20VC：优先认购权之争：Valar Atomics起诉Day One Ventures | Meta携Muse Wonder重返AI前沿 | NVIDIA瞄准300亿美元Reflection交易，Jev以75亿美元估值融资8.7亿美元

他身边也有过其他人。

He's had others.

要花一百亿请一个人，还让他留下来拼命干活，可不是件容易的事。

It's not easy to pay someone $10 billion and have them stick around and work that hard.

说起来容易做起来难，更难的是找到一个你真正能信任、从言行举止上都信得过的人。

It's easier said than done, and it's even harder to find someone you can truly trust in that body language.

这是让这一切运转起来的关键，对吧？

It's key to making this work, right?

我知道他身边还有其他人，像纳特·弗里曼、丹尼尔·格罗斯这些人，但找一个得力助手真的很难。

I know he's got all the other folks, the Nat Freemans and the Daniel Grosses and all these guys, but hard to get a right-hand person.

确实难。

It's hard.

所以我觉得这才是没被讲出来的故事。

So I think that's the untold story.

就像罗里说的，这可能对规模来说是个很好的交易。

To Rory's point, this may be a bit of a great deal for scale.

结果可能会是这样。

It may turn out.

当时发生的时候，我们几乎都笑了。

We kind of almost laughed at it when it happened.

不是真的笑，对吧？

Not literally, right?

结果可能会成为下一个WhatsApp或Instagram。

May turn out to be the next WhatsApp or Instagram.

但更有趣的是，我对那个我无法使用的模型并不怎么印象深刻。

But the more interesting, I wasn't that impressed with the model I can't use.

我很难太兴奋，对吧？

Hard for me to get too excited, right?

不过，它只会让Muse应用变得更好，这一点非常有意思。

The fact that it's just going to make the Muse app better, though, is super interesting.

世界变化太快了，但我认为Meta有很大机会拥有一个超级AI产品。

The world changes so fast, but I think there's a material chance Meta has just a monster AI product on its hands.

投入的资金量太大了。

The amount of investment into it is so high.

能力很强，自主性也很高。

The capabilities are so high, the autonomy is so high.

现在你还要把它增强10倍甚至50倍。

And now you're going to make it 10 or 50 times more powerful.

这可能会比ARVR眼镜那些东西做得更好。

This could do better than the ARVR goggle stuff.

现在Muse和Instant的竞争对手数量真是惊人。

The amount of Muse and Instant competitors right now is pretty astonishing.

好的。

Okay.

大科技公司开始收购了，伙计。

Big tech goes shopping, baby.

现在有两个传闻在流传。

We have two rumors that are circulating.

当然，我们不传谣言，所以我想先聊聊第一个。

Obviously, we don't deal in rumors, so I'd love to start with the first one.

Reflection，可能有些人不知道，这是一家在美国做开源模型的优秀团队，背后有红杉资本支持。

Reflection, which for those that don't know, is a fantastic team building open source models in the US, backed by Sequoia.

我记得上一轮融资大概在250亿美元左右，现在有传言说英伟达要以300亿美元的价格收购他们。

I think the last round was at $25 billion or so, and the rumor is that NVIDIA is buying them at a $30 billion price.

这还是在英伟达以120亿美元收购Poolside之后，Poolside也是做开源模型的。

This is on the back of them buying Poolside for $12 billion, which is in the open model game as well.

各位，我们怎么进一步分析黄仁勋进军开源模型领域这件事？

Guys, how do we analyze Jensen further going into the open model space?

说到进一步，你是在提醒我们他们已经收购了Poolside。

And by further, just to be clear, you're reminding us that they already bought Poolside.

他们已经有自己的Nemotron模型了。

They already have their Nemotron existing model.

我觉得答案就是，我们之前也聊过好几次了，如果你做芯片生意，而你的大客户里有些是闭源模型公司，从长远来看，如果你有一大笔钱要投入这个领域，你当然希望更多的钱留在你这里买芯片，而不是留在模型公司那边，对吧？

And I think the answer is, as we've discussed several times, if you're in the chip business and some of your biggest customers are closed source model companies, zooming out, if you've got a pile of money that's going to go into that collective pool, it's in your interest that more of it stay with you and the chips, and less of it stay with the model companies, right?

所以从定义上讲，你当然喜欢开源模型，而且只有一个芯片供应商，就像桌子另一边，模型公司喜欢闭源模型，而且每个人都自己搞芯片供应商。

So by definition, you love open source model and only one chip provider, just like on the other side of the table, the model companies love closed source models and everyone having their own chip provider.

所以我们是朋友，是最好的朋友。

So it's we're friends, we're the best of friends.

我们互相借个一千亿美元左右，但我们也在密切关注你，如果你在利润率上贪心，我们得能反击。

We lend each other $100 billion or so, but we're watching you closely, and we want to be able to fight back if you get margin greedy.

所以这一切都说得通。

So it just all makes sense.

你觉得这笔交易能成吗，Rory？

Do you think it'll go through, Rory?

当你手里有五万亿美元，想花三百亿买什么，就能花三百亿买下来。

When you've got $5 trillion, if you want to buy something for $30 billion, you can buy something for $30 billion.

但问题是，在Groq那件事之后，会不会有阻力，这种收购式招聘的模式现在还行不行，会不会有人反对？

Whether or not there'll be that whole pushback after the Groq thing, will the structure of an acqui-hire work now, or will there be pushback?

谁知道呢？

Who knows?

现在世界局势挺紧张的。

The world is pretty busy right now.

接下来的两个月，美国政府可不会闲着。

It's not like these next two months are going to be idle time for the US government.

所以谁知道呢？

So who knows?

但我觉得——

But I think that-

Rory，如果你是董事会成员，你会同意出售还是拒绝？

Rory, if you're on the board, would you say sell or would you say reject?

或者我会直接卖出。

Or I'd hit the bid.

因为这些公司都很棒，但作为独立业务，如果你有开放权重模型，你很大程度上依赖推理来变现，你会参与其中吗？

Because they're amazing companies, but the standalone business, if you have an open-weight model, you're relying a lot on the monetization is inference, are you in that?

我一直觉得开放权重很令人兴奋，但它在商业上怎么运作呢？

I've always thought open weight is really exciting, but how does it work monetarily?

如果出现这种溢价，对收购方来说价值非常高，收购方的市值也非常高。

If you're getting these kind of excess prices, the value to the acquirer is very high, and the market cap of the acquirer is very high.

这可能是千载难逢的机会。

This might be a moment in time that doesn't come back.

说实话，我也经历过这种情况。

It's also, honestly, I have been on the other side of it.

拒绝英伟达压力很大，因为你知道他们会直接跟你竞争，还会扶持你的竞争对手。

It's very stressful to say no to NVIDIA, knowing that they're going to directly compete with you and promote your competitor.

就像，"我们正在做这件事。

Like, "We're doing this.

我们投资500亿、1000亿，打造我们自己的开源权重模型，美国本土开源。

We're investing 50 billion, 100 billion in our own open-weights model, US-based open.

我们会继续做下去。

We're going to continue to do it.

这不是一次性的。

This isn't a one-off.

你有机会拿到七百万或九百万，但如果你不接受，我们会变得更大更强。

And you have a chance to take seven or nine million, but if you don't, we're going to be a bigger and better competitor."

事情没那么简单。

It's not as simple.

你可以拒绝，但确实有点吓人。

You can say no, but it is a little scary.

当你拒绝的时候，确实有点吓人，因为他们会收购其他人，然后变得越来越大。

It is a little scary when you say no, and they're going to buy other people and get bigger and bigger.

这不会改变大局，但作为创始人，当你收到那种“由我们的合作伙伴构建”的带着微笑的威胁时，确实会让人心里有压力，对吧？

It doesn't change the game, but as a founder, it certainly weighs on you when you get that "build by our partner" sort of threaten with a smile message, right?

是的，所以除非你很有信心，否则我会接受。

Yeah, so I would take it unless you're confident.

你得手里有必胜的牌。

You've got to have the winning hand.

