加文·贝克如何投资人工智能，以及泡沫下一步将走向何方

加文·贝克是人工智能领域最活跃的投资人之一，但几乎没人听说过他。

Gavin Baker is one of the most prolific AI investors that almost no one has heard of.

过去二十年，他一直在投资一些如今家喻户晓的AI巨头，而那时它们还默默无闻。

He spent the last twenty years investing in some of the biggest AI companies before they became household names.

他是英伟达的早期投资者，也投资了Cerebras——这家公司最近刚上市。

He was an early backer of NVIDIA, as well as Cerebras, which IPO ed very recently.

他有一个非常明确的观点：人工智能并没有处于泡沫中。

And he has a very concrete thesis, which is AI isn't in a bubble.

恰恰相反。

It's quite the contrary.

它正处于一个超级周期。

It's in a super cycle.

他通过分析算力、晶圆和训练令牌这些AI基础设施的核心指标，发现了关键的瓶颈和限制，并提出了一个简单的核心观点。

He says that by looking at the Watts, wafers, and tokens, the infrastructure of AI, he's identified some of the key bottlenecks and constraints, and he has one simple thesis.

在人工智能领域，回报最高的领域是电力、能源和芯片制造。

The biggest returns that you can get in AI is in electricity, power, and silicon fabrication.

这跟SaaS软件服务一点关系都没有。

It's got nothing to do with SaaS software as a service.

也跟Anthropic或者OpenAI这样的聊天机器人无关。

It's got nothing to do with chatbots such as Anthropic or OpenAI.

所有这些都源自半导体，也就是构建整个AI产业的‘镐头和铁锹’。

It all filters downstream from semiconductors, the picks and shovels that build the entire AI industry.

他为此投入了41亿美元。当大多数人认为AI行业是个泡沫时，他却认为这是投资AI基础设施的世代性机会，并且他在基金里明确表达了这一观点。

And he's been expressing that interest to the tune of $4,100,000,000 While most people are calling the AI industry a complete bubble, he thinks that it is a generational buying opportunity specifically for AI infrastructure, and he makes the thesis very clearly in his fund.

如果你听到他所说的这些限制因素，关于AI基础设施的这些说法，你会发现听起来相当耳熟。

And if you hear these constraints that he's talking about, this AI infrastructure, you realize that this kind of sounds pretty familiar.

我们以前就听过这种观点，因为在这档节目里，我们多次提到过一位叫利奥波德·阿申布伦纳的投资者。

We've heard this thesis before, and that's because on the show many times we've covered a investor by the name of Leopold Aschenbrenner.

他已经活跃了三年。

He has been around for three years.

他刚发布了最新的13F文件，其中很多核心理念都和这位投资者高度一致。

He just published his most recent 13 f, and he has a lot of the same core philosophies.

但这里有个关键区别。

Now here's the difference.

Leopold 一直活跃了大概三年吧？

Leopold's been around for, what, three years?

Gavin 却已经干了二十多年了。

Gavin's been doing this for twenty plus years.

当我们拿这两个人比较时，Leopold 管理的资产规模几乎是他的三倍。

And when we're comparing these two, Leopold actually has almost three times the assets under management.

不过在节目开始前，我们的制片人 Luke 提醒我一句话。

But there's this great quote that I was reminded of from our producer Luke before the show.

这么说吧。

It's like, okay.

你或许能在一年内超越沃伦·巴菲特，但你能连续几十年都赢过他吗？

You can beat Warren Buffett over a year, but can you beat him over multiple decades?

而 Gavin Baker 的业绩记录表明，他对这个投资观点可能有着完全不同的看法。

And Gavin Baker has a track record that proves that he might have a slightly different outlook on this investment thesis.

不知道的人可能不知道，加文·贝克是这家叫交易管理公司的创始人。

For those who don't know, Gavin Baker, he is the facelift founder of this company named the trades management.

他们是一家投资基金，加文已经专注投资英伟达二十多年了。

They are an investing fund, and he has spent 20 years investing in NVIDIA.

埃贾兹，如果你二十年来一直持有英伟达的股票，他还能继续工作简直是个奇迹——这回报率也太惊人了。

Now, Ejaaz, if you've invested in NVIDIA for twenty years, it's a miracle he's still working because those are some pretty incredible returns.

我们最近看到的一些早期成功案例，比如Cerebras公司，他很早就参与投资了。

Some of the early wins that we've recently seen include companies like Cerebras, which he was a very early investor in.

Cerebras你可能还记得，上市时估值高得离谱。

Cerebras, if you'll remember, just IPO ed for an ungodly amount of money.

还有这家公司，Astera Labs，情况也一样。

Same with this company, Astera Labs.

另外还有很多你可能没听过的公司，我们这一期会一一盘点，跟着他的投资组合，看看他对人工智能行业未来走向的判断，以及他认为哪些领域最有潜力。

And there's a lot of other companies that I don't think you've probably heard of before that we're gonna cover in this episode as we walk through his portfolio and his guidance on where he sees the AI industry going in terms of investment and where all the opportunity lies.

那问题就来了，他到底在投什么？为什么投这些？

So then the question becomes, what is he investing in, and why is he investing in it?

如果我们看看他最近的13F报告，也就是他的基金名称，管理的资产规模大约有40亿美元。

Well, if we look at his recent 13, which is the name of his fund, they have about $4,000,000,000 worth of AUM.

我们来看看他的一些重仓股，分析一下这些公司到底解决了什么问题。

We look at some of his biggest positions and unpack what some of these companies actually solve.

这其实指向了加文在多个访谈中提到的瓶颈，也就是AI未来发展的关键所在。

It kinda points towards the bottleneck that Gavin references in a bunch of his different interviews as to where AI is going.

他持有不少公司的大量股份，但很多人可能没听说过，因为这些公司看起来一点都不酷。

So he has some pretty sizable positions in some companies that a lot of people may not have heard of purely because they're quite unsexy.

他在这家公司上的仓位接近9%。

So he has an almost 9% position.

他在一家叫Astera Labs的公司上持有约10%的基金仓位。

So 10% position of the fund in this thing called Astera Labs.

