20VC：SpaceX以600亿美元收购Cursor | Stripe押注OpenRouter投入80亿美元 | Anthropic实现首次盈利及迈向6000亿美元营收背后的数学逻辑 | Lovable和Higgsfield完成巨额融资

罗里说得对。

Rory is right.

Stripe 在收购方面确实很擅长。

Stripe actually appears to be very good at acquisitions.

这也是他们快速拓展加密货币业务以及其他领域的方式。

It's how it accelerated into crypto and otherwise.

他们在这方面很有经验。

They're good at it.

换个角度说，也许他们自己应该更擅长搭建这些系统吧？

The flip side you could say is maybe they should be better at building these themselves, right?

这是比较挑剔的说法。

That's the grouchy version.

你们怎么不自己搞呢？

Why didn't you build it?

但如果你擅长并购，而这笔交易只占你市值加现金的5%，而且你明天就要拿到手，那这就很合理了。

But if you're good at M and A and this is 5% of your market cap plus cash and you want it tomorrow, it makes sense.

不过你得擅长并购才行，然后才能这么做。

You have to be good at M and A though, and then you do it.

但反过来说，我很喜欢OpenRouter。

The counterpoint is I love OpenRouter.

大概六个月前我在节目里聊过它。

I talked about it on the show like six months ago.

我是它的客户，也是用户，它真的很棒。

I'm a customer, I'm a user, it was brilliant.

它就像11labs那样，属于那种部署起来特别容易的软件。

It was one of these pieces of software like 11labs, which is instantly easier to deploy.

就是很优雅。

It's just elegant.

它就是个很漂亮的软件，但挺小众的。

It was just a beautiful piece of software, but it's pretty niche.

怎么说呢，Jason？

In what sense, Jason?

就是纯粹的好奇心。

Just a genuine curiosity.

好的。

Okay.

据我了解，OpenRouter 在开发者工具方面确实很强，它能让你很轻松地挑选模型。

So OpenRouter, as I understand it, and I think it's right, is really strong in sort of developer type tools where you want a simple way to pick a model.

明白了，因为你可以随意选择任意模型。

Okay, because you can pick any model.

你不需要搞那些花里胡哨的。

You don't need to get on fireworks.

你也不用去配置什么。

You don't need to set up anything.

它在聊天机器人方面表现真的非常强，尤其是那种不要求完美输出的场景。

And it's really, really, really strong with chatbots where like they don't have to be perfect.

比如跟我那个数字版哈里或者罗瑞对话时，你不需要完美的输出，对吧？

When you're talking with my digital Harry or digital Rory, you don't need perfect outputs, right?

你可以根据可用性在模型之间切换，这就是它们的两个优势领域。

You could route between models based on availability and those are their two niches.

现在我们来谈谈需要大量推理的B2B工作流，这些工作流必须准确无误。

Now let's talk about workflows with a lot of reasoning for B and B when it has to be accurate.

你会把范围缩小到一两个模型，因为不能有模型漂移。

You're going to down spec to one or two models because you can't have model drift.

你不能在Kimmy、Quinn、46和Fable之间来回切换。

You can't be routing from Kimmy to Quinn to 46 to Fable.

突然之间，你的B2B工作流本来必须完美，却因为所有这些模型而出现漂移。

And all of a sudden, your B2B workflow that has to be perfect drifts from all of them.

即使只是从一个Opus模型切换到另一个，也会出现漂移。

It drifts even just going from like one Opus model to another.

你会看到漂移。

You see drift.

你必须进行质量保证、重新验证、修复和测试。

You have to QA it, requalify it, fix it, test it.

所以对于高推理能力的模型，大家追求的是接近前沿水平的输出效果，对吧？

So for high reasoning models, people do frontier esque outputs, right?

没人会在一堆模型里来回切换。

People don't rotate through 11 models.

我觉得 OpenRouter 并不适合这种场景。

And I don't think OpenRouter is the right product for that.

这完全没问题。

And that's fine.

他们也是这么想的。

They get that too.

但我觉得Stripe的逻辑是，嘿，听着，地球上任何一笔交易，都能抽走百分之二点几的手续费，对吧？

But I think Stripe is, hey, listen, any transaction on planet Earth, can take two point something percent of, right?

但这套模式在OpenRouter身上行不通。

But it's not going be true for OpenRouter.

它属于一个细分领域。

It's a niche.

这是一个非常棒的细分产品。

It's a wonderful niche product.

但在路由这个领域，大家都在做，Databricks在做，Replit在做，Lovable在做，Vercel也在做。

But in the world of routing, which everybody does, Databricks does, Replit does it, Lovable does it, Vercel does it.

这是一个细分产品，但有两个非常好的细分市场。

It's a niche product with two really good niches.

但对Stripe来说，风险在于他们最终会拥有一个细分市场，一个成功的细分产品。

But this is the risk to Stripe is that they end up owning a niche, a successful niche product.

但这不是他们的风格，他们的基因不是做细分市场的。

And that's not their their DNA is not niche.

这真的不是他们的基因。

It's just not their DNA.

首先，我觉得你说得对，因为我们内部确实在这个领域纠结过，这正是我们焦虑的地方。

First of all, I do think that's fair because we internally agonize about this space, and that was exactly what we're angsty about.

你的分析完全正确。

And your framing is exactly correct.

咱们把积极的一面说清楚。

Let's spell out the positive.

积极的一面是，只要前沿模型今年想从你身上赚到一千亿美元，而你又是企业客户，你就得有个备选方案，至少得让事情保持透明。

The positive trend is as long as you have the frontier model trying to extract $100,000,000,000 in revenue from you this year and you're an enterprise, you're going to want a plan B, at least to keep the thing honest.

所以你会需要某种运行机制。

So you are going to want some kind of running.

但让我有点共鸣的是你说的，Jason，还记得云计算刚起步的时候，人们都说，哦，我想多云部署。

But it's what you said that resonated with me, Jason, a little bit, which is remember when cloud was starting, people were like, oh, I want to be multi cloud.

但多云部署真的很难。

It's really hard to be multi cloud.

现在呢，也许我想多模型，但可能我只想要两三个模型，所以我不需要这种路由功能。

Here, maybe I want to be multi model, but maybe I only want two or three models, and therefore I don't need this kind of routing functionality.

这就是风险所在。

That is the risk.

如果你的企业客户决定，我需要在三个模型之间切换，而不是十个，那你的价值就下降了。

If your enterprise customer decides, I need to flip between three models but not 10, then your value here goes down.

我觉得，从积极的角度来看，如果你能在选择一大堆模型的基础上，把所有选项都标准化，然后努力把模型商品化，那你的价值对企业客户来说就会提升。

I would imagine the the the positive spin is your value here goes up to the enterprise if you can build on top of just picking a whole bunch of normalization of all those options and try and commodify the model.

所以这就是那种紧张点。

So that that's the kind of tension point.

你越能做到这一点，越能服务那些不太在意细节的人，你的生意就越好。

The more you can do that and the more you can service the people who don't care all that much, the better your business.

但你说得对，如果摩根大通说，我不仅仅想要Anthropic，但我不会去评估十个模型，我只想用poolside作为我的备选方案，然后把其他的外包给别的公司。

But you're right, if JPMorgan says, I want more than just Anthropic, but I'm not going to qualify 10 models, I'm just going to work with poolside as my plan B, and then offload the rest to something else.

那你说得对。

Then you're right.

那你就只能占据一个小众市场，拿不到那部分收入。

Then you have niche and you don't get that revenue.

是的，比如这周，Rippling作为一家B2B公司，发布了他们使用哪些模型的观点，对吧？

Yeah, like for example, like this week, Rippling posted their view as a B2B player of what models they use, right?

而且他们什么都有。

And they had it all.

他们说，在整个Rippling的体系里，我们只看了两个最适合我们的东西。

And they said in the world across Rippling, we looked at two things that were best for us.

Opus 4.8。

Opus 4.8.

