2027年最佳AI股票指南已上线！

但现在他们又重新有了竞争力。

But now they're once again competitive.

接下来我们看下一张图表，就能看到加速是从哪里开始的。

And we can see where the acceleration begins as we go to this next chart.

看看OpenAI，他们的收入从11月的190亿增长到6月的330亿，增长率不错，但这可能不足以支撑他们需要的估值。

Looking at OpenAI, they went from 19,000,000,000 in revenue in November up to 33,000,000,000 in June, which is a good growth rate, but it's probably not going to get them the valuation they need.

不过到了8月底，年化收入似乎已经加速到了480亿。

Well, as of late August, the run rate appears to have accelerated to 48,000,000,000.

所以从380亿到480亿，增长了……

So from 38,000,000,000 to 48,000,000,000 growth of...

我觉得大概有50%，这主要得益于我之前那张幻灯片里展示的Codex的快速加速。

I think that's around 50%, and it's due to the rapid acceleration in codex in the chart that I showed on the prior slide.

再看看Anthropic。

We look at Anthropic.

Cloud Code在2026年初就获得了主流采用。

Cloud Code hit mainstream adoption at the beginning of twenty twenty six.

当时，他们的年化收入运行率是90亿美元。

At the time, their annualized revenue run rate was 9,000,000,000.

到七月底，已经涨到650亿美元了。

At the end of July, they're up to 65,000,000,000.

所以，再次强调，正是这些工具，加上先进的编程能力，才让像OpenAI和Anthropic这样的公司实现了收入的飞速增长。

So, again, having these harnesses, thanks to advanced coding capabilities, is what is allowing these companies, like OpenAI and also Anthropic, to see massive revenue acceleration.

这就是让……

It's what allowing...

这就是让公司们现在能在整个系统里构建AI的原因。

It's what's allowing companies themselves to build AI throughout all of their systems right now.

所以这是基础。

So this is at the base.

就像我说的，如果你只是去ChatGPT问问题，AI并没有根本性的改变。

Like I said, if you are someone who just goes to chat GPT and ask questions, AI hasn't fundamentally changed.

它变得更聪明了，希望如此，也更有能力了，但它并没有出现那种反映在收入增长率上的飞跃式变化。

It's gotten smarter, hopefully, and more capable, but it hasn't seen the step change that's being reflected in this revenue growth rates.

要理解为什么这些公司突然愿意花这么多钱，把整个未来都押在AI上，你得明白这个代理式变革，这就是2026年的主线故事。

And to understand why all these companies are suddenly willing to spend so much and bet their entire future on AI, you have to understand this agentic step change that has been the story of twenty twenty six.

好的。

Okay.

所以，我希望我已经把2026年讲得很清楚了。

So I hope I've covered 2026 well.

接下来我们要看的是未来会发生什么，2027年及以后的情况。

What we're gonna look at now is what's going on into the future, what's going on in 2027 and beyond.

而且我觉得会有一些非常关键的行业，会定义AI的下一个阶段，现在屏幕上就显示了几个。

And there's gonna be some really key industries that I think are gonna define the next phase of AI, and we've got a couple of them on the screen right now.

首先，我刚才讲的2026年所有事情，最终的结果是什么呢？

So first off, everything I just talked about in 2026, what has been the, you know, end outcome from that?

你看，SanDisk，是市场上的领先股票。

Well, you see, SanDisk, leading stock in the market.

Moderna，你可以说它跟AI关系更间接一点，但它确实是一家利用AI进行个性化医疗的公司，是第二好的股票。

Moderna, you could say it's a little more AI adjacent, but it's definitely a company uses personalized medicine, thanks to AI, is the second best stock.

我们看2026年至今表现最好的10只股票，其中9只都是因为AI涨起来的。

When we look at the top 10 stocks in 2026 so far, nine out of 10 are due to AI.

这足以说明今年AI的影响力有多大。

This is this is how powerful AI has been this year.

而且在那段初期，大概三月份的时候……

And a lot of in that initial period in kind of March...

从三月末到六月，我们看到了很多这样的涨幅，因为市场当时正在消化智能体AI带来的这种巨大变化。

The end of March through June where we saw a lot of these gains, as, again, the market was ingesting the step change that was happening thanks to agent AI.

所以问题来了，还有哪些重大变化是市场还没真正理解的？

So the question is what what step changes are still misunderstood by the market?

2027年还会发生什么？

What what is still going to happen in 2027?

我觉得今天我们会聊几个领域。

I think a few areas we're gonna cover today.

主权AI。

Sovereign AI.

基本上，企业将如何能够创建……

Basically, how enterprises are going to be able to create...

他们想如何使用所有这些模型，以及如何保护他们最有价值的数据。

How they want to use all of these models and how they protect their most valuable data.

AI网络。

AI networking.

我在开头谈了很多关于内存的内容，因为理解为什么内存带来了像SanDisk这样的回报真的很重要，截至8月底，其年初至今涨幅达到529%。

I talked a lot about memory at the beginning because it's really important to understand why memory led to returns like SanDisk, 529% year to date as of late August.

美光，上涨了219%。

Micron, up 219.

希捷、西部数据。

Seagate, Western Digital.

内存一直是重头戏。

Memory's been the story.

内存和存储一直是2026年的主线。

Memory and storage has been the story of 2026.

网络越来越成为缓解内存瓶颈的一种方式，而AI网络的增长率依然非常惊人。

Networking is increasingly going to be a way to reduce the memory bottleneck, and the growth rates of AI networking remain outstanding.

所以我们要聊聊网络方面的一些关键主题，还要聊一只股票——Semtech，它仍然是我最喜欢的参与这个趋势的方式之一。

So we're gonna talk about some of the key themes across networking, and we're gonna talk about a stock semtech that remains one of my favorite ways to play this trend.

我们要聊聊表后电力。

We're gonna talk about behind the meter power.

我们都知道电力瓶颈的问题，建设所有这些数据中心会带来巨大的电力需求。

We all know about the power bottleneck that building out all these data centers is gonna create enormous power demands.

如果你依赖电网，第一，电网的建设速度根本跟不上。

And if you're going through the grid, number one, grid power can't be built fast enough.

第二，即使一切都通过电网供电，也会对消费者电价产生下游影响，这真的很难承受，尤其是在美国这个中期选举年。

And number two, even if everything was going through the grid, you're gonna have these downstream impacts on consumer pricing that is gonna be really untenable, especially as a, you know, midterm political year in The United States.

要让数据中心上线，解决方案就是在现场发电，而不一定依赖电网作为主要电源。

The solution to be able to get data centers online is going to be able to generate power on-site without necessarily relying on the grid as a primary source.

所以我们要聊聊一些可能从这个趋势中受益的公司。

So we're gonna talk about some companies that could benefit from this trend.

我们还会聊聊数据中心内部的电力输送瓶颈。

We're also gonna talk about power delivery bottlenecks inside data centers.

英伟达正在大规模转型，改变数据中心内每台服务器的供电方式。

NVIDIA is making a massive transition into how power is delivered to each one of the servers inside data centers.

我们会讨论一个关于Vicor的股票想法，它可能会成为大赢家。

We'll talk about a stock idea in Vicor that might be a huge winner here.

最后，我们会聊聊那些制造机器的公司，也就是制造芯片机器的公司。

And finally, we're gonna talk about the companies making the machines to actually make the machines, the machines to make chips.

这叫做半导体设备，过去十五年来，它一直是投资中最赚钱的领域之一。

It's called semiconductor equipment, and it's been one of the most lucrative spaces in investing across the past fifteen years.

但很多人回避它，因为它确实很复杂。

A But lot of people avoid it because it is really complicated.

我们来分解一下半导体设备中的一些领域，这些领域可能在2027年带来非常可观的回报，因为届时会有更多供应涌入，试图赶上对各种AI芯片的需求。

Workout break down some some areas in semiconductor equipment that could lead to really phenomenal returns in 2027 as a lot more supply moves in to try and catch up with this demand for basically AI chips of all kinds.

那我们就开始吧。

So let's get started with this.

首先，我想谈谈主权AI。

First off, I wanna talk about sovereign AI.

你在这里看到的是，虽然AI竞赛追求最智能的模型，但最智能、最强大的模型使用率并不总是很高。

What what you can see here is, well, the race for AI is the smartest possible model, the usage of the smartest and most capable models isn't always high.

我们看底部的Fable五。

We see Fable five right here on the bottom.

它的使用率非常低。

It has very low usage.

