Apple WWDC, Siri AI, And SpaceX Data Centers | The Brainstorm EP 135

大家好，欢迎回到《Brainstorm》的又一期节目。

Hello everyone and welcome back to another episode of The Brainstorm.

今天人手有点不够。

We're down a man today.

没有萨姆，就我和布雷特·温顿，咱们来聊聊苹果和WWDC。

No Sam, it is just myself and Brett Winton and we're talking Apple and WWDC.

说不定这次Siri真能变聪明了。

Maybe we finally have a smart Siri.

谁知道呢？

Who knows?

布雷特，咱们俩得小心点。

Brett, you and I have to be careful.

这几年我们一直在为这个名字较劲。

We've been battling on this name for the last few years.

今天萨姆不在，不过我觉得今天能聊的东西还挺多的。

We don't have our mediator, Sam, today, but I think there's a lot to

我们开始吧。

dive in.

幸好我们不在同一个房间里。

It's a good thing we're not in the same room.

不然我们肯定又吵起来了。

We'd probably Yeah.

没错。

Exactly.

是的。

Yeah.

你和我一直对苹果在人工智能上的长期方向看法截然相反，这一直是咱们争论的焦点。

It's always been a hot button for both you and I taking really the opposite side of where we think Apple will be long term when it comes to AI.

今天，我觉得终于有了一些关于他们AI计划的最新进展。

Today, I think we finally got some incremental news around what they plan to do with AI.

在我看来，这些信息完全说得通。

And to me, it fits.

这其实正是他们过去几年早就该做的事。

It fits into what they should have been doing for the last few years.

这是第一次，我们看到他们开始以一种外界普遍认为苹果早就该采取的方式来应对AI。

It's the first time we're starting to see them address AI in the way that I think everyone in the world outside of internal at Apple thinks they should have been addressing it.

那我们直接切入正题吧。

So we'll dive right in.

首先是Siri的改名。

First, name change for Siri.

Siri现在改名叫Siri AI了。

Siri is now being called Siri AI.

如果有人还记得的话，它的后台架构也彻底更新了。

It is getting a whole new back end architecture if people remember

那个名字。

that name.

改名？

Change?

就是把AI加在名字后面而已。

It's the same name with AI on the end of it.

是的。

Yeah.

把AI加在后头。

With AI on the end.

这倒是挺新的。

Well, that's that's new.

这样也行。

That's Okay.

按苹果的标准来看，这已经算是很大的变动了。

That's a significant for app for by Apple standards, that's pretty significant.

我只是确认一下。

Just making sure.

明白了。

Okay.

继续说。

Keep going.

关于我们能得到什么，之前在这档节目里聊过，苹果会提炼谷歌的Gemini模型。

Terms of what we're getting, there was news and we've talked about it on this show that Apple would be distilling Google Gemini models.

今天他们确实提到了这一点。

And so they did talk about that today.

他们是把苹果自己的基础模型和Gemini模型结合起来，提取Gemini系列模型中实用的部分，融入苹果的架构和芯片中。

This is taking Apple foundation models with Gemini models and pairing the two, so distilling what was useful from the Gemini suite of models and bringing it to the Apple architecture and silicon.

现在你可以同时利用设备端的计算能力，他们也有自己的私有服务器，所以会在设备端处理和云端处理之间做些权衡，决定哪些功能在设备上运行，哪些要传到云端。

Now you can have both on device compute and they also have their own private servers, So there's a bit of toggle between what will happen on device and which devices will be able to do this and then also what's going to be routed externally to the cloud.

但很明显，Siri即将迎来一次重大升级——现在它能进行更深入的对话，深度整合你的个人数据，你可以直接查询自己的短信、图片、照片等等，Siri现在能访问这些内容了，我觉得这尤其重要，至少从我个人角度看是这样。

But it's very clear that Siri looks like it's going to be getting a major upgrade, as in now you can have deeper conversations, it is deeply integrated into all of your personal context, so you can query into your own messages, images, photos, you name it, Siri now has the ability to access that, which I think has been a big point, at least from from my standpoint.

这不是他们两年前就宣布过的东西吗？

Isn't this what they announced was it two years ago?

不是。

No.

就是说，

And it's the

他们两年前宣布的和今天这个相比，现在更贴近整个市场的方向了，而两年前他们说的那些，更像是个演示模式——比如搞了个ImageGen，结果只能生成卡通人物，根本不是那么回事。

same would say what they announced two years ago versus today is much more in line with where the rest of the market is versus what I think they announced two years ago was, like, playground mode, where it was like, here, we have ImageGen, and it can only do cartoon characters or No.

他们两年前推出的正是这些东西。

Those are the things they launched two years ago.

但两年前我记得他们推出Apple Intelligence的时候，苹果股价涨了快百分之十。

But two years ago, I remember there being Apple Intelligence, and Apple, the stock was up, I don't know, high single digits percent.

大家都说，他们终于开窍了，这些才是要真正上线的功能。

People are like, they finally get it, you know, and this is the features they're going to ship.

结果呢，大部分功能根本没真正上线。

And then they didn't effectively ship most of those features.

对他们来说，实现起来太难了。

It turns out it was too hard for them.

所以他们是不是重新调整了方向？现在看来，或许更有机会把这些功能真正推出，但本质上还是在重复当初那套说法？

So haven't they gone back and maybe now they have a better chance of actually launching these features, but they're basically announcing the same stuff again?

还是说不是。

Or is No.

就像是

There like

我觉得这是在他们之前宣布的内容和最终推出的一些功能基础上做的延伸。

I would say that this is building off of what they had previously announced and some of the features they ended up rolling out.

iOS 里已经内置了一些带有 Apple Intelligence 功能的模块。

So there are certain features within iOS that have Apple intelligence baked in.

现在他们把这些能力进一步释放出来，充分发挥了新 Apple 基础模型的潜力，而这些模型是经过 Gemini 精炼过的。

Now it's taking that and really kind of, I would say, like unleashing the capability to the fullest extent of their new Apple foundation models, which have been distilled by Gemini.

