20VC：Anthropic 提交上市申请 | 企业界陷入代币预算恐慌 | Cognition 融资 10 亿美元，估值达 260 亿美元 | 阿波罗警告私募股权软件回报将灾难性 | 996 工作伦理：作秀还是创业现实？

确实如此，对吧？

Isn't it, though?

说实话，我也没参与过几次这类讨论。

Listen, I don't claim to be I've only been in the conversations a few times.

我可不敢说自己是专家。

I don't claim to be expert.

当然，这很明智。

Of course, it's smart.

从风投或初创公司的角度看，这种稀释程度根本没法预测。

The dilution is as viewed from, like, a venture perspective or start perspective is unknowable.

我的意思是，这根本没法衡量。

I mean, it's it's tie able.

完全看不出来。

It's tie it's unseeable.

但如果他们真觉得回本这么快，为什么不直接发债，避免股权稀释呢？

But if they really thought that payback was so quickly, wouldn't you issue debt and have no dilution?

因为八百亿的稀释就是八百亿的稀释。

Because 80,000,000,000 is still 80,000,000,000 of dilution.

对吧？

Right?

要保持中立的话，他们最终还是得花八百亿现金把那些股份买回来。

To be to be neutral, they're going to ultimately have to repurchase 80,000,000,000 of cash to get those shares back.

对吧？

Right?

如果你觉得能在合理时间内还清，那为什么不发债呢？

So wouldn't you do debt if you thought you could get pay it back in any reasonable amount of time?

首先，他们可能根本不想花八百亿。

Well, first of all, it may be that they don't want to spend 80,000,000,000.

他们也许想花两千亿，照样可以加杠杆。

They might want to spend 200,000,000,000, and they might lever the same.

他们可能会说，八百亿债务，再加八百亿股权，这样你就能背上一千二百亿的债务，还不影响信用评级。

They might say 80,000,000,000 of debt, another 80,000,000,000 of equity, which will say, and you'll take on a 120,000,000,000 of debt without engendering a credit rating.

所以首先，他们两种方式都可以用。

So first of they could do both.

对。

Right.

这是其中一个观点。

That's one comment.

对。

Right.

没错。

Yeah.

而且，还说不准什么时候能回本。

And then, you know, it's not clear the payback.

我的意思是，说到底，这事儿也不明朗。

I mean, again, to give it is not clear.

我们以后可以聊聊，到底有没有可能真正回本。

We can talk at some point about is the payback there in the end at all.

但即使如此，这些AI数据中心的回报周期通常也是两到三年。

But even when it is, the payback on these AI data centers is normally two to three years.

现在埃隆通过与Anthropic的存储合作，表现远超这个水平——只要合作持续一年出头，他就能把所有投入都赚回来。

Now Elon has massively outperformed that with his storage deal with Anthropic, where he's getting all his money back if the deal lasts a year and a bit.

不过一般来说，回报并不是没有，三四年的时间也算相当长了。

But normally, the payback is not it's not nothing, and three or four years is a fair amount of time.

谷歌是全球第二大盈利公司。

So Google is the second most profitable company on the planet.

他们完全可以合理地借到所需的一切资金。

They could borrow all they wanted to within reason.

我的意思是，他们背了大约七百亿美元的债务，完全有能力借到任何需要的钱。

I mean, have some debt, I think about $70,000,000,000 of debt, so they could have borrowed all they want.

我只是觉得，如果你打算在未来四年每年花三千亿美元，而回报还没回来，那保持一个强劲的资产负债表会更明智。

I just think it's smart to have a strong balance sheet if you find that you might wanna spend 300,000,000,000 a year for the next four years before it comes back.

这确实能让你在债务市场上，不管哪周哪月遇到什么波动，都能稳得住。

Well, certainly, insulates you from any any colds or or flus you get in the debt market on any given week or month.

至少你不用操心债务市场的波动，虽然也会有小恐慌。

At least you don't have to worry about the vagaries of the debt markets, do have micropanics.

他们确实会经历小恐慌。

They do have micropanics.

我觉得我们该拿个奖，表彰在三期节目中对Anthropic的讨论时间最短，这对我们来说还挺厉害的。

I think we deserve a a milestone slash award for the shortest time given to an Anthropic section in a trio show, which which is impressive for us.

我话题跳得有点快，抱歉啊。

I I'm jumping around, forgive me for it.

我们聊了很多关于上市公司的事，尤其是SaaS公司，往往不被资本市场看好。

But we've talked a lot about public companies, and often it's been a tough conversation with SaaS not being appreciated by public companies.

我们看到了Snowflake。

We saw Snowflake.

我们看到了MongoDB。

We saw MongoDB.

我们看到了Salesforce。

We saw Salesforce.

这是很长一段时间以来最好的财报，各家表现都非常出色，股价全线飙升。

Best earnings in a significant amount of time, all of them did very well, and we saw stock surges across the board.

这标志着SaaS的末日结束了吗？

Is this the end of the SaaS pocalypse?

你们是怎么分析这个情况的？

How did you guys analyze this?

从公司层面看，我觉得杰森的那套逻辑挺管用：要么重新加速增长，要么绑定AI支出，那些爆发的公司两者都做到了，而像Zscaler这样故事混乱的，股价就跌了。

Let's there's there's the company specific stuff where I think the kind of Jason rules apply, which is either you reaccelerate or you attach to AI spend, and the guys who exploded did both, And the guys like Zscaler who had a messy story went down.

所以是单个公司的具体情况。

So individual deals.

如果看整体，这里发生的事其实挺有意思，经历了一轮巨大的反转。

If you look overall, what's happened here is you've had it's interesting it's been a massive round trip.

我只看World Cloud这个ETF，它代表SaaS行业，跟QQQ或者SPY对比一下。

I look at just World Cloud, which is the ETF that is the SaaS industry versus the QQQ or SPY.

一个月前，它还跌了30%。

What you basically have is 30% down of a month ago.

然后我们聊过这个。

And then we talked about this.

记得吗？你当时还催我，让我赶紧买点股票。

Remember, we talked about you were giving me shit, have to buy some stock.

所以我买了WorldCloud。

So I bought WorldCloud.

我当时真没搞清楚该选哪几只。

I couldn't figure out which ones.

我有足够的时间好好想想。

I had enough time to think.

于是我买了WorldCloud，现在涨了百分之二十五到三十。

So I bought World Cloud, and it's up 25, 30%.

但今年到头来，还是没赚没赔。

But we're just back flat on the year.

这才是有趣的地方。

That's the fun fact.

截至昨晚，过去一个多月，它们的表现一直优于半导体板块。

As of yesterday night, for the last, I think, just over a month, had outperformed semis.

有意思的是。

Fun fact.

但今年到目前为止，半导体板块大幅跑赢了SaaS，因为SaaS才刚回到持平水平。

But for the year, semis have murdered SaaS to the upside because SaaS has just gotten back to flat.

就连今天，你也看到股价明显回落了。

And now, even today, you've seen a significant bounce down.

至于具体个股，我还是听Jason的，但总体来看，这说明之前的炒作太过了，后来大家才反应过来：这些公司根本不会归零。

To me, I'll defer to Jason on the individual ones, but overall what it says is the narrative got way overdone, and then people looked up and said, These things aren't going to zero.

