SpaceX IPO、《寓言 5》、AI 资本支出更新与市场动态，嘉宾：Gavin Baker、Andrew Fox 与 Clark Tang | BG2

这事儿跟你们两位都有关。

It's either of you guys.

我觉得，如果你仔细想想 Cursor 对 xAI 意味着什么，我们其实已经有例子了——一旦真正踩上那条帕累托前沿，收入就能飞速增长，Anthropic 就是活生生的例子。

So I don't think I think if you think about those variables around what Cursor could mean for xAI, and we do have an existence proof that once you really get on that Pareto frontier, revenue can scale rapidly, and it's called Anthropic.

而且很明显，编程这块的需求真的特别大。

And there does seem to be an exhaust There seems to be a lot of demand for coding.

我觉得没错，John Masad 发过一篇特别有意思的东西。

And I do think, I'm John Masad posted something very interesting.

Replit 的创始人。

Founder of Replit.

Replit 的创始人。

The founder of Replit.

Replit，没错。

Replit, yeah.

他管这叫‘苦涩教训的近亲’，认为编程可能是通向通用人工智能和超级人工智能最快的一条路。

He called it Bitter Lesson Adjacent, that coding may be the fastest path to AGI and ASI.

因为如果你编程能力很强，就可以写代码让模型去做任何事。

Because if you're really good at coding, you can write code, if a model's good at coding, to do anything.

我觉得这是个非常关键的观点，编程未来还会继续非常重要。

I think that's a profound point, and I think coding is going to continue to be very important.

如果你想想这个变量，比如星链V3实现直连手机，再想想他们能多快部署地面算力，或者能不能做到。

I think if you think about that variable, if you think about Starlink direct to cell enabled by Starlink v3, and you think about how quickly they can or cannot bring on terrestrial compute.

我觉得轨道计算对上市估值不是必需的，但确实很重要。

I don't think orbital compute is necessary for the IPO valuation, but it's certainly important.

也许换种说法是，你可能觉得我们会比实现轨道计算更快抵达ASI。

Maybe another way to say it is you may think we're going to get to ASI faster than we're going to get to orbital compute.

这可能会让我们从300智商提升到400、500，甚至更高。

That may take us from 300 IQ to 400 IQ, 500 IQ and beyond.

还有能力把它扩展到消耗全球GDP的10%。

And the ability to scale it up to consume 10% of global GDP.

但也许我们接下来该聊这个了。

But maybe that's where we should move next.

不，不是的。

No, no.

我觉得在轨道计算方面，Foxy和Art Clark能很好地把数据梳理清楚。

I think out of orbital compute, think Foxy would be great, Art Clark, to lay out the math.

对。

Yeah.

从基本原理来看，Clark有一张非常清晰的图表，展示了每吉瓦的成本和每吉瓦的美元花费。

From first principles, Clark has this great chart on the gigawatts it costs, the dollars per gigawatt.

很好。

Great.

请你详细讲讲这个经济逻辑。

Walk us through the economic case.

对。

Yeah.

没错。

Yeah.

说到轨道计算对投资是否关键，我觉得其实不是。

So, I mean, on this point of is Orbital key to investing here, I don't think it is.

我想先说一点，目前市场对人工智能业务的预期，对应的变现率是多少？

And the first point I'll make is what are the implied monetization rates based on expectations today for the AI business?

你提到的那个1600亿美元的数字，就是大家经常在讨论的那个。

I think you threw out the $160,000,000,000 number that's been linked out there that people are talking about.

按这个数字推算，人工智能业务每吉瓦每年的变现额大约是140亿美元。

The implied monetization rate on that number is something like $14,000,000,000 per gigawatt per year for the AI business.

他们刚在2022到2023年间和Anthropic签了合同。

They just signed Anthropic at '22 to '23.

最近又和谷歌签了50吉瓦的协议。

They just signed Google at 50.

没错。

Right.

对吧？

Right?

所以我觉得，即使不靠太空项目，只在地面投资人工智能领域，也一样值得期待。

So I think you can invest behind the AI business terrestrially and still be excited about it.

至于太空项目，

With orbital

这一点很重要。

important point.

如果他们能拿到土地和电力，那确实让人兴奋。

Excited about it if they can get the land and the power.

没错。

Right.

但我觉得，对大多数投资者来说，

But but I mean, I think for most investors

是的。

Yes.

对吧？

Right?

他们更容易理解SpaceX在地面业务上是怎么赢的。

They get they have an easier time getting their head around how SpaceX wins terrestrially.

比如，他们能不能搞到土地、电力和芯片？

Like, can they go get land, power, and chips?

这个问题的答案可能性很高。

The answer to that is high probability.

是的。

Yes.

对。

Yeah.

好的。

Okay.

我们说的是，按他们目前的变现速度，光靠地面业务就能达到外面流传的那些数字，根本不用先假设他们未来会在轨道数据中心上拉开差距。

And what we're saying is at the rate they're monetizing that, that gets you to the numbers that are being leaked out there before you even have to take the leap of faith that they're going to extend the lead with Orbital data centers.

