AI刚刚拯救了SpaceX：创纪录的146亿美元季度、首次盈利以及特斯拉合并

我很兴奋。

I am excited.

四个月前，SpaceX进行了历史上规模最大、最昂贵的IPO，华尔街几乎有一半人都不看好他们。

SpaceX four months ago had the largest, most expensive IPO in history, and half of Wall Street basically wrote them off.

他们称这家公司被严重高估，然而昨晚，他们公布了季度财报，单季度营收高达146亿美元。

They called the company massively overvalued, and yet last night, they reported their quarterly earnings and a massive $14,600,000,000 in revenue in a single quarter.

这比他们去年一整年的收入还要多。

That is more money than they made in the entirety of last year.

但最有趣的是他们是怎么赚到这笔钱的。

But what's most interesting is how they made that money.

SpaceX，顾名思义，是一家太空公司，但他们大部分收入却来自人工智能。

SpaceX, as its name suggests, is a space company, yet they made the majority of their money through AI.

他们的业务线在过去三个月里翻了三倍，首次超过了星链。

Their business line has tripled over the last three months, eclipsing Starlink for the first time ever.

他们之所以能做到这一点，是因为他们已经成为全球第一的云服务商，为包括Anthropic和谷歌在内的全球一些最大的实验室和公司提供计算服务，每月销售数十亿美元的计算资源。

And the reason why they've been able to pull this off is because they've become the number one neocloud, the number one provider of compute to some of the biggest labs and companies in the world, including Anthropic and Google, selling billions of dollars worth of compute per month.

而且事实证明，算力是一门极其赚钱的生意。

And it turns out compute is an incredibly lucrative business.

马斯克预计，到2027年，也就是明年，年经常性收入将达到5000亿到1万亿美元。

Elon Musk is projecting $500,000,000,000 to $1,000,000,000,000 of ARR next year in one year, 2027.

但别光听他说。

But don't just take his word for it.

摩根大通也支持这一观点，并给出了每股500美元的新目标价。

JPMorgan also backs this up and has given them a new price target of $500 per share.

我们刚才提到了很多大数字，我仔细研究了整个财报，还有更广泛的战略，我真的很想深入聊聊，因为其中涉及SpaceX和特斯拉的合并。

There's a lot of big numbers that we're throwing around here and I dug into the entire earnings report, but also the wider strategy, which I'm really excited to jump into because it includes a merger between SpaceX and Tesla.

这两家公司合并有很多理由，最终可能让他们成为全球第五大最有价值的公司。

And there's a lot of reasons why these companies should merge and eventually make them the fifth most valuable company in the entire world.

所以，让我们在今天的节目中深入探讨这一切，本期节目由高通赞助。

So let's get into all of this on today's episode brought to you by Qualcomm.

四个月前，我们做了一期关于SpaceX IPO的节目。

Now four months ago, we made an episode covering the SpaceX IPO.

那时候情况看起来有点不太妙。

Back then things looked a little bit precarious.

他们以每股135美元的价格上市。

They IPO ed at $135 per share.

股价后来涨到了，我记得是171美元，然后接下来几周又跌了超过50%。

The stock then sent to, I think it was $171 and then it proceeded to dump more than 50% over the subsequent few weeks.

华尔街当时非常看空，这也很合理，因为SpaceX没法证明他们到底怎么赚钱。

Wall Street was incredibly bearish and for good reason, because SpaceX couldn't demonstrate how exactly they were going to make money.

太空显然是个很好的愿景，但他们还有很长的路要走，得证明火箭测试成功，才能做到一天多次往返发射火箭。

Space was obviously a really good vision, but they had a long way to go to kind of like prove the rocket tests and get to the point where they could launch a rocket back and forth multiple times per day.

但过去几个月发生了一些事情，改变了一切。

But something happened in the last couple of months that changed all of it.

那就是埃隆·马斯克的一个简单转变：他们收购了xAI，并且拥有了比任何AI公司，甚至任何公司都多的算力。

And it was a simple pivot by Elon Musk, which was they acquired xAI and they acquired the most compute out of any AI company or just company in general out there.

他们买下了最大量的GPU，现在这些GPU都放在他们的三个核心数据中心里，分别是Colossus一号、二号和三号。

They acquired the largest amount of GPUs, which they now house under their three core data centers, Colossus one, Colossus two, and Colossus three.

这些GPU原本是用来支持croc的，但马斯克还没能完全搞定。

Now these GPUs were meant to serve croc, except that Elon Musk couldn't quite get it to work just yet.

与此同时，他有了个聪明的想法。

And in the meantime, he had a bright idea.

他决定把这些算力租给需要的人。

He decided to rent that compute out to people who needed it.

比如Anthropic，现在每月付给他们12.5亿美元，还有谷歌，每月也付大约10亿美元。

The likes of Anthropic, who are now paying them $1,250,000,000 per month and Google, who's also paying about $1,000,000,000 per month.

马斯克觉得，因为他比其他任何公司都能更快地扩展算力，所以可以收取溢价。

And Elon figured that because he was the quickest to be able to scale compute quicker than any other company out there, he could charge a premium.

在录制这段视频的时候，股价已经上涨了大约9%，但我想深入看看季度报告里的实际数据，以及这些数据真正说明了公司的情况和未来的发展方向。

Now at the time of recording, the stock is up about 9%, but I want to get into the actual numbers from the quarterly reports and what that actually tells me about the company and where they're headed.