Astera Labs可以理解为连接各个GPU的通信层。

Now Astera Labs can be described as kind of like the connectivity layer between GPU.

想象一下数据中心，GPU就像是引擎，负责模型的预训练、后训练，还有推理任务。

So if you imagine a data center, you have GPUs, they're kind of like the engine that can like pre train, post train your model, and also like inference your model.

但为了让这些GPU正常工作，它们需要处理大量的数据传输。

But in order for these GPUs to work, they need to traverse a bunch of different data.

它们得在彼此之间来回传递数据。

They need to send data between themselves.

同时还得从存储这些数据的各个内存芯片里读取信息。

They also need to access data from all these memory chips that they store this data on.

要实现这一点，就得有个类似管道系统的东西。

Now, in order to access this, you need some kind of a plumbing system.

我这里说得比较笼统，毕竟我对这些技术细节也不精通。

And I'm being very high level on this because I'm not gonna pretend that I know the intricacies of all of this.

Astera Labs 这家公司，本质上就是解决这个问题的。

Astera Labs is a company that essentially fixes that.

随着AI集群扩展到成千上万块芯片，瓶颈不再只是GPU本身，而是数据传输的窗口——什么时候传什么数据、怎么及时访问对的数据，这正是Astera Labs所构建的管道系统要解决的核心问题。

The problem it solves is as AI clusters scale to hundreds of thousands of chips, the bottleneck stops becoming GPU specifically, but it starts becoming that transfer window and sending the right data at the right time, accessing the right data at the right time is what the plumbing system that Astera Labs builds.

在准备这期节目之前，我根本没听说过Astera Labs，但我记得另一个类似的公司叫Cerebras Labs，差不多六个月前Gavin还提到过它，考虑到AI发展的速度，这已经算挺久以前的事了。

Now, I haven't heard of Astera Labs until we started researching for this episode, but I remember another company being the exact same case called Cerebras Labs, which is what Gavin was talking about almost like, I think like six months ago, which is quite long given the relative scale or timeline of AI.

接下来我听说他们上市了，估值大概是多少来着？

And then the next thing I know is that they IPO ed for like, what was it?

大概六百亿美元，上市之后股价又涨了百分之四十。

Like $60,000,000,000 And it's just been like, it's up 40% since the IPO.

这些趋势都指向同一个方向。

So just kind of point towards these different trends.

Astera Labs 可能正是这个领域里值得关注的公司。

Astera Labs might be something significant on that horizon.

对。

Yeah.

Cerebras 是他最早期的投资之一。

Cerebras is one of his earliest investments.

他早在 Cerebras 创立初期就加入了，这意味着他早就看好这个方向。

I mean, was in Cerebras very early in the company's lifetime, which means he's bet on this thesis for years.

他还长期押注过其他几家公司。

There's also a few other ones that he's been betting on for a long time.

要说 NVIDIA，作为这个领域的标杆，二十多年一直持有并始终坚信，这确实很了不起。

I mean, NVIDIA being that flagship one, being involved in NVIDIA for twenty plus years is pretty incredible, and still having conviction all the way through is impressive.

他最近上了两个播客，我听了他谈自己对 NVIDIA 的持仓。

He would Gavin was recently on two podcasts where I was listening to him speak about his NVIDIA position.

很明显，他相信 NVIDIA 能持续保持高利润率和强劲的需求。

It's very clear that he believes that they're gonna be able to maintain these profit margins and maintain the demand as well.

这意味着他看好 NVIDIA 有望逼近一万亿美元的市值。

Which means he's putting NVIDIA on a clear path to getting close to $10,000,000,000,000 in market cap.

但现在还只走了一半的路。

It's only halfway there right now.

还有几个值得关注的公司，比如美光，我们在上一期聊过，强烈推荐你回头去看看，能帮你理清 AI 投资生态里这些公司的位置。

A few other noteworthy mentions are Micron, which we discussed on a previous episode, which I would highly recommend going to watch in terms of the AI investment stack and where all these companies lie.

美光是全球最大的内存制造商之一，关于它有个惊人的数据。

Micron is one of the largest memory makers and a crazy statistic about Micron.

一年前，它的市值还不到一千亿美元。

A year ago, it was sub $100,000,000,000.

就在这段录音时，它的市值刚刚突破了一万亿美元。

And as of recording this, it just eclipsed a trillion dollars in market cap.

一年内涨了十倍，这充分说明了内存问题有多关键。

It got a 10 x in a single year, and it's a testament to how important that memory problem is.

接下来是一些没那么出名但挺有意思的公司，Ejaaz，我想特别跟你提一个，我觉得你会喜欢——Unity Software。

Now perhaps some of the less noteworthy companies that are interesting, Ejaaz, there's one I wanna highlight for you in particular, because I feel like you're gonna like this one, is Unity Software.

如果你不太了解，作为玩家我太清楚了，Unity是一款游戏引擎。

And for those that don't know, Unity I mean, I know this very well, as a gamer, Unity is a game engine.

很多热门游戏都是用这款三维渲染软件开发的。

There's a lot of popular games that are built using this three d rendering software.

那为什么投资AI的人会关注Unity呢？它不就是做我玩的游戏的吗？

So why would someone who's investing in AI be investing in Unity Software, the thing that makes my video games?

答案就在于这个三维游戏引擎。

And the answer is the three d game engine.

Unity其实是个三维世界建模工具，它对物理规律、世界运行方式，还有材质和光影都有非常深入的理解。

Unity is a world model builder, and it has a really deep understanding of physics and the way that world works and the understanding of textures and lighting.

当这些人工智能公司试图打造通用人工智能时，他们其实是在研发类人机器人。

And when these AI companies are trying to build AGI, they're trying to build humanoid robots.

其中很重要的一部分，就是模拟虚拟环境和虚拟数据集，让这些机器人能在里面接受训练。

A big part of that is simulating virtual environments and virtual datasets that allow these robots to be trained in.

而Unity恰好就是其中最出色的平台之一。

And Unity just so happens to be one of the best ones.

