20VC：万亿目标或落空：新的风投游戏 | Coreweave抛售引发恐慌 | Town估值达10亿美元，Anduril估值900亿美元：收入乘数失控 | 通用大西洋瞄准IPO，Modal以157.5亿美元估值融资7.5亿美元

顺便说一句，从策略上讲，我觉得Rory你也会同意。

For what it's worth, just tactically, I think Rory, you'd agree.

Chamath的推文是，‘就按1.5万亿的估值去融你需要的钱，然后继续前进’，对吧？

Chamath's tweet was, "Just raise whatever you need at 1.5 trillion and move on," right?

如果你能按2万亿融到，但那会给你的资金需求带来压力，那只是估值问题。

If you can raise it two trillion but it's going to stress your capital needs, it's just evaluation.

所以从绝对数字来看，一切都很好，对吧？

So in absolute terms, everything's great, right?

这在各个层面都是前所未有的。

This is unprecedented on all levels.

团队很棒，应用也很棒。

Great team, great application.

我怀疑的是，提醒一下，IPO之后会非常动荡，因为回到Dev在节目上提到的观点，开发者并不忠诚，而且自从这个节目开播以来，市场份额的变化，Anthropic、OpenAI、OpenWeights、Jev之间的变化，对吧，Jev在45天前甚至还不存在。

What I suspect, well, a reminder is, it's going to be very choppy after the IPO because going to Dev's point when he was on the show, developers are not loyal, and the amount of market share since this show started, the amount of change in Anthropic versus OpenAI, versus OpenWeights, versus Jev, right, which didn't even exist, what, 45 days ago.

这是一个赢家，但在增量增长和市场份额方面，变数很大，这份收入的可预测性有限。

This is a winner, but the amount of variability in the incremental growth and in the market share, there's only so much predictability in this revenue.

这里的ARR其实也就那么多，我觉得如果市场反应太热烈的话，可能会带来一些压力。

There's only so much ARR in the ARR here, and I think depending on how effusive the markets are, it may create stress.

如果明年年底Anthropic的市场份额比现在下降10%，我也不会感到意外，而且从绝对数字来看，它依然可能表现很好。

It wouldn't surprise me if at the end of next year, Anthropic's market share is down 10% from where it is today, and it still could be great on an absolute basis.

但市场份额下降几个百分点完全不会让人震惊，不只是token方面，收入方面也一样。

But down percent wouldn't be a shocker at all, not just in tokens, but in revenue.

这种情况完全有可能发生。

Easily could happen.

你看，不管OpenAI是70%还是50%，那都是很大的份额。

Look at, despite whether OpenAI's at 70 or 50, it's a lot.

你说下降，是指市场份额下降吗？

When you say the sh-- you mean the share is down?

对，市场份额可能会下降，对吧？

Yeah, market share could be down, right?

所以当你构建模型的时候，不只是DCF模型，还有收入预测模型，你可能会和乐观情况、最佳情况有较大偏差，而这应该是意料之中的。

And so when you're building your models, right, not just DCF models, but your top-line models, you could be materially off from your good case, right, your best case, and it should be expected.

说白了，这比一般的经常性收入业务要吓人一点，对吧？

To state the banal, is a little scarier than the average recurring revenue business, right?

这就是大家纠结的地方。

And that's what people are wrestling with.

我怎么考虑呢——如果主要玩家之间的市场份额持续变动，那下一步通常就是，我想保住我的市场份额，所以我会考虑价格。

How do I think about-- If you have continued market share shifts between the key players, then the next shoe to drop is typically, well, I want to keep my market share, so I'll look at price.

而且我觉得，大家都在从博弈论的角度琢磨，这未来的损益表会是什么样？

And I think, everyone's trying to work out from the game theory perspect- what does that look like as a P&L going forward?

这也是为什么，回到我一直说的那个老生常谈，只要顶层的终端用户需求还在，大家都能赢。

Which is why, going back to the boring thing I keep saying is that provided the end user demand at the top is there, everybody can win.

我们都喜欢那种大家都能赢的生意，对吧？

And we all love businesses where everybody can win, right?

如果需求开始放缓，然后价格竞争就变成了——真的，你为了从对手那里抢份额、保持增长率而不得不做的事，那情况可能会变得相当艰难。

If that starts to slow down and then price competition becomes not a-- literally the thing you got to do to take it from the other guy to keep the growth rate going, then it could get pretty tougher.

你觉得山姆以1.4万亿的估值融资，是不是特别有策略性，就是为了某种锚定效应和向市场传递信号？

Do you think Sam was super strategic raising his supposed round at 1.4, just in terms of some form of anchoring and message to market?

我觉得当你需要钱的时候，你就得拿钱。

I think when you need money, you take money.

我不知道是不是有什么高深的策略，比如1.4和1.5在传递信息上有什么不同。

And I don't know if there's a big sophisticated, oh, 1.4 as a messaging versus 1.5.

我觉得事情很简单，我们在计算上花了很多钱。

I think it's as simple as, we spent a lot of money in compute.

我们的IPO没做成。

We didn't get our IPO done.

一直挺好的。

Been good.

我们把孙正义的钱都花光了。

We spent all of Masa Son's money.

最好再以某个价格拿个40亿。

Better get another 40 at a price.

有什么限制条件吗？

What are the constraints?

估值必须高于上一轮，否则消息就不好看了，而且还得低于实际价值，好给上涨留出空间。

It's got to be up on the last round, otherwise, the messaging is shit, and it's got to be down below reality to leave room for the upside.

数字的位数是有限的——虽然数字是无限的，但在1和2之间，只有有限个一位小数的数字。

There's only a finite number of digits between-- Well, there's infinite number of digits, but there's only a finite number of one decimal plus digits between one and two.

这个故事背后有几个因素。

There are a couple of elements beneath this story.

其中一个，或者说在Anthropic之前一周，是Replit把OpenAI设为默认模型。

One of which, or beneath Anthropic is a week, was Replit made OpenAI its default.

Jason，我对此很感兴趣。

Jason, I am intrigued here.

你对此有什么看法吗，尤其是考虑到你是Replit的重度用户。

Anything from you on that, especially given your Replit power user status.

我觉得那是个无聊的媒体炒作。

I think it was a stupid media story.

我们确实每天使用Replit大约14到15个小时。

We do use Replit about 14 to 15 hours a day.

基本上，Replit 里面有自己的路由器，对吧？

Basically, Replit has its own router in it, right?

它有自己的模型选择器。

Its own model selector.

根据你所在的类别，它对你可能是不可见的，也可能是可见的。

It can be invisible or visible to you, depending on what category you're in.

而在 Replit 里，OpenAI 的模型在输出质量差不多的情况下更便宜，所以他们就直接把它设为默认了。

And OpenAI models are just cheaper for the same rough quality output in Replit, so they just made it the default.

TechCrunch 写了一篇文章，说他们不再和 OpenAI 合作了，这其实是不对的。

And TechCrunch miswrote an article saying they were no longer working with OpenAI, which wasn't true.

他们只是把默认模型从 Opus 和 Fable 换成了 Astra。

They just flipped the default from Opus and Fable to Astra.

那只是默认设置，我们都在路由，对吧？

That's just the default, and we're all routing, right?

我们都在以某种方式开放路由。

We're all open routing one way or the other.

只是如果你是Replit、Lovable、Base44或V0，你有这些额外的动力，因为你的普通用户只是想要高质量的输出……而且他们非常倾向于为你把质量和价格的关系调整到最佳，对吧？

It's just if you're Replit or Lovable or Base44 or V0, you have these extra incentives because your average user just wants to have a high-quality output... put, and they're very incented to get this quality price thing just right for you, right?