你现在可以直接在聊天里提问，我觉得这很重要。

So you're able to query in into your chats, which I think is a big deal.

你还能直接把信息推送到日历里。

You're able to push out directly to your calendar.

现在所有功能都深度整合了，不管是苹果自家的应用，还是第三方应用——当然，得看开发者愿不愿意接入这些能力。

So everything is, like, deeply integrated now with both their first party and also some third party apps, depending on if the developer wants to use some of these capabilities.

就目前已公布的内容来看，范围其实很有限，跟他们计划后续发布的相比差远了。

In terms of what has already been announced, I would say it's very limited in scope compared to what they plan to announce.

我最担心的是，这其实又回到了你之前说的，布雷特，他们两年前就宣布了这件事，结果我们还得等好几个月才能用上。

My biggest concern, and I think it gets back to your point, Brett, which is they made this announcement two years ago, and then we had to wait a few months to get access to it.

结果一上手，感觉也就那样，没啥惊喜。

And then it was pretty underwhelming.

不过今天我看了之后，真有点激动得不行。

Well, I was watching this today and I was getting super excited.

我心里想，天啊，这次我终于要对一回了。

I'm like, oh my god, I'm finally gonna be right.

布雷特只能干坐着说：‘天哪，尼克，你这次真说对了。’

And Brett's gonna have to sit there and be like, Oh no, Nick, you were right.

可就在最后，像那种药品广告一样，突然冒出一大串免责说明：这功能要到秋天才上线，因为还在测试，而且不支持这些语言，还有X、Y、Z一大堆限制。

And then right at the end, like any good drug commercial, it was like this massive disclaimer of this won't be launching until the fall because of beta and it won't be available in these languages and X, Y, and Z.

所以又是老一套，各种推迟，真正能对公众开放的时间一拖再拖，这

So it was like all of these delays again in terms of when this will actually be available to the public, which

我觉得苹果在这方面面临着一个挑战。

I think there's a challenge that Apple faces in that.

这不仅是苹果的问题，也是整个AI行业面临的难题。用AI做一个能完成预期80%功能的原型其实挺容易的，但要把它打磨好，做到剩下的15%，把错误率控制在5%以内，那就难多了。

And it's a challenge not just for Apple, but across all of AI, where delivering a prototype that can do 80% of what you anticipate it doing is quite easy with AI, but then cleaning it up so it does the last, call it even like 15%, leaving 5% as error is quite hard.

所以这对管理层来说是个考验，他们往往会说：这个演示差不多能用了。

And so I think it's a challenge for management teams and they say, Oh, this demo is almost working.

于是他们就认为：没问题，这事儿肯定能成，我们只要再稍微调优一下，就能彻底搞定。

Therefore, we can say, Yes, this is going to happen, and we'll just clean it up and we'll get it all the way there.

而且要把这些收尾工作全部做完，其工作量其实是之前所有工作的三倍。

And that cleanup, get it all the way there is actually 300% more work than everything that's already been done.

这几乎可以说是AI系统运作方式带来的一个特性，不仅是在苹果这里，这种特性很容易让那些预测者产生误判。

And that's almost like a function of the way AI systems work that can lead prognosticators astray, not just in Apple land.

你想一想，在无人驾驶出租车这块，我们之前不是说嘛，这事儿就在眼前了，马上就能搞定。

I mean, think about how long on the robotaxi side we've said, hey, this is like right there, we're almost there.

我觉得当你真正进入落地实施的阶段，你会发现世界变得极其复杂琐碎，这就让事情变得很有挑战性。

And I think that as you get into the real implementation space, the world is massively micro complicated in a way that makes it challenging.

我认为苹果这家公司以交付完整无缺的成品著称，因此特别容易受到这个问题的影响——因为他们会觉得，这产品还有这么多明显不行的地方，不能发布。

And I think that Apple as a company that prides itself on delivering the entirely completed packaged product to people is particularly vulnerable to that, where it means that they'll hold back from shipping because it's like, well, has all these things that we know don't work.

所以我们就没法上线。

And so then therefore we can't ship it.

这就让他们在前沿领域推进起来特别困难。

So it makes it challenging for them to try to operate at the frontier.

是的。

Yeah.

不对。

No.

我同意。

I agree.

而且我觉得，两年前那次发布让我有点犹豫，他们今天展示的东西，等到秋天正式推出时，可能根本不是当初演示的那个样子。

And I think, you know, the the the release two years ago gives me some hesitation that what they announced today, once they officially roll it out in the fall, won't actually be what they demoed today.

所以你说得对，我们确实得等等看，不过我觉得

So to your point, like, I think we are definitely in a wait and see mode, but I do

但我认为，不只是苹果，所有人都在看Anthropic的成果，看Codex，心想：这才是我们想要的。

think But I also think not just Apple, but everybody also is looking at what Anthropic has delivered and looking at Codex and saying, Oh, that's what we need.

我们需要至少接近这样的东西。

We need something at least approaching that.

但我还觉得，过了一年，由于这些模型的快速进步，整个格局又会变。

But I also think over the course of a year, the playing field's gonna shift again because of the rate improvement of these models.

就连苹果，我看过他们的演示，那个人说：‘你能帮我查一下这场演出的票吗？’

So even Apple, I saw them demo this, the person saying, hey, can you look up like tickets for this?

结果系统回答：‘票要到那个日期才开售。’

And it's like, well, the tickets won't drop until this date.

然后它说：‘请在票开售那天给我设个提醒。’

And it's like, set a reminder for me on that date when the tickets will drop.

但其实，人们期待的并不是这样的功能。

And it's like, but that's not actually what people will come to expect.

人们不会想：‘哦，现在我的设备会弹个提醒，我得再打开设备，然后说：好吧。’

It won't be, oh, now I'm gonna get pinged as a reminder on my device, so I have to go back to my device and be like, okay.