这是Anthropic目前最强大的模型，但它的使用率仍然很低，因为相对昂贵。

This is the most capable model at the current time from Anthropic, but its its usage remains low because it's relatively expensive.

这里的情况是，公司们将不得不创建自己的混合方案……

And and what's going on here is companies are going to have to create their own mix of...

他们越来越多地在所有工作流程中使用AI。

They're increasingly using AI across all their workflows.

他们想依赖单一公司吗？

Do they wanna rely on a single company?

他们想要开源吗？

Do they want to have open source?

他们想怎么处理自己的数据需求？

How do they want to handle their own data needs?

这导致了开源模型的巨大增长……

And this is leading to a massive surge in open source...

应该说，是开放权重模型。

Open weight, I should say, models.

截至7月1日，开放权重模型的市场份额仅为27%，这是来自一家名为Vercel的公司的数据。

Open weight models as of July 1 were just 27% market share of tokens, and this is data from a company named Vercel.

最近，在8月25日，这个数字达到了55%，使用量翻了一番。

And recently, in August 25, that number hit 55%, so a double in usage.

这是一个非常重要的趋势，因为在这个领域占据优势的公司开始看到他们的股价飙升。

And this is a really important trend because the companies that have really advantaged positions in this space are really starting to see their share prices soar.

所以，Snowflake是我们之前在AI投资者播客中推荐的一家公司。

So Snowflake is a recommendation that we've had in the AI investor podcast.

就是……

It it...

他们最近发布了财报。

They recently report earnings.

我记得第二天股价涨了差不多25%。

I believe that they were up almost 25% the next day.

原因是，他们将成为那些想保护自己数据、又能整合各家模型的公司的分析平台。

And the reason is because they're going to be an analytics platform that companies that want to protect their own data and be able to mix in models across all these companies are going to look to.

市场开始意识到这些顶尖公司的重要性。

The best of breed companies, it's starting to dawn on the market.

这些公司不一定会被颠覆，反而会因为AI的应用而看到更大的机会。

But these companies aren't necessarily going to be disrupted, but they're actually going to see their opportunities really expand from the usage of AI.

所以，这个领域的领先公司最近回报都很不错，这也让投资者在AI领域除了硬件公司之外，还能分散投资到同样在增长的软件公司。

So, again, the leading companies in this space have seen excellent returns recently, and this allows people to diversify from, if you're looking at AI, just hardware companies to actually software companies that are growing as well.

Snowflake是我们推荐过的，表现非常出色。

Snowflake's a recommendation we've had that's done exceptional.

Cloudflare 是我们推荐的另一家公司，它继续展现出，基本上，作为智能体时代收费站的未来潜力。

Cloudflare is another company we recommend that continues to see, basically, an increasing future of how they can be a tollbooth for the agentic era.

这两家公司都相当贵。

Both those companies are pretty expensive.

我们最近推荐的一家公司是 ServiceNow，这家公司股价大幅下跌。

One recommendation we recently made is ServiceNow, and this is a company that sold off dramatically.

我相信在它们的峰值时，它们下跌了超过60%。

I believe at their peak, they're down more than 60%.

但这家公司将负责智能体时代的编排层，目前它们的AI收入不到公司总收入的10%，但它们预计到本十年末这一比例将增长到30%以上。

But the company is going to handle the orchestration layer for this agentic era, and their AI revenue today is less than 10% of company revenues, but they expect that to grow to more than 30 by the end of the decade.

这家公司定位非常精准。

This is just a company that is extremely well positioned.

相对于该领域的其他一些投资标的，它相对便宜。

It's relatively cheap relative to some of the other plays in the space.

这向你展示了在过去，你知道，我们看待2024年、2025年AI趋势的方式。

And it shows you how in the past, you know, we look at the the trends for AI in 2024, 2025.

其实都是这些名字，英伟达、AMD，还有那些……

It was really all these names, NVIDIA, AMD, all the...

还有美光这些，所有做硬件的公司。

All these Micron, all these companies making the hardware.

软件领域也在兴起一股新的投资热潮。

There is a burgeoning trade that's opening up amongst software names as well.

很多投资者过去几年在这些公司上吃了不少亏，可能不太理解现在正在打开的机会，但我相信这个机会会持续下去。

And a lot of investors who saw these companies just get burned for years on end don't necessarily understand the opportunity that's opening up, and that's one that I believe is going to continue.

随着公司越来越把AI作为运营的核心。

As companies increasingly put AI at the center of their operations.

他们需要软件编排层，这些名字你绝对值得考虑。

They're going to need the software orchestration layer, and these are some names that you should definitely consider.

接下来我们聊聊AI网络。

Now let's move across to AI networking.

我认为这也是2027年的一个关键趋势。

This is another key trend, I believe, for 2027.

这里的增长率真是惊人。

The growth rates here are astounding.

我是说，我们可以看看。

I mean, we can look.

2023年，也就是AI真正起飞的那一年，AI网络市场规模大约是100亿。

AI networking in 2023, which was the year that AI really kicked off, was about 10...

一个100亿美元的市场。

A 10,000,000,000 market.

预计到本十年末，这个数字将增长到2450亿美元。

That's expected to grow to 245,000,000,000 by the end of the decade.

从2022年，也就是ChatGPT那一年，到2030年，这个市场的复合年增长率接近50%。

Between 2022 when, again, the ChatGPT year to 2020 to 2030, this is a market growing at nearly 50% compounded annual growth rate.

正如我之前提到的，这确实是一个能解决这个计算……

And as I've mentioned earlier, it's really a place that allows solving for this com...

我应该说，某种程度上的内存瓶颈的地方。

I should say this memory bottleneck in a way.

所以你看……

So you...

AI网络涉及三个方面。

You've got three things with AI networking.

第一，纵向扩展。

Number one, scale up.

也就是说，一个服务器机架里能放多少块GPU，让它们像一个整体、一个芯片那样运作，同时共享内存等资源。

And this is how many GPUs can be put in a server rack to effectively act as a single domain, a single chip while sharing resources like memory.

对英伟达来说，他们正努力把每个机架的GPU数量从72块提升到576块。

For NVIDIA, they're trying to push from 72 GPUs in a rack to 576.

这听起来可能没什么大不了的，但实际上影响很大。

This this might not sound like it has massive implications, but it does.

因为，再说一遍，能让越多的GPU像一个芯片那样协同工作，就能共享越多的内存。

Because, again, the more GPUs you can get to act effectively as a single chip, the more you can share memory.

这样一来，你就会减少在内存上的投入，而把更多资金用在AI网络上，让所有这些芯片作为一个整体协调运作。

And what you're going to do is you're going to try spending less on memory and more on AI networking to get all these chips to act as one coherent unit.

然后是横向扩展，也就是在一个数据中心里连接十万甚至更多的芯片。

Then you're gonna have scale out, and that's where you're going to have a 100,000, even more chips that are connected within a single data center.

然后是跨区域扩展，也就是把不同地区的数据中心连接起来，继续推进这个……

And then you're gonna have scale across, And that's where you're gonna be actually linking data centers across a region to to basically continue pushing this...

芯片能力的极限，你会……

The limits of what you can do with chips where you'll...

单个数据中心可能只能承受这么多电力。

You you can maybe only have so much power as single data center site.

但如果在一个50英里的范围内有10个数据中心，你其实可以把它们全部连接起来。

But maybe if you've got 10 data centers within a 50 mile region, you can actually link them all together.

而且，这又不断给AI网络带来更多需求。

And, again, this all just continues putting more demands on AI networking.

所以我想重点介绍的股票是Semtech，如果长期收听AI投资者播客的朋友会知道Semtech，因为我在2027年已经推荐过两次了。

So the stock I wanna feature for this is Semtech, which if anyone is a longtime listener of the AI investor podcast, they'll know Semtech because I've now recommended it twice in 2027.

但这里的关键卖点是这个。

But but here's the key pitch here.

Semtech有几个大的趋势。

Semtech has a couple major trends.

第一，AI网络的复杂性正在大幅增加。

Number one, the complexity of AI networking is increasing substantially.

这其实很像当年iPhone和智能手机发布时的情况，那是上一个主导性的技术趋势。

This is really akin to what happened with the release of the iPhone itself and smartphones, the last dominant technology trend.

当时的情况是，第一代iPhone在2007年推出。

What you had was you had the first iPhone come out in 2007 at the time.

我甚至不确定3G网络当时是否已经普及。

I don't even know if three g was necessarily established network.

我觉得第一代iPhone甚至都不支持3G。

I I don't even think the first iPhone had three g.

