#59 - 我们来聊聊AI泡沫 - Nat Rubio-Licht

好，Nat。

Alright, Nat.

我们来聊聊泡沫吧。

Let's talk about bubbles.

聊聊泡沫。

Let's talk about bubbles.

最近关于泡沫的讨论少了很多，但我认为还是有很多误解存在。

A lot of the bubble talk has died down a lot lately, but I think that there's still a lot of misconceptions out there.

还有人觉得泡沫会像大爆炸一样破掉，毁掉一切。

There are still people that think that it is going to pop in a big fiery explosion that's gonna ruin everything.

还有一部分人觉得

And then there are people

这全都是假象。

that This is all a mirage.

全都是假象。

It's all a mirage.

一切都是假的。

Everything's fake.

还有的人认为泡沫根本不存在，AI领域的一切都坚不可摧，这其实

And then there are people that think that the bubble isn't real at all and that everything in AI is ironclad, which is

也是不对的。我们得承认，事实并非如此。

simply We know not it's also.

现实情况往往介于两者之间。

The reality is somewhere in the middle.

但我认为澄清一些误解很重要。

But I think it's important that we clear up some misconceptions.

对。

Yeah.

你知道，关键在于这里涉及很多利益，对吧？

You know, the the thing is that there's a lot at stake, right?

有很多资金都投入到了人工智能领域。

There's a lot of money that's been invested in AI.

对。

Yeah.

有很多人在上面押上了自己的职业生涯。

There's a lot of people investing their careers in it.

很多公司在上面投入了巨额预算。

There's a lot of companies investing huge amounts of budget in it.

而且我们也知道，在风险投资等领域，粗略估计，大约有一半的风险投资都流向了AI。

And then we know like in terms of venture capital and all that, like roughly, you know, by some estimates, half of all of venture capital is going to AI.

这是一笔惊人的资金，可以说是史无前例的巨量资源在押注这个东西。

It's an incredible amount of money, like an unprecedented historic amount of resources betting on this thing.

所以我觉得我们得意识到，其中肯定有一些部分会效率低下，也就是说那些地方可能就是泡沫所在，对吧？

And so I think we know we have to believe there are pockets of this that are gonna be inefficient, meaning that's probably where there are bubbles, right?

那些泡沫最后肯定得破灭，或者缩小，但同时又因为人们感觉到，他们好像在找替罪羊，你明白吗？

Those bubbles are gonna have to, they're gonna have to pop or deflate or whatever, but also because people sense that they're sort of looking for boogeyman, you know, right?

他们就是在四处张望，说这个有问题，那个有问题。

They're like looking and saying like this thing or that thing.

这可以理解，因为我们知道这里面有些部分会耗尽能量。

It's understandable because we know that there's parts of this thing that are gonna run out of energy.

这就像自然过程，让百花齐放，最终你不得不剪掉一些花。

It's like the natural process that you let a thousand flowers bloom, eventually you're gonna have to cut down some flowers.

或者有些花

Or some flowers are

就会枯萎凋零。

just gonna wilt and die.

或者它们就会枯萎凋零。

Or they're just gonna wilt and die.

它们会自然消亡。

They're gonna die a natural death.

我们还能用什么比喻呢？

What other metaphors can we use?

气泡、花朵，

Bubbles, flowers,

都是小叮当那种。

all Tinker Bell.

小叮当。

Tinker Bell.

对。

Yes.

对。

Yes.

所以这就是小叮当的比喻。

So here's the Tinker Bell metaphor.

说来

Hit

听听。

me.

在小叮当里，孩子们或者那些我一时想不起的角色，要让小叮当活着，你就得相信有精灵。

In Tinker Bell, the way that children or whoever the characters in Tinker Bell are that I'm forgetting conveniently right now, The way to keep Tinker Bell alive is that you have to believe in fairies.