它是个N-one模型，但用他们的框架训练得很好。

It's an N-one model, but it's well trained with their harness.

然后还有价格、性能和速度这些因素。

And then there's like there's price, performance and speed.

然后我觉得他们选了GPT 5.5 medium之类的。

And then I think they picked whatever GPT 5.5 medium or something.

他们说，其他的对现在的Rippling来说不值得。

And they said the rest isn't worth it for Rippling today.

不过这种情况可能在六十天、九十天内就会改变。

Now that could change in sixty, ninety days.

所以他们一次只选两个，然后还得管理这些模型的输出。

So they down spec to two at a time, and then they have to manage the outputs from these.

你可以针对这一组工作流进行优化，对吧？虽然它们在另一种场景下可能耗时较长。

And you may tune one set of workflows here, right, that are long reason in another.

即便你像 Rippling 那样规模庞大，同时管理十二个模型也太多了。

And even if you're rippling scale, managing 12 models is too much.

如果你是做开发者工具的，让用户自己选，那也行，对吧？

If you're a dev tool and let people pick, so be it, right?

这对 OpenRouter 来说是个好消息。

That's great for OpenRouter.

或者如果你想在自己的产品里加个备用方案，OpenRouter 在这方面绝对是满分。

Or if you want to build into your own product a fallback, OpenRouter is a 10 out of 10 for this.

比如说，万一某个服务挂了，对吧？

Like, let's say something's down, right?

OpenRouter 会自动切换到备用方案。

OpenRouter automatically falls back.

不过我觉得它算是个细分领域的产品，但这个细分市场可能非常大。

But I think it's a niche product, but it could be a massive niche.

对，我记得那些对话，因为你说得对。

Yeah, I'm remembering the conversations, because you are right.

Stripe核心产品的好处在于所有支付都是平等的，而Visa是所有支付的轨道。

The great thing about the core Stripe product is all payments are equal, and Visa is the rails for everything.

也许是这样。

That might be the case.

我会去看看那篇文章，因为其中一个重要问题是，企业能对封闭的基础模型公司施加多大的定价压力，以及他们如何施加这种压力，因为我认为这影响很大。

I'm going to check out that post because one of the big questions will be how much pricing pressure enterprises can put on the closed foundation model companies, and how do they put that pressure on because I think it impacts a lot.

抱歉，哈里，你继续说。

Sorry, Harry, go on.

五年后，这笔收购会被认为是成功的吗？

In five years' time, will this be considered a successful acquisition or not?

来打个赌吧。

Bets on prediction.

我觉得会像收购Scale那样成功。

I think it'll be like the scale acquisition.

这会是某个更大事业的起点。

It will be the start of something that gets bigger.

不管这个品牌五年后还在不在，甚至这个产品五年后还在不在，我觉得这个产品五年后可能都不存在了。

Whether this brand exists or whether even this product exists five years, I don't even think this product will exist in five years.

但我认为有很高的概率，超过51%的概率，它会发展成Stripe 20%到30%的收入来源。

But I think there's a high chance, more than 51% chance it builds into a 20 or 30% revenue stream for Stripe.

这就够了。

That's enough.

但五年后，OpenRouter作为Stripe的一部分还会存在吗？

But does OpenRouter as part of Stripe exist in five years?

我敢跟你打赌，赌什么都行。

I'll bet you dollars to donuts.

五年时间太长了，而且这是个非常小众的产品。

Five years is so much time and it's such a niche product.

这个产品本身，如果它还存在的话，它会藏在Stripe顶部下拉菜单的深处，大概要往下翻11层才能找到，因为它会被整合进整个令牌管理平台，也就是他们的TMP。

This this product itself, if it does exist, it'll be deep in a drop down menu on the top of Stripe 11 layers down because it'll be subsumed into the whole token management platform, their TMP.

我不知道。

I don't know.

我觉得Jason的回答让我很有共鸣。

I think Jason's answer resonates with me.

如果成功了，那就会被看作是一次扩大市场规模的举措。

If it works, it'll be seen as a TAM expansion play.

现在Stripe有意思的地方在于，他们正在进行的收购，很明显是那种“我们不玩这个领域”的姿态。

What's fun about Stripe right now is they're doing that acquisition, which is very much a, we don't play in this space.

让我们在新领域里插上一面旗。

Let's put a stake in the new ground.

同时，他们又在谈收购PayPal，这又完全是“我们已经掌控了这个领域”的姿态。

And at the same time, they're talking about a PayPal acquisition, which is very much, We own this space already.

把这些人买下来，并入我们已有的业务，然后靠整合大赚一笔。

Let's buy these guys, fold them into what we already have, and just make a shit ton of money consolidating.

说实话，我觉得这是个挺聪明的策略。

And actually, I think that's a clever strategy.

我觉得他们这手牌打得挺巧妙的。

I think they're actually playing a very clever hand.

五年后他们能做成一个巨大的AI路由业务，概率大概三分之一吧。但要是真做成了，那就是第二波大增长。

There's probably a one in three chance that they have a massive AI routing business in five years, but if they do, that's a big second leg.

同时呢，如果PayPal那笔交易谈成了，就像Jason刚才说的，只要你在并购和整合方面够厉害，通常就能比较清楚地看到：你能保住现有收入，砍掉整个G&A（一般及行政费用），还能因为拥有了消费者钱包——这是Stripe没有的——而形成一个更强大的双边网络，再加上核心业务的整合收购。

While at the same time, if they get the PayPal deal done, that's the kind of deal you have a high, to Jason's point, if you're good at M and A and good at consolidation, you probably have a high degree of visibility that you keep those revenues, that you remove the entire G and A, you get more of a two sided network because you have consumer wallets, which Stripe doesn't have, and you've done core consolidation acquisitions.

这两件事如果能同时搞定，对提升企业价值来说，真的非常有趣。

Doing them both together, provided you can pull them off, is super interesting in terms of building enterprise value.

而且他们全是在私下里操作的。

And they're doing it all private.

再说回刚才那个观点，他们做的这些并购案规模都堪比上市公司级别，而且还是在非上市状态下成功完成的。

Again, back to the comment, doing what looked like public company size M and A and pulling it off while private.

我知道他们在OpenRouter那笔交易里，让投资者接受了Stripe的股票作为支付手段。

I know they got the investors to take Stripe stock in the OpenRouter deal.

我觉得其中一部分是用股票支付的，而PayPal那笔交易结构更复杂，估计也花了不少心思琢磨。

I think some portion of it was stock, and the PayPal deal is more complex and probably requires more thought.

但话说回来，能完成一个我认为是400亿到500亿美元的交易，还有一个70亿美元的交易，而且都是私下发行股票完成的，这确实挺厉害的。

But again, being able to do what is, I think, a 40,000,000,050 billion dollars deal and a $7,000,000,000 deal issuing paper while private is pretty impressive.

Stripe的企业发展团队圣诞节真该拿奖金。

Stripe's corp dev team need a bonus at Christmas time.

他们今年可忙坏了。

They are busy this year.

他们今年确实忙，但谁不忙呢？

They are busy this year, but isn't everybody?

我们刚才提到了基础模型利润率压力的问题。

We mentioned margin pressure on foundation models.

Anthropic在第二季度收入115亿美元的情况下首次实现盈利。

Anthropic turns its first profit on $11,500,000,000 of q two revenue.

对Dario来说，业务正在好转。

The business is getting better for Dario.

这周Gavin Baker也提到Dario，说他相信他们会是最后一家私有公司。

This is also in a week where Gavin Baker said about Dario saying he believes that they will be the final private company.

你看到这个消息了吗？

Did you see this?

看到了。

We did.

咱们还是把夸张的未来预期和当下的事实分开来看吧。

And let's separate the hyperbole and the future from the facts and the present.

他们能赚到钱并不奇怪。

It's not surprising they're making money.

你回想一下去年，他们去年做了45亿美元，而且我觉得他们的毛利率从前一年的负值变成了正30%左右。

If you just go back to last year, they did $4,500,000,000 last year, and I think their gross margins went from negative the year before to positive 30 or something like that.