但在后续的每一代产品中，我们不断推动智能手机的网络技术从3G到4G再到5G。

But what we did for successive generations was we kept pushing the networking technology for smartphones from three g to four g to five g.

而由此产生的结果是，智能手机时代最成功的公司并不是苹果，这可能会让人感到意外。

And what that created was the most successful company in the smartphone era wasn't Apple, which might be shocking.

苹果是最占主导地位的，现在他们是一家市值四万亿美元的公司。

Apple was the most dominant, and they are now a $4,000,000,000,000 company.

但如果你想赚最多的钱，你应该买的第一只股票是一家叫Avago的公司。

But the number one stock you should have bought if you wanna make the most money was a company named Avago.

而Avago做的事情……

And what Avago does...

过去和现在都在做的，实际上他们现在就是博通，就是随着网络技术从3G到4G再到5G变得越来越复杂，他们制造了那些滤波器。

Did and still does today, they are now Broadcom actually, was they built the filters that as networking technology, three g to four g to five g, kept getting more complex.

这需要越来越多Avago正在制造的射频芯片内容。

It required more and more of this radio frequency chip content that Avago was building.

那么，Semtech在这里也有类似的机会，因为我将在屏幕上放一张图表，你们会看到。

Well, Semtech has a similar opportunity here because I'm going to put a chart on screen that you're going to see.

我们将比过去更快地经历网络代际的更迭。

We're about to move through networking generations much faster than we have in the past.

我们将在2026年和2027年开始看到800G的速度提升。

We're going to start seeing speeds of 800 g rising in 2026 and 2027.

但到了2020年，这将被1.6太比特取代。

But in 2020, that's going to give way for 1.6 terabit.

所以我们看到数据中心网络技术的速度持续提升，并且以更快的速度变化，到2030年，我们将看到下一代网络速度技术——3.2太比特——的增长率爆发。

So we're seeing the speeds for networking technology in the data center continuing to increase and continuing to shift at a much higher rate to where by 2030, we're going to see an explosion in the growth rates for the next networking speed technology, which is 3.2 terabits.

这为什么重要呢？

And why is this important?

嗯，Semtech认为，每个光收发器，也就是这些网络连接，他们的机会将从现在的水平增长十倍，直到我们达到3.2太比特这一代。

Well, Semtech believes that their opportunity per each of these optical transceivers, these these networking connections, is going to grow from tenfold from where it is today up to when we get to this 3.2 terabit generation.

实际上，他们有点像当年Avago在iPhone时代移动网络中的角色，只不过现在是针对今天的人工智能网络机会。

Effectively, they serve as somewhat of an analog to what happened to Avago back in the iPhone generation with mobile networks for today's AI networking opportunity.

所以随着我们在数据中心看到更多高速光模块，这将需要Semtech带来更多收入。

So as we continue seeing more optics in the data center at higher speeds, it's going to require more revenue from Semtech itself.

其次，他们有一项技术叫有源铜缆。

Second, they have a technology called active copper cables.

正如我在这里写的，整个市场预计将以98%的复合年增长率增长...

As I put here, this whole market's expected to grow at 98% compound...

2026年到2030年间的复合年增长率，这非常惊人。

Compounded annual growth rate between 2026 and 2030, which is phenomenal.

你几乎看不到哪个市场能有这么高的增长率，但他们和谷歌在供应这些有源铜缆方面有着特别紧密的合作关系。

You almost never see a market growing at this growth rate, but they have a specifically attractive relationship with Google supplying these active copper cables.

随着谷歌扩大其TPU项目，也就是他们的定制芯片，尤其是在接下来一年里和联发科这样的合作伙伴一起，这一点会变得非常重要。

And this is going to really matter as Google scales up its TPU program, its its custom chips, especially with a partner named MediaTek across the next year.

所以你会看到，在2026年末到2027年，Semtech会因为有源铜缆这个业务而实现巨大的增长。

So what you're going to see for Semtech is across 2026, the tail end of it, and 2027, massive growth rates from this active copper cable opportunity.

之后，在2028年、2029年、2030年，你会看到随着网络速度的提升，机会会大量涌现。

And then after that, in 2028, twenty twenty nine, twenty thirty, you're going to see a massive boom in opportunities from these networking speeds increasing.

所以这家公司……

So this is a company...

大多数人不知道Semtech是做什么的。

Most people don't know what Semtech is.

这可以说是现存最不起眼、最冷门的技术之一了。

This is this is one of the most boring backwater technologies has existed.

但你知道还有什么曾经是无聊的冷门技术吗？

But you know what also used to be a boring backwater technology?

内存。

Memory.

你知道现在每个投资者都在追什么吗？

You know what every single investor is chasing today?

内存带来的回报，以及它凭借独特的技术地位所获得的这些回报。

The returns from memory and the unique technology position it employed that allowed it to receive these returns.

这就是为什么我认为Semtech在未来五到七年内会处于有利位置。

And that's why I think Semtech is positioned across in the next five to seven years.

即使AI的增长速度开始放缓，Semtech自身及其产品的增长速度也会显著高于整个AI市场。

Even if we see growth rates for AI beginning to moderate, the growth rates for Semtech itself and the products it provides is going to be significantly above the broader AI market itself.

这就是我们这里三个股票推荐中的第一个。

So that's that's the first of three different stock recommendations we have here.

另一个趋势，这是我们2027年的第三个趋势，是表后市场。

Another trend, this is our third trend for 2027, is behind the meter.

这里……

Here...

这里有个大新闻，就在德克萨斯州。

Here's here's the whopper right here in Texas.

他们有466.5吉瓦的电网接入申请。

They have 466.5 gigawatts of grid connection requests.

其中90%是数据中心。

90% of those are for data centers.

他们批准了8.9吉瓦，但目前只有3.9吉瓦在运行。

They've approved 8.9 gigawatts, and they've only, have 3.9 operating today.

这就是底线，显示了在德克萨斯州接入电网的竞争有多疯狂。

Here's here's the bottom line, which shows just how crazy the drive is to be able to connect to the grid in Texas.

向德克萨斯州的ERCOT申请接入的容量达到466.5吉瓦，这……

The request to connect to ERCOT in Texas at 466.5 gigawatts, that is...

我脑子里正在算这些数字。

I'm doing the numbers in my head here.

这比纽约市消耗的电力多36倍。

It is 36 times more than the power that's consumed by the city of New York.

是东京地区的8倍。

It's eight times the Tokyo region.

而且，令人惊叹的是——

And and here's what's amazing.

想想旧金山，所有那些AI公司总部所在的城市。

You think about San Francisco, the city where all these AI companies are headquartered.

德州新电网连接的申请量大约是旧金山本身年用电量的500倍。

The request into Texas for new grid connections are about 500 times the annual power usage from San Francisco itself.

我……

I...

我过去在这个播客，也就是AI投资者播客上说过，Eden，全球数据中心设备供应商，他们表示美国的积压订单有307吉瓦。

I've said in the past on this podcast, on the AI investor podcast, I should say, Eden, which is a global supplier of data center equipment, they've said that The US backlog's 307 gigawatts.

按照2025年的建设速度，这相当于15年的积压量。

At twenty twenty five build rates, that's a backlog of fifteen years.

所以这里……

So here...

问题就在这里。

Here's the problem.

我们打算把这个图表放在这里，虽然它有点断断续续的。

We're gonna we're gonna put it, I guess, this chart right here where it's kind of broken up.

但你看屏幕底部，自2000年以来，美国的发电量基本持平，现在需要大幅增长才能满足这些电力需求。

But you see at the bottom of the screen, US power generation is roughly flat since the year 2000, And now it needs to grow significantly to be able to handle these these these power needs.

那么，这里的策略是什么？

So what is the play here?

策略是什么？

What is the play?

如果你是埃隆·马斯克，想明年建到10吉瓦的规模，而到目前为止，德克萨斯州只从数据中心接入了3.9吉瓦的电力。

If if you are Elon Musk and you want to maybe build up to 10 gigawatts next year, and so far, Texas has only connected 3.9 gigawatts to the grid from from data centers.

你打算怎么做？

How are you gonna do this?

嗯，埃隆·马斯克有个计划，他的计划是在数据中心现场自己发电，用喷气涡轮机和天然气发电设备、天然气发动机之类的。

Well, Elon Musk has a plan, and his plan is to generate power capacity at the sites themselves using things like jet turbines and natural gas generation, natural gas engines.