我觉得这个比喻非常贴切，正好反映了我们当下对人工智能的看法。因为大家普遍被一种乌托邦式的愿景所说服，认为AI能够彻底改变我们的社会，包括我们的工作方式、生活方式、医疗保健以及基础设施等方方面面。

And I think that that is an apt metaphor for where we're at with AI because we are sold on this very utopian vision of what AI has to offer, that it can completely transform our society in the way that we work and the way that we live and our health care and our infrastructure and everything.

但前提是我们要相信它，允许它存在，并且每个人都得认同这一点。

But only if we believe in it, only if we allow it, only if everybody buys in.

但我认为，我们最终得到的支持，可能并不像我们预期的那样；至少，目前这些实验室所获得的支持程度，恐怕也没那么高。

But I think that we're not going to get the buy in that we are thinking or that the, that, at least that We're certainly the not labs getting are

确实如此。

it right now.

就像

Like

我们现在还没弄明白。

We're not getting it right now.

在60位公众人物中，

The narrative, among the 60 public

有1%的人并不相信这一点。

1% of people don't believe in.

对的。

Yeah.

或者对AI持负面看法。

Or have a negative view of AI.

从我们的数据来看，其他数据也是这样。

From our data and from other people's data too.

所以我们知道，受众以及更广泛的美国人群中，有61%的人对AI持负面看法。

So we know that 61% of people, in our audience and in the broader, in, in The U S, have a negative view of AI.

对。

Yeah.

所以我们经常在《The Deep View》讨论，AI在更广泛的公众中正失去舆论优势。

So, I mean, we talk about it a lot on The Deep View that AI is losing the narrative among the broader public.

对。

Yeah.

但在那些投资AI、从事AI工作、以AI为事业基础的人中间，他们可是全都全力投入的。

But among the people who are investing in it, working in it, you know, basing their careers on it, man, they are all in.

我们在参加各种会议或者其他场合时就能看到这一点，大家真的非常投入，觉得这会改变一切、改变未来、改变所有事情。

Like we see it when we go to conferences, when we go to other places, like they're so in, like this is changing everything, this is changing the future, this is changing the things.

你说的那个点正是如此，这种热情推动了投资，也促使人们转行进入这个领域。

It's that part, that's the part that you're getting at, that like that's sort of lifting up the investment, the people moving careers to get into it and all of that.

这就形成了一股向上的力量，在推动着整个行业的发展。

And that is like, you know, this sort of upward force that's driving it.

但如果他们不再相信了，那就麻烦了。

But if they stop believing, uh-oh.

对，没错。

Yes, exactly.

哎呀。

Uh-oh.

不过我想正因为如此，我觉得结果会处于中间状态。

But I guess how much this is why I think it's gonna fall somewhere in the middle.

到底需要多少人相信，大家才会真正接受呢？

How much belief needs to happen in order for it to like, who's who needs to buy in?

谁需要买账？我知道，因为有个误解，觉得这些公司赚不回他们投入的全部资金。

Who needs to like, I know that because these there's this misconception that these companies are making they're they're they're not going to be able to make all the money that they're spending.

我确实认为他们已经花了很多钱。

And I do think that they are have spent a lot of money.

他们花了很多钱在建设上，花了很多钱在算力上，花了很多钱。

They're spending a lot of money in the build out, spending a lot of money on compute, they're spending a lot of money.

我说的这些公司，指的是那些大型AI实验室。

And when I say these companies, I'm talking about the the big AI labs.

对。

Yeah.

花大量钱来提供算力。

Spending a lot of money to serve up compute.

其中大部分用在训练上，而不是推理上。

A lot of that is going to training, however, not to inference.

还有很大一部分是在建设数据中心。

And a lot of that is also going to building out data centers.

所以我觉得成本会降下来，对吧？

So I think that costs are going to come down, but right?

另外还有一部分原因是，他们的大部分收入并不是来自订阅服务。

And also the other part of it is that most of their money isn't coming from the subscriptions.

大部分收入其实来自企业合同。

Most of their money is coming from enterprise contracts.

没错吧？

Correct?

没错。

That's right.

他们的收入

Their revenue

收入，他们带来的

The revenue, they're bringing

主要来自企业客户。

in, you're is coming from the enterprise.