现在帮我找到网站，再做点别的事。

Now find me the site and do other stuff.

你应该说，一开售就帮我把票买了。

Instead, you wanna say, when that drops, buy me the tickets.

是的。

Yeah.

这正是我想的。

That's exactly what I thought.

我当时就想，天哪。

I was like, oh my god.

天哪。

Oh my god.

他们真会这么做。

They're gonna do it.

他们真会这么做。

They're gonna do it.

然后我设了个提醒，心里还嘀咕：别闹了。

And then it was, set a reminder and I was like, come on.

是的。

Yeah.

我们的能力早就远远超过这个阶段了。

We've like, we're way past this in terms of capability.

但说到底，这还是苹果的风格——他们不是在重新发明，也不是在创造什么新东西。

But this is I again, it's like this is very much in Apple fashion where they're not reinventing or they're they're not inventing anything new.

他们只是在调整，让它更好地融入苹果的生态系统。

They're kind of reinventing and making it fit to the Apple ecosystem.

不过我得说，还有一个重大变化让我觉得，他们至少在朝着正确的方向迈出了步子。

I will say the one other major change that I think should give us some hope as to, okay, they are making the step in the right direction in terms of capability.

Siri 现在会有一个独立的应用程序。

Siri now will have its own app.

当你提问或聊天的时候，至少有个地方可以回溯查看，就跟我们现在用的聊天机器人应用差不多。

So when you are making these queries and you're having these conversations, there's at least a place to go back to, which is very akin to what we're used to with the chatbot apps now.

所以在我看来，他们终于开始朝着正确的方向迈出了步伐。

So that to me seems like they're starting to make the right steps in the right direction finally.

但两年前，情况可不一样，那时候大家觉得：嗯，还行吧。

Whereas, I think two years ago, it was, okay.

他们确实在进步，但谁也不知道最终会走向哪里。

They're making the right steps, but who knows what direction they're gonna ultimately take.

我跟你说，你可能记错了，两年前他们刚推出这些功能时，人人都说：苹果终于来了。

I'm telling you, maybe you misremember two year two years ago, when they did this, everybody was like, Apple's arrived.

苹果智能就是一切。

Apple intelligence is it.

他们能卖光所有iPhone了。

They're gonna sell all their iPhones.

真的。

Thing.

明白了。

Okay.

是的。

Yeah.

我只是觉得

I I just think that

我好像又掉进同样的陷阱了。

I'm falling into the same trap.

也不一定。

Well, not necessarily.

我的意思是，我觉得他们这次承诺的东西更谨慎了，我觉得是这样。

I mean, I think I think they are being a little bit more measured in what they're promising here, I think.

而且我还是觉得，他们更有可能兑现一些已经承诺的功能。

And I still so I I do think they're more likely to deliver on some of the things that they've promised.

比如跨整个生态系统的搜索功能，先不谈AI，现在的搜索其实有很多奇怪的问题。

Like, even search across their entire ecosystem, leave aside AI, search is like doesn't work very well in weird ways.

就算你用桌面的聚焦搜索，有些内容能搜到，有些就是搜不出来。

Even if you do spotlight search on your desktop, some things show up and others don't.

感觉他们好像说要彻底重构整个系统。

It almost feels like they said they're re architecting all of it.

我觉得他们确实该这么做。

I think they need to.

而且之前

And was a

今天特别强调了这一点。

big point of emphasis today.

他们确实提到，苹果全家桶里的搜索体验实在不够好，现在也承认了。

They did talk about, I think they've come to terms with the search experience across the Apple suite of products isn't the best.

所以他们花了不少时间，还专门发布了一个新功能——搜索功能将全面升级后端。

So they did spend time and that was actually announced as like a separate product unveil that search is getting essentially like a new back end.

应该能从根本上改善。

It should start to end.

但如果做不到，问题就出在各个模块的协同上，因为据我所知，苹果智能会依赖这个搜索系统。

And if it doesn't work right, this is where all of the pieces fitting together matter because at least from what I understand, Apple intelligence will be using that search.

没错。

Yes.

这搜索功能会像一个工具，帮你精准找到需要的内容。

It'll be like tool use of that search to like find the stuff you need to find.

如果这个环节出问题，Apple Intelligence 就真不智能了。

So if that is broken, then Apple Intelligence won't be that intelligent.

对。

Right.

他们确实把这打造成一个统一的AI界面，但本地运行的模型明显是苹果自己开发的。

And I do think that there's a they presented as this as a unified kind of like AI surface, but the on device models are clearly Apple developed.

我觉得这跟Gemini模型是两回事。

And I think that's separate from the Gemini models.

我担心的是，本地处理、私有云，还有他们转向谷歌云服务这些不同系统之间怎么协调。

And I worry about on device versus their private cloud versus stuff they're going out to GCP for Google Cloud Services or whatever that stands for even.

所以最终呈现出来的体验，可能会有明显的断层。

So there's actually going to be weird kind of like seams in what they deliver.

这就像微软和OpenAI的合作，他们明确表示，必须全部自己来做。

And that's the, if you look at like what Microsoft did with OpenAI, they've basically said, we have to do it all ourselves.

我们不能依赖外部公司，来打造真正适合我们应用的体验。

We can't rely on another in order to deliver the experience that really, really works for us within the context of our apps.

至于他们能不能成功，我们拭目以待，但很明显，他们决定要自己投入资源发展这项能力。

And we'll see if they succeed, but that's clearly like they decided we need to spend this capital ourselves on our own capability.

苹果还没下定决心要这么做，至少没下这么大的注。

Apple has not decided to do that, at least in a very big way.

他们觉得，这部分可以外包，那部分可以自己做。

They've said, Oh, we can outsource this portion and kind of like in source this portion.

我觉得这会是个问题。

And I think that's a challenge.

你觉得他们会推出专门针对AI的硬件形态吗？

And I think that there's a Do you think that they are going to launch an AI specific form factor?