只要不会归零，就有现金流价值，那它们的实际价值应该比现在更高。

If they're not going to zero, they have cash flow value, so they probably are worth more than that.

但这并不意味着它们有长期的上涨动力。

It doesn't mean that they have kind of long term momentum.

我可没盯着World Cloud说：这个月我赚了30%。

I'm not looking at World Cloud and saying, I made 30% this month.

明年我还要再赚30%。

I'm going to make another 30% next year.

那我就先不说了，谢谢。

I'm probably, Thank you very much.

说得真好。

That was great.

至于具体公司，我就不多说了，想听听你的看法。

And then on the individual, I'd kind of chase it over to you, your thoughts on some of the individual companies.

说到媒体，如果你只看最近三十天，确实不错。

You know, the media, I mean, if you just viewed it the last thirty days, it's great.

看你用哪个指数，我选的那组云和软件股，今天录完这段话后，今年已经涨了5%，又是不错的一天，对吧？

Depending on what basket you use, I have a slightly more optimistic basket, but my basket of cloud stocks, software stocks is up 5% this year after today when we record this another great day, right, in this run.

但纳斯达克涨了21%，半导体更是涨得离谱，快到三位数了。

But Nasdaq's up 21%, and I think semis are in the triple digits or close.

过度调整是结束了，但根本问题还在那儿。

So great that the overcorrection is over, but the fundamental concerns are all there.

对吧？

Right?

我觉得我们唯一学到的是，我从来没搞懂过那种所谓的‘ spatocalypse ’为什么会引发全面恐慌。

I think our only learning is that there was a there was a I never understood the total panic of the spat spocalypse.

总不能全靠感觉写代码吧。

It couldn't all be vibe coding.

对吧？

Right?

虽然哈里的节目里大家都想搞感觉编程，但其实他们都在用正经工具。

Notwithstanding Harry's show where they all wanna vibe code, they're all serum.

这说不通。

It didn't make sense.

但像座位缩减这类深层问题，还有Gartner说今年AI软件支出将增长60%，这意味着其他地方必须砍预算，这些问题都没变。

But the meta issues of seat contractions, of the fact that AI software spend according to Gartner will be up 60% this year, so that means it's got to be cut somewhere else.

真正重要的是，我们不再处于自由落体状态了，因为自由落体只会引发全面恐慌。

Those issues haven't changed, and I would say we're no longer what's great is we're not in free fall because free fall leads to panic everything.

恐慌性抢购，什么都恐慌。

Panic buying, panic everything.

但即使在复苏之后，Atlassian 依然估值是年经常性收入的四倍，HubSpot 是 3.8 倍，Salesforce 也是四倍。

But even with the reacceleration, you know, Atlassian still at four x ARR, you know, HubSpot 3.8 x, Salesforce four x.

我知道 Rory 完全不同意我，但我真觉得，如果没有十倍甚至更高的回报，投资根本没意义。

So I know Rory really disagrees with me here, but I just don't think investing really works without 10x or higher outcomes.

我知道它们都是上市公司，也都很成熟了，但我认为 SaaS 领域的恐慌潮已经过去了。

And I know they're public and I know they're mature, but I think the SaaS the panic part of the SaaS pocalypse is over.

我们就像往常一样，反应过度了，现在可能又过度集中于半导体股票，而这些股票本身也有季节性波动。

Like, we overcorrected as we always do, and we may be over indexing on semiconductor stocks today that that are will be seasonal as well.

但那些根本问题并没有消失。

But the issues haven't gone away.

只是我们对这些问题爆发的时间点反应过度了。

It's just we we overpanicked on the timing of them.

其实，我们看法是一致的。

Actually, are in agreement.

我觉得我们说的是一回事。

I think we're saying the same thing.

这真是难得一见，整个行业都被压得低得离谱。

I'm this was a rare occasion where, literally, an entire sector was discounted to the point where it made no sense.

一个月前，我还能挑出一堆股票，觉得它们便宜得简直不合理。

A month ago, I was able to identify a bunch of stocks where I could say, hey, they're so cheap that it's silly.

现在你得反过来问，经过这轮调整后，这些股票里哪些真有复苏或者AI故事可讲。

Now you actually have to say which of these stocks, given that we've repriced, of these stocks has a genuine reacceleration or AI a catch story.

这话说起来就难多了。

That's a harder message.

我的意思是，确实有几只表现特别突出，但总体来说，还是挺难找的。

I mean, I think some of the ones who killed it do, but overall, it's still pretty tough.

所以我觉得咱们想法一致，杰森。

So I think we're in sync, Jason.

我觉得有意思的是，你看。

I think the interesting thing is there is look.

几乎所有人都出现了温和的估值回升。

There was a modest reacceleration of multiples for almost everybody.

对吧？

Right?

就连受创最重的Mondays和Atlassians，它们的估值也从谷底反弹，只是从一文不值涨回了平平无奇的水平。

Even the hardest hit Mondays and Atlassians, their multiples all bounced off the hard deck and just went back to crappy from worthless.

但今年到目前为止，真正值得记住的一点是，时间已经足够长了，那些真正受益于AI的上市公司——我们姑且称它们为成熟的上市公司——已经开始看到成效。

But the real the only real learning of this of the year to date is that it's been long enough now that the public companies let's just call them mature the mature public companies that are benefiting from AI are seeing the boost.

没错，半导体股确实涨了。

So, yes, semiconductors are up.

但之前杰夫·劳森上这档节目时说过，他最近没怎么关注Twilio，但他很确定公司会从AI中受益，因为AI代理比普通代理更依赖我们的语音服务。

But, you know, when Jeff Lawson was on this show, he said, I haven't run Twilio in a while, but I'm pretty sure we're gonna benefit from AI because agents in AI just need to use more of our voice than other agents.

虽然花了大概一两个季度才显现，但他确实说对了。

And it took a quarter or two, but he's right.

今年股价已经涨了57%。

Stock's up 57% this year.

增长率从大概百分之四五涨到了百分之二十。

It's gone from, I think, four or 5% growth to 20.

Okta，就是你爸以前用的那个企业认证系统，今年涨了57%，差不多56%，对吧？

Okta, you know, which was which was your dad's enterprise auth system, up 57% this year, 56%, right?

Datadog，人人都在用的那个工具，对吧？

Datadog, which everyone uses, right?

所有AI领域的领军企业都在用，今年涨了一倍。

Every AI leader uses, up 100% this year.

所以我们这个显而易见的教训就是：给了六个月时间，所有做智能代理产品、代理需要更多这类服务的软件公司，股价全涨了。

So our mistake in the captain obvious learning is, hey, look, we gave it six months and every software leader that were Agentic products and agents need more of it is up.

不管他们在X上怎么说，那些只靠人工使用的工具，业绩反而在下滑。

And no matter what they say at x, the ones that only humans use is kind of down.

就算它们从低谷反弹了，也还是这样。

Even if they bounced off the hard deck, right?

市场根本不再需要更多的按人头计费的许可证了。

There's just not an appetite for more human per seat licenses.

就连Salesforce，增长重新加速了，靠的也是扎实的努力。

And even Salesforce, which reaccelerated growth, it was all through hard work.

靠的是Agent Force。

It was through Agent Force.