对。

Yeah.

确实是大头。

Like big time.

所以，企业客户是唯一真正需要买单的人群吗？

So are the enterprises the only ones that really need to buy in?

如果是这样的话，他们为什么还要花那么多精力去影响公众舆论呢？

And if so, why are they spending so much time trying to sway the public narrative?

故事线。

Narrative.

为了吸引消费者，去争取他们，是的。

To get consumers, to get the, yeah.

所以因为

So because

他们需要让很多人参与进来，因为这些人是员工，这些人将来会创办公司。

they need to get a lot of people to buy in because those people are employees, those people are gonna start companies.

那些人会创办初创公司。

Those people are gonna start startups.

那些人都是学生，将来他们要影响到自己当前和未来工作公司对 AI 的接受。

Those people are students who are eventually going to influence AI adoption in their companies, in their future companies.

所以他们需要所有人都参与进来，因为这也会整体带动企业端对 AI 的采用。

So they need everybody because that's also going to sort of lift this general, you know, adoption from the enterprise.

但我说句明白话，至少在可预见的未来，企业依然会是 AI 和 AI 采用的主要推动力。

But I think make no mistake, at least for the, you know, the coming future, enterprise is going to continue to power, AI and AI adoption.

企业的投入怎么样，AI 的采用就会跟着怎么样。

As enterprise spend goes, like so will AI adoption.

我觉得这就是为什么你会看到 OpenAI 在 2026 年特别拼命转向企业，直接停掉了 Sora 和一大堆面向消费者的东西。

I think that's why you've seen OpenAI pivot so hard toward enterprise in 2026 and like stop doing Sora and a lot of the consumer facing stuff.

对。

Yes.

但他们并没有完全放弃消费端的业务。

But they're not completely done with the consumer facing stuff.

当然。

For sure.

对我来说，ChatGPT Health 感觉更像是一个面向消费者的产品。

Like ChatGPT Health to me feels like a consumer thing.

是的。

Yep.

ChatGPT 感觉像是面向消费者的产品。

ChatGPT feels like a consumer thing.

对。

Yeah.

比如，我妈今天给我发短信，问我帮她弄手机上的东西，因为她快七十了，有时候不太会用手机。

Like my mom texted me today and she asked me for help with something on her phone because my mother is in her late sixties and sometimes doesn't know how work her phone.

我就说，哦，不好意思。

And I was like, oh, I'm sorry.

我现在没法帮你。

I can't help you right now.

我就是，我挺忙的。

I just like, I'm busy.

这让你听起来像个不孝的女儿。

And that makes you sound like a terrible daughter.

不过，那不是什么紧急的事，真不好意思。

However, it wasn't an urgent I'm so sorry.

对不起，妈妈。

I'm sorry, mom.

这让我听起来很糟糕。

It makes me sound terrible.

总之，后来她又发短信给我。

Anyways, but then she texts me later.

她说，哦，我自己搞明白了。

She's like, oh, I figured out how to do it.

然后她发来，说ChatGPT知道怎么做。

And then she sends, chat GPS knows how to do it.

我就说，

And I'm like,

哦，哦，妈。

oh, oh, mom.

太厉害了。

Amazing.

她显然是去问了ChatGPT帮忙解决某个问题。

She clearly went to chat GPT to help her with something.

对。

Yeah.

不过，我妈那种人，根本不会去创业。

So, however, my mom is not someone that is gonna start an enterprise.

我妈她呀

My mom is

她肯定不愿意每个月花20块钱。

She's not gonna pay $20 a month.

我妈根本不会每月花20美元。

My mom's not even gonna pay $20 a month.

对。

Yeah.

那服务这类受众的好处到底是什么呢？

So what is the benefit of serving that audience?

没错。

Yeah.

我的意思是，面向的是更广泛的公众。

I mean, it's the broader public.

用的人越多，大家越熟悉它，迟早他们都需要一些服务。

If you get more people using it, more people familiar with it, you know, they're gonna need services at some point.