按照这个趋势，我估计年底大概能到40%。

On track, I think end of the year roughly 40.

当你有像40%这样不错的营业利润率，而且年收入从45亿一下子跳到季度100亿时，这意味着你有40亿的毛利，对吧？

When you have decent operating margins like 40% and you go from $4,500,000,000 in a year to $10,000,000,000 in a quarter, That means you have 4,000,000,000 of gross margin, right?

而这仅仅是在两个季度里实现的。

And that's literally in two quarters.

你根本来不及在利润表下方增加费用来阻止自己赚钱。

You can't add expenses below the line fast enough to stop yourself making money.

所以这是不可避免的。

So it's inevitable.

我是说，对，他们增长了12倍，这意味着如果增速继续稍微提升一下，他们的毛利大概会达到14倍，而且趋势一直在上升。

I mean, yeah, they 12 x growth, which means they probably 14 x gross margin if it continued to increase even slightly, and the trajectory has been increasing.

你不可能在六个月内把人员数量或者表外培训成本翻14倍。

You're not going to 14x headcount or below the line training costs in six months.

所以是的，我完全没想到他们能实现综合运营。

So yes, I'm totally not surprised they are operating in composite.

我们在年初算过账，得出的结论就是这样。

We had run numbers at the start of the year that came to that conclusion.

有意思的是，随着他们继续扩张，如果像之前提到的那样从埃隆那里购买昂贵的算力设备，而且两个月内价格就大幅上涨，我怀疑他们在IPO时会不会把持续盈利作为基准预期。

The interesting thing will be as they continue to grow, as they buy that expensive compute from Elon, if you remember, that has a big price increase two months in, I doubt they will forecast for their IPO a base case of continuing profitability.

我可能错了，但这笔利润并不让我意外。

I could be wrong, but this profit didn't surprise me.

说实话，这表现太惊人了。

I mean, it's amazing performance.

营收数据也漂亮得惊人。

It's amazing revenue.

只要毛利率 decent，收入高就能掩盖其他几乎所有问题。

I mean, revenue with any kind of decent gross margin cures almost all else.

我觉得还有个问题，眼看 IPO 迫在眉睫，Anthropic 到底能拿出什么样的数字来过关呢？

I think the other question is, as we gear up for an IPO, could be very imminent, right, is what numbers does Anthropic get away with, Right.

你们有表外负债，对吧？

You've got off balance sheet liabilities, right?

你们还有大量的承诺支出。

You've got massive commits.

你们的股权激励可能史无前例地高。

You've got probably stock based compensation like we've never seen in the history of mankind.

所以如果给你们的数字加上星号和斜杠标记，那结果绝对会惊掉下巴，对吧？

So if you get asterisks and daggers on your numbers, they will be jaw dropping, right?

如果他们必须完全核算这些，而且其中一些是非公认会计原则的，这些表外项目。

If they have to fully account for that, and some of that's non GAAP, these off balance sheets.

如果他们像可怜的Wix或其他公司一样因为股票补偿而被猛烈抨击，每个人都会写出可怕的负面影响，对吧？

If they're going to be hammered like a poor Wix or someone for SBC, and everyone's going to write out the horrific downside, right?

但我觉得大家都会忽略那些数字里能看到的各种技术性负面因素。

But I think everyone's going to look through all the nerdy negative things you could see in the numbers.

他们只会选择无视。

They're just going to ignore it.

对。

Right.

但我确实觉得被无视很重要。

But I do think it's important that it get ignored.

我觉得对Anthropic来说，被无视很重要。

I think it's important for Anthropic it get ignored.

而且我觉得那些破事都不会有影响，用个专业术语来说。

And I think none of that shit will matter, to use a technical term.

唯一重要的就是增长率，还有27、28年的预期收入。

The only thing that matters will be the growth rate and the '27 28 projected revenue.

只要收入到位，其他一切都不是问题。

Provided the revenue comes, everything else will be fine.

如果收入到位，那你需要那些表外的东西，而且你有收入去买它。

If the revenue comes, then you'll need the off balance sheet stuff and you'll have the revenue to buy it.

换句话说，所有这些表外的东西，基本上就是，我承诺两年后从你那里买一大堆算力，因为如果我的收入再增长10倍，我就需要那么多算力。

In other words, all these off balance sheet stuff are basically, I promise to buy a whole load of compute from you in two years' time because if my revenue grows 10x for two more years, I'm going to need all that compute.

如果收入增长，你就需要算力。

Well, if the revenue grows, you need the compute.

你也很乐意拥有它。

You're happy to have it.

事实上，你还会坚持要拿到它。

In fact, you're insisting you get it.

如果收入放缓，那你就不需要算力了。

If the revenue slows down, then you don't need to compute.

情况会变得很棘手。

It all gets hard.

所以几乎所有事情最终都会归结为你未来两三年打算写多少数字。

So almost everything is going to boil down to what number do you want to write for the next two or three years.

然后就像你说的，股权激励这块没人会在意，因为大家担心股权激励是因为像Workday这种成熟公司，我们分分钟就能覆盖掉这部分成本。如果你每年给中层管理人员发500块的股票来让他们出勤上班，他们心里其实已经把这部分RSU算进薪酬包里了，他们会想：我拿了400块现金和500块的RSU。要是你停止给他们RSU，他们就会要求更高的现金薪酬。

And then as you said, the stock based comp, no one's going to care because the reason you worry about stock based comp is because in a steady state like Workday, we can cover that in a second, if you're giving someone $500 every year to show up and be a middle manager, they're probably mentally putting those RSUs into their comp, and they think to themselves, I paid 400 in cash and 500 in RSUs, And if you stop giving them the RSUs, they're going to want cash.

所以归根结底，这还是个现金数额的问题。

So it really is a cash number.

所以在成熟的企业里，关注股权激励费用（SBC）是完全合理的。

So in a mature business, it's totally correct to worry about SBC.

但这里的股权激励费用数字会非常巨大，因为所有这些员工都拿到了期权，结果发现价值远超他们最初的预期。

But the SBC numbers here are going to be huge because all these people got grants, then it turned out to be worth way more than they ever thought.

对，有个很典型的例子，有人23年入职时拿的是百万美元的薪酬包，四年后身价涨到了5100万。

And, yeah, the classic example, someone who was hired with a million dollar package in '23 ended up making 51,000,000 four years later.

但这并不意味着你得给下一个入职的人开出5100万的价码。

That doesn't mean you'd have to pay the next guy 51,000,000.

如果那个人拿到了当初签约时的一百万，那才是真正的经济性股权薪酬。

If that person had gotten the million they signed up for, that's all the real economic stock based comp would take.

另外的五千万纯粹是运气好。

The other 50,000,000 is just dumb luck.

你运气好，这不是常态。

You got lucky, it's not a run rate.

我其实觉得在高速增长的公司里，忽略一部分股权激励成本并将其正常化是可以的。

I actually think it is okay in a hyper growth company to look past a good slug of the SBC and normalize it out.

反过来，在成熟公司里就不行了。

And conversely, it's not okay in a mature company.

在Workday或Salesforce，股权激励是实打实的钱，是人们在花真金白银。

That SBC, stock based comp, in Workday or Salesforce, that's real money that people are spending.

这有点不公平，因为你相当于给了高速增长的公司一个免费通行证，但它们确实有这个通行证。

And it's a little bit unfair because you're kind of giving the hyper growth company a free pass, but they get a free pass.

你拿到免费通行证，这和我们之前说Cursor的情况是一样的。

You get a free pass, and it's the same thing we said about Cursor.

只要收入增长，你在利润率、表外项目以及股权激励这块都能拿到豁免。

You get a free pass on margin, you get a free pass on off balance sheet, you get a free pass on SBC provided revenue go up.

一旦收入停止增长，那就全都不作数了。

Once revenue stop go up, all bets off.

一旦收入停止增长，那就全都不作数了。

Once revenue goes up, all bets are off.

