20VC：Instinct以100亿美元估值融资10亿美元 | AMD以82亿美元收购李飞飞的World Labs | Meta挖角MongoDB首席执行官 | 贝索斯基金募资57.5亿美元 | Oura撤回IPO，Nubank瞄准80–120亿美元收购Monzo

但我的一个合伙人主持那次会议时，随口问了个关于风险投资的问题，说风险投资接下来会怎么样？

But one of my partners who was doing the meeting asked just a tangential question on venture, what's going to happen in venture?

然后他就轻声说，就像经济学家那样，"回报会高度分化。

And he just quietly said, as economists do, "Returns are going to be highly skewed.

随着AI的发展，波动性会加大，你们很多人会失败。"

Variance is going to go up with AI, and many of you will fail."

然后他就直接跳到下一个问题了。

And then just moved on to the next question.

你们有些人会发财，但很多人会失败。

Some of you will get rich and many of you will fail.

好吧，泰勒，谢谢你帮我们理清了这一点。

Well, thank you for clarifying that, Tyler.

杰克，你说得对。

Jack, you're right.

这就像波动性最大的时期，这也就说明，即使你下注时看到了上行空间，有足够的资金做20次投资才是关键，这就是为什么我不理解，我们之前聊过这个，也是为什么你说得对，Instinct虽然价格高，但仍然是早期阶段的赌注。

It's like the highest variance time, which speaks-... to even when you're doing these bets with the upside, having enough to do 20 is key, which is why I don't get the, we've talked about this before, it's why you're right about Instinct being an early-stage bet despite the price.

现在还没到可以讨论集中度的时候。

It's not yet at the stage where you can have the concentration discussion.

真正在讨论集中度的人，其实是在说OpenAI和Anthropic最后两轮IPO前的融资。

The folks having the concentration discussion are really talking about OpenAI and Anthropic at the last two pre-IPO rounds.

但你说得对，即使是100亿的估值，如果还是现在这个情况，你也不会想投30%进去。

But you're right, even at a 10 billion pre, if it's where it is today, you don't want to put 30% in.

我也觉得这一轮不是让我担心的。

I also think this is not the round that worries me.

如果你看看Instinct的那些用户群，虽然我没见过，但我猜使用量肯定很惊人。

If you were to look at the Instinct cohorts, which I haven't seen, but I imagine the usage is incredible.

我猜使用的深度是我们以前从没见过的。

I imagine the depth of usage is like we've never seen before.

让我担心的是，三周内有三轮融资，中间没有实质性的进展，也没有数据表明有什么变化。

What worries me is when you have three rounds in three weeks with no material movement in between and no data to suggest there's been anything different.

这才是让我担心的。

That's what worries me.

那些说什么"才14个人就估值100亿"的人，就跟当年有人说"一个公司凭什么值10亿"一样。

The people who say, "Oh, 10 billion, 14 people," it's just like people who said, "A billion for what?

当年说Instagram值10亿太离谱的人就是这批人。

For Instagram was stupid."

这就相当于，对，你今天的表现可不怎么明智啊。

It's like, yep, you don't look that smart today.

当时脸书花160亿美元收购了WhatsApp。

16 billion for WhatsApp.

没错。

Exactly.

但你知道我最喜欢什么吗？

But you know what I love?

我喜欢的是，我确实请到了你的合伙人Ev Randall上节目，他说：'啊，我们会回头看Andreessen、General Catalyst和Lightspeed这些机构。'

I love the fact that I actually had your partner, Ev Randall, on the show, and he said, "Ah, we're going to look back at Andreessen and General Catalyst and Lightspeed."

他还转发了这条推文，这显得特别谦逊，完全不像风投圈的风格，他当时说：'也许我之前说Andreessen在这轮不会做到5倍回报是错的。'

And he retweeted it, which is so humble and non-VC, where he was like, "Maybe I was wrong to say that Andreessen won't do a 5X in this vintage."

因为马丁·卡萨多先生让我们显得很没面子，他搞了OpenRadar，昨晚又跟李飞飞搞了World Labs，然后还有Cursor。

Because Mr. Martin Casado is making us look bad because the man has had OpenRadar, obviously last night had Fei-Fei with World Labs, and then he had Cursor.

所以今天的头条新闻是，AMD以82亿美元的全股票交易收购了李飞飞的World Labs。

So for the news story of the day, AMD buys Fei-Fei Li's World Labs for $8.2 billion in stock.

这是这家新实验室成立两年半以来的首次重大退出。

That first big neo-lab exit two and a half years into the company's journey.

咱们先聊聊这位创业者吧。

Let's focus on the entrepreneur first.

我觉得这对李飞飞和她的团队来说很棒。

I think it's great for Fei-Fei and the team.

我只见过她一次。

I've only met her once.

我觉得她人特别好，特别谦虚。

I thought she was awesome, very humble.

我很喜欢她的书，如果你读过的话。

I love her book, if you read it.

这是一本非常感人的书，讲的是移民到美国的故事。

A very moving book about being an immigrant to the US.

我自己也是移民，所以深有体会，而且也是从一无所有开始的。

I was myself, so I know it, and kind of came up from nothing.

有人指出，这不是一夜成名的故事。

As someone pointed out, it's not an overnight success.

她辛辛苦苦搞出了ImageNet，整个AI复兴就是被她那个项目点燃的，因为2012年ImageNet比赛的冠军是伊利亚和另外两三个人，他们的模型用神经网络直接碾压了其他所有人。

Cranked through ImageNet, kicked off the whole damn thing because the entire AI resurgence was really kicked off when that ImageNet project that she built, when the winner of that was Ilya and two or three other people, I think in 2012, where their model using neural nets just kicked ass and blew past everyone else.

那是第一次让人看到，深度学习这项技术会彻底颠覆一切。

And it was the first hint that this technology, this deep learning technology, was going to blow everyone away.

她搭建了这一切，却没有靠它赚钱，后来在斯坦福教书，在谷歌工作，然后中年辞职，两年前开始创业，结果大获成功。

And she kind of set that up, didn't monetize that, has worked at Stanford, worked at Google, and then midlife quit, did this two and a half years ago, and nailed it.

所以我觉得这真是个了不起的故事。

So I think it's just an awesome story.

这又一次提醒我们，又一个极其成功的移民故事，来到美国，赚得盆满钵满。

Just a reminder, yet another wildly successful immigrant story, coming to the US, making a ton of money.

所以我为她感到兴奋。

So I'm stoked for her.

风险投资的事我们稍后再谈，但这确实是个很棒的故事。如果AI史上最大的繁荣都发生了，而李飞飞却没拿到一张大支票，那也太说不过去了。

We can come to the venture money later, but it's just a great story, and it would be pretty criminal if we had the biggest boom in AI history, and Fei-Fei Li didn't get a big ass check.

所以我觉得因果报应已经恢复了，挺好的，对吧？

So I think karmic justice has been restored, and it's good, right?

把支票兑现吧。

Cash the check.

嗯，就一件事。

Well, just one thing.

AMD是你爷爷的投资之一，对吧？

AMD was one of your grandpa's investments, right?

他和他在仙童的伙伴们，或者我不太清楚具体细节——但这可是个老古董了。

He and his buddies from Fairchild, or I don't know what the exact-- But this is an oldie moldy.

今年涨了279%，市值达到一万亿。

It's up 279% this year to 1 trillion.

所以我都不知道他们是要继续做3D世界模型，还是就只是成为英伟达模型团队的对应部门。

So I don't even know whether they're going to continue the 3D World models, or they're just going to be their counterpart to NVIDIA's Modal team.

但不管怎样，就为了那8%的市场份额，8%？

But whatever it is, for 8% of that market, 8%?

找一个世界级的团队来确保那279%的涨幅能继续下去，如果团队选对了，那这钱花得值，对吧？

Getting a world-class team to make sure that 279% run continues, it's cheap if it's the right team, right?

就像杰克说的，这是个关键时刻，对吧？

It's a moment in time, to Jack's point, right?

如果AMD今年只涨了3%，他们才不会花80亿美元呢。

If AMD was up 3% this year, they ain't going to be spending $8 billion.

但这次涨势确实惊人，可你得维持住，对吧？

But this is a run that it's epic, but you've got to maintain it, right?

他们必须得做到英伟达之下的第二名，或者不管目标是什么。

They've got to be number two to NVIDIA or whatever the goal is.

而且这是个越跑越快的习惯性轨迹。

And it's a habit trail that keeps going faster.