靠的是非有机的收购。

It was through inorganic purchases.

马克和团队在电话会上说，他们第一次把业务拆成了两个垂直领域。

And Mark and the team came on the call, and they said and they split the business up into two verticals for the first time.

我想起来了，具体叫什么我给忘了。

I think it I forgot what they're called.

一个是Agentic或者Agent Force，另一个是其他的。

Agentic or Agent Force and the others.

他们说，软件业务的增长会长期停留在个位数。

They said the software business will be perpetually in single digit growth.

这已经是不错的结果了，但其他部分都是两位数增长，而且正在以12%到13%的速度上升。

That's as good as it's to get here, but we're double digits in the rest, and that's growing 12 or 13%.

我们做了分拆。

We bifurcated.

我们看清了趋势：真正有前景的是AI驱动、以智能代理为核心的的产品，而传统的按人头收费的软件正在逐渐被淘汰。

We saw what made sense, which is AI fueled, agentic focused products, but the classic human per seat software really is dying.

还没完全消失。

It's not dead.

但根本没人愿意买这种东西了。

But no one wants to buy this crap.

事实上，他们还在砍掉这些业务，因为得把钱省下来买那些该死的token。

In fact, they're cutting it because they gotta come up with money for all these goddamn tokens.

人人都想要这些token，Rory和Harry。

Everyone wants these damn tokens, Rory and Harry.

我们根本不需要为那些根本不用项目管理软件的人再买一个用户席位。

You know, we're just we just don't need another human seat for folks that that don't do any work on a project management software.

这事儿说到底，数学上就是这么算的，对吧？

It just that has to be the math, right?

如果你的AI软件要增长60%，那就不可能不砍掉相当一部分其他业务。

You can't grow 60% AI software without cutting some material amount of the rest.

所以我们现在看得清清楚楚。

So we we are seeing it.

但其实我们早该想到这一点。

But we should have known this.

我们早就该抛开自己当初的赌注，只看谁跌得最少。

We should have all forget about the way I've made my bets, just who fell the least.

那时候正是最低谷。

That was at the Nadir.

对吧？

Right?

我当时押的是跌得最少的那一个。

I bet on who fell the least.

我们本该全都押Jeff Lawson的赌注，就是那样。

We should have all made the bet the Jeff Lawson bet, which is, okay.

到底谁会真正从Agentic中的智能代理里受益？

Who's gonna honestly benefit from agents in Agentic?

这其实没那么复杂。

And and it's not that complicated.

对吧？

Right?

当然是Twilio和Datadog了。

Of course, it's gonna be Twilio, Datadog.

我从来没想过Okta，但如果当初我用Claude时再聪明一点，早就该看出来了，这其实很明显。

I never would have thought Okta, but I I if I'd if I'd been a little smarter with Claude, we would have figured it out because it's it is an obvious one.

如果所有做身份认证的私营公司都在爆发式增长，那Okta也理应如此。

If if if everyone doing auth that's a private company is blowing up, they should blow up too.

对吧？

Right?

我觉得真正的挑战在于，到底要达到什么样的规模才适合上市，就像杰森说的，流动性又该是什么样子。

I think the real challenge is more a question of what does critical mass look like to go public, to Jason's point, what does liquidity look like, all that.

你们经常提到 Replit。

You guys often talk about Replit.

这是个典型的例子，一家在 AI 出现前就存在的公司，巧妙地搭上了这波趋势，一下子飞起来了。

There's an example of a company that was pre AI, that brilliantly attached to the trend and just got a ton of lift.

其实有很多办法可以借上这股势头。

There's stuff you can do to get lift.

只有一个标准：你是不是跑得更快了？

There's only one test: are you going quicker?

如果你跑得更快了，那就说明你搭上了顺风车。

If you're going quicker, you've gotten Lyft.

只要搭上了顺风车，你就没问题了。

If you've gotten Lyft, you're fine.

我觉得新一波以 LLM 为核心的 AI 应用生态里，Superbase 就是其中一个例子吧？

Can imagine lots of parts of the new LLM first AI harness software stack isn't it Superbase?

它把 Postgres 数据库用得特别巧妙，跟各种功能无缝结合，增长也特别快。

It's the Postgres database that's just brilliantly co attached to everything it's growing quickly, too.

这种情况难免会发生，但关键在于要有敏锐的产品管理能力，确保你紧紧跟上未来的趋势。

So there will be incidents of that, but it requires deft product management and making sure you're attached to the future.

如果做不到，你不会立刻被淘汰，但你会面临一个问题：怎么创造价值，又怎么把价值变现——这也会引出其他几个话题。

If not, you're not going away quickly, but you just have a question of how do you create value and how do you realize value, which will segue to a few other discussions.

我们之前提到过那种爆发式增长和软件开发的事。

We we said about kind of explosive growth and then software creation.

Cognition这家公司融资了十亿美元，估值高达260亿美元。

Cognition raised a billion dollars at a $26,000,000,000 valuation.

他们的核心产品Devon，年经常性收入达到了4.92亿美元。

Devon, their core product, hit 492,000,000 in ARR.

增长非常惊人，还拿下了不少全球顶尖大企业作为客户。

Incredible growth with some mega customers, some of the largest enterprises in the world.

Jason，我觉得你刚才说的那番话很有道理。

Jason, I thought your statement here was was a good one.

如果Cursor的年收入是30亿美元，那定价60美元算便宜吗？

Was Cursor at 3,000,000,000 ARR cheap then if that was priced at 60?

你怎么看待Cognition这一轮融资的增长？

And how do you reflect on the growth, this Cognition round?

我很想听听你的看法。

I'd love your thoughts.

要是今天我能更懂Cognition就好了。

I wish I was more of an expert on Cognition today.

我们刚开始做这个播客的时候，我就说过，我投资组合里两个智商最高的CTO，早早就开始用Devon了。

When we started this pod, I did say that probably the two absolute highest IQ CTOs in my portfolio were using Devon in the early days.

那时候，任何工具都还很不成熟。

That when when things were still pretty crappy before for anything.

但这个想法，我觉得依然非常有吸引力。

And the idea, I think, is still super compelling.

现在我们有了具体的数据。

Now we have some metrics.

至少这个理念是：你的工程师效率提升了十倍、百倍，甚至一千零一倍，这太棒了。

The idea at least is, hey, it's great that your your engineers are 10 or a 100 or 1.001 x more productive.

更有趣的是，你完全可以拥有一个自主的AI工程师。

What's far more interesting is that you can have an autonomous AI engineer.

这才真正有意思。

That's much more interesting.

我记得我投资组合里最聪明的那位CTO，早期就用过Devin和Slack，他只要吩咐Devin去处理一些任务，然后它自己就会完成提交。

And I remember that probably the smartest CTO in my portfolio just in the early days ran Devin and Slack, and he would just tell Devin to go do these things and come back and make the commit on their own.

当然，一年前那时候的水平肯定一般，毕竟底层模型还没那么强。

And I'm sure it was pretty mediocre in like, you know, a year ago or something like that just because of the nature of the underlying models.

但对我来说，这个愿景比什么Claude代码工具都更吸引人。

But the vision to me is actually more compelling than all this Claude code crap.

与其事无巨细地指挥代理，不如直接交给这些自主的工程师去搞定。

It's much more compelling to just tell the agent what to build or not tell the agent, have these autonomous engineers that do it.