所以我特别想跟你提一下Unity，Ejaaz，作为世界模型的代表——一家以游戏引擎闻名的公司，完全有可能成为人工智能领域的重要玩家。

So I feel like that one I want to highlight specifically for you, Ejaaz, as the World Model Maxi, there is a clear path in which a gaming company that is known for its gaming engine becomes a pretty serious player in the world of AI.

是的。

Yeah.

世界模型的核心理念其实很简单。

The the whole thesis behind world models is pretty simple.

目前的AI模型或者大语言模型，都是通过文字、通过书籍来理解世界的。

It's AI models or LLMs currently understand the world through text, through books.

这就像是一个学生整天泡在图书馆里，却从未真正接触过现实世界。

It's kind like a student sitting in a library, but it doesn't actually have experience of the real world.

世界模型恰恰解决了这个问题。

World models basically unlock that.

这就像是把游戏里的角色放进一个模拟环境里，让它真正理解现实世界的物理规律。

It's like putting a game character into a simulated environment and understanding the physical reality of how the world works.

如果我放下这部手机，或者踢一脚球，会发生什么？

If I if it works, if I drop this phone, if I kick a ball, like, what happens?

接下来会有什么连锁反应？

What are the next consequential steps?

你该怎么做？

What do you do?

世界模型有效弥补了这个短板，但目前真正能大规模实现这一点的公司寥寥无几。

World Models effectively fix that, and there are very few players that have, like, built this out at scale.

我觉得现在领先的应该是谷歌的Genie三号。

I think currently the leader is probably Google with Genie three.

他们最近发布了一个叫Gemini Omni的新模型，确实能在规模上做到这一点，但离理想状态还差一点。

They actually released a a new model called Gemini Omni recently, which kind of like does this at scale, but it's not quite like where it's meant to be.

它还没迎来自己的高光时刻。

It hasn't quite had its chattypety moment.

我特别欣赏盖文的一点是，他有一种两极化的策略。

What I like about Gavin in particular is he kind of has this barbell.

我不知道你有没有注意到，乔什，他挺老派的，觉得大家迟早需要显卡。

I don't know if you noticed this, Josh, where he's kind of old school and he's like, people are gonna need GPUs.

大家迟早需要内存。

People are gonna need memory.

我会投资那些最大的厂商，比如美光和英伟达。

I'm gonna invest in the biggest players, Micron and NVIDIA.

但他同时又很有前瞻性，觉得这才是未来趋势所在。

But then he has this kind of forward looking thing where he's like, I think that's where the puck is gonna go.

所以我打算投资Cerebras，因为我觉得推理能力会变得极其关键。

And so I think let me invest in Cerebras because I think inference is gonna be super important.

我还要投资Unity，因为我觉得世界模型将是训练机器人和未来大模型的未来方向。

And then let me invest in Unity because I think world models are gonna be the future of how we train robots and future LLMs.

所以他采用了一种双端策略。

So he has this kind of like barbell approach.

另外，我注意到他的投资组合里还有两家公司。

Now one thing that I see in his portfolio over here as well is this well, there's two companies actually.

其中一家叫Positron，专门做推理芯片。

It's this company called Positron, which kind of creates inference chips.

如果你觉得这跟Cerebras很像，那确实很相似。

Now if that sounds similar, it's very similar to Cerebras.

他们做的就是这一套东西。

That's exactly what they do.

这背后是盖文最近几次访谈里反复提到的一个观点：AI模型的基础设施，特别是训练环节，正从预训练转向更注重后训练的方向。

And it's around this entire thesis, which Gavin has spoken about on his recent interviews, which is the infrastructure stack, specifically the training stack for AI models, is moving from pre training to something more focused on post training.

如果你一直关注AI领域，应该也能感觉到这种转变正在发生。

Now if you've been involved in the AI sphere generally, you kind of had an idea that this shift was happening.

但盖文是彻底押注在这个方向上的。

But Gavin is all in on this thesis.

就算你有了一个模型，它仍然需要理解新出现的信息和数据。

And if you have a model, it still needs to understand new information that comes in, new data.

它得不断更新自己。

It needs to update it.

光靠当初训练时用的那一组数据，可没法让它一辈子都变天才，它还得持续学习新东西。

Just because you pre train it on a on a specific dataset doesn't mean it's gonna be a genius for the rest of its It still needs to learn new information.

这个过程发生在后训练阶段，而且非常依赖推理计算。

That happens in the post training layer and it requires a lot of inference.

再者，如果你希望AI能像人一样更深入地思考问题——比如接收到新信息时，会想：这个角度合理吗？会不会有别的理论更适用？

Secondly, if you need the AI model to actually think about the problem more, you know, in the same way that like we take in new information, we're like, I wonder if this angle makes sense or if another thesis maybe applies.

这叫推理能力。

That's known as reasoning.

而实现推理，同样需要大量计算资源。

You need a lot of inference.

目前估算，单是推理这一块的成本或商业机会，就比预训练所消耗的算力高出五到十倍。

Now the estimates are the cost or the revenue opportunity from inference alone is worth around five to 10 x more than the amount of compute that is being put into pre training.

所以，AI实验室和芯片厂商突然开始转向这个方向。

So AI labs and chip makers are suddenly making this big shift.

你已经看到英伟达推出了好几款专为推理优化的GPU，支持智能体式的应用。

You've seen NVIDIA create a bunch of different GPUs that are aligned to inference to allow agentic kind of exposure.

我们看到加文也通过他针对推理领域的多项投资，体现了这一点。

And so we're seeing Gavin express this through his different investments on inference alone.

最后我要提的一点是，加文之前和我聊过一个特别有意思的观点，就是在录音前我们还讨论过中国的情况。

And the final point that I'll make is Gavin made a really cool point, and we're talking about this before we started recording, Josh, on China specifically.

在AI竞赛上，一直有种说法是中国 vs 美国。

It's been very much like China versus The US when it comes to the AI race.

中国有个很独特的优势：能源充足，而且能大规模扩大芯片制造能力。

China has a very unique kind of setup where they have an abundance of energy and the ability to scale chip manufacturers.