而且现在的情况是，随着所有建数据中心的人都意识到需求有多大，大家同时都明白，他们得从依赖电网转向自己发电，尽可能获取所有资源来现场发电。

And and what's happened is as it becomes clear to everyone who's building all these data centers just how much demand there is, Everyone's realizing at the same time, they're going to have to shift from connecting to the grid to getting all the supplies possible to generating power on-site.

我屏幕上正好有这个图表。

I've got the chart right here on screen.

但六个月前，大约70%的项目计划接入电网，30%是表后发电。

But six months ago, about 70% of projects were planning to connect to the grid and 30% behind the meter.

现在，仅仅六个月，这个比例就变成了差不多五五开，这些数据来自供应链内部的人。

Today, in a shift of six months, that splits closer to fifty fifty, and and that data comes from people inside the supply chain itself.

所以我们会看到，公司会迅速转向考虑这些表后发电的解决方案。

So what we're going to see is a rapid shift towards companies looking at these behind the meter solutions.

那么，哪家公司能从中受益呢？

So what is a company that could benefit from this?

嗯，FTAI Aviation绝对是一个我认为投资者需要考虑的名字，因为它...

Well, FTAI Aviation is definitely one name I think investors need to consider because it...

这真的很奇怪，如果2023年有人告诉你AI会变得这么厉害，你肯定猜不到。

It's such a weird thing that if if I had told you how big AI was going to be back in 2023, you would have never guessed.

你绝对想不到，哦，我得去买喷气发动机。

You would have never guessed, oh, I need to go buy jet engines.

但这就是经济中这些相互关联的领域运作的方式，现实往往比小说更离奇，当我们把这些趋势推演下去的时候。

But this is the way that these interconnected areas of the economy work that the reality is often stranger than fiction when we when we play out these trends.

就像我之前说的，能够发电的一个关键方法基本上就是用喷气发动机来发电，然后用它们为这些大型数据中心供电。

And like I said earlier, one of the key ways to be able to generate power is essentially taking jet engines to generate power and using them to generate power for these large data centers.

而这正是FTAI Aviation将要做的。

And this is what FTAI Aviation's going to do.

正如我提到的，SpaceX现在有1.4吉瓦的计算能力。

SpaceX, as I noted, they have 1.4 gigawatts of compute today.

他们预计会增加到……

They're expected to go up to...

他们正努力在2027年达到10吉瓦，就像你在屏幕上看到的。

They're trying to build up to 10 gigawatts in 2027 as, again, you can see on the screen.

那他们打算怎么做呢？

How how are they gonna do this?

他们会寻找表后解决方案。

Well, they're gonna look for behind the meter solutions.

因为表后方案能让你再次绕过电网。

Because what behind the meter does, it allows you to once again skip the grid.

而且第二点，还能帮你绕过很多其他瓶颈。

But number two, allows you to skip a lot of other bottlenecks.

如果你有这些喷气发动机在现场发电，那通常还能帮你跳过很多其他积压的环节，比如特定电压的变压器、变电站等等，这些都有好几年的积压，因为你不需要再走电网这条流程了。

If if you've got these jet engines that are essentially generating power on-site, what it allows you to often do is also skip a lot of the other backlog components, like transformers of certain voltages, electrical substations, all these other things with years of backlogs because you're not having to once again go through this grid pipeline.

你只需要在现场发电，然后直接供应给现场使用。

You are simply generating power that you're going to be delivering on-site.

所以，你就不需要去碰那些环节了。

So, again, you're not needing to go into those areas.

问题是，这些转向电力需求的喷气发动机公司，积压的订单量相当大。

The problem, the backlog for these jet engine companies that are converting towards power needs is substantial.

这个领域里排名第一的公司是GE Vernova。

The the number one company in this space is GE Vernova.

他们目前每个季度大约出货三吉瓦的涡轮机，而积压订单高达116吉瓦。

They're currently shipping about three gigawatts of turbines per quarter against a backlog of a 116 gigawatts.

这意味着，即使他们的产能有所增长，积压订单也要排到2031年。

So that means that their backlog now stretches into twenty thirty one even with some growth in in what they're able to produce.

这就是FTAI Aviation登场的地方。

Here's where FTAI Aviation comes in.

这家公司的老本行基本上就是……

This is a company that their legacy is essentially doing...

他们是一个租赁和维修喷气发动机的平台。

They're a platform for leasing and repairing jet engines.

他们意识到可以转型进入这个市场，而且能够创造目前不存在的产能，因为积压订单已经排到2031年了。

And and what they realized is that they could make a pivot towards entering this market, and they would be able to create capacity that currently does exist because the backlogs go to 2031.

我们在这里看到，正如下面的图表所示，今年的收入预计将从约39亿美元增长。

And what we're seeing here, as you can see in the chart below, is the revenue is expected to grow from about 3,900,000,000 this year.

随着他们扩大规模，预计到2029年收入将超过100亿美元。

And as they scale up, it's expected to hit more than $10,000,000,000 by 2029.

所以有趣的地方在这里。

So here's what's interesting.

这家公司目前的估值大约是2029年预期盈利的10倍，而GE Vernova则是这个倍数的两倍。

This company is currently at about 10 times their expected 2029 earnings versus GE Vernova, which is twice that multiple.

所以他们处于一个非常有吸引力的行业。

So they're in a very attractive industry.

这些计划有大量的产能可以吸收。

There's a lot of capacity to absorb with the the plans.

你知道，我之前给你看过那些超大规模数据中心运营商的计划。

You know, I showed you the plans of all these hyperscalers earlier.

好吧，再说一次，他们需要找到解决方案，而FTAI最近签了一份大合同，据信是跟SpaceX签的。

Well, again, they're going to need to find some solutions, and FTAI recently signed a massive contract, which is believed to go to SpaceX itself.

还有，对FTAI来说，一个潜在的威胁是，埃隆·马斯克认为这个行业太关键了。

And, again, a potential threat to FTAI is that Elon Musk sees this industry as so essential.

他打算开始自己制造喷气涡轮机，以便为他的数据中心供电。

He's going to begin trying to build his own jet turbines to be able to power his data centers.

所以一方面，这是一个威胁。

So on one hand, that is a threat.

另一方面，他认为这非常关键。

On the other hand, he sees this as so essential.

这本身就说明了市场的状况。

It says something about the market itself.

另一个好处是，如果未来FTAI的销售额真的下降，那么你至少有一个庞大的安装基础。

And another beautiful thing is that if the sales actually decline for FTAI in the future, well, one thing is you get a large install base.

而且这些设备还会有维护合同。

Well, you're gonna have maintenance contracts on these as well.

所以你还是能获得一些经常性利润。

So you do get some recurring profits behind.

所以我认为这家公司真的非常令人兴奋。

So this is a company I think is really exciting.

不过，我也推荐过这个领域的其他一些公司。

I've recommended some other companies in this space, though.

我……

I...

我从来没推荐过Bloom Energy，但我认为它是表后市场最具颠覆性的公司之一。

I've never recommend Bloom Energy, but I I see it as one of the most disruptive companies in the behind the meter space.

你会在下面看到他们的股票代码和标志。

You'll you'll see their ticker and logo below.

Solaris能源基础设施。

Solaris Energy Infrastructure.

这是另一家SpaceX依赖的公司，因为他们有极高的建设速度目标。

This is another company that SpaceX is relying upon for their extremely ambitious build rates.

他们向数据中心出租移动燃气轮机。

They rent mobile gas turbines to data centers.

再说一次，你知道，他们提供的是全套服务。

Again, you know, they're a complete package.

它涵盖了发电、配电、排放控制、储能、运营和维护。

It spans generation, distribution, emission control, storage, operations, and maintenance.

这是一家转型的公司。

It's it's a company that pivoted.

很多这样的公司都在转型。

So many of these companies are pivoting.

他们看到了……

They are seeing...

你可能会说，这些公司正在转向一个即将达到顶峰的市场。

You know, you could say, well, these companies are pivoting to something that is about to enter its peak.

但如果你了解当前对AI的旺盛需求，以及未来几年电网在解决电力需求方面的局限性，你就会明白这对这些公司来说有多赚钱，而且能持续多年。

But if you understand the insatiable demand happening right now for AI and you understand how limited the grid is going to be in terms of solving power needs across the coming years, you see how lucrative this can be for for these companies for years.

对吧？

You know?

这更像2024年初的存储器股票，而不是现在的情况，因为现在这个局面已经基本明朗了。

It is something more akin to memory, these stocks in memory at the beginning of twenty twenty four than where they are today where this situation has largely played out.