所以我当然愿意拿出市值的8%来让好日子继续下去。

So I'd certainly invest 8% of my market cap to keep the good times going.

而且我觉得这只是个开始，后面还会有更多。

And I think it's the first of a few.

这个话题我们聊过好几次了。

We've talked about this a couple of times.

你看这些投资，说实话，我觉得那些世界实验室类的公司，不是特指世界实验室，而是那类公司，单独来看，不太可能像Anthropic和OpenAI那样有那种发展轨迹，就是很快从一项惊人的AI技术变成能赚钱的聊天产品，然后天哪，又变成能大赚特赚的编程工具，对吧？

You look at these investments, and candidly, I don't think the World Lab companies, not the World Lab in particular, but those World Lab-type companies, would on a standalone basis have anything like the trajectory of Anthropic and OpenAI, where there was a very quick path from here's an amazing AI technology to here's chat and that can monetize, and then, oh my God, here's coding and that can monetize like a mother, right?

我觉得那条路不会那么清晰，这也是为什么过去几年我对这些交易一直有点担心。

I don't think the path would've been as clear, which is why I've been a bit nervous about those deals over the last couple of years.

但现在回头看，我觉得已经很明显的一点是，所有大型基础模型公司可能都在市场上寻找收购某种机器人基础模型的故事。

But in retrospect, the thing I think that's now become obvious is I think all the big foundation model companies are probably in the market to acquire some kind of robotic foundation model story.

就像你说的，AMD只是想跟NVIDIA保持相关性。

I think that, as you say, AMD just wants to be relevant to NVIDIA.

所以我觉得如果出现大量收购兴趣，我也不会感到意外。

So I think that it wouldn't surprise me if there was a ton of acquisition interest.

现在还有100家Neo实验室，所以你必须在10家获胜者中，但我认为如果市场继续维持，未来6到12个月会有很多大规模收购。

There's still 100 Neo labs, so you got to be in the 10 that win, but I think there will be a bunch of these big ass acquisitions over the next six, 12 months if the market continues to hold.

你觉得呢？

Do you?

因为我有一个非常聪明的法国合伙人保罗，他天生就非常愤世嫉俗，他写了一篇关于102家Neo实验室的报告，这些实验室总共筹集了超过700亿美元。

Because I invest with a brilliant French partner, Paul, who is incredibly cynical just by nature of being French, and he wrote this report on 102 Neo labs, $70 billion plus raised.

我的问题就是，其中有多少能被收购？

And my question is just how many of them can get acquired?

当只有10到12家的时候，是的，那还好。

When there were 10 to 12, yeah, it'll be okay.

102家？

102?

我们都能有下行保护吗？

We all going to have downside protection?

当所有Neo实验室刚开始起步、趋势起飞的时候，很容易就说："这完全是疯了。"

It was very easy when all the sort of Neo labs got started and the trend took off to just say, "This is total insanity."

而且你说得对，Harry，我们现在还不知道整个事情的净资产负债表会是什么样。

And to your point, Harry, we don't yet know what the net balance sheet of the whole thing will be.

但我们确实开始看到像World Labs和其他一些公司的数据点，表明有真正的兴趣。

But we are definitely starting to see data points like World Labs and many others, where there is real interest.

而且我认为，现在的情况和我见过的任何时候都不一样，也许历史上也有过这样的时刻，但现在就是有很多收购方。

And I think part of what's so different now than certainly any time I've ever seen, maybe there have been other points in history like this, but there are just a lot of acquirers.

有10家公司有能力并且愿意做100亿美元的收购，这真的太不一样了。

There are 10 companies that can do $10 billion acquisitions and want to, and that's just so different.

而且这比上市容易多了，也更快，不用经历那些漫长的周期和焦虑，这真的改变了局面。

And it's much easier than going public, and it's quicker, and you don't have all of these long cycles and the hand-wringing, and that has really changed things.

当然，Martin的表现非常出色，这也表明在这些基础设施实验室类型的公司里，还有很大的发展空间。

And obviously, extremely impressive run by Martin, and it's just shown that in these infrastructure lab-type companies, there's a lot of room to run.

所以，显然，我并不比其他人更清楚事情会如何发展。

So I don't know how it will go, obviously, any better than anybody else.

但在这一点上，我不太愿意过于怀疑。

But I'm hesitant to be too skeptical at this point

考虑到现在的情况，确实如此。

point given what's happening.

我同意这一点，而且我觉得当你面对那100家公司时，是的，可能只有10到20笔收购，但这就是人们用“共识”这个词来表达负面意思的地方。

I agree with that, and I think when you deal with the 100, and yes, there might only be 10 or 20 acquisitions, but this is where people use the word consensus in a negative sense.

但毫无疑问，学历和资历很重要。

But there is no doubt that credentialism matters.

比如你请到了李飞飞。

Like you got Fei-Fei Li.

我都能写出那份新闻稿了。

I could write that press release.

整个AI领域的开创者。

Originator of the whole AI thing.

这有点像OpenAI和Anthropic里的顶尖人才。

It's a little like the quality folks at OpenAI and Anthropic.

所以我觉得另一部分在于，还有Leo Labs，要确保你拥有的是纯粹的技术资产，而不是一个粗糙的团队。

So I think the other part of this is, and then Leo Labs, making sure you have the pristine tech asset versus the gritty team.

我在应用层做的很多事情，我其实很喜欢那些接地气的团队。

A lot of the stuff I do in the apps level, I love my gritty teams.

我觉得像构建世界模型这样的东西，你需要有经过验证的技术成功，因为这既是它需要交付的东西，也是它需要具备的，才能成为理想的收购对象。

I think for something like building a world model, you need to have proven technical success because that's both what it needs to deliver and frankly what it needs to navigate and be the desired acquisition candidate.

我不确定整体上是不是这样，但这只是我观察到的情况。

I'm not sure that's not true overall, but it's just something I've observed.

我昨晚和一位全球顶尖的CIO聊天，他直接跟我说：“老实说，兄弟，你真的觉得现在没有十亿美元规模的基金还能玩转风投吗？”

I was talking to one of the best CIOs in the world last night, and he just said to me, "The honest takeaway, dude, you really think you can play in venture with less than a billion-dollar fund now?"

坦白说，当我看到这里的很多公司，像Fei-Fei那家，第一轮融资就差不多是6500万。

And candidly, when I look at many of these companies here, the first round for this and Fei-Fei was like 65.

前几天我的团队跟我说：“我们找不到低于一亿美元的项目。”

I had my team say to me the other day, "We can't find anything under 100 million."

我说：“哇，种子轮价格真贵啊。”

I said, "Wow, seed prices are expensive."

他们说：“不，不，是一亿美元的融资规模。”

They said, "No, no, 100 million round size."

稍微反驳一下，我觉得这项技术奇怪的地方在于，我也一直在想这个问题，作为一个9亿美元的基金规模。

Pushing back a little, I think the weird thing about this tech is, because I've been thinking about this a lot too, as a $900 million fund size.

你有两种类型的公司。

You've got two kinds of businesses.

有一种公司，你可以在不到4亿或5亿美元的情况下推出产品，比如像NeoLab那样的大规模融资。

You've got businesses where you can ship a product on less than $400 or $500 million, like huge NeoLab raises.

但你也必须承认下面这一点，这也是事实。

But you've also got to admit the following, which is also true.

有些公司利用NeoLab、OpenAI和Anthropic的技术，用相对较少的资金就推出了产品。

You've got companies that use that NeoLab and that OpenAI and Anthropic technology that are shipping products on relatively little capital in.

你们在Higsfield。

You guys are in Higsfield.

你们会遇到很多这样的人，他们说：“我们拿了1000万，但只花了300万就推出了产品，然后客户就来了，天哪，我们银行里还有500万，但我们还是会去融5000万，因为我们可以。”

You run into a bunch of these guys who are like, "Well, we took $10 million, but we shipped the product for 3 million bucks, and then the customers took off, and shit, we still got 5 million bucks in the bank, but we're going to raise 50 anyway because we can."

所以并不是所有事情都需要1亿美元来启动一些GPU。

So it's not like everything is $100 million just to spin up some GPUs.

有很多公司用一千万美元就能走得很远，然后可能再融一个亿，但那是出于别的原因，因为有能力融，对吧？

There's a whole bunch of $10 million got us a long way, and then maybe we're raising 100, but for different reasons, because we can, right?

所以情况并不一致。

So it's not consistent.

这几乎是两个不同的世界，但也说得通。

It's almost two different worlds, and it kind of makes sense.