不过你也说说这方面的情况。

But take us there on that too.

当然。

Sure.

对。

Yeah.

所以，关于轨道项目，我觉得关键在于两级可重复使用。

So so, look, with Orbital, I think the key thing is two stage reusability.

对。

Yeah.

而且是快速的两级可重复使用。

And beyond that, rapid two stage reusability.

对。

Yeah.

目前星舰已经证明了他们能成功让助推器返回着陆。

So today with Starship, they've shown that they can successfully re land the booster.

是的。

Mhmm.

第二级的情况，我们今年晚些时候就能看到了。

The second stage, we'll see what happens later this year.

我觉得他们打算明年就把这一级回收回来，实现重复使用。

I think they're attempting to bring that back and then make it reusable by next year.

对于轨道计算来说，两级可重复使用最关键的是，每公斤的成本会大幅下降。

But the thing that's important about two stage reusability when it comes to the economics for orbital compute, right, is the cost per kg comes down significantly.

从猎鹰火箭的每公斤1500美元左右，降到每公斤2.5美元，甚至更低。

We're talking about going from $1,500 per kg on Falcon, somewhere in that range, to $2.50 per kg, something lower.

火箭回收次数越多，价格就越低。

And the more that you can reuse the rocket, the more that price comes down.

对吧？

Right?

因为你只是在分摊发射器的成本。

Because you're just depreciating the cost of the launch.

最终，成本会趋近于燃料本身的费用。

And eventually, you asymptote to the cost of the fuel.

没错。

Right.

假设火箭能一直重复使用的话。

Assuming you can use a rocket for forever.

对。

Yes.

对吧？

Right?

但要真正实现这一点，还得花上很长时间。

Which will take a very long time for us to to really achieve that.

到了那时候，每公斤的成本就能降到远低于250美元。

But and at that point, we're talking about something well south of 250 per kg.

然后你再看看这些AI卫星的规格。

So then you look at the specs of these AI satellites.

你懂的，做得真不错

You know, did a great

是的

Yeah.

他前几天介绍的那个卫星规格，那个载荷简直太惊人了。

That pod was incredible that he laid out the other day, the specs on the satellites.

特别棒，因为我觉得他们终于给大伙儿展示了怎么实际设计这样的卫星，还有卫星多重、一艘星舰一次能装多少颗。

It was really great because I think they are finally showing people here's how you could viably design one of these satellites and how heavy is the satellite, how many could you fit into a Starship launch.

算一算账，每艘星舰的发射能力大概能达到五兆瓦。

And when you back into the numbers, you get to something like five megawatts of capacity per Starship launch.

没错。

Right.

一艘星舰能运载一百公吨。

There's a 100 metric tons in one of those Starships.

所以你可以倒推一下，每吉瓦的发射成本到底有多少。

So you can back into the math of how much will it cost Right.

把这些卫星送入太空，每吉瓦要花多少钱。

Per gigawatt to launch these satellites into space.

对。

Right.

把这种计算能力发射到太空去。

Launch this compute into space.

在考虑坏掉的显卡、故障的卫星这些情况之前，你算出来的数字是这样的，但这些事迟早都会发生。

And the math that you get to before you account for things like bad GPUs, bad satellites, right, these will all be things that happen.

但算下来，把这种卫星送入太空，每吉瓦的资本支出大约是50亿美元。

But the math you get to is it's about $5,000,000,000 per gigawatt of CapEx to put these in space.

对。

Right.

作为对比，地面上的发电设施——发电机、变压器、外壳、输电系统——现在每吉瓦的成本大概是200到250亿美元。

For comparison, terrestrially, talk about the switch gears, the generators, the transformers, the shell, getting the power, that today is about 25 20 to 25,000,000,000 per gigawatt.

所以，我们在材料成本的一半上，实现了五倍的成本降低。

So we're talking about a five x reduction in cost on half of your bill of materials.

这对数据中心来说，是个惊人的数字。

For the data center, which is a huge number.

简单来说，现在在地上部署一吉瓦的算力，成本高达六百亿美元。

Just very simply, I mean, to say it costs $60,000,000,000 to put a gigawatt on the ground today.

其中大约三百五十亿是用于GPU和做训练与推理的硅芯片，另外二百五十亿是土地、机房、电力和冷却的费用。

We'll call it 35 of that are the GPUs and the silicon that's doing the training and the inference, and 25,000,000,000 is the land, the shell, the power, and the cooling.

我推测，这些部分的成本可能会持续上涨，所以那二百五十亿未必能降下来。

I would hypothesize that those elements are probably going to be inflationary, so that 25,000,000,000 may not go down.

而在太空中，电力和冷却几乎是免费的——我这里说的‘太空’，指的是不用花钱买地。

And because space, power, cooling are effectively free in space, and when I say space, I mean land.

太空中根本没有土地这一说，

There's no land in space,

但确实有空间可用。

but there is But there's a space.