这一点，目前美国还做不到。

That's something that The US currently struggles with.

所以他们大量把生产外包给台湾的台积电。

That's why they outsource a ton of stuff to the likes of TSMC on Taiwan.

他主要想表达的是，中国有机会打造一套与众不同的基础设施或芯片，因为他们的重心完全放在推理上，跟美国的方向很不一样。

And what he basically explains is that China has a unique opportunity to create infrastructure or chips specifically that are gonna look very different to what The US is creating because they're focused so much on inference.

所以可以说，盖文通过投资美国的推理基础设施，正在美国这边带头布局。

So Gavin, you could say, is leading the charge in The US through his investments on building or taking a bet on The US infrastructure setup for inference.

我觉得这未来可能会是一个巨大的机会。

And I think that could be a huge opportunity in the future.

值得注意的是，这个押注并不仅仅是为了博取上涨。

And it's worth noting that this bet also isn't only for the upside.

他们还在一个叫QQQ的ETF上持有大量的看跌头寸。

There is a large put position here in an ETF named QQQ.

对不太熟悉的人解释一下，QQQ追踪的是纳斯达克100指数。

Now for those who are not familiar, it covers the Nasdaq 100.

它是一篮子股票组成的基金。

It's a basket of stocks.

这是美国交易量第二大的ETF，表现一直非常出色。

It's the second most traded ETF in The United States, and it's been performing incredibly.

2023年涨了55%。

In 2023, it was up 55%.

2024年涨了25%。

2024, 25%.

2025年涨了20%。

2025, 20%.

而2026年到现在，已经涨了17%。

And so far in 2026, already up 17%.

所以作为一只指数基金，QQQ的表现实在太出色了。

So QQQ, as an index fund, has been doing incredibly well.

操作起来很简单。

It's easy.

它就是由前100大股票组成的一个投资组合。

It's a basket of the top 100 stocks.

加文的意思是，他在对冲这只基金。

Gavin is saying, I'm shorting against that.

我觉得这个会下跌。

I think that's gonna go down.

我觉得你弄错了。

I think you are wrong.

这说明他非常看好人工智能领域，尤其是那些在解决关键瓶颈的龙头企业。

And what this tells me is that he believes strongly in the AI the AI play in the sense that he's going to invest in the key makers who are solving these bottlenecks.

但从整体市场情绪来看，他并不特别乐观。

But as a market wide general sentiment, he doesn't appear to be very bullish.

这是一种对冲策略，用来防范下行风险——即使市场出现不利波动，虽然人工智能依然表现强劲，他还能靠QQQ来对冲。

And this is a hedge against that that downside protection, where if the market does start to fall apart in ways that are less favorable, even though AI still wins, he has that hedge with QQQ.

现在我们可以把这些关键点拆解成几个他认为在投资中最重要的瓶颈。

Now we can kind of break this down into a few bottlenecks that he believes are going to be most important when it comes to investing.

这些是他认为人工智能未来发展所必需的核心要素。

These key things that he is looking at in terms of what the world of AI is going to need as we progress forward.

真正的限制到底在哪里？

What are the actual constraints?

它们长什么样？

What do they look like?

那你怎么投资呢？

And then how do you invest?

你怎么把这些转化为实际的资金，投到公司里去赚钱？

How do you convert these into actual dollars that you could put into companies to earn you money?

一共有四个。

And there's four of them.

第一个是垂直领域的小型语言模型。

The first one is verticalized small language models.

如果你想想一般的大型语言模型，比如你用的Claude和ChatGPT，它们都是通用型的。

If you think about LLMs in general, like the chatbots that you talk to such as Claude and ChatGPT, they're generalized LLMs.

它们对世界有广泛的理解，能回答各种具体问题，但要训练一个专门针对某个领域或特定问题的模型，完全是另一回事。

So they have a wide understanding of the world in context, and they'll be able to answer like specific questions, But they it's another thing training a model around a specific vertical or a specific problem that you're trying to solve.

这些具体的问题到底出现在哪里？

Where do these specific problems exist?

主要是在那些专注于解决特定问题的企业，或者已经在某个细分领域建立起独特优势的公司里。

Well, in enterprises that are deep on solving a particular problem or companies that have made or formed a niche in that particular subsector.

而垂直化的小型语言模型，恰恰就是为了解决这个问题。

Now, verticalized small language models basically address exactly this.

它们属于前沿模型，但经过高度优化，能在企业专属数据上高效运行，或者直接在本地设备上运行。

They're frontier models, but highly optimized towards running efficiently on specific enterprise data or locally on device.

我们之前也聊过本地运行的模型，主要是因为你的手机或日常设备里存着大量高度个人化的内容，但你并不想把这些数据交出去，公司也没法轻易拿到。

Now, we have spoken about on device or locally run models before, purely for the case that there's a bunch of data in your phone or the devices that you use that are highly personable to you, but you don't necessarily want to give up and companies don't necessarily have access to that.

比如医疗记录、财务信息这些。

For example, medical records, financial details.

我看到OpenAI推出过一个金融AI助手，能访问你的银行账户，却不能真正操作，因为里面涉及太多你不想泄露的个人信息，像社保号、银行账号之类的。

I saw OpenAI release a financial AI agent that can get access to your bank account, but it can't actually act on it because there's a lot of personal identifiable information that you don't wanna share, such as your social security number, banking details, etcetera.

而本地运行的模型，或者这些小型语言模型，正好能解决这类问题。Gavin坚信，这类技术未来会大放异彩。

Now locally run models or these SLMs can solve that kind of a problem, and Gavin is making a huge bet that these are gonna become huge in the future.

我注意到，Gavin对苹果公司尤其看好。

One company that I've noted that Gavin is hugely bullish on is Apple.

虽然他并没有表现出投资兴趣，但他认为苹果会成为主要的设备制造商之一，让这些本地运行的模型能在他们的设备上跑起来。

Although he doesn't express an investment interest, he knows or thinks that Apple is gonna be the device maker, one of the major device makers that allows for these locally run models to run on their devices.