前面那些融了一百亿的大公司，他们带来的技术突破，让其他人能用相对少的钱做出惊人的事情。

The big technology lift up front from the guys raising $10 billion is what's setting everyone else up to do amazing shit on relatively little.

是的，我觉得人才成本比以前高太多了，因为人才有很多其他选择，所以你再也不能只融两三百万美元就开干了。

Yes, I would say that just bunny talent costs are so much higher than they've ever been because of a lot of the alternative options for the talent that you can't do a $2 to $3 million round anymore.

你知道这有个挑战是什么吗，Harry？

You know what the challenge to that, though, Harry?

我同意这个数字，而且Rory上周也提过，我想听听Jack的想法，就是自2010年以来，风投领域的名义通胀率大概是2.5倍。

I agree with the math, and Rory made the point last week, and I want to hear Jack's thoughts, that I think nominal inflation's like two and a half X in venture since 2010.

我有时候会把名义和实际搞混。

I sometimes get my nominal and non-nominal confused.

确实如此。

All true.

但如果你看看那些还没被发现的潜力项目，我觉得种子轮，或者叫它预种子轮、预预种子轮、甚至叫受孕后轮，不管怎么叫，本质上还是那回事——三到四个人，撑过18个月，需要多少钱？

But if you look at the undiscovered gems, okay, I think seed rounds, and maybe they're called pre-seed or pre-pre-seed or post-conception, but whatever we call it, they're still what does it take three to four folks to get 18 months down the road?

其实是一样的。

It's the same thing.

而且如果你运气好，能从哪拿到一百万美元的免费代币，或者从OpenAI、Anthropic拿到五十万来创业，再加上其他各种资源，那么几百万美元还是能走很远的，特别是如果你愿意在旧金山合租一个月八千美元的公寓。

And if you're coming out and you're lucky enough to get a million dollars of free tokens from whatever, a half million from OpenAI or Anthropic to start a company, you get all these other things, you can still get pretty far for a couple million bucks, especially if you're willing to all share in a $8,000 a month apartment in San Francisco.

所以我不是说那个数字不对，但事实是，我觉得现在花两三百万美元能做成的事，和十年前差不多。

So it's not that I disagree with the math, but the truth is you can do as much, I think, for 2 to 3 million bucks as you could 10 years ago.

而且如果除了演示日之外，没人急着给你投钱，那么种子轮的自然资金规模可能还是那么多——我今天把术语搞混了，但就是种子轮。

And if you don't have folks dying to give you capital outside of demo day, that may still be the natural atomic amount of capital for, I get the terms wrong today, but a seed round.

应该还是两三百万美元。

It still should be 2 to 3 million bucks.

对，在某些市场是这样，不过我这人有点爱较真。

Yes, in certain markets, just at the risk of being a pedant.

是的，在某些市场里，你可以借助其他一切来推出产品，但进入NeoLab领域至少要几亿美元，而进入半导体领域则需要5亿美元。

Yes, in certain markets where you can ship a product leveraging everything else, but it's a couple of hundred million minimum to enter the NeoLab space, and it's 500 million to enter the semiconductor space.

我只是想说，杰克的观点里有个奇怪的地方，就是融资方式千差万别，你得好好理清。

I'm just making a point is that the weird thing to Jack's point, there's such a range of different financing things you got to get your head around.

你想投1亿美元给一家半导体公司，结果还没流片估值就20亿了。

You want to put 100 million in a semiconductor company to still pre-tape out at 2 billion.

让我想想。

Let me think about that.

然后明天，再看看给一家软件公司500万美元的种子轮，也许他们就能推出产品，获得消费者收入，仅此而已。

And then tomorrow, let me look at $5 million for a seed round for a software company where maybe they can get a product out the door and consumer revenue on nothing more.

哈里开场说的那个观点，我基本同意，就是说早期融资已经变得非常极端，原因有两个。

The characterization that Harry opened up with, I mostly agree with, which is basically that the initial rounds have gotten so extreme, and there's two reasons.

一个就是像你们说的，有些实验室，少于2亿美元什么都做不了，所以那就是第一轮融资。

One is that there's like you all are saying, there's these labs where you can't do anything for less than 200 million, and so that's the first round.

另一个原因是，有一批创始人，在硅谷圈子里人脉很广，他们不一定需要去融5000万美元的种子轮，但他们能融到。

The other reason is there is a set of founders that are very in the Silicon Valley network that don't necessarily need to raise a $50 million seed, but they can.

所以他们跳过第一轮，跳过以前那个六百万到四千万的轮次，干脆就不融了。

And so they skip the first round, they skip the six at 40 or whatever that round used to be, and they just don't raise it.

或者他们可能只融三十万，先撑一个月，然后再去融大钱。

Or maybe they raise $300,000 just to get a month in, and then they raise the money.

所以在很多赛道上，我觉得那种轮次基本上已经消失了。

And so in many lanes, I think that round has kind of evaporated.

所以我觉得市场上有一批人，传统的种子投资，就是写三百万到六百万的支票，占百分之八到十五的股份，那种模式我觉得已经完全行不通了，或者说根本不存在了。

And so I think there is a cohort of the market where traditional seed investing, where you're going to write 3 to $6 million checks by 8% to 15%, where I just think that is fully broken/just isn't there anymore.

话虽如此，市场上还有另一部分，Jason，我记得我刚开始做Lattice的时候，大概十二年前或者十年前，就经常看你的博客。

All of that said, there is another part of the market which, Jason, I remember reading your blogs way back when I was starting Lattice- ... 12 years ago, or 10 years ago, whatever.

当时很多内容就是帮我搞明白风险投资的数学，而且那是个小基金。

And a lot of it was just trying to get my head around what the venture math was, and it's a small fund.

你买一家公司百分之十的股份，希望它值个十亿或二十亿。

You buy 10% of a company, you hope it's worth a billion or two.

你就能拿回一到两亿。

You get 100 to 200 million.

这种事情在一个篮子里做几次就行了。

You do that a couple of times out of a basket.

我其实觉得这种模式可能还存在，但它的存在方式并不是我们在推特和头条上看到的那种。

I actually think that that might still exist, but the way that it exists is not in the things that we're reading about on Twitter and in headlines.

这些轮次是那些因为某种原因暂时没法进行那种融资的公司。

It is these rounds where at the moment they, for one reason or another, can't go do that round.

他们要么市场不对，要么公司形态不太合适，然后他们要么转型，要么获得的增长比你预期的要好，然后公司就慢慢变大了。

They're either not in the right markets or the shape of the company isn't quite right, and then they somehow pivot, or they get more traction than you thought, and it just kind of gets bigger.

但这个过程是慢慢发生的。

But it happens the slow way.

我们现在已经习惯了这些故事在一两年或三年内就上演。

We have now gotten accustomed to all of these stories playing out in a year or three years.

我们说的是Instinct。

We're talking about Instinct.

就像这些故事发展得很快，它们就变大了。

It's like these stories play out so quickly, they get big.

但即使你拿像Lagora这样的公司来说，显然我知道Harry很了解这家公司，这些故事仍然在短短几年内上演。

But even if you take a company like Lagora, which obviously I know Harry knows well, and these stories still play out in a short number of years.

而且我们都习惯了这一点。

And we've all gotten used to that.

但风险投资的另一部分需要10到20年的时间，软件公司只是慢慢积累。

But there is another part of venture that takes place over 10 to 20 years, where a software company just kind of compounds slowly.

所以，我确实认为今天有一个30轮融资在发生，到2041年，那家公司会在警察局、消防部门或图书馆的市场中悄悄增长吗？

And so do I think that there's a three at 30 round happening today, where in 2041, that company is just going to have quietly compounded in the market of police or fire departments or libraries?

是的，我可能这么认为。

Yeah, I probably do.

所以这只是市场的不同部分，而且确实存在很大的分化。

So it's just a different part of the market, and it has really had a big dispersion.

我只是不认为那是一个行业。

I just don't think that's an industry.

我认为那只是一个异常现象，而我们所在的才是一个行业。

I just think that's an anomaly, and we're in an industry.

我可不想赌三年后可能发生的转型，那可能会让公司变得不伦不类，然后再重新走上正轨。

And I don't want to bet on a potential pivot three years down the road that might lead to a misshapen company that then becomes shapen again.

我可赌不起那种慢吞吞的转型。

I can't bet on slack pivots.

杰克，每次我想跟哈里讲这个故事，他都会直接把我怼回去。

Jack, every time I try and pitch the story to Harry, he clubs me just like that.

那都是胡扯。

It's bullshit.