所以打个比方，长远来看，比如你妈妈需要找个度假的地方。

So for, you know, broadly the bet is like eventually say your mom needs a service to, look for a place to go on vacation.

这时候就有一个像她那样的聊天机器人，专门做旅行推荐，比如她喜欢的酒店或者旅行社之类的。

And there's a chat bot like her, travel, you know, her favorite, you know, travel hotel or travel agent or whatever.

它可能会说，哦，我们有个这个选项。

It says like, oh, well we have this option.

你可以用这个。

You can use this.

要是她之前用过 ChatGPT，对吧？

And if she's used chat GPT, right?

那一下子就会觉得，哇，太棒了。

Like then all of a sudden it's like, oh great.

有次我试了一下，它回复了我，我也就搞明白了。

I used that that one time when that one answered my message and I was able to, I was able to figure it out.

对吧？

Right?

所以我的意思是，赌注在于，如果更多人习惯用这些东西，将来会有很多应用场景，甚至能发展成企业级业务，对吧？

So it's like the, the bet is that if more people get used to using this stuff, like there are gonna be a lot of uses that eventually are also businesses and we could call them enterprise in some fashion, right?

那这跟泡沫有什么关系呢？

So then how does this connect back to the bubble?

所以关于泡沫，问题是这样的，你在Instagram上看到了这个，然后我们就在讨论它。

So the bubble, the thing is, you saw this thing on Instagram, and then we were talking about it.

对。

Yeah.

这是我们想聊这个的原因之一。

This is one of the things that made us wanna talk about this.

我们就在想，天啊，关于这个泡沫的看法真是五花八门。

We're like, man, the views on the bubble are all over the place.

有些人说，泡沫已经破了，根本就没存在过。

Some people are like, the bubble is dead, it was never a thing.

然后你又看到，你知道的，就像你在Instagram上看到的那样。

And then you have, you know, like you saw on Insta.

对。

Yeah.

就是那种感觉，它就在眼前，然后这就是为什么这一切都要崩盘的原因。

Where it's kind of like, here it is, and this is why all of this is gonna collapse.

彻底崩盘。

Crash and burn.

对。

Yeah.

这一切都会彻底崩盘。咱们稍微聊聊你看到的内容，以及为什么我们会聊起这个话题。

All of this is gonna So crash and talk a little bit about what you saw and why we were chatting about it.

我看到一个视频，里面的人说AI的经济账根本算不通。他举了个例子，说哪怕有人花两百美元在ChatGPT上——抱歉，我说错了，是Chewy T。

I saw this video of somebody who was talking about how the economics of AI don't make any sense because they brought up that if even though someone spends $200 on a chat I'm sorry, not chat Chewy t.

Claude的订阅，最高档的那个，每月服务费要八千美元。

A Claude subscription, the maximum subscription you can get with Claude, it costs $8,000 a month to service that subscription.

他们就是这么说的。

That's what they said.

确实是这样说的。

That is what they said.

不过说实话，他们并没有提供数据来源。

I don't I they did not cite their sources to be fair.

对。

Yeah.

不过，这个人提到的经济账未必算得过来。

However, the economics of that don't necessarily the economics that this person mentioned don't necessarily add up.

首先，那八千美元里很大一部分其实是花在训练上，而不是推理。

For one, a lot of the money is spent on a lot of that $8,000 is spent on inference sorry, on training rather than inference.

实际上，推理的成本比训练一个大型强模型要低得多。

Inference is actually cheaper than training a big powerful model.

对。

Yeah.

他们刚才说的，让我先插一句哈，就是

What they were talking about, just as, you know, interrupt you for one second, is like

不用客气。

No, please.

你可以去查一下，如果按 token 计费的话。

The fact that, you know, you can look and you can see if you pay per token.

对。

Yeah.

你知道，当你买那个20美元的订阅或者200美元的订阅时，你会得到一定数量的令牌。

You know, when you buy that $20 subscription or the $200 subscription, you get a certain number of tokens.