我们不知道未来会怎样，但Salesforce宣布他们有自己的模型，这些模型基于NeMo Tron构建，并且是与NVIDIA合作开发的。

And we don't know where it's going to go, but Salesforce announced they've got their own models that are built on NeMo Tron and built with NVIDIA.

你得预测18个月后NVIDIA把所有软件工程师和整个团队都押在这上面，情况会变成什么样，对吧？

You've got to predict where this is going to be in 18 months with NVIDIA putting all of its software engineers and all of its team behind it, right?

这不仅仅是一家硬件公司，更是一家软件公司。

This is a software company, not just a hardware company.

你可以想象，软件、硬件和模型之间的护城河会变得如此强大，以至于12到18个月后，为什么你还会想用NVIDIA以外的其他开源权重或开放模型呢？

You could imagine that the moat here between software, hardware model is so powerful that in 12 or 18 months, why would you want to use any other sort of open-weight, open model other than NVIDIA's?

今天这看起来不太合理，但我能想象18个月后情况会是这样。

Today, it doesn't seem logical, but I could imagine it being the case in 18 months.

这组合简直完美。

It's such a perfect combination.

这会让Thinking Machines在440亿美元的融资轮中变得更有吸引力还是更没吸引力？他们现在正在筹这一轮。

Does it make Thinking Machines more or less attractive at 44 billion, which is the round that they're raising now?

你在这一轮里有优先认购权吗？

Do you get pro rata rights in the round?

我们心里有数吗？

Do we have a sense?

我们清楚吗？

Do we know?

这是个很好的问题。

That's a very good question.

让我查一下我的——

Let me just check my-

不然的话，我就大幅降价了，就像基思说的那样。

Otherwise, I'm lowering my price substantially, to Keith's point.

只要有一个更大的收购方愿意出更高的价，或者有一个明确清晰的商业模式就行。

As long as there's either a larger acquirer to get pay more or an obvious clear business model.

这两种情况都有例子。

There's examples on both sides of the table.

有些公司是最后才成功的，早期的人都卖了，后来赚了很多；也有些公司早期的人赚了很多钱，之后进来的人却后悔还留在里面。

There's companies who were last and made a ton after all the early guys sold, and there's companies where the early guys made a ton of money and then everyone after that's kind of sorry they're still in it.

这完全取决于具体情况，定个死规矩没什么用。

It's all so situation-dependent, it's useless to have an ironclad rule.

你只需要搞清楚这里玩的是什么游戏。

You've just got to know which the game here is.

再说，我还是坚持我的看法，我觉得Thinking Machines的收入增长得不错，但要做这个判断，你得对收入和收入的可见性有非常清晰的认识。

And again, I go back to, I believe Thinking Machines has grown nicely in revenue, but you've got to have a lot of clarity on revenue and revenue visibility to make that call.

现在，我们就等着看吧。

Now, we're just going to see.

有意思的是，我们是周二录制，周四发布，而周三Anthropic要开投资者会议，我猜第三季度的收入数据会在24小时内泄露出来。

It's interesting, we're recording on Tuesday, publishing on Thursday, and on Wednesday, Anthropic's doing its investor meeting, and I assume the Q3 revenue numbers will leak within 24 hours.

所以，这些公司并不是没有大额收入。

So it's not like there aren't big revenues attached to these businesses.

它们创造了前所未有的收入。

They're the biggest revenues ever seen.

但唯一的问题是，结果会怎么样？

But the only question is, how will it check out?

对于下一层级的公司，比如Thinking Machines，结果又会如何？

How will it check out for one level down, like Thinking Machines?

我们卖掉Memo的原因很简单，就是搞不到足够的算力，负担不起，跟不上这场军备竞赛。

The sort of pool side why we're selling Memo that we just couldn't get enough compute, we just couldn't afford it, we could not keep up with the arms race.

事情没那么简单，不是44亿对多少亿的问题。

It's not as simple as 44 billion versus whatever.

你甚至能不能拿到足够的算力来保持竞争力？

Can you even get the compute to be competitive?

我不知道。

I don't know.

你需要比我更接近内幕的人来回答，但我的直觉是，不行。我的直觉是，你根本拿不到足够的算力来保持竞争力。

You need someone that's even closer to it than I am, but my gut is no. My gut is you can't get enough compute to be competitive.

所有资源都被抢光了，对吧？

Everything's being taken, right?

连Anthropic都在找Meta，说：“Meta，能不能分我们一些算力？”对吧？

Even Anthropic's reaching out to Muse, saying, "Meta, can we get some of your computes," right?

也许他们会把算力给Thinking Machines，我不知道，但这场战争才刚刚开始。

Maybe they'll give it to Thinking Machines, I don't know, but that war has just begun.

你连内存都拿不到，更别说算力了。

You can't even get memory, let alone compute.

这不仅仅是软件的问题，对吧？

It's not as simple as software, right?

所以我也说不准。

So I don't know.

该开投资委员会了。

It's time for investment committee.

杰森，今天的议程有Jev和TypeSafe，母公司刚在推出Jev几周后，以75亿美元的估值完成了8.7亿美元的A轮融资。

Jason, on the agenda for today, we have Jev and TypeSafe, the parent company, raises an $870 million Series A at a $7.5 billion valuation weeks after launching Jev.

安德森领投，红杉参投，这是史上最大的A轮融资之一。

Andreessen led, Sequoia participating, one of the largest Series A rounds ever.

杰森，我们要以75亿美元的估值参与这轮融资吗？

Jason, will we be participating in this round at $7.5 billion?

听着，显然，我们能进第二轮已经很幸运了，因为现在有很多钱想以100亿、120亿甚至150亿的估值进来。

Well, listen, obviously, we're lucky to get into the second close because there's plenty of money that wants to get in at 10, 12, even 15 today.

所以我建议，我们连全额配售都拿不到，我建议我们投入基金能投入的最大金额，因为它的估值已经超过这个了，对吧？

So I recommend, we can't even get our full allocation, I recommend we put in the maximum amount we can out of the fund because it's already valued more than this, right?

那个估值已经过时好几周了。

That valuation is several weeks stale.

而且自从安德森正式握手敲定交易以来，Jev的使用量已经翻了两番。

And Jev usage has quadrupled just since Andreessen actually shook hands on the deal.

所以我们别无选择。

So we have no choice.

我们必须全力以赴，尽可能多拿股份。

We have to be all in and get every share we can.

我要求所有早期投资者放弃他们的按比例配售权。

I'm asking that we can get all the earlier investors to waive their pro rata.

这是我唯一的条件。

That's my only requirement.

但我们想尽可能多地投入每一股。

But we want to put in every share we can.

我们担心竞争吗？

Are we worried about competition?

我们担心Cloudflare已经推出了竞争产品吗？

Are we worried that it appears that Cloudflare already has a competitive product?

我们担心OpenAI似乎也有竞争产品吗？

Are we worried that OpenAI appears to be having a competitive product?

我们担心Anthropic甚至可能免费提供这个吗？

Are we worried that Anthropic could even give this away for free?

我们担心亚马逊已经表示他们在开发自己的Jev版本吗？

Are we worried that Amazon has suggested they're building their own version of Jev?

我们确实担心，但这样的增长正是风险投资的意义所在，伙计们。

We are, but growth like this is the point of venture, boys.

如果我们不这么做，那还不如待在家里。

And if we don't do this, we should just stay home.

我们每12到13个月就募集一次新基金。

We're raising new funds every 12 to 13 months.

这些都是巨额基金。

These are massive funds.

我们必须把它们投给赢家。

We have to deploy it into winners.

而且说实话，我们有多年的数据证明，不是每个赢家都能赢，但只有投给赢家才能赢。

And to be clear, and we have years of data on this, not every winner wins, but you can only win by being in the winners.

这就是一个赢家。

This is a winner.

一年后它还会是赢家吗？

Will it be a winner in a year?

我不知道。

I don't know.

它是当今最大的突破性赢家之一。

It is one of the massive breakout winners today.

我们仍然能以75亿美元的投资额进入。

We're still getting to invest at 7.5 billion.

我们大概可以把账面估值做到150亿，然后跟我们的LP说我们都是天才。

We can probably do a paper mark up to 15 and tell our LPs we're geniuses.