如果未来真是这样，我们就可以想象一个场景：也许你每天并不需要再和Claude这样的模型交互。

Now in a world where that is the future, you can start thinking of a world where maybe Claude doesn't need to be the model that you need to interface with every day.

你可能更需要一个基于你个人数据训练的专属AI助手，而这正是这些小型语言模型最终会演变成的样子。

Maybe you need a personalized AI agent trained on your own data, and that's what these SLMs end up eventually becoming.

这是广义上的说法，比如你可以在自己的手机上运行它；而很多企业则会部署这些高度优化的专用模型，用它们的专有数据进行训练，从而更好地推广产品或提升销售。

Now that's the generalized version of it where, you know, you can run it on your own phone, but a bunch of enterprises will run these highly optimized and specialized models to train on their proprietary data, which ends up helping them sell a product or market a product much better.

天啊。

Oh, man.

苹果在这方面真是占尽了优势。

Apple's in such a good position there.

确实太好了。

So good.

我已经等不及WWDC了。

I can't wait for WWDC.

快了。

It's coming.

对。

Yes.

再过几周就是苹果的开发者大会了，到时候他们会发布所有这些新的AI软件，展示它们如何与硬件整合。

We are just a few weeks away from Apple's developer conference where they're gonna unveil all of this new AI software that's coming, and what that looks like integrating the hardware.

这会是巨大的突破。

That's gonna be huge.

我们会全程跟进报道。

We will be covering that.

我特别期待聊这个话题。

I'm so excited to talk about that.

接下来这个关键点，是自主基础设施。

In terms of the next pillar of this, it is sovereign infrastructure.

我们一直都在说，比特的传输速度远超原子的移动速度。

We always talk about this, that the speed of bits is so much faster than the speed of atoms.

当你思考AI基础设施时，模型的质量已经呈指数级提升。

And when you think about AI infrastructure, the quality of models has gone purely exponential.

每瓦特能产生的智能、每个令牌能承载的智能，都在持续飙升。

The amount of intelligence we could generate per watt, the amount of intelligence per token, is up into the right only.

但物理部署的速度并没有以同样的速度增长，因为这本身就是个瓶颈。

What isn't up into the right at nearly the same rate is the speed of physical deployment, because that itself is the mode.

要把那些极其复杂的硬件投入实际使用，难度非常大。

It's very difficult to take hardware that is incredibly complicated.

我们说的是晶体管精度已经达到原子级别，却要在基础设施本就吃紧的世界里大规模部署。

We're talking about transistors that are are down to the atomic level of precision and deploying them at scale in a world in which our infrastructure is already suffering.

随着电动车的快速普及，电网早就承受了更大的压力。

I mean, with the acceleration of electric cars, the grids have already been feeling a little bit more of a strain.

它们几乎已经满负荷运转了。

They're kind of at max capacity.

现在又多了个能源问题。

Now they have the energy problem.

现在他们又面临芯片短缺的问题。

Now they have the chip problem.

加文坚信，基础设施的建设难度极高。

Gavin is very strongly betting on the fact that infrastructure is hard.

这事儿少则几个月，多则几年才能完成，而他押注的是那些能把周期压缩到几周的人。

It's going to take many, many days to months to years to do, and he's betting on the people that can compress that into weeks.

所以，物理部署的速度本身，就是护城河。

So the speed of the physical deployment itself is the moat.

他正在仔细筛选那些能最快完成部署的公司。

He's kind of narrowing in on who the companies are that are able to deploy as fast as possible.

一想到这个，我第一个蹦出来的就是SpaceX——他们这么快就建好了Colossus，然后租给Anthropic，以后肯定还会租给其他公司。

When I think about this, my first thought is SpaceX, and how quickly they've been able to build Colossus, then rent that out to Anthropic, and I'm sure companies in the future.

但这个基础设施板块，正是他重点关注的核心之一。

But that infrastructure pillar is one of the key ones that he's looking at.

我觉得所有人都明白。

I think everyone.

我们也看了利奥波德的投资组合。

We looked at Leopold's portfolio as well.

那是其中的核心部分。

That was a core component of that.

真的很难把东西做出来。

It's just it's really hard to build things.

谁能做出东西，谁就能卖个好价钱。

And whoever can build things, they can sell it for a lot of money.

SpaceX现在收入最高的业务，其实是他们对外出租的数据中心。

SpaceX, their largest line item now in terms of revenue is the data center that they're renting out.

跟火箭根本没关系。

It has nothing to do with rocket ships.

这足以说明这个领域有多重要。

I think that's a testament to how important this pillar is.

所以他看重的是速度，觉得这很关键，但成本也很重要。

So it's speed that he cares about, and he thinks that it's important, but it's also like the cost.

对吧？

Right?

他老提到一个指标，就是每瓦性能。

He keeps referencing this metric, which is performance per watt.

每瓦。

Per watt.

没错，每瓦。

Yeah, per watt.

他这里说的，是AI实验室这类公司越来越关注每度电能生成多少个token。

And what he's talking about here is companies are increasingly Companies being AI labs are increasingly caring about how many tokens can you generate per watt, right?

你想啊，今年光是这五家公司，就在GPU和算力上砸了几十亿甚至上万亿美元，就为了给这些系统供电，当然得追求性价比，特别是当你的规模已经大到像那些超大规模云服务商那样时。

Because if you think about spending billions and trillions of dollars, which is what like five companies are currently spending this year alone on GPUs and compute to kind of like power these things or electricity to power these things, you want to get a lot of bang for buck, especially if you're scaling to the size that most of these hyperscalers are doing.

举个例子，如果我问Claude和ChatGPT同一个问题，Claude回答我只要两分钱，ChatGPT却要一美元，那我大概率还是会选Claude，哪怕它在理论上只有ChatGPT95%的智能。

So if you think about it, if I prompt Claude and if I prompt ChatGPT and Claude gives me an answer that cost me 2¢ and ChatGPT gives me an answer that cost me $1, I'm probably gonna end up using Claude even if it's like hypothetically, like say 95% of the intelligence hypothetically that ChatGPT has.