有大量的空间可以利用。

There's a lot of

你是在说，把一吉瓦的算力送入太空，只要三十亿美元，还能大幅降低运营成本。

You're space in talking about putting a gigawatt into space for $30,000,000,000 and having lower operating costs.

现在，动态成本和600亿美元的通胀成本相比，那300亿美元里可能有50亿是会随着时间下降的。

Now, dynamic versus $60,000,000,000 that's inflationary, and that $30,000,000,000 that 5,000,000,000 may be deflationary over time.

但我们真正需要考虑的是可靠性和维护问题。

But what we need to consider is the reliability and the maintenance.

谁都会算这笔账，但只要这些太空卫星不会以惊人的频率出故障，数学模型就成立。顺便说一句，我们知道GPU会过热，激光器也会失效。

Everybody can do the math, but as long as these satellites in space aren't failing at an astronomical rate, the math maths has to And by way, we know GPUs melt and lasers fail.

我们在数据中心里早就见过这种事，尤其是在大规模训练的时候。

We know this happens in data centers, particularly during big training runs.

GPU确实会过热。

I mean, GPUs melt.

所以，只要可靠性和维护水平没有大幅下降，一旦我们实现星舰V3的可重复使用，再进一步实现快速重复使用，这笔账就说得通。

So as long as the reliability and maintenance is not dramatically lower, the math is there once we have reusability and then rapid reusability for Starship v3.

当你看这个问题时，我们已经讨论过星链了，很自然就会想到，星链未来会直接面向终端用户提供服务。

When you look at this, okay, so we went through Starlink, and we said, like it just stands to reason we're gonna have direct to sell on Starlink.

这些假设听起来都挺合理，不难理解。

Like the assumptions there are, know, again, seem like you can get your head around.

至于建造地面数据中心，从这几笔交易来看，很明显埃隆会让星链或SpaceX在这一块做大很多。

Then when it comes to building terrestrial data centers, again, not a hard one to think that based on these couple deals that Elon is going to build a much bigger, Starlink is going to build or SpaceX is going to build a much bigger business there.

而且这还附带一个期权，未来可能让成本进一步降低。

And then you have this call option on space that would drop the price even further.

我们还没提到他们的商业模式。

The one thing we haven't talked about is their model.

没错。

Right.

我觉得这一点挺让人意外的。

And I find this surprising.

没错。

Right.

六个月前，xAI还在和别人竞争。

Six months ago xAI was competing.

当时表现不错，但最近几个月他们做了一件大事——收购了Cursor。

Were doing pretty well but they've done something dramatic over the course of the past couple months which is they bought Cursor.

对。

Right.

Cursor公司已经有七八百人了，从收入角度看表现已经非常出色。

Cursor is 700, 800 people was already doing incredibly well from a revenue perspective.

我们自己的预测是，他们今年年底的收入可能达到一百亿美元。

Our own projections were that they could exit this year at up to $10,000,000,000 of revenue.

所以他们的增长速度非常快。

So they were growing very fast.

他们是领先的编程助手之一。

One of the leading coding agents.

但他们还有一支顶尖团队，完全有能力打造前沿级别的模型，只是受限于算力。

But they also had this incredible team with the potential to really build a frontier level model, but they were compute constrained.

突然间，他们被X收购了，而X拥有海量的算力资源，现在他们可以尽情训练了。

So all of a sudden they get bought by x, x has massive compute that they can now train on.

当我思考人工智能领域的收入时，如果看模型里的这项数据，从一百亿涨到一千五百亿，其中很大一部分会来自他们核心的Weave类业务。

And when I think about the revenue in AI, if I look at that line item in the models, having it go from $10,000,000,000 to $150,000,000,000 yes, a lot of that will be the core weave type business that they have.

但问题是，其中有多少是真正由来自 Cursor 的新团队驱动的 xAI 核心业务？

But the question is how much of that is going to be the core xAI business that's really powered by the new team from Cursor?

对此你有什么看法，Gavin？

So any thoughts on that, Gavin?

目前来看，Composer 2.5 在十二天前就已经在性能上全面领先了。

Right now, so Composer 2.5 was Pareto dominant twelve days ago.

它是基于 Kimi K2.5 基础模型训练出来的。

It was trained on the Kimi K2.5 base model.

现在

Now

情况是这样的，Grok 4.3，那个拥有1.5万亿参数的模型正在训练中。

what's happening is the Grok 4.3, 1,500,000,000,000 parameter model is training.

根据缩放定律推测，这可能会成为一个更优秀的基础模型。

One would hypothesize, based on scaling laws, that that might be a better base model.

然后，Cursor的数据被注入到了预训练过程中，而不仅仅是用于强化学习。

And then the cursor data is being injected into the pretraining process, not just reinforcement learning.

我们等着看吧，我觉得一旦这个结果出来，会是一个非常关键的指标。

And we'll see, and I think that is going be a very important data point when that comes out.

我只是想提醒大家，一旦你能在帕累托曲线上多个点上取得突破，只要有算力，就能快速扩张。

And I just think everyone should keep in mind that once you are at multiple places on that Pareto curve, if you have compute, you can scale