我说我们把每一股都买下来，把优先认购权给堵住，然后尽可能多地做创始人的二级市场交易。

I say we buy every share, block the pro rata, and do as much founder secondary as we can.

那个模仿秀最可怕的地方就是它简直太真实了。

The terrible thing about that parody is it's accurate as hell.

完全同意。

Totally.

那根本就不是模仿秀。

It's not a parody at all.

它让你意识到，没错，你确实可以用三分钟的片段来做这份工作。

It makes you realize, yes, you can do this job in three-minute soundbites.

你能给我讲讲Jev的上行空间吗？

Can you paint the upside for me on Jev?

它怎么才能变成一家极其成功、规模巨大的公司？

How does this become a wildly successful, massive company?

五千亿美元的愿景在哪里？

Where's the $500 billion vision?

不过我跟你说，从战术层面、从构建者层面来看，就是回到评估和Meta那些东西，Muse Turbo，这些模型没有一个能达到90分或100分。

I'll tell you, though, just at a tactical level, at a builder level, right, which is that no models are particularly, going back to the evals and the Meta whatever, the Muse Turbo, none of these hit 90s or 100.

所有这些模型都会犯很多错误，出很多问题，明白吗？

All these models make many errors and many issues, okay?

它们不可靠。

They're not reliable.

这周我能跟你说出无数的问题。

I could tell you endless issues from this week.

Jev是一个分类器，我们现在已经进入分类器2.0，我们在A阶段，它确实能用，但可能只有70%到80%的精心提示匹配能正确。

Jev is a classifier, which we're now in classifier 2.0, we're in A, and it works, and maybe it gets 70% to 80% of carefully-fully prompted matches right.

Cloudflare团队作为独立团队，还会继续投入那么多精力吗？

Will the Cloudflare team keep putting as much effort in as a standalone team?

亚马逊、OpenRouter，所有这些公司会一样努力吗？还是说这些会成为小众产品，收入不够，他们没时间投入，然后Jev这些人只是让分类器做到以前做不到的事情，获得极高的解析率，然后他们就能扩展它？

Will Amazon, will OpenRouter, will all these people try as hard, or will these be niche products with not enough revenue that they don't have time to invest in, and that these guys at Jev just make classifiers do something they couldn't do before, get incredibly high resolution rates, and then they can expand it?

现在你可以问相对复杂的问题，但只能得到二进制的答案。

Right now you can ask relatively complicated questions, but you only get binary outputs.

这就是为什么它效果这么好，对吧？

That's why it works so well, right?

罗里还是哈里，谁是我的更好的投资人？

Rory or Harry, who's my better investor?

罗里，哈里，哈里还是罗里。

Rory, Harry, Harry or Rory.

他们会变得越来越好，越来越好，而且我不知道即使现在容易的时候，Cloudflare和亚马逊的团队能不能追上。

They're going to get better and better and better, and I don't know that even if the Cloudflare Amazon teams can catch up today when it's easy.

现在要复制六个月前的OpenRouter很容易，对吧？

It's easy to build what OpenRouter was six months ago today, right?

但他们会不会持续投入呢？

But are they going to maintain that investment?

我觉得作为风投，我们必须做出这个有压力的投资决定，那就是一个来自OpenAI、非常强大且只专注这一件事的核心团队，会从今天看似平起平坐的状态中脱颖而出。

I think as VCs, we have to make this stressful investment that a core team that is very strong out of OpenAI, that is doing nothing else, will pull ahead from the seeming parity you can achieve today.

我其实愿意打这个赌。

I'm actually willing to make that bet.

因为我经常用Jev，我已经能看出来，如果它一直不变，我可能会换掉它，对吧？

Because I'm using Jev a lot, and I can already see that if it never changed, maybe I would swap it out, right?

但如果它比其他的都好，那我肯定不会换。

But if it gets better than the rest, there's no way I would change.

哪怕只是稍微好一点，你也不会再用别的了。

Even just somewhat better on its evals, you would never use anything else.

所以我现在还是会打这个赌，这个团队会专注做这一件事。

So I would still make that bet today, that this team will do nothing else.

而且我觉得推理成本不会把他们拖垮。

And I don't think the inference costs are going to destroy them.

我觉得750这个模型推理成本很低。

I think the 750, this is a low-inference model.

训练成本我不清楚。

I don't know the training costs.

我觉得750在这个领域已经能走得很远了。

I think 750 gets you pretty far in this space.

说真的，我觉得它比Thinking Machines走得更远。

For real, I think it gets you further than it gets you to Thinking Machines.

它可能给你两三年的发展空间。

It might get you two or three years of runway.

所以我谨慎乐观地认为，一个专注的团队，而不是Cloudflare那四个聪明的孩子，能够保持领先，达到以前不可能达到的准确度水平。

So I'm cautiously bullish that a focused team, rather than four smart kids at Cloudflare, can maintain its lead and get to accuracy levels, to summarize, that were never possible before.

这就是关键所在。

That's the unlock.

Jev又快又便宜，明白吗？

Jev is super fast and super cheap, okay?

而且它的准确度足以用于你大约10%的AI用例。

And it is accurate enough to use for maybe 10% of your AI use cases.

如果他们让它变得超级准确，那就会改变游戏规则。

If they make it hyper accurate, it changes the game.

这就是理论上的情况，然后回到数字上来，如果它占你总量的20%，我们大概两周前说过这个。

And that's the theory, and then just to get back to the numbers, if it's 20% of your total, we said this I think two weeks ago.

如果前沿AI推理的总市场是2000亿美元，一两年内最好能达到这个数，不然我们都有麻烦，那么20%就是400亿美元。

If total AI frontier inference is a $200 billion market, which it better be in a year or two or we're all in trouble, then 20% of that is $40 billion.

如果价格下降80%，你还能增加80亿，再增加80亿，如果他们拿到一半，那就是一个40到50亿美元收入的公司，可能值500亿美元。

And if the price of that goes down by 80%, you still add eight billion, and eight billion, and if they get half of it, that's a $4 or $5 billion revenue company, which is probably worth $50 billion.

这就是理由。

That's the case.

每一句话都说得通。

Every one of those sentences is credible.

我觉得最大的问题是其他竞争。

I think the biggest question is the other competition.

但杰森说得对，业务是存在的。

But to Jason's point, the business is there.

让我担心的是，在为这些价格辩护时，常常会说，流动性优先权只有11亿，而且这是个战略资产，团队又好，所以下行保护就是他们会被一个好的收购方买走。

What worries me in every justification for a lot of these prices is, wow, the liq pref's only a billion one, and it's such a strategic asset with a good team that the downside protection is they'll just get bought by a great acquirer.

我每次都能听到这种说法。

I just hear this every single time.

也许这是真的，但并不是所有东西都有被收购作为下行保护，对吧？

And maybe that's true, but not everything has a downside protection of being acquired, does it?

在它失效之前确实如此，这话说得有点轻巧。

It does until it doesn't, and that's meant to be glib.

但在一个拥有12家市值超过万亿美元的公司、而且它们都热衷于收购、都在拼命争分夺秒地建设公司、尤其是法律窗口还开着的情况下，肯定会有很多收购发生。

But in a world where you've got 12 companies above a trillion dollars, all of whom are acquisitive, all of whom are in a wild tearing hurry to build companies where time is money, and especially if the legal window is open, then there's going to be a lot of acquisitions.

这个逻辑是成立的。

And the logic is true.

如果市场放缓，那么所有这些事情都会同时走向糟糕的境地。

If the market slows down, then all this stuff will go in a bad place at the same time.

所以，在它不成立之前，它都是成立的。

So, it'll be true until it's not.

我觉得这种说法有点夸大其词了。

I think it's overstated.

我得说，现在这个清算优先权在Jev这里，比2021年那一千个独角兽的一千个清算优先权值钱多了。

I will say today it feels like a liquidation preference in Jev is much more valuable than 1,000 liquidation preferences in 1,000 2021 unicorns.

那些里面很多最后都会一文不值。

Many of those will end up being worthless.

所以我们夸大了清算优先权的价值，对吧，就是为了让交易能通过委员会。

So we overstate the value of the liquidation preference, right, to get the deal through the committee.