比如200美元的那个，你得到大约8000个令牌，在这种情况下，他们看了一个图表，显示“哦，那是价值8000美元的令牌”。

Like with the $200 one, you get like 8,000, in this case, they looked at one of these charts that shows like, oh, that's $8,000 worth of tokens.

他们犯的错误是，他们以为“哦，Anthropic 提供这些令牌的成本是8000美元”，但事实并非如此。

And the leap that they made that wasn't right was that like, oh, it costs Anthropic $8,000 to serve those tokens, which is not the case.

因为如果你按令牌付费，就像企业那样，如果你使用所谓的API，那么你最终会支付更多。

Because if you were paying per token, like enterprises do, if you pay for what's called the API, then you know, you end up, you pay a lot more.

对。

Yeah.

但当你支付20美元订阅费给消费者时，你不需要支付那么多，为什么？

But you don't pay that when you're, pay the $20 subscription to your consumer, why?

因为大多数人根本用不到那么多，对吧？

Because most people never get anywhere close to that, right?

对。

Yeah.

这只是一小部分，非常小的一部分。

It's such a very, very small fraction

这就好比提供自助餐厅无限畅吃的服务，但你根本不可能把整个自助餐都吃完。

of It's like offering an all you can eat buffet, and then you're not gonna eat the whole buffet.

大多数人顶多吃个沙拉之类的东西，对吧？

Most people are gonna get a salad in one thing, right?

那顿饭，他们花了50美元，但食物只值7美元左右。

And that meal, they paid $50 for like $7 worth of food if

没错。

they Exactly.

确实。

Would

对吧？

Right?

大多数人最后都是这样。

Most people ends up being like that.

所以这个推理是有缺陷的，但你和我对它感兴趣，因为你看，这确实存在，而且Instagram上的人甚至都在讨论它有多火。

And so it was like this flawed reasoning, but you and I were interested in it because it's like, look, this is out there, and people on on Instagram were even just talking about how It had

很多点赞，很多分享。

a lot of likes, a lot of shares.

没错。

Exactly.

这就是为什么AI，你知道，普通人都在谈论为什么AI是泡沫，它最终会自我崩溃。

This is why it's all this is why AI you know, this is like normal people just talking about like why AI is a bubble and it's all gonna collapse on itself.

对。

Yeah.

而且我认为这是整个行业的普遍看法，认为这些都是骗人的。

And I think that that is the general perception of the industry that it's all snake oil.

是的。

Yeah.

我也不会说里面完全没有骗人的东西。

And I'm not gonna say that none of it is snake oil.

就像我们刚才说的，对。

As we were saying Yeah.

现在确实有很多公司会倒闭，或者说是那些跟风模仿的、都在干同样事情的公司，你懂的。

Today, there are plenty of companies that are going to burst and pop or plenty of companies that are model rappers or that are all doing the same thing of, like, you know

我们是代理的控制平台。

We're control plane for agents.

没错。

Exactly.

我跟你说，怎么

Telling you, how

好多人来找我，想让我投资他们的项目，说我们

many people come to me and pitch me on like, we're We have a few

控制你的AI。对。

control your AI Exactly.

确实，每家公司都在做类似的事情，很多公司做的内容都差不多，但也有很多公司在做一些真正很酷的东西。

And it's like, yes, every company is doing There's a lot of companies that are doing the exact same thing, but there is a lot of companies that are doing really cool stuff.

对。

Yeah.

毫无疑问，这里确实有价值，所以我认为最终我们会找到一个中间点。

And there is undoubtedly value here, which is why I think that we are going to land somewhere in the middle.

但具体怎么做呢？这个中间状态到底长什么样？

But how do we what does that middle look like?

对，我觉得这也是我们之前聊过的一个点，就是那个‘小叮当效应’。只要大家还信，这个泡沫就能撑住，对吧？

Yeah, I think that this is one of the things that we were talking about as well, is where we got to the Tinker Bell thing, which is that the bubble, as long as people believe it, right?

支撑它的正是人们的信念。只要足够多的人相信这是下一个风口，它就能继续转下去，对吧？

Like that's what props it up is people believe, enough people believe this is the next big thing that keeps it going, right?