不像那些普通人自己配置OpenRouter，还希望自己能弄对。

Unlike the average idiot pre-configuring OpenRouter themselves, hoping they get it right.

这些人实际上——看Replit、Lovable、Base44、V0做什么更有意思，因为他们有动力去把路由做好，对吧？

These guys are act-- It's more interesting to watch what Replit, Lovable, Base44, V0 do because they have an incentive to nail the routing, right?

他们有非常强的动力。

They have a very strong incentive.

他们只能部分地把这个优势传递给最终用户，对吧？

They can only partially pass it on to the end user, right?

所以我们可以眯着眼，从这些人的做法中窥见未来。

So we can actually squint and see the future based on what these guys are doing.

他们说的是，听好了，Amjad的意思是，质量已经足够接近了，而价格对我们来说今天便宜了30%到40%。

And what they're saying is, listen, what Amjad's saying is the quality is sufficiently peer level here, and the price is 30% to 40% better for us today.

我们为什么还要用Fable和Opus呢？

Why would we use Fable and Opus?

你知道吗？

And you know what?

下周，他可能会在另一个活动上，上台直接跟Anthropic谈市场份额的事。

Next week, he may be at a different event and get on stage and just go into this market share thing with Anthropic.

情况可能会变。

It may change.

情况可能会变。

It may change.

他们可能还没完全想清楚怎么用Sonnet 5.5和Haiku 5.5，而且他们还有Kimmy和其他模型。

They may not have fully thought through how to use Sonnet 5.5 and Haiku 5.5 and they have Kimmy and others.

但如果Kimmy真的那么厉害，你知道吗？

But if Kimmy actually was so great, you know what?

那些公司，比如Replit，就会完全用它了，这可能是讨论中更有意思的一个发现。

All those guys, Replit would be using it entirely, which is maybe the more interesting learning that was under discuss.

因为路由也会经过他们，但实际上并没有。

Because it will route through them, too, and it doesn't.

阿姆贾德没有上台说“我们会把所有用户都引导到Kimmy”，对吧？

Amjad didn't get up on stage and say, "We'll route everyone through Kimmy," did he?

他说的是“我们要从Claude转到OpenAI，因为效果一样好，而且便宜很多。”

He said, "We're going from Claude to OpenAI because it's just as good and materially cheaper."

这说明了什么？

What's the takeaway from that?

说明Kimmy并没有大家想的那么好吗？

That Kimmy is not as good as people think?

还是说——

That-

实际上，他没有看到Kimmy给他那十亿美元收入、给他的客户带来什么好处。

In practice, not seeing the benefits for his billion dollar of revenue, for his customers.

他没看到好处，不然他早就自动切换过去了。

He's not seeing the benefits, or he would just auto route it.

Replit和Lovable的客户对成本非常敏感。

The customers are very cost-sensitive at Replit and Lovable.

不只是中小企业对成本非常敏感，因为他们想用每月25美元的价格获得所有服务，甚至还有免费的LLM，但企业客户也非常在意成本，因为他们不想为产品团队去开发原型而产生大量不固定的账单。

Not just the SMBs are very cost-sensitive because they want to get everything for like $25 a month or they even have free LLMs, but the enterprise customers are very cost-sensitive because they don't want these massive variable bills for the product teams going off and building prototypes.

这是一个超级成本敏感的模式，对吧？

It's a super cost-sensitive model, right?

托管、数据库、LLM。

Hosting, database, LLMs.

如果你放慢脚步看看他们在做什么，他们每个都有大约十亿美元的收入，对吧？

If you slow down and see what they're doing, and they're each at, like, a billion in revenue, right?

这就像一个小实验室，我们应该多加关注。

It's like a little lab that we should be paying more attention to.

这比那些乱花钱、分配不当、还指望能成功的傻瓜们更有意思。

It's a more interesting lab than the idiots spending all the money and misallocating it and hoping it works.

那是一个很好的商业模式，但我们从这些家伙在优化高质量输出和低成本方面的大规模投资中学到更多。

That's a great business model, but we learn much more from the massive investment these guys are making in optimizing high quality output for low cost.

也许其他人也在做同样的事，每个都有十亿美元的收入，但我觉得没有多少人在为终端用户做这件事。

Maybe others are doing this, each at a billion in revenue, but I don't think that many people are doing it for end users.

我觉得这个观点很好，因为，是的，这是一个非常客观的计算测试，关于价值——你怎么看待你为了支付顶级价格却得到低于顶级智能而留下的价值，以及你真正需要什么？

I think it's a great point because, yes, it is a very objective calculating test of what value-- How do you think about the value you're leaving on the table by paying top-tier prices for less than top-tier intelligence, and what do you actually need?

因为有了十亿美元，你就有足够的时间去认真思考这个问题。

Because with a billion dollars, you've got enough time to really think about it.

杰森，你是这些工具的最佳用户和分析者。

Jason, you are the best user and analyzer of these tools.

在Anthropic的《本周发生了什么》中，Claude Code Review已经发布了。

Under the Anthropic What Happened This Week, Claude Code Review is out.

我们看到Boris Cherny在推特上提到了这件事。

We saw Boris Cherny tweet about it.

我很想听听你的想法。

I'd love your thoughts.

哪些初创公司会被这个摧毁？

Which startups are destroyed by this?

你觉得怎么样？

What did you think of it?

我不想乱说，因为变化太大了，对吧？

I don't want to talk out of my rear because it's so much change, right?

Greptile现在特别火。

I mean, Greptile is super hot.

他们这周刚融了资，对吧，750亿。

They just raised whatever this week, right, 75 billion.

我觉得Rory在CodeRabbit，也可能我记错了。

I think Rory's in CodeRabbit, or I could be wrong.

他们是重要的玩家，对吧？

They're an epic player, right?

有这么多优秀的人。

There's all these great folks.

他们都有各自的优势。

They all have benefits.

我觉得在很多情况下，他们的优势在于能跑在LLM上，这非常有价值，对吧？

They all have the benefit of being, I think, in many cases, of being able to run on LLMs, which is very valuable, right?

让一个LLM检查另一个LLM的工作，让OpenAI检查Claude，或者反过来。

Having one LLM check another LLM's work, having OpenAI check Claude or vice versa.

所以这种应用很多。

So there's a lot of it.

就像软件行业自古以来一样，专注单一领域的公司有很多优势，对吧？

Just like there have been since the dawn of time in software, there's a lot of benefits to having pure plays, right?

但这也总是让人担心，因为这是Anthropic的核心。

But it's also always scary because this is the core of Anthropic.

Claude Code是我们讨论过的Anthropic引擎的核心。

Claude Code is the core of Anthropic's engine we've talked about.

而做代码审查，做所有这些流程，实际上可以说比写代码更重要。

And making code review, making all of this process, it's actually arguably much more important than building code today.

你知道我们现在花多少时间在审查代码上吗？因为LLM太擅长帮你完成95%的工作了。

You know how much time we spend now reviewing the code because the LLMs are so good at getting you 95% of the way there.

你跟任何工程师聊，他们都会说：‘哦，我在外面吃晚饭，然后我给Harry做了那个功能。’

You talk to any engineer, they're like, "Oh, I'm out at dinner, and I built the feature for Harry.

我在烧烤，我的手机就在给哈里写功能。

I'm having a barbecue, and my phone is building the feature for Harry."

你确实写了哈里要的功能，但接下来整整三天你都得花在测试和修bug上。

And you did build the feature that Harry asked, but you're going to have to spend the whole next fucking three days testing it and fixing it.

所以这事儿挺大的。

So it's a big deal.

这事儿到底有多大？

How big a deal is it?

