Elon Musk on Terafab, US Chip Production, and the AI Compute Race
打开互动全文版(中英对照 + 朗读 + 问答)→埃隆·马斯克解释美国芯片产能仅占全球约 10%的原因,披露太瓦级 Terafab 计划,并警示亚洲地缘政治风险。
Elon Musk explains why the US produces only about 10% of the world's chips, unveils his plan for a terawatt-scale Terafab, and warns of geopolitical risk in Asia.
在准备这次对谈时,我脑海里浮现的画面之一,是《E.T. 外星人》里那个男孩骑着自行车从月亮前飞过,还带着一台用来联络、用来「打电话回家」的装置。这显然就是 Starlink 的前身。看来那招没成功。埃隆凡事都遵循一个原则——第一性原理:每件事都要用对的方式去做,不是最快的方式、不是最省事的方式、不是最便宜的方式,而是最好的方式,而且每件事都要从头重新审视。所以如果有人替他的某家公司做一笔投资,衡量标准永远是:这是我们能做到的最好吗?最快吗?最有效吗?我想说的一点是,他是个建造者。他大概是我所知道的美国唯一一个真正在造东西的人——而且造的不是普通的东西,是 1000 万平方英尺、500 万、2000 万平方英尺的厂房。现在他又在考虑、或者说准备为 Terafab 建一座 1 亿平方英尺的厂房。从来没有人做过这种事。没有任何人。连类似的都没有。那么问题当然是,做这样的事会带来规模效益——你采购、生产的效率都会远高于一个小体量的实体。所以我想问:你是怎么做到的?为什么从来没有别人尝试过类似的事?我们正在努力让芯片制造重回美国。我们需要它。我们需要电力,我们需要芯片。除了你,没人在做这件事。为什么?
When I was preparing for this conversation, one of the things I visualized was E.T. going across the moon with a boy on a bicycle and a contraption to communicate, to call home. Obviously a predecessor to Starlink. So I guess that didn't work. So Elon uses a principle, first principles, about everything should be done the right way—not the fastest way, not the quickest way, not the cheapest way, the best way—and everything is looked at all over again. So if someone makes an investment on his behalf for one of his businesses, it's always viewed in terms of: is this the best we can do? The fastest we can do? The most effective? So one thing I would say is that he's a builder. He's probably the only builder I'm aware of in the United States who builds stuff—but builds stuff that's not just regular stuff, like 10 million square foot buildings, 5 million, 20 million. And now he's got a 100 million square foot building for Terafab that he's thinking about or going to build. And no one else has done this. No one. Anything like it. And so the question is, of course, when you're doing such a thing, you get scaling benefits. You get to buy things, you produce things much more effectively than you could if you were in a smaller entity. So the question is: how do you manage to do this, and why didn't anyone else ever try to do anything like this? We're trying to get silicon to come back to the United States again. We need it. We need power. We need silicon. No one's doing this except for you. How come?
嗯,这么说其实不太准确,因为台积电(TSMC)和三星(Samsung)都在美国建晶圆厂。台积电正在亚利桑那州建一大片晶圆厂,我记得其中一些已经投产了。三星在得州也有已投产的晶圆厂,而且正在为特斯拉建——应该说已经建好了——一座专门生产特斯拉自研 AI5 和 AI6 芯片的晶圆厂。但和亚洲的规模相比,美国的芯片制造量非常小。即便按最乐观的假设,把所有已宣布要在美国建的晶圆厂都算上,也仍然只是亚洲的一小部分——大概 10% 的量级,乐观估计 10%。大多数人应该都知道,台湾在芯片生产上承担了极其繁重的工作。台积电是一家了不起的公司——真的很惊人,它做的芯片种类之多、规模和产量之大,都令人难以置信。三星也是一样,我觉得美光(Micron)在存储方面也做得很好。所以芯片绝大部分是在亚洲制造的。我真正担心的是,那个地区始终存在一定的地缘政治风险。我认为这种风险某个时候可能会成为现实;到那时,如果我们在美国没有大规模造芯片所需的全部环节,我们可能就完全没有芯片,或者只有很少——也就是那 10% 左右。再往前看,越过中期可能出现的各种地缘政治颠簸——就算不考虑地缘危机,我们造芯片的速度也实在太慢了。如果你想做到每年一太瓦(terawatt)的芯片产量——这正是我们的目标,也是它叫 Terafab 的原因。有时候我说「tera」,人们会想到 Terafab 有三种可能的含义:一是「terror」(t-e-r-r-o-r,恐怖),有人就会问:「你干嘛要建一座恐怖工厂?」然后是「terra」(t-e-r-r-a),意思是地球。但它其实是「tera」(t-e-r-a),就是太瓦的那个「太」。不过可以说三种意思都沾边——它大得吓人。全世界今年大概会生产 20 GW 的 AI 芯片,也许 25 GW。明年,看你信谁的预测,大概是 35 到 45 GW。可要是想做到一太瓦,那就需要大约明年全球产量的 30 倍。
Well, that's actually not quite accurate, in that TSMC and Samsung are building fabs in the US. So TSMC is building a large collection of fabs in Arizona. Some of them are, I believe, operational. And Samsung has some operational fabs in Texas and is building—or has built, I should say—a fab for Tesla that will be dedicated to the AI5 and AI6 chip designs that Tesla's done. So, but relative to what's being done in Asia, the amount of chip fabrication in the US is very small. And even if you take the best-case assumptions of all the chip fabs that have been announced that will be built in America, that's still a tiny fraction—perhaps on the order of 10%, optimistically 10%—of what's in Asia. So, as I think most people know, Taiwan does an incredible amount of heavy lifting on chip production. TSMC is an amazing company—truly incredible, the sheer number of a variety of chips and the magnitude and the volume is incredible for what TSMC does. Same thing for Samsung, and I think Micron does a great job too on the memory side. So it's overwhelmingly built in Asia. So my concern really is that there is always some geopolitical risk in that region. And I think that risk may come to fruition at some point, in which case, if we don't have all the pieces necessary to build chips at scale in America, then we may not have any chips at all, or very few—the 10% number type of thing. And then there's the—maybe looking ahead, past any sort of medium-term geopolitical bumps in the road—we're simply, even if you exclude geopolitical crisis issues, the rate at which we're making chips is simply too slow. So if you want to get to, say, a terawatt a year of chip production, which is our goal—that's the reason it's called the Terafab. Sometimes people, when I say 'terra,' there's three possible meanings to Terafab: one is 'terror' (t-e-r-r-o-r), which I think sometimes people like, 'Why would you want a terror-fab?' Oh, and then there's 'terra' (t-e-r-r-a), which just means Earth. But it's actually 'tera' (t-e-r-a), as in terawatt. So but arguably all three apply. So it's terrifyingly large. So the rest of the world is making—we'll make this year around 20 GW of AI chips, maybe 25. And then next year, depending on whose projections you believe, it's maybe around 35 to 45 GW of AI chips next year. But if you want to do a terawatt, that's now—you need 30 times what current global production will be roughly next year.
只给我们自己用。我们是为自己建的。
For just us. We're building that for us.
嗯,是的,我们在建——当然我们也可以卖给别人,但基本上我们是在一座工厂里建出相当于全年产量 30 倍的产能,你会觉得它的产值能抵得上全世界——抵得上全球第 20 大经济体的 GDP。就一座工厂。
Well, yes, we're building—I mean, we could sell it to other people too, but basically we're building 30 times what the annual production is in one plant, which you think is going to do as big production as the GDP in the world—as a GDP for the 20th largest country. One plant.
是啊,说不定还能进前十。
Yeah, it might even make the top 10.
所以,显然,这样的计划——我觉得如果失败了,大家会说它疯狂;如果成功了,就会说它大胆。所以某种程度上取决于结果。我认为它至少在一定程度上会成功。
So now, obviously, you know, a plan like this—I think if it fails it'll be crazy, and if it succeeds it'll be audacious. So it sort of depends on the outcome. I think it will succeed at least to some degree.
希望如此。
Hope so.
所以,是的。是的。我是说——
So yeah. Yeah. I mean—
这要花一万亿美元。长期来看可能还不止。
It's going to cost a trillion dollars. Probably more than that over time.
但我们不会立刻需要那么多钱。工厂会分期建设。而且我们正在研究改变芯片制造工艺的方法,提升芯片生产效率,就像我们在很多其他领域做过的那样。举个例子:现在一片晶圆在晶圆厂里,超过一半的时间只是在被搬运和等待。也就是说,芯片在晶圆厂里一半以上的时间什么都没做——要么在被挪来挪去,要么在等机器。所以我认为芯片制造基本上还有很大的改进空间。
But we won't need that money immediately. We'll be built in phases. And we are looking at ways to change the manufacturing process for chips to improve the efficiency of chip production, as we've done in many other arenas. Like, just one example: right now the amount of time that a wafer in a chip fab spends simply being transported and waiting is more than half the time. So more than half of what a chip is doing in a chip fab is nothing. It's either being moved around or waiting for a machine. So I think there's basically a lot of potential for improvement in chip fabrication.