你知道，我们现在正想进入像表后计量这样的趋势。

You know, we are looking to get into trends like behind the meter right now.

在我看来，市场对这些公司未来几年建设过程中可能看到的增长率理解得很不够。

When the market, in my opinion, poorly understands the kinds of growth rates these companies could see as we look at the build out across the coming years.

最后，我想说，如果你甚至没在看像FTAI这样的公司，你知道，它转向喷气发动机用于发电，还有Bloom Energy。

And lastly, I'll just say, if you weren't even looking at companies like FTAI for its, you know, its pivot towards jet engines as, you know, power generation, Bloom Energy.

如果你对Solaris Energy不感兴趣，它的股票代码是SEAI……

If you weren't interested in Solaris Energy, which is s e a I...

股票代码是SEI，你甚至可以看看天然气公司，因为很多这种表后数据中心开始运营，会给本来就很紧张的天然气供应带来额外压力，而与此同时，天然气还会面临更多的全球出口需求，比如已经建成的液化天然气项目。

S e I as a ticker, you could even look towards natural gas companies as a lot of these behind the meter, you know, data centers begin operating, it's going to be incremental pressure in very tight natural gas supplies right at a time that natural gas is going to see more global exporting things to love, liquefied natural gas that's been created.

所以你可以看看像WBI、EXE、EQT这样的公司。

So you could look at companies like WBI, EXE, EQT.

再说一次，这些都在屏幕底部。

Once again, all on the bottom of the screen right now.

所以这就是……

So this is this is...

我觉得这确实是表后市场。

You know, I I do think this is behind the meter.

如果你真的深入了解AI，这个话题已经讨论了两三年了。

It's it's been something that if you are really in the weeds of AI, it's been talked about for two or three years.

但真正加速、需求变得非常紧迫的时候，是从2027年开始的。

But where it really hits is acceleration and the need for it becomes really acute truly begins in 2027.

所以我对这些机会感到非常兴奋。

So I'm I'm really excited about some of these opportunities.

另一个在2027年开始成形的有趣趋势是，英伟达正在推动电力输送方式上史上最大的架构变革之一。

Another really interesting trend that is beginning to take shape in 2027 is NVIDIA's pushing for one of the largest architecture shifts ever in how power is delivered.

我们经常谈论向数据中心供电。

We talk a lot about delivering power to data centers.

嗯，那只是其中一方面。

Well, that's one side of it.

另一方面是电力在数据中心内部是如何实际分配的，而这一领域很少有人关注，因为这里有着巨大的增长潜力。

Another side of this is how power is actually delivered inside the data center, and that's an area far too few people have been thinking about because it is an area with extreme amounts of growth.

所以我们看到的是，英伟达正在推动，基本上，他们想要……

So what we see here is NVIDIA is pushing that, basically, they want to...

目前，电力是以大约54伏的电压输送到服务器的。

Currently, you're going to take power at about 54 volts to servers.

他们想把这个电压提升到800伏，并且想省去交流电和直流电之间的很多转换步骤。

They wanna take this up to 800 volts, and they wanna eliminate a lot of conversion steps between alternating current and direct current.

他们想把直流电直接送到服务器。

They wanna take direct current directly to the servers.

那么，这会带来哪些下游影响呢？

Now what are the downstream implications of this?

首先，这会提高效率。

Well, number one, it's gonna improve efficiency.

这会让数据中心消耗更少的电力。

It's gonna make data centers consume less power.

这很好，因为正如我们刚才在上一部分看到的，电力是多么大的一个瓶颈。

That's that's good because, again, we just saw in the last section how much of a bottleneck power is.

但另一件事是，这会大幅减少铜的使用量。

But another thing is gonna lead to significantly less copper usage.

NVIDIA估计，对于一个千兆瓦的数据中心，这种转变将减少20万磅的铜需求。

NVIDIA has an estimate that for a gigawatt data center, this transition will lead to 200,000 fewer pounds of copper needed.

但对股票的其他影响是，你基本上会看到很多收入从那些制造专用配电单元的公司转移到芯片本身，也就是电源管理芯片。

But the other implications on stocks is going to be you're going to essentially move from a lot of revenue going to companies that make these specialized kind of power distribution units into chips themselves, power management chips.

而且现有的电源管理芯片将使用全新的化合物。

And the kinds of power management chips that exist are going to use entirely new compounds.

它们将不再使用硅，而是使用碳化硅、氮化镓等材料，有些公司在这方面有专门的技术。

Instead of silicon, they're going to use things like silicon carbide, gallium nitride, and there's companies with specific expertises in these spaces.

那么，我们在这里关注哪些公司呢？

So, you know, what are some of the companies that we're looking at here?

很多公司在四月、五月和六月都有过一波上涨，但最近几个月，随着许多动量股在夏季回落，它们也出现了大幅下跌，这让它们的估值更具吸引力。

Well, a lot of them have seen runs throughout kind of April, May, and June, but a lot of them have seen significant falloffs, in recent months as many momentum stocks hold off across the summer, which makes their valuations far more attractive.

这些公司通常是...

These are often companies...

很多这类用于高电压的电源管理芯片，以前主要只用在电动汽车市场上。

A lot of a lot of these kind of power management chips for higher voltage were only previously applicable to EV markets.

所以这些公司经常被当作电动汽车公司来定价和看待。

So these companies are often priced and thought of as EV companies.

但随着这个800伏的转变发生，你看……

But as this 800 v shift happens, again, you...

我们只会看到……

We're only gonna see...

我在之前的幻灯片里提到，2026年数据中心采用800伏的比例估计是2%，2027年是37%，2028年是59%，然后2029年是74%。

I've got in the prior slide, 2% data centers at 800 v estimate in 2026, 37% in 2027, 59% in 2028, and then '70 four percent 2029.

所以实际上，未来几年我们会看到这种转变大规模发生，这将为这些电源管理芯片公司带来蓬勃的收入，比如英飞凌、意法半导体，这些都是推荐标的。

So we're effectively going to see a widespread shift happening across the coming years, and that's going to lead to booming revenue for these power management chips for companies like Infinion, ST Microelectronics, which is a recommendation.

如果你看一个更具投机性的想法，比如纳微半导体，这又是一家投机性公司。

And when you look at a more speculative idea, something like Navitas, which, again, this is a speculative company.

你……

You...

你会想只建一个小仓位，但这家公司可能会从这个转型中获得显著收益。

You'd want to position to only a small position, but a company that maybe could see significant benefits from this transition.

但有一个我之前在AI投资者播客里没讨论过的电源管理股票，就是Vicor。

But one stock idea that I haven't discussed before on the AI investor podcast from power delivery is going to be Vicor.

它的股票代码是VICR。

And this is ticker v I c r.

这家公司，相对于我们之前看的一些其他公司，可能稍微贵一点。

It's it's a company that, relative to some of the others that we would be looking at, is maybe a little bit more expensive.

但它所拥有的，是可能解决下一个由距离引起的瓶颈。

But what it what it has is it might solve kind of the next bottleneck caused by distance.

这真的很重要，因为我们想想高带宽内存。

And that's really important because we think about high bandwidth memory.

再说一次，这种用于AI服务器的高带宽内存，其收入从2025年的239亿美元增长到2027年预计的1730亿美元。

Again, this high bandwidth memory, this memory that's used in AI servers, high bandwidth memory went from $23,900,000,000 in revenue in 2025 to an estimated $173,000,000,000 in 2027.

增长真是惊人。

Just phenomenal growth rates.

高带宽内存领域的龙头股包括美光，过去一年涨了大约700%。

And and the leading stocks in high bandwidth memory are companies like Micron, which across the past year is up about 700%.

还有SK海力士，涨了540%。

SK Hynix, which is up 540%.

这些都是我们推荐过的股票。

Again, these are recommendations we've made.

高带宽内存的作用，简单来说，就是把内存尽可能靠近计算芯片。

And what high bandwidth memory does is, basically, it puts memory as close as possible to the compute die.

内存距离只有大约5毫米，而传统内存的距离是30到55毫米。

Puts memory about five millimeters away, where conventional memory sits 30 to 55 millimeters away.

所以，把内存靠近计算芯片，能提高每瓦特的带宽。

So this this putting memory closer to the computing die leads to more bandwidth per watt.

在AI时代，这绝对是至关重要的。

And in an AI world, this is just absolutely essential.

这是一个非常重大的概念。

This is an absolutely massive idea.

然后我们接着往下看，我们之前聊过的下一个趋势，就是共封装光学。

And when we go down and look at the next trend that we've talked about, co package optics.