因为关键是，你可以多问几次，最终花更少的钱也能得到同样的结果。

Because the point is you can prompt it even more and eventually get to the answer for a lot less of cost.

所以成本成了一个大问题。

So cost becomes a huge thing.

就在这一周，微软和优步宣布，他们正在大幅减少对云计算的依赖，因为他们的年度预算在短短四个月内就被严重削减了。

This week alone, Microsoft and Uber announced that they are effectively reducing their exposure to clawed code because their annual budgets kind of like got sequestered hit in about four months.

所以说，获得这种智能服务的成本真的非常重要。

So the point is like the cost of getting access to this intelligence matters a lot.

你看看加文的投资组合，比如Cerebras、Positron，还有Astera Labs，我注意到他总是在找那些非常细分的基础设施瓶颈，然后下单一注。

And you see this across Gavin's investment portfolio with Cerebras, with the likes of Positron and with Astera Labs, which basically like, what I've noticed is he identifies these really niche infrastructure bottlenecks, and he basically makes one simple bet.

他说，只要这家公司能把那个问题解决掉，让每瓦特的性能达到某个特定水平，AI实验室就很可能大量采购他们的GPU或者其他产品。

He's like, yeah, if this company solves that, the performance per watt gets to this specific level, which means that AI labs are probably gonna buy more of these GPUs or more of these companies or more of these things.

最终，这些他押注的公司就会解决掉某个关键瓶颈。

And you end up with a bottleneck being resolved by one of these different companies that he's betting on.

所以他的核心理念其实很简单，虽然执行起来非常复杂——我就专注在基础设施层面的AI瓶颈上。

So his thesis actually is quite simple, although it's quite complex, which is I'm just gonna focus on the AI bottlenecks at the infrastructure level.

只要我能找到一家公司，能把性能提升到这个程度，把每个token的成本压到这个水平，那我就相信，这家公司一定会变得很有价值，要么上市，要么被高价收购。

And if I can identify a company that can increase performance by this amount, that can make tokens this amount cheaper, then my bet is those companies are gonna be valuable and will either IPO or get acquired for a large sum of money.

如果你想跟着盖文投资，这几个关键人物你得了解。

And the thinkers to know for the section, for those looking to copy trade Gavin.

Astera Labs、Cerebras、SciFi 和 Positron，这四家公司都很重要。

Astera Labs, we have Cerebras, we have SciFi, and we have Positron.

这四家公司在这个领域里起着决定性的作用。

Those are the four companies that are really critical in this sector.

接下来是第四块，也是最后一块，涉及能源和太空。

Now the fourth and final is a combination of energy and space.

就像我们之前说的，地面电网严重限制了能源供应，想新增电力太难了。

Because, I mean, like we talked about in the previous point, the terrestrial grid very much limits energy, and it's very difficult to earn new energy.

有数据显示，高达40%的新数据中心都遭到强烈反对。

I think a statistic like 40% of new data centers have very strong petitions against them.

有人游说抵制，也有人上街抗议。

People are lobbying against them, people protesting.

大家都不希望这些数据中心建起来，阻力真的很大。

They do not want these data centers to There's a lot of kind of resistance to them.

我们解决这个问题的办法有两个方面。

And the way that we solve this is twofold.

一是直接在设备里发电。

One is creating energy out of the box.

这是一种便携式能源，可以带到数据中心，用小型装置就能供电。

It's it's kind of portable energy you could bring to these data centers, power them with a smaller phone.

像Bloom Energy这样的公司，Leopold对它们非常看好。

That's companies like Bloom Energy who Leopold is very bullish on.

但另一方面，还有轨道计算这个部分，这也是Gavin最近重点转向的方向。

But then there's also the orbital compute part of this, which is what Gavin has really shifted his focus to.

这个领域目前最大的公司，当然是SpaceX。

Now the first largest company in this sector, of course, is SpaceX.

他们是唯一能打通太空运输通道、把实际载荷送入轨道、把大量设备和数据中心部署到太空，并产生足够算力与能源的公司。

They're the only one who is capable of being the highway to space to delivering actual payload to orbit, to getting mass and data centers in large orbit that can generate enough intelligence and power that can funnel this.

我觉得这里说的‘SpaceX’，其实指的不只是SpaceX本身。

I think the SpaceX is a little bit bigger than just SpaceX.

我挺惊讶的，他明明觉得这个领域前景巨大，可投资组合里太空类股票却没怎么配置。

I was surprised to see there wasn't more allocation of space stocks in his portfolio, given the fact that he believes this is such a huge industry.

也许他们的看法是，现在还太早了，SpaceX才是打开这个行业的关键，他们现在主要在密切观察星舰第三次试飞的情况。

Perhaps their reality is that it's just too early, and that SpaceX is the linchpin to unlocking this industry, and just kind of closely evaluating Starship v three launches.

我们上周刚进行了一次星舰发射。

We had a Starship launch last week.

表现非常出色。

It performed very well.

如果没有星舰正常运作，我们就没法在太空中获取能源。

Without Starship functioning, we don't get energy in space.

也没法把服务器机架送入轨道。

We don't get racks to orbit.

这是必须的，因为需要运送的载荷实在太大了。

It is required because the amount of payload is so large.

所以毫无疑问，SpaceX是最值得关注的公司。

So I'm sure SpaceX is the one to watch.

还有很多其他相关企业会受到这波影响。

There's a lot of other second order companies that can be impacted by this.

不过我觉得我们还是得回到大家最常问的那个问题：这跟当年的互联网泡沫有什么不同？

But I think we wanna get to the end of this with with the question that a lot of people are going to be asking, which is, why is this not just the .com bubble again?

加文被问过很多次这个问题。

And Gavin was asked this question many times.

他的回答很有说服力，我倒是挺信他的。

He has some pretty strong answers, and I kind of believe him.

他提出的论点确实让人信服。

It seemed his case that he made was pretty convincing.

明白了。

Okay.

我觉得他的观点可以这样理解：

So the way I I think of his case is as follows.