那些人为什么不自己去改?如果他们的工厂里芯片一半时间都停着不动,是什么阻止了他们改进自己的系统?
Why don't these people do it themselves? If they have plants just sitting with chips half the time just not moving, what's to stop them from improving their own systems?
嗯,不是这样。其实你去参观晶圆厂就会看到,晶圆一直在大量移动,到处飞快地穿梭。也有等待,但晶圆上的原子以有用的方式被改变的时间不到一半。我会说有时候只有大概 20% 的时间,那片晶圆上的原子在发生有用的变化。他们之所以不愿在晶圆厂上做冒险的尝试,是因为这些设施要投入巨额资本,如果手里已经有一套行之有效的配方,他们就不想冒险尝试新东西。所以芯片行业普遍的偏好是:照上次的做法重复一遍,做些小改动,不冒任何风险。所以我们正在得州建一个研究实验室,就在奥斯汀的得州超级工厂(Gigafactory Texas)。这个实验室会尝试许多不同的芯片制造方法。
Well, no. Actually, if you visit a fab, you'll see that the wafers are moving a lot. They're zipping all over the place. And waiting, but the amount of time that atoms are being changed on the wafer in a way that is useful is less than half the time. And I'd say sometimes it's maybe 20% of the time that something useful is happening to the atoms on that wafer. The reason that they don't want to try something adventurous or with chip fabs is that these things cost a tremendous amount of capital, and if they have a recipe that works, they simply don't want to risk something new. So generally the preference in the chip industry is to repeat what they did last time with minor changes and not take any risks. So we're building a research lab in Texas, just in Austin, at Gigafactory Texas. That research lab is going to try many different approaches to chip fabrication.
其中许多方法不会奏效,但我确信有些会。然后我们会把这些改进转化到 Terafab 本身,也就是量产用的 Terafab,位于德克萨斯州 Grimes 县,靠近德州农工大学。那就是我们将扩大 AI 芯片生产规模的地方。
Many of those approaches won't work, but some of them I'm sure will work. And then we will translate those improvements to the Terafab itself, the production Terafab, which is in Grimes County in Texas near Texas A&M. And that's where we'll scale up production of our AI chips.
我认为奥斯汀研究实验室有趣的一点是,我在想我们是否有能力做逻辑、内存封装和光刻,这样我们就能设计新芯片,并在同一个屋檐下制造整个芯片、测试它,实现非常快的迭代周期。通常需要一年的事情,我想我们可能缩短到一个月。
I think one of the things that's going to be interesting about the research lab in Austin is that I wonder if we will have the ability to do logic, memory packaging, and do the lithography so that we can design new chips and under one roof make the whole chip, test it, with a very fast iteration cycle. Something that would normally take a year, I think we may be able to reduce to a month.
所以对于任何给定的技术努力,进步的速度取决于你能做多少次迭代以及在这些迭代之间能取得多少进展。因此,如果你能提高迭代速度,以及每年或每单位时间能做多少次迭代,并且在任何给定技术中每次迭代之间的进展,你将比任何人都快得多地加速进步。如果你能快一个数量级地迭代,作为一阶近似,你将快一个数量级地推进你的技术。
So for any given technology endeavor, the pace of progress is determined by how many iterations can you do and how much progress can you make between those iterations. So if you can improve the speed of iteration and how many iterations you can do per year or per unit time and the progress between each iteration in any given technology, you will accelerate your progress at a rate far greater than anyone else. If you're able to iterate an order of magnitude faster, you as a first approximation will advance your technology an order of magnitude faster.
所以大多数人——商业中存在巨大的惯性,因为如果有人努力花钱但没成功,那对他们来说是个污点。如果成功了,而他们想争取企业的最高职位,那只是意味着他们必须更努力工作。所以如果成功,他们更努力;如果失败,他们就失去了争取最高职位的机会。
So most people are—there's a tremendous amount of inertia in business because if someone makes an effort to spend money and it's not successful, then that's a black mark on them. And if it is successful and they're trying to get the top job of the business, then it just means that they have to work harder. So if they're successful, they work harder, and if it fails, they're finished for their opportunity for the top job.
Elon 在他的企业中建立了一种文化,总是关于带着想法来找我,然后如果成功——如果你认为成功,我相信是这样——那么我们会抛弃之前所做的,重新开始。总是关于做最好的事情。
Elon has established a culture in his businesses where it's always about come to me with ideas and then if it's successful—if you think it's successful and I believe that's the case—then we'll throw away what we've done before and start all over again. It's always about doing the best thing.
你的文化中让我感兴趣的一点是,你介绍我认识了 陈立武(Lip-Bu Tan),他是英特尔负责人,现在将对 Terafab 有兴趣。所以我和他待了一段时间,他来拜访我,我说:‘和 Elon 一起工作是什么感觉?’他说:‘嗯,其他人来参观我在俄勒冈的设施,他们走一个半小时就走了。Elon 来了,他坐在那里——他待了四个半小时,他一个接一个地面试工程师,质疑一切。’
One of the things that's interesting to me about your culture is that you introduced me to Lip-Bu Tan, the person who's in charge at Intel who's now going to have an interest in Terafab. So I spent some time with him and he comes to visit me and I said, 'So what's it like to work with Elon?' He says, 'Well, other people come to visit my facility in Oregon and they walk around for an hour and a half and they leave. Elon comes and he sits there—he was there for four and a half hours and he interviews one engineer after another after another, questioning everything.'
所以一切都与速度有关。然后我去德克萨斯州博卡奇卡参观他们正在那里建造的汽车工厂——当时还在建设中——他们把我介绍给该设施的负责人,经理,一个巨人。当然,对我来说每个人似乎都是巨人,但他真的是个巨人。他之前是福特的生产主管。所以我说:‘那么,福特和特斯拉有什么区别?’他说:‘嗯,当我在福特时,如果有变更,我必须经过六个不同层级才能批准变更。在这里,如果我需要变更,我只需走上楼,Elon 就在那里。’
So everything he has to do with speed. And then when I went to Texas to Boca Chica to visit their car factory that they were building there—while it was under construction—they introduced me to the head of that facility, the manager, a giant man. And of course, everyone seems like a giant to me, but he was a really giant guy. And he had been previously the head of production at Ford. And so I said, 'Well, so what's the difference between Ford and Tesla?' He said, 'Well, when I was at Ford and I had a change, I had to go through six different levels to get a change approved. When I have to have a change here, then all I do is walk upstairs and there's Elon.'
所以当工厂在建时,Elon 不是某个遥远的人。他在那里。他甚至有一辆 Airstream 房车。他住在 Airstream 里。你知道任何其他大公司的高管——可能是世界上最大的公司——会住在 Airstream 里,以确保在建造设施时出现的所有问题都能快速得到回答吗?所以这是惊人的文化,惊人的人,以及招募的惊人的人。在那里工作的人,他们就是聪明人。在那里找工作比进哈佛还难。当然,我不知道为什么人们现在还会去哈佛。
And so when a plant is under construction, Elon's not some distant guy. He's there. He even has an Airstream. He lives in an Airstream. Do you know any other executives of any big companies, potentially the largest company in the world, going to live in an Airstream to make sure that all the questions that are taking place in making a facility are going to be answered fast? So it's amazing culture, amazing person, and amazing people who's recruited. So the people who work there, they're just brilliant people. It's like harder to get a job there than get into Harvard. Of course, I don't know why people would go to Harvard anyway anymore.
是的。
Yeah.
嗯,但顺便说一句,Airstream 房车其实相当不错。所以我不会说我喜欢——不,没那么糟。老实说,我有点喜欢我的小 Airstream。但它实际上就停在我们建造的发电厂旁边——这个发电厂发电超过 1 吉瓦。而且可能达到 1.5 到 2 吉瓦的内部发电厂产量,大约只需通常建造时间的五分之一。所以大约一年建成,而通常需要大约五年。
Well, but by the way, Airstreams are actually pretty nice. So I wouldn't say I'm like—no, it's not that bad. Honestly, I kind of like my little Airstream. But it's actually parked right next to the power plant that we built—this power plant that's producing over a gigawatt. And probably it'll go to one and a half to two gigawatts of internal power plant production in about a fifth of the time that such a thing would normally be built. So built in about a year as opposed to normally about five years for that.
并不是有人必须上楼来找我。没有楼上。所以我们在孟菲斯的设施没有上层。我有时一天走遍整个场地两、三、四、五次。所以整个场地,发电厂和所有建筑。我还坚持至少亲自做一些手工工作。所以我拉过蟒蛇大小的电力电缆——它们非常重且长。我做过光纤布线。我做过电气开关。我参与过发电厂管理软件。我尝试至少做一点每样事情,这样我不会对实际做这些工作的人的痛苦麻木,并且我对困难以及可以做什么来改善情况有一些了解。所以并不是我坐在办公室里,有人必须来找我。我在现场走动,主动问他们问题。
And it's not as though somebody has to come to me upstairs. There is no upstairs. So we have no upper floor at our Memphis facility. I walk the entire site sometimes two, three, four, five times a day. So the entire site, the power plant and all the buildings. And I also insist on doing at least some of the work myself manually. So I've pulled power cables the size of an anaconda—they're incredibly heavy and long. And I've done fiber optic cabling. I've done electrical switches. I've worked on the power plant management software. I try to do at least a little bit of everything so that I'm not numb to the pain of the people who are actually doing it and I have some understanding of the difficulty and what can be done to improve the situation. So it's not as though I'm sitting in an office and somebody has to come to me. I'm walking the floor asking them questions actively.
