NVDA是AI与机器人基础设施的主导供应商；鉴于这一地位，当前估值显得过低，但AI支出崩塌构成重大风险。
逐段跳播
Nvidia 上涨 2.5%。然而，到目前为止，我们还没有触及 Nvidia 的 227。我以为我们会测试那个水平。我们最终收在 222 美元。我们没有达到 227。我原本希望到周五能测试那个突破。
这一点我错了。我们没有到那里。我们只到了 222。
Nvidia up 2.5%. However, so far, we haven't hit 227 on Nvidia yet. I thought we would test that level. We ended up at $222. We didn't make it to 227. I was hoping to test that breakout by Friday.
I was wrong about that. We didn't get there. We only made it to 222.
我今天早上说过，我们不太可能达到 227。所以，我试着跟进，说：“好吧，看起来它开始了。”“走弱之类的。”我们仍然在 222。但我们下调评级是对的。今天我们在说的就是观望。
I said this morning that it was unlikely we'd make it to 227. So, I tried to follow through and said, "Okay , it looks like it's starting." " Weakens or something like that." We're still at 222.
But we were right to downgrade. Wait and see is what we were saying today.
这笔钱中的绝大部分，可能大约 4000 亿美元，可能都流向了 Nvidia。Nvidia 毫无疑问是当前局势的造王者。在整个 AI 生态系统中。
the vast majority of that money, probably around $400 billion of it, probably goes to Nvidia. Nvidia is without a doubt the kingmaker of what's happening right now. In the entire AI ecosystem.
所以，一个芯片模型每秒实际处理的符号数量，在 RTX 5090 或 Blackwell 6000 上大约是每秒 36 个符号。
So, the practical number of symbols per second for a chip model would be something like 36 symbols per second on an RTX 5090 or a Blackwell 6000.
至于 Vera Rubin，它实际上是……我的意思是，这都是指一位天文学家。Vera Rubin 是一位天文学家。Vera 是 CPU，Rubin 是 GPU。所以，我们这里重点关注 Rubin。
As for Vera Rubin, which is actually... I mean, this is all a reference to an astronomer. Vera Rubin was an astronomer. Vera is the CPU, and Rubin is the GPU. So, we'll focus on Rubin here.
Rubin 每秒处理的符号数量大约是 RTX 6000 的 10 倍。这实际上意味着，即使你有更多的内存，你的符号处理能力也远远超过内存千兆字节数的差异。
Rubin processes about 10 times more symbols per second than an RTX 6000. This actually means that even though you have much more RAM, your symbol processing power far outweighs the difference in memory gigabytes.
这里那里。我的意思是，想想看。Rubin 有 288GB。简化一下计算，288 除以 96 正好等于 3 倍。所以，这个——我点这里——这个拥有 RTX 6000 编解码器的 3 倍。
而且它拥有，或者说，配备了，视频内存。它几乎是 3 倍，或者说 3.3 倍，相比 H100。但它的实际性能，鉴于其带宽高得多，实际上是 A100 的 10 倍。按每秒编解码器计算，大约是 H100 的 8 倍。
Here and there. I mean, think about it. The Rubin has 288GB. To simplify the math, 288 divided by 96 equals exactly 3 times. So, this one here— I'll click here—this one has 3 times the RTX 6000's codecs.
And it has, or rather, is equipped with, video memory. It has almost 3 times, or rather 3.3 times, compared to the H100. But its performance on a practical basis, given that its bandwidth is much higher, is actually 10 times that of the A100.
On a codec-per-second basis, it's about 8 times that of the H100.
所以，这意味着这些芯片实际上对内存的依赖正在减少。现在，我不是说高带宽内存不好。我只是说，随着这些 GPU 的带宽增加、效率提高，它们在技术上需要更少的空间。它们需要更少的内存，因为它们能如此快速地处理这么多任务。
So, what this means is that these chips are actually becoming less reliant on RAM. Now, I'm not saying that high-bandwidth memory is bad. I'm just saying that as the bandwidth of these GPUs increases and they become more efficient, they technically need less space.
They need less RAM because they can handle so many tasks so quickly.