不过接下来它可能会开始动摇，这其实是自然规律。

Where it's gonna start to falter and like, it's gonna be natural.

肯定会经历一段幻灭的低谷期。

There's gonna be a trough of disillusionment.

这是高德纳的那个曲线，对吧？

There's the Gartner's thing, right?

就是那个期望膨胀的高峰期。

Like the peak of heightened expectations.

对。

Yeah.

然后还有那个幻灭的低谷。

And there's the trough of disillusionment.

我们绝对正处于那个巅峰的高度。

Like we are definitely in the height of the peak.

你觉得我们还没到顶吗？

You don't think we've already reached?

我觉得我们百分之百正在滑向那个幻灭的低谷。

I mean, I think we're sliding toward the trough of delusionment 100%.

对。

Yeah.

对吧？

Right?

你看看现在美国的人和十二个月前相比，他们对AI的看法变得更差了。

You look at people today in The US versus twelve months ago, their opinion has gotten worse about AI.

是的。

Yeah.

那是不是意味着我们正在滑向幻灭的低谷？

So would that mean we're sliding toward the trough of disillusionment?

我们肯定还没到谷底，但当人们不再相信，当更多人不再相信的时候，就是一些泡沫开始破裂的时候。或者如果，你知道，如果有数据出来说，这家公司每个客户花了10万美元，却只向客户收取25美元，那人们就会开始传播更多像我们之前讨论的那种视频，然后人们就会说，天哪，这整个东西就像个传销，对吧？

We're definitely not at the bottom though, but when people stop believing, when more people stop believing, that's when some of these sort of bubbles start to pop or anybody, if somebody, you know, if data comes out and it's like, oh, this company was spending a $100,000 per customer and they're only charging customers $25 Well then like people start to go more videos like the one that we were talking about, start to go and then people are like, oh man, this whole thing is like a pyramid scheme or something, right?

就是那种事情会发生。

Like that's the stuff that will happen.

对。

Yeah.

然后发生的每一件小事，对吧？

Then like every little thing that happens, right?

大家开始觉得，哎，这就是泡沫。

People start getting more like, oh, there it is, it's a bubble.

对。

Yeah.

我们之前就有点那种感觉，你还记得吗？去年年底的时候差不多就是这样。

We were sort of at that, remember like at the end of last year, we were sort of at that.

当时所有人都在谈论泡沫。

The bubble was like all people were talking about.

我记得，确实。

I remember, yeah.

然后智能体出现了，炒作周期又开始了一轮。

And then like agents happened and then it's like the hype cycle started all over again.

是的，感觉我们暂时从这件事上分心了。

Yeah, it does feel like the, we've been momentarily distracted from this.

但我觉得所有关于泡沫的讨论，很大程度上都是被循环融资这种想法带起来的。

But I think all the bubble conversations were largely fueled by this idea of circular financing as it related Yes.

关于数据中心交易的那一万亿美元承诺。

To the trillion dollars of commitments for data center deals.

不过，我们现在到了这个点，OpenAI和Anthropic都急需算力。

However, we are now coming to this point where both OpenAI and Anthropic are like desperate for compute.

他们的算力不够用。

They don't have enough compute.

他们的算力确实不够。

They don't have enough compute.

关于Mythos，有个不太为人知的小秘密，就是他们说‘我们把它收回来了，因为它太强大了’。

That's the dirty little secret about Mythos was like, oh, we pulled it back because it's just too powerful.

好吧。

Okay.

但部分原因是他们根本没有足够的算力来运行它。

But part of it was that they just didn't have the compute to run it.

这很明显。

Like, it's pretty clear.

这就是他们为什么跟SpaceX合作，还跟其他公司合作的原因。因为他们急需拿到算力资源，不然的话，他们的客户端会遇到速率限制的问题，根本没法提供服务。

That's why they partnered with SpaceX and they're partnering with all of those other companies because they're like, we need to get our hands on compute right now because our, we're running the rate limit issues with our, with our clients and we're not able to serve.

对。

Yeah.