我猜肯定比那个没什么用的Claude Design（或者随便叫啥）要大得多，对吧，那个也没干掉Figma。

I have to imagine it's a bigger deal than the nothing burger, which was Claude Design or whatever the heck it's called, right, which didn't kill Figma.

但到底有多大呢？

But how much?

这真的很难预测。

It's so hard to predict.

如果你是竞争对手的员工，那确实挺吓人的。

It's scary if you're working at a competitor.

总是让人害怕，对吧？

It's always scary, right?

必须得这样。

It just has to be.

不管你在推特上怎么说，总是让人害怕。

No matter what you say on Twitter, it's always scary.

就算它让你损失30%或40%，你也只能做得更好。

Even if it memes you 30% or 40%, you just got to be better.

我们上周说过一件事，我现在仍然坚持，独立AI应用的最佳市场，我把编程也算作一个应用，那些增长率最高的也最接近太阳。

One of the things we said last week, and I stand by it, the best markets for standalone AI apps, and I put coding as effectively an app, the ones with the highest growth rates have also been closest to the sun.

它们最接近基础模型。

They've been closest to the foundation model.

事实上，我今天刚发了一条推文，说这是伊卡洛斯时刻。

In fact, I just did a tweet on that today about it's the Icarus moment.

伊卡洛斯别无选择。

Icarus had no choice.

他想取暖，就得飞得离太阳近。

He wanted to get warm, he got to fly close to the sun.

我觉得这些大的应用场景很有意思。

I think these big use cases are interesting.

比如代码生成、代码审查、法律这些领域。

It's like code generation, code review, legal.

它们都跟基础模型贴得很近，每六个月你就会经历一次恐慌时刻，心想：“天哪，他们真要这么干了”，对吧？

They're all very close to the foundation models, and every six months, you have the terror moment of, "Oh my God, they're going to do it," right?

这就是身处那个位置的坏处。

And that's the downside of being in that real estate.

但好处是，它们是标志性的，是这项技术最出色的应用场景。

But the upside in being in that real estate is they are marquee, they are superb use cases for the technology.

正因为如此，CodeRabbit 这些公司的发展速度是我们前所未见的。

And as a result of that, CodeRabbit, all these companies are growing faster than we've ever seen.

如果你想避开这种风险，哈里，那你就回到你之前吹嘘的那些平淡无奇、年增长三倍的软件公司去吧，对吧，那些没人愿意投资的公司，对吧？

If you want to avoid that risk, Harry, then you go back to your dull little three X year-on-year software companies that you were bullshitting on about earlier, right, that no one wants to invest in, right?

你不能两头都占。

You can't have it both ways.

你不能坐在那儿说：‘哦，我想要无限快的增长速度，完全跟基础模型没有区别，哪怕我们之所以能有这该死的增长速度，就是因为基础模型解锁了新能力。’

You can't sit there and go, "Oh, I want infinitely fast growth rate, utterly undifferentiated from foundation models, even though the only reason we're getting the fucking growth rate is because the foundation models have unlocked new capabilities."

这根本说不通。

It just doesn't make sense.

这不可能。

That's not a thing.

我这辈子想成为很多角色，但这事儿成不了。

I mean, I'd like to be a lot of things in my life, but it ain't going to happen.

我不知道从风投回报或收入角度看是不是这样，但我确实认为代码审查是当今软件领域最重要的一个类别。

I don't know if this is true from a venture outcome or revenue, but I actually think code review is the single most important category today in software.

绝对是最重要的类别。

By far, the most important category.

但这并不意味着它能带来最大的财富。

It does not mean it will make the largest fortunes.

这并不意味着会出现一个价值六千亿美元的代码审查工具。

It does not mean there'll be a 600 billion code review cursor.

但我的意思是，你不能靠感觉写出自己的SAP或Salesforce，但我们现在构建代码的速度是前所未有的，而且代码质量已经相当不错了，但它从来都不是为了生产环境而设计的。

But what I mean is you cannot vibe code your own SAP in Salesforce, but the rate at which we're building code is unprecedented, and the code now is pretty darn good, but it's never production intent.

这是大家都忽略的一点。

This is what everyone misses.

它从来都不是生产就绪或面向生产的。

It's never production-ready or production intent.

所以如果我们退一步，更广义地定义代码审查，它是构建过程中最重要的环节。

And so if we step back a minute and define code review more broadly, it is the most important category in building.

这太重要了……每个人都在Claude Code或Cursor上每天工作10、12、14个小时，随便什么。

It is so important ... everyone is building 10, 12, 14 hours a day in Claude Code or Cursor, whatever.

但审查、处理这些代码并使其达到生产标准所需的时间，仍然比编写代码多出一个数量级，而我们还没有解决这个问题。

But the amount of time it takes to review and process it and make it production intent is still an order of magnitude more than building, and we have not figured out.

我们假装这项工作已经完成了。

We're pretending that this work is complete.

其实不是这样。

It's not.

如果我能找到最好的下一代真正的代码审查工具，或者最好的团队，我会下十个赌注，因为我们太需要它了。

If I could find the best next-generation true code review, probably the best team, I'd make 10 bets because we need it.

而且它们可能都会被Anthropic干掉。

And they may all get killed by Anthropic.

我们拭目以待，但这比代码生成重要得多。

We'll see, but it's much more important than code generation.

这又回到了Amjad的观点。

This is back to Amjad's point.

我可以在Astra和Fable之间切换，或者其他什么工具。

I can rotate between Astra and Fable and whatever.

它们都很好，对吧？

They're all great, right?

也许Kimi如果能赶上也不错，但我怎么把它应用到生产环境中呢？

And maybe Kimi if it catches up, but how am I going to get it into production?

同意。

Agreed.

从大局来看，很明显我们正在开发的软件比以往任何时候都多，可能多出好几个数量级。

Big picture, it's pretty clear that we're making way more software than we've ever made, by probably orders of magnitude.

看看Anthropic关于就业的那份报告，还有他们最近关于失业的论文。

Just looking at the Anthropic thing they did on employment and recent paper they did on unemployment.

所以，基于这一点，整个软件开发生命周期都会发生变化。

So given that, the whole software development life cycle is going to change.

因为在智能体时代，如何成功部署软件可能会变得非常不同。

Because how that gets deployed successfully is probably going to be very different in an agentic world.

所以我同意。

So I agree.

我觉得有一大堆相关的诱人机会，不仅仅是代码审查，还有端到端的内存管理、测试、沙箱环境——就像我们之前那样，你和我都有投资。

I think there's a whole bunch of compelling related opportunities, not just around code review, but the end-to-end memory management, testing, Sandbox is how we were, you and I have an investment.

所以，是的，我上周说过，现在再说一遍，编码是LLM价值的最大来源，对吧？

So yeah, I just think, I said it last week, I'll say it again, coding is the mother load of LLM value, right?

这是所有事情发生的地方，因此，围绕它会有很多相关的机会。

It's where it all happens, and therefore, there's going to be a whole bunch of ancillary opportunities around that.

而且你知道，这些机会可能不是250亿美元那么大，所以你们可能会错过，但如果它们能带来1000万美元的成果，我会接受我那份小支票，并且非常高兴。

And you know, they mightn't be $25 billion, so you guys might pass, but if they're a $10 million outcome, I will take my humble check and be very glad.

也许我不应该把它叫做代码审查，因为罗里的观点是，这样可能显得太狭隘了，但从构建到生产是最大的类别。

Maybe I shouldn't call it code review because Rory's point, it might be construed too narrow, but build to prod is the biggest category.

完全同意。

Absolutely.

从构建到生产。

Build to prod.

你怎么从它已经构建好、能运行、在开发环境中没问题、看起来很棒，

How do you go from it's built, it works, it works in dev, it looks great.