Anthropic 要达到2028年2000亿美元的ARR目标，以及随后一年6000亿美元的目标，在用量上需要达到什么水平？

What would it take in usage for Anthropic to hit the $200,000,000,000 in ARR plan for 2028 and then 600,000,000,000 in in the next year?

简单点说，就是看全球有多少知识型工作者，对吧？然后看有多少人能接受订阅模式。

The simple version is how many knowledge workers are in the world, alright, how many folks can take a subscription.

就算乐观估计，大概有十亿人吧？

Being generous, is it a billion human beings, right?

假设 Anthropic 拿下百分之百的市场份额，单价两百美元，那总收入就是两千亿。

You know, if Anthropic has 100% market share at $200, that's 200,000,000,000.

要是 Anthropic 能拿到百分之三百的市场份额（注：此处原文可能为口误或夸张说法，意指远超100%的潜在价值或包含其他维度），那就是六千亿了。

If Anthropic has 300% market share, that's 600,000,000,000.

我也不太确定，罗里可能想得更多一些。

I don't know, Rory thought more.

六千亿这个数字看起来挺复杂的，但其实我们对人工智能的需求才刚刚开始。

The 600,000,000,000 seems complicated, but our demand for AI has only just begun.

两百亿这个数字还是能想象得出来的，对吧？

You can see 200,000,000,000, which is a number, right?

一旦涉及到六千亿这个量级，那就真的很难把握了，因为很少有人会去关注这么大的数字。我最近也在这方面做了一些研究，你懂的。

Once you start getting the 600,000,000,000 number gets really hard because no one ever looks at the big numbers, and I've just been doing some work on this, right?

从来没人停下来看看那些大数字。

No one ever steps back and looks at the big numbers.

你说全球有十亿知识工作者，这纯属扯淡。

You said a billion knowledge workers in the world, absolute bollocks.

说句实在话，美国通常占据全球软件预算的50%，因为我们拥有全球一半的高端知识工作者。

Hard nose comment here, US is typically 50% of the world's software budget because we're 50% of the world's high end knowledge workers.

我们的GDP占全球的25%。

We're 25% of the world's GDP.

所以，如果支出至少跟GDP挂钩，那也就是美国的4倍。

So at a minimum of spend tracks GDP, it's only 4x The US.

但每家软件公司通常都是美国的两倍。

But every software company is typically 2x US.

为什么？

Why?

因为世界其他地方买不起我们用的软件，他们更穷，人多但工资低，软件也少。

Because the rest of the world can't afford the same software we do because they're poorer and they have more people at lower wages and less software.

这就是为什么我们去欧洲时网络那么差。

That's why we have crappy Internet when we go to Europe.

所以，说实话，最直白的看法是，你可能要把美国知识工作者的支出翻一倍。

So the truth is, the hard nosed comment is this, you probably take The US knowledge worker spend and double it.

美国有8300万知识工作者，还有大约8600万体力劳动者。

There are 83,000,000 knowledge workers in The US, and then roughly 86 physical labor workers.

这就是你的起点。

So that's what you start with.

然后你开始逐步缩减，上周末莉莉还在算这笔账呢。

And you start cutting it down, and I've Lily was doing the math this weekend thinking about it.

你会发现缩减的速度非常快。

You start cutting it down real quickly.

事实上，知识工作者这个概念涵盖了医疗行业的所有人。

The truth is, you know, knowledge workers includes everyone in health care.

我觉得我们不可能取代护士的工作。

I don't think we're gonna you know, we're not gonna replace the nurses.

这里面包括老师。

It includes the teachers.

对吧？

Right?

最理想的情况是这样的：美国大概有180万人从事编程工作，再加上质量保证和其他相关岗位，总共有大约500万人从事软件相关的活儿，比如系统管理之类的。他们每年的总收入加起来大约是6000亿美元。

The sweet spot, the sweet spot of the whole damn thing is there are about 1,800,000 people doing coding in The US, and then including then QA and all the other there's around 5,000,000 people that do software related shit, systems admin stuff, all the rest of that, and they get paid in total grossing up about 600,000,000,000 a year.

2000亿美元意味着你要替换掉其中三分之一的人。

200,000,000,000 means you're replacing a third of them.

那可真不少。

That's a lot.

我们之前说过，最重要的一个比例——我问过你觉得是什么，Jason——就是在稳定状态下，工资支出和AI支出的比例是多少？

Remember we said this before, the single most important ratio, and I asked you about what you thought it was, Jason, is in a steady state, what's the ratio of salary dollars to AI dollars?

因为如果是工资的50%，你很容易就能达到2000亿，但6000亿在编程领域很难，你达不到那个数。

Because if it's 50% of salary dollars, you can easily get to $200,000,000,000 $600,000,000,000 is hard in coding, where you can't get there.

如果是10%，那整个行业要达到2000亿也很难。

If it's 10%, then it's hard to get $200,000,000,000 across the whole thing.

所以归根结底，在稳定状态下，软件能占多大比例，因为如果软件是采用率最高的尖刀，你觉得每花10万美元请一个工程师，或者20万美元请一个工程师，你会花10万在AI上，5万在AI上，还是20万在AI上？

So it really boils down to, in the steady state, how much software because if software is the tip of the spear in terms of max adoption, what do you think for every $100,000 you spend on an engineer, $200,000 you spend on an engineer, are you going to be spending 100 ks on AI, 50 ks on AI, or 200 ks on AI?

这就是关键数字。

That's the number.

对，我们正在测试。

Yeah, we're testing it.

过去六十天，每一家扩张中的公司都在真的削减AI预算。

The last sixty days are every single scale up is capping their AI budget for real.

不只是优步这样的公司。

It's not just Ubers of the world.

大家都在限制预算，因为增长确实是指数级的，对吧？

Everyone's capping it because it's grown truly exponentially, right?

大家都把预算控制在每年600万、800万，对吧？

Everyone's capping it at 6,000,000 a year, 8,000,000 a year, right?

我觉得最终会落在每个工程师相当于100美元的水平。

I think it'll land at $100 per engineer equivalent.

我觉得我们会给每个顶尖工程师10万美元的代币，然后相应地，我们会把开发团队的规模缩减40%。

I think that's what we'll I think we'll give each of our best engineers $100,000 of tokens, and in return, we'll cut the size of our dev teams 40% effectively.

实际执行起来可能不会完全这样，但大致就是这个方向。

It won't exactly work out that way, but close enough is how it's going to work out.

所以这里100美元是用来运行推理的，24小时不间断，同时跑10个代理。

So there's $100 here for running inference 20 fourseven with 10 agents in parallel.

顺便说一句，我其实也同意这个看法，这和我心里的模型差不多，而且这指向一个总数，假设不只是开发引擎。

For what it's worth, I actually agree with that's that's about with my mental model too, and that points to a total and let's assume it's not just dev engine.

咱们也把系统管理员、QA人员算上。

Let's give the sysadmins, the QA guy.

对所有人都做同样的事情。

Let's do the same thing for everybody.

对所有人都一样。

Same thing for everybody.

你拿200美元的工资，包括所有福利，再加上100美元的AI，但我们裁掉30%的人。

You get $200 of wages, fully loaded, including all the benefits, and $100 worth of AI, but we cut 30% of you.

这算下来在美国大概是一个2000亿美元左右的市场，挺吓人的。

That turns out to be terrifyingly about a 200,000,000,000 plus or minus market in The US.

明年吗？

Next year?

不，是永远。

No, ever.

嗯，那是Anthropic对明年的估计。

Well, mean, that's Anthropic's estimate for next year.

我知道，我的意思是，如果把所有人数都算上，再套用Jason Matt的算法，美国这边大概能到两千亿，那全球估计得挣扎着冲到三千五百亿。

I know, my point is this, if you count all the heads and apply the Jason Matt, you get 200,000,000,000 in The US, which probably means you struggle to get $350,000,000,000 worldwide.

这是英镑换算过来的数。

That's the pound.