所有那些想要用我们服务的人。

All the people that want us.

我说情况很明显，意思是大家应该还记得他们之前出现过很多次宕机吧。

When I say it's pretty clear, mean, that, like, remember they had lots of outages.

对。

Yeah.

记得他们发布Mythos之后，在美国政府介入之前吗？

Remember after they released Mythos, before the US government, you know, had them.

是的。

Yeah.

然后重新提交，结果他们突然开始签下一堆算力协议。

Resend it, that all of a sudden they started signing all these compute deals.

对。

Yeah.

是的。

Yeah.

就是你提到的那个。

That you mentioned.

所以这有点像是一种恐慌性的举动，对吧？

And so it was like a little bit of a panic move, right?

他们当时没有足够的算力来运行。

They didn't have the compute to run it.

然后突然间，OpenAI 去年一直在争取的那些算力资源，之前看着简直不可思议，现在突然之间

And all of sudden all the compute that OpenAI was lining up last year, all of sudden, which looked insane, now all of sudden

看起来就有点先见之明了。

It looks a bit more prescient.

感觉稍微合理一点了，嗯，这样讲得通了。

It looks a little more like, okay, that makes sense.

还有件事我们也得聊聊，就是刚过去的财报季，谷歌、亚马逊和微软都发布了财报，这些云服务公司全都从中大赚了一笔。

And then we also had the thing we should talk about too, is we just came through earnings season and Google, Amazon, and Microsoft released their earnings and the cloud companies are just in bringing from all of this, you know

当超大规模云服务商的日子真不错。

It's good day to be a hyperscaler.

现在是做超大规模云服务商的绝佳时机。

It's an amazing time to be in hyperscaler.

对。

Yeah.

它们的增长都是两位数的，都是这样。

They are up, you know, double digits all in all cases.

我想说亚马逊，抱歉，亚马逊。

Think Amazon At Amazon, sorry.

亚马逊市值首次突破三万亿美元。

Amazon hit a $3,000,000,000,000 market cap for the first time ever.

它是有史以来第五家达到这个水平的公司。

Only the fifth company ever to cross.

以前从来没有人做到过。

To ever do that.

对。

Yeah.

是的。

Yeah.

对。

Yeah.

所以没错。

So yeah.

这让我想到，最终的决定权还是在企业客户手里。

It's and I it I it does make me think that I think that where the buck stops is with the enterprises.

因为归根结底，很多钱确实是从那里来的。

Because at the end of the day, yes, that is where a lot of the money is coming from.

模型提供商和云服务商的价值也主要是在那里体现的。

That is where a lot of the value is being derived for the model providers, for the cloud providers.

就是这些企业，现在把投入的资源全都吃掉了。

That's, those are the people that are eating up everything that's being put down right now.

我确实也在想，企业会不会发展到某个阶段？我觉得他们现在已经开始有这个趋势了，特别是在限制 token 用量这方面。

And I do wonder if the enterprises are going to get to a point, and I think they're actually starting to approach that right now with this whole idea that we're tugging the reins on token maxing.

对，没错。

Oh, yeah.

明白了

Got the

我想关键问题是，企业会不会自己产生幻灭感？这种幻灭感又会引发怎样的多米诺骨牌效应呢？

whole I guess the question is, are enterprises going to become disillusioned on their own, and how is that going to cause a domino effect?

是啊，这点必须得聊聊。因为这在今年春夏季的各大会议期间一直是个热门话题。

Boy, yeah, we have to talk about that because that's been a huge theme of all of the, like, spring summer conference season.

对。

Yeah.

就是这些活动上大家都在讨论这个。

Is it all these events.

所以你去了微软的Build大会，而Snowflake去了Databricks。

So you went to Microsoft Build and I Snowflake went to Databricks.

一切都关乎效率。

It's all about efficiency.

数据加AI。

Data plus AI.

全都关乎效率。

All about efficiency.

还有Confidence Computing峰会，以及其他活动。

Confidence Computing Summit, others.

而这些就是我们一直在讨论的企业。

And that's all the enterprises we're talking about.