但今天它们确实有实际价值，对吧？

But they have real value today, right?

我只是觉得，与其要2021年的那些，我宁愿要这个清算优先权。

It's just I'd rather have this liq preference than quite a few from 2021.

好的，Jason，你提到了2021年。

Okay, you said 2021 there, Jason.

2021年，有34家十角兽公司。

2021, there were 34 decacorns.

现在，2026年，我们创纪录地有42家初创公司成为了十角兽。

We now have a record 42 startups that have become decacorns in 2026.

就像辛普森动画里说的，到目前为止2026年就是这样，对吧？

As they say in the Simpsons cartoon, that's 2026 so far, right?

还剩一个季度。

With a quarter to go.

对，2021年有34家十角兽公司，对吧？

Yeah, there was 34 decacorns in 2021, right?

所以换句话说，现在的十角兽公司比上次情况糟糕透顶的时候还多。

So in other words, it's more decacorns than the last time things went horribly wrong.

这是最明显的描述。

That's the first obvious description.

换句话说，因为如果你回头看那些十角兽公司，Mira在2021年就是一家十角兽公司。

In other words, because if you look back on those decacorns, Mira was a decacorn in 2021.

Airtable也是一家十角兽公司。

Airtable was a decacorn.

我觉得它们不会一样，但值得提一下。

I don't think they'll be the same, but just worth pointing out.

我觉得更有趣的统计数字是这个。

I think the more funny statistic is this.

在公开市场上，总共只有98家十角兽公司。

There are, in the public markets, a grand total of 98 decacorns ever.

如果你按十年来看，有个吓人的词。

And if you look at them by decade, there's a scary word.

就是说，从今年年初算起，把私人和公开的十角兽加起来，过去五十年里我们大概每十年产生20家左右，上下浮动。

It's like, and you take the private and the public decacorns as of the start of this year, we roughly generated 20 plus or minus decacorns per decade for the last five decades.

而我们今年产生了，你刚才说多少？

And we generated, what did you say?

不到一整年，就产生了42家十角兽。

42 decacorns in less than a full year of this decade.

当然，有些公司的市值还在不断上涨，这些因素都对你有好处。

Now, there is some, obviously, market caps continually go up, all those things point in your favor.

但想想每十年才20到30家，而不到一年就冒出42家，这确实令人震惊。

But it is stunning to think of 20 to 30 decacorns per decade, and then 42 decacorns in less than a year.

而且只有98家上市公司。

And that there's only 98 public companies.

我刚查了一下。

I just looked it up.

超过100亿美元的有98家，超过1000亿美元的还有30家。

98 above $10 billion, another 30 above $100 billion.

所以自美国资本主义诞生以来，科技领域总共有130家公司市值达到或超过100亿美元，而今年又增加了42家。

So 130 companies in total worth $10 billion or more since the dawn of, in tech, since the dawn of American capitalism, and a further 42 this year.

说到点子上了，你基本上是在说我们现在处于泡沫之中。

Scratching the obvious itch there, you're basically saying that we're in a bubble.

我不知道我是不是在说——你是想让我承认我们处于泡沫之中。

I don't know if I'm saying-- You're trying to make me say I'm in a bubble.

我只是觉得，嗯，我们有优先股，就像Jason指出的那样，所以我们都没问题。

I just think that, well, we've got the pref, as Jason points out, so we're all fine.

但可以肯定的是，十角兽的产生速度比我们之前看到的任何时期都要快，包括过去五十年公开市场的所有数据。

But what you can say for certain is the production rate of decacorns is higher compared to anything we've seen before, including five decades of data in the public markets.

这绝对是自工业革命以来最具经济变革意义的事件——我们是不是有点太超前了？

It definitely, either this is the most economically transformative event- Since the Industrial Revolution, are we getting a wee bit ahead of ourselves?

你来选。

You pick.

罗里，你会选哪个？

Which would you pick, Rory?

也许更好的答案是：

Maybe the better answer is this.

我其实想回到优先股那个答案，对吧？

I'm actually going to go back to the pref answer, right?

有点好笑的是，你得记住所有这些私营公司的估值，那个估值是由持有清算优先权的人设定的。

It's kind of funny because you have to remember all these private company valuations, that valuation is set by someone who does have a liq pref.

一旦你有了清算优先权，并且你相信它存在，你就可以对自己说，不像普通股那样每天交易，你可以说它是1倍加期权价值。

And once you have a liq pref and you believe it exists, you can say to yourself, unlike a common stock where it trades every day, you can say it's a 1X plus option value.

而现在的问题是，期权价值很高，因为波动性很大。

And the thing right now is option values have high value because there's a lot of vol.

看起来这个十年里会有三家公司市值达到一万亿美元。

Some apparently three companies in this decade will become worth a trillion dollars.

很多公司会达到一千亿美元，哪怕只有十家。

Many will become $100 billion, even if it's 10.

说这家公司有非零概率变得超级厉害，而我的下行风险只是1倍，这是合理的。

It's rational to say there's some non-zero probability this company becomes freaking amazing, and my downside is a 1X.

具体来说，我不认为所有这些股票如果都在公开市场交易，都能达到100亿或更高的估值。

To make it concrete, I don't think all these stocks would trade at 10 billion or more if they all traded on the public markets.

这还是没有回答我的问题。

That's not giving me an answer.

我觉得回答了。

I think it is.

好吧，我们都在赌这是真的，对吧？

Well, we're all betting this is true, right?

Menlo在募集上一只基金的时候算过一笔账。

Menlo had some math when they raised their last fund.

我记得他们募集了一个30亿美元的基金，要在接下来的12个月里投出去，他们赌的是至少会有四笔超过250亿美元的退出，才能让这个基金回本，对吧？

I think it was a $3 billion fund to deploy the next 12 months, that they're betting they'll have at least four $25 billion-plus exits to make the fund, right?

这就是数学。

That's just the math.

不复杂，对吧？

It's not complicated, right?

他们能像钟表一样准时做到这些。

That they can do that like clockwork.

这就是他们的赌注。

That's the bet.

AI的投入和最近几个月的退出案例证明了这一点。

The AI spend and the recent exits of the last X months justify it.

但历史证据让这个赌注显得可疑。

Historical evidence makes it suspect.

所以你是活在当下，还是以史为鉴？

So are you living the moment or are you living as a student of history?

我只是觉得在风投行业，做历史的学生没什么太大好处，虽然我们节目里就有一位，我对此很感激。

I just don't think in venture there's a lot of upside in being a student of history, even though we have one starring on the show, which I'm grateful for.

但我不确定这能不能帮你拿到大额的收益分成。

But I don't know if it gets you the huge carry checks.

实际上，当我看这些数据的时候——是我一个同事为了别的事情整理的——我觉得真正需要做的是，你对自己说，好吧，如果外推一下，你假设需要多高的增长率才能让这个数学成立？

And in fact, as I looked at this data, which one of my colleagues was doing for something else, I think what I actually realize you need to do is then you say to yourself, okay, if you extrapolate, what level of growth are you assuming to make this math work?

因为，是的，杰森，你就是这样。

Because, yes, Jason, you are.

换句话说，因为每年、每个十年看起来都比上一个十年更大，除了70年代，因为股票在涨。

In other words, because every year, every decade it looks bigger than the last decade, with the exception of the 1970s, because stocks go up.

但问题是，你提到过，四个250亿以上的退出，而在一个世界里——

But the question is, you made a comment, four 25 billion-plus outcomes, and in a world-

每个基金。

Per fund.

每个基金，对。

Per fund, yeah.

而且历史上超过100亿的也就只有130家。

And there are only 130 above 10 billion ever.

所以超过250亿的可能也就70家。

So there's probably only 70 above 25.

所以你得再找到四个新的。

And so you got to get four new ones.

这是可行的，但这里你容错率很低，兄弟。

It's doable, but you can't miss a lot here, dude.

但这就是安德森做得对的地方。

But that's what Andreessen got so right.

人们，在我们开始投资的时候，或者可能在我投资生涯之前，早期人们嘲笑安德森，说他出价过高，做得不对，还说“我们必须投中所有赢家”。

People, when we started our investment, well, and maybe pre my investing career, but people mocked Andreessen in the early days for overpaying and doing this wrong and saying, "We have to be in all the winners."