然后你还得把范围扩大到软件之外。

And then you've got to go beyond software.

显然软件之外也有收入来源，但远不如软件领域那么肥沃。

And there is obviously revenue beyond software, but it's nowhere near as fertile.

而且这个比例不会那么高。

And the percentage isn't going be anywhere near as high.

对。

Yeah.

不过挺有意思的，我算出来的数字正好吻合。你看那些增长数据，样本里最顶尖的那1%，也就是明显偏向科技前沿人群的那部分人，每月花7000刀，而中位数才花100刀左右。

But it's funny, it's exactly the number I come out with because you see the ramp data that says, the top 1% of their sample, which in turn obviously is a biased sample of tech forward people are spending 7 ks, and then the median is spending like 100.

这差异也太大了。

It's amazing the dispersion.

七千乘以十二，一年也就八万四。

And seven ks times 12 is only 84 ks.

所以，就算是最顶尖的那1%，在科技投入上最舍得花钱的那群人，人均也就花这么多，而且这还是所有员工里的1%。

So the top 1% of the most curated group you can imagine in terms of tech spend is spending, and that's 1% for all employees.

所以，最乐观的支出上限，也就是工资的50%。

So that's the pointy edge of the most optimistic spend, is $0.50 of salary dollar.

我觉得我投的那些公司里，最好的那些，增长最快的，尤其是2022、2023年之前成立的那些，能达到10万美元。

I think we're going to get to $100,000 in the investments I've made that are the best ones, the ones growing faster, especially ones that are pre-twenty twenty two, '23.

他们有一个参照系。

So they have a frame of reference.

他们的交付速度确实快了两到三倍。

They literally are shipping two to three times faster.

只是最近才这样。

Only recently.

只是最近才这样。

Only recently.

去年那说法有点扯，对吧？

That was kind of bullshit last year, right?

人们会那么说。

People would say that.

但那都是装样子，像是在刷存在感。

But it was all performative like token maxing.

上周我开了两次董事会，他们都已经把今年的路线图定完了。

I had two board meetings in the last week where they finished the roadmap for the year.

都已经排到2027年了。

They're into 2027.

这些是我增长最快的两家公司，但不是那种刚起步的新公司。

These are my fastest growing, two fastest growing, but not brand new companies.

他们路线图都定完了。

They finished the roadmap.

2027年的路线图都已经规划得很深入了。

They're well into the 2027 roadmap.

所以你会花一百块给团队做这件事，但加起来就是好几百万。

So you're going to spend $100 on your team to do that, but it's adding up to so many millions.

这让人压力山大。

It's overwhelming.

所以我真觉得这个十万的额度挺合理的。

So I do really think there's some this 100,000 makes a lot of sense.

你可以解释一下为什么需要更多人，或者大家会逐步增加预算，但我认为这会成常态，你们会给团队设上限，毕竟这是CFO们会做的事。

You could justify more or less people will ratchet it, but I think it'll be it'll be the new normal and you'll cap your team and it'll all be that's just what the CFOs will do.

一百亿美元的推理算力，你就能雇这么多工程师。

A $100 of inference and you get to hire this many engineers.

但说他们真的把2027年的路线图落实了，这可不是做样子，也不光是公关手段。

But the idea that they're pulling their twenty twenty seven roadmaps in, it's not just performative, it's not just PRs.

你肯定想往里面投钱，直到达到它能发挥最大效用的极限，对吧？

You want to invest in that up until the maximum where it works, right?

不过那些绝对数值确实变得非常惊人了。

But the absolute numbers are just getting really big.

是的。

Yes.

我得说，兄弟，如果你不是这样，你今天就在输。

I got to say, man, if you're not that way, you're losing today.

如果你在2026年8月还没有深入规划2027年的路线图，在智能体时代，你的团队就不足以应对今天的竞争。

If no, if you're not into your 2027 roadmap, deep into it by August 2026 in the agentic world, your team is not good enough to survive today.

这是你最后的机会去做出改变。

This is your last chance to make changes.

听着，如果你是OpenRouter，你甚至没有2027年的路线图。

Listen, if you're if you're OpenRouter, you don't even have a '27 roadmap.

这甚至不重要，因为你每天都在重新制定计划。

It didn't even mean it because you're just remaking it day by day.

但如果你还在用老一套的方法，每个季度完成这么多，每个月完成这么多，而没有规划到2027年，你会输给竞争对手。

But if you're running the classic playbooks of these I can get this much done each quarter, this much done each month, and you're not into 2027, you're going to lose to the competition.

你必须诚实面对。

You've to be honest.

你们对2027年的规划深入到了什么程度？

How deep into '27 are you?

还不够深。

Not deep enough.

是的，Jason总是抛出一些让人后背发凉的言论，让我回去后反复琢磨。

Yeah, Jason always gives me these terrifying sound bites that I go back and think about.

因为我们做过这项调查，我们这类人比较讲究事实依据，所以我们调研了所有公司，发现情况跟之前提到的分布类似：有些公司全力投入，有些公司在采用，但人均支出还是明显偏低。

Because we did this survey, we tend to be fact based people, we did the survey of all our companies, and we saw similar to the ramp dispersion, which is some companies all in, some companies adopting but still dramatically less spend per head.

我记得平均值，但当时低得多。

I remember the average, but it was dramatically less.

我过去没做的一件事，现在回想起来其实应该做、也会去做，就是回去梳理一下：如果你能建立起产出与支出合理性之间的强相关性，那你就要对那些落后的团队说，你们只会越来越落后。

What I didn't do, and actually now that I think about it, I should do and I will do, is go back and say, if you could touch a strong correlation, which you believe you should be able to, between output and can you justify the spend?

对于那些拖后腿的，你直接告诉他们，你们迟早会被甩在后面。

Then you should be saying to the laggards, you're just gonna fall behind.

如果这个数字涨到20亿到25亿之间，你会考虑买吗？

If it goes out at 2 to 2,500,000,000,000, would you be a buyer?

首先，我想说清楚。

First of I wanna be clear.

我不认为软件市场本质上已经走到了尽头。

I don't think the software market is definitionally the end of the time.

我觉得普通知识工作者不会把工资的50%花在这上面，但他们会花一个相当可观的比例。

I think the average knowledge worker won't have 50% of salary and think, but they'll have a meaningful percentage.

所以，总可寻址市场（TAM）远不止开发者，因为还有律师，但我觉得律师不会花那么多。

So the TAM is significantly bigger than just developers because you have lawyers, but I think lawyers won't look.

那个年收入200万美元的合伙人级别的知识工作者，不会花掉20万美元的token费用。

The K and E guy who's pulling 2,000,000 a year as a partner isn't gonna be doing 200 k's worth of tokens.

他们肯定不会花掉100万美元的token费用。

It's definitely not going be doing $1,000,000 worth of tokens.

律师宁愿死也不会用100万美元的token来替代100万美元的实拿工资。

A lawyer would die before they give $1,000,000 of tokens instead of $1,000,000 of take home pay.

所以，这个市场比软件市场大，但没有任何地方能像软件这样处于最佳位置。

So the market is bigger than software, but there's nowhere else that's such a sweet spot as software.

所以我不想把自己限制在200。

So I don't want to be limited to 200.

但我觉得真正棘手的是，我肯定想成为第一个上市的公司，而不是第二个，特别是如果你有接近盈利或者盈利波动的情况。

But I think the really challenging thing, I definitely want to be first out rather than second out in terms of going public, especially if you have some kind of near profitability story or bouncing around profitability.

我认为，作为Anthropic在秋天上市，带着“嘿，我们曾经盈利过，好吧，我们现在又不盈利了，但我们是企业市场的赢家”这样的故事，远比明年上市更有吸引力。到那时，Anthropic和公开市场的增长率可能都开始放缓，如果OpenAI再试图进入市场，那他们的处境肯定会更加艰难。

I think it's a far more attractive strategic position to be going out as Anthropic in the fall with, hey, we've been profitable, okay, we're unprofitable again, but we're the winner in the enterprise, then going out next year where maybe the growth rates have started to slow both for Anthropic and the public markets, And if you open AI trying to access the markets, then I definitely think they're in a strategically more challenging situation.