太棒了。所以让我们谈谈现在新闻中的数据中心的,人们——
It's amazing. So let's talk a little bit about data centers which are now in the news and people—
是的,热门话题。
Yeah, popular topic.
如何交友和影响他人,你知道,数据中心——对数据中心的抵制程度相当高。
How to make friends and influence people, you know, data centers—the pitchfork level for data centers is pretty high.
是的,我会这么说。但如果你想要 AI,你就必须有数据中心。所以让我们先在这个星球上谈谈它们。
Yeah, I'd say so. But if you want to have AI, you got to have data centers. And so let's talk about them first on this planet.
所以 Elon 认为阻碍数据中心建设的障碍是拥有足够的芯片,实际上是足够的算力。所以他所做的是——很多人——我们有 Grok,所以他处于 OpenAI、Anthropic 的前沿——那些人是他的伙计。然后当然他不再喜欢 OpenAI 了,但是——
So Elon believed that the roadblock that would be in the way of data center construction would be having enough chips, enough compute actually. And so what he did is that—a lot of people—we have Grok, so he was at the forefront of OpenAI, of Anthropic—those guys are his guys. And then of course he doesn't like OpenAI anymore but—
是的,嗯,我确实将 OpenAI 创办为一个开源非营利组织,现在它却是一个为了最大利润而封闭的实体,所以它与我最初创办的完全相反。
Yeah, well I did start OpenAI as an open source nonprofit and now it is a closed for maximum profit entity so it's literally the polar opposite of what I started.
对。
Right.
然后 Anthropic 是由七个离开 OpenAI 的人创立的,因为他们不信任 Sam Altman,有充分的理由。
And then Anthropic was founded by seven people that left OpenAI because they didn't trust Sam Altman for good reason.
他们公开说过这件事,这不是秘密,所以实际上 Anthropic 是 OpenAI 的一个分支。所以是啊,我大概当初就该把 OpenAI 直接做成一家普通的 C 型公司,而不是非营利组织。但我当时以为这会是一件造福世界的公益事业。我没料到它有一天会变成一家闭源的营利性公司。
They said this publicly, it's not a secret, and so effectively Anthropic is an offshoot of OpenAI. So yeah, I guess I should have just started OpenAI as a regular C-corp instead of a nonprofit. But I thought this would be something that would be a philanthropic thing for the world. I didn't anticipate it would one day turn into a for-profit that is closed source.
然后你离开的时候,因为你对方向感到失望,他们给了你一份所有权权益,你说不,我不要,于是你就走了。
And when you left, because you were disappointed in the direction, they offered you an ownership interest and you said no, I don't want it, and so you left.
嗯,是的,我的意思是,我当时很困惑,一个非营利组织怎么会有商业利益。这在我看来说不通。公司初期几乎所有的资金都是我出的。这是我的主意,我给它起的名字,招募了核心人员,还把我所知的一切都教给了他们——但除此之外,我什么都没做。
Well yeah, I mean, I was confused as to how you could have a commercial interest in a nonprofit. That didn't seem to make sense to me. I provided almost all the capital in the beginning for the company. It was my idea and I named it and recruited the key people and taught them everything I know — but apart from that, nothing.
回到数据中心的话题。所以你以为,把所有这些 AI 投入运营的关键在于我们需要数据中心,而人们想建它们,通常建一个要花大约两年半,成本大概 600 亿美元——建一个吉瓦级要 500 或 600 亿美元。而那些为超大规模云厂商建好再出租的人,为了拿到资金,他们得借钱或以某种方式融资,这就意味着他们必须签长期租约,至少五年才能拿到钱。所以五年租约下人们付的租金,对 600 亿美元的投资来说大概是每年 120 亿美元。而我们不这么做,我们自己建。我们建是因为我们自己要用,因为我们是垂直整合的——为我们的车、我们的机器人、我们的火箭、为一切。而且因为我们需要它们,我们就能承担在没拿到别人合同的情况下先把它们建起来的风险。
So to go back to data centers. So you thought that the idea about putting all this AI in operation was that we need data centers, and people want to build them, and it takes them ordinarily about two and a half years to build one, and it costs maybe $60 billion to build it — 50 or 60 billion to build a gig. And then the people who build them for hyperscalers and rent them, in order for them to get the capital, they need to borrow it or get it in some way that they have to have long-term leases, five years at least, to get them. So the rents people pay in five-year leases are maybe $12 billion a year for a $60 billion investment. What we do instead is that we build them. And we build them because we have the use for them, because we're vertically integrated — for our cars, for our robots, for our rockets, for everything. And because we need them, then we can take the risk of building them without knowing that we have contracts for other people.
但实际情况是,我们建好了,而它们——他们建要两年半,我们建只要 8 个月、9 个月;他们花 600 亿,我们花 350 亿。而算力需求如此之大,那些建了数据中心的人,机房是黑的,因为他们没电把它开起来。我们会把算力租出去,直到我们自己需要用它来跑 Grok,我们会以每年大约 500 亿美元的价格租给他们。所以我们用 350 亿美元的投资,每年拿到 500 亿美元。所以 SpaceX 的营收即将爆发式增长。往上走。
And what happens though is we have them built and they're not — so it takes them — it's taken us four months. It has now taken us eight or nine months to build. So it takes them two and a half years, takes us 8 months, 9 months; cost them 60 billion, cost us 35 billion. And the demand for compute is so great that people who built centers, they're dark because they don't have the power to turn them on. And we'll rent them to compute until we need it for ourselves for Grok, and we'll rent it to them for somewhere around 50 billion a year. So we're getting 50 billion a year for a $35 billion investment. So the revenues for SpaceX are about to explode. Going up.
是的,它们在爆发式增长。
Yes, they are exploding.
我觉得是的,我们今年第二季度到第三季度的营收增长,我想超过了 80%。我不确定第三季度到第四季度的确切数字会是多少,但大概在 70% 左右。第四季度到第一季度可能也差不多。要大规模地把算力上线是极其困难的。我认为外面有很多投机者,以为你只要订一些 GPU,把采购订单交给某家公司去建数据中心,然后就能坐收现金。而现实是,这难到疯狂。
It's — I think yeah, our Q2 to Q3 revenue increase was, I think, over 80% this year. I'm not sure what the exact numbers will be Q3 to Q4, but probably something like 70% Q3 to Q4. And I would say maybe it's something similar, you know, Q4 to Q1. It is extraordinarily difficult to bring on compute at scale. And I think there's a lot of speculators out there who think you can simply order some GPUs, hand a purchase order to some company to build the data center, and just rake in cash. And the reality is it's insanely hard.
有个梗我觉得挺好笑,好像叫 Jose 梗,你可能见过——一个人在路边挖坑,旁边有八个人在看他挖。所以它的意思是,只有一个人在干实际的活,然后八个人在看这个人干活。那么,除非你增加这个国家里 Jose 的数量,否则完成的工作量并不会增加。所以增加更多只会给 Jose 下采购订单的人,并不会让更多事情发生——只有更多 Jose,才有更多工作。所以我立志做 Jose,SpaceX 也想做 Jose。我们干实际的活。这非常关键。
So there's a meme that I think is quite funny, which I think it's called the Jose meme, where — you may have seen it before — there's one guy digging the hole in the road and there's like eight people watching him dig the hole. So now in that meaning, like there's one person doing actual work and then eight people watching the guy do the work. Now, unless you increase the number of Joses in the country, then the amount of work done does not increase. And so adding more people who simply issue purchase orders to Jose does not result in more things occurring — only if there are more Joses, more work. So I aspire to be Jose, and SpaceX wants to be Jose. We do actual work. This is a really big deal.
而随着美国大规模扩张和建设,坦率地说,从事实际体力工作的人并没有大量涌入,但下采购订单的人却大幅增加。这意味着真正干活的人被摊得极其稀薄。所以全国各地有大量半途而废的数据中心项目,因为到这个时候,花生酱抹得太薄,你几乎尝不出味道了。
And with the massive expansion and construction in the United States, frankly it's not as though there's been this massive influx of people who do actual work at a physical level, but there has been a massive increase in the number of people who issue purchase orders. So that means that the people doing actual work are spread incredibly thin. And so you have a large number of partially completed data center projects throughout the country, because at this point the peanut butter is spread so thin you can barely taste it.