到底怎么把它弄到生产环境里去？

How the hell do I get that thing into prod?

这才是现在能赚大钱的地方。

That's where the big money's going to be made now.

框架构建的钱已经赚到了，而且它们很棒。

The money's already been made in frameworks for build, and they're great.

让Sam和Dario他们每天去争吧，我们都受益，但我怎么把这玩意儿弄到生产环境去？

And let Sam and Dario and friends duke it out every day, we all benefit, but how do I get this crap into production?

我想转到我觉得最有趣的一轮融资，Modal以157.5亿美元的估值完成融资，年化收入超过10亿美元。

I want to move actually to one of the most interesting rounds I thought was Modal closes at a $15.75 billion valuation, run rate surpasses a billion dollars.

这是由Accel领投的7.5亿美元D轮融资。

It's $750 million Series D led by Accel.

我前几天刚采访了Eric。

I just actually interviewed Eric the other day.

他在演播室里表现真的很棒。

He was amazing, actually, in the studio.

我们怎么看待这轮融资，以及它对更广泛的推理市场意味着什么？

How do we think about this round and what it means in terms of the wider inference market?

我觉得这比CoreWeave好。

I like it better than CoreWeave.

我更喜欢为这个世界创造价值，而不是单纯地租一堆GPU机架。

I like adding value to the universe rather than just paying some outsource racks of GPUs.

我更喜欢这个，尽管就像Rory说的，竞争看起来无穷无尽，对吧？

I like this one better, even though to Rory's point, the competition appears to be infinite and limitless, right?

所以你得有点胆量，还得有风险投资的支撑才能做这件事。

So you better have some cojones and venture growth to do it.

不过是的，至少我喜欢为世界创造价值。

But yeah, I like at least adding value to the universe.

Modal有意思的地方在于，他们看起来，他们大概六——他们一直都被开发者认为很强，但我觉得去年年中到年底，他们的收入大概在6000万到1亿之间，对吧？

The interesting thing about Modal is they appear, they were like six-- They were always perceived strong by developers, but I think they were like 60 or under 100 million mid-late last year, right?

那种增长速度简直惊人，对吧？

And that level of acceleration is just amazing, right?

看到的时候真的让人震惊。

It's stunning when you see it.

所以，首先的大背景就是，天哪，推理市场正在爆发。

So, the first big picture is, oh, my God, inference is exploding here.

是的，他们对开发者非常友好。

And yes, they are very developer-friendly.

他们有一大堆产品。

They have a whole bunch of products.

所以，这就是要点。

So yes, that's the takeaway.

我还在想另一件事，这个市场能有多大？

The other thing I was just thinking about it is, how big could this market be?

你看那四五个主要的推理服务提供商，把它们加起来，年化收入大概在50到60亿美元左右，这已经很多了。

I think you look at the four or five key inference providers, and you add them all up, and maybe they're plus or minus $5 or $6 billion in run rate revenue, which is a lot.

但你想一下，它们都在服务所有开放权重的模型，而现在，OpenAI和Anthropic的收入加起来可能已经超过1000到1100亿美元了。

But if you think about it, they're all serving all the open-weight models in a world where right now, and it's not like with like, the sum of the OpenAI and Anthropic revenues is probably north of 100, 110 billion right now.

所以按金额算，还不到5%。

So by dollars, it's less than 5%.

这正好说明了这里的机会有多大。

So it just points to the opportunity here.

回到Jason提到的Replit那个人，想想Replit的Amjad坐在那里想：“我是不是要给Claude Code送钱，让他们在这个token上赚70%的毛利？”

Going back to Jason's Replit guy, think about Amjad at Replit sitting there going, "Do I want to give money to Claude Code and give them a 70% gross margin for this token?

还是说我要在Modal上买，让他们在这个token上赚30%的毛利？”

Or do I want to buy it on Modal and give them a 30% gross margin for this token?"

如果token差不多的话，你知道你想怎么做。

Well, if the tokens are roughly the same, you know what you want to do.

所以是的，这确实说明了那里的上升空间。

So yeah, it does point to the upside there.

完全同意，确实如此。

Completely agree it does.

我一直对推理市场无限扩大的规模感到震惊。

I'm constantly shocked by the infinite expanding size of the inference market.

绝对。

Absolutely.

不，这太惊人了。

No, it's stunning.

这是我今天知道的一件事，真希望我12个月前就知道，因为如果当时我的想法更开阔一些，我12个月前的做法可能会有所不同。

It's the one thing that I know today that I wish I'd known 12 months ago because I would've done things slightly differently 12 months ago had I been more expansive around my mindset here.

好吧，那你得比我聪明，才能明白有多少收入会流到OpenAI和Anthropic之外。

Well, so you had to be smarter than me to understand how much this revenue could be captured outside of OpenAI and Anthropic.

一年前，我还没聪明到能真正想清楚这一点。

A year ago, I wasn't smart enough to really think through this.

我当时觉得流到外面的部分不会太多，所以没觉得这有多有意思。

I didn't think enough of it would be captured outside of it, so I didn't think it was that interesting.

但这些推理服务提供商能值3000亿、4000亿、5000亿美元吗？

But are these inference providers going to be worth 3, 4, $500 billion?

我总觉得这是两件事之间的赛跑。

I always think it's a race between two things.

你可以戴上MBA的帽子，说开放权重。

You can put your MBA hat on and say open weight.

回到Jason对CoreWeave的评论，所有新云厂商都在说：“天哪，让我理清楚。”

Going back to Jason's comments on CoreWeave, all the neo clouds are saying, "Oh, my God, let me get this straight.

所有的建设工作都是我们做，所有的算力都是我们买，所有的资金都是我们借，然后这些家伙跑过来，从我们这儿买，再卖给开发者，赚得盆满钵满。

We do all the building work, we buy all the compute, we borrow all the capital, and then these bozos come in, buy from us, sell it to developers, and make a shit ton.

那我们为什么不自己干呢？

Why don't we do that?"

所以，你戴上MBA的帽子，可以想象这里的竞争会激烈得多，对吧？

So, you put your MBA hat and you can imagine a lot more competition here, right?

然后你又会想，这样的估值恐怕很难实现。

And then you say to yourself, "I doubt you'd get those kind of valuation."

但另一方面，当你有这么高的增长率时，其他什么都不重要了。

On the other hand, when you have this kind of growth rate, nothing else matters.

它能掩盖很多问题。

It covers a multitude of sins.

我猜今年的利润率已经悄悄上升了，对吧？

My bet is the margins have crept up this year, right?

确实如此。

They have.

Fireworks现在已经有30多倍了，很快就要到50倍了。

Fireworks in mid-30s now, and they'll be hitting 50 pretty soon.

就算保持在30多倍，公开市场也能接受。

Even mid-30s if it stays there, the public markets can tolerate that.

关键是最终你能得到的估值倍数，对吧？

It's just the multiple you'll ultimately get on all of it, right?

这样的人很多。

There's plenty of folks like that.

所以不管最后结果如何，都没问题，对吧？

So it's okay wherever it ends up, right?

对Accel来说，以160亿的估值敲定很重要，但我觉得无论怎样都行。

It's important for Accel probably nailing it at 16 billion, but I think it works either way.

你觉得推理领域的三大超大规模服务商是Base10、Fireworks和Modal吗？

Do you think the three kind of hyperscalers of inference are Base10, Fireworks, and Modal?

如果简单看谁赢的话，你可能是对的。

Probably in some simplistic who's the winner, you're probably right.

这三家是最大的，但也许我想说的是，这不像超大规模云服务商那样，其他人都无关紧要。

Those are the three biggest, but maybe the statement is this, I don't think it's like the hyperscalers where no one else mattered.