我们确定会主导机器人领域的一个。我对此毫不怀疑，因为它们已经遍布机器人领域。而人们不相信。人们甚至不去想它，或者在我看来，他们看不到它。听起来很简单，但就是 Nvidia。因为 Nvidia 实际上就是大多数这些机器人运行的头脑和系统。想想这一点有点疯狂。
One we know for sure is going to dominate robotics. I have no doubt about it because they're already everywhere in robotics. And people don't believe it. People don't even think about it, or in my opinion, they don't see it.
It sounds simple, but it's Nvidia. Because Nvidia is literally the brain and the system that most of these robots run on. And that's kind of crazy to think about.
但让我们先了解一下 Nvidia 在幕后在做什么。Nvidia 获利，是因为这些机器人必须被训练。它们在哪里训练？在 Robin 芯片上。我们刚刚看到同一家公司 InScale，我为此做了一个单独的视频，签署了一项 35 亿美元的计算协议，最终安装 10 万个 Vera Robin GPU。
我猜他们需要更多钱，对吧？那是一款疯狂的芯片，Nvidia 从中获利，因为 InScale 想把这些芯片授权给像 Figger 这样的公司，以便它们训练机器人，或者把它们租给像 Figger 这样的公司。
这样它们就能训练自己的机器人。所以，Nvidia 在训练领域获利。
But let's understand for a moment what Nvidia is up to behind the scenes. Nvidia profits because these robots have to be trained. Where are they trained? On Robin chips. We just saw that same company, InScale, which I made a separate video about, sign a $3.5 billion computing deal that ends with the installation of 100,000 Vera Robin GPUs.
I guess they'll need more money, right? That's a crazy chip, and Nvidia is profiting from it because InScale wants to license those chips to companies like Figger so they can train or lease them to companies like Figger.
So they can train their robots. So, Nvidia profits in the training space.
但然后你必须转向设备端处理。Nvidia 也有这个；人们现在在 CUDA 上训练，在 ROS 软件 2 上运行机器人。他们在 Omniverse 平台上的人工世界中训练它们。这些机器人实际运行的芯片是 Thor 和 Oren，它们是机器人的计算机和大脑。
所以，你有边缘芯片。他们实际上拥有用于训练的 CPU 和 GPU 封装。他们拥有用于与机器人通信的 ConnectX 和 BlueField DPU。他们拥有定制的机器人软件。
他们拥有人工智能训练的 Omniverse。他们拥有用于训练的 CUDA，以及运行 Oren 和 Thor 边缘的那个。太不可思议了。
But then you have to move on to on-device processing. Nvidia has that too; people are now training on CUDA and running robots on ROS software. 2. They train them in artificial worlds on the Omniverse platform.
The chips these robots actually run on are Thor and Oren, which are the robots' computers and brains. So, you have the Edge chip. They literally have the CPU and GPU packages for training.
They have ConnectX and BlueField DPUs for communication to and from the robot. They have the custom robot software. They have the Omniverse for artificial intelligence training.
They have the CUDA they train on and the one that runs the Edge on Oren and Thor. It's incredible.
这个循环把你带回 Nvidia。这就是为什么在我看来 Nvidia 的股票这么便宜显得很疯狂。但它确实在提振所有人。
This loop brings you back to Nvidia. That's why it seems crazy to me that Nvidia's stock is so cheap. But it's really lifting everyone up.
所以，你最终资助的正是那些让这一切成为可能的 Nvidia 芯片。这不错。这相当令人印象深刻。
And so you end up funding the very Nvidia chips that make this stuff possible. It's not bad. It's pretty impressive.
最大的风险是 AI 支出会以某种方式崩溃，人们会在还没猎到那头该死的熊之前就卖掉熊皮。如果他们走进森林却找不到任何熊，就不会有资金，因此他们无法卖出熊皮。他们无法交付产品。他们无法训练机器人。
Nvidia 拿不到钱。算力会耗尽。OpenAI 和 Anthropic 得不到他们需要的算力。问题就从这里开始。Nvidia 拿不到钱。
The big risk is that AI spending will somehow collapse, and people will sell bear pelts before they even hunt the damn bear. And if they go into the forest and don't find any bears, there will be no funding, and therefore they won't be able to sell the pelts.
They won't be able to deliver their product. They won't be able to train their robots. Nvidia won't get paid. Computing power will run out. OpenAI and Anthropic won't get the computing power they need.
And that's where the problems start. Nvidia won't get paid.
观察点
这个频道对 $NVDA 的历次判断
Meet Kevin 在这只股票上共 21 条判断，这里只列相邻的几条。