在AI网络里，也有类似的逻辑，就是我们需要把距离缩短到毫米级别。

Well, in AI networking, there's a similar dynamic that we need to basically get the distance down to millimeters.

我们看到，共封装光学这个市场，预计到2026年会达到1亿美元。

And what we see with copackaged optics is this market is expected to be a $100,000,000 in 2026.

基本上现在几乎为零，但会扩大到150亿美元。

Essentially nothing, but expand to 15,000,000,000.

那增长率是多少呢？

So what's that growth rate?

那是相当高。

It's a lot.

那是相当高。

It's a lot.

到2030年，大概能增长150倍左右。

It's a 150 x or so by by by 2030.

而且再次强调，距离只有毫米级别，因为目前主流技术是这些可插拔的光模块，它们插在备份服务器上，然后需要把光信号转换成电信号，再传输大约15到30厘米。

And once again, the distance is millimeters because the current dominant technology is these pluggable transceivers that plug into the backup servers, and then you have to convert from a light signal to an electrical signal and run about 15 to 30 centimeters.

而共封装光学做的就是，再次把距离缩短到毫米级别，让电信号转换尽可能短，这样能带来巨大的效率提升。

Well, what copackage optics does is, again, it takes it down to essentially a distance of millimeters where you're going to have basically just this electrical transition be as as small as possible, and that's going to have incredible efficiency gains.

我们看看在共封装光学领域占据主导地位的公司，比如Lumentum，过去一年涨了617%。

And when we look at the companies with dominant positions, copackage optics that we have right here, Lumentum, up 617 the past year.

Coherent涨了218%。

Coherent, up 218%.

大家都在找能解决这个技术瓶颈的公司。

You know, people are looking for the companies that can solve this technology bottleneck.

很多投资者都在找那些供不应求的瓶颈环节。

We, you know, so many investors are looking for bottlenecks of where where things are sold out.

但真正要找的是下一个技术瓶颈是什么。

What you're actually looking for is what's the next technology bottleneck?

下一个架构变革是什么，哪些公司在这方面最有优势？

What's the next architectural shift, and what companies are best positioned for that?

而且我们可能在这里遇到了电力输送本身的问题。

And we might have something here happening with power delivery itself.

所以Vicor拥有的是他们的垂直电力输送模块。

So what Vicor has is their vertical power delivery modules.

我试着把这个移一下。

I'm gonna try and move this here.

其实，我们先退出这里，我把这张幻灯片移一下，这样你能看得更清楚。

Actually, let's exit here, and I'm gonna move this slide so you can see it better.

但Vicor要实现的，基本上就是把电力输送进一步靠近芯片本身。

But what what Vicor is going to enable is that they're going to basically move, again, power delivery closer to the die itself.

而且，这将会把一些东西从……

And, again, this is going to take something that is moving from...

你做了这些先进的封装。

You do these advanced packaging.

芯片上的空间越来越宝贵，最终你会遇到这样的情况：你需要一个尽可能靠近芯片的电力输送单元。

The space on chips is getting more and more premium, and you're going to wind up at a situation where you're going to need a power delivery unit that's as close as you can to the die.

而这项技术正是Vicor所主导的。

And this is a technology that Vicor is dominant in.

随着我们看到更高电压直接输送到芯片，这将变得越来越关键。

And as we see higher voltage delivered directly to chips, this is going to become more and more essential.

所以现在的Vicor，价格有点贵，尤其是相对于我们之前讨论过的其他一些股票。

So Vicor today, it's a little expensive, especially relative to some of the other stocks that we've talked about.

但如果技术趋势真的发生转变，我相信它会遵循我们在高带宽内存和光学领域看到的模式——谁拥有最好的技术，能够尽可能靠近计算芯片并提供电力，我认为这家公司就是Vicor。

But if it's technology, where we're seeing this shift happens, I believe is going to follow what we saw in high bandwidth memory, what we saw in optics, where it's who has the best technology for being able to essentially get as close to the compute die as possible and empower delivery, I think that company is Vicor.

如果这真的实现，我认为这家公司有巨大的上涨空间，类似于我们在内存和光学领域看到的情况。

And if this does play out, I think this is a company with significant upside similar to what we've seen in areas like memory and areas like optics.

那么这是不是有点不对称，要么成功要么失败？

So is this a little bit more asymmetric where it either plays out or doesn't?

我觉得确实是这样。

I think definitely that is that is the case.

但如果真的实现了，我认为这个趋势带来的上涨空间会相当可观。

But if it does play out, I think the upside is relatively substantial for this trend.

那我们继续往下看。

So let's move on.

我们还有一个2027年的趋势，我想稍微聊聊半导体设备。

We've got one more trend for 2027, and I just wanted to talk a little bit about semiconductor equipment.

如果你一直在听这个播客，也就是AI投资者播客，你会知道我们相信未来几年半导体设备的增长速度会超出预期。

If you've been listening to this podcast, the AI investor podcast in the past, you know that we believe that the growth rate of semiconductor equipment is going to outpace expectations in the coming years.

有意思的是，这个领域在未来几年里位置非常好，而且它算是过去十五年投资中隐藏的赢家。

And what's what's so interesting about this is this is a space that's extremely well positioned for the coming years, and it's a space that's been kind of the hidden winner across the past fifteen years of investing.

当我们看表现最好的股票时，现在屏幕上就有，你会看到像英伟达、康福特系统、特斯拉、博通、美光这些名字。

When we look at the top performing stocks, which I have on screen right now, you see names like NVIDIA, Comfort Systems, Tesla, Broadcom, Micron.

这些名字里有很多投资者可能都知道。

Many of those names are things that investors would likely know.

但继续往下看，还有泛林集团、科磊、应用材料。

But you continue down the list, and you've got Lam Research, KLA Corp, Applied Materials.

关键点在于，供应制造芯片的设备，其实一直是地球上最好的行业之一，只是不太引人注目。

The key idea here is supplying the equipment to make chips has been kind of sneakily one of the best industries on Earth.

你知道吗？

You know?

这里又没有苹果。

You you don't see an Apple here again.

也没有亚马逊。

You don't see an Amazon.

大多数投资者熟悉的名字都不在这里。

You don't see all the names that most investors know.

这些公司其实是在幕后默默耕耘。

These companies kind of toil in the background.

但随着科技在经济中占比越来越大，那些制造芯片设备的公司，嗯，这真的非常复杂。

But as technology becomes an increasing part of the economy, the companies that make the systems to make the chips, well, it's incredibly complex.

这个领域门槛很高，很难进入。

It's it's an area that it's it's hard to get into.

而一旦你在制造芯片设备的某个特定垂直领域有了专长，通常就能长期保持优势。

And when you have an expertise in a specific vertical area of building systems to make chips, you generally are able to hold on to the advantages for long periods of time.

所以，你知道，我觉得当我们展望未来几年的关键机遇时，就是我刚才在上一部分提到的电力传输本身，也就是先进封装。

So, you know, I I think when we're looking at what the key opportunity for the coming years is, it's something I just got into in the last section talking about power delivery itself, that advanced packaging.

你知道，过去芯片的性能取决于……

You know, in the past, how capable chips were would be...

从七纳米到五纳米再到三纳米，就是不断缩小晶体管本身的尺寸。

You'd be going down from seven nanometers to a five nanometers to three nanometers or just increasingly shrinking the size of the transistors themselves.

但未来我们要关注的是，我们已经足够小了，这种缩小能力对芯片性能提升的作用会减弱，而如何封装芯片将变得更加重要。

But where we're going to be looking in the future is we are small enough that this shrinking capability is going to be less of the progression of chip capabilities and how you actually package chips.

再说，如何……

Again, how you...

如何把元件紧密地放在计算核心旁边，如何堆叠内存。

How close you put things to a compute die, how you're able to stack memory.

所有这些领域都将越来越成为驱动力。

All of these areas is going to increasingly become the driver.

所以，那些在封装技术方面有专长的公司将成为最大的赢家。

So the companies with expertise in how to do this packaging are going to be the biggest winners.

而且，这也让一些新玩家有机会进入这个领域。

And, also, it allows some new entrants into this space.

所以你会看到一些公司，它们在自己擅长的细分领域里占据主导地位，然后站稳脚跟。

So you're going to be able to see companies that have their new kind of niche areas that they dominate in establish themselves.

就像我之前提到的，过去十五年里，排名前十五的股票中有三只，每只都涨了超过五千倍。

And as I showed saying that three of the top 15 stocks across the past fifteen years, each with 5000% returns or above.