Jason，这是一轮很大的融资，但你在投资委员会上决定以157亿美元投Modal，这太容易了。

Jason, this is a big funding round, but it's too easy for you in the investment committee to decide that you're probably going to do Modal at $15.7 billion.

所以如果我们开投资委员会，还有另一轮已经宣布的，就是Town以10亿美元的估值。

And so if we call an investment committee, there was another round that was announced, which was Town at a billion dollars.

我记得是1.08亿美元的B轮。Jason，我们要用成长基金来做这个吗？

I think it was $108 million Series B. Are we going to do this from our growth fund, Jason?

Harry，你知道的，我后悔错过了Modal，让Excel那帮人钻了空子。

Well, look, Harry, as you know, I regret passing on Modal and letting the Excel guys sneak in there.

他们很激进，而且在AI投资方面很多地方都落后了，所以我的意思是，他们周末就把交易做了。

They're aggressive, and they're behind in many ways on the AI investments, so I mean, they did the deal over the weekend.

我要求一个工作日的时间，但在Modal那里没得到。

I asked for one business day, and I didn't get it at Modal.

Modal从各方面来看，都很便宜。

Modal was, by all intents, it was cheap.

对，五个月从30亿降到10亿。

Yeah, I mean, $3 billion to $1 billion in five months.

到年底，我们最终会达到14亿。

At the end of the year, we end up at $1.4 billion.

明年，50亿。

Next year, $5 billion.

即使利润率永远达不到40%，我们在这笔交易上也能赚钱。

Even if the margins never hit the 40s, we make money on this deal.

市场崩盘时，这毫无疑问。

There's just no question what the markets crash.

Town我就不投了，Harry，如果你和Rory以及团队其他人想做，我没意见。

Town I'm going to pass on, Harry, and if you and Rory and the rest of the team want to do it, I'm fine with it.

现在的10亿就相当于以前的1亿，对吧？

A billion today is the new 100 million, right?

10亿只是经典B轮融资的入场门槛。

A billion is just the entry point for a classic B round.

以前我们三个人见面的时候，B轮投资人常说，只要不超过一亿就行，因为经典的B轮投资目标是要十亿美元以上的IPO。

In the old days, when the three of us met, Series B investors used to say, as long as it's 100 million or less, it didn't matter because the classic Series B wanted a billion-dollar-plus IPO.

他们想要十倍回报，所以一亿、八千万都行。

They wanted their 10X, so 100 million, 80 million.

他们不在乎，对吧？

They didn't care, right?

所以现在十亿就是以前的一亿。

So a billion's the new 100 million.

Town这家公司没什么数据。

Town has no data.

收入有点意思。

It's sort of interesting revenue.

团队很聪明。

It has a smart team.

这纯粹就是一场赌注。

This is just a bet.

这就像是一个老派的A轮或B轮项目，只不过估值到了十亿，还打了个蝴蝶结。

This is just an old classic Series A, Series B with a bow tie at a billion.

如果这是你未来12个月想做的十个项目之一，那就直接做吧。

And if this is one of the 10 you want to make in the next 12 months, just do it.

我就把它当成一个筹码，不需要参加剩下的会议。

I just consider it a chip, and I don't need to be at the rest of the meeting.

但这里的信息量确实不够，让我提不起兴趣，不管是看好还是不看好。

But there's just not enough here for me to get excited either way.

一群来自Plaid的人。

Bunch of guys from Plaid.

那是个好项目，但如果是今天，我们连Plaid都不会投。

I mean, it's a great one, but we wouldn't even do Plaid today.

我甚至看不到Plaid能退出到250亿，所以我根本不会把它拿到投资委员会上去。

I don't even see a $25 billion exit for Plaid, so I wouldn't even bring it to the IC.

现在预测它的走向还为时过早，对吧？

It's too early to predict, right, where it's going to go.

现在说还为时过早。

It's too early.

这只是一个经典的赌注。

It's just a classic bet.

那我们就不做了，杰森？

So we're not doing it, Jason?

好消息是，团队预测未来12个月只需要大约5亿美元的推理计算。

Well, I mean, the good news is the team only predicts it needs about 500 million of inference over the next 12 months.

这比Instinct团队需要的50亿美元好多了。

So that's much better than the Instinct team, which needs 5 billion.

所以从某种意义上说，我喜欢这个项目，因为我们比同行更保守一些。

So in a sense, I like it for the fund because we were a little more conservative than our peers.

所以我们只需要凑一些SBV，再投入5亿美元来支持推理计算。

So all we need to do is round up some of our SBVs and put in another 500 million to fund the inference.

所以从这个角度看，还是有积极的一面的，对吧？

So in the sense, there's a positive here, right?

因为我很庆幸我们没投Instinct，虽然我们看起来像傻瓜，因为我不知道谁会来给这个大家伙投钱。

Because I'm glad we passed on Instinct, even though we look like fools, because I don't know who's going to fund this bad boy.

至少Town，再说一次，他们未来12个月只需要5亿。

At least Town, again, they only need 500 million over the next 12 months.

所以如果我们是为了保护下行风险，那就干吧，各位。

So if we're about protecting our downside, let's do it, guys.

我就是不知道。

I just don't know.

你宁愿投Instinct的300亿，还是Town的10亿？

Would you rather do Instinct at 30 billion or Town at a billion?

我只想知道我们能不能以20的估值重新进入Modal。

All I want to know is maybe we can get back into Modal at 20.

我们周末刚错过了那个机会。

We just blew that one over the weekend.

Excel那帮人，他们就是太激进了。

The Excel guys, they're just aggressive.

我觉得我们应该把这段视频存档，两年后，当你意识到十亿美元仍然是一笔巨款时，我们可以再回来看这段讨论。

I think that we should put this clip in the vaults, and two years from now, when you realize that a billion dollars still really is something, and we can just revisit this discussion.

当我们随口说出十亿美元这样的数字时，听起来好像不算什么。

Whenever we're throwing blithely around numbers like a billion dollars, I mean, yeah, it doesn't sound like a lot.

这个国家只有四五千个亿万富翁。

There's only 4,000 or 5,000 billionaires in the country.

这是真金白银，对吧？

It's real money, right?

至于三百亿，那更是极其稀少了。

And as for 30 billion, it's pretty damn scarce.

但好吧。

But okay.

你知道我昨晚写LP报告时在想什么吗？

Well, you know what I was thinking as I was writing my LP report last night?

这正好说到Harry清单上的下一个点，Anduril估值九百亿，对吧？

It goes to the next point on Harry's list, Anduril at 90 billion, right?

我在想Town和其他这些公司的时候，听着，Anduril显然是一家划时代的公司，对吧？

And what I was thinking about with Town and all these is, listen, Anduril is clearly a generational company, right?

早早进入了下一代国防领域，显然很深入，有合同，什么都有，对吧？

Got in early on next-generation defense, clearly deep, has the contracts, has everything, right?

但这家公司真的值40倍营收吗，考虑到它的利润率和其他情况？

But is this thing worth 40 times revenue with the margins and everything it has?

我觉得核心问题是，我们做的很多轮融资，也许Modal不算，是不是和公开市场的可比公司脱节得太厉害了，以至于要么需要这些大型公开公司重新调整估值，就像Anthropic那样，要么有些交易根本说不通，哪怕公司本身很优秀。

And it's like, I think the meta-question is, are a bunch of the rounds we're doing, and maybe Modal isn't one of them, are they so disconnected from public comps that we either need these massive public comps to re-correct, right, in a way, and have the Anthropic do, or some of these deals just make no sense, even though the companies are great.

Anduril在900亿估值下说得通吗？

Does Anduril make any sense at 90 billion?