我倒是愿意相信，因为我们在这样的公司上做过一些很成功的投资。

And I want to believe in it because we've made some magnificent bets on companies like that.

但我也同意，目前的市场格局是80%的快节奏，20%的慢节奏，这很合理，因为2022年出现了一个断层，之前的一切都过时了。

But I do agree that the table at the moment is 80% the fast action table and 20% the slow action, which makes sense because in 2022, there was a discontinuity, and everything before that became obsolete.

所以，从定义上讲，现在这个阶段，也就是四年后，任何公司都必然是快节奏的，因为它们都是从22年开始的。

So by definition, anything at this stage, four years in, is fast action by definition because it started in '22.

而且，如果你想走那种模式，你的业绩数据在很长一段时间内都会很难看。

Also, if you want to go for that model, your numbers will be crap for quite a long time.

而且我们总是忘了这是个机会成本游戏，人们可以把钱投到你们的Ceragos和Benchmarks这样的公司里。

And we always forget that we're in an opportunity cost game where people can put money into your Ceragos and your Benchmarks of the world.

而你们公布的数字简直是前所未见的。

And while you guys post numbers that are just never before seen.

杰克，我昨晚还在给切坦发消息，说：“天哪，快停吧。

I was messaging Chetan last night, Jack, being like, "Oh, God, just stop.

最新那只基金也太夸张了。”

The latest fund is just ridiculous."

我们就是机会成本。

We're an opportunity cost.

你进不了那些基金，但你可以努力去争取。

You can't be in those funds, but you can try and be in those funds.

所以要在别人都没注意到的地方下功夫。

And so trying to be smart where no one else is.

同意。

Agreed.

巨大的吸力声，基本上钱就是信号，价格就是信号，价格在发出信号，大家都往这边来。

The giant sucking sound, basically money is a signal, price is a signal, and price is sending a signal, everybody go right here.

然后所有人都会往这边来，因为这就是价格的作用。

And everyone will go right here because that's the job of price.

哈耶克先生会很高兴的。

Mr. Hayek would be happy.

关于安静复利的问题，就像杰克说的，在我看来，作为一个从2012年就开始推销安静复利的人，问题在于它们就是不稳定。

The problem with the quiet compounder, to Jack's point, in my view, as someone who's pitched quiet compounding since 2012, is they're just not stable.

你必须更快地构建更多的软件，对吧？

You have to build so much more software so much more quickly, right?

杰克和我都是Owner的董事会成员，那是一家收入超过一亿美元的火箭公司。

Jack and I are both on the board of Owner, which is a rocket ship north of 100 million in revenue.

但看看他们今年得建多少该死的软件，对吧？

But look how much effing software they have to build this year, right?

有多少人能搞定这件事？

How many folks can pull that off?

有多少人能筹到资金呢？

How many folks can raise the capital?

而且就算竞争对手多十倍，对吧？

And even if you have 10 times more competitors, right?

我还有另一家公司，营收一亿。

I have another company, 100 million.

上次董事会上他们放了一张竞争对手的幻灯片。

They put up a competitor slide at the last board meeting.

其中八九个我都没听说过。

I never heard of eight or nine of the ones.

我原以为我们只有一个竞争对手。

I thought we had one competitor.

结果发现我们有九个，对吧？

Turns out we have nine, right?

营收都到一亿了。

At 100 million.

所以并不是我不愿意相信。

So it's not that I don't want to believe.

先别管筹集LP资金更难这件事，那是Harry的观点。

Forget about the fact that it's harder to raise LP capital, which is Harry's point.

我觉得大家对这种不稳定性反思得不够，对吧？

This instability is something that I think people under-reflect on, right?

这就是问题所在。

That's the problem.

太不稳定了。

It's so unstable.

太不持久了。

It's so undurable.

换个角度说，对。

Put it another way, yeah.

如果你做的是复利投资，那相应的回报应该是低风险。

If you're making a compounding play, the quid pro quo should be low risk.

而且如果世界是这样的，科技环境变化如此之大，你得到了复利增长，虽然不是超高速增长，但风险水平却一样高，那这本质上就是个不划算的游戏。

And if the world is such that the tech environment is changing so much that you get the compounding, not the hypergrowth, but you get the same level of risk, that by definition is a suboptimal game.

是的，对我来说，我同意哈里观点的一个最主要原因就是，现在人才争夺战简直疯狂到了极点。

Yes, and sort of to me, one of the most dominant reasons why I do agree with Harry here is it's just like it's a talent war in the most insane way right now.

这些机会对优秀人才来说实在太有吸引力了。

It's just these opportunities are just too enticing for great people.

这正好是个完美的过渡。

Which is a perfect segue.

完美的过渡，霍伊。

Perfect segue, Howie.

来吧，你懂的。

Come on, figure it out.

你是在暗示杰克给我铺好了路，罗里，让我轻松得分——

Are you suggesting that Jack teed me up there, Rory, for a slam dunk-

是的。

I am

……而我却没有接住？

... and I didn't take it?

但你确实没接住。

But you were missing it.

哈里，我也没看到，而且我也算是个播客主持人，所以这确实有点难。

Harry, I didn't see it either, and I also try to be a podcaster, so that was a hard one.

杰克，我觉得我们就是比罗里慢半拍，对吧？

Jack, I think we're just a bit slower than Rory, okay?

奥德里斯科尔教授，他反应很快。

Professor O'Driscoll, he's very sharp.

我们应该注意——

We should be mind-

他反应很快。

He's very sharp.

停。

Stop.

别扯那些没用的了。

Stop the bullshit.

跟观众和听众们说清楚吧，事情是这样的：MongoDB是一家独立上市的上市公司，市值高达200亿美元，非常成功。这位CEO上任还不到九个月，就收到了你们的邀约，立马接盘，跳槽去Facebook/Meta负责企业部门。结果MongoDB股价一天之内跌了20%，然后Dev又回来接手了。他不是创始人，但之前当了很长时间的CEO，现在又回来掌舵了。

For the viewers, our listeners, we should just say, look, what happened here is the CEO, the chief executive officer of a standalone independent public company, MongoDB, a very successful $20 billion market cap public company, who just took the job less than nine months ago, got an offer from yous, hit the bid, moved over to Facebook/Meta to run their enterprise division, and Mongo stock dropped 20% in one day, and then Dev stepped back in, who'd been not the founder, but the CEO for a long ass time, and is now back running it.

但杰克的意思是，金钱的影响力实在太大了。

But to Jack's point was, the money can exert such a powerful influence.

这笔钱大到足以让这个人放弃最高职位，去Facebook上班，我猜那个offer肯定诱人到无法拒绝。

It was powerful enough to persuade this guy to quit the top job, go work at Facebook, because I assume the offer was just earth-shatteringly compelling.

杰克，你的观点就是，当你手握热门股票，有势头的时候，你就能开出让人无法拒绝的条件，想挖谁就挖谁，这就是个例子。

And that's your point, Jack, is that when you have the hot stock, when you have the momentum, you can make people offers that just allow you to take whatever talent you want, and this is an example of that.

你觉得真的是钱的问题吗？

Do you think it is the money?

如果你一年赚5200万，也许吧，我不知道，一年5200万美元，但我想如果你一年赚5200万，再赚1.2亿，那当然不错。

When you're earning 52, maybe, I don't know, 52 million bucks a year, but I imagine if you're earning 52 million bucks a year, earning 120 million bucks a year, that's nice.

但我觉得对他来说，当他看到这个机会时，他会想：“天哪，我有机会产生更大的影响了。”

But I think for him, when he looks at this, he's like, "Holy shit, I have the chance to impact far greater."

不是吧。

Nah.

我觉得他原来的薪酬是5200万美元，而扎克伯格给他开出了5亿美元的薪酬，让他去负责企业业务。

I think it was a $52 million package, and he was offered a $500 million package by Zuck to run Enterprise.

十倍，轻轻松松。

10X is easy.

那得付出什么代价？

What does it take?

5.2亿，然后就这样，对吧？

520 million, then done, right?

没人愿意从CEO的位置上退下来，去当什么企业产品和代币部门的负责人。

No one wants to step down from CEO to be the chief executive enterprise products and tokens officer.

你觉得他一年能赚5亿美元？

You think he's paid 500 million bucks a year?

对，就是十倍。

Yeah, it's 10X.

就是十倍。

It's 10X.

我直接挂买单。

I'm hitting the bid.

挂买单。

Hit the bid.