他们就是算了这笔账。

And they just did the math.

是的，你必须投中所有赢家。

Yes, you have to be in all the winners.

而且即使是很多赢家，你也会因为出价过高而赚不到钱，但超级赢家会弥补这一切。

And even a lot of the winners you won't make money in because you'll overpay, but the super winners will make up for all of it.

这很简单。

It's just basic.

那些嘲笑他们的人，结果完全错了。

And all the yahoos mocked them, and they were completely right.

为赢家多付钱的代价，就是能进入超级赢家。

The price of overpaying for a winner is being in the super winner.

说到超级赢家，Jason，你想回到投资委员会吗？

Well, talking of super winners, Jason, do you want to come back to the investment committee?

我们可以做一个DeepMind。

We can do a DeepMind.

Demis Hassabis，传奇人物Demis Hassabis，几个月前从Thrive融资，估值大约100亿美元。

Demis Hassabis, the legend that is Demis Hassabis, raised at about $10 billion, give or take, from Thrive matter of months ago.

现在，他的公司Isomorphic Labs正在洽谈，估值在400到500亿美元之间。

Now, Isomorphic Labs, his company, in talks to raise it between $40 to $50 billion price.

我们要这么做吗？

We going to do this?

一代天才，就要付出一个时代的价格。

Generational talent is generational pricing today.

尽管我非常相信Jev团队和TypeSafe团队，但他会组建一个自人类诞生以来在人工智能领域从未见过的团队。

As much as I believe in the Jev teams and the TypeSafe teams, he's going to pull together a team like's never been seen since the dawn of mankind in artificial intelligence.

也许他会失败。

Maybe he fails.

也许他已经受够了。

Maybe he's had enough of it.

也许他在伦敦的生活太享受了，太多高档俱乐部了。

Maybe he's enjoyed life in London too much, too many posh clubs.

但我认为我看到了那股热情。

But I think I've seen the fire.

这是一场令人害怕的赌注。

It is a scary bet.

我们有清算优先权，没人能再组建这样的人才团队了。

We have our liquidation preference, and no one will be able to assemble a set of talent like this.

这是挑选最顶尖人才的最后机会，也是我们的使命。

This is the next and last chance to cherry-pick the best of the best of the best, and this is the job.

这家公司有意思的地方在于，和大部分收入来自基础模型、靠智能自动化替代劳动力不同，这里的价值捕获不是替代劳动力。

The interesting thing about this company will be that unlike the vast bulk of the revenues that come from the foundation models where it's intelligence automation, where the value capture is replace labor, here, the value capture is not replace labor.

而是真正做出科学发现。

It's actually make science discoveries.

因为当你发现新分子时，如果只是自动化那5亿美元的科研投入，付给科学家的工资，那根本不够。

Because when you discover, when you do a molecule, if you automate the $500 million of science spend that you pay science wages to, that's not really going to cut it.

你真正想要的是，当你创造出200亿美元的药物时，能按版税拿到其中的10%。

What you really want to do is when you create the $200 billion drug, you want to get 10% of that on a royalty basis.

所以，随着模型吞噬掉所有可以直接替代的知识工作者岗位，它们最终必须通过用智能创造全新事物来实现变现。

So it's kind of the, as the models devour all the knowledge worker dollars that are available to them in terms of direct replacement, at some point, they're going to have to monetize via creating whole new things with their intelligence.

而且现在这听起来可信多了，尤其是考虑到他们最近在数学方面的工作量，正如一位数学家指出的，解决斯托克斯-纳维方程之类的问题，其实赚不了多少钱。

And it's a lot more credible today just given the amount, the math, the recent work they did on math, which as one mathematician correctly pointed out, you don't actually get paid a whole ton for solving Stokes-Navier, whatever it is.

但如果你能发现下一个GLP类药物，那就是大赢家，因为那些药真的很赚钱。

But if you can discover the next GLP drug, that's a winner because those things do money.

所以这值得关注。

So it will be interesting to see.

我以前说过，药企在还要做一期、二期、三期临床试验的情况下，很难拿到有意义的专利费。

I have in the past said it's hard to get the drug companies to extract meaningful royalties when they've still got to do phase one, phase two, phase three.

而且很多这类公司最后自己都变成了药企。

And many of these companies end up being drug companies down the line.

看看这里会发生什么会很有意思。

It'll be interesting to see what happens here.

但如果你能不只是从劳动力成本里榨取价值，而是真正拿到那部分收益，那可是一大笔钱。

But it's a big pile of money if you can extract the value, not just from a labor percentage.

对，对，我觉得几年后我们都会用上十种类似GLP-1的药。

No, no, I think we're all going to be on 10 GLP-1-like drugs in a few years.

而且我觉得我们从来没真正了解过Anthropic那个湿实验室的完整情况。

And I think we never really got the full story on Anthropic's wet lab.

我觉得这里的潜力比大家知道的还要大。

I think the potential here is even greater than is let on.

我觉得药物发现的潜力可能非常巨大。

I think the potential for drug discovery is probably massive.

而且我们正从那些怪人微剂量用药，发展到发现我们都会有一批药物，能大幅延长寿命、提高生活质量，而他们可以快速推进、拥有知识产权，让别人去做那些辛苦的工作。

And we're going from weirdos microdosing to finding out we're all going to have a set of drugs that dramatically extend longevity and quality of life, and they can radically accelerate it, own the IP, and have other people do the hard work.

也许AI真的能相当快地重塑整个行业。

Maybe AI really can reboot the entire industry fairly rapidly.

当然，不是全部，不是临床试验，也不是其他部分，但天哪，如果能提前五年推出一个GLP-1药物。

Of course, not all of it, not the trials, not the rest, but man, if you could get a GLP-1 out five years faster.

Jason，你真的相信这个吗？

Jason, do you actually buy that?

我不是在讽刺。

I don't mean that cynically.

你是说药物那部分？

The drug stuff?

是的，我也是最近才相信的。

Yeah, only very recently did I buy it.

我之前觉得那都是些怪人、布莱恩·约翰逊之流，还有卖假货的。

I thought it was wackos and Bryan Johnsons and people selling crap.

是什么让你改变了想法？

What caused you to change your mindset?

因为GLP-1的医学证据太充分了。

Because the medical evidence for GLP-1s is so profound.

对心脏病，对很多其他疾病都是如此，而且我觉得我们定制药物来延长寿命的能力，那才是巨大的市场。

For heart disease, for so many other things, and I think our ability to tailor drugs to extend longevity is the massive market.

这比推理还厉害。

It's better than inference.

而且你随便跟一个前沿的医生聊聊，他们都会跟每个人说，比如你髋关节有问题，就说“别把另一边也弄坏了”。

And if you talk to any just physicians at the cutting edge now, they're all saying to anybody, like your hip or whatever, "Just don't hurt the other one."

再过几年，我们能实现的寿命延长幅度会大得多，所以我觉得，虽然旧金山的那些怪人、微剂量服用者和布莱恩·约翰逊让我一度怀疑，但压倒性的证据表明，几乎每个人都应该用GLP-1，对吧？

In another couple of years, the amount of longevity we're going to be able to achieve is so much higher that I just think the wackos or the micro-dosers of SF and Bryan Johnson made me skeptical, but just the overwhelming evidence of the benefits that almost everyone should be on a GLP-1, right?

而且我们会越来越擅长搞清楚这些东西。

And the fact that we're going to get better and better at figuring all this stuff out.

我们就是一堆癌症、心脏病和其他问题的集合体，解决这些重大问题将极大地延长寿命。

We're just walking sets of cancer and heart disease and other things that having material impacts will radically extend longevity.

在充分识别分子并了解其作用机制之前，还有一整套步骤要走，之后还要经历各种不同的生物阶段。

There's a whole sequence of steps before you even have adequately good molecule identification with some understanding of how it works, and then you still have all the different bio phases.

所以我说，我确实认为要从中获取价值会比你想的更难，但我很高兴他们在尝试。

So as I say, I do believe it'll be harder to capture value than you think, but I'm really glad they're trying.