对吧？

Right?

另一方面，也许这并不算特别惊人，毕竟这个领域投的钱实在太多了。

On the other hand, you know, there's just maybe it's not that impressive because there's just so much money in the space.

也许它只是把模型重新包装成一个更有趣的流程，如果运行得还行，那其实也没那么惊人。

Maybe if it works reasonably well, alright, and it's just a way to repackage the models for one workflow that's more interesting, maybe it's not as impressive as it looks.

但这个愿景，我觉得是这个领域里最吸引人的。

But the vision, I think, the most interesting vision in the space.

我们为什么要让平庸的销售员变得更强？

Why do we want to empower mediocre sales reps?

不，我们要用AI来自动化他们。

No, we want to automate them with AI.

平庸的工程师也是一样。

Same with mediocre engineers.

干脆把他们淘汰掉吧。

Let's get rid of them, man.

留下顶尖的，其他的都交给Devin。

Let's have the best ones, but all the rest, let's have Devin.

Devin从不争辩。

Devin doesn't argue.

Devin可不只是想做那些像你顶尖工程师才碰的有趣问题。

Devin doesn't only wanna work on interesting problems like most of your best engineers.

这事儿没意思。

It's not interesting.

我不会加入的。

I'm not gonna join.

我刚跟一个人聊过，他拒绝了Anthropic的offer。

I was talking to someone that that turned down an an an off right Anthropic.

他说这工作不够有趣，但还是想参与其中。

Just said it wasn't interesting enough, but they wanted to work them on.

他就是不想去搞那些基础应用的边边角角。

They just didn't wanna work on, like, little edges of basic application.

太无聊了。

It was boring.

对吧？

Right?

那就让Devin来干吧，老兄。

So let's have Devin do it, man.

我希望他能成为所有项目里最成功的那个。

So I hope it's the biggest winner of all.

这是一个变化极快的市场，领先者随时可能易主。

This is very much a market where leads change hand at a furious pace.

每一个估值上万亿的公司，包括谷歌，都恨不得把你吃掉。

So every one of the companies that's worth a trillion bucks and Google probably has a every one of the companies worth a trillion bucks is gonna wanna eat your lunch.

这风险很高，但回报也巨大，祝他们好运。

So it's a high risk, huge market return, and good luck to them.

我太喜欢了。

I love it.

今天早上我跟一位顶尖的CTO聊了聊，他说，他们公司相当于给每个员工都发了张信用卡，还说额度不限。

I was speaking to one of the best CTOs this morning, and he said, we've just the analogy is we've just given a company credit card to every employee and said there's no limits.

想花就花吧。

Spend away.

这就是如今的令牌预算机制。

And that's the Token Budgeting today.

我想问你的是，如果有了严格的预算管控，我们是不是对市场规模估计得太高了？

And and my question to you is, if there is rigor and budget instilled, are we dramatically overestimating market size?

我觉得我们并没有高估市场规模。

I don't think we're overestimating market size.

不过，咱们还是祈祷一下吧。

So but but let's pray.

这事儿挺有意思的。

It's funny.

上周我们还聊过这个，我当时真想发条推文，因为录节目那天周二，我就在想，说这些东西可能没有投资回报率，听起来好像有点大逆不道。

We talked about this last week, and I was gonna tweet this because literally on Tuesday when we record, I was like, it may seem her heretical to say that maybe this stuff isn't got an ROI.

可等到周四内容发布时，突然之间满屏都是‘天啊，投资回报率’这类文章。

By the time the thing came out on Thursday, there'd been explosion of these, oh my god, ROI articles.

所以正好赶上了舆论觉醒的临界点。

So it was right on the cusp of the Zeitgeist where people finally woke up.

昨晚我一直在想这件事。

And I was thinking about it last night.

一切都说得通了。

It all makes sense.

去年十一月或十二月，Claude 推出了那个神奇版本的 Claude Co 模型。

Sometime in November, December, Claude produced the magical version of the model of Claude Co.

那个模型真的特别好用。

That just works.

到了 2026 年初，他们调整了计费方式，改成了按使用付费。

In early twenty twenty six, they kind of changed the pricing model so you have to pay as you go.

结果第一季度，所有人都疯狂地用了起来。

And it's like literally everyone cranked in Q1.

我几乎能想象，全美每个 CFO 的办公室里，都有人正在做应计会计处理。

And I can almost imagine in every CFO's office in the land, someone was doing accrual accounting.

到了五月中期，他们突然发现：天哪，我们以前是按这个数字估算账单的，现在应计金额居然错了十倍。

And by mid May, they suddenly realized, oh my god, You know, we used to estimate our bill based on this, and suddenly we're 10 x wrong on our accrual.

就在那一刻，全美所有公司都突然醒悟了。

And, literally, the penny dropped simultaneously across the entire corporate US.

我们叫他们在第一季度猛用，结果真就猛用了。

We told these guys to crank in q one, and fuck me, they cranked.

一看，我们整个预算都花光了。

And it looks like we spent our entire budget.

这事真是普遍发生了，谁都没躲过去。

And literally, it universal, and it's happened.

所以一切都说得通了。

So it all makes sense.

我就是这时候发现这个问题的。

This is about the time I've discovered it.

以前别人都没提，我先把这个事摆上台面，现在我换个角度说说好处。

Now to your point, I was raising the issue when people weren't, but now I'm going to take the positive side.

只要生成的代码质量高、真有用。

Provided the code that's generated is good and useful.

我能想象有人会暂时叫停这件事。

I can imagine a scenario where people put a pause on it.

我看到优步今天宣布，给每个人每月发1500美元，这跟我们目前看到的平均支出水平差不多。

I saw I think Uber announced today they're just going to give everyone $1,500 a month, which is about the right amount on what we're seeing in terms of averages.

我不会说这是‘合适’的金额。

I won't say the right amount.

这属于价值判断。

That's a normative statement.

比我们之前看到的平均支出略高一点，他们这是想给每个人设个上限，好控制开销。

It's about a little above what we saw the average spend was, so they're basically going to cap everyone and try and get control of the spend that way.

从中期来看，对你有利的是，前沿模型的成本并没有下降。

The thing you have in your favor over the medium term is the frontier models aren't getting cheaper.

我想把这一点说清楚。

I want to be clear on that.

实际上，它们的成本还在微微上涨。

They're getting actually slightly more expensive.

但今天最先进的模型，一年后成本会降到原来的五分之一到十分之一，因为它已经不再是前沿模型了。

But the model that is frontier today will in a year be five to 10x cheaper because it won't be the frontier model anymore.

所以，如果你现在从这种过度支出中获得了价值，未来十二个月你可能会放缓投入。

So if you're getting value from this egregious spend today, you might slow down for the next twelve months.

只要你跟得上这个价格走势，等它不再是前沿模型时，就能以更低的价格使用，依然能从中获得收益。

But as long as you stay on that pricing curve, what will then be not a frontier model but an older model will be available at a cheaper price and you will be able to continue to get value from it.

这么说吧，我不觉得大家会突然醒悟，说：天啊，我们再也不花钱在这上面了。

That's a long winded way of saying, I don't think you wake up and go, oh my god, we're not going to spend any money on this.