我觉得他们只是认命了。

I think they've just capitulated to it.

当然，你说得对，谁都想当第一个。

Of course, you want to be first to your point, right?

我认为OpenAI不得不整顿内部，高管层动荡不断，不过最近三十天似乎表现不错，对吧？

I think OpenAI has had to get their house together, more executive turmoil, Apparently a great last thirty days, right?

但上半年，他们的速度比之前的小竞争对手还慢，对吧？

But first half of the year, slower than its previously junior competitor, right?

他们不得不做很多事来表明：听着，我们会第二个上市，然后我们会有个对比，对比就是这样。

And they've had to do so much to say, listen, we're going to go public second, and then we're going to have a comp out there, and the comp is what it is.

而且我们可能不会像SpaceX和Anthropic那样有那么多炒作，世界也不会因此终结。

And we may not trade with the hype that SpaceX and Anthropic did and the world will not end.

就是说，我们会以一个非常精确的数字上市。

Like, we will trade at a very precise number.

我们会知道我们以什么价格上市。

We will know what we're going to go out at it.

如果我们以1.3万亿美元的估值上市，世界也不会完蛋。

And the world will not end if we trade at 1,300,000,000,000.

我觉得他们已经不再担心这个了，因为归根结底，Rory说得对，先上市确实更好，但从长远来看，这并不重要。

I just think that they've given up on worrying about that because ultimately, Rory's right, it's much better to be first, but in the long run, it doesn't matter.

对吧？

Right?

如果你不需要资金，那就顺其自然吧。

You're just if you don't need the capital, it just is what it is.

但是，Jason，这就是关键所在。

But, Jason, that's the sentence.

这就是我的观点。

This my point.

但再深入一点。

But push a little.

这就是关键所在。

That is the sentence.

地球上没有哪两家公司比这两家更需要资金了。

There are no two companies on the planet that need more capital than these guys.

但在你需要资金的世界里，当老二是很痛苦的。

In a world where you do need the capital being second sucks.

我同意，从整体上来说，你说得对。

Because I agree, in general, you are correct.

但这其实无所谓。

It doesn't matter.

两家公司上市时间相差一年左右，十年后大家根本不在乎谁先谁后了。

It's two companies go public plus or minus a year, and a decade later, no one cares.

这些年来我们确实见过这种情况。

We've definitely seen that over the years.

这个案例的难点在于，两家公司仍然有巨大的资本需求，动辄数十亿甚至上百亿美元。

The thing that's challenging in this particular case is both companies still have enormous, many 100,000,000 billion dollar capital needs.

在这种情况下，我更愿意……我是在问你，这有点不公平。

In that situation, I would much prefer to be I do I'm putting in ask you to unfair there.

我觉得你说得对。

I think you're right.

但问题是，比如说，你随便选个数字。

But the thing is, let's say and pick your number.

假设Anthropic以2万亿美元的估值上市，其实这真的不重要。

Let's say Anthropic is public at $2,000,000,000,000 it really doesn't matter.

OpenAI在上市前，就能以低于其隐含估值的价格出售股票。

OpenAI is going to be able to sell stock at a discount to its implicit valuation before it goes public.

资本仍然足够。

There's still enough capital.

假设它们两家公司的隐含估值都是两万亿美元，对吧。

Let's say they're both worth $2,000,000,000,000 right, implicitly.

OpenAI明年就能以1.8万亿的价格卖股票。

OpenAI is going to be able to sell stock next year at 1.8.

大家都会买的。

People will do it.

尤其是如果你作为CEO在自己公司里没有股份，那打个折卖也无所谓。

And especially if you have no stock as CEO in your own company, it's okay to sell at a small discount.

对，同意。

Yes, agree.

而且你看，我不是在危言耸听，但我觉得有意思的是，如果你是市值较小的那家公司，但你的资本需求更大——现在OpenAI就是这样，因为他们的资本需求目标更激进。

And look, I'm not catastrophizing here, but I think the interesting thing is if you're the smaller market cap company and you have the bigger capital need, which right now OpenAI does because they have a more ambitious capital need target.

那你更愿意当那个市值1.5万亿、只需要融资1000亿的公司，还是那个市值1万亿、却需要融资3000亿的公司？

Now would you prefer to be the guy trading at 1,500,000,000,000 who only needs to raise 100,000,000,000 or the guy trading at 1,000,000,000,000 who needs to raise 300,000,000,000,000?

到某个时候，这些问题就会变得棘手。

At some point, these things become troubling.

对，价格能解决所有市场问题。

And yeah, price clears all markets.

这可能是我们见过的有史以来最好的新技术市场。

This is the best new technology market we've seen ever, perhaps.

如果你是这个市场的创始人，即使现在排名第二，你也会吸引资本，只是你不知道具体条件是什么。

And if you are the founder in that market and even now the number two, you're going to be at track capital, but you just don't know the terms under which it happens.

回到我刚才说的，我觉得你会后悔今年没能进入资本市场。

And going back to my comment, I think you will regret not being able to access the capital markets this year.

不不，我当然同意。

No, no, of course I agree.

我也不想花太多时间在这上面。

And I don't want to spend too much.

我只是想说，媒体和社交媒体把这件事炒得很大，对吧？

My only point is the media and social media make a big deal out of this, right?

谁先上市，谁表现更好。

Who goes public first and who does better?

我觉得山姆和OpenAI团队已经说了，这就是我们的命运。

I just think Sam and the OpenAI team have said this is our fate.

他们明天就能上市，对吧？

They could go public tomorrow, right?

有足够多的人愿意买这些股票来支持上市。

There is enough people to buy these shares to go public.

他们决定，虽然这不完美，但这是目前最好的选择，我们会接受这种不确定性，这又不是世界末日。

They've decided that while this isn't perfect, right, this is the best on the board and we're gonna live with the doubt, like it's not the end of the world.

比如，不可能今晚就解决所有问题。

Like, can't solve every problem tonight.

他们得解决更大的问题，对吧？

They gotta solve bigger problems, right?

牌已经这样了，对吧？

The cards are the way they are, Right?

杰森，你提到了管理团队的人员流动问题。

Jason, you said about management team churn.

最近离职的，可能有些人不知道，是首席营收官Denise Dressor，接替她的是Dali Rajic。

The churn for those that don't know most recently was Denise Dressor who was the CRO who left, and Dali Rajic has replaced her.

可能有些人不知道，Dali是最受尊敬的首席营收官之一。

For those that don't know, Dali, he's one of the most respected CROs.

他在Whiz的时候简直是个天才。

He was a freaking master at whiz.

而且我觉得业内最厉害的首席营收官或者销售负责人就是Chad Petes，他真的是顶尖中的顶尖。

And I think the best CRO or sales leader in the business is Chad Petes, says he's the best of the best.

所以今天我对他们的Codex和企业部门更有信心了。

So I'm feeling a little bit more confident for their codex and enterprise division today.

对。

Yeah.

只是变化太多了。

It's just a lot of change.

内部情况我也不太清楚。

I don't know anything inside.

我觉得就是格雷格·布罗克曼接手了，对吧，然后把那个搞销售的高手请来了。

I just think Greg Brockman took over, right, and brought in the whiz guy.

就是受够了Salesforce那套东西。

Just had enough of the Salesforce crap.

对不对？

Right or wrong?

其实，你要是看看整个AI行业，好多Salesforce的高管都被挖过来了，对吧，来帮忙。

Actually, if you look across all of AI, a ton of Salesforce executives have been recruited, right, to come in and help.

你可以笑话这事儿。

And you can make fun of it.

就像我以前也笑话过，当年Salesforce招Oracle高管，因为他们老拿Oracle开涮。

Like, I used to make fun of how back in the day Salesforce hired Oracle executives because they took shots at Oracle.