我的意思是，你怎么定义未来一年内这种形态？

Mean, even how would you define that over the next year?

你怎么看这个问题？

How do you think about that?

对。

Yeah.

我不确定会不会有专门针对AI的设备。

I don't know if it will be AI specific.

我个人觉得

I take

的

the

我们现有的硬件生态已经足够，完全可以围绕AI来优化，根本没必要从头再造。

personal stance that we have the hardware ecosystem today that would be you can kind of purpose build it with AI in mind, but you don't need to reinvent anything.

我不喜欢戴项链之类的配件，也不喜欢额外戴什么首饰。

I'm not a fan of pendants, I'm not a fan of wearing some extra piece of jewelry or what have you.

我觉得有手表、AirPods、眼镜和手机，就已经能获取所有需要的上下文信息了。

I think with the watch, with AirPods, with glasses, with a phone, you can capture all of the context you'd need.

所以，外接硬件的唯一目的就是这个。

And that's really the only purpose to have external hardware.

一是为了有更好的语音输入设备，二是能捕捉你看到和听到的环境信息。

It's one, to have a better input device for voice, and then to be able to capture, you know, context from what you're visually seeing and what you're also hearing.

至于是否需要专门针对AI的硬件，我相信他们最终肯定会推出一些产品。

So in terms of do they need something AI specific, I'm sure they'll end up launching something.

但在我看来，你身上戴的这些设备，其实已经是最适合他们用途的了。

But to me, I think you're you're wearing the pair of, hardware that is probably best for what they will do.

也许他们会加个摄像头，这已经被传得挺多了，用来采集一些视觉数据。

And maybe they throw a camera on it, which I think has been rumored so that they can capture some visual data.

但除此之外，更重要的问题是，未来会往哪个方向发展，以及每家公司各自的起点在哪里。

But outside of that, I think the more important question is where this is headed and where the starting points for each one of these companies are.

如果你想想Anthropic和OpenAI，他们是从头打造一个全新的操作系统开始的。

If you think about Anthropic and OpenAI, they're starting from we're building the new operating system.

而苹果则在另一边说，我们已经有现成的硬件基础了。在我看来，这两方最终会慢慢靠拢，Anthropic和OpenAI迟早会通过授权，或者自己开发硬件，来获取更多个人场景的数据。

Apple is over here saying, well, we have the existing hardware footprint, and they're both kind of headed towards each other, in my opinion, where eventually, in my opinion, Anthropic and OpenAI will either license or try to build out their own hardware to be able to capture more of that personal context.

而苹果这边，他们正在朝这个方向走：得打造一个全新的操作系统，免得那些新兴软件公司彻底颠覆现有的应用生态。

Whereas on Apple side, they're moving towards, Okay, we need to build out this new operating system so that we don't have these new software companies come and completely disrupt the app ecosystem.

没错，我的意思是，苹果推出一个Siri应用，这根本算不上是在构建新操作系统。

So to Yeah, I mean, that's clearly not Apple doing like a Siri app is not building a new operating system.

但我认为

I do think

它是动态嵌入到这个新系统里的，位于顶层。

it's dynamically embedded into the new island, it sits atop.

如果你仔细想想他们今天发布的内容，它确实是在顶层运行的。

If you were to really think about what they announced today, it does sit op.

它甚至有自己的专属应用，方便你回溯之前的对话记录。

Like, it has its own app just so you can go back and reference past conversations.

但它的位置确实高于其他聊天机器人，因为你只需要说‘嘿，Siri’，而不用说‘嘿，Chatty Bitty’或者‘嘿，Claude’。

But it does sit atop other chatbots from just the mere fact that you can say, hey, Siri, instead of, hey, chatty bitty or hey, Claude.

对吧？

Right?

现在 Siri 经常会误触发，我都搞不清楚怎么才能让它别再弹出来。

As of now, Siri accidentally triggers for me, I'm like, you know, can't even figure out how to get it off my screen.

拜托了，真受不了了。

I'm like, come on, please.

是啊。

Yeah.

也许他们能从这点出发，把体验做得更好。

So that maybe they can net improve from that.

我倒是想说，你有没有注意到，现在有人在卖3D打印的小配件，用来撑着笔记本电脑，方便随身带着走。

But what I was thinking is like, I don't know if you've seen, but people are now selling little three d printed doodads to keep your laptop open as you're carrying it around.

这样你的代码或者Claude的程序就能一直跑着了。

So your codecs or your Claude code can keep running.

这说明我们真的需要一种全新的设备形态。

Now that's a sign that we need a new form factor.

这不是简单的个人化设计问题，而是整个电脑的使用方式都该变了。

And it's not a new form factor that's a personal form factor, but it's like, even how computers work now needs to change.

苹果的AI硬件形态其实做得很好，比如Mac Mini和Mac Studio，现在都卖到缺货了。

Apple has a great AI hardware form factor, which is the Mac Mini and the Mac Studio, and those are sold out till whenever.

如果几年前我替苹果做战略规划，我会说：你们得让开发者在你们的生态里用AI变得特别简单，能轻松把AI模型集成到应用里。

And I think that, like if I were playing corporate strategies for Apple a few years ago, I would've argued, well, look, you need to make it very easy for people developing within your app ecosystem to use AI to develop and to kind of like trigger kind of like AI models into your apps.

但现在我觉得，他们应该说：这些Mac Mini会成为很多人运行个人智能代理的底层平台。

Now I think they need to say, Hey, these Mac Minis are going to serve as the substrate for a lot of kind of like people running their own personal agents.

咱们得想个办法，让手机能安全地远程指挥这些代理，把两者无缝又私密地连起来。

Let's figure out a way that we can securely use the phone to command those agents from distance and like actually tie that together in a way that's safe and private.