我们的半导体设备公司，可能会在这里诞生一些新的赢家。

Our semiconductor equipment companies, we could see the birth of some of the new winners here.

所以，我特别看好的几只股票，都在我们的投资组合里，比如Onto Innovation、Camtek，还有一家更大的公司，可能涨幅没那么大，但风险也低一些，就是Lam Research。

So some of the stocks I really like here that are in our portfolio, Onto Innovation, Camtek, A larger one that may be less upside but less risk would be a Lam Research.

而且，随着我们进入2027年、2028年，大家会越来越清楚，要支撑起这些数据中心的规模，需要多少半导体设备。

And, again, as as we come into, you know, 2027, 2028, and and it becomes clear how much more of the semiconductor equipment is needed to be able to build out the ambition around how big these data centers are.

我觉得这些公司的增长率会远远超过华尔街的预期。

I I think the growth rates for these companies is is going to be significantly in excess of what Wall Street expects.

所以，你知道，这些趋势可能听起来有点专业，比如电表后端、电力输送、半导体设备这些。

So, you know, this one, you know, I I know these are all kind of in the weeds trends in the sense that, you know, behind the meter and and power delivery and semiconductor equipment.

它们现在不像内存那样广为人知。

They're they're not as well known today as memory.

但再说一次，你只需要回顾一下。

But, again, you just need to go back.

如今，投资领域搜索量最高的东西是DRAM ETF和内存等。

Today, the most searched for thing in investing is areas like DRAM ETF and memory.

两三年前，这看起来会非常无聊。

Two or three years ago, that would have seemed deeply boring.

非常无聊。

Deeply boring.

而今天，它成了投资者高度关注的领域。

And today, it's an area of intense interest among investors.

我觉得我们在那里有一些下一步。

And I think we have some next steps there.

我在上一部分讲了五个。

I covered five of them across the last section.

除了更多内容之外，我还给你推荐了那三只重点股票。

I also gave you those three kind of featured stock ideas in addition to a lot more.

所以，你知道，我觉得这个领域未来会受到越来越多的关注。

So, you know, I think I think this is just an area that is going to see a lot of increased attention.

关注。

Attention.

所以，让我们……

So let...

让我们把目光放得更远。

Let's look beyond.

我刚才讲了一些2027年的大话题。

So I just covered some of big topics in 2027.

让我们再看看更远的未来。

Let's let's look beyond.

2028年及以后会发生什么呢？

What what happens in 2028 and beyond?

大局是什么？

What what is the big picture?

有哪些趋势现在刚开始，也许它们在2027年还不会看到催化剂和加速，因为它们还有更多工作要做，但这些趋势可以改变世界。

What are some trends that are beginning today that maybe they're not going to see the catalyst and the acceleration in 2027 because they still have a little bit more work happening, but there are trends that can change the world.

我们在这里会讲一些这样的趋势。

We'll we'll cover some of those here.

所以我觉得问题……

So I I think question...

我想讲的第一个问题是我所说的3450亿美元的问题。

The first question I wanna cover is what I'm calling the $345,000,000,000 question.

当你报道AI时，最常被问到的问题就是：这是泡沫吗？

And this is, you know, the number one question you get asked when you're covering AI is, is it a bubble?

这是每个人都想知道的。

That's what everyone wants to know.

这是泡沫吗？

Is it a bubble?

而泡沫从哪里开始……

And where the bubble begins...

问题在于，像谷歌这样的公司会不会支出超过它们的能力范围？

Question begins is, are companies like Google going to be spending more than they're capable of?

如果AI需求放缓，它们就会陷入超支的困境，不得不迅速收缩，市场就会像互联网泡沫那样自我崩溃。

And if the demand for AI slows down, they're going to be caught in a situation where they were spending essentially beyond their means, and they're gonna have to pull back extremely rapidly, and we're going to see the market basically implode on itself like the .com bubble.

这就是问题所在。

That is the question.

我在这里列出的3450亿美元，就是我认为这些大型AI公司明年（2027年）计划在资本支出上的投入，与它们预计的经营现金流之间的差距。

And so this $345,000,000,000 question I have outlined here, that is the gap between what I believe most of these large AI companies are going to try to spend next year on CapEx in 2027 and what they're going to earn in operating cash flow.

谷歌明年，华尔街预计其经营现金流为2665亿美元，而它们内部的雄心是支出接近3750亿美元。

Google next year, Wall Street has them at 266,500,000,000 in operating cash flow, and their internal ambitions are to spend closer to 375,000,000,000.

那它们打算怎么做呢？

So how are they going to do this?

嗯，首先，我觉得可以从几个角度来看这个问题。

Well, number one, I think there's a few ways that we can look at this.

赚更多的收入。

Earn more revenue.

现在的情况是，华尔街很可能低估了这些公司能赚多少钱。

What's happening right now is Wall Street's likely underestimating how much revenue these companies will earn.

你知道，这些公司都……

You know, these these companies all...

尤其是亚马逊、微软、谷歌，它们都有云计算部门，而谷歌的云计算部门现在正以82%的速度增长。

Especially an Amazon, a Microsoft, a Google, they have cloud computing units, and Google's cloud computing unit's now growing at 82%.

如果对AI的需求持续下去，它们会赚更多的钱。

If demand for AI continues, they're going to earn more money.

所以，也许明年它们的运营现金流不是2665亿美元。

So maybe maybe they'll earn instead of 266,500,000,000 next year in operating cash flow.

可能更接近3300亿或3500亿美元。

Maybe it's closer to 330,000,000,000, 350,000,000,000.

这样一来，那个差距就没那么大了。

And suddenly, that gap isn't so large.

第二，我觉得我们会看到更多融资工具。

Second, I think we're gonna see a lot more financing vehicles.

英伟达最近宣布了一个5000亿美元的特别目的载体，基本上就是把计算这些DAS小时变成一种资产类别。

NVIDIA recently announced a $500,000,000,000 special purpose vehicle, essentially turning computing these DAS hours into an asset class.

博通正在洽谈为其自己的平台筹集1000亿美元。

Broadcom is in talks for a 100,000,000,000 for its own platform.

我觉得我们会看到公司做更多融资工具，基本上就是把计算变成一种资产类别。

I think we'll see companies do a lot more financing vehicles, essentially turning compute into an asset class.

第三，他们可以发行债务或股权，我们在AI投资者播客上已经讨论了很多。

And third, they can issue debt or equity, which we've talked a lot about on the AI investor podcast.

但是，你知道，最近的债务发行继续被超额认购。

But, you know, recent debt issuances continue to be oversubscribed.

需求非常旺盛。

There's an incredible amount of demand.

像伯克希尔·哈撒韦这样的公司，巴菲特仍然是董事长，正在带头购买谷歌的债务，这表明像巴菲特这样的人有多么相信这家公司的质量，即使他们正在发行债务来继续资助这些大规模的建设计划。

A company like Berkshire Hathaway, with Buffett still has its chairman, is leading the charge to buy Google's debt, which shows how much quality someone like a Buffett believes this company still has even as they are issuing debt to continue funding these large build out plans.

所以这仍然是他们继续获得更多运营现金流的一种方式。

And and so that that continues to be a way that they can continue getting more operating cash flow.

然后我觉得这就引出了下一个问题。

And and then I think this leads into the next question.

到2028年到2030年，AI市场会有多大？

How how big will the AI market be in 2028 to 2030?

因为当我们看这张图时，它显示了这些超大规模企业每年各自的花费。

Because when we look here is a chart that shows how much each of these hyperscalers are spending by year.

你知道，我在整个演示中多次展示过这张图，但它确实显示了起飞的程度。

You know, I've displayed this several times throughout this presentation, but this really shows how much of the takeoff.

如果你觉得2023年到2025年已经很疯狂了，你只需要看看图表上2023年到2025年之间的这段时期。

You know, if we think if you think that 2023 to 2025 were crazy, you you only need to look at the period here on the chart between 2023 and 2025.

与接下来我们要去的地方相比，这根本不算什么。

It's absolutely nothing compared to where we're headed next.

而且，我们还能继续增长吗？

And, again, can we keep growing?

2027年会是AI的顶峰之年吗，还是我们能继续增长超越它？

Is 2027 a peak year for AI, or can we keep growing beyond it?

我觉得有几件事值得考虑。

You know, I I think a few things.

第一，OpenAI和Anthropic。

Number one, OpenAI and Anthropic.

像这样的领先公司能产生多少收入？

How much revenue are leading companies like this going to be able to generate?