但你需要有人懂怎么扩大规模。

But you need folks to know how to scale.

Salesforce现在市值多少？

What is Salesforce at?

年营收跑到了450亿，还是500亿？

45,000,000,000 run rate, 50,000,000,000 run rate?

我的意思是，Anthropic现在都已经超过这个数了，对吧？

I mean, Anthropic's past that now, right?

OpenAI也早就超过了。

OpenAI is past that.

所以你不能光招年轻人。

So you don't want to hire kids.

你得找个懂怎么打比赛的人。

You want to hire someone that has some idea how to play.

所以 Salesforce 也就那样，对吧？

So Salesforce is about it, right?

但退一步说，我宁愿要一个来自 Wiz、贴近技术、身处高度竞争环境的人，而不是整天问你要买多少 Slack 席位。

But if you step back from it, I'd rather have someone from Wiz that is close to technology, that is in a hypercompetitive space rather than asking how many seats of Slack you want.

这完全是两种不同的市场打法。

It's just a very different go to market motion.

对吧？

Right?

确实很不一样。

It's very different.

杰森，你说过，如果产品还没达到年终目标，而且还没到2027年就已经落后了。

Jason, you said if you have not already hit your end of year goal in terms of products and you're not well into 2027, you're behind.

我是在做一些假设。

I'm making assumptions.

我不觉得Workday已经像你们两家公司那样，在2027年就已经达到那些目标了。

I don't imagine Workday's quite at the cutting edge like two of your companies at 2027 already hitting those goals.

而且银湖资本对Workday提出了430亿美元的私有化收购要约，这是有史以来最大的SaaS收购之一。

And Silver Lake circles a $43,000,000,000 take private bid for Workday, one of the biggest SaaS buyouts ever.

我们这里有两位业界最顶尖的SaaS专家。

We've got two of the best SaaS minds in the business here.

两位，你们怎么看这件事？

Guys, what should we take from this?

SaaS并没有死，这是最大的公司之一，也是最大的收购之一。

SaaS isn't dead, one of the biggest firms, one of the biggest buyouts.

消息公布后，股价涨了18%。

The stock popped 18% afterwards.

哇。

Wow.

我觉得，光说SaaS没死这个结论太简单了。我觉得更值得关注的是，一个非常懂财务、极其精明的买家愿意下注，以一个有约束的价格买下这家公司，然后加杠杆，最终获得回报，因为它的收入足够稳定，能在五年内还清债务，而且倍数保持合理稳定，再转手卖掉，赚个20%左右的内部收益率。

I think what you can take from because the SaaS isn't dead thing is just too simplistic, I think what you can take is a very financially oriented, wildly savvy buyer is willing to bet money that they can buy this at a constrained price, lever it, and generate a return because the revenues are sticky enough to allow them to pay down the debt over five years and with reasonable multiple stability, sell it on and make a 20% IRR plus or minus.

我算过这笔账。

I ran the numbers.

这就是他们的赌注。

That's the bet.

所以它没死，但也不是什么特别让人兴奋的事。

So it's not dead, but what it is not is wildly exciting.

这其实说明，这个行业已经进入成熟期了。

What it says is, Sat, this is the mature phase of an industry.

这不再是关于疯狂增长，甚至也不是关于无节制的增长。

When it's not about wild growth, it's not even about untempered growth.

实际上就是有人说，这个东西年增长率是13%。

It's literally about someone saying, this thing is growing at 13% year on year.

我们可以以大约5倍营收、16倍过去十二个月的EBITDA的价格买下它。

We can buy this thing for roughly 5 times revenues, 16 times trailing EBITDA.

我们可能会用两到三倍或四到五倍的EBITDA来加杠杆，但这会是一笔很大的股权投入。

We'll probably leverage it two or three times and four or five times in EBITDA, but it's going be a big equity check.

然后你运行LBO模型，你说，保持35%的营业利润率五年，用掉所有现金，这大约是每年100亿美元的收入。

And then you run the LBO model and you say, you keep at a 35% operating margins for five years, you use all that cash, it's roughly $10,000,000,000 a year in revenue.

所以每年大约有30亿美元的现金，你用来还债和利息，只要买得合适，你可以在四五年内赚到20%的回报，差不多翻倍。

So it's like $3,000,000,000 a year of cash, you pay down the debt and the interest, and provided you buy right, you can make 20% and almost a 2X over four or five years.

我看这笔交易，我有点纠结，因为首先我觉得Silver Lake非常聪明。

I look at that deal and I go, I'm torn because first of I think Silver Lake are wildly smart.

当你运行敏感性分析时，这非常有趣。

It's very interesting when you run the sensitivities.

如果你多付了20%，回报率就会降到百分之十几。

If you pay 20% too much, it dips down into the mid teens.

这几乎和风险投资完全相反。

It's almost the exact opposite of venture.

在风投交易中，如果你投对了项目，你付多少钱几乎不重要。

In venture deals, if you're in the right thing, it almost doesn't matter what you paid.

具体细节可以看Cursor，也可以看OpenRouter。

You see Cursor for details, see OpenRouter for details.

这恰恰相反。

This is the exact opposite.

这是精细的金融工程。

This is fine, precise financial engineering.

如果你在价格上错了20%或30%，你的内部收益率就会从20%——这个规模下完全可以接受——降到百分之十几，那样的话你就会后悔做了这笔交易。

If you're wrong by 20%, 30% on price, you know, your IRR dips from 20, which is totally acceptable at scale, to low teens, in which case you wish you hadn't done the deal.

我能问个问题吗？

Can I ask a question?

对于四到六年的持有期来说，这确实是精细的金融工程。

Precise financial engineering for a four to six year hold period.

六年前，Chatuchi BT还不存在。

Six years ago, Chatuchi BT didn't exist.

在这样一个变化飞快的世界里，你还能做这种精细的金融工程吗？

Are you able to do fine precise financial engineering in a world where we move so fast?

我觉得系统记录本身不是关键，它可能是个护城河，但并不是增长的通行证。

I don't think system of record is I think it's a moat, but I don't think it's a ticket to growth.

这一点我觉得非常重要，而且X上的每个人都搞错了。

This is, I think, super important, and it's something that everyone on X gets wrong.

拥有一个像Workday那样的系统记录当然很好。

It's great to have a system of record, which Workday has.

即使有AI和语言模型的帮助。

Even with AI and LMs help.

客户很难流失，或者说你根本不想流失。

It's very hard to churn or you just don't want to churn.

但这绝对不意味着我想在那个供应商身上花更多的钱。

But it sure as hell doesn't mean I want to spend more money with that vendor.

那是他们面临的挑战。

That's their challenge.

但这绝对意味着那五年比95%的情况要可预测得多，比我们喜欢的可怜的Monday或其他公司要可预测得多。

But it sure as hell means the five years are far more predictable than 95%, than poor Monday, which we love, or others.

我们完全不知道Monday甚至HubSpot五年后在中小企业市场会是什么样子。

We have no idea where Monday or even HubSpot will be in five years at the SMB level.

我们基本知道Workday会是什么样子。

We know pretty much where Workday's going to be.

对吧？

10, right?

所以我觉得这种增长与留存的关系被误解了。

And so I think this growth versus retention is misunderstood.

这笔交易里还有一点上升空间，我不知道这是不是Silver Lake考虑的一部分。

There is a little bit of upside in this deal, which may I don't know if it's part of Silver Lake's calculation.

CEO回来了。

The CEO came back.

其中一位联合创始人回来了。

One of the co founders came back.

阿尼尔也回来了。

Aneel came back.

他回来了。

He came back.

他在AI时代来临之前，趁着形势好的时候，雇了他的继任者。

He hired his successor when times were easy just before AI.

他找了一位非常擅长细节运营的联合CEO，就像我们在UiPath的朋友丹尼尔那样，然后他意识到，如果今天去上班，就回来了。

He brought in a great knobs and dials co CEO, you know, like our friends at Daniel at UiPath and others and realized if I go to work today, came back.