其中一个主题就是控制。

One was like control.

我们怎么能更好地控制这一切呢？

How can we like better control this?

因为他们总是会担心这个问题。

Because they're, they're always gonna be worried about that.

但更大的问题是成本。

But then the bigger one was like cost.

他们觉得在这上面花太多钱了，得把成本控下来。

They're like, we're spending too much money on this and we've gotta get it under control.

对。

Yeah.

那肯定意味着几件事。

So that means a few things for sure.

他们还在犹豫投资回报率，因为如果他们心里有数，觉得这东西确实能带来不少收入，

That they're still not sure about the ROI because they, if they were and they were like, okay, we're using this and it's driving a lot of revenue for us.

他们就会继续加大投入，就像那些超大规模云厂商在做的那样。

They'd be like, keep pushing the gas, which is what the hyperscaler are doing.

他们的情况是，连机器都买不够，根本跟不上现在的需求。

They're like, we can't even buy enough machines to scale up to the demand we have right now, the hyperscalers.

但真正在拉动需求的企业，对吧？

But like the enterprises who are the ones actually buying the demand, right?

他们就像是举了白旗。

They're like throwing up the white flag.

他们说，我们在AI上花得太多了，得想办法解决这个问题。

They're like, we got, we're just spending too much on AI and we need to figure out, we need to figure this out.

我知道，确实。

I know, yeah.

所以，你之前问了我一个问题，因为我跟你提过我遇到的那家公司。

So, but then you had this question for me, because I told you about this company that I met.

对。

Yeah.

我们录这期节目的时候正在参加AI4大会。

We're AI4 as we're recording this and the AI4 conference.

那家公司真的很棒，但他们告诉我，他们有很多工程师在token上的花费，超过平时1.5倍。

And there was this company on that, like really, really great company, but they had told me, they shared with me that they had not a small number of engineers who were spending more, they were spending 1.5 X in tokens.

太疯狂了。

Crazy.

他们的总薪酬。

Their total compensation.

太离谱了。

Crazy.

所以咱们假设一下，如果他们年薪是一万或者两万的话，咱们就说两万吧，毕竟现在工程师的工资都很高。

So if we'll just say, if they were making, you know, a 100 or $200 let's say they're, they're making, because engineers are expensive this day, we'll say we make 200 ks.

这意味着他们正在产生，

That means that they're making,

就是他们在令牌上的花费达到了三十万美元，三百

that they're spending $300,000 300

千美元，按年化计算，这是他们的总支出。

ks in tokens per, you know, annualized, Their total.

然后你就问，这投入产出比是多少？

So, and you were like, what is there, what's the ROI?

他们吗？我当时就想着，他们跟我说的是他们不关注单位经济。

Like, are they And thinking I was like, well, what they've told me is like, they're not focused on the unit economics.

他们说，他们看到了非常明显的加速。

They're like, they're seeing so much acceleration.

对。

Yeah.

而且他们还是盈利的。

And they're still profitable.

所以你问，那投资回报率呢？

And so you're like, so what's the ROI?

那些是什么？我表示我也不知道。

Like what are the, and I'm like, I don't know.

我觉得那是一个非常特殊的情况，是个特例。

I feel like that is also a very, that's a special case.

而且我觉得他们几乎算是运气好的那一批人。

Like I feel like they're almost the lucky ones.

我觉得这种情况不会适用于所有人。

Like I don't think that is gonna be the case for everybody.

我觉得我们最后可能会落在一个中间地带。

And I think actually that's where we land somewhere in the middle.

确实会有那种情况，有些企业觉得在 token 上花多少钱都行，因为 AI 能帮他们加快交付、创造价值之类的。

There are gonna be cases like that where there are enterprises that are like, we can spend as much money on tokens as we want because AI is helping us ship and bring value and whatever.

那家公司也不是上市公司，所以

That company is not public too, so that

我明白。

I know.

对。

Yeah.

所以他们完全不在乎。

So that, like, they're like, we don't care at all.

他们想花多少钱就花多少钱。

We could just spend as much money.