不同之处在于,在 SpaceX,我们确实习惯于做极其艰巨的实体工程,比如造巨型火箭和庞大的发射场。特斯拉也非常擅长大规模建设,而地球上除中国之外,最擅长实体硬件的两家公司大概就是 SpaceX 和特斯拉。我甚至不知道第三名会是谁。我觉得就算把中国算进来,特斯拉可能也是世界最好的,而中国已经非常厉害了。值得一提的是,特斯拉 Model Y 是中国最畅销的电动车。
The difference is that at SpaceX we actually are used to doing incredibly challenging physical work in terms of building giant rockets and vast launch complexes. And Tesla is also very good at large-scale construction, and probably the two best companies at physical hardware on Earth outside of China are SpaceX and Tesla. I don't even know who would be number three. And I think even Tesla might be the best in the world, even including China, where China is incredibly good. It's worth noting that the Tesla Model Y is the bestselling EV in China.
嗯,你在中国教了他们很多。
Well, you taught them a lot in China.
这就像在亚洲打赢一场陆战——出了名的难。但我们实际上差不多做到了,Model Y 是中国最畅销的车。你想想看。而且我告诉你,中国的车企超级强。
This is like winning a land war in Asia — classically difficult. And yet we kind of are, actually, in that Model Y is the bestselling car in China. Think about that for a second. And let me tell you, Chinese car companies are super good.
是啊。福特的那个人去了那里。他说他被中国汽车的质量震撼到了。
Yeah. The person from Ford went there. He said he was blown away by the quality of the cars in China.
是的,部分是这样。中国很棒。顺便说一句,我爱中国。去那里很受启发。
Yes, partly. China's awesome. I love China, by the way. It's inspiring to go there.
嗯,大多数在中国建厂的人,都必须让中方持有中国业务相当大一部分的所有权。
Well, most people who build in China, they have to have Chinese ownership of a significant portion of the operation in China.
在中国通常你只被允许持股 49%。而特斯拉——他们太想让我们去,所以允许我们 100% 持股。
49% is what you're allowed to own in China normally. And Tesla — they wanted us to be there so badly that they let us go 100% ownership for us.
是的。
Yes.
那花了好几年时间跟中国官员会面。实际上,政府里有些人真的冲我吼,唾沫横飞那种,说我们必须跟一个中国伙伴“结婚”。我说,我们只是个婴儿公司,我觉得婴儿不该结婚,因为抱着婴儿走红毯看起来会很滑稽。他们说,你不是婴儿。我说,不,我们是婴儿,婴儿就是不该结婚,太年轻了。所以最后他们说,好吧,行。
That took quite a few years of meeting with Chinese officials. And I actually — some of the people in the government actually literally yelled at me, like spitting type of thing, that we had to get married to a Chinese partner. And I'm like, well, we're just a baby company and I don't think babies should get married, because that would look funny going down the aisle with a baby. And they're like, you're not a baby. And I'm like, no, we're a baby, and babies should just not get married. It's too young. So then eventually they said, okay, fine.
我们会修改法律,然后你就可以拥有工厂。而大多数时候,我们只是需要能够快速行动、做出决策。如果你总是要问别人才能做决定,那你怎么搞,你知道,如果你实际上没有握着方向盘,就很难创新。嗯,如今因为自动驾驶,你不需要手握方向盘了,但这个比喻在未来就不成立了。
We'll change the law and then you can own the factory. And mostly it was like we just need to be able to move fast, you know, and make decisions. And if you've got to always ask someone else for permission to do the decisions, then how you going, you know, it's difficult to innovate if you don't actually have your hands on the steering wheel. Well, these days you don't need your hands on the steering wheel because of self-driving, but that metaphor is not going to make any sense in the future.
所以我们说的是 SpaceX 的收入会大幅增长。我记得去年是 150 亿或 160 亿,而今年到 26 年底将达到约 1000 亿美元的年化收入。然后明年我们认为会在 5000 亿到 1 万亿美元之间。所以我们从 150 亿增长到大约 1000 亿年化,再到 5000 亿到 1 万亿。而这真的只是刚刚开始。而真正酷的是,把我们的数据中心出租给竞争对手,基本上他们为我们花了 350 亿的东西支付 500 亿,而且我们还会因为出租我们仍然拥有的数据中心而获得源源不断的现金收入。所以我们相当于免费得到了它们,对吧?
So we're talking about revenues going up a lot for SpaceX. It was I think 15 or 16 billion last year and this year it's going to be ending the year 26 at a hundred billion dollars or so of annualized revenues. And then next year we think it's going to be between 500 billion and a trillion. So we've gone from 15 billion to somewhere around a hundred billion annualized to 500 billion to a trillion. And that's just really just getting started. And what's really cool about selling renting renting our data centers to competitors that basically they're paying us 50 billion for something cost us 35 and we're going to have collections of cash coming in for renting data centers that we still own. So we got them for free, right?
嗯,我是说,算是吧。我想就其中几点评论一下。所以我们正朝着在 12 月实现超过 1000 亿美元的年化收入目标前进。显然 12 月已经很近了,所以我有充分的动力。除非发生某种自然灾害之类的事情,否则我们会达到那个目标。然后根据我们明年的计划,看起来很有可能我们明年 12 月的年度经常性收入将超过 5000 亿美元。现在这只是我的,这不是承诺。这只是我认为会发生的事情。这是我最好的猜测。我只是想区分这是对做某事的承诺,还是我认为会发生的事情。我认为会发生的是明年 12 月的 ARR 达到 5000 亿美元。而我只是说 1 万亿并非不可能。意思是以前是不可能的,现在不再是了。可能性大概是 2% 左右。但有趣的是它不是 0%。这才是重点。而且概率会随时间增加。所以我的生物神经网络对数字超级智能对全球经济影响的预测是大约增长 20% 到 30%,这是在不考虑机器人技术的情况下。我认为大约增长 20% 到 30%,这大致相当于全球经济增加 25 万亿美元。而且这会非常快地发生。
Well, I mean, sort of. I think just to comment on a few of those things. So we are tracking to have an annual rate of revenue that is in excess of a hundred billion in December. Obviously December is pretty close so I have good incentive. So unless there's some sort of natural disaster or something we will meet that target. And then based on the plans that we have next year, it seems quite likely that our annual recurring revenue in December of next year will exceed 500 billion. Now this is just my, this is not a promise. This is just what I think will happen. This is my best guess. I just sort of want to distinguish between this is like a promise to do something versus what I think will occur. What I think will occur is an ARR in December of next year of 500 billion. And I simply said a trillion is not out of the question. Meaning it was previously out of the question and it is no longer out of the question. It's maybe 2% or something like likelihood. But it's interesting that it's not 0%. That's the interesting part. And the probability is increasing with time. So my, as a sort of what my biological neural net predicts for digital super intelligence in terms of its effect on the global economy is about a 20 to 30% increase, this is in the absence of taking robotics into account. I think around 20 to 30% increase which would be roughly call it a $25 trillion increase in the global economy. And this will happen very fast.
所以那是 130 万亿美元的全球经济。你认为数字 AI 会使其增长 25% 到 30%。
So that's $130 trillion global economy. And you think that digital AI is going to increase that by 25 30%.
20% 到 30%。是的,这是我的猜测。
20 to 30%. Yes, that's my guess.
好的。那么,你猜它会增加多少经济?
Okay. So well, what's your guess for how much would increase the economy?
我想大概是 20% 到 30%。我的意思是,我认为可能会更高。但它可能会使白领工作的生产力翻倍。
I think probably 20 to 30%. I mean I think like and it could be higher than that. But it probably doubles the productivity of white collar work.
所以那只是数字的。不是物理的。那是数字的。数字 AI。
So that's just digital. That's not physical. That's digital. Digital AI.
物理方面是我们将会占据主导地位的。我认为。
Physical is something that we're going to be dominant in. I think.
是的。嗯,思考以机器人形式存在的物理 AI 的方式是,任何给定的经济体都是其中的人数乘以人均平均生产力。所以如果你有一个极其高效的人形机器人,可以 24/7 工作。所以它每周工作 168 小时。它不度假、不上厕所、不需要吃东西等等。而且它的运行速度比人类快。你可以预期每周生产力大约是一个人的五倍,因为一个人可能每周工作 40 小时,显然需要休息,而且移动速度不如机器人快。所以我说人形机器人的最低生产力预期是人类的五倍,然后,比如说拿美国人口来说,有多少人在上学或退休,嗯,大约一半的人,大约一半的人在上学或退休,然后在工作年龄的人中,不是每个人都工作或从事高生产力就业,所以剩下的大概,我不知道,美国有 1.2 亿人具有净正的商品和服务产出。那么你说,需要多少机器人才能匹配那个生产力?嗯,我猜你需要大约 2000 万到 3000 万个机器人来匹配美国的商品和服务产出。而我们会建造那些。建造机器人比建造汽车便宜。
Yes. Well, the way to think about physical AI in the form of robotics is that any given economy is the number of people in it times the average productivity per person. So if you have a humanoid robot that is extremely productive and can work 24/7. So it's working 168 hours a week. It doesn't take vacation or bathroom breaks or need to eat food or anything like that. And it's the speed at which you can operate is faster than human speed. You can expect roughly a five, a productivity per week that is on average five times better than a person roughly because a person's like maybe working 40 hours a week and does need to take breaks obviously and doesn't move as fast as a robot. So I say a minimum productivity expectation for a humanoid robot would be five times that of a human then of the, say take the population of the United States say how many people are either in school or retired well that's about half of the people about half of people are either in school or retired and then of those who are working age not everyone works or is engaged in highly productive employment so then you're left with perhaps around I don't know 120 million people in America that would have a net positive output of goods and services. So then you say well how many robots then would be needed to match that productivity? Well, I guess you'd need about 20 to 30 million robots to match the goods and services output of the United States. And we would build those. It's less expensive to build a robot than to build a car.