2000年互联网泡沫时期，靠的是借贷驱动的。

In the .com bubble in the February, it was debt fueled.

那时候，人们借了巨额资金去支持那些未经验证的构想和产品，而这些产品根本没人关心，也没人用。

You had people borrowing a hell of a lot of money for unproven theses and products that people didn't actually care about or use.

但如果你拿这个和他所说的当前AI超级周期相比，仅OpenAI和Anthropic两家公司，今年就有望实现两千亿美元的年经常性收入。

Now if you compare this to the current AI super cycle as he describes it, just from OpenAI and Anthropic alone, they're on track to reach $200,000,000,000 of ARR this year.

仅仅是两家公司。

Just two companies.

这可不是凭空捏造的钱。

And this isn't made up money.

这些钱是通过合同实实在在签下来的，大部分还是预付款，来自大量企业客户和零售客户，预付比例高达四到六成。

This is money that they've signed through contracts that are already prepaid in large part, think, 40 to 60% from a bunch of enterprise and retail customers that are funding this.

所以，这是真金白银在流动。

So this is real money exchanging hands.

再来看GPU这类计算机设备，别光看那些模型实验室，咱们看看基础设施端，那些从英伟达采购硬件的公司。

Now, if you look at the GPU computer, so let's like not look at the Model Labs, let's look at like the infrastructure, the people who are buying the goods from NVIDIA.

谷歌、微软、亚马逊和Meta，全都是用自己的现金储备在买单。

Google, Microsoft, Amazon, and Meta are all paying from their own cash reserves.

所以他们根本就没借过一分钱。

So they also haven't borrowed any money at all.

他们只是在用自由现金流来投入。

They're just spending free cash flow on this.

亚马逊的自由现金流刚刚用完了。

Amazon just came to the end of their free cash flow.

如果他们开始借钱，那我们才真该担心了。

Now if they start borrowing money, then we can start to get worried.

但关键是，他们也没加杠杆。

But the point is they are not levered up either.

而且，这可是全球顶尖的五家公司，论市值和规模，它们无疑是世界上最聪明的企业之一。

Also, this is like five of the top companies in the entire world who are arguably like some of the smartest companies because they are where they are in terms of like market cap value and size.

回到互联网泡沫时期，当时是一大堆名不见经传的小公司，融了大笔钱，却花得毫无道理。

So the fact is like back to the Dotcom bubble, you had a bunch of like no name companies that had raised a ton of money that were spending money in ways that didn't really make sense.

而这一次，却是全球最顶尖的几家公司，在用完全不加杠杆、不靠压缩利润来支撑的资金进行投入。

In this cycle, you have like some of the smartest companies in the world spending money that isn't levered at all, that isn't kind of like compressed into like various different margins.

而且我们最近几周在之前节目中讨论过的所有季度报告都表明，利润正通过这些不同的变化不断优化，模型也在持续进步。

And all the quarterly reports that we've spoken about on previous episodes very recently over the last couple of weeks have proven that profit is optimizing through a bunch of these different movements, that models are progressing.

它们变得越来越智能。

They're getting more intelligent.

所以加文的核心观点是，这根本不是互联网泡沫，因为我们没有靠借贷融资，而且我们所讨论的瓶颈本质上是受物理资源限制的。

So the single argument that Gavin has is this isn't a .com bubble because we're not levered on money, but also because the bottlenecks themselves that we're speaking about is constrained by physical atoms.

买一堆内存芯片、买一堆GPU是一回事，但尼古达根本不可能超卖GPU。

It's one thing being like, okay, I'm gonna buy a bunch of memory chips and I'm gonna buy a bunch of GPUs, but it's not like NVIDIA can oversell GPUs.

美光也不可能超卖AI内存芯片，因为他们根本没有足够的产能来做到这一点。

It's not like Micron can oversell AI memory chips because they just don't have the chip production facilities to be able to do that.

因此他的简单逻辑是：如果市场根本无法供过于求，那就谈不上泡沫。

And so his simple argument is it's not a bubble if you can't oversupply the entire market.

我们受限于工具不足，根本没法开展这项工作，而他正是在投资这些工具。

We are constrained by the fact that we don't have enough picks and shovels to do the thing, and that's what he's investing in.

还有一个非常关键的点。

And there's one great point.

如果你把这个页面往下稍微滚一点，就能看到两千到三千亿美元这个数字。

If you scroll down just a hair on this artifact, you can see the 2 to $3,000,000,000,000 number.

Gavin认为，如果台积电能供应得上，英伟达今年和明年完全有可能卖出两千到三千亿美元的GPU。

This is something that Gavin believes is a reality that NVIDIA could sell 2 to $3,000,000,000,000 of GPUs this year and next year if only TSMC could supply them.

所以他觉得，台积电其实是判断市场是否过热的关键环节。

So he's saying that TSMC is actually one of the major linchpins of bubble territory.

我来解释一下原因，因为我觉得这个点特别有意思。

And I'll explain why for a second, because I found this really interesting.

如果台积电真要满足这些公司对芯片的庞大需求，那需要投入巨额资金。

If TSMC were able to supply the amount of demand that is required from these companies to actually provide them with that many chips, it will cost them a tremendous amount of money.

而且，Ejaaz，如果你再往上滚一点，就能看到收入和资本支出的对比数据。

And in fact, Ejaaz, if you scroll up just a little bit more, you could see the earned income to the CapEx.

那里有个柱状图。

There's like this kind of bar chart there.

是的。

Yeah.

就在这里。

Right here.

你可以看到，AI的资本支出和运营现金流相差不大。

Where you could see AI CapEx is not much different than the operating cash.

到目前为止，这些公司赚的钱还足够支撑它们的建设投入。

And so far, companies are generating enough cash to fund the build out of these things.

假如台积电明天就跑去对英伟达说：其实我们一夜之间就能把产能翻三倍。

In the case that TSMC came to NVIDIA tomorrow and said, actually, we could triple our capacity overnight.

英伟达肯定不会拒绝，立马就会砸下巨额资金去采购这些芯片。

NVIDIA wouldn't say no, and they would start spending an unbelievable amount of money to buy those chips.