我觉得随着人们意识到自己到底花了多少钱，采用速度很可能会明显放缓。

I think the pace of adoption might slow markedly as people realize quite how much they've spent.

我有两个增长最快的被投公司，说他们今年的预算早就花光了。

I had two of my fastest growing portfolio companies saying they already blew through their budget this year.

所以这不只是大公司的问题，对吧？

So it's not just the big guys, right?

我有两个增长最快的公司。

Two two of my fastest growing companies.

我那些增长最慢的公司都没这么说。

None of my slowest growing companies have said that.

我那些增长最慢的投资组合公司，没有一家说：‘我们把预算用光了。’

Not not a single of my slowest growing portfolio companies have said, guys, we burnt we burnt through all the tokens.

我不确定这是因果关系还是相关性，但有两家这么说，确实挺有意思。

I don't know if it's causation or correlation, but it was interesting that two of them said that.

从实际角度来看，当然得控制成本。

I think that at a practical level, of course, there has to be cost containment.

对吧？

Right?

我们不是所有公司都像那些刚融了五千万、只有六个人的初创企业，花多少钱都无所谓。

We're not all startups that just raised 50,000,000 with six people where it doesn't matter.

对吧？

Right?

成本控制是必须的。

There has to be cost containment.

而且很多人在大量浪费令牌。

And folks are massively wasting tokens.

他们真的在大量浪费这些令牌。

They're massively wasting them.

所以人们在开发时必须更加谨慎地考虑如何使用。

So it has to come that people have to be more more thoughtful with what they build.

实际上，编码中用掉的大部分令牌都用在了测试环节。

The vast majority of tokens that are used in coding are in QA anyway.

而不是在生产环境中。

It's not in production.

我们可能需要更仔细地思考一下这种方式，或者不这么做。

We may have to be more thoughtful about how we do that or not.

但我觉得，如果我们不那么冲动地做事，也没关系。

But I think it's Okay if we have we don't shoot from the hip as often.

我觉得，放慢我们开发新功能的速度，也无妨。

I think it's Okay if we slow down the number of features that we build.

我觉得这也没那么严重。

I don't think it's the end of the world.

对吧？

Right?

话虽如此，现在越来越多的人同时使用多个模型。

Having said all that, more and more folks are using multiple models at the same time.

比如，我就不多提Replit了。

For example, not to talk about Replit too much.

现在Replit已经自动这么做了。

Replit does it automatically now.

如果你有个复杂的功能，Replit会用Claude Sonnet来构建，而不是OpenAI，为的是省钱，然后让Codex来检查成果。

Replit, if you have a complex feature, Replit builds it in Claude Sonnet, not OpenAI, it builds it in Sonnet to save money, and then it has Codex come in and check the work.

它同时用这两个模型，但你根本察觉不到，也不需要去管，因为主要面向的是非技术人员。

It has both of them, and you don't even know this you don't even check it because it's mostly for nontechnical users.

有个挺有意思的观察：如果你想了解成本敏感度，不妨研究一下Replit和Levelable，它们现在正承受着客户带来的巨大压力。

One interesting learning, if you want to learn about cost sensitivity, actually study Replit and Levelable because they're under huge pressures right now from their customers, massive pressures.

所以如果你想看看未来，我们可以看看我一个项目里优步那位同事怎么说的。

So if you want to see the future, we can we can look at what the Uber dude said in one of my portfolios.

但对我来说，更有趣的是观察Rapid和Lovable如何彻底改造他们的平台，因为那多花的几美元，会直接导致底层客户流失。

But it's much more interesting to watch how, to me, how Rapid and Lovable are evolving their platforms radically because a couple extra dollars there leads to churn at the bottom of their customer base.

所以他们极度关注成本控制，但业务依然在运转。

So they're radically focused on cost containment, and yet they still run.

我不太清楚LevelUp的情况，但它们确实同时运行两个模型，不是所有事情都用，但遇到复杂任务就会启用。

I don't know about LevelUp, but still runs two models, not for everything you do, but for anything complicated.

现在架构师角色由Codex担任，负责检查Sonnet的输出，效果非常好。

The architect agent now is Codex that it brings in to check Sonnet, and it works really well.

用Codex来核对Sonnet甚至Opus生成的内容，威力简直惊人。

Like it's incredibly powerful to check your work in Sonnet and probably Opus with Codex.

这真的特别有用。

It's incredibly useful.

每次它都能发现问题。

It's every single time it finds issues.

每次都能发现问题。

Every single time it finds issues.

太强大了。

It's so powerful.

所以这是走向另一个方向，对吧？

So that's going the other way, right?

随着多智能体系统越来越普及，任务也越来越复杂，我们自然会想要……

Is that as we go more and more multi agent and we ask them to do more, of course, we're going to want to.

所以我们得设预算，但这种张力只会越来越强。

So we have to have budgets, but the tension is only going to grow.

我觉得我们还没说到罗里的观点，你刚才是问了我的看法。

And I don't think we've yet to Rory's point, you just asked my opinion.

对吧？

Right?

我觉得在这档节目里，我们学到的是——而且这跟节目刚开始时已经不一样了——其实根本没那么重要。

I think in the course of this show, what we've learned, and it has changed since the beginning of the show, is it really doesn't matter.

这情况可能在九十天内就变了。

And this could change in ninety days.

但到目前为止，老平台便宜、普通模型便宜都没用，因为我们根本不需要它们。

But so far, it really doesn't matter if older platforms are cheaper, normal models are cheaper because we don't want them.

当然，有些场景我们还是需要的。

There are use cases where we want them.

但整体来看，整个生态我们都押注在最好的上面。

But overall, as an ecosystem, we are all in on the best.

我们都选了Opus。

We are all in on Opus.

这才是我们真正想要的。

That's what we want.

所以我们没怎么享受到AI带来的成本下降，反而在往涨价的一边掏钱。

So we're not benefiting as much from the deflationary benefits of AI, and we're paying into the inflationary side.

我再说一遍，我不是想唱衰，实际上我整体是乐观的。

Again, I don't want to sound negative because I'm actually a net positive.

所以我想再回过头来谈一谈这个。

So I want to come back to that.

但首先，我觉得从宏观角度看，这对模型提供商来说是个好消息。

But first of all, I think the zoom out comment is this: At some enormously high level point, this is validatingly good news for the model providers.

现在大约有3%到5%的科技预算被用掉了，大家突然意识到：天啊，我们真在花这么多钱，但没人说要彻底砍掉。

What's happened is somewhere around 3%, 5% of tech spend, everyone noticed, oh my god, we're spending this, and no one said, we're going to cut back to zero.

这意味着你已经确立了一个市场规模可能在5000亿到1万亿美元之间的新类别，现在各公司都在走流程，琢磨怎么挤出这笔钱。

What this means is you've just established a category that probably has a market size of half 1,000,000,000,000 to $1,000,000,000,000, and everyone is now going through the corporate process of saying, how do I find that money?

钱从哪儿来？

Where do I find it elsewhere?

怎么安排这笔支出？

How do I manage it?

怎么控制上限？

How do I cap it?

但他们没人说：别花了。

But what they're not saying is, Stop it.

你已经打造了一个规模巨大的新领域。

So you've built a category that's huge.