先看一些关键数字：营收146亿美元，比上个季度增长了161%。

So some of the headline numbers, $14,600,000,000 of revenue compared to last quarter, which is up a 161.

这绝对是个巨大的数字。

Absolutely huge.

但这里最重要的点是，这主要来自人工智能。

But the most important point to take away here is that it largely derived from AI.

仅人工智能板块就带来了大约76亿美元的收入，比上个季度增长了830%。

The AI segment alone pulled in around $7,600,000,000 which is up 830% from last quarter.

这些数字背后有几个值得注意的点。

Now there's a few things to take away from these numbers.

第一，它已经超过了星链的收入。

Number one, it's eclipsed Starlink revenue.

这很重要，因为星链一直是SpaceX的主要收入来源。

And this is a big deal because Starlink has for a while been the major revenue earner for SpaceX.

而且这次是大幅超越。

It's eclipsed it in a pretty big way as well.

第二，这正式标志着SpaceX历史上首次实现盈利。

Number two, this officially gives SpaceX the first time in its history of achieving a profit.

正如你所料，埃隆正在加倍押注这一点。

And Elon is, as you can expect, doubling down on this.

这个季度的资本支出是196亿美元，他下个季度的目标是350到400亿美元。

This quarter's CapEx spend was $19,600,000,000 and he's aiming for 35 to $40,000,000,000 next quarter.

但我觉得最吸引人的是，这些投入会如何转化为公司的收入，因为预测数字相当可观。

But what I find most fascinating is how this is going to translate into revenue for the company because the projections are pretty hefty.

到今年年底，SpaceX计划上线大约3吉瓦的计算能力，这已经很厉害了。

By the end of the year, SpaceX plans to have around three gigawatts worth of compute online, which is impressive.

但更厉害的是明年的预测，马斯克估计他们会有大约10到20吉瓦的计算能力。

But what's more impressive is the projection for next year where Elon estimates they'll have around 10 to 20 gigawatts worth of compute.

当然，财报还有其他亮点，包括他们的AI卫星网络StarMind，他们公布了很多令人兴奋的细节，还有TeraFab，如果建成的话，那将是史上最大的芯片制造厂。

Now there were obviously other highlights of the earnings, including StarMind, which is their AI satellite network, which they're launching loads of exciting details there, as well as TeraFab, which is the largest chip manufacturing plant that will ever exist if it comes to light.

但最核心的问题是，一家太空公司怎么就成了全球第一的神经云服务商？

But the number one story here is how did SpaceX become the number one neocloud in the world when they're a space company?

答案很简单，就是马斯克，以及这家公司是全球最会扩张的实体扩张者。

Well, the answer simply is Elon and the company is the number one scaler physical scaler in the entire world.

神经云这门生意其实很好理解，对吧？

The business of being in neocloud is very easy to understand, right?

你主要从英伟达买GPU，顺便说一句，SpaceX和埃隆是英伟达GPU最大的买家之一，甚至可能就是最大的。

You buy GPUs mainly from Nvidia and by the way, SpaceX and Elon is one of the, if not the biggest purchaser of NVIDIA GPUs.

所以一旦你有了这些GPU，就把它们放在一个叫数据中心的大仓库里。

So once you have those GPUs, you house them in a warehouse known as a data center.

你把各种线都插好，确保温度合适，确保有足够的电力和能源来供应这些GPU。

You plug in all the different cables, make sure temperature is regulated, make sure there's enough power and energy to feed these GPUs.

你把这些GPU上线，然后你决定做两件事中的一件。

You bring those GPUs online and then you either decide to do one of two things.

要么你自己用，训练和推理你自己的模型，要么你把它租给竞争对手或者那些需要同样推理和计算能力的人，从中赚一大笔钱。

You either use it for yourself to train and inference your own models, or you rent it out to competitors or people who want that same inference and compute and make a hell of a lot of money out of it.

现在，埃隆和SpaceX用Grok尝试了第一种选择，听着，他们进展不错，但还不够快，不足以证明投入所有这些钱到计算上是值得的。

Now Elon and SpaceX tried the first option with Grok and listen, they're making good progress, but not quick enough progress to warrant investing all of this money into compute.

所以埃隆灵光一现。

So Elon had a light bulb idea.

他想，我就把多余的算力卖出去，赚足够的钱来继续资助我想做的其他事情，说不定还能赚一笔。

He was like, I'll just sell any excess compute that we have to make enough money to still fund all the other things that I want to do, and maybe even I'll turn a profit.

然后他往旁边一看，发现OpenAI和Anthropic在算力上线方面真的很吃力。

And then he looked over the fence and he realized, Open Air and Anthropic are really struggling to bring compute online.

他们看了看Meta，觉得Meta做得还不错。

They looked at meta and they said, they're not doing that bad.

再看看谷歌，自己搞TPU，他心想，祝你好运吧。

Looked at Google building their own TPUs and more like, good luck with that.

我干脆直接从最好的GPU制造商和供应商——黄仁勋的英伟达那里买，这样我就能最快搞定。

I'm just going to buy from the best maker and provider of GPUs, Jensen Huang, Nvidia, and then I'll be the quickest one to do it.

这样我就能在所有人中建立起巨大的领先优势，拥有大量多余的算力，要么用来训练世界上最顶尖的模型，要么直接卖给竞争对手。

And I'll gain such a huge marginal lead on everyone that I'll have so much excess compute to either decide to train some of the best models in the world, or I'll just sell it to my competitors.