是的。最初,我会说差不多。但随着时间的推移,会少得多。
Yes. Initially, I'd say about the same. But then over time, it would be much less.
所以一个机器人的成本是 3 万,朝着 2 万迈进。
So 30,000 cost for a robot on the way to 20.
是的。我们会出租它们还是出售它们,或者我们会怎么做?
Yeah. And we would rent those or we would sell them or what would we do?
我们有想法了吗?我想大概是出租它们。
Do we have an idea yet? I think probably rent them.
继续关于数据中心的话题。那么如果人们真的对社区里的数据中心感到不满,既然这些对我们来说如此有利可图,有没有办法我们直接说‘好吧,你们社区的公用事业账单将保持不变,我们会补贴它,我们会利用那个补贴,我们还会给你们 5000 万美元用于学校和医院。’我们有没有这样的项目,我们去社区说,是的,我们会在你们的社区帮助你们,我们会给你们更多工作。我们真的是一家创造就业的公司。没有人像我们这样创造就业。有没有任何方式,我不认为我们做过,我不认为我们的行业在解释这是对人们有益而非有害方面做得很好。
To stay on the data centers. Then if people are really upset about data centers in their communities, is there any way that since these are so profitable for us? Why wouldn't we just say, 'Okay, here your utility bills in your community will stay flat and we'll subsidize it and we will use that subsidy and we will also give you $50 million for your schools and hospitals.' Do we have any programs like that where we have gone to communities and say yes, we are going to help you in your communities and we'll give you more jobs. We're really a job creating company. No one creates jobs the way we do. Is there anyone any way I don't think we've done I don't think our industry's done a good job about explaining how this is the benefit of people not harming them.
是的。嗯,实际上在孟菲斯地区,我认为我们实际上相当受欢迎。奇怪的是,受欢迎的地方是有数据中心的地方,而不受欢迎的地方是没有数据中心的地方,他们认为数据中心会有坏影响。但在孟菲斯,你知道孟菲斯因为 SpaceX 而繁荣。餐馆满座,酒店满房。我们有时支付是人们之前收入的两倍。
Yeah. Well, actually locally in the Memphis area I think we're actually quite popular. It's actually people it's oddly the popularity is there where the lack of popularity is where there aren't data centers and they think the data center will have a bad effect. But in Memphis you know Memphis is booming because of SpaceX. The restaurants are full, the hotels are full. We're paying at times twice the prevailing wage that people were earning before.
为什么人们不写这个?为什么没有关于这个的报道?
Why don't people write about this? Why aren't there stories about this?
有一些报道会影响到——我们也帮助了当地学校,县和市获得的税收是巨大的。所以我认为这种受欢迎程度的问题只是暂时的担忧,因为数据中心确实带来了巨大的繁荣。如果你和镇上的人交谈,他们都能感受到,因为他们的生意兴隆。
There are some stories that affect— We've also helped out with local schools, and the tax revenue that the county and the city are earning is tremendous. So I think this popularity thing is a temporary concern because there actually is tremendous prosperity that's brought by the data centers. And if you talk to the people in town, they all feel it because their businesses are booming.
那么就去太空建数据中心吧,这显然是解决方案,这样人们就不必再担心它们出现在自己的社区里了,如果他们不想的话。
So then go to data centers in space, the obvious solution, so people don't have to worry about them in their communities anymore if they don't want to.
对我们来说,把它们放在太空更便宜。而且每次我们推出新一代,它们都会比之前更强大。
And for us, it's less expensive to put them in space. And every time we put up new generations, they become more and more powerful than they had been before.
这不是——大家都说,“嗯,那还得等 10 年。”而埃隆说,“不,我们已经在开始了。谷歌已经用我们为他们运行的卫星进行了测试。效果很好,所以我们将在 27 或 28 年开始部署它们。”
Isn't that— And everyone says, "Well, it's 10 years away." And Elon says, "No, we're starting. Google has already tested with our satellite that we've run for them. It's been effective and so we're going to be starting to put them up in '27 or '28."
多久能产生影响?
How long before impact?
是的。所以,目前人们已经测试了 NVIDIA GPU。我们发射了一颗搭载 Nvidia GPU 的卫星,最近还发射了一颗搭载谷歌 TPU 的卫星,据我所知,它们在太空运行良好。我们预计将开始发射首批 StarMind 计算机,这是我们与 Nvidia 共同设计的,可能——我只是猜测——在明年第三季度。我们会从慢开始,但随后可能在 28 年提升到可观的量,然后在 29 年达到非常大的量。这确实——为了达到每年太瓦级甚至可能每年 200 吉瓦级的规模,你真的需要去太空,因为地面上的电力根本不足以现实地达到那些年度水平。作为一个参考,美国的平均电力消耗略高于 500 吉瓦。所以如果你看美国的峰值功率能力,大约是 1.2 太瓦,但平均电力消耗略高于 500 吉瓦。所以你可以看到,如果你每年做 200 吉瓦的 AI,那相当于每年整个美国电网增加 40%,这根本不可行。而当你把太阳能板放到太空,因为太空基本上总是阳光明媚,太阳能板可以始终直接面向太阳——它们不会与太阳成角度——而且没有大气衰减,所以你在太空能获得多 30% 的太阳能。一直阳光明媚,没有季节性。你不需要——你可以用非常小的电池组,因为你不需要撑过夜晚。一旦每吨入轨成本降低,太空中的智能成本就会大大低于地面。所以星舰,因为它将是第一枚完全可重复使用的火箭,将使进入太空的成本降低 100 倍或更多。我认为实际上极限情况下——可能在两三年内——每吨入轨的成本将低于从加利福尼亚到澳大利亚的空运成本。
Yeah. So, yes, there's currently— people have tested NVIDIA GPUs. So there's a satellite with an Nvidia GPU up there that we launched, and there's a satellite we recently launched with a Google TPU, and these are operating—to the best of my knowledge, they're working fine in space. We expect to start launching the first StarMind computers, which we've co-designed with Nvidia, probably—I'm just guessing here—but in the third quarter of next year. And we'll start off slow, but then it will ramp to significant volume probably in '28 and then very, very large volume in '29. And this really—in order to get to the terawatt per year scale or even, I'd say, maybe the 200 gigawatt per year scale, you really need to go to space because there just isn't enough power on the ground to realistically get to those levels per year. Just as a frame of reference, the average power consumption in the United States is just over 500 GW. So if you look at the peak power capable in the United States, it's roughly 1.2 terawatts, but average power consumption is just over 500 gigawatts. So you can really see that if you're doing say 200 gigawatts a year of AI, that would be like doing a 40% increase in the entire US power grid every year, which is simply not feasible. And when you take solar panels and you put them in space, because it's essentially always sunny in space and the solar panels can always face the sun directly—they're not off angle to the sun—and you don't have atmospheric attenuation of solar power, so you get 30% more solar power in space. It's sunny all the time. There's no seasonality. You don't really need—you can get away with a very tiny battery pack because you don't need to last through the night. The cost of intelligence in space is dramatically less than on the ground once you have a low cost per ton to orbit. So Starship, because it will be the first fully reusable rocket, will lower the cost of access to space by a factor of 100 or maybe more. I think actually in the limit—maybe within probably 2 or 3 years—the cost of a ton to orbit will be less than the cost of air freight from California to Australia.