如果从这个角度讲，对我们这些在硅谷的人来说显而易见，但那些说这一切都会消失、觉得这很荒唐的悲观派，错了。

So if I'm on topic, it's obvious to us because we're in the valley, but the doomers who say it's all just going to go away because it's silly, no.

美国企业界已经承认：我们确实花上了这笔钱。

Corporate America said, We're spending this kind of money shit.

不喜欢也没用。

Don't like it.

花得是多了点。

We're spending too much.

但我们现在得想办法把这笔钱用好。

But we're going to have to have a plan to spend it.

这简直是极大的认可。

It's enormously validating.

如果能看看杰森怎么管理这笔开支，会很有意思。

It would be interesting to see Jason as part of managing that spend.

我们是不是还都会想要每样东西都买最贵的型号？

Will it still be true that we all want the most expensive model for everything?

我知道我有些做应用的公司，会用多个模型，便宜的活儿就用开源的。

I know I see some of my app companies find a way to use multiple models, use open source for the cheaper stuff.

对他们的应用来说是这样，那开发环节呢？

For their application, but for their development?

是不是有些场景你只需要一个模型，但其他情况你可以把任务分出去？

Is it a question that there are things where you want one but not two foundation models, but are other things that you offload?

因为我觉得，再说一遍，非前沿模型的成本一直在降，但前沿模型的成本却在涨。

Because I don't think again, going back to what I said that the cost of off frontier models keep going down, but the cost of frontier models keep going up.

我刚说了一句积极的话，但我觉得企业界不会说：太好了，前沿模型越来越贵，我们正合意。

And I just made a positive statement I don't think corporate America is saying, Yay, the cost of the frontier model is going keep going up, and we're good with that.

所以我觉得，现在财务部门终于盯上我们了。

So I think we've now been found by finance.

他们盯着CEO问：喂，你一月一号还说想用多少就用多少呢。

They're looking to the CEO and saying, Dude, on January 1, he said, Use all you like.

这事儿肯定会很精彩。

It's going to be amazing.

现在都五月十五号了，问题来了。

And now it's May 15, and we've a problem.

咱们得想办法解决。

Let's figure it out.

我不觉得他们会说，不管什么情况都继续用前沿模型，不是按每个token算钱，而是按每次调用、每次运行的成本来看，价格一直在涨。

I don't think they're going to say, keep using the frontier model for everything, where not on a price per token basis, but on a price per pass, on a price per run basis, it's been going up.

我觉得我们现在得开始搞清楚成本和边际成本了。

I think we're now going to have to discover costing and marginal costing.

杰森，你不觉得开源对开发预算会有显著影响吗？

Jason, do you not think that open source will have a meaningful impact specifically on development budgets?

你明明已经区分了不同场景。

You you clearly identified different scenes.

也许吧。

Maybe.

我只是觉得我们把两种不同的情况混在一起了，要是能有确切的数据就好了，我们搞混了应用程序用的模型和软件开发用的模型。

I just think I just think we're confusing the these narratives of two things, and I wish I had the exact number, but we're confusing the models used for by applications versus models used for software development.

至于应用程序方面，我的意思是，现在OpenRouter简直爆了。

And for applications, anyone I mean, this is OpenRouter blowing up.

这已经是全面铺开了。

This is everything.

大家都在优化。

Everyone's optimizing.

就算你没刻意优化，其实你也在优化。

Even if you're not optimizing, you're optimizing.

对吧？

Right?

你用Opus做高端任务，一次差不多就一块钱。

You know, everything high end you do in Opus is like a buck.

明白吗？

Okay?

而且可能还会更高。

And it could be more.

可能是几美元。

It could be dollars.

而在低端，只要五毛钱。

And at the low end, it's 50¢.

你不可能每次聊天都只收五毛或者一美元。

You can't do 50¢ for a chat for every single chat or a dollar.

所以这种情况可能会变，确实有些工作流程需要大量的推理和思考，但我觉得我们搞混了，开发者其实压力很大。

So this could change, and there are certainly workflows where you need massive amounts of inference and thought, but I think we're confused into developers are under pressure.

这些可怜的人，就像哈里在节目里说的，他们得加班到晚上九点、每周六天，真是不容易。

These poor guys, as Harry has in the show, they gotta work nine nine six, these poor guys.

他们周六才不会用个烂模型呢。

They're not gonna use a crappy model on Saturday.

算了，我不干了。

Like, I quit.

这我可做不到。

I can't no.

如果你告诉我不能自己选模型，我直接就辞职不干了。

Literally, I would quit as a developer if you told me I could not use the model of my choice.

我真会辞职。

I would quit.

这不值得我花时间。

It's not worth my time.

你是说我就得用那个，我不是说这种情况不会发生，但反正我是会辞职的。

Like, you're literally telling me I've got to use so I just I'm not saying it won't happen in some use cases, but I would quit.

所以对我来说，最值得琢磨的是，把这些都串起来看——现在我们谈到预算了，其实就是谈成本的问题。

So the more interesting thing to me from all this here's the thing that's interesting to me to tie it all together, then maybe we can move I think now that we're hitting budget discussions, just discussions.

对吧？

Right?

现在CIO们越来越参与进来了，这事儿也不只是Uber一家，虽然Uber那故事被夸大了，有点儿道听途说，但今年这情况肯定会蔓延到各处。

And now that CIOs are more involved, and now it's it's I think it's not just Uber, even though I think the Uber story was a little blown up and a little apocryphal, but it's gonna happen everywhere over the course of this year.

对吧？

Right?

预算总得有来源。

The budget has to come from somewhere.

我真的觉得，到今年年底，我们在工程和产品上会更倾向于用模型而不是人。

I really do think by the end of the year, we're gonna choose tokens over humans for engineering and product.

我们现在正处于临界点，你得做出选择，这跟我以前在大科技公司时的经历完全一样。

We're at the margin, you're gonna go in and you're gonna say because this is certainly the way it worked for me in the old days when I was at a big tech company.

你今年的预算是一千万。

Your budget this year is 50,000,000.

不再是四百人、两百人的编制，你今年的预算可能是五千万、一亿、甚至四亿，用于工程产品开发。

Instead of your budget being four a 100 heads, guys, or 200 heads, your budget this year is 100,000,000 or 200,000,000 or 400,000,000 for EPD, engineering product development.

我以前在Adobe的时候，全都是人力。

When I was worked at Adobe, it was that was all humans.

我们根本不在乎其他成本，因为那些都不重要。

Like, we ignored all the other costs because they didn't matter.

EPD部门，你们的预算以前就只是人头数，对吧？

EPD, your budget was just head count, right?

每个人成本都一样，一年三万美元。

Everyone costs the same, $100 $3,000 a year.

管他是行政经理还是顶尖人才，为了简单起见，都一视同仁。

It didn't matter if they were an office manager or your top because just to keep it simple.

现在好了，2027年的预算讨论要复杂多了，你的预算就是这么多。

Okay, now we're going to have much more sophisticated budget discussions going into 2027, which is your budget is this much.

你们自己决定怎么花这笔钱，各位负责人。

You decide where you want to spend it, leaders.

我会坐在那儿想，我是要再招二十个平庸的工程师，还是把无限额度的代币给我的顶尖人才？

I'm going to be sitting around, I'm going to say, do I want to have another 20 mediocre engineers on my team, or do I want to give my best guys unlimited tokens?