你得明白，马斯克在这方面是个高手，这可不是夸张。

And you have to understand, Musk is a master at this, and that's not an exaggeration.

他每个月能上线20万块GPU。

He's bringing online 200,000 GPUs per month.

这相当于每月投入约131亿美元的重型设备，来支撑整个AI资本支出的建设。

That is around $13,100,000,000 worth of heavy metal machinery to fund this whole AI CapEx build out online per month.

没有其他公司能做到这一点。

No other company is doing this.

当被问到他是怎么做到的，他的回答特别搞笑。

And when he was asked about how he was able to do this, his answer was pretty hilarious.

他说，我每周有四天待在那里。

He said, I'm spending four days a week there.

我就睡在制造车间的地板上，待在孟菲斯的Colossus 12和3个数据中心，努力把所有东西都上线，确保一切搞定。

I'm sleeping on the manufacturing site floor, and I'm at the Colossus 12, And 3 data centers in Memphis, trying to bring all of this online and making sure it gets done.

更疯狂的是，这还不是最难的部分。

What's even crazier is that's not even the hardest part.

其实最简单的部分就是买GPU并让它们上线。

The easiest part actually is buying the GPUs and bringing them online.

难的是你从哪里找土地、电力和电力来启动所有这些设备？

The harder part is where do you get the land, power and electricity to turn all of these things on?

因为我在之前的节目里提过，现在AI竞赛中一个关键的制约因素和瓶颈不是买GPU。

Because I've mentioned this on previous episodes, one of the key constraints and bottlenecks in the AI race right now is not buying the GPUs.

关键是能不能把它们启动起来。

It's switching them on.

关键是能不能给它们通电。

It's powering them on.

关键是能不能获得足够的电网电力，并且突破所有监管上的繁文缛节，来给你的GPU供电。

It's getting the electrical grid energy and enough of it and surpassing all the red tape of regulatory burdens to feed your GPUs.

亚马逊这周其实说了，他们有价值数十亿美元的GPU就那样放着积灰。

Amazon actually said this week that they have billions of dollars worth of GPUs that are just collecting dust.

微软也有同样的问题，因为他们还没拿到许可证。

Microsoft has the same issue as well because they haven't got the permits.

他们还没有涡轮机。

They haven't got the turbines.

他们还没有电网来解决这一切。

They haven't got the energy grid to try and figure this all out.

这是埃隆·马斯克过去一年一直高度关注的事情。

This is something that Elon Musk has been hyper focused on for the last year.

我跟你说个疯狂的事实。

And I'll tell you a crazy fact.

他们改造了SpaceX的火箭车间，用来制造涡轮机，用于发电机，临时发电机，这些发电机在电表后面或者数据中心现场运行，以克服申请电网接入的障碍或挑战。

They have repurposed rocket workshops at SpaceX to build turbines to use in generators, makeshift generators that they run behind the meter or on-site at these data centers to overcome the obstacles or challenges of requesting electrical grid access.

这就是埃隆和SpaceX在与竞争对手下的三维或四维棋，这让他们在计算能力上领先第二名竞争对手十二到十八个月，我认为第二名是Meta。

This is the three d or rather four d chess that Elon and SpaceX are playing against their competitors, which have allowed them to get a twelve to eighteen month lead on compute against their nearest second place competitor, which I think is meta.

所以正是因为埃隆痴迷于这种垂直整合，他才能够实现最大的壮举之一，这可能让SpaceX在不久的将来有机会成为前沿AI实验室，或者顶级前沿AI实验室之一。

So it's because of this vertical integration that Elon has just obsessed over that he's been able to pull off one of the biggest feats that may give SpaceX a chance to be a frontier AI lab or one of the top frontier AI labs in the very near future.

但他拥有所有最先进的算力，对吧？

But he has all the sexist compute, right?

昨天的节目讲的是Core Weave，它也有很多算力，但人们质疑这些东西是否真的有实际需求。

And yesterday's episode covered Core Weave, which also had a lot of compute, but people were questioning whether there was actually any real demand for these things.

埃隆昨天在电话会议上回答了这个问题。

Elon answered that on the call yesterday.

他说，我们每天都会接到很多次敲门，来自一些全球最大的公司，他们请求使用我们的算力。

He said, we get a knock on the door so many times every single day from some of the biggest companies in the world asking for access to the compute.

他说，如果他现在的算力能再多好几倍，这些算力早就被抢购一空好几轮了。

He said, if he had multiple times the amount of compute that he has online today, he would have been able to sell that compute several times over.

但最疯狂的是，SpaceX 对他们卖出的每一瓦特数据都收取溢价。

But the craziest part about this is that SpaceX is charging the premium price on every single Watt of data that they sell.

所以现在，我觉得他们每瓦特每年的价格在50到60美元之间。

So right now, I think sells between $50 to $60 per Watt annualized a year.

如果按他们明年10吉瓦的保守估计来算，那就达到了我之前提到的4000亿美元的年经常性收入。不过，别光听我说。

Now, if you scale that out to the conservative estimate that they have of 10 gigawatts next year, that brings you to the $400,000,000,000 ARR estimate that I mentioned earlier on, but don't take my word for it.

Anthropic 每个月付给他们12.5亿美元。

You have Anthropic paying them $1,250,000,000 a month.