我同意杰森的看法，因为你本来要说Mongo不刺激，而Meta企业业务会很刺激。

I'm just going agreeing with Jason, because you were going to go down some kind of Mongo's not exciting and Meta Enterprise would be exciting.

我觉得这没错，但反过来，当过CEO的人不会回去不当CEO。

I think that's true, but conversely, no one who's been a CEO goes back to not being a CEO.

这实在太难了，对吧？

It's just so damn hard, right?

所以答案肯定是那个报价太诱人了。

So the answer is it must've been just a compelling offer.

我是说，老兄，我们见过尼克·克莱格，他当过副首相，后来去给马克·扎克伯格打工……

I mean, dude, we saw Nick Clegg, who was deputy prime minister, go and work as Mark Zuckerberg's...

我是说，是的

I mean, yes

所以，你知道，那就是——

So like, you know, that's-

是的，我在努力。

Yes, I'm trying.

是的。

Yes.

说话小心点。

Be careful what you say.

我只是尽量不说尼克·克莱格、我们英国的戴维·卡梅伦、我们的马克·扎克伯格的坏话，但这太难了，所以我干脆就不说了。

I'm just trying to not be mean about Nick Clegg, our England, our David Cameron, our Mark Zuckerberg, and it's just too hard, so I'm just going to let it slide.

明白了。

Got it.

不过，我们应该从中得到什么启示呢？

What should we take from this, though?

实际上，现在连CEO这个角色，面对金钱时离职也变得很正常了？

That actually even the role of CEO is one where departure is normal now in face of money?

你应该听听杰克说的话。

You should take what Jack said.

这又回到了你之前聊的那个话题，因为他从你关于机会成本的评论接了过去。

It gets back to the conversation you were making, because he segued it off your comment on opportunity cost.

市场在发出信号，说唯一值得待的地方就是这些异常火爆的AI公司，而且它是通过价格来传递这个信号的，人们也在对价格做出反应。

The market is sending a signal that the only place to be is in these extraordinarily hot AI companies, and it's sending that signal via price, and people are responding to price.

对，对我来说，我觉得就是钱，还有那种关注度，那种时代潮流，就像白热化的中心，我觉得这对人来说太有吸引力了。

Yeah, to me, I think it is the money, and it's also the-- it's the attention, and it's where the zeitgeist, and it's like the white hot center, and I think that is so alluring to people.

这里还有一件完全独立的事情，就是这件事发生的方式，我离得太远，完全不了解细节，我觉得这本身就是一个独特的情况。

There's an entirely separate thing here, with kind of the way that this all went down that I'm not close enough to at all to know the details of, and I think that's kind of its own unique beast.

但总的来说，我就是觉得它太有吸引力了。

But in general, I just think that it is so alluring.

钱是其中一部分，但不仅仅是钱。

And the money is part of it, but it's not just the money.

而且这也是所有头条新闻里的东西，是我们都在用的产品，是我们所有家人都在谈论的话题，我觉得这真的太集中了。

It's also the thing in all of the headlines, it's the products that we're all using, it's the thing that all of our families are talking about, and I think that it's just so concentrating.

关于它，有很多短期的负面因素。

There's all of these short-term things that are negative about it.

我想说的一件事，从宏观角度来看是积极的，那就是人才流动率可能比我们见过的任何时候都要高得多。

One of the things that I will say, just kind of this is zoomed out, broad positive, is that the much higher transition rate of talent that I think we've ever probably seen.

其中一个好处是，现在人才感觉不那么被困住了。

One of the benefits of it is that talent feels very unstuck right now.

历史上有很多时候，优秀的人才被困在那些你可能不希望他们待的地方。

You have all of these times in history where great talent gets very, very stuck in places that you might not want it to be.

我们现在可能正处于市场流动性较高的时期之一，优秀的人确实在流向最重要的机会，这从社会层面来看可能是非常积极的，即使在这些短期情况下你会挠头想：'这到底是怎么回事？'

And we are probably in one of the higher liquidity moments in the market where great people are in fact going to the most important opportunities, which I think is probably on some societal level very positive, even if in these short-term situations you scratch your head and you're like, "What's going on here?"

你说得对。

You're right.

我往往关注钱，但你说得对，不仅仅是钱。

I tend to focus on the money, but you're right, it's not just that.

这就是时代精神。

It is the zeitgeist.

这一点我同意你和哈里的看法，是的。

I'll give you and Harry that point, yeah.

这也是这里比东方好的地方。

It's what's good about here even versus the East.

我不记得了，是不是有人在东海岸或英国休花园假？

I can't remember, was it someone on garden leave, either in the East Coast or the UK.

加州的好处就是，你完全没有这些限制。

The great thing about California, you got none of that.

就像我过个马路，第二天就能去新公司上班。

It's just like, I go across the street, the next day I start in a new company.

你在干什么，哈里？

What are you doing, Harry?

该开投资委员会会议了。

It's time for the investment committee.

你已经做过一次了。

You already did one.

你在整点的时候为Instinct做了一个很好的展示。

You did a great one for Instinct at the top of the hour.

哦，不，那不是投资委员会。

Oh, no, that wasn't an investment committee.

不，这周的投资委员会是关于Jev的。

No, the investment committee this week is Jev.

Jev正在市场上以100亿美元的估值融资。

Jev is in the market to raise at a $10 billion price.

Jason，欢迎来到会议室。

Jason, welcome to the room.

你能介绍一下，在种子轮大约2亿美元之后一周，我们是否应该以100亿美元的估值做这笔交易？

Can you please present whether we should be doing this deal or not at $10 billion a week after the seed at about $200 million?

当然，我们应该做。

Well, of course, we should do it.

价格是多少，100亿美元？

What's the price, $10 billion?

一开始我觉得应该投基金规模的5%，但看了最近20家风投的表现后，我决定建议投到基金规模的30%。

At first thought we should do 5% of the fund, but after the last 20 VC, I've decided to recommend up to 30% of the fund.

Jev已经占了OpenRouter流量的17%。

Jev's already 17% of the traffic on OpenRouter.

通过Vercel路由器的流量占了20%。

It's 20% of the traffic through Vercel's router.

当然，很多人会模仿它，就像很多人会模仿Instinct和其他公司一样。

Many people will, of course, copy it, just like many will copy Instinct and others.

但我们已经到了这样一个阶段：尽管Neo Labs很令人兴奋，Anthropic即将进行的IPO和OpenAI也很令人兴奋，但这些AI成本是不可持续的。

But we are reaching the point where, as exciting as the Neo Labs are, as exciting as Anthropic's pending IPO and OpenAI are, these AI costs are unsustainable.

不管Sonnet 5.5和最新的Opus是否更便宜，每天花10到12个小时在这些成本上都是不可持续的，而且竞争门槛只会越来越高。

It doesn't matter if Sonnet 5.5 and the latest Opus is cheaper, it is unsustainable to spend these costs 10, 12 hours a day and the competitive bar just goes up.

Jev的价格是它们的七十分之一，速度快100倍。

Jev is a 70th of the price and 100 times faster.

而且，就像我的好朋友杰克·奥特曼说的，变化的速度太快了，甚至可以说是爆炸性的，也许明年Jev就不是赢家了，但这正是我们必须做的投资。

And listen, as my good friend Jack Altman says, the pace of change is so fast or exploding so fast, maybe Jev isn't the winner next year, but this is exactly the kind of bet we have to do.

我们有百分之百的下行保护。

We have 100% downside protection.

不管怎样，总会有人把X OpenAI的这个团队挖走的。

Someone's going to scoop up this team out of X OpenAI no matter what.

在OpenRouter上占有17%的市场份额。

17% market share on OpenRouter.

我的意思是，尽管上周我极力推荐Instinct，但这周，这是我的项目。

I mean, as much as I advocated Instinct last week, this week, this is my deal.

我全力押注Jev。

I'm all in on Jev.

动用基金的30%。

30% of the fund.

我知道这有风险，但我想获得所有权。

I know it's risky, but I want to get the ownership.

而且我确实相信有50倍的上涨空间，能达到5000亿美元，所以我把决定权交给Rory。

And I do believe there's a 50x upside to $500 billion, so I defer into Rory.

我觉得我们可以投基金规模的20%到30%。

I say we do, well, 20% to 30% of the fund.

Jack，我得提醒你，不知道为什么，Harry就喜欢这种金句，而且播客上效果很好，所以我也没什么好说的。

Jack, I should warn you, for some reason, Harry likes this kind of soundbite, and it plays well on the pod, so who am I to judge?

Jack，你太懂我了。

Jack, you know me so well.