所以 NASA 说 2010 年进入轨道的成本是每公斤 18,500 美元。那是因为火箭只用一次。然后当我们发明了可以反复使用的猎鹰 9 号火箭,成本降到了我认为是 2,500 美元,然后降到 1,500 美元
So NASA said that in 2010 the cost to get to orbit was $18,500 a kilogram. That was because you use a rocket one time. And then when we came up with a rocket Falcon 9 you could use over and over again, it became I think it was $2,500 and then it went to 1500
现在它将降到 100 或 15 或 50。我认为我们能达到每公斤低于 100 美元入轨。所以现在重要的是要认识到完全可重复使用是多么深刻。对于猎鹰 9 号,我们大概有 80% 的可重复使用性。助推器回来着陆。你可能看过猎鹰 9 号助推器着陆的视频,然后包含卫星的整流罩或鼻锥——那些整流罩 halves 通过降落伞落在海洋中,我们重新飞行那些,但我们不重新飞行上级。所以我们无法回收猎鹰 9 号的上级,而上级的成本大约是一架中型喷气机的成本。所以这意味着猎鹰 9 号每次飞行的最低成本是一架中型喷气机。现在,这仍然比其他任何东西好得多。但可以肯定地说,如果每次你飞到某个地方,你都必须扔掉一架中型喷气机,那就不会有太多的航空旅行。你会说,我想我还是开车吧。所以现在有了星舰,整个系统都是为了可重复使用而设计的。不仅仅是可重复使用,而是极其快速的可重复使用。用塔臂——我们称之为筷子——抓住助推器和飞船的原因,有点像《空手道小子》里的宫城先生抓苍蝇,只是把它变大。原因是火箭从发射台起飞。火箭由塔臂放置在发射台上,然后当它从太空返回时,被同样的塔臂抓住并放回发射台。所以你不仅有完全可重复使用性,还有极其快速的可重复使用性。你不可能有比降落在起飞地点更快的可重复使用性。原则上,这将允许我们每小时重新飞行火箭,巨大的助推器,超重型助推器,
and now it's going to go to 100 or 15 or 50. I think we get to under $100 a kilogram to orbit. So now it's important to just appreciate how profound full reusability is. So with Falcon 9 we have, I guess, probably around 80% reusability. So the booster comes back and lands. You've probably seen the videos of the Falcon 9 booster landing, and then the fairing or nose cone that contains the satellites—those fairing halves land by parachute in the ocean and we refly those, but we do not refly the upper stage. So we're unable to recover the upper stage of Falcon 9, and the cost of the upper stage is roughly the cost of a medium-sized jet. So that means the minimum cost per flight for Falcon 9 is a medium-sized jet. Now, that's still much better than anything else. But it's safe to say that there wouldn't be a lot of air travel if every time you flew somewhere, you had to throw away a medium-sized jet. You'd be like, I think I'll drive. So now with Starship, the entire system is designed for reusability. And not just reusability but extremely rapid reusability. The reason to catch the booster and the ship with the tower arms—we call them the chopsticks—it's kind of like Mr. Miyagi in Karate Kid catching the fly, but just make it bigger. And the reason for that is that the rocket takes off from the launch pad. The rocket is placed on the launch pad by the tower arm and then it is caught by those same tower arms and placed back on the launch pad when it returns from space. So you don't just have full reusability, you have extremely rapid reusability. You really couldn't get faster reusability than landing back where you took off. And this in principle would allow us to refly the rocket, the giant booster, the Super Heavy booster, once an hour,
每年 10,000 次发射。
10,000 launches a year.
是的,最终可能更多,但是的,每小时超过一次发射。所以当我年轻的时候,人们会排队观看每一次——现在很少看火箭。而我们可以每小时一次。甚至不再上报纸了。
Yes, probably more in the end, but yes, more than a launch an hour. So when I was young, people would line up to watch every—now very rarely a rocket. And we can do one an hour. Doesn't even make newspapers anymore.
是的。嗯,这将非常——我是说,你真的会想看这个。这个很酷。
Yes. Well, this will be quite—I mean, you really want to watch this one. This one's cool.
它是有史以来最大的飞行物体,也是人类制造的任何种类中推力最大的物体。所以是的,这东西在起飞时重超过 5,000 吨,它直直地快速上升。然后那个巨大的东西回来,从太空飞回来,就像起飞后 5 分钟,助推器以某种方式下来。希望是以好的方式,但它以某种方式下来。所以兴奋是保证的。成功不总是保证的,但兴奋肯定是。
It's the largest flying object ever made and it's the highest thrust object ever made by humans of any kind whatsoever. So yeah, this thing weighs over 5,000 tons on liftoff and it's going straight up, fast. And then that giant thing comes back, flies back from space, like 5 minutes after liftoff that booster is coming down one way or another. It's going to, hopefully in a good way, but it's coming down one way or another. So excitement is guaranteed. Success is not always guaranteed, but excitement certainly is.
当你看到它返回并降落在起飞的地方时,那真是太酷了。
It's very cool when you see it come back and land in the same place where it took off.
是的。
Yeah.
筷子。
Chopsticks.
是的。看起来太疯狂了。当我第一次向团队提出这个想法时,他们看我的眼神就像我疯了一样。但我说,看,我认为这是可以做到的。从物理学的角度来看,没有什么能阻止它发生。我能想到的唯一比用塔臂捕获稍微好一点的方法,就是直接降落在发射台上。但那样的话,如果出了差错,你就会损坏发射台,修复起来需要大量工作。所以,如果我们只用机械臂捕获,那会好一点,因为如果出了问题,我们不一定会砸坏发射台。但用机械臂捕获和直接降落在它起飞的发射台上,差别只有几分钟。
Yeah. It looks crazy. When I first proposed this to the team, they looked at me like I had lost my mind. But I'm like, look, I think this can be done. There's nothing that's preventing it from happening from a physics standpoint. The only thing I can think of that would be slightly better than catching it with the tower arms would be literally landing on the launch pad itself. But then if you get something wrong, you damage the launch pad and that would take a lot of work to fix. So if we just catch with the arms, that's a little better because we don't smash the launch pad necessarily if something goes wrong. But it's a difference of minutes between catching with the arms and simply landing on the actual launch pad it took off from.
现在,比助推器更难的是捕获飞船。而且飞船,我们必须还要做很多以前从未做过的事情。从来没有人制造过真正可重复使用的轨道热防护罩。航天飞机的热防护罩在每次飞行之间需要 9 个月的维修。所以,尽管航天飞机的轨道器部分确实返回了,但他们在每次飞行中都会丢失主燃料箱,那是主要的机身或主要结构,所有东西都附着在上面。但他们能回收轨道器,只是需要 9 个月的热防护罩翻新才能再次飞行。所以,从来没有人制造过完全可重复使用的轨道热防护罩。这是一件极其困难的事情。这是我们十年来一直在努力的事情。基本上,它太难了。我认为这是人类尝试解决的最困难的技术问题之一。但我们有一支非常有才华的团队,他们非常努力地工作了很长时间,我们能够迭代。回到迭代。我们能够进行许多次迭代,而航天飞机无法真正迭代。航天飞机上总是有人,所以他们不能冒任何风险,因为可能会有人丧生。而对于星舰,我们承担了许多风险,以探索实现热防护罩和整个飞行器完全可重复使用所需的物理原理。
Now the much harder than the booster is catching the ship. And the ship we've got to have also there's a lot of things that have to be done that have never been done before. No one has ever made a truly reusable orbital heat shield. The shuttle heat shield required 9 months of repair between flights. So even though the shuttle orbiter portion did come back, they'd lose the main tank which is the primary airframe or the primary structure really to which everything attaches on every flight. But they'd get the orbiter back but it would take 9 months of refurbishment of the heat shield before they could fly again. So no one has ever made a fully reusable orbital heat shield. It's an incredibly difficult thing to do. And it's something that we've been working on for a decade. Basically, it's so hard. One of the hardest technical problems I think humans have ever tried to solve. But we've got an incredibly talented team that has worked incredibly hard for a very long time and we've been able to iterate. Go back to iteration. We've been able to do many iterations whereas the shuttle couldn't really iterate. They always have people on board the shuttle, so they couldn't take any risks because somebody might get killed. Whereas with Starship, we're taking many risks in order to explore the physics of what's needed to achieve full reusability of the heat shield and the entire vehicle.
所以我们实际上已经多次在海上实现了飞船的精确软着陆。如果那里有塔架,飞船就会被捕获。所以这也许是一个有趣的点:如果飞船以为它正在与模拟塔架着陆,而那里真的有塔架,它就会被捕获。现在我们想非常确保,当我们把它带回实际的塔架时,一切都会正常,这样我们就不会砸坏塔架,也不会让碎片像雨一样落在人们头上,那肯定会影响我们的受欢迎程度。所以我们想对飞船的可靠性有高度信心。我认为我们终于到了认为值得用 42 号飞船尝试的地步。希望这是有预言性的。你知道,42 是一个特殊的数字。我们可以尝试用塔架捕获。所以这就是我们大约一周后要做的事情。我们都支持它。
So we've actually achieved precise soft landings of the ship several times out at sea in the ocean. Where the ship, if there had been a tower there, the ship would have been caught. So that's maybe an interesting point to make that if the ship thought it was landing with a simulated tower and had a tower been there it would have been caught. Now we want to make very sure that everything's going to work when we bring it back to the actual tower so that we don't smash the tower or rain debris on people's heads which will definitely affect our popularity. So we want to be highly confident in the reliability of the ship. And I think we're finally there where we think it is worth with ship 42. Which hopefully is prophetic. You know, 42 is kind of a special number. That we can try to catch up with the tower. So that's what we're going to do in about a week from now. We're all rooting for it.