这可能会引发新一轮的变动，不管是AI裁员还是AI替补，跟之前那些找借口裁员的把戏完全不是一回事。

And that may fuel the real, whether they're AI layoffs or just AI backfills, it may fuel another wave of this, which is very distinct from the click up whatever excuse for getting fit.

到年底，这可能是个理性的选择。

This may be a rational choice at the end of the year.

钱就这么多，我选代币。

There is only so much money, and I'll take tokens.

我宁可要代币，也不要B类。

I'll take tokens over a b.

在SaaS这边，我们早就做决定了。

We've already made that choice at SaaS.

我们把所有的B类都砍掉了。

We got rid of all our b's.

我们更愿意拿代币。

We'd much rather have tokens.

这挺有意思。

It's curious.

所以接下来的问题是，你得明确一个比例分配方案。

So therefore, the next question is you have to come to some kind of opinion what the percentage mix is.

大家都会被迫做出一些艰难的选择。

People are going to be pushed into some kind of hard choices.

这会迫使人们形成一种信念。

And then it will force a belief.

你真相信他们能带来20%的提升吗？

Do you really believe that they give you a 20% lift?

如果需求没有变化，那实际上就是25%的提升。

In which case, assuming no net change in demand, you have 25% lift.

花80美元在人身上，20美元在代币上。

Spend $80 on people and $20 on tokens.

你觉得五五开的话，能带来100%的提升吗？

Do you feel it gives you 100% left at fiftyfifty?

一旦你有了明确的预算和交付目标，就不得不量化这些投入了，我觉得这才是接下来要发生的事。

It will force quantification once you have a dollar budget and a set of deliverables, which I think is the next shoe to drop.

确实会这样。

It will.

但更重要的是，在竞争更激烈的公司里，顶尖人才不会容忍得不到他们想要的东西。

But even more than that, more competitive companies, the best people won't tolerate not getting what they want.

所以你是打算这么做的，对吧？

So you're to so you're going look.

我觉得到了2027年，工程和质量保证团队会被彻底重组。你可能会想：我有六个QA工程师，十个，都很优秀。

And so what what's going to I think what will happen going into 2027 is an engineering QA departments will get destroyed because you'll be like, I got six QA engineers, I got 10, and and they're great.

但说实话，我宁愿用两个QA加上AI工具，再把客户成功部门里没划进财务预算的那些人，全换成AI来处理客户入职和支持工作。

But do I really like I'd rather I'd rather go to two with tokens and and whatever's left in my my customer success department that I didn't put into FDs, I'm just going to get rid of them for tokens to manage customer onboarding, customer support.

我会把所有边缘岗位、名单最末尾的那些职位都砍掉。

I'm going to get rid of all the marginal roles, the the ones at the bottom of that of that list.

我愿意拿这些岗位换AI工具。

I'm gonna sacrifice for tokens.

这事儿太简单了。

It's so easy.

根本不用多想，其实特别直接。

You don't it's it's really simple.

那些在第一轮裁员中没被剔除、还在边缘摇摆的人，这次也会被AI取代。

The ones on the bubble that weren't already cut in the first wave will be cut for tokens.

挺有意思，看看你是不是对的。

It would be interesting to see if you're right.

你打算做波测试吗？

Are you going to wave testing?

你打算用代码做波操作，就是得好好想想。

Are you going to wave code with you know, like, have to think about

能少做就少做，剩下的交给模型去完成。

all to whatever the minimum you can do and have and have the models do the rest.

你会自然而然地做出这种选择，而不是随便应付一下。

You will just inherently make that choice rather than someone that's just Okay.

我跟你说，我们

And I will tell you, we

有过

had

这个FDCCO客户成功峰会，是我们SaaS或AI的年度活动。

this FDCCO CS summit our SaaS or AI annual event.

来了几百位领导者。

Was a couple 100 leaders.

几乎人人都在谈，他们正在淘汰Gainsight和其他所有工具和团队。

Almost everyone was talking about how they're getting rid of Gainsight and all the rest and all their teams.

今年的主题就是这个。

That was the topic this year.

我们正打算把所有人员和软件都取消，因为直接在智能代理层面处理更高效。

We're getting rid of all the people and all the software because it's more efficient to just handle this at the Agentic level.

无论是老员工还是新来的，都一样。

This everyone, old and new.

抱歉。

I'm sorry.

这是客户支持领域的事。

That's in customer support.

是客户成功。

Customer success.

是的

Yeah.

我只是说，这些岗位分布在不同部门，年底都会被裁掉。

I'm just saying these are all roles on different lines that will just get cut at the end of the year.

这些部门里剩下的人员，我不会裁掉我最优秀的工程师。

Whatever's left in these departments, I'm not gonna cut my best engineers.

我更不会裁掉我本来就比以前更小的销售团队。

I'm not gonna cut my actually smaller sales team than it used to be.

对吧？

Right?

我不会裁掉核心团队，但我更希望让Token来处理客户来电。

I'm not gonna cut my best folks, but I'd rather have Token's do the inbound call.

我更希望让Token来处理三千万级别的交易。

I'd rather have Token's handle the three k deals.

我更希望让Token来做质检或者客户服务。

I'd rather have tokens do QA or CS.

杰森，恕我直言，我其实更认同你之前说的那番话，换句话说，我不同意你的看法。

Jason, with all due respect, I think I liked what you said earlier, which is a nice way of saying I disagree with this.

在应用层面，比如客户支持，Token的成本相对于总成本来说微乎其微。

At the app level, for like customer support, I think the cost of tokens is so low relative to the total cost.

这点开销根本微不足道。

It's in the noise.

我觉得在客户支持这类事情上，Token的使用量不会成为关键因素。

I don't think token intensity for something like customer support will be a significant factor.

你以后回过头看，不会说：‘啊，这真是花了巨量的Token。’

You won't look back and say, That's the huge amount of tokens.

我只是想搞清楚，从最初开发、代码审查到测试，整个产品和研发流程中，工程师的总薪酬和Token的花费大概会是什么比例？先不考虑其他部分，单说这一块，你怎么看？

I'm just trying to understand the end to end product and development lifecycle across initial engineering, code review, tests, all that, what do you think the split will be between dollars paid to engineers in total and dollars paid to tokens, just in that area, leaving out

我也不确定。

I I don't know for sure.

我只是想说，咱们先回过头看看。

I'm just saying, listen, I don't me let go back in time.

假设我还在Adobe，手下有400号人，行吧？

Let's imagine I was back at Adobe, I had 400 people in my little BU, okay?

年底的时候，预算固定，我们围坐一圈，决定下一年缩减到300人，再把相当于100个人的工作量交给AI来处理。

And at the end of the year, I had it was a fixed budget and we got around the room and we decided we wanted to go into next year with 300 people and the equivalent of 100 humans of tokens.

明白吗？

Okay?

再加100个AI的工作量。

Another 100 of tokens.

我就想这么安排。

I wanted it.

你觉得怎么样？

Right?

如果我现在有固定预算，我肯定会这么干。

This is very this is what I would do today if I was fixed.

就这样。

Okay.

所以现在我会裁掉100个人。

So now I'd go now I'm gonna get rid of 100 people.