说到底，我就是个剪片子的，行了吧？

At the end of the day, I'm a clip monkey, okay?

你就是个剪片子的。

You're a clip bitch.

我想说，但回到正题，有意思的是，这件事你算算基本账，其实不算离谱。

I would say, but going back, the funny thing is, in this case, you can do some basic math and it's not crazy.

我们上周就聊过这个。

I mean, we talked about this last week.

现在每天的支出大概是一千亿美元。

Spend right now today is roughly $100 billion.

你分析一下，其中20%跟Jev有关。

And you do the analysis and 20% of that is relevant to Jev.

所以那就是两百亿美元的可及收入。

So that's $20 billion of accessible revenue.

假设他们五比一压缩，那就变成四十亿美元的可及收入。

Say they compress it five to one, that gets you to $4 billion of accessible revenue.

而且我们看到最惊人的是，开发者的采用速度快得惊人。

And what we've seen that's been amazing is the developer adoption's been lightning fast.

我的意思是，有传言说他们的数字到处都是，一千亿美元的年化运行率，而且考虑到代币几乎不值钱，那使用量可真是海量，对吧？

I mean, rumor has it they're at numbers all over the place, $100 billion run rate, and given that the tokens are half nothing, that's a shit ton of usage, right?

你确实可以很快达到十亿美元的收入线，基本上就是拿已经花出去的钱，每美元省下80美分。

You could credibly get to a billion-dollar revenue line relatively quickly here by literally taking money that's already been spent and saving 80 cents on the dollar.

所以Jason说的完全没错。

So Jason's comments aren't wrong at all.

我是说，他得学会，如果每周花掉基金的30%，三周后就没钱了，但除此之外——

I mean, he has to learn that if you spend 30% of the fund every week, in three weeks you're out of action, but other than that-

我们是Benchmark嘛。

Well, we're Benchmark.

我们只要给LP发一封邮件，就能搞到新基金。

We just go back to the LPs with one email and we have a new fund.

我正想说，Jason会想办法周转一下，最后总能搞定的。

I was going to say, Jason will just recycle something and it'll all work out somehow.

对，我们会周转一下。

Yeah, we'll recycle something.

你能在三周内搞定吗？

Can you flip it in three weeks?

我们还把Martin招进了团队。

We also recruited Martin to the team.

他会来帮忙的。

He's going to help.

我们就直接循环利用吧。

We'll just recycle.

恩斯坦，你能给LP发封邮件，说我们要成立新基金吗？

Enstein, can you please email LP's new fund, please?

好的。

Yes.

我是说——

I mean-

我们投资了基金的150%。

We invest 150% of the fund.

另一轮大融资是Modal估值翻三倍达到150亿美元，Base10在谈260亿美元。

Another big round is Modal triples to $15 billion and Base10 talks at $26 billion.

事实证明，有一段时间你回头看，也许我们现在还处于那个时期，但确实有一段时间，回头看，投资的最佳答案就是把所有钱都投到实验室里。

It has turned out that there was a period where you look back, and maybe we're still in that period, but there was certainly a period where you look back and the right answer with investing was just put it all in the labs.

直接买实验室。

Just buy the labs.

每一轮融资的时候，大家都说“太贵了”，但正确的做法就是继续买实验室。

And every round, people were like, "It's expensive," but the correct answer was just keep buying the labs.

最近18个月左右，事实证明，正确的做法就是继续买推理服务。

It has now turned out in the last 18 months or whatever, the correct answer was just keep buying inference.

有Modal、Baseten、Fireworks、Fal、Together这些公司。

You have Modal, Baseten, Fireworks, Fal, Together.

全都成功了。

It's just all worked.

我的合伙人Eric Vishria有句话，我最近经常引用，因为确实如此，那就是：一切都会成功的。

My partner, Eric Vishria, had this line that I've been kind of quipping a lot lately because it's true, which is that, it's all going to work.

当时大家问“是这个还是那个？”

It was like, "Is it this or that?"

Eric就说：“兄弟，全都算上。”

And Eric's like, "Dude, it's all of it."

这并不意味着每家公司都会成功，也不是每个领域都会成功，但总的来说，有大量的事情都在顺利发展。

And it doesn't mean that every company's going to work, not every sector's going to work, but in general, a crazy number of things are working.

推理服务一直是押注实验室以外所有领域的一个非常好的方式。

Inference has been a really great way to get an index bet on everything outside the labs.

杰森刚才提到的观点，也是这些推理公司表现如此出色的一个重要原因，那就是成本实在无法持续。

Jason just made the point, which is a big part of why these inference companies are doing so well, which is that the costs are just not sustainable.

所以一方面，你看到的是成本太高。

So you have that on one side, which is it's too expensive.

另一方面，你看到的是这样一种动态，我觉得我们在很多领域已经跨过了某个门槛，而且这种情况越来越多，就是智能饱和，也就是说现在智能已经足够好，可以完成那些任务了。

And on the other side, you have this dynamic, which I think is we have now crossed sort of the threshold in a lot of areas and more and more happening, where you get sort of like intelligence saturation, where it is now good enough to do the thing.

举个简单的例子，你的纳税申报要么正确，要么不正确。

To take a simple example, your tax return is filed correctly or it's not filed correctly.

一旦你正确申报了，再往这个问题上投入更多智能就没有意义了。

Once you have filed it correctly, throwing more intelligence at that problem doesn't do you any good.

如果你的工作是用锤子把钉子敲进去，你的锤子已经够好了，就算换成一把镶金的、价值三万美元的锤子，你的工作也只是把钉子敲进去而已。

If your job is to hit a nail in with a hammer, your hammer's good enough, making it a gold-encrusted hammer that costs $30,000, your whole job was just to get the nail in.

所以像这样的任务越来越多，结果就是你会看到更多开源，也会看到这些推理公司的兴起。

And so we have more tasks like that, and as a result, you're going to see more open source, and as a result, you're going to see these inference companies.

而且显然，开源表现非常出色，虽然我不记得是谁刚才提到的。

And there's obviously been, I can't remember who just mentioned it, but open source is doing tremendously well.

顺便说一句，这种成本动态并不意味着开源就能完全胜出。

By the way, this whole cost dynamic does not mean that open source just runs away with it.

那些大实验室显然也会推出自己模型更便宜的版本，而且我觉得人们往往忽略了他们在很多方面的成本优势。

The labs are obviously also going to offer much cheaper versions of their own models, and I think people forget how cost-advantaged they are in a lot of ways.

他们在算力获取方面有很强的结构性优势。

Their access to compute is structurally very strong.

他们在用户获取方面也非常有优势。

Their access to users is very strong.

他们有很多不同的盈利方式，因此在某些情况下可以合理地补贴某些成本。

They have a lot of different ways to make money, and therefore, can subsidize certain costs in certain situations as it makes sense.

所以这并不意味着开源会主导市场，但确实意味着开源将成为市场的重要组成部分，而这也就意味着会有优秀的推理公司出现。

So it doesn't mean that open source will dominate, but it does mean that open source is going to be a big part of the market, and that means you're going to have great inference companies.

所以我看好推理领域。

So I'm long inference.

我觉得开源在市场份额上已经到顶了。

I think open source has reached its maximum as a market share.

我觉得它会继续往下走。

I think it's going to keep going down.

但即使如此，即使它下降50%，而总消费量增长10倍，你仍然会得到——

But even if it does, even if it goes down by 50% and total consumption goes up by 10x, you still get a-

对，还是有好投资机会的。

Yeah, there's still good investments.

我不是说推理需求不是近乎无限的，但我觉得开放权重已经达到顶峰了。

I'm not saying that there's not an almost infinite amount of inference, but I think we've reached peak open weights.

那么，杰森，为什么？

Well, Jason, why?

抱歉，帮我理解一下。

Sorry, help me understand.

有两个原因。

Two reasons.

杰克说中了一点，还有第二点。

Jack hit one of them, and there's a second one.

这两点都很清楚。

They're crystal clear.

第一，归根结底，Anthropic和OpenAI完全可以自己决定非旗舰模型的定价，而且他们有很多直接竞争的方式。

One is, at the end of the day, Anthropic and OpenAI are just deciding what they want to price their non-max frontier models, and they have many ways to compete directly.

他们完全可以做到想多便宜就多便宜。

There's no reason they can't be as cost-competitive as they want to be.

现在，Anthropic的数据看起来不错，如果排除那7000项，毛利率有80%。

Now, Anthropic's got its numbers if you exclude 7,000 things that has 80% gross margins.

好吧，我们可以调侃一下这个。

Okay, we can make fun of that stuff.