所以星舰是这里的关键突破。顺便说一句,我注意到那个负责星舰运营的年轻人,他当时对着一群穿着运动夹克、西装和白衬衫的观众讲话,而他上台时穿着工程师靴、蓝色牛仔裤、格子衬衫,没刮胡子,头发很长,他开始讲话,他说‘这个 f,那个 f,这个 f’。然后我看着他,我说,‘天哪,那是我 40 岁时的样子。真是个好人。’
So starship is the great unlock here. Just as an aside, I was watching the young person who was heading the starship operation and he was speaking to a group of people in the audience wearing sport jackets and suits and white shirts and he came on and he was wearing engineer boots and blue jeans and a plaid shirt and unshaven and really long hair and he starts talking and he said, 'F this and f that and f this.' And then I'm looking at him, I'm saying, 'Oh my god, that's me when I was 40. Really good guy.'
是的,骑摩托车。
Yeah, drives motorcycles.
所以这是巨大的突破,但埃隆做的所有事情,一开始你可能不理解,然后你会说,‘哦,哇,我明白了。他可能一直都在考虑那个。’所以当你想到可重复使用的火箭时,你知道,当谈到进入太空的成本如此之高,只有那些能进入太空的人,那是 18,000 美元,唯一能进入太空的是政府,为了保护他们的国家,或者可能是一家电视公司,花费 20 亿美元购买一颗卫星,以便向一个地方广播,而不是其他地方。所以,通过如此大幅地降低成本,突然间你使星链成为可能,星链现在有 10,000 颗卫星,正在向 20,000 颗迈进,这是我以前的想法,现在他说不,不会是 20,000,而是 100,000。为什么是 100,000?因为过去我们认为星链产生的收入将来自消费者、政府和企业。但现在他说,智能体式(agentic)将是之前想象的好几倍,而且每次你
So this is the great unlock and but all the things that Elon does, we might not make sense to you at the very beginning when you hear about them and then you say, 'Oh, wow. I see. He probably was thinking about that all along.' And so when you think about a reusable rocket, you know, when you talk about so much cost to get to space, the only ones who can go to space that was $18,000, the only ones who go to space are governments to protect their countries or it can be a television company who spends a $2 billion on a satellite so it broadcasts to one place, no one else. And so at bringing down the cost so dramatically then all of a sudden you enable Starlink and Starlink you have 10,000 satellites now on the way to 20,000 and that's what I thought and now the says no it's not going to be 20,000 it's going to be 100,000. Why is it going to be 100,000? because we thought in the past that the revenues he'd be generating from Starlink were going to be from consumers and government and enterprise. And now though he says that agentic is going to be multiples of what we thought about before and every time you
哦,是的。是的。嗯,我的理解是,实际上智能体式互联网流量已经超过了人类流量,如果我读到的 Cloudflare 数据正确的话。而且我相信 Cloudflare 预测,在 5 年内,智能体式流量将比人类流量多一千倍。如果有什么的话,我认为那还低估了智能体式 AI 流量的数量。
Oh yeah. Yeah. Well, my understanding is that actually agentic internet traffic already exceeds human traffic if I read that correctly from Cloudflare. And I believe Cloudflare predicted that in 5 years agentic traffic will be a thousand times more than human traffic. And if anything I think that is underestimating the amount of agentic AI traffic.
一千倍在多少年内?五年。
1,000 times in how many years? Five.
是的。
Yes.
所以我们一直认为星链是一个可以在 10 年内实现 1 万亿美元收入、7000 或 8000 亿美元 EBITDA、价值 14 或 15 万亿美元的业务。整个 SpaceX 现在价值不到 2 万亿美元。所以那是一个机会。但现在我认为我们可能真的低估了 SpaceX 和星链的价值。但因为你现在增加了这么多卫星,我们现在有了太空中心数据中心。
So we've been thinking about Starlink as being a business that could do a trillion dollars in revenues and make seven or $800 billion in EBITDA in 10 years and be worth 14 or 15 trillion. All of SpaceX is now worth less than two trillion. So that was one opportunity. But now I think we're probably really light on how much SpaceX how much Starlink can be worth. But because you're adding so many more satellites now we got space centers data centers.
是的。是的,我的意思是,按照任何正常商业标准,星链的收入增长非常迅速。你知道,在大规模下,你大约每年收入翻倍。而且你知道,我记得那曾经是件大事。不久之前,每年收入翻倍、高毛利率和正现金流通常是一件令人印象深刻的事情。尽管我不得不说,在 AI 领域,它确实超过了那个。所以在 AI 领域,你有每年数量级增长的潜力。
Yeah. Yeah, I mean Starlink revenue is growing very rapidly by any sort of normal business standards. You know, like you're roughly doubling revenues per year at large scale. And you know, I remember when that used to be a big deal. It wasn't that long ago that you know, doubling revenues with high gross margin and positive cash flow every year would normally be an impressive thing. Although I have to say with AI, it does exceed that. So with AI you have order of magnitude per year growth potential.
如果你想想我刚才说的那些话,我讲了一堆听起来极其疯狂的事情,但我认为我们终会实现。有时候我在时间框架上有点乐观,但最终我总是能兑现。另外,我并不总是迟到。媒体总是报道我迟到的时候,却从不报道我提前的时候。实际上我经常提前。媒体不写正面文章;他们忽略我提前的时候,只印我迟到的时候。所以看起来我总是在迟到,但事实并非如此。我目标是让我的预测处于第 50 百分位,也就是说,就预测的时间框架而言,它们基本上是一半对一半错,而事情的发展也大致如此。
If you think about the things I just said, I say a bunch of things that sound extremely crazy, but I think we'll get there. Sometimes I'm a little optimistic on the time frame, but I always deliver in the end. Also, I'm not always late. The press always reports on when I'm late, but they never report on when I'm early. Actually, I'm early quite often. The press doesn't write positive articles; they ignore the times I'm early and only print the times I'm late. So it seems like I'm always late, but that's actually not the case. I aim for my predictions to be at the 50th percentile, which means they are basically half right with respect to the predicted time frame, and that's about how things have turned out.
我们现在有 11,000 颗卫星在轨,主要是第二版卫星。我们刚刚开始发射第三版卫星。第三版的能力是第二版的 10 倍以上,我们将发射 10 倍数量的卫星。所以 Starlink 在——我不知道——两三年内产生的带宽将是目前的 100 倍。你说每吉比特每秒的收入是多少?会很高。然后你说你传输多少吉比特?那是个非常大的数字。而来自 AI 和机器人的巨大需求远远超过人类需求,大约是一千倍。仅通信收入的起点就会高得惊人。但如果你看 AI 部分,如果我们每年向太空发送 100 吉瓦的 AI 算力,那么每吉瓦算力的价值是多少?即使降到 10 美元,那每年也会多出 1 万亿美元。累积起来,每年都会增加 1 万亿美元,因为卫星不会蒸发。所以你每年都会堆叠 1 万亿美元的收入,每五年就是 10 万亿美元。如果按 50 美元算,那就是 5 万亿美元。
We have 11,000 satellites up there right now, primarily the version two satellite. We've just started launching the version three satellite. Version three has over 10 times the capability of version two and we'll be launching 10 times as many of them. So the amount of bandwidth that Starlink will produce in, I don't know, two or three years will be 100 times more bandwidth than it currently offers. You say, what's the revenue per gigabit per second? It's going to be high. And then you say, how many gigabits are you transferring? It's a very big number. And with massive demand from AI and robots, that far exceeds human demand by like a thousand. Just the starting communications revenue will be something staggeringly high. But if you look at the AI stuff, if we do say 100 gigawatts a year to space of AI compute, well, what's the value per gigawatt of that compute? Even if it went down to $10, that would be a trillion dollars a year extra. And cumulative, it would be so every year would be another trillion dollars because the satellites don't vaporize. So you'd be stacking a trillion dollars a year of revenue per year, and every five years $10 [trillion]. And if it was say at 50, you'd have five trillion.
电信运营商每年是怎么做的?电信运营商怎么可能用他们现有的基础设施满足服务需求?举例来说,如果他们拥有频谱这种公共资源,为什么政府,为什么政府会允许这种情况继续下去,而市场需要比他们所能提供的多得多的容量?
How are telcos going about it every year? How are telcos ever going to meet demand for the service with the infrastructure that they have? And if they have spectrum, for example, which is a public good, why does the government, why would the government allow that to continue to be the case when there's a need for a lot more capacity than they could possibly provide?