Town在10亿估值下说得通吗？

Does Town at a billion make any sense?

而Braze只是整个公开GTM领域最好的退出案例，也才6倍营收。

I mean, when Braze was just the best exit in the whole public GTM category at six times revenue.

Monday才2倍。

Monday's at two.

HubSpot也是两倍。

HubSpot's at two.

Klaviyo也是两倍。

Klaviyo's at two.

两倍营收。

Two times revenue.

增长率大概百分之二十多，对吧？

Growing 20-something percent, right?

我不是说它们完全一样。

I'm not saying they're the same.

增长点在这里。

The growth is here.

但我确实在想，有些——我的意思是。

But I do wonder if some-- Here's my point.

如果估值倍数和它们最终会达到的水平脱节了，这些倍数。

If the multiples are disconnected from where they will land, the multiples.

而且这还只是风投的情况，对吧？

And that's just venture, right?

我的看法是这样的。

I mean, my comment would be this.

我觉得这几乎肯定是事实。

I think that's almost certainly true.

估值确实和实际脱节了，唯一能让这种脱节看起来合理的就是增长率。

The multiples are disconnected, and the only thing that kind of makes that plausible is the growth rate.

你看，如果你是一家成熟的国防承包商，年收入20亿美元，你最多也就值80亿。

Look, if you're a mature defense contractor doing $2 billion in revenue, you're worth $8 billion, best case.

但如果你是一家年收入20亿美元的国防承包商，过去几年收入翻倍，而且你觉得未来五年能跟政府或五角大楼搭上线，那也许你可以按2倍、4倍、8倍、20倍来算，这样估值就说得通了。

If you're a $2 billion revenue defense contractor, and you've doubled for the last couple of years, and you think you've got a line to the administration or the Pentagon for the next five years, then maybe you go 2, 4, 8, 20, and you justify the valuation.

所以也不是不可能，但就像你说的，Jason，这很大程度上是在赌未来，对吧？

So it's not impossible, but to your point, Jason, it's leaning in a lot into the future, right?

确实，赌得很大。

A lot.

而且这些估值，你可能会在某天醒来时发现，这家很棒的公司只是定价错了。

And all these valuations, you could wake up one day and decide amazing company's just priced wrong.

因为有很多因素可能会帮你解围，你刚才顺带提了一个，美国市场现在的估值已经很高了，从这里再大幅提升倍数来救你的可能性很小，我猜。

Because of all the things that are likely to bail you out, just because you mentioned one other in passing, it's highly unlikely that the US market, which is trading at an-- You're not going to get huge multiple expansion from here that saves you, I would guess.

这是不是又回到了杰森之前提到的那个点，就是需要转移大量资金？

Does it not go back to the point that Jason made earlier about the need to move large amounts of money?

实际上，如果你需要转移十亿或二十亿美元，我完全同意你的看法，从这里开始你会看到倍数压缩。

And actually, if you need to move a billion or two billion, I agree with you totally on the multiple compression you're going to get from here.

但有没有可能它是一家两千亿美元的公司？

But is there a chance it's a $200 billion company?

对，不是不可能。

Yeah, not impossible.

所以那是2.2到2.5倍。你从十亿移到了三十亿。

So it's a 2.2, 2.5 X. And you moved a billion to 3 billion.

我只是好奇。

I just wonder.

而且，听我说，做这个收购要约的人，都是非常精明的投资者，对吧？

And listen, whoever's doing this tender, these are very sophisticated investors, of course, right?

我们别取笑他们。

Let's not make fun of it.

但我确实在想，我们今天是不是把收入倍数推得太高了。

But I do wonder if we are just taking revenue multiples too far today.

我给你举另一个例子。

I'll give you a different example.

比如Profound或者Harry's和Pique，对吧？

Like Profound or Harry's and Pique, right?

这些公司都很有意思，对吧？

These are interesting companies, right?

但天哪，你看公开市场的可比公司，交易价格是收入的两倍。

But my God, when you look at the public comps are trading at two times revenue.

好吧，你可能会说：“不不，这些是搞智能体的。”

Okay, and you can say, "No, no, these are agentic guys."

对，这些是增长很快的智能体公司。

Okay, these are agentic players who have high growth.

但当增长放缓到70%或60%的时候呢？

But when the growth, even-Just slows to 70 or 60.

它们的估值会像Cloudflare那样，还是比Braze更好？

Will it trade at a Cloudflare revenue, or will it trade at a better than Braze revenue, right?

Braze到今天为止估值是收入的3倍。

Which until today was trading at three times revenue.

增长这块我理解。

I get the growth side.

这就是我们这行的工作。

It's the job we're in.

但我确实在想，那些低倍数公司的AI版本，会不会过早地就定在了同行的估值水平上。

But I do wonder if the AI versions of low multiple companies will settle too early into the multiples of their peers.

我也不确定，对吧？

I just don't know, right?

因为我觉得它们正在受到惩罚。

Because I think they're being punished.

我觉得整个营销板块是公开市场里表现最差的板块，这一点毫无疑问，对吧？

I think the whole marketing category is the worst public category, without question, right?

它们并没有在消亡。

They're not dying.

HubSpot 和它的同行们并没有在消亡。

HubSpot and friends are not dying.

Klaviyo 也没有在消亡。

Klaviyo is not dying.

它们的估值已经低于三倍营收了。

They're trading at below three times revenue.

我很难让投资委员会对它们感到兴奋。

I would have a hard time getting the IC excited about it.

Jason，说到估值倍数，以及公开市场如何看待估值倍数，在哪些地方被认可，在哪些地方不被认可。

Jason, speaking about multiples and how public markets think about multiples, where they're appreciated, where they're not.

对我来说，一个特别有意思的事情是伯克希尔加码了Bending Spoons。

A fascinating one for me was actually Berkshire doubling down on Bending Spoons.

显然，它买了3.3亿美元的新股，导致股价大幅上涨。

Obviously it bought $330 million of new shares, caused a massive bump in the stock price.

而Bending Spoons收购的资产，传统上都是有点棘手的，增长放缓，不是那种明星资产。

And the assets that Bending Spoons is buying are traditionally slightly challenged assets, slowing growth rates, not the darlings.

但现在公开市场给Bending Spoons这家控股公司的溢价非常高。

And then the price premium that's being placed on Bending Spoons, the holding company, right now in public markets is enormous.

我就是很好奇，各位。

I'm just intrigued, boys.

我们是怎么看待这件事的？

How did we think about this?

完全没想到。

Did not expect it.

我觉得这挺合理的。

I think it makes sense.

我觉得Bending Spoons就像是软件界的伯克希尔，对吧？

I think Bending Spoons is like the Berkshire version of software, right?

就像喜诗糖果或者可口可乐那样。

It's like See's Candy or Coke.

他们不是买来再转手卖掉。

They're not flipping them.

这些是控股公司。

These are holding companies.

沃伦·巴菲特想永远持有喜诗，对吧？

Warren Buffett wanted to hold See's forever, right?

而Bending Spoons在说：'听着，我们会很谨慎。'

And Bending Spoons is saying, "Listen, we're going to be diligent.

我们每年看上千个交易机会，或者类似的数量。

We look at a thousand deals a year" or whatever you have.

哈里，你们每年是看5000个交易机会吗？

What is it, Harry, 5,000 deals a year?

我们打算每年做五六个项目，然后找到那些能长期持有的优质软件资产，尽可能长时间地持有它们。

We're going to do five or six, and we're going to find enduring software assets that we're going to hold for as long as we can hold them.

如果你退一步看，我觉得优质资产的情况正在改善。

And if you actually step back from it, I think quality assets is improving.