谷歌每个月付给他们9.65亿美元。

You have Google paying them $965,000,000 a month.

就算10亿吧。

Call it a billion.

然后我觉得 Reflection AI 也承诺了63亿美元，就为了从马斯克那里拿到尽可能多的算力。

And then I think you have reflection AI that committed 6,300,000,000 to get as much compute as they can from Elon Musk.

最疯狂的是，尽管他把所有这些算力都卖给了他们，他仍然有剩余的算力来训练一个前沿AI模型。

The craziest part about this is even though he's selling them all of this compute, he's still able to have excess compute to train a frontier AI model.

我觉得今年年底会推出的下一个模型将是一个10到15万亿参数的模型，我告诉你，这需要大量的算力。

I think the next model that is gonna come out by the end of this year is going to be a 10 to 15,000,000,000,000 parameter model, which let me tell you requires a lot of compute.

他之所以能做到这一点，是因为他通过出售这些算力的短期租约来实现的。

And the way he's been able to do this is by selling short term leases on this compute.

通常，如果你是Anthropic或OpenAI，你会想和算力提供商签订长期协议，这样他们就能保证有足够的算力来训练和运行所有模型。

So typically, if you're Anthropic or OpenAI, you wanna sign a long term agreement with a compute provider so that they can guarantee that they have enough compute to train and serve all their models.

埃隆·马斯克，因为他严格来说是个竞争对手，而且他想构建自己的AI模型，所以他说，听着，我可以把算力卖给你，但只签短期合同，比如三到六个月，或者九到十二个月。

Elon Musk, because he's technically a competitor and he wants to build his own AI model is saying, listen, I'll sell you the compute, but only for short term contracts like three to six to nine to twelve months at a time.

而且我还会为此向你收取溢价。

But I'm also gonna charge you a premium to do that.

这就是他能够做到的事情，因为所有这些客户都在支付那个价格。

And that's what he's been able to pull off because all of these customers are paying that price.

所以需求超过了可用的算力，而物理上的限制正在阻碍人们将这些GPU上线。

So the demand is in excess of the amount of compute that is available and physical atoms are constraining people for bringing these GPUs online.

但埃隆·马斯克找到了绕过很多这些问题的方法，而且到目前为止，他是唯一一家能做到这一点的公司。

But Elon Musk has found a way to circumvent a lot of this, and he's so far the only company that has been able to pull this off.

所以，有几件事取决于这个计划是否真的可行。

So a few things are contingent on this actually being realistic.

第一，埃隆能像现在这样持续增加算力，但他在这里会遇到很多阻力。

Number one, Elon is able to continually bring on compute in the manner that he has, and he's going to run up against a lot of friction here.

他会遇到电网限制。

He's going to run up against grid constraints.

他会遇到制造限制，比如他正在自己制造燃气轮机之类的，但这些东西迟早会达到产能上限。

He's going to run up against manufacturing constraints like cool that he's building his own gas turbines and stuff, but that's going to reach capacity at some point.

现在需求还会存在吗？

Is the demand going to still exist right now?

他可以收取溢价，因为市场上可用的算力不多。

He can charge a premium because there's not much compute available.

但我想，如果市场饱和了，或者另一个竞争对手找到了像SpaceX目前这样大规模扩张的方法，那会发生什么？

But what happens when I guess the market becomes flooded or maybe another competitor figures out a way to scale in the same way that SpaceX has done so far.

而且最重要的是，首先制造这些GPU本身就有限制。

And on top of all of that, there's just constraints in manufacturing these GPUs in the first place.

我提到过，SpaceX计划在未来三个月内投入350到400亿美元的资本支出。

I mentioned the 35 to $40,000,000,000 CapEx spend that SpaceX alone is planning to spend over the next three months.

所以我们就看看情况如何发展吧。

So we'll see how that goes.

我们看看埃隆能收购多少并上线多少。

We'll see how much Elon is able to acquire and bring online.

这里肯定还有一些危险信号，但如果以这个季度的财报来看，前景还是乐观的。

There's still definitely some red flags here, but if this quarter's earnings is anything to go by, it's looking optimistic.

所以如果你纯粹从投资角度来看，我的看法是，我们比三个月前要乐观得多，但还是得谨慎行事。

So if you're looking at this purely from an investment perspective, my own take on this is we're looking a lot more optimistic than we did three months ago, but still tread cautiously.

记住，SpaceX的股票解禁时间表是我们见过的所有公司中最疯狂的。

Remember, SpaceX has the craziest share unlock schedule of any other company that we've ever seen.

通常你会看到一年的锁定期，然后所有股票开始解禁。

Typically you see a one year cliff and then everything starts to unlock.

解禁期在公司上市后几周内就开始了，而且他们还有那种奇怪的股价目标层级，如果达到某个估值目标，就会有更多股票解禁。

The cliff started within a few weeks of the company IPO ing and they have like kind of these, this weird hierarchy of price targets where if a valuation target is hit, then more shares unlock.

这一点需要注意。

It's something to be aware of.

事情并不完全像新闻报道的那样。

It's not just all what the news reports.

所以一定要谨慎，因为显然向市场大量抛售这些多余的股票，会造成购买需求不足的问题，从而难以维持当前的价格。

So definitely tread carefully because obviously flooding the market with all this excess amount of shares is going to create the issue of having enough purchase demand to keep the price where it's at.