但他们确实有能力想多竞争就多竞争，对吧？

But they have the ability to be as competitive as they want to be, right?

杰克说得对。

And Jack's right.

确实，到了某个阶段，最重要的就是完成任务，但即便如此，我觉得事情也没那么简单。

For sure, there's a certain point where all that matters is resolving a task, but even that I don't think is quite that simple.

但他们可以给刚发布的Sonnet Plus 5.5定价。

But they can price Sonnet Plus 5.5, which just came out.

我刚测试过。

I just tested it.

它只便宜了大概20%，但这是在一周内便宜了20%。

It's only like 20% cheaper, but that's 20% cheaper in one week.

他们可以做到30%，如果愿意的话，40%也行。

They can do 30, they can do 40 if they want.

这是第一个原因。

That's reason number one.

第二个原因是，天哪，大家真的不希望Anthropic和OpenAI用他们自己的第一方数据来训练。

Reason number two is, boy, people really don't want Anthropic and OpenAI training on their own first-party data.

我刚从Dreamforce回来，跟你说，我知道那里有很多人穿着西装外套和不合身的西装什么的。

I just got back from Dreamforce, and I got to tell you, I know there were a lot of blazers there and ill-fitting suits and stuff.

那里没人想用开源模型。

No one wants to run on open source models there.

至少中国那边的，没人。

At least China-based, nobody.

我聊过的人里，没有一个觉得舒服的，明白吗？

Not a single person is comfortable with it that I talked to, okay?

创始人另说，但客户没有。

Founders aside, but no customers.

所以我觉得就这两个趋势来看，我们已经到顶了，对吧？

So I think just those two trends mean we've hit the peak, right?

就像杰克说的，这么说确实有点夸张。

To Jack's point, it's ridiculous to say it.

OpenAI和Anthropic的人都不傻。

No one at OpenAI and Anthropic's stupid.

他们可以调整各种参数，想怎么做就怎么做，保持成本竞争力，而且，开源是有它的道理，但高层的那种焦虑只会越来越严重。

They can twist the knobs and dials and do what they want and be as cost-competitive and just, there are arguments for open, but this anxiety at the C-level is only going to increase.

这种担忧只会越来越严重，而且随着安全问题变得越来越突出，这种担忧还会加剧，就像Astra 5.1因为安全问题被撤回一样。

It's only going to increase, and it's only going to increase as security becomes a bigger issue, and it's only going to increase as Astra 5.1 was pulled back for security concerns.

而且，我知道这听起来有点老套，但这就是企业界的现实。

And listen, I know it's fuddy-duddies, but it is the real world in enterprise.

我问过他们。

I asked them.

我接触过的人里，没有谁愿意在生产环境中使用开放权重模型。

No one wants to use an open weight model on the floor that I talked to.

一个都没有。

Nobody.

所以我觉得市场份额已经见顶了。

So I just think the market share has peaked.

我不是说这不重要。

I'm not saying it's not material.

你问的是开放权重，还是特指中国的开放权重？

Did you ask open weight or did you say Chinese open weight?

因为我觉得这里面有好几件事在同时发生——

Because I think there's a bunch of things going on-

我明白你的意思。

I get your point.

我简化一下，我们现在用的这些模型，绝大多数都是中国出产的，对吧？

I'm simplifying that the vast majority of these models that we're consuming today are China origin models, right?

而且那些人里面，有些也在往不开放权重那个方向走。

And where some of those guys are also moving towards not being open weight themselves.

不过没错，我觉得这就是为什么那个poolside和Nvidia的事有意思。

But yes, I do think that's why the whole poolside Nvidia thing is interesting.

要是能有个低成本的美国本土替代品，那会很有意思。

It would be interesting to have a low-cost US-based alternative.

再说回Jeff说的，它虽然不开放权重，但它是美国的，而且它不是大语言模型，但对某类使用场景来说，它就是个直接竞争对手，对吧？

And going back to Jeff, it's not open weight, but it is US and it's not an LLM, but for a certain class of use case, it is a direct competitor, right?

我不是说它们就完了，这是回应Jack的观点。

I'm not saying that they're destroyed, to Jack's point.

我就是觉得他们已经到顶了。

I just think they've peaked.

他们的市场份额已经到顶了。

They've peaked in market share.

我觉得会降下来，而且我也不觉得这会伤到Modal或者Baseten或者其他公司，但是我觉得今年那些看起来疯狂的市占率图表，虽然有很多来源，但我觉得OpenAI和Anthropic在这里有很多结构性优势，未来几个月他们会利用这些优势，对吧？

I think it will come down, and I don't even think it's going to maim Modal or Baseten or others, but I think these market share charts that have seemed crazy this year, there's been a lot of sources, but I think there's a lot of structural benefits that OpenAI and Anthropic have here, and they're going to take advantage of it in the coming months, right?

而且OpenAI刚说他们在企业领域达到了700亿。

And OpenAI just hit 70 billion, they said in enterprise.

那边势头很猛。

It's a lot of traction there.

顺着这个说，虽然我也不知道答案，但如果你在前沿有优势，你就不会浪费时间在二流的东西上。

Pushing on that, and again, I don't know the answer, is that provided you have opportunity at the frontier, you won't waste time with the second-string stuff.

某种程度上，就像你说的，如果你觉得在前沿上已经接近极限了，你可能更愿意花时间去做报税这种事儿。

To some extent, you're probably more likely to spend time on, as you say, doing the tax return if you feel you're asymptotic on some of the frontier stuff.

因为你边际上分配的是算力，如果你是OpenAI或者Anthropic，我很想知道这个数学怎么算：你是拿下一块GPU去做一个能治愈癌症的生物学模型，还是拿去做杰克的报税，那是个有限的任务，其实可能也不是那么有限，因为你已经测过基准了。

Because the thing that you're allocating at the margin is your compute, and if you're OpenAI or Anthropic, I would love to know what the math is like on do you take this next chunk of GPU and build a model for biology that can cure cancer, or do you take it and do Jack's tax return, which is a pretty finite, probably not a pretty finite task, actually, now that you benchmarked it.

这其实是个无限的任务，但在AI的范畴内，它又挺有限的。

It's a pretty infinite task, but within the context of AI, it's pretty finite.

这里面的权衡取舍真的很有意思。

And super interesting set of trade-offs there.

我觉得还有几点反应，一是关于非美国模型，确实有很多焦虑。

I think-I think a couple other reactions are, one is that to the point of non-American models, I do think that there's a lot of anxiety there.

我也觉得我们开始看到，而且会继续看到很多企业自己微调模型，基于开源权重，用推理公司来打造自己的模型，然后自己运行。

I also think we are starting to see, and we'll continue to see a lot of enterprises post-train their own models, and draft off of open weights and use inference companies to make their own models and then run them themselves.

所以我不知道这会怎么影响格局，但我确实认为在美国我们会看到更多这样的情况，而且这会产生一定的影响。

And so I don't know how that will factor in, but I do think we will start to see more of that in the US, and I think that will have some amount of impact.

另外，你可以试着退一步，模糊视线，看看这里最重要的一件事是什么？

The other thing is you can kind of-- I often try to just step back, blur my eyes, and what's the one thing here?

我觉得总的来说，最终还是要看谁拥有算力。

And I think one of the things you could look at here in general, it's all just going to come down to who's got the compute.

如果你看看所有的推理云，我觉得大概有千兆瓦级别的规模。

If you look across all of the inference clouds, I think it's on the order of a gigawatt or something like that.

我觉得OpenAI和Anthropic，各自大概是个位数百分比。

And I think OpenAI and Anthropic, high single digits each.

所以你可以说，也许一个比另一个更高效一点，也许一个的token效率更好，一个定价略有不同，但在某种程度上，这最终还是会归结为所有计算机一直在运行，以及谁拥有它们，我觉得这将成为方程式中极其重要的一部分。

And so you could say maybe one's a little bit more efficient than the other, maybe one's got better token efficiency, one prices a little bit differently, but on some level, this will also just come down to all of the computers firing all the time and who owns it, I think is going to turn out to just be a dominantly important part of the equation.

是的。

Yes.

算力份额可能大致对应你的token份额，再稍微调整一下，可能就对应你的收入份额。

Compute share probably proxies to your token share, probably proxies with a little adjustment to your revenue share.

至少在两倍以内之类的。

At least within 2x or something like that.

对，差不多。

Within, yeah.

明白了。

Got it.

确实有道理。

That does make sense.