Vimeo，算是他们留着的一个不太好的项目。

Vimeo, kind of a dog that they kept.

Evernote，也不太理想。

Evernote, kind of a dog.

现在他们通过提价让这些项目盈利了，对吧？

Now they made them work with price increases, right?

这两个交易他们都能赚钱，对吧？

They're going to make money off both those deals, right?

Eventbrite，也不太理想。

Eventbrite, kind of a dog.

像Miro和Airtable，我觉得会更好一些。

Like Miro and Airtable, I'm going to argue better.

所以我觉得伯克希尔的意思是，'听着，你们要把这些变成一堆喜诗糖果和可口可乐。'

And so I think Berkshire's saying, "Listen, you're going to make these into a bunch of See's Candies and Coca-Colas.

人们觉得你们没法让它们长久，但你们会让这些持续十年甚至更久。

And folks think you can't make them enduring, but you're going to make these endure for a decade or longer."

也许我错了，但我觉得他们最终可能看到了某些相似的特性，能让这些成为长期资产。

Maybe I'm wrong, but I think they ultimately probably see some similar attributes that they can make these long-lasting assets.

对。

Yes.

就像喜诗糖果一样。

It's just like See's Candy.

拿一盒1960年代的喜诗糖果和今天的喜诗糖果，放在一起看，它们看起来几乎一模一样。

Get a See's Candy from the 1960s and get a See's Candy from today, and you look at them side by side, they pretty much look like the same thing.

我觉得2080年的印象笔记产品可不会是这样。

I don't think that'll be the case for the Evernote product of 2080.

不，但我的意思是，这可能是软件领域最接近的了，就是Bending Spoons的模式，因为他们会保留这些资产，继续投资它们，对吧。

No, but my point is it may be as close as you get in software, the Bending Spoons model, because they're going to keep these and keep these assets, right, and invest them.

我不是说它跟喜诗糖果一模一样。

I'm not saying it's the same as a See's Candy.

我觉得这个概念上跟伯克希尔·哈撒韦是一样的，但可能在一个不太适用的行业里。

I describe it as it's conceptually the same idea as Berkshire Hathaway, but perhaps in an industry where the idea doesn't apply.

或者就像简街把所有的钱都投给那个叫什么来着，利奥·阿皮切拉。

Or maybe it's like Jane Street putting all the money into whatever his name, Leo Apicella.

伯克希尔会说：“这跟我们做的差不多。

Berkshire's like, "This is close to what we're doing.

这不是我们的核心业务。

It's not our core.

我们让那些意大利人去搞，然后我们拿一个重要的股份，我们觉得他们有点东西。

We're going to let the Italian guys do it, and we'll take a meaningful position, and we think they have something.

我们觉得他们能保持这种惊人的增长水平，因为他们一年能看上千个资产，而且我们能以维斯塔做不到的方式来维护它们。”

We think they're going to be able to maintain this incredible growth level because they will have a thousand assets to look at at a year, and we'll be able to maintain them in a way Vista can't."

各位，你们也可以引导一下。

Guys, you can also guide.

我们可以聊聊General Atlantic准备IPO的事。

We can do General Atlantic set to IPO.

我们也可以聊聊特斯拉的新款Roadster。

We can do Tesla, the new Roadster.

那些车用推进器把自己抬起来的视频，简直不可思议。

Unbelievable videos of cars that literally elevate themselves with thrusters.

我们还可以聊聊Benchmark在台上承认，如今在风投圈里，竞争这个词有点站不住脚。

We can do Benchmark on stage conceding that competition, eh, it's a bit of a flimsy term in venture these days.

你有没有特别想聊的方向？

Any predilection as to where you want to go?

我想聊37 Signals那个，但是……

I wanted to do the 37 Signals one, but-

我们可以聊37 Signals。

We can do 37 Signals.

David Hanema Hanson，他是一位非常出色的赛车手，拥有最丰富的赛车收藏。

David Hanema Hanson, who is an incredible race car driver, has the most extensive library of race cars.

37 Signals 自2020年以来盈利了3.7亿美元。

37 Signals made 370 million in profit since 2020.

自力更生也没那么糟糕，对吧？

Bootstrapping ain't all that bad, hey?

对，我正想说，回到我们一开始聊的话题，这些软件公司，没有一个风投在乎。

Yeah, I was going to say, going back to the conversation at the start, all of these software companies, none of the VCs cares.

如果你自己留着，赚了3.7亿美元，我觉得你肯定没问题。

Well, if you hold onto it yourself and pull out $370 million, I think you'll be fine.

那真是个了不起的成绩。

That was a great result.

我知道很好，但是——

I knew it was good, but-

我觉得那不对。

I think that was wrong.

不管怎么说，我觉得方向是对的。

For what it's worth, I think it's directionally correct.

我觉得有意思的一点是，他们说他们做10%的利润分享，对吧？

I think the point which was interesting is they say they do 10% profit sharing, right?

所以如果你做3700万的利润分享，那就是10%。

So if you do 37 million of profit sharing, that's 10%.

Rory可以帮我算一下这个数学。

Rory can help me with the math.

我觉得那意味着总利润是3.7亿。

I think that would suggest overall profits were 370.

我不觉得它能赚那么多钱。

I don't think it could make that much money.

我猜它一年大概赚6000到7000万，差不多6000万左右，对吧？

I'm guessing it does 60 to 70, 60-ish million a year, right?

即使有4000万能到账上——

And even if 40 million can go to the bottom line-

4000万乘以6。

40 times six.

你说得对，240。

You're right, 240.

可能差不多吧？

It may be in the zone, right?

可能稍微高一点。

It may be a little high.

但你看那些车队什么的，显然他们一直在赚这么多钱。

But when you look at the fleets of cars and all that, obviously they've been making this kind of money.

我跟你说说我的一些想法。

I'll tell you a couple thoughts I had.

一个是，在这个时代值得反思的事情。

One is it's something to reflect on in the era.

我觉得我们不会有很多一个人的十亿美元初创公司，但团队可能会小得多，对吧？

I don't think we're going to have a lot of one-person billion-dollar startups, but we could have much smaller teams, right?

我想到的关于AI时代的一个例子是，天哪，Zapier本该是这样的。

The one I thought about for the AI era is, man, Zapier should have been this.

大家总是拿Zapier作为风投界一个典型的成功案例。

People throw out Zapier as The one and done example in venture.

对风投来说，这当然很好，对吧？

Like and for VCs, it's great, right?

你从YC出来，就融了一轮资。

You raise one round of whatever out of YC.

我记得Bessemer那边有过一轮低调的二手股权交易，但基本上就一轮风投，现在他们的营收已经达到3亿美元了。

I think there was a quiet secondary round from Bessemer, but really one venture round and you get up to th- they're at 300 million in revenue today.

增长可能放缓了，以现在的公司形态可能不会上市，但这是一家了不起的公司，做出了划时代的成就。

Growth has probably slowed, probably not going to IPO as its current company, but a great company that did something generational.

如果他们当初一分钱都没融，就我们三个人把那3亿里的2亿分了，会怎么样？

What if they raised nothing and the three of us just split 200 of the 300?

这样也不算差。

Like you could do worse.

对，我觉得这样已经相当不错了。

Yeah, I think you could do a lot worse.

GA那些人可能赚得更多，但你已经不错了，对吧？

I mean, the GA guys may make more, but you could do worse, right?

所以我几乎在想，这种一次性融资模式，我以前还挺喜欢的，会不会在新时代变成一个陷阱。

So I almost wondered if this one and done, which I used to be a fan of, is going to be a trap in the new age.

不融资反而更好，因为一次性融资很难把分红做好。

None and done j- is much better because o- it's tough to do dividends right in one and done.