但埃隆还提到了其他一些与业务其他部分相关的事情，其中一件大事就是特斯拉和SpaceX的合并。

But there's also a number of other things that Elon brought up pertaining to other parts of the business, including one big one, the merger between Tesla and SpaceX.

关于这个的传闻早在SpaceX上市之前就开始了，但埃隆终于对整件事发表了评论，这让我对这件事发生的可能性感到非常乐观，可能今年年底就会发生。

The rumor mill around this started well before SpaceX even IPO ed, but Elon finally commented on this entire thing, which gives me a lot of positivity about the potential of this happening arguably maybe by the end of the year.

那么，让我们进入今天由高通赞助的节目。

So let's get into it on today's episode brought to you by Qualcomm.

埃隆对此说的关键一句话是，应股东的要求，两家公司都成立了特别委员会来讨论合并的可能性。

Now, the key sentence Elon delivered on this was by the request of shareholders, both companies have formed special committees to discuss the potential of a merger.

所以现在特斯拉和SpaceX的股东和董事会成员大概都在碰头，讨论这次合并会是什么样子，以及是否真的可行。

So you have shareholders and board members presumably from Tesla and from SpaceX that are now huddling and having conversations about what this merger would look like and whether it's actually feasible.

但有一个关键点，埃隆因为明显的利益冲突，不能参与这些会议。

But there's one important point Elon, because he has obvious conflicts of interest cannot be involved in these meetings.

当被问到是否认为这是个好主意时，他的回答就是简单的“是”。

And when he was asked whether he thought this was a good idea, his answer was simply yes.

而且这个“是”是有充分理由的。

And it was yes for a very good reason.

因为当我想到特斯拉这家汽车公司和SpaceX这家太空公司之间的协同效应时，SpaceX在做一些AI相关的事情，还搞了个什么星链卫星网络，这之间的联系似乎不太明显。

Because when I think about the synergy of Tesla, a car automotive company, and SpaceX, a space company that's kind of doing AI stuff and creating some kind of nebula StarMind satellite network, the connection doesn't seem quite clear.

所以让我尽力解释一下那些显而易见的协同效应。

So let me try and do my best at explaining what the obvious synergies are.

第一，SpaceX正在努力让进出太空的运输成为最便宜的事情。

Number one, SpaceX is trying to make transport to space in and out of space actually the cheapest possible thing alive.

所以基本上他们会真正掌控把东西送入太空的高速公路。

So basically they'll effectively own the highway of getting things into space.

但接下来的自然问题是，那又怎样，你想在太空里做什么？

But then the next natural question is, so what, what do you wanna do in space?

太空里有什么产品？

What are the products in space?

首先，就是星链。

Well, number one, it's Starlink.

星链基本上是一个分布式网络，为地球上以及可能地球以外的任何设备和任何人提供互联网连接。

Starlink is basically a distributed network for connectivity to the internet for any device and any human on earth and maybe off of earth.

所以这里的想法是，它是一个网络，将互联网从太空向地球上的所有盲区发射，这样你就能连接到最快、最便宜的互联网。

So the idea here is it's a network that would beam internet down to earth across all dead zones so that you can be connected to the fastest, cheapest internet ever.

这样你就不必依赖那些彼此之间存在各种问题的电信基础设施公司了。

And so that you don't have to be reliant on telecom infrastructure companies that have myriad different issues between themselves.

但接下来还有星智。

But then you have StarMind.

星智是他们的AI卫星网络，目前正在与英伟达合作定制建造，而且它也将是TeraFab项目的一个大产品，我稍后会讲到这个。

StarMind is their AI satellite network, which is currently being custom built with Nvidia and also what's going to be a large product from the TeraFab thing, which I'll get to in a second.

但这里的想法是，你可以发射一颗专门的AI卫星，让它飞向太阳。

But the idea here is you'll be able to launch a specialized AI satellite launched out towards the sun.

它利用太阳能，在太空中进行AI模型的推理或训练，这听起来完全疯狂，因为有很多限制，对吧？

It harnesses the energy of the sun and is able to inference and or train AI models up in space, which sounds completely crazy because there's a lot of constraints, right?

有辐射问题。

There's the radiation issue.

还有，它到底怎么在太空中存活并得到维护？

There is the, how the hell is it going to be survived and maintained out of there?

还有一个大问题，就是这些GPU之间如何通信。

There's the big question around how these GPUs talk to each other.

在过去一年里，SpaceX一直在深入思考这个问题，他们已经想出了一些非常有说服力的原型。

Over the last year, SpaceX has been thinking very deeply about this, and they've been able to figure out some really convincing prototypes.

但当然，最大的疑问是，他们能否大规模实现这一点。

But of course, the big question mark is whether they'll be able to do this at scale.

根据埃隆提到的时间表，我们将在未来十二到十八个月左右，也就是明年，看到这些版本的实现。

According to the timelines that Elon has cited, we'll be able to see versions of this come alive next year over the next twelve to roughly eighteen months.

但这里的愿景是，你会像星链一样，有一组AI卫星在太空中利用太阳能，非常像卡尔达舍夫二级文明，而且成本尽可能低。

But the vision here is you would have similarly to Starlink, a network of AI satellites out there harnessing the power of the sun, very Kardashev type two civilization at the cheapest possible rate.

现在，如果你想要获得这种能源，首先，你会遇到很多繁琐的审批程序。

Now, in mind, if you want to get your hands on that kinds of energy, one, you run into a lot of red tape.