也许他们该好好利用自己独特的硬件优势——未来可能很多人家里都会一直开着一台小型服务器，专门跑自己的智能代理。

And maybe you need to like basically lean into their actually, really differentiated hardware capability there, where suddenly we're gonna go to, you know, potentially people having little servers running in their homes all the time to run their own agents.

如果他们真能把个人和这些代理的交互界面做得特别好，那会是个巨大的突破。

And, like, their kind of like personal interface to those agents could be a really, really big, if they did it, like I think they could do that well.

这甚至会彻底改变整个操作系统：你不再直接操作软件，而是指挥那些在设备上运行的智能代理。

And it would be a way that kind of like the entire operating system changes where you're no longer, you're really just orchestrating the agents that run on your device.

然后他们再慢慢推出自己的一套代理服务，让它们在平台上运行起来。

And then from there, they begin to try to develop out kind of their own agent sets that run on it and stuff.

就像那些以前搞比特币挖矿的热水器公司一样。

Yeah, it's like all of those companies that were building out water heaters that would mine Bitcoin.

现在不是挖矿了，而是运行智能代理，公司反而付钱给你。

Instead of mining Bitcoin, you're just running agents and the company is paying you for that.

对。

Yes.

或者它只是为你自己服务。

Or it's serving your own purposes.

没错。

Right.

我觉得这个智能代理领域，他们的隐私优势和你自己运行代理，确实是个不错的结合。

And I do think that that agent space, there is like their privacy angle and you running your own agents.

这里面有挺实用的组合。

There is like a useful mix there.

即使你和他们

Even if you And they

他们有自己的定制芯片，而且正如你所说，他们在算力方面确实有优势，但这一点可能还没被充分认识到。

have their own custom silicon and I think they do have, as you point out, an edge on the compute side that I don't know is fully appreciated yet.

也许等到人工智能代理进入下一个发展阶段，成为个人层面主要的算力承担者时，这个优势才会被市场和其他参与者真正注意到。

And maybe when we get to that agentic, that next phase of growth where agents are doing the primary amount of compute, even on a personal level, maybe that gets more recognized by the market and other participants.

是的。

Yeah.

说到硬件基础设施，上周末SpaceX又公布了关于地面数据中心的新消息。

Well, speaking of hardware infrastructure, there was additional SpaceX terrestrial data center news over the weekend.

之前有报道说，Anthropic从SpaceX租用数据中心，每吉瓦的租金略高于300亿美元，具体数字要看怎么算。

So previously, it had been reported that Anthropic was renting data center from SpaceX for around, it's a little more, depending on how you do the math, more than $30,000,000,000 per gigawatt.

这对SpaceX来说非常划算，因为每吉瓦的数据中心建设成本还不到300亿美元。

This is quite good for SpaceX because per gigawatt, they're spending less than 30,000,000,000 to build the data centers.

也就是说，一年就能回本。

So, one year payback period.

谷歌也在向SpaceX租用GPU，租金更是高达每吉瓦500多亿美元。

Well, Google is also renting GPUs from SpaceX and they're renting GPUs for more than $50,000,000,000 per gigawatt.

实际上，讽刺的是，以前包括我们在内的很多人都以为，SpaceX一旦上市，头五年肯定全靠星链，重点是星链的盈利能力和资本回报率。

And so actually, ironically, kind of like people, including us, had always thought, well, actually the first, call it even five years of SpaceX, once it's a public company, it's going to be all about kind of Starlink and the profitability of Starlink and the great return on capital characteristics of Starlink.

但他们现在却搞起了地面数据中心业务，而且因为已经开始赚钱，看起来会非常赚钱——他们建一个数据中心，按五年寿命算，每吉瓦成本大概290亿美元。

But here they have this on the ground terrestrial data center business that because they're monetizing, it's going to look really profitable, as in they will have spent, call it $29,000,000,000 per gigawatt to build a thing over a five year life.

算下来，每年成本大约560亿美元，另外还要花十亿美元来运营。

That's like, call it $56,000,000,000 in costs per year, and there's another billion to run it.

所以他们每年能收500多亿美元，而成本才60亿美元左右。

So, they're going to get 50 plus billion against 6,000,000,000 per year in costs.

像谷歌这种合同的边际利润率，毛利估计能到80%以上。

The marginal profitability of kind of that Google contract is going to look like it's high 80s gross margin.

这就让我有点困惑，不知道他们打算怎么看待这个业务，跟Grok和xAI那边怎么区分。

And so that is, it's actually a little bit, I'm a little puzzled how to think about how they're going to treat that business relative to kind of their grok xAI business.

是的，我有个问题想问一下。

Yeah, have a question off of this.

埃隆之前回应了你做的分析，你指出这项业务的盈利能力，他还转发了你的推文，说：‘别激动，冷静点。’

Given Elon responded to some of the work you did showing and highlighting the profitability of this business and basically followed up on your tweet and said, Hey, calm down.

这只是短期的。

This is only short term.

结果一周后，他又签了一个大单。

And then a week later, he goes and signs another major contract.

那你长远来看是怎么想的？

Exactly how are you thinking about the long term?

具体说说Grok这块，假设这个业务根本不存在，因为Grok像OpenAI、Claude和Gemini那样成功，真正受限的是算力，而不是反过来。

Specifically on the Grok piece, assuming this business doesn't happen if Grok is as successful as OpenAI and Claude and Gemini, where they're compute constrained, not the other way around.

假设因为Cursor的收购，Grok开始上线了，那这个Neo Cloud业务会怎么样？

So assuming Grok starts to come online because of the Cursor acquisition, what happens to this Neo Cloud business?

这得看你所说的‘Grok上线’是什么意思。

Well, it depends on what you mean by Grok coming online.

我觉得他们的需求远远超过服务能力。

So I think the More demand than they can service.

对。

Yes.

是的。

Yes.

但那样的话，收入会从前面产生。

But then it would be monetized in front.

所以你应该把他们的吉瓦算作分配在四个不同类别里。

So you should think of like their gigawatts as being allocated to four different categories.