其他公司也得借钱来买这些芯片，这样一来，资本支出泡沫就会迅速膨胀，和企业的实际运营现金流越拉越大。

Other companies would then have to borrow money to fund the purchase of all these chips, and therefore, we would start to see that CapEx bubble really start to grow and separate from the operating cash of these companies.

但因为整个产业链都存在供应瓶颈——内存、芯片制造、能源，尤其是台积电在芯片供应上的限制，我们根本没法这么快地扩张产能。

But because there are these supply constraints across the board, we have them in memory, we have them in chip making, we have them in energy, particularly as it relates to TSMC with the chips, we're not actually able to build out that fast.

所以，台积电实际上成了抑制这个泡沫加速膨胀的关键瓶颈。

And therefore, TSMC is creating a blocker on the rate of acceleration of this bubble.

只要台积电的产能一直受限，只要三星和其他芯片制造商短期内无法真正抢占市场份额——这几乎不可能，毕竟要花上极其漫长的时间，难度也极高——那么当前的增长速度就还算可持续。虽然感觉很快，但需求依然远远超过供给，因为我们根本来不及扩建产能。

And so long as TSMC stays limited and constrained to the amount of chips that they could produce, and so long as companies like Samsung and other chip makers don't actually overtake that market share, which seems improbable because it's gonna take an incredibly long time, and it's very difficult to do, then we're at a pretty sustainable rate of growth where it feels fast, but there's still this overwhelming amount of demand that can't be satiated because we just can't build it out fast enough.

只要这个局面维持下去，我觉得目前还算稳得住。

And so long as that dynamic stays intact, I think we're probably good for now.

不过呢，还有另一件事。

Well well, there's also this other thing.

对吧？

Right?

你可能会觉得需求会保持稳定，但实际情况根本不是这样。

Because, like, you could assume that, like, demand stays static, but that also doesn't happen.

对AI相关产品的需求正在呈指数级增长，远远超过了我们当前所有芯片的生产能力。

You have an exponentially increasing demand side for all this AI stuff, which is outpacing the production supply that we have for all of these different chips.

所以在我看来，这个观点唯一可能被推翻的情况，就是有人奇迹般地复制出ASML，然后突然冒出一大堆ASML的竞争对手。

So like the only way that I see this thesis kind of being unproven is if somehow someone miraculously recreates ASML and we suddenly have like ASML competitors all around.

顺便说一下，如果有人还不知道，ASML生产的设备每台价值四亿美元，而台积电和所有主要芯片制造商都离不开这些设备。

By the way, those of you who don't know, ASML produces these $400,000,000 machines, which basically TSMC and every major chip fab manufacturer needs.

ASML只有一支团队在挪威制造这些设备，生产周期特别长，现在订单已经排到了五年后。

And ASML is only one team in, I believe, Norway that creates these things, and it takes ages, and they are backlogged for, like, five years right now.

或者如果我们能开发出另一种不需要那么多GPU、也不那么依赖内存的大型语言模型，但目前根本看不到这种迹象。

Or if we recreate a different type of LLM that doesn't require as many GPUs or doesn't require as much memory, but we're just simply not seeing that.

今天我看到一条新闻，SK海力士——NVIDIA显卡最主要的内存供应商——出了大动静。

I saw a story break today about SK Hynix, which is the number one memory manufacturer and supplier for NVIDIA GPUs.

他们在AI内存领域是绝对的领头羊，现在谷歌和微软等公司正在争相向他们提供高达五百亿到一千亿美元的订单。

They are basically the top dog when it comes to AI memory, and they're currently courting, I think it's like $50 to a $100,000,000,000 worth of offers from the likes of Google and Microsoft.

仅仅是这两家公司，就愿意掏钱提前锁定未来三年的产能，为的是确保三年后的供应不会断货。

So two companies alone just to pay for the equipment that they need to scale up future supply over the next three years to lock in supply that they're gonna create three years from now.

可见这些大公司对内存的需求有多迫切。

This is how desperate some of these big companies are for memory.

而这还只是AI硬件中的一个细分领域。

Just and this is one subsector of the AI component.

但SK海力士的态度很明确：不接。

And SK Hynix is instead just saying, no.

我不能向你保证我们的供货情况。

I don't wanna give you guarantees of our supply.

我们干脆直接涨价。

Instead, we'll just hike up the price.

顺便说一句，他们的运营利润率高达70%，这在半导体行业简直是闻所未闻。

They have, like, 70% operating margin, by the way, which is just unheard of in semiconductor companies at all.

所以加文全力押注也就说得通了，因为这看起来不像是泡沫。

So it makes complete sense why Gavin is just going all in because it doesn't look like a bubble.

表面上看，可能确实像泡沫。

It might seem like a bubble.

市场可能会对此做出反应。

The market might react to it.

我们今天录播前刚看了下股票组合，全都在跌，但这只是短期波动，因为大方向很明确——我们迟早需要更多GPU。

We opened our stocks portfolio today before we started recording and everything was, like, down, but it's just reactionary because the directional goal of all of this is that we're just gonna need more GPUs.

我们需要更多半导体芯片，但现在的供应根本不够。

We're gonna need more semiconductor chips, and we don't have enough supply.

现在能生产这个的厂家太少了。

We don't have enough manufacturers for this.

所以总结一下，就是电力和晶圆。

So in conclusion, Watts and wafers.

就这两点。

That's it.

这是两大瓶颈。

Those are the two brick walls.

这两个因素限制了我们无法加速发展。

Those are the two limiting factors Mhmm.

正是这些限制条件，阻止我们过快扩张。

The limiting constraints that are going to prevent us from accelerating too fast.

只要电力和晶圆持续保值、需求旺盛、供应紧张，未来就大有可为。

And so long as those Watts and wafers stay valuable, stay in high demand, and stay limited in their supply, there will be good times ahead.

如果你只想快速了解加文的投资组合，我来给你念几个最大的持仓。

Now if you are looking for a TLDR on Gavin's portfolio, I will read out some of the largest holdings for you.