其次，这东西的成本高得离谱。

Two, the cost for this thing is crazy.

第三，它由太多不同的机构管理，处理起来非常麻烦。

And three, it's governed by so many different parties that it's just a pain to deal with.

所以埃隆的解决方案是，把GPU发射到太阳附近，全天候利用所有能量，以最低的成本，然后与地球上所有的数据中心、计算提供商和能源提供商竞争。

So Elon solution to this is project the GPUs out towards the sun, harness all that energy 20 fourseven for the cheapest rates possible, and then compete with all the data centers, compute providers and energy providers down on earth.

这里的想法是，没有竞争。

And the idea here is that there is no competition.

如果埃隆能把GPU运到那里，建立那个网络，并真正利用所有能量，他们最终会提供有史以来最好的服务。

If Elon is able to transport these GPUs out there, create that network, and in fact, harness all that energy, they will end up with the best offering ever.

如果有人想竞争，他们必须通过SpaceX。

If anyone wants to compete, they need to go through SpaceX.

所以这是一个疯狂的垂直整合体系，既能通过星链提供互联网连接，又能为任何AI模型提供最便宜的算力推理和训练。

So it's this crazy vertically integrated stack that would help create connectivity to the internet via Starlink and connectivity to the cheapest compute inference and training for any AI model.

这可以是Grok，但也可以是埃隆·马斯克和SpaceX目前正在向其出售算力的所有竞争对手。

And that could be Grok, but that could be also for all the competitors that Elon Musk and SpaceX is currently selling compute to.

想象一下未来的Anthropic。

Imagine Anthropic in the future.

他们可不想在地球上花大价钱买那些很烂的算力。

They don't wanna spend crazy prices for very crappy compute on Earth.

他们直接付钱给埃隆就行了。

They would just pay Elon.

埃隆可以定价更便宜，但利润率反而更高，因为他的成本比任何竞争对手都低。

And Elon can charge a cheaper price, but still make wider margins because he has it at a cheapest cost versus any of his competitors.

好的。

Okay.

那这和特斯拉有什么关系呢？

So what does this have to do with Tesla?

特斯拉制造了所有这些设备，我特意用“设备”这个词。

Well, Tesla creates all the devices and I use the term devices very specifically.

你可能觉得特斯拉汽车就是一辆车，确实如此，但它也是一台电脑，一台与其他电脑、其他特斯拉汽车相连的电脑。

You might think that a Tesla car is a car and very much it is, but it's also a computer, a computer that is connected to other computers, other Tesla cars.

目前，全球有数百万辆特斯拉汽车在行驶。

And we have currently millions of Tesla cars roaming the earth.

其中很多都具备完全自动驾驶能力，这意味着它们能比地球上99%的人类驾驶得更好。

A lot of them have full self driving capability, which means that it can autonomously drive better than 99% of humans on earth.

所以，这里的想法是，它周围有摄像头和传感器。

So the idea here is it has cameras and sensors around it.

它是一台功能完备的电脑。

It is a computer that is fully functioning.

现在就可以在车里运行Grok或前沿的Grok AI模型。

You can run Grok or the frontier Grok AI models in the car right now.

你可以把它连接到你的任何设备、任何应用上。

You can connect it to any of your devices, any of your apps.

这就形成了一个连贯的系统，几乎就像钢铁侠里的贾维斯一样。

Creates this coherent system, almost like Jarvis from Ironman.

所以，你需要物理基础设施和设备来支撑所有这些人工智能。

And so you need physical infrastructure and devices to be able to serve all the AI.

而为了服务所有这些人工智能，为了把这些不同的软件产品连接到人类，连接到我们正在交互的东西，你需要互联网。

And in order to serve all that AI and in order to connect all these different software products to humans, to what we're interacting with, you need internet.

所以，所有这些特斯拉、所有这些擎天柱机器人、这些人形机器人，以及特斯拉制造的任何其他设备，一方面可以在地球上相互连接，另一方面可以连接到全球最便宜的人工智能推理服务提供商和互联网，这是合理的。

So it makes sense that all of these Teslas, all of these Optimus robots, these humanoid robots, any other device that Tesla builds can be one connected to each other on earth and two connected to the number one cheapest AI inference service provider and the internet.

如果你是SpaceX，别忘了你还有那个叫xAI的东西，而且你正在训练自己的人工智能模型。

And if you're SpaceX, remember you have that thing called xAI and you are training your own AI models.

那么，训练最好的人工智能模型需要什么呢？

So what do you need to train the best AI models?

不仅仅是计算能力。

It's not just compute.

还需要数据。

It's also data.

看看特斯拉，特斯拉有所有的摄像头和传感器，全天候从世界各地特斯拉运营的地方收集数据。

If you look over at Tesla, Tesla has all the camera and sensors that are pulling in data 20 fourseven from around the world wherever Tesla operates.

所以把这些数据输入模型，对未来的模型迭代会非常有价值。

And so feeding that data into the model can be incredibly valuable for future model iterations.

但还有很多体力劳动需要解决。

But there's also a lot of manual labor that needs to be figured out.

人类，说到底，是有限的。

Humans, for all intents and purposes, are limited.

你需要从零开始培训他们。

You need to train them from scratch.

你需要确保他们有能力在数据中心、AI模型周围进行建设和操作。

You need to make sure they're qualified to be able to build and do the thing around data centers, around AI models.