是的。好吧,我们要跳过那个问题吗?我认为 SpaceX 凭借已获得的频谱就能走得很远。但在某个时候,如果我们能比传统电信运营商更有效地利用频谱,我们直接收购那家传统电信运营商,把频谱的使用效率提高 10 倍。我们一直在为数据中心考虑这一点,所以最终 Starlink 的规模可能是 1 万到 10 万颗卫星。每 5 年我们替换现有卫星时,替换上去的容量是脱轨降下的卫星的 10 倍。所以我们将提供数千颗卫星的巨大容量。是的,我们仍然看到卫星能力有数量级提升的机会。所以我认为 Starlink 第四版的能力将是第三版的 10 倍,尤其是远距离激光链路。我们目前的激光链路大约是每秒 100 吉比特。下一代激光链路将是每秒 10 太比特。好 100 倍。然后我们会装上 10 个。
Yeah. Well, should we pass that question? I think SpaceX will be able to get very far with the spectrum that it's acquired. But at a certain point, if we're able to make much better use of the spectrum than say a legacy telco is, we could just acquire that legacy telco and use that spectrum 10 times more effectively. So we've been thinking that for the data center, so we're thinking about 10,000 to 100,000 for Starlink ultimately. And every time every 5 years when we replace the satellites that exist, we're replacing them with 10 times the capacity of what we're taking down deorbiting. So those are thousands of huge amounts of capacity we're going to be providing. Yes, we still see opportunity for order of magnitude improvements in the satellites' capability. So I think Starlink version 4 will have 10 times the capability of version three, especially with the long-distance laser links. Our current laser links are around 100 gigabits a second. Our next generation laser links will be 10 terabits a second. 100 times better. And then we'll put 10 of them on.
我们时间有点紧,但还有几件事。当人们担心数据中心和 AI 时,中国的人并不担心数据中心太多、电力太多。在中国,他们的电力大约是我们的三倍。他们每年新增的电力,大概 300 吉瓦,而我们总共有 550 吉瓦。他们一年就新增 300。
So we get a little short of time but there's a couple things more. So when people worried about data centers and about AI, people in China aren't worried about having too many data centers and too much power. So in China, they have about three times the amount of power that we do. And they're adding as much power a year, what, 300 gigawatts, compared to our 550 that we have in total. They're adding 300 a year.
我不确定那是不是稳定状态的电力数字,但肯定在增长。我不认为有那么高。我觉得比那个略低。但这是个巨大的数字。所以我认为中国目前的发电量大约是美国的 3 倍,而且趋势是到 4 倍,至少达到美国总发电量的 4 倍。而工业产能粗略地说与你的电力产出成正比。然后芯片方面,他们目前落后,但正在快速前进。
I'm not sure if that's a steady state power number but there are certainly increases. I don't think it's quite that high. I think it's a little lower than that. But it's a massive number. So China at this point I think is roughly three times the power output of the United States and it's trending to four to get to at least four times the total power output of the United States. And industrial capacity to first approximation is proportionate to your electricity output. So, and then chips, they're behind for now but they're moving forward fast.
几年前 SpaceX 的一位高管告诉我,她认为我的孙辈说普通话和 SpaceX 之间只差一件事。你相信吗?
And one executive at SpaceX a number of years ago told me that she thought that the only thing standing between my grandchildren speaking Mandarin and with SpaceX. Do you believe that?
嗯,我认为无论如何学普通话都有一些好处。这是一门非常有趣的语言。但是的,我认为在与中国竞争方面,从发电的角度以及你为经济活动利用了多少数字智能来看,我认为与中国竞争的唯一方式是把智能送上太空。这样我们每年可以带来几百吉瓦,甚至可能每年 1 太瓦。所以如果我们能每年向太空发送 1 太瓦的智能,那么在那种情况下,除非中国也能每年向太空发送 1 太瓦,否则我们的经济产出将超过中国。当然,那是个非常大的数字。那意味着我们每年做的数字超级智能是美国当前平均用电量的两倍。美国平均用电量是 500 吉瓦;1 太瓦是 1,000 吉瓦。所以如果我们每年做 1 太瓦,那就是每年向太空发送相当于美国当前平均用电量两倍的数字超级智能。这些基本上是惊人的数字。
Well, I think there's some merit to learning Mandarin either way. It's a very interesting language. But yeah, I think in terms of competing with China, from an electricity generation standpoint and just overall how much digital intelligence have you harnessed for economic activity, I think the only way to be competitive with China is to take the intelligence to space. So there we can bring on a few hundred gigawatts a year, and potentially a terawatt a year. So if we're able to do a terawatt a year of intelligence to space, then in that scenario unless China can do a terawatt a year to space, then we will have an economic output in excess of China. But that's a very big number, of course. That's saying we do twice as much digital superintelligence per year as there is currently the average power usage of the United States. So average power usage of the United States is 500 gigawatts; a terawatt is 1,000 gigawatts. So if we did a terawatt per year, that would be doing twice the current average power usage of the United States to space with digital superintelligence every year. These are staggering numbers basically.
那么还有两个问题。
So two more questions.
为什么我们需要这么多电力?为什么我们需要这么多芯片?对于那种智能、那种电力和芯片的需求会是什么?
Why do we need so much power? Why do we need so many chips? What's the demand that's going to come for that intelligence, for that power and chips?
需求将会是——这听起来有点循环——但对智能、对数字超级智能的需求,将来自数字超级智能本身。
The demand will be—and this is going to sound somewhat circular—but the demand for intelligence, for digital superintelligence, will come from digital superintelligence.
那是为了医疗目的,那是为了——我的意思是,那到底意味着什么?
Which is for medical purposes, which is for—I mean, what does that mean?
这意味着超级智能将远远满足所有人类需求。将会有全民高收入,全民极高收入,全民医疗保健,比今天地球上最好的医疗保健还要好,你可以希望得到一座城堡或任何你想要的东西,而数字超级智能可以作为一个副业为你制造出来。
It means that superintelligence will saturate all human needs by far. There will be universal high income, universal extremely high income, universal medical care that is better than the best medical care available on earth today, where you could wish for a castle or whatever you want basically, and the digital superintelligence can make it for you as a side project.
所以如果你把视野拉远,你可以把文明的进步看作是那个文明利用了多少能源。为了展示我们的太阳系还有多少增长空间:即使我们把文明利用的能源增加一百万倍,我们仍然只会使用远少于太阳能量百万分之一的能量。增长空间巨大。而且你显然会完全满足地球上任何人可能希望的任何东西。到那时,真正是数字超级智能为数字超级智能服务。我们的经济将显得古怪,我们的货币体系将显得像用贝壳交换货币。这就是超级智能眼中我们的货币体系的样子。
So if you zoom out a lot, you can think of the progression of civilization as how much energy has been harnessed by that civilization. And just to show you how much room for growth there is in our solar system: even if we increase the amount of power harnessed by our civilization by a factor of a million, we would still be using much less than a millionth of the sun's energy. That's a lot of room for growth. And you'd obviously completely saturate anything anyone could ever wish for on Earth. And at that point, it's really the digital superintelligence serving the digital superintelligence. And our economy will seem quaint, and our money system will seem like exchanging seashells for currency. That's what our monetary system will seem like to superintelligence.
严格来说,那确实是过去发生过的事情。
Technically that's a thing that used to happen.
嗯,货币曾经是人们互相给予的闪亮金属。比如,我给你这块闪亮的银片,然后你有一些闪亮的金片,你必须把这些闪亮的金属片从一个地方搬到另一个地方来进行交易,这真的很繁琐,坦率地说,而且容易被抢——如果你有金币或银币的话。但过去人们就是这样支付东西的。长途购物一定极其困难,因为你必须把所有的银和金运过海洋,船沉了,你甚至不知道发生了什么。这不像你可以用手机打电话给他们之类的。
Well, money used to be shiny metal that people would give to each other. Like, I'll give you this shiny silver piece of metal, and then you have some shiny gold pieces of metal, and you have to carry these shiny bits of metal from one place to another in order to do transactions, which is really onerous, frankly, and easy to rob—if you got gold coins or silver coins. But that used to be how people would pay for things. It must be incredibly difficult to buy things long distance, because you've got to transport all the silver and gold over the ocean and the ship sinks and you don't even know what's going on. It's not like you can call them on your cell phone or something like that.
所以显然,能和你永远聊下去对我来说很有趣,但我没时间了,所以只有最后一个问题。那么,当你像我一样 83 岁时,你希望人们怎么看你?你工作这么疯狂的时间,你知道,我认为没有人会像你这样工作只是为了积累金钱。所以你认为你拥有的任何东西最终都会被征很高的税,但你有很多钱最终会捐给基金会,你的基金会。你希望人们怎么看你?当你 83 岁时,你希望人们怎么看你?
So obviously it'd be fun for me to talk to you forever, but I'm out of time and so have just one last question. So, when you're 83 like I am, how do you want people to think about you? So you're working like these crazy hours you're making, you know, nobody would work the way you work to just accumulate money, I think. And so you think that whatever you have is going to be highly taxed at the end, but you have a lot of money ultimately going to foundation, your foundation. What is—how do you want people to think about you? How will you want people to think about you when you're 83?
嗯,我想我希望——有认知和现实。我的意思是,无论认知如何,我希望现实是,我为推动文明进步做出了重大贡献。你知道——是的。
Well, I guess I'd like—there's perception and reality. I mean, I would like the reality, whatever the perception is, but I'd like the reality to be that I made a significant contribution in the furtherance of civilization. That you know—yeah.
非常感谢你今天来到这里。
Thank you so much for being here today.
谢谢。
Thank you.