CUDA platform for Nvidia is critical because that also addicts them to Nvidia chips and Nvidia hosting, whether that occurs at Nvidia or whether that occurs at AWS or anywhere else.

CUDA 非常關鍵，讓大家留在 CUDA 平臺對 Nvidia 來說至關重要，因為這也會讓他們離不開 Nvidia 的晶片和 Nvidia 的託管服務，無論是在 Nvidia 本身還是 AWS 或其他地方。

So Jensen Huang is highly strategic and figures it out here.

所以 Jensen Huang 很有策略眼光，在這裡找到了方法。

Tamura on Facebook who finds a lot of interesting papers and explains some of these deals as well.

Tamura 在 Facebook 上分享，她會找到很多有趣的論文，也會解釋一些交易。

And she puts it this way, that in the competitive landscape now.

她是這樣說的，在目前的競爭格局中。

Google owns layers but not the broad industry lock in.

Google 擁有層級，但沒有廣泛的產業鎖定。

AMD is a hardware challenger, but a weaker ecosystem.

AMD 是硬體挑戰者，但生態系較弱。

Gravity OpenAI is a model leader, but a buyer of compute and not an owner of it.

Gravity OpenAI 雖然是模型領導者，但卻是運算資源的買家，而不是擁有者。

Meanwhile, she says Nvidia plus GROK could unify deterministic real time inference, flexible training and GPU throughput DPU offloads.

同時，她說 Nvidia 加上 GROK 可以統一確定性即時推論、彈性訓練和 GPU throughput DPU 卸載。

And so it's really interesting just to understand the fuller strategic scope of these companies.

所以瞭解這些公司更全面的策略範圍，真的很有趣。

And yeah, kind of like what we were saying, dominating the platform as.

對，就像我們剛才說的，主導平臺。