各位，我想聊一个可能比平时早一点的话题，我们有个叫Makka的公司，就是做机器人数据、物理世界数据的，这些数据通常用于机器人领域，它从红杉资本融了6000万美元的B轮。

Guys, I want to talk about one which may be a little bit earlier than usual, but we had Makka, which is the Macaw or the surge for robotics data, physical world data that can be implemented in, most often, robotics, raises 60 million Series B from Sequoia.

我提这个的原因是，我们在欧洲有MicroAGI，它很快就要冲到1亿美元了。

The reason I wanted to bring it up is that we have MicroAGI in Europe, which has posted race to 100 million very quickly.

还有Makka，它的年经常性收入已经远超1亿美元了。

There's Makka, which is well over 100 million in ARR.

还有另一家供应商，我忘了名字，好像是Exosoft之类的，也超过了1亿美元。

There's another provider that I'm forgetting the name of, Exosoft or whatever it is, that's again over 100.

而且我们看到，数据提供商的规模超过一亿美元的比比皆是。

And we just see the sheer volume of data providers over 100 million.

我们应该从中读出什么、推断出什么、分析出什么呢？

What should we read, infer, analyze from this?

从整体数据来看，你可以推断出，Anthropic和OpenAI在训练数据上花费了大量资金，其他三家想成为基础模型的公司也是如此，而且整个行业都在努力满足这一需求。

You can infer from data in general that Anthropic and OpenAI spend a large amount of money on training data, as do the other three wannabe foundation models, and there's an entire industry that suffers to take care of that.

这是一个有趣的行业，因为只有五六个大型竞争对手，他们显然都想向企业销售产品，但绝大部分收入来自相对少数的大型企业。

It's an interesting business because you only have five or six large competitors, and they all obviously want to go on and sell to enterprises, but the vast, vast bulk of the dollars come from a relatively small number of enterprises.

这些公司增长非常快，因为基础模型公司需要数据。

Those companies are growing very quickly because the foundation model companies want data.

他们现在就需要，而且愿意为此大笔投入。

They want it now, and they're willing to spend big to get it.

所以他们能够从零迅速增长到数亿，有些甚至达到数十亿。

So they've been able to go from zero to hundreds, and in the case of some, billions very quickly.

而在基础模型的情况下，有趣的是，显然有一个最终的收入来源在为此买单。

And the interesting thing in the foundation model cases, there's clearly an end revenue source that's paying for that.

在机器人数据这块，一切都说得通。

In the case of robotics data, it all makes sense.

道理是一样的。

It's the same analogy.

有一堆公司在做各种机器人模型，挺有意思的。

There's a bunch of companies really interesting doing various forms of models for robotics.

显然，这还处于部署生命周期的早期阶段。

It's obviously early in its deployment life cycle.

所以你现在看到的是早期增长的开支，有1000万、1亿美元的交易，这些可以从机器人公司拿到风险投资，还有基础模型公司，因为他们也想进入机器人领域。

So you've got the early burgeoning spend, you've got the 10 million and the $100 million deals, and you can get that on venture dollars from the robotics companies and from, yeah, and foundation model companies too because they want to be in robotics.

我觉得要把它扩大到10亿或20亿，得等到实际部署的时候，而时间点还不确定。

I think what's going to scale that to a billion or 2 billion is when the deployments happen, and that's still unproven in terms of timing.

因为人形机器人，我见过GXO，那是一家物流公司，我的一家公司跟他们有合作。

Because the humanoids, I saw a GXO, which is a logistics company that one of my companies deal with.

他们那边有个人给我发了个帖子。

They sent me a post, one of the guys there did.

那篇帖子的总结就是，42个人形机器人试点项目，零个真正部署。

And the summary of the post was, 42 POCs of humanoids, zero deployed.

然后还有一些关于宝马和Figure合作的讨论，我们又是说得天花乱坠，但人形机器人根本没上线。

And then there was some thread about the BMW relationship with Figure where, again, we talked a big game, but the humanoids aren't out there.

所以我确实觉得机器人领域存在部署问题，这个问题会向上游传导，最终影响数据业务的支出，直到部署层面的瓶颈被突破。

So I do think there's a deployment issue in robotics which will filter back upstream to ultimately spend on data business until such time as the unlock happens at the deployment level.

但就目前而言，确实是趁着阳光正好，赶紧晒草。

But for now, definitely making hay while the sun is definitely shining.

我们是不是要按投资委员会的意见来做Makka？

Are we doing Makka from our investment committee points?

那你得先完成红杉第二轮融资，还要满足监管要求的20倍估值提升或者他们觉得合适的条件，所以不行。

Well, you have to do the second half of the Sequoia round that the regulatory 20X step-up or whatever it is they deem to be appropriate, so no.

好吧，说实话，我还没聪明到能真正搞明白。

Well, listen, I'm not smart enough to really know.

我确实觉得这是个有意思的布局，因为显然现在对机器人的兴趣已经达到了历史最高点，对吧？

I do think it's an interesting play in the sense that obviously the current level of interest in robotics is beyond an all-time high, right?

而且他们都需要训练数据，对吧？

And they all need training data, right?

你可以对其中一些人持怀疑态度，但对高质量训练数据的需求，哪怕只是花钱让人穿戴传感器去工作，也是高得离谱。

And you can be skeptical of some of these folks, but the demand for high-quality training data, even if it's just paying people to wear sensors and do work, it's just off the charts.

我在YC演示日遇到了一位创始人。

I met a founder at YC Demo Day.

他不是这一届YC的，只是我在别的地方认识的，他把公司转型成专门做一种非常细分的训练数据，结果60天内就赚了3000万。

He wasn't in this class of YC, he was just there that I knew from faster stuff, and he tilted his company into doing just a very niche version of training data, and he did 30 million in 60 days.

他说问题在于，他们根本不想做他要求做的那些事情。

And he said the problem is they just don't want to do all the stuff that he's asked to do.

他说：“我们有3亿的需求。

He's like, "We have 300 million of demand.

做这个细分产品需要很多国家的很多专业人才。

It just takes so many specialized humans in so many countries to do this niche product.

我们拒绝了270个，只接了30个，因为我们不想做，”对吧？

We're turning away 270 out of 30 because we don't want to do it," right?

所以如果你能搞定那个二维组合，既能提供数据标注，又能提供人力，还有所有其他东西，那你就找对方向了。

So if you could nail that two by two of being able to provide the data labeling, the humans, and the everything, you have the right one.

也许我更愿意做这个，而不是投资一堆机器人，那些机器人可能会把贝西阿姨扔出门外，或者在抓锅的时候一拳砸穿我的厨房。

Maybe I'd rather do this than invest in a bunch of robots that are going to throw Aunt Bessie out the door or smash their fist through my kitchen when they're trying to grab the pots.

我肯定对家用机器人持怀疑态度。

I'm certainly skeptical of the residential ones.

也许这才是我们最安全的投资方向。

Maybe this is the safest place to put our money.

天哪，我宁愿做这个，也不愿碰任何机器人公司。

Oh my God, I would way rather do this than any robotics company.

但讽刺的是，如果机器人公司不成功，这些东西根本就不会存在。

But the absurdity of that is these things aren't going to exist unless the robotics companies are successful.

所以我不同意。

So I disagree.

嗯，不是，我是说我不知道哪家机器人公司会成功。

Well, no, because I'm saying I don't know which robotics company will be successful.

我不知道最终会是什么形态，也不知道会用在哪个领域，但我相信机器人技术一定会发展起来。

I don't know what form factor, I don't know what vertical, but I believe that robotics will be.

所以我想投身于基础设施，这样无论哪家公司成功，我都能参与其中。

So I want to be in the infrastructure that allows any of them to be.

但我要反驳一下，这就像说我想做Scale.ai或者Merco这样的公司，而不是Anthropic，虽然它们都是很棒的公司。

But pushing back on that, that's like saying I want to do scale.ai or Merco, which are great companies, instead of Anthropic.

如果让我选，我会选Anthropic。

If I had a choice, I'd do Anthropic.