剩下的人里，凡是支持岗位表现一般的，直接走人。

So I'm immediately whoever's left in support that isn't great, gone.

明白吗？

Okay?

我会把整个客户支持团队全砍了，只留一个团队负责人。

Whoever's I would get rid of my entire CS team except my head of CS.

功能测试团队我也大部分裁掉，只留下最顶尖的那几个。

I would get rid of this most of the all of my functional QA team, and I would just leave the smartest guys.

我之所以这么做，就是因为需要腾出人力预算，真的。

Like, I would just get rid of all of those people because I need the tokens, man.

我觉得年底的时候这事就会发生，因为大家自然会做出这样的选择。

I think it will happen at the end of the year because people will make that choice.

你这话的意思是，从400人减到300人，把100个岗位换成人力预算，相当于缩减了三分之一左右。

What you're saying but the the math to that, 400 to what you're basically saying is the remaining 300 people you took 100 heads and replaced them with tokens, which implies roughly a 33%.

每个拿到二十万或三十万薪酬的工程师，大概都有一十万的代币。

Every engineer who's getting $200 or $300 has roughly $100 in tokens.

数学上就是这么算的。

That's what the math would be.

当然。

Sure.

但销售人员用的销售工具可不便宜。

But salespeople will have sales applications that aren't cheap.

每个人都会用上智能代理。

Everyone will have agents.

我不是不想，我是说，我其实想，但你总不肯答应。

That's I don't want I mean, I would but you keep refusing to.

我想就专注在智能代理上，因为我觉得销售和客服的数学模型会不一样——那边可能用更多应用，但每个应用消耗的代币更少，当然也可能用第三方应用。

I'd like to just keep it to Agentic because I think that the math will be different in sales and customer support because I think there'll be more app with less token intensity, but maybe third party apps.

我觉得工程团队这边才更有趣。

I think engineering is where it's more interesting.

我记不清了，Uber 或者谁今天宣布，给每个工程师每月固定 1500 美元，也就是每年 18,000 美元。

I can't remember what's at Uber or someone announced today they're going to keep it to, as I say, 1,500 per engineer per month, which is $18,000 a year.

大概算百分之十吧。

Call it roughly 10%.

这应该只是个初步估算。

And that's probably a first pass swipe.

也就是说，一个年薪 20 万的工程师，能拿到 10% 的代币预算。

That says on a $200,000 engineer, they're getting 10% token budget.

你意思是，一个年薪 20 万的工程师，能拿到 33% 的代币预算？那岂不是有六万六千？

You're implying on a $200,000 engineer, you're getting 33% token budget, which would be, what's that, 66,000?

我觉得 Uber 并不是顶尖的软件公司。

And I don't think Uber is one of the greatest software shops out there.

他们就两个主要产品。

I think it has two products.

我想说的是：今年这个问题肯定会引发激烈争论。

My point is this: this is the question that's going to get litigated this year.

最惊人的是，即使只有10%，你也能达到Anthropic和OpenAI那样的发展轨迹。

The amazing thing is you can get to the Anthropic and OpenAI trajectory on even 10.

这就是我觉得这事很隐蔽的原因。

This is why I think it's secret.

是的，我觉得我们在谈两回事。

Yeah, think we're having two different discussions.

你说得完全对。

You're absolutely right.

我觉得哪怕只有10%也足够推动他们的增长了，对吧？

I don't think it will necessarily Even just 10% is enough to fuel their growth, right?

我觉得可能比10%还要高。

I think it may be higher than 10%.

如果真是33%，那就意味着两件事：IPO时不管什么价格都得买，但之后会非常艰难，因为你看到的可能是每三个、每四个岗位——按你的说法，不只是整个工程和产品开发团队，而是几乎全被替代了，还不算由此带来的额外增长。

And if it's 33, then two things are true: buy at any price in the IPO, just any price and then it's going to be pretty tough because you're going to see one in three, one in four, per your construct, it wasn't just across the entire engineering product development stack get replaced, plus or minus any growth that comes from that.

所以，明确说一句，根据我跟人聊的情况，我不觉得会高到那个程度。

So yeah, I'm not convinced it is as high as that, just to be clear from my conversations.

但我得承认，我不是专家，也不清楚工程领域当前最前沿的令牌效率到底是什么水平。

But I also admit I'm not an expert, and I don't know what absolute state of the art token efficiency in engineering looks like.

但我清楚一点：这是我最想弄明白的数字，也是我通常问每一位工程副总裁的第一个问题——你是怎么看待这个问题的？

But I do know one thing: this is the number that I most want to understand the next and it's the first question I usually ask all my VPs of engineering: How are you thinking about it?

哪些做法是有效的？

What's working?

哪些地方行不通？

What's not?

因为这个数字将决定：一千亿美元的公司，是不是已经定价到位，哪怕放缓一年消化一下也无妨？

Because this is the number that will determine: Is $1,000,000,000,000 a fully priced company that could slow down a little bit for a year while digest?

还是说，天啊，根本没人会停下来喘口气？

Or is it, Oh my god, no one's even going to pause for breath?

我们只会不停地推出新产品。

We're just gonna keep rolling this shit out.

它会吃掉三分之一的工程薪资，两年内就能把估值推到四万亿美元。

It's gonna eat one third of engineering salaries, and that's gonna get you to 4,000,000,000,000 by two years from now.

有可能。

It might.

我觉得这种像优步那样限制token的做法，只是暂时的。

I just think that this idea of capping tokens like this Uber thing, I just think it's a transitory thing.

这根本不是一种基础服务。

It's not a utility.

它不像电力，也不像办公室工位的密度。

It's not just electricity or or or our density for desks in our office.

现在这么做挺不错的。

I think that's a great thing to do now.

但到2027年、2028年底，你应该让各部门负责人自己做选择。

By the end for 2027, 2028, you should give department leaders a choice.

在好公司里，他们会选token。

And they're gonna choose tokens in good companies.

他们会更倾向于选择token而不是

They're gonna choose tokens over the

深层结构。

deep layers.

我觉得接下来会发生的是，你虽然给了他们选择权，但实际操作中，AI产品这种设计实在太狡猾了， CFO们迟早要被逼得抓狂。

And I think what's gonna happen, you're you're gonna give them a choice, but then practically I mean, the thing that is so insidiously clever about the AI products that CFOs are gonna be tearing their hair out.

它让你作为普通员工，能显得效率高得离谱，还能把一堆琐碎重复的活儿全甩掉。

It's a product that allows you, as an individual worker, to be and look wildly more efficient and have it take away a whole bunch of your grunt work.

而这些成本，根本不用你来承担。

And the cost isn't borne by you.

你是愿意花两小时死磕PPT，一点点打磨，还是直接输入Claude，说‘帮我做个这样的PPT’？

Which would you prefer, crank for the next two hours on a PowerPoint to get it just right or type it into Claude and say, make me a PowerPoint that does this, this, and this?

尤其是当这么做完全不用额外花钱的时候。

And especially if there's no trade off cost.

所以没人愿意接受限制，但每个CFO都希望有这些限制。

So no one's going to want the restrictions, but every CFO is going to want the restrictions.

围绕这个问题的博弈，会变得非常激烈。

And the dynamic around that is going to be huge.