说到调整杠杆，OpenAI重新开放了它暂停的200美元套餐，这个套餐之前因为最新的Sasha模型算力不足而暂停，但现在200美元能买到的东西减半了。

Speaking of tweaking the levers, OpenAI reopens its $200 plan that it paused because it ran out of compute for the latest Sasha model, but halves what $200 buys.

所以Harry，你说得对。

So Harry, you're right.

所以这其实是在讨论同一个话题，我觉得这在某种程度上是这一切的根本问题。

So this is getting at the same topic, which I think is kind of on some level the fundamental equation of all of this.

我觉得这非常重要。

I think it's extremely important.

其他我们不知道的部分，也是不断变化的变量，就是每个任务需要多少token？

The other parts of it that we just don't know that are also updating variables continually are, how many tokens do you need per task?

所以随着模型变得更好，它们在消耗token方面变得更聪明了。

And so it's like as the models get better, they get much smarter at consuming a token.

然后另一个问题是，每个token的效用是多少？

And then the other is, what is the utility per token?

所以确实，消耗的token更多了，但每个token的成本更低了。

So it is true that there are more tokens being consumed, they cost less dollars.

而我们不知道的另一面是，比如，你需要多少token才能报完税。

And what we just don't know on the other side of it is, how many do you need to file the tax return, for example.

而这个问题对所有这一切来说都是个开放性问题。

And that's just an open question to all of it.

所以，你能从一吉瓦电力中获取更多智能吗？

So can you get more intelligence out of a gigawatt?

当你试图去理解这个问题时，你会发现它有多难。

And when you try and get to grips with that, you realize how hard it is.

因为你知道我是个技术宅。

Because you know I'm a geek.

我回去试着搞清楚，token成本在下降，token数量在上升，还有token效率。

I went away and tried to figure out, you've got token cost going down, number of tokens going up, token efficiency.

而事实是，你在乘三个数，每个数的误差范围都大得离谱。

And the truth is you're multiplying three numbers, each of which has an error bar that's pretty damn large.

如果你有任何一点学术诚实，最后你只能承认：我真的不知道。

And if you-- any intellectual honesty at the end you kind of go, "I just don't really know."

唯一能观察到并且真正有实际意义的事情，就是那些在分配资金的人的购买决策。

The only thing you can observe that's actually actionable is the buying decisions of people who are allocating their money.

因为我算不清楚三个大数字相乘到底意味着什么。

Because I can't figure out multiplying three big numbers together what it means.

但去年有个买家在Tropic上花了五亿美元。

But some buyer spent half a billion dollars last year on Tropic.

你得假设他们算过这笔账，并且从中获得了价值。

You got to assume they ran the numbers and are getting value from it.

所以最后，我还是回到我之前说的，所有这一切的代理指标，杰克，你明白，关键就在于，人们分配预算时说，在边际上自动化这个任务是合理的，值得花五亿到五十万美元去做，不管具体是多少。

Which is why in the end, I go back to what I said, the proxy for all this, and Jack, you out with the rubber hits the road is, people allocating budget saying at the margin automating this task makes sense and it's worth spending half a billion dollars to half a million dollars, whatever it is to do.

因为除此之外，真的很难说服自己你把数学算得足够清楚，能理解它，或者更重要的是，能预测12个月后的情况。

Because other than that, it's just super hard to really convince yourself you've done the math well enough to understand it, or more importantly, to predict it 12 months out.

这就是为什么如果我能知道一件事，我会一直回到我反复说的那个观点，我想知道那两家公司的第三季度数据。

And that's why if I could know one thing, I go back to my comment over and over again, I would know the Q3 numbers for those two companies.

长期来看，这并不重要，但短期来看，我认为这非常重要，从动量角度来看，以及这些预算如何继续扩大。

And long term, it doesn't matter, but short term it matters a shit ton is my opinion, just from a momentum perspective, and how those budgets are continuing to expand.

你知道吗，纯粹为了好玩，有个我正在做的应用叫SaaStr Connect，我刚刚自己跑了个评估——不知道能不能叫评估，反正我是在刚出的Sonnet 5.5上跑的。

You know what it's worth just for fun, there's an app I'm trying to finish called SaaStr Connect, and I just ran an eval my own-- I don't know if you can call them evals, but I ran it on Sonnet 5.5, which just came out.

输入token，比之前高了42%。

Input tokens, 42% higher than before.

输出token，高了44%。

Output tokens, 44% higher.

所以谁知道Jack的报税是不是真的能便宜多少。

So who knows if Jack's tax return's getting all that cheaper.

不过，这也有好处，对吧？

Now, there's benefits from it, right?

它通过了更多的盲测，对吧？

It passed more of the blind tests, right?

质量提升了，这也是意料之中的。

Quality went up, which is what you'd expect.

成本大概降了10%，不是30%，因为用了更多GPU，但token多了40%。

Cost went down about 10%, not 30% because more GUs, but plus 40% more tokens.

我就是觉得这些东西太难预测了，兄弟。

I just think the stuff's hard to predict, man.

从5.5开始涨42%已经很多了，对吧？

42%'s a lot to go from 5.5, right?

而且那只是Jason做他自己的任务这么细的层面，要预测一般人的任务难上10倍、100倍，因为我不认识那些人，也看不到他们的token消耗，所以你就得看看那些Joseph们在做什么。

And that's at the level of granularity of Jason doing Jason's task, and it's 10X, 100X harder to say tasks in general from people I don't know or I don't see their token spend, which is why you just got to look at what the Josephs are doing.

今天我们有个新闻，跟那些几万亿美元的IPO相反，Oura撤回了他们的IPO，原本计划是160亿美元的。

Today we have news on the flip side of the incredible multi-trillion dollar IPOs, we have Oura pulling their IPO, which was planned for a $16 billion.

Jack，我每次一说到新话题，我就得看Rory的脸色。

Jack, I have to-- whenever I basically say a new topic, I read Rory's face.

他总是一脸痛苦，但我得判断他痛苦到什么程度。

And it's always miserable, but it's just the extent of misery that I have to judge.

我承认。

I'll own it.

你说得对。

You're right.

我们其实已经展示过远征项目了——

We actually have shown expeditions-

哦，不，你是不是持有一些Oura的股票？

Oh, no, you've got some of Oura, don't you?

抱歉。

Sorry.

哦，真的很抱歉。

Oh, I'm so sorry.

我们会专门为你开一个众筹页面，Rory——

We'll set up a GoFundMe page just for you, Rory-

……就为了让你能筹点钱弥补一下损失。

... just so you can get some shekels in your pot to make up for it.

你们真是太好了，Harry。

You're all sweet, Harry.

我知道你非常关心我。

I know you care so deeply.

但你看，我评论一下，是的，我确实有股权，但并没有积极参与，所以没有任何内幕消息。

But look, just comment, yes, I'm in the position of having an ownership interest, but not in any way being actively involved, so not having any insider information.

但我得承认，这件事让我很意外。

But I will admit this surprised me.

真的非常意外。

It super surprised me.

Oura原本计划进行IPO。

Oura had planned to do an IPO.

他们本来已经进展到很后面了，原定这周三定价。

They were a long way down, meant to price this week, Wednesday.

他们一开始信心满满。

They went into it feeling very strong.

事实上，顺便说一下，他们的承销商是摩根士丹利、高盛和摩根大通。

In fact, for the record, they have Morgan Stanley, Goldman Sachs, and JP Morgan.

都是大牌。

All the people.

那时候，你想要谁都能拿到，对吧？

There was no people left that you couldn't have, right?

所以并不是说他们请的都是些外行。

So it's not that they had the dummies here, people.

第二点，这家公司是盈利的，规模很大，消费者也都知道它。

And then the second thing is, it's a profitable, it's a big company, consumers know it.

这种事情按理说应该很容易做成。

It's the kind of thing that should be very doable.

第三件有意思的事是，他们最初提交文件的时候，最大的投资者之一Forerunner——我觉得他们非常聪明——说他们要卖掉全部股份，而我已经做这行很久了。

And the third interesting thing is, when they first filed, one of the largest investors, Forerunner, who I think are super smart, said they're going to sell all their position, which I've been doing a long time.

我从来没见过有人在IPO时能把自己所有的股份都卖掉。

I'd never seen someone in an IPO being able to sell all their position.

所以我当时想，这真是大胆的决定，得给他们点赞，因为消费电子行业确实很难做。

So I was like, hmm, ballsy call and give you credit because your consumer electronics is hard.

但关键是，他们一开始就明确说了“我们打算这么做”。

But the fact that a priori they'd said, "We're going to do this."