如果你能选赢家，那就选赢家，对吧？

If you can pick the winner, pick the winner, right?

因为做模型的公司，从定义上讲，规模肯定比那些卖东西给模型公司的公司要大。

Because the model companies are by definition bigger than the companies that sell to the model companies.

抱歉，我没太听懂你这个类比。

Sorry, I didn't quite understand that analogy.

这有点像说我想做Scale而不是Lagora，因为我不知道谁会在应用层胜出，但我知道模型肯定需要不断进步。

It's a bit like saying I want to be in scale, not Lagora, because I don't know who's going to win the application layer, but I do know that models are going to need to get better.

这个类比更贴切，也更容易理解。

That's a better comparable, which would be understandable.

我个人——

I personally-

我明白。

I understand.

你话里的隐含意思是，你认为每家机器人公司都有自己的基础模型，像我们所在的Journalist Pi和Skilled这样的公司。

What you're saying implicit in that is you believe there's a statement in that, that each robotics company has its own foundation model, and the companies like we're in the journalist Pi and Skilled.

如果Journalist Pi和Skilled成为机器人领域的主导基础模型，那这就跟我之前说的基础模型观点一致了。

If journalist Pi and Skilled become the dominant foundation models for robotics, then it is equivalent to my foundation model comment.

但如果每家机器人公司都得自己开发专属模型，那模型层就会变得非常碎片化。

But if every robotics company has to build its own specific model, then yes, then it becomes a Very fragmented market at the model layer.

这我就不确定了。

And I don't know.

这绝对是我投资机器人领域时，投入不少却最看不清的一个方向。

It's definitely one of the areas where I have a fair number of robotics investment and the least amount of clarity.

好的，各位。

Okay, boys.

还有很多话题要聊。

There's a lot else.

你们想聊什么，我来引导。

Guide me in terms of where you would like to go.

我们可以聊聊Anthropic把Haiku的价格砍了90%，这标志着和OpenAI的价格战进一步升级。

We can talk about Anthropic cutting the price of Haiku by 90% as the price war with OpenAI further intensifies.

Jason，我挺想听听你对这个的看法，毕竟你用它们用得那么频繁。

Jason, I'm relatively intrigued to hear your thoughts on that, given how intensely you use them.

你看，我知道我们在节目里已经聊过很多这类话题了。

Look, I know we've talked a lot about this kind of stuff on the show.

但我要说，这个话题其实聊得还不够。

I will say it's under-discussed.

这些模型，Haiku 5.5和Sonnet 5.5，它们都很棒，而且在我看来，它们真的是LLM和AI领域里第一个真正带来通缩效应的东西。

These models, Haiku 5.5 and Sonnet 5.5, they're great, and they're really, from my perspective, the first thing that's actually been deflationary with LLMs and AI.

其他东西都没有降价，因为不管你怎么说token价格在降，我们消耗的token数量太多了，根本享受不到这个好处，对吧？

Nothing else has been deflationary, because no matter how many you talk about token prices falling, we've been consuming so many more tokens, we're not getting the benefit of it, right?

我们投资的公司里，没有哪家看到AI成本下降的，对吧？

None of our portfolio companies are seeing their cogs fall in AI, are they?

确实没有，对吧？

Not really, right?

我们都消耗了更多的token，用了更多最先进的模型，所以根本没看到什么降价的好处。

We've all consumed so many more tokens and used more and more of the state-of-the-art frontier models that we haven't seen inflationary benefits.

在我看来，这些是第一批真正降价的。

These, to me, are the first ones that are deflationary.

Haiku，基本上，在节目之前，一直都有那个N减一的模型，它不是最前沿的，你只能拿到去年的旧款，来配今年的新款。

Haiku, basically, since before the show, there's always been the N minus one model, which isn't quite the frontier one, and you get last year's Cadillac and fins for this year's shiny new.

那还算值得。

It was sort of worth it.

但现在，Haiku 5.5，简直是个笑话。

But now, Haiku 5.5, which is a joke.

如果你用Haiku 5.5，开发者会笑话你的。

Developers would laugh at you if you used Haiku 5.5.

在很多用例上，它几乎和Sonnet一样好，但价格只有它的十分之一或二十分之一。

It's almost as good for a lot of use cases as Sonnet, and it's a 10th or the 20th the price.

这太震撼了。

It's like shock.

太便宜了。

It's so cheap.

Haiku 5.5是第一个推理模型，你可以调节推理量，而且几乎不用在乎成本，就像Jev说的那样。

Haiku 5.5 is the first reasoning model, and you can dial up how much reasoning you want, that you can almost not care about what it costs, like Jev.

以前我们没遇到过这种情况，这对我来说很管用。

We haven't had this before, that worked for me.

以前我们没有一个推理级别的模型，让你不用为每个token、每块钱操心，因为它只有Sonnet 4.5价格的十分之一。

We haven't had a reasoning level model where you didn't have to stress every token, every dollar, because it is a 10th the price of what Sonnet 4.5 was.

价格只有十分之一。

It's a 10th the price.

所以我觉得这挺颠覆性的。

So I think it's disruptive.

顺便说一句，Sonnet 5.5也挺有意思的。

Sonnet 5.5 is interesting as an aside.

至少在我做的测试里，它几乎和之前的Opus模型一样好。

It seems, at least in the testing that I've done, it's almost as good as the Opus models before.

但有意思的是，它只比之前的Sonnet便宜大约10%，但速度快了一倍。

But what's interesting, it's only about 10% cheaper than the previous Sonnet, but it's twice as fast.

而且这些评估很少提到速度。

And these evals don't talk enough about speed.

这就是为什么很多这类东西没什么用，比如Groq和其他一些，因为它们太慢或者会超时。

This is why a lot of these things are useless, like Groq and others, because they're too slow or they time out.

速度，如果你的聊天能快一倍，那可是大事，对吧？

Speed, if you can have your chat go twice as fast, it's a big deal, right?

而Sonnet 5.5速度快一倍，便宜10%，虽然不多，但更好。

And Sonnet 5.5 is twice as fast, 10% cheaper, not that much, and better.

速度也带来了通缩效应，因为你能更快地完成更多事情。

Speed is also deflationary because you can do more and more quickly.

所以对我来说，这两个加上杰文斯，这是我们第一次看到通缩。

So for me, these two plus Jev, the first time we've seen deflation.

是的，你会有杰文斯悖论，所以我们当然会使用更多。

And yeah, you're going to have your Jevons paradox, so we're going to use even more, of course.

但这是第一次你可以直接进去，换掉模型，在一周内获得更快的速度和更低的成本。

But this is the first time you could literally go in, swap out models, and have much more speed and pay a lot less in a week.

这很令人兴奋，我觉得对B2B领域来说尤其如此，因为我们很多B2B应用还在做这些权衡，对吧？

It's exciting, and I think it's most exciting for B2B stuff, where we're all still making these trade-offs for a lot of B2B apps, right?

我们在努力让解决方案变得更好。

We're making it how good can we make the resolution?

我们能多好地解决这个问题？

How good can we resolve this issue?

我们能多好地完成这个工作流程？

How good can we do this workflow?

如果速度快一倍，而且有些情况下价格只有原来的十分之一，那下次董事会上我们肯定得看到不少成果。

If it's twice as fast, and in some cases, a tenth the price, we sure better see a lot of results at the next board meeting.

伙计们，别再找借口了。

No more excuses, boys.

这是我见过的第一个在实用层面上具有通缩效应的东西。

It's the first one at a practical level I've seen that's deflationary.

我觉得接下来有趣的是，不管叫它杰文斯效应还是价格弹性，因为价格弹性就是说，价格按比例下降时，总的美元支出是上升还是下降。

And I think the next interesting thing will be, call it Jevons, call it price elasticity, because price elasticity simply says, for a proportional reduction in prices, does the net dollar spend go up or down?

显然，他们押注的是高度弹性，支出会下降，但最终你会花得更多。

And the bet, obviously, they're making is it's highly elastic, and it goes down, and you end up spending more.

我们拭目以待。

And we'll see.

这显然就是他们的赌注。

That's obviously the bet.