第一，基础设施即服务，目前他们已经出租或售出了0.8吉瓦，而到2026年第一季度末，他们建成了整整一吉瓦。

One, infrastructure as a service, which right now they've basically sold 0.8 gigawatts for infrastructure as a service or rented it out versus end of Q1 twenty six, they had a full gigawatt built.

也就是说，他们已经出租了80%的算力，明白了。

So they've rented out 80% of their compute, Okay.

第二，付费推理服务，他们向客户收取订阅费，提供访问权限。

Two, monetized inference where they're charging somebody a subscription and they are giving them access.

我们认为，目前每吉瓦的付费推理服务能带来大约300亿美元的收入。

And we think that monetized inference can currently, you can get around $30,000,000,000 per gigawatt for that.

这是一门非常赚钱的生意。

It's a very good business.

第三类是未变现的推理服务，OpenAI 有一大块，XAI 也有很大一部分——比如，X 平台的高级用户或者普通用户都可以向 Grok 提问。

Three, unmonetized inference, which is OpenAI has a slug of this and XAI clearly has a slug of this, where in XAI's case, a premium X user or random X user, can ask Grok questions.

我可以使用 Grok Imagine 来生成内容。

I can use Grok Imagine to generate stuff.

他们确实在这些地方消耗了大量算力，但我并没有直接为此付多少钱。

They're spending gigawatts on that, but I'm not paying directly for it in any profound way.

所以他们指望的是，要么靠广告赚钱，要么将来把我转化成付费订阅用户。

And so they're counting on either advertising or converting me into subscriptions down the road.

然后第四类，是用于模型训练的算力。

And then three, like gigawatts for training.

这是在为未来模型的能力做投资，希望将来这些能力能变成能变现的算力。

And that is investing in the future capability of the model that you hope is going to be monetized gigawatts.

从短期看重股东回报、或者只关心眼前现金流的角度来看，问题在于：埃隆签了那么多赚钱的合同，却把收入都拿去投在训练上，赌的是 Grok 未来的能力，而那现在还不存在。

So from the, short term focused shareholder perspective or immediate cash flow focused shareholder perspective, the concern would be Elon has signed all of these great contracts and he takes away generating revenue to spend it on training, which is on a bet on future capability of Grok that's not there yet.

如果只是说，‘嘿，Cursor 的需求确实很大’，我觉得这没问题。

I think it would be fine if it's like, hey, there's lots of demand for Cursor.

Cursor 需要更多算力，因为它的毛利率应该做到百分之七十五左右。

Cursor needs more capacity because Cursor should be pricing on the order of 75% gross margin.

所以，即使你知道这种基础设施即服务的业务迟早会消失——毕竟合作伙伴总会去找别的供应商——但你还是愿意接，因为最终能把它转化成能带来收入的业务，还能增强你的竞争优势。

So you'd happily take infrastructure as a service business that you know is going to evaporate anyway because those partners will find other providers and convert it into business that generates you revenue and kind of increases your competitive position.

问题是，一旦 Cursor 被拉进来，他们很可能会砸一大笔钱用在 Cursor 数据的训练上，指望能提升模型能力，但你短期内根本看不到现金流。

The issue is once Cursor is roped in, like they'll have probably an opportunity to throw a lot of money at training against the Cursor data on the bet that that improves the model and improves their capability, but you won't see the immediate term cash flow.

这大概就是我目前的想法。

And so I think that's the roughly as I'm thinking about it.

我觉得 Anthropic 和 Google 这两个合作都是短期的。

I think both these Anthropic and Google deals are short term.

就像他说的，双方都有权随时终止合同。

Short term, as he says, it's like he has a right to cancel, they have a right to cancel.

他们现在急需算力，因为 Anthropic 想在上市前把收入增长推到最高。

They need immediate compute because in Anthropic's case, they want to have the revenue growth as high as possible as they get into an IPO.

而且他们能开出天价每瓦特的费用，因为现在大家花钱根本不在乎成本。

And they're able to actually charge ultra premium amounts per gigawatt because people are spending kind of like paying no attention to the costs right now.

对Anthropic来说，这笔支出反而带来了收入。

And so for Anthropic, this is revenue generating for them to pay this cost.

但他们显然在规划未来的产能，不打算继续以这个价格租用。

But clearly, like they are trying to capacity plan down the line and it does not include continuing to rent at this price.

所以我估计，每年会有大约一半的收入流失，直到逐渐归零，然后假设SpaceX只能按市场行情提高每吉瓦的收费。

So I have it bleeding, like 50% of it falling off annually until it bleeds down, and then assume that SpaceX can only charge more market rate per gigawatt for their compute.

即便如此，对SpaceX来说这依然非常有利，因为他们能更高效地建设基础设施。

Even that for SpaceX looks really good because they can build the infrastructure more efficiently than others.

你知道吗，我觉得他们建造这些数据中心的成本还不到每吉瓦300亿美元。

So there's also this, you know, I mean, I think that there's a, they spent less than $30,000,000,000 per gigawatt building these data centers.

而市场上其他公司建造同样的数据中心，成本要高达每吉瓦450亿到550亿美元。

The market trying to build these data centers spends 45 to 55,000,000,000 per gigawatt.

所以，他们其实完全可以只做基础设施服务，比其他新兴云服务商赚得更多，规模也更容易做大。

So actually, they could just be an infrastructure as a service company and be like way more profitable than the other neo clouds and scale and create a big business that way.

但我认为他们不会从战略上这么做。

But I don't think they're going to strategically.

你觉得他最终会用CursorGrock吗？

You think he'll come in with CursorGrock eventually?

会的。

Yes.

如果他们只做这一单不到一年，其实等于白得了这个数据中心。

And you could, in some ways, can interpret their actions right now if they do this business for like less than a year, then they've gotten the data center for free.

所以他们现在是按边际成本在运作，出租算力只是为了覆盖资本支出。

So, they're operating on this at the margin, we're renting out capacity just to pay for the CapEx.