这是一种稀缺资源。

It is a scarce resource.

但想象一下，如果你能至少自动化体力劳动这一部分，你可以通过Optimus人形机器人来实现。

But imagine if you could automate at least just the physical labor side of things, you could do that via the Optimus humanoid robots.

当然，我这么说，是拉长了时间线来看这些事情的实现。

Now, of course, I'm speaking at a very great length of time in terms of like making these things happen.

要实现这些，可不是件容易的事。

It's not going to be easy to pull off.

但核心想法是，如果未来世界上有数亿甚至数十亿台人形机器人，能自动化大量体力劳动，而你又能掌控算力，甚至掌握数据，用来打造全球最顶尖的人工智能模型，再把这两者结合起来，你实际上就拥有了一个超人类系统，能推动人类文明远远超越今天的高度。

But the idea here is if you end up in a world with hundreds of millions, maybe even billions of humanoid robots that can automate a lot of the physical labor, and then you can own the compute and maybe even the data that you can deliver to create the best artificial intelligent models in the world, and then you can pair them together, you end up with effectively what is a superhuman system that can help advance humanity well beyond where we're at today.

但还有一样东西在拖后腿，那就是芯片。

But there's one more thing holding them back, chips.

所有这些都需要物理硬件来承载它们的“大脑”，无论是训练和推理AI模型，还是驱动人形机器人或汽车。

You need physical hardware to house the brain for all of these different things, whether it's the AI model to train and inference it, or whether it's the humanoid robot or the car to power it.

谁来制造这些芯片呢？

Who's gonna build those chips?

可能是英伟达，也可能是SpaceX和特斯拉合并后的产物，最近我们看得很清楚，就是TeraFab。

Could be NVIDIA or it could be whatever the merger of SpaceX and Tesla is, which we've seen very clearly most recently, TeraFab.

TeraFab基本上是人类最雄心勃勃的尝试，要在全球打造一座专门为AI、算力和各类设备服务的最大芯片工厂。

TeraFab is basically the most ambitious effort to try and build the biggest chip plant specifically for AI and compute and devices in general on earth.

你可以想象，如果我们真的拥有一个十亿机器人的世界，不管是汽车还是人形机器人，然后我们需要足够的算力来驱动所有这些设备，那你就需要数十亿颗芯片。

As you can imagine, if we do have that world of a billion robots, whether that's cars or humanoids, and then we need enough compute to power all those things, you're going to need billions of chips.

而目前，就算你把英伟达和其他芯片制造商，比如英特尔或AMD，加在一起，也远远不够。

And right now, if you combine Nvidia or any of the other chip fabs such as Intel or AMD, you're not gonna get anywhere close.

而且，如果你是一家需要大量生产这些东西、依赖这些芯片的公司，你可能不想依赖那些可能会抬高价格或者做其他事情的公司。

And if you are one company that is reliant on producing a lot of these things and needing those chips, you may not want to rely on those companies who could price you out or whatever.

所以，解决这个问题的最好办法就是垂直整合，这也是为什么这个计划如此雄心勃勃。

So the best way to figure this out is to vertically integrate it, which is why this is such an ambitious effort.

他昨天在谈到第一阶段时提到，第一阶段投入168亿美元，在2030年前每月生产10万片晶圆，但在高峰期，十年内每月生产100万片晶圆，这还将创造大约3000多个工作岗位。

And he commented yesterday on the first phase where he said $16,800,000,000 in the first phase, a 100,000 wafers a month before 2030, but at its peak, 1,000,000 wafers a month within ten years, which will also create around 3,000 plus jobs.

目前正在推进中。

It's currently doing that right now.

现在，我想在节目最后让自己脚踏实地，冷静一下，因为我想把话说清楚。

Now, I want to end this episode by bringing myself down to earth and grounding myself for a bit because I want to be clear here.

这件事有很多可能出错的地方。

There's a lot of ways that this could potentially go wrong.

我这里列出的资本支出是350亿到400亿美元。

My list here is CapEx of 35,000,000,000 to $40,000,000,000.

这笔钱从哪里来？

Where are you getting that money from?

我上次看特斯拉的资产负债表时，他们大约有430亿美元。他是打算自己出资，还是打算再融资？

Last time I looked at Tesla's balance sheet, they had about $43,000,000,000 Is he going to self finance it, or is he going to raise more money?

有传言说，他现在正试图在Anthropic IPO之前再筹集300亿到500亿美元的资金。

Rumor has it is he's trying to raise more money right now to the tune of 30 to, I think, $50,000,000,000 before the Anthropic IPO.

我不知道他进展如何，但关键是他想尽可能多地筹集资金来支撑这笔资本支出。

I don't know how far he's got along with that, but the point is he's trying to raise as much money as he can to fund that CapEx.

这仍然是一个很大的隐患。

That's still a big red flag.

他可能无法实现这个目标。

He may not be able to pull this off.

如果他做不到，那就会推迟，甚至可能完全终止所有这些努力。

And if he doesn't, it's going to delay or at worse, stop any of these efforts entirely.

第二点，从GB300系列过渡，这可是他们的主要收入来源，每月新增20万在线用户，转向Vera Rubins可能在技术上要难得多。

Number two, transitioning from GB 3 hundreds, which is where the bread and butter is bringing on 200,000 online every month to Vera Rubins might actually be way more difficult technically.