如果产品需求缺乏弹性，你可不想降价，因为那样就会面临经典问题：高固定成本、需求增长缓慢、价格还更低，这简直糟糕透顶。

You would not want to be cutting prices in a demand inelastic product, because then you'd have the classic problem of high fixed costs, low demand expansion, and lower prices, which is terribly, horribly bad.

所以这里的赌注，当你看到这些降价时，实际上就是这些东西在需求方面非常有弹性。

So the bet here, as you see these price cuts, is in fact that this stuff is very elastic in terms of demand.

我猜我们很快就能在S1里看到结果。

And I assume we'll see it in the S1 soon enough.

听着，这些大语言模型的会计处理会很有意思。

Listen, all the accounting for these LLMs is going to be interesting.

但我想，当你使用Haiku或Sonnet，也就是基本上用N减一模型时，唯一的成本就是推理，对吧？

But I assume when you're using Haiku or Sonnet, when you're using essentially the N minus one model, the only cogs is going to be inference, right?

你基本上可以把训练成本一笔勾销，因为你可以把那些都算到前沿模型上，比如Fable和Opus，对吧？

You can basically write off the training cost because you can assign that all to the frontier models, to Fable and Opus, right?

如果真是这样，那基本上就是双赢的局面。

Assuming that's the case, it's sort of a win-win all around.

因为我相信Anthropic的推理利润率相当高。

Because I believe that Anthropic's inference margins are quite high.

我确信这一点。

I believe it.

但我相信他们实际的毛利率仍然相当糟糕，因为训练成本太高了。

But I believe their real-world gross margins are still pretty crappy because of training costs.

所以我觉得看看他们怎么拆分这些成本会很有意思。

So I think it'll be interesting to see how they break this out.

但我觉得这很令人兴奋。

But I think it's exciting.

不过是的，我们只会用得更多。

But yeah, we'll just use more.

会有很多这样的情况，就像你不用付热水费，就可以洗个二十分钟的澡。

There's going to be a lot of stuff where it's like taking a 20-minute shower when you don't have to pay for the hot water.

你永远都不想出来，对吧？

You never get out, right?

提醒一下，我同意我们会用得更多，也同意你的评估，但当你把价格砍掉90%时，光说我们会用得更多是不够的。

Just as a reminder, and I agree we will, and I agree with that assessment, but when you cut prices by 90%, we're going to use more doesn't cut it.

我们得用十倍以上才行。

We got to use 10 times more.

否则，需求弹性小于1，你的实际收入就会下降。

Otherwise, demand elasticity is less than one and your actual revenue goes down.

我确实认为我们会用10倍那么多，因为接下来会发生的是，你不会再只是试一下，而是会全部用上。

I do think we'll use 10 times as much because what's going to happen, instead of sampling something, you'll just do it all.

这正是我同事在提到使用Jev时说的。

That was what my colleague said on using Jev.

你会在数据库的每一场战争中都用到它。

You'll use it in every war of the database.

所有地方。

Everything.

关键是速度，Matt。

And the key is speed, Matt.

我们讨论速度不够多。

We don't discuss speed as much.

Jev真的很快。

Jev is so fast.

而且Haiku 5.5实际上比之前的版本慢，但根本没法比。

And Haiku 5.5 is actually slower than the one before, but it's not even comparable.

Sonnet快了两倍。

Sonnet's twice as fast.

因为更快，你可以多用一些。

You can just use more of it because it's faster.

如果慢的话，并行使用就有限了。

If it's slower, you can only use so much in parallel.

不过，我同意你的观点。

But yeah, I agree with your part.

你就什么都用它。

You just use it for everything.

这对美国企业的开源意味着什么？

What does this mean for open source in US enterprises?

这是不是让它天然没那么有吸引力，你会想，‘天哪，我可以用Haiku和Anthropic，便宜90%’？

Does this make it inherently less attractive, where you're like, "Hell, I can get Haiku and Anthropic for 90% less"?

这确实让开源失去了吸引力。

That really reduces the attractiveness of open source.

我也是这么想的。

That's why I thought.

我不记得是Dev还是Jack说的，但我说过，你可以想象明年情况会完全反转，然后整个OpenRouter的事情都会改变，人们并不忠诚。

I don't know whether it was Dev or Jack, but I said you could imagine it would all flip around next year, and then the whole OpenRouter thing would change, and people are not loyal.

不管对错，没人愿意用至少是中国背景的开源权重模型。

Right or wrong, no one wants to use at least China-based open weights models.

企业里没人愿意用它们。

No one in the enterprise wants to use them.

所以当你有一个本土解决方案，最好来自一个已经通过认证且值得信赖的供应商，我不是说这一定会发生，因为DeepSeek等公司有很多聪明人在工作。

So when you have a domestic solution, ideally from a vendor that's already been qualified and trusted, I'm not saying it will happen because there's a lot of smart people working at DeepSeek et al.

有些是最顶尖的。

Some of the best of the best.

但你可以想象，这只是一个时间点，12个月后，它们会变成我们投资的一些公司使用的小众模型，但在企业中几乎不存在。

But you could imagine that this being a moment in time, and then 12 months from now, they're niche models that are used by some companies we invest in, but really in the enterprise are almost nonexistent.

再说，我觉得Salesforce在交易完成前就让Fin把它撤掉，这很能说明问题。

Again, I think Salesforce making Fin rip it out before the deal closed is telling.

这只是一时的现象。

It's a moment in time.

也许不会发生。

Maybe that won't happen.

但人们就是不想要这些中国的开源权重模型。

But people don't want these China open weight models.

他们就是不想用。

They just don't want them.

但他们会有Reflection。

But they'll have Reflection.

这就是为什么OpenRouter可能——如果NVIDIA收购了Reflection，那么Reflection加上Poolside再加上其他所有东西，最终会占50%的使用量和15%的收入。

That's why the OpenRouter could, maybe if NVIDIA buys Reflection, then Reflection plus Poolside plus everything ends up being 50% of the usage and 15% of the revenue.

我一点也不意外。

It wouldn't surprise me at all.

Jason，这个日程里还有什么让你感兴趣的事？

Jason, what else tickled your fancy from this schedule?

我本来不太想提，但Bending Spoons把Airtable的人基本都裁了。

I almost don't want to bring it up, but Bending Spoons laying off basically all of Airtable.

不意外，对吧？

Not a surprise, right?

但说实话，这事发生的时候我写了，根本没人关心。

But man, I wrote this when it happened, and no one cared.

我当时就说：“各位，大家都在谈钱和ARR收入。”

I'm like, "Guys, everyone's talking about the money and the ARR revenue."

我还说：“你们知道几乎所有人都要失业了吗？”

And I'm like, "You realize almost everyone's losing their job?"

我没看到谁在乎大家都要失业这件事。

I didn't see anyone give a rat's arse about the fact everyone's going to lose their job.

我就觉得，Vimeo的人全被裁了，Evernote的人全被裁了，Miro的人也要全被裁了，我几乎没看到有人发推文关心一下。

I'm like, everyone lost their job at Vimeo, everyone lost their job at Evernote, everyone's going to lose their job at Miro, and I barely saw a tweet about anyone giving a rat's ass.

我并不是在评判，因为这就是资本主义。

And I'm not even judging because this is capitalism.

但是，当你把公司卖给Bending Spoons的时候，你得明白你的整个团队都会走人。

But man, when you sell your company to Bending Spoons, you just got to realize your entire team's going to be gone.

创始人欠团队什么吗？

And do founders owe anything to their teams?

董事会欠团队什么吗，还是他们只对Cedars-Sinai负责，追求最好的回报？

Do boards of directors own anything to their teams, or do they just owe their best returns to Cedars-Sinai, right?

你有没有道德义务，不把公司卖给一个几周内就会裁掉80%团队的人？

Do you have any moral obligation to sell to someone that's going to lay off 80% of your team within weeks?

他们是利益相关者，尤其是像Miro和Airtable这样现金流为正、还在增长的公司，对吧？

They are stakeholders, especially if you're cash flow positive and growing, like Miro and Airtable, right?

他们两家银行里都有数亿美元，而且还在增长。

They both have hundreds of millions in the bank and growing.

他们重要吗？

Do they matter?