从这个角度看，它其实是自我造血的。

And so, it's self funding in that way.

一旦把每一份算力都租出去了，剩下的就留着给自己的前沿训练用，再把后续扩建的算力租出去，用来补贴前沿训练的开销。

And then once we've rented out each bit of capacity, then we're reserving for or we're taking the frontier for our own training and then reserving the long tail of stuff that we've built out or renting the long tail stuff that we've built out to pay for that frontier training.

这样一来，这笔投入就不再是持续烧钱，而是基本收支平衡了。

So it could be a way that the line, rather than costing them cash over time, it's basically cash neutral.

最终，他们的目标是和Anthropic、OpenAI竞争，靠合理的定价赚钱，等太空中的数据中心或者太空服务器机架开始规模化，就能获得极强的变现能力。

And that would be, ultimately, it's trying to get to a spot where they're competitive with Anthropic and OpenAI, monetizing that healthy rate, and then their data centers in space or really server racks in space begins to scale, and then they have great monetization characteristics on that.

但我们还是走着瞧吧。

But we'll see.

是的。

Yeah.

历史上有没有哪家公司，在上市前业务模式发生过如此巨大的转变？

Has there ever been another company in history where the underlying business has changed so much leading up to its IPO?

因为就在一个月前我们聊这个的时候，情况还完全不一样。

Because it seems like even a month ago when we were talking about this, it was a materially different story.

如果再往前推一年，那时候我们讨论的这些事，要是跟投资人提，他们大概会说：‘嗯，祝你好运。’

And then if you were to rewind the clock a year, it's like some of the stuff we're talking about now, if you had pitched investors a year ago, they would say, Okay, yeah, good luck.

或者：‘这事儿短期内根本不可能实现。’

Or, Yeah, that's not happening anytime soon.

但现在看来，它的业务模式跟一年前，甚至一个月前，都已经完全不一样了。

And now it just seems like a fundamentally different business than it was even a year ago, and especially a month ago as well.

没错，如果回溯一年，或者看看我们当时的公开资料，其实是一年多以前了。

Well, yes, a year if you rewind the clock, or if you look at our publications, it was a little more than a year ago.

我们根本没料到SpaceX会涉及AI算力，那时候他们还没收购XAI。

We didn't anticipate AI compute at all for SpaceX, and they hadn't bought XAI yet.

而且我觉得，他们自己都没意识到这是个机会。

And kind of they, I don't think they recognize that as an opportunity.

你完全可以基于Starlink的预期来评估整个公司的价值。

And you can underwrite the entire business based on expectations for Starlink.

所以，他们现在选择上市，主要是因为AI算力这个领域潜力巨大，急需大量资金投入。

So while kind of like the reason they're coming public is because there's so much opportunity to deploy capital because AI compute is such a big opportunity.

但如果没有这个机会，他们可能根本不会上市，因为Starlink自己就能赚钱，先完成卫星星座建设，再慢慢推进火星计划。

If you didn't have that, they probably wouldn't have come public because then Starlink can self fund, and they would have built out that constellation and then gone from there to kind of the Mars.

但有了AI算力，这个机会的规模和影响力就完全不一样了，资金投入的需求也变得极其迫切。

But the scope and scale of the opportunity is so much bigger with AI compute and really the need to deploy capital.

这就是他们现在要融资的原因。

That's why they're raising.

而且说实话，我们一直觉得，基础设施即服务迟早会成为他们业务的一部分。

And, I mean, we always anticipated that infrastructure as a service would and could be a part of the business.

从一开始，我们就预测AI计算业务会把他们三分之一的计算资源对外出租。

Like the way from the start we modeled AI computers, they were going to rent out a third of their space compute resource.

有了对财务数据的深入了解，我们现在发现，星链的变现能力比我们原先预估的要高得多。

With access to the underlying financials, now we can tell, one, Starlink monetizes at a higher rate than we had previously modeled.

而且，我们本来就知道他们在数据中心方面很在行，但没想到他们在建设数据中心时居然这么高效、这么省资本。

And then two, they are, I mean, we knew they were good at data centers, but they are very, very good or very capital efficient at building data centers.

所以，如果他们把地面业务做成基础设施即服务，其盈利能力远超我们之前的预期。

And so the profitability of the on the ground business, if they do it as infrastructure as a service, is much higher than we would have anticipated.

实际上，这个业务相当复杂，很多核心指标之前都受保密协议限制，我们没法公开讨论。

So it's more like the business is complex and many of the underlying metrics had been under restrictions on our ability to share them and NDA.

等到公司上市后，我们才得以拆解业务，发现这些特性原来这么有意思。

And so then upon kind of coming public, then it allows us to pull apart the business and be like, oh, this is a really interesting characteristic, this is really interesting.

至于收购XAI这件事，我不确定有没有哪家公司在IPO后六个月内就收购了规模这么大的公司。

And the XAI acquisition, I don't know if a company has acquired another entity that large within six months of its IPO.

那我给你一个非常简短的回答。

Then a very, very short answer to you.

是的。

Yes.

至于Cursor，它只是对XAI的一个小小补充。

And then, well, Cursor's like a little it's a little small add on to x a x a

才600亿美元。

I Only 60,000,000,000.

第二点是，

Two, it's

根本不值一提。

not even worth talking about.

不过我觉得它是个很棒的资产。

Well, I mean but it's I think it's a great asset.

我觉得现在越来越清楚了，前沿模型领域可能最终会剩下三到四个主要玩家。

I mean, I think that there is you a know, it it seems more clear, maybe, that this is consolidating down to like three, four players going after kind of the frontier model space.

而把这个机会看作一个总估值在15万亿到20万亿美元之间的市场，SpaceX拥有别人不具备的独特优势去参与其中。

And recognizing that as a, call it, collective 15,000,000,000,000 to $20,000,000,000,000 enterprise value opportunity, SpaceX has its own advantages that allow them to go after that, that others do not.