一次性融资也能做好分红，但难度更大。

You can do them right in one and done, but it's, it's tougher.

和Benchmark团队一起做分红更难，对吧？

It's tougher to do dividends with the Benchmark team, right?

他们其实对分红不太感兴趣，是不是？

They're not really that interested in the dividends, are they?

你不能直接把投资者买断吗？

Can you not just buy out investors?

如果你是Zapier，你就直接说：‘各位，这是五倍的估值。’

If you're a Zapier, you just go, "Listen, guys, it's a five X multiple.

我们会以15亿美元的价格把你们买断。

We'll, we'll buy you out at a $1.5 billion price.

你们想要流动性嘛。

You want the liquidity."

可以。

You can.

这确实会变得很……我是说，这种情况不常发生……也许罗里能举些例子。

It gets ex- I mean, it doesn't happen that... Maybe Rory can come up with some examples.

确实会发生，但用自由现金流来做成本很高。

It does happen, but it gets expensive to do it from free cash flow.

就算你有很多钱，也没那么简单，对吧？

Even if you have a lot of it, it's not that simple, right?

确实不简单。

It's not.

但也不是做不到，对吧？

It can be done, right?

Buffer就做到了。

Buffer did it.

Expensify也做到了，但他们做的时候，投资者的回报都挺一般的。

Expensify did it, but they all did it at pretty mediocre outcomes to the investors.

我觉得Redpoint在Expensify上几乎没赚到什么钱，但可能比他们坚持到IPO还是要好一些，对吧？

I think Redpoint made like nothing on Expensify, but probably more than if they'd held IPO, right?

这事儿确实挺难的。

It's, it's tough.

我觉得现在“零和”比“一和”更好。

I like none and done better than one and done now.

嗯，这事儿难就难在，你唯一能支付的方式……你看，如果他们只占公司一小部分股份，那一两年的现金流就能搞定，但如果你想买断一个占30%股份的投资者，而且想用六倍的现金流去买断，那按定义来说，你就得占用整整两年的现金。

Well, the reason it's tough is because the only way you can pay... Look, if they own a small percentage of the company, then, you know, one or two years cash flow can take care of it, but if you're trying to buy out an investor that owns 30% of the company, and you want to buy it out for six times cash flow, by definition, you're tying up two whole years cash.

所以如果用债务来做这事儿，风险就大了。

So it gets risky if you're doing it with debt.

所以你说得对，这些交易确实很难谈。

So you're right, these, they're hard deals to negotiate.

但是你看，如果你需要一两百万美元才能起步，你会先做这笔交易来启动。

But look, if you needed a million or $2 million to get going, you'd do the deal to start.

但总的来说，我们想说的是，有很多软件公司可能不会让人特别心动，但它们能产生现金流，对创业者来说非常有价值，我觉得这很棒。

But yeah, what w- the big picture we're saying is there are a number of software companies that mightn't set anyone's little heart aflutter, but are kicking off cash and are extraordinarily valuable to their entrepreneurs, and I think that's great.

这也说明，不是所有事情都要通过风险投资的视角来看，非要“要么赚250亿，要么死磕到底”。

And it just show... Not everything has to be through the lens of venture, "I want to make 25 billion or die trying."

这些公司能在这里赚钱，真的很好。

And it's really good that these guys are making money here.

确实。

It is.

根据我的经验，当我拿它来比较时，我之所以选Zapier，是因为他们做到了3亿美元，算是他们自己那一代的杰出公司，但可能不会上市。

In my learning, when I compare it to, I just picked Zapier as, because they were at 300 million and a generational company in their own way, but probably won't IPO.

嗯，我以前觉得这种“一次性退出”就是理想境界，对吧？

Um, it's like I used to think this one and done was a nirvana, right?

我们三个人聚在一起，想创办一家公司。

The three of us get together, we want to start a company.

我们是第一次创业。

We're first time founders.

我们没有那么多钱。

We don't have the money.

我们从二十家风投那里融了几百万的种子轮，然后就一直把公司做大，对吧？

We raise a couple million from 20 VC seed or whatever, and we just keep growing our company, right?

而且压力也不大。

And it's not a lot of pressure.

没人……

No one...

我们不需要有董事会成员，但我们三个当时确实没有资源去做这件事。

We don't have to have board members and, but we don't, we didn't, the three of us didn't have the resources to do it.

我以前觉得那是个很好的结果。

And I used to think that was a great outcome.

现在在AI时代，我觉得那是个陷阱，因为利润分红没那么容易了。

Now in, in the AI era, I think it's a trap because you can't dividend out the profits as easily.

你不可能那么轻易就变成37signals那样的公司。

You can't become a 37signals that easily.

他们当时拿了贝索斯或其他人的钱。

Now they, they took money from Bezos or others.

我相信他们后来搞明白了，对吧？

I'm sure they figured it out, right?

嗯，但我有点怀疑，在某些方面，不融资是不是比融一轮更好。

Um, but I almost wonder if no venture rounds is, is the way to go rather than one in some ways.

Rory，你怎么看GA上市这件事？

Rory, how do you feel about GA IPO-ing?

我们还会看到更多公司上市吗？

Are we gonna see many more firms IPO?

我觉得你得达到一定的规模才行。

I think you have to be a critical mass.

我刚才也在看这个。

You know, I was looking at it there.

你得有八亿美元左右的费用收入，对吧？

You have to be 800 million plus or minus a billion of fee income, right?

这才是真正的问题所在。

That's the real question is.

你看，这些公司，退一步讲。

Look, these companies, step back.

显然大家都知道那些公开上市的私募股权公司，比如KKR、凯雷、阿波罗、TPG，然后问题是，从增长和风险投资的角度来看，你能做这么多这样的公司吗？

There's a bunch of, obviously everyone knows publicly held PE-backed companies like, you know, KKR, Carlyle, Apollo, TPG, and then the question is, can you do a bunch of these for the kind of growth and venture perspective?

而制约因素就是费用收入，通常公开投资者会给你很多管理费收入的信用，但对附带权益却很少给信用，这在风险投资的角度来看真的很奇怪，但不管怎样，要达到一个有意义的市值，因为他们最多按收入的四到五倍交易。

And, you know, the, the, the binding constraint is the fee income, and then typically, the public investors give you a lot of credit for the fee, management fee income and very little credit for the carry, which is really odd in a venture conte- in a venture perspective, but be that as it may, to get to a market cap that matters, because they trade at four or five times revenues best case.

所以你可能需要有八亿美元的费用收入。

So you're probably looking at having to have 800 million in fee income.

现在你对自己说，按2%算，那是多少？

Now you say to yourself at 2%, what's that?

50乘以8亿——50乘以8。

50 times 800 mil- 50 times eight.

是的，你实际管理的资产得有几十亿美元，甚至更多。

Yeah, you got a couple hundred million dollars, well more of actual AUM.

所以这是可行的。

So it's doable.

安德森·霍洛维茨可能很容易就能达到这个目标，但这就是你在这里看到的情况。

You know, Andreessen probably will g- could get there easily from here, but it's, that's, that's kinda what you're looking at here.

不过，也许你可以用更少的资产管理规模就达到目标，因为风险投资一方面收费稍高，另一方面附带权益更有意义。

Now, maybe you'll get to be able to go on a lower amount of AUM because venture, A, get paid slightly more, and B, has carry that's more meaningful.

所以最大的公司能做到，但能做到的不多，这就是答案。

So the very largest companies will do it, but not many is the answer.

这纯粹是……

And that's just purely...

我，我刚刚是从财务指标的角度来回答的。

I, I was a very financial metrics answer.

让我再说一个完全不同的观点。

Let me make a s- totally separate comment.