现在，NVIDIA或者黄仁勋昨天上台了，或者可能是今天早上。

Now NVIDIA or Jensen Huang went on stage yesterday or maybe it was this morning.

是在欧洲，德国，柏林的GTC大会上。

It was in Europe, in Germany, in Berlin at GTC.

他上台说，这是最容易上线的GPU。

And he went on and said, this is the easiest GPU to bring online.

所以它不仅是最智能的，也是最容易的，大约需要五分钟，没有电缆，自带冷却系统。

So it's not only the smartest, but the easiest it takes around five minutes, no cables, it's self contained cooling system.

不过我觉得这不完全是真的。

Now I don't think that's strictly true.

显然他可能说得比较宽松，但根据历史经验，下一代GPU一旦生产出来，实际上线还需要一段时间。

And obviously he's probably being very lenient in his explanation, but if history is anything to tell us, it takes a while to actually bring the next gen GPUs online once they've actually been produced.

所以这可能会进一步推迟事情。

So this may further delay things.

最后一点，显然你得对埃隆的估计大打折扣。

And then the final thing is obviously you have to take Elon's estimate with a huge grain of salt.

一方面，摩根大通预测明年收入将达到2540亿美元，这对SpaceX来说确实是个巨大的数字，他们之前也预测过巨额收入，但一直没法证明。

You have JP Morgan on one hand predicting $254,000,000,000 of revenue next year, which granted is huge for SpaceX who previously forecasted a massive amount of revenue, but had no way of justifying it.

而另一方面，埃隆的保守估计是明年5000亿到1万亿美元。

Then versus Elon's conservative estimate of $500,000,000,000 to $1,000,000,000,000 of next year.

很多情况还有待观察。

A lot of it remains to be seen.

不管怎样，我都会坐在这儿，拿着爆米花，看这一切怎么发展。

I'm going to be sat here regardless with my popcorn watching all of this unfold.

我只想说，认真讨论这些数字真的有点疯狂。

And I just want to say it is crazy to be talking about these figures at length realistically.

就算倒退五到十年，三个月赚150亿美元几乎是闻所未闻的，当时肯定只有一两家公司能做到。

If we wind back even five to ten years, the notion of making $15,000,000,000 in three months was almost unheard of, definitely reserved for just one or two companies at the time.

而现在我们看到，新公司也能做到这一点，它们经历了最昂贵的IPO，或者像Anthropic和OpenAI这样的私营公司，据说每季度能赚数百亿美元，而且它们自己也可能在未来几个月内进行重磅IPO。

And now we're seeing this happen to new companies that pop up after the most expensive IPOs or even private companies such as Anthropic and Open Air who are rumored to be making tens of billions of dollars per quarter that are gonna be having blockbuster IPOs themselves over the next couple of months.

我听说Anthropic可能这周末就要公开提交上市文件了。

I heard that Anthropic potentially files publicly at the end of this week.

简直太疯狂了。

Just insane.

那这期节目就到这里。

So that is the episode.

财报是昨晚发布的。

The earnings dropped last night.

我熬了一整夜，希望这些内容对你有帮助。

I was up all night, and hopefully this was helpful to you.

如果你喜欢这期节目，记得给我们点个赞。

If you enjoyed this episode, give us a thumbs up.

其实这里有很多新朋友，很多新注册订阅、在看我们内容的人。

There's actually a lot of newbies here, a lot of new people that have signed up and subscribed and are watching our content.

非常感谢，欢迎你们。

Thank you so much and welcome.

建立这个频道、制作所有这些内容，真是一段旋风般的旅程，但我每天都在努力带来关于AI和前沿科技领域所有事情的最佳观点、新闻和见解。

This has been a whirlwind of a journey building up this channel and building up all this content, but I'm trying to bring the kind of best takes news insights on all the things happening in AI and frontier tech every single day.

最后我想留给大家的一点是，昨天我做的那个节目挺有意思的，讲的是Core Weave，它也是一家新云服务商，正在努力证明自己的估值合理，结果把整个股市都拖下水了，而SpaceX做的事情其实差不多，只是规模大得多。

And one final point I'll leave you with is it's funny making the episode that I did yesterday, which was about Core Weave, who's also a neocloud struggling to justify their valuation and tanking the entire stock market versus SpaceX that is doing arguably the same thing just at a much larger scale.

也许他们有资金支持，但他们的营收预测更疯狂，而且市场反应积极得多，因为他们创造了那么多收入。

And maybe they have the money to back it, but they have crazier forecasts for revenue and there's a much more positive reaction because they're generating so much revenue.

需求显然是存在的。

The demand is obviously there.

人们愿意为这种通勤买单。

People are paying for this commute.

人们想要构建AI模型，为他人提供服务，并在此基础上开发不同的产品。

People want to be building AI models and serving it to others and building different products off of it.

所以现在存在一种奇怪的二元对立，市场很谨慎。

So there's just this weird dichotomy that exists right now where the market is tentative.

他们担心我们处于泡沫之中。

They're worried that we're in a bubble.

然后SpaceX证明了至少在最近三个月里，收入这块是非常可观的。

And then you have SpaceX kind of proving that the revenue aspect, at least in the interim or over the last three months is very much there.

所以接下来会很有趣，我们下期节目见，本期节目由Hawkham赞助播出。

So it'll be exciting to see all of this play out and I'll see you on the next episode of In the Loop brought to you by Hawkham.