但他们相比其他人也有短板。

And they have weaknesses relative to the others.

所以他们现在得赶紧想办法。

So they are having to kind of like scramble.

埃隆曾经说过，他有一个策略，能逐步让我们获得最多的算力，这确实没错。

Elon at some point said, basically, I have a strategy that will get us access to the maximum amount of compute over time, which is true.

接下来我只要别在中期阶段出错就行，我觉得我们现在就处在这一阶段。

And then I just need to not mess up the mid game, which is I think the space we're in now.

要实现到2030年代中期算力突然爆炸式增长的目标，还有大量执行工作要完成，这正是他必须掌控的。

And there's a lot of execution to get to kind of like that, call it mid 2030s, suddenly space compute is really scaling in a crazy way that he has to conduct.

说到前沿模型这块，依我看现在主要有五家巨头。

It seems like to your point on the frontier model side, there's five major players.

分别是SpaceX AI、OpenAI、Anthropic、谷歌和Meta，这五大玩家。

There's SpaceX AI, there's Open AI, there's Anthropic, there's Alphabet, and then the last one is Meta, the AI five.

在持续超越方面，你甚至可以说其中一些公司已经

In terms of continually leapfrogging, and you can say that some of them have even

如果把Meta算上，那可能还得把微软和DeepSeek也加进来。

Well, mean, you're going to layer in Meta, then you'd probably have to layer Microsoft and then you have to put in DeepSeek.

我觉得你应该把企业级AI和消费级AI分开来看，企业级是有一批公司在抢市场，消费级可能就更容易出现一家独大。

I think you should think of collection going you should think of it there's like this enterprise AI opportunity that there's a collection going after, then there's the consumer AI opportunity, which is maybe it's more winner take all.

这时候苹果就说了，我们也有自己的打法。

And this is where like Apple is like, yes, we have our angle on this.

显然谷歌在盯这个市场，Meta也在积极布局。

And clearly, Alphabet's going after that, clearly Meta's going after that.

而OpenAI也有自己的布局。

And then OpenAI has its exposure.

说实话，XAI在这块儿的优势没那么明显。

And I would say XAI has like less of an angle there, frankly.

不过咱们走着瞧吧。

But we'll see.

长远来看，这些公司之间重叠的部分，可能会有一两家最终在两个领域都胜出。

Like, over time, it does seem like the overlap between those players, there's probably going to be one or two that manages to win at both.

但这大约十五万亿美元的企业价值，只是从企业端来看的。

But the 15 ish trillion enterprise value opportunity, that's just looking at the enterprise side.

所以，在消费端，你的业务可能有九千亿美元的收益空间。

So on your work, it's something like $900,000,000,000 accruing to the consumer side.

这么说的话，消费端或许还另有一个九万亿美元的机会值得去开拓。

So, you could say maybe that's worth another $9,000,000,000,000 as a separate thing to go after.

好了，时间到了。

All right, we are up against time.

最后再讲一点。

One last point.

十三天前，埃隆以四千亿美元的价格把推特私有化了。而就在五天后，也就是本录音录制时预计的时间，推特将重新以X的名义上市，预期IPO估值高达一万七千五百亿美元。

Elon took Twitter private thirteen twenty days ago at $44,000,000,000 In five days time, expected from the time of this recording, Twitter will officially again X will officially be a public company again at the price of 1,750,000,000,000 as an expected you know, IPO price.

我真没料到会这样，但事实就是如此。

So didn't think that would happen, but here it is.

说真的，有人提出过一个观点：从一千亿美元增值到一万亿，比从一千亿涨到一万亿美元更容易，而后者又比从一百亿涨到一千亿更容易。

Well, it is it is you know, I think that there is a someone made or was making a claim that going from 1,000,000,000,000 to 10,000,000,000,000 in value is easier than going from 100,000,000,000 to $1,000,000,000,000 in value, which is itself easier than going from 10 to 100.

我在想，之所以会这样，是因为想象一下，幂律曲线正在变得更陡——如果你处在前三四名，那价值就会直冲云霄，对吧？

And I was thinking, like, the reason that would be is imagine the power law curve is steepening, where it's like, if you're in the top three or four positions, it just like shoots to infinity, right?

如果你处在中间地带，那就危险了，会被压进长尾里。现在有几百家公司值几十亿美元，但只有寥寥几家能值上百万亿美元。所以，这其实也说明了为什么X要和XAI、SpaceX深度整合。

If you're where you don't want to be is the middle like this middle ground where you just get, like, collapsed into the long tail, and there's like, you know, 500 companies worth whatever, like a few billion dollars, and then a few companies worth $100,000,000,000,000 And so that's, one, an argument for kind of X layering in with XAI, layering in with SpaceX.

整个融合的逻辑就是：未来可能会有那么几家公司，价值达到十万亿、甚至百万亿美元，但就那么几家而已。

It's kind of like the entire convergence thesis is like, hey, there's going to be some companies that are, yeah, 10,000,000,000,000, 100,000,000,000,000, but it's only going to be a few.

所以你必须站在那条幂律曲线的左侧，抓住拐点，不断往上走，而不是被拖进长尾里沉没。

And so you have to be on the left side of that inflection and of the power law curve to keep extending up rather than getting pled down into the tail.

对。

Yeah.

如果你看看那些市值超过一万亿美元的公司，再看它们的收入增长，会发现它们的增长率，比大多数公司的总市值还要高得多。

If you look at the cohort of companies that are a trillion dollars plus, and then you look at just their revenue growth, like they have best in class revenue growth on revenue basis that are the size of most companies market cap, you know.

所以就像你说的，它们在各自领域里已经突破了临界点，开始疯狂地集中资源和权力。

So it's like to your point, like they have hit some sort of escape velocity in whatever sector that they operate in, where they seem to be just consolidating power.

正因如此，收入、利润，最终是市值，都会一路飙升。

And as a result of that revenue, profits and eventually market cap shoots up.