Anthropic在IPO过程中提到，到2020年底，他们的年化收入（ARR）将达到约2000亿美元。

Anthropic, as part of their IPO process, has point towards exiting 2020 about $200,000,000,000 in what you call annualized revenue, ARR.

他们的一些风投说这个数字更接近2500亿，而一些第三方风投，比如我在播客里提到过的Gavin Baker，声称这个数字可能达到4000到5000亿。

Some of their VCs have said that's closer to 250,000,000,000, and some third party VCs like a Gavin Baker, who I've talked about on this podcast, have claimed the number could be 400 to 500,000,000,000.

这些公司能否达到他们的宏伟目标，将很大程度上说明这个市场能有多大。

Where these companies hit on their lofty targets will say a lot about how big this market can go.

因为，再说一次，如果OpenAI和Anthropic在明年年底前收入超过2500亿，他们基本上就变成了像微软一样规模的公司。

Because, again, if an OpenAI and an Anthropic are hitting numbers beyond 250,000,000,000 in terms of their revenue by the end of next year, they have essentially become companies the size of Microsoft.

这些公司的规模只比谷歌或苹果稍微小一点。

Companies the size that are just slightly trailing behind a Google or an Apple.

而这将大大推动AI行业本身的收入增长。

And that's going to significantly fuel how much revenue is coming into AI itself.

我们来看看数据，未来几年需要关注的关键数字是这样的。

When we look at the numbers, here's the numbers to know in the coming years.

8450亿美元。

845,000,000,000.

这是2026年这些超大规模企业预计的支出。

That's what's expected to be spent by these hyperscalers in 2026.

1.37万亿美元是他们预计的……

1,370,000,000,000 is their expected...

他们在数据中心上的投入。

How much they're gonna spend on data centers.

这是2027年的资本支出预算。

Their outlay is for capital expenditures in 2027.

还有1.446万亿美元，这是他们目前预计在2028年的支出。

And 1,446,000,000,000, That's how much they're currently expected to spend in 2028.

那么像英伟达CEO黄仁勋这样的人认为他们能花多少呢？

Now what does someone like NVIDIA CEO Jensen Huang believe they can spend?

他把2030年的数字定在了每年大约3万亿到4万亿美元。

Well, he's put the number by 2030 at closer to 3,000,000,000,000 to 4,000,000,000,000 annually.

如果我们真的达到那个数字，今天的大多数公司，比如英伟达、博通，这些大型AI公司，回头来看会显得非常便宜，因为它们基本上是按照增长率从2027年的高点开始放缓来定价的，我们稍后会详细讨论这一点。

If we did hit that number, most of the companies today, an NVIDIA, a Broadcom, these large AI companies are going to look extremely cheap in retrospect because they're essentially priced, and we'll get more into this in just a minute, that growth rates are stalling from where they hit in 2027.

我们怎么能从这里继续增长呢？

How could we continue growing from here?

这真的取决于这些模型变得有多好。

Well, it really comes down to how good these models get.

我们有一些潜在的突破，而且实验室本身也发出了很多信号，表明他们即将迎来下一波AI增长。

We have some potential breakthroughs, and there's been a lot of signals from the actual labs themselves that they're about to see a next wave of AI growth.

一个领域是持续学习。

One area would be continual learning.

现在的模型，它们……

Today's models, they're...

它们基本上是在训练后就固定了能力，它们获得记忆的方式就像我们之前讨论的那样。

They essentially train in their frozen in capability, and they get ideas like memory from from what we've talked about earlier.

基本上，这些技巧让现在的AI模型在某种意义上非常低效，它们需要使用KV缓存来记录你之前的对话，并且在生成每一个新词时访问这个KV缓存。

Basically, these hacks that AI models today in a sense are very inefficient, and they they need to use KV cache to essentially take notes from your prior conversation and and access that KV cache when it's generating every single new additional word itself.

持续学习会带来的是……

What continual learning would do would...

它能让模型真正从经验中更新自己。

It would allow models to actually update from experience.

所以这不仅会让它们能力大幅提升，而且想想看，这里会产生的切换成本。

So it would make them not only significantly more capable, but think about this think about the switching cost that would exist here.

如果持续学习成为现实，而你主要使用OpenAI或Anthropic，那么模型拥有的记忆会突然变得非常有价值。

If continual learning is a concept and you're using predominantly OpenAI or Anthropic, suddenly, the memory that the models have would become significantly more valuable.

所以你得想想，现在像NVIDIA这样的公司，突然更积极地收购Hugging Face，还有收购那些开源模型公司，这背后的意义有多大。

So you have to think about how much of today's rates amongst companies like NVIDIA suddenly getting far more involved buying Hugging Face and and buying, you know, companies in open source models.

他们有多少是在预见到下一代模型开发的方向，并且知道这些切换成本会有多高呢？

How much is them maybe seeing where this next generation of model development is going and knowing how sticky these switching costs could be?

你知道，像英伟达这样的公司在财报电话会议上。

You know, this is a company like NVIDIA on their earnings call.

他们对AI公司的未来以及像OpenAI、Anthropic这样的公司的收入机会一直非常乐观。

They have been nothing but bullish about where the future of AI companies is going and the revenue opportunity for companies like OpenAI, Anthropic.

而且我确实认为，这很大程度上是因为他们看到了AI模型下一个转变的方向。

And I do believe a big part of that is where they see the next shift for AI models coming from.

另一个领域，我这边记着，第二点，长期自主任务。

Another areas, I've got right here, number two, long horizon agentic tasks.

你知道，这又是，因为目前LLM的设计使得这些自主循环非常消耗内存。

You know, this is, again, today, because because these agentic loops are so memory intensive with the current design of LLMs.

如果我们能大幅提高它们的效率，我们就能处理持续数小时甚至数天的任务。

Well, if we are able to make them significantly more efficient, we could have tasks that could go for hours or days.

那将会是……

It would be...

你看，我这里有个任务，今天你说，怎么迁移数据库并获得建议？

You know, I've got here, but a task today, you say, how do I migrate a database and get suggestions?

一个长期代理任务就是，你对一组代理说，我想迁移一个数据库，然后他们会工作好几天，回来告诉你，我们已经迁移了数据库。

A long horizon agentic task would be, you say, to a group of agents, I want to migrate a database, and they would work for days, and they would come back and say, we've migrated the database.

我们已经测试过了。

We've tested it.

我们修复了所有错误，并且已经完全部署。

We fixed all errors, and it's fully deployed.

这就是你看到这对工作流程影响多大的地方，以及AI如何成为工作流程的核心，尤其是对于广泛的知识工作，也就是我的第三点，更广泛的知识经济。

This is this is where you see how much this impacts workflows and where AI becomes the center of workflows, especially for broad knowledge work, which is, you know, my third point, the broader knowledge economy.

我在前面的演示中花了很多精力谈论编码的增长，以及它如何推动了2026年AI的惊人增长。

I I took a lot of pains to talk about the growth of coding earlier in this presentation and how much that's driven the incredible growth of AI across 2026.

编码之所以成为AI真正革命化的第一个领域，并且完全改变了编码在公司中的部署方式，是因为编码是可验证的。

Well, the reason that coding was kind of a first area that AI really revolutionized it and, you know, has completely changed how coding is is deployed across companies is coding is what you call verifiable.

你写代码，然后看它是否有效。

You you write code, and you see whether or not it works or not.

数学是可以验证的。

Math is verifiable.

有很多领域都是可以验证的。

There's there's a lot of domains that are verifiable.

但很多广泛的知识工作并不是可验证的。

A lot of broad knowledge work is not verifiable.

它需要像判断力这样的东西。

It requires ideas like judgment.

但有很多迹象表明，AI在这些不可验证的领域正变得越来越擅长。

But there's a lot of signs that AI is becoming extremely more adept at these nonverifiable domains.

如果真是这样，AI的潜在市场就会大幅扩大，因为像美国、日本这些发达经济体，它们的核心基础就是更广泛的知识经济。

And if that happens, the total addressable market for AI increases substantially because, you know, advanced economies like The US and Japan and, you know, these developed markets, their core foundation is the broader knowledge economy.

就是知识经济。

It's the knowledge economy.

如果AI突然能更广泛地应用于这些知识领域，能用上代理式AI，能带来实实在在的改进和巨大的生产力提升，那它的价值就会比现在翻十倍。

And if AI is suddenly much more applicable to this broader knowledge area, again, where you can use agentic AI and where it can have meaningful additions and massive productivity boost, you know, it it 10 x's from the levels today.