所以我觉得银湖资本并不打算对这笔交易进行彻底改变。

So I don't think Silver Lake is planning on on a deal like radically changing it.

但如果他真这么做了，那确实有上升空间。

But I think if he does, there's real upside to that.

也许不是他们20%的内部收益率，而是可能带来颠覆性的改变。

Maybe instead of their 20% IRR, it could be a game changer.

如果他打造出Workday的智能体版本，至少创始人重新掌舵来推动这件事。

If he creates the agentic version of Workday, they at least have the founder back in the saddle doing it.

如果我是银湖资本，我会觉得更有底气，因为还有上升空间。

And that would make me feel a lot better if I were Silver Lake that I have upside.

但那绝不会出现在基础假设里。

But it wouldn't be in the damn base case.

杰森，你对基础假设的界定完全正确。

Jason, you framed the base case exactly correctly.

也就是5.3倍乘以20，换句话说，这意味着财务软件公司愿意为年增长13%的系统记录软件支付5倍营收的价格。

It's 5.3 times 20 In other words, what this says is financial minds will pay five times revenues for system of record growing at 13%.

任何不是系统记录、增长没那么快的，价格就得相应调整。

Anything that's not a system of record, anything that's not growing as fast, price accordingly.

因为你说得对，你不可能把同样的杠杆应用到一个待办事项、任务管理、项目管理或网站建设软件上。

Because you're right, there's no way you'd apply the same kind of leverage to, for example, a To Do or a task management or a project management software or a website building software.

这让你对顶级LBO收购的基准水平有了概念。

What this gives you is a sense of what the baseline is for best in class LBO takeouts.

如果像Airtable这种弯曲勺子的情况让你明白没有那种系统记录会怎样，那你只能拿到2.7倍。

If the air table bending spoons give you an idea of what it is if you don't have that kind of system of record, you get 2.7.

如果你勉强盈利，而且像Jason说的，在一个你无法预测五年后的领域，那你只能拿到2.7倍。

If you're vaguely profitable, and in a space where, as Jason says, you can't predict five years, you get 2.7.

而Workday的例子说明，如果你有30%的营业利润率，增长虽然一般，但你有系统记录，能让你真正相信未来五年，那运气好的话，你能拿到5.3倍的营收。

And what Workday says is if you've got 30% operating margins, modest growth, but you have a system of record where you really can believe in the next five years, then if you're lucky, you get 5.3 times revenues.

这就是现在的买卖价差。

That's the bid ask spread right now.

对比一下OpenRouter的情况，他们可能会拿到，我觉得，过去十二个月的营收大概在正负100左右。

Contrast that with the game for OpenRouter where they're going to get, I think a trailing revenue is plus or minus 100.

他们会拿到过去营收的70倍。

They're going get 70 times trailing revenues.

你更想玩哪个游戏？

Which game would you prefer to play?

我觉得Workday对PE来说比其他目标公司更有吸引力，因为它某种程度上是个封闭的系统记录。

Workday has something that makes it a better deal for PE, I think, than anybody else on the target list, which is that it is a somewhat closed system of record.

现在Salesforce在外面拼命干活，因为他们是个非常开放的平台。

Now Salesforce is out there working their frigging tails off because they are a muchly open platform.

你明天就可以在Salesforce上面构建自己的智能体。

You can build your own agents on top of Salesforce tomorrow.

很多热门的GTM初创公司都是建立在Salesforce之上的。

A lot of the hot GTM startups are built on top of Salesforce.

他们不一定只用Salesforce，但它是开放的。

They're not necessarily only on Salesforce, but it's open.

试试在Workday上开发吧。

Try building on Workday.

那可不容易，对吧？

It ain't so easy, right?

就像LinkedIn一样，对吧？

It is like LinkedIn, right?

它故意几乎不开放。

It is intentionally barely open.

所以这也有不利的一面，对吧？

So there are negatives to that, right?

但这也意味着你会在整个生态系统中捕获更多的预算，比其他系统要多。

But it also means you're going to capture more budget overall in your ecosystem than you would for others.

所以它比开放生态系统更能抵御智能体对你增长的冲击，对吧？

So it has more of a buffer against agentic damage to your growth than an open ecosystem has, Right?

开放系统现在也有弊端。

Open has negatives today.

所以我想要的是记录系统，不可流失，完全封闭。

And so I would want system of record, churn impossible, closed AF.

我想要最封闭、不可能流失的系统，因为记录系统之所以不那么好，是因为你需要你的系统记录。

I want the most closed system that can't churn because the reason system of records aren't that great is because you need your system record.

但如果你稍微开放一点，而且你自己或第三方能做出更好的智能体，那么价值就会流向智能体，即使记录系统还在。

But if you're remotely open and you can produce a better agent yourself or a third party, the value will extract to the agent even if the system of record is retained.

但是Workday太封闭了，所以他们占了优势，对吧？

But Workday is so closed, they've got a leg up, right?

Jason，Salesforce的开放程度怎么样？

Jason, how open is Salesforce?

他们就像Shopify一样是个收费的守门人，但他们相当开放。

They are a toll keeper like a Shopify, but they're pretty open.

Shopify和Salesforce都挺开放的。

Shopify and Salesforce are pretty open.

我们三个人，明天就能用OAuth发布一个Salesforce应用。

The three of us, we can use OAuth to ship a Salesforce app tomorrow.

为了证明这一点，Harry，很快说一下，你看，有很多公司，甚至在LLM出现之前，比如Gong、Outreach、SalesLoft，它们实际上都是建立在Salesforce平台上的。

Just to prove that, Harry, really quickly, that, look, there's a bunch of companies, even in the pre LLM world like Gong, Outreach, SalesLoft, that are all effectively built on top of the Salesforce platform.

在Workday上，你很难轻易说出类似的公司。

You can't name the equivalent with any ease in Workday.

虽然有一些，但比某些规划工具难多了，不过基本上大多数都是这样，这也说得通。

There's a few, but it's much harder than some of the planning tools, but pretty much most and it makes sense.

在财务核算系统里，所有数据最终都会被总账和会计系统这股引力吸进去。

Within the financial accounting system, everything gets sucked into the gravitational pull that is the GL and the accounting system.

所以我同意。

So I agree.

说得好，Jason。

It's a good point, Jason.

不管这个生态里有多少资金，只要策略得当、眼光独到，Workday 基本上能拿下大头。

Whatever dollars are in that ecosystem, if they're careful and shrewd, Workday will get most of them.

另一方面，如果他们太贪心，投入不够，客户就会觉得，天哪，这产品都没什么进展了。

And on the other hand, if they get too greedy and they don't invest enough, then the customers start thinking, oh my god, this is just not advancing.

过了五年，也许我确实需要更高级的代理式工作流。

Over five years, maybe I do need more of this agentic workflow on top.

也许小客户开始评估Natsuite，看看下一代的产品。

Maybe the smaller customers start evaluating Natsuite, start evaluating the next generation.

哪怕是最小的客户群体。

Even at the very small end.

有Willis这样的，有campfire这样的，还有像他们那样的人。

You've got the Willis, you've got the campfires, you've got the people like that.

你不能在自己的生态里贪得无厌，逼得别人开始想往外搬。

You can't be such a greedy bastard in your ecosystem that you incent people to start trying to move out.

但Silver Lake很聪明，Neil也很聪明。

But Silver Lake are smart, and Neil's smart.

你完全可以把它做成一个盈利的、自给自足的体系。

You could have this be a profitable, self contained universe.

但记住，最激动人心的版本是五年内还清所有债务，然后以20%的回报翻倍你的钱。

But remember, the most exciting version of that is you pay down all the debt in five years and you double your money for 20%.

这已经是现在最好的情况了。

That's as good as it gets now.

那可是很大一笔钱。

It's on a lot of money.

你可能要投入大约200亿到300亿美元的股权资金，因为你不可能无限举债。

You're probably putting in plus or minus a 20,000,000,000 to $30,000,000,000 equity check because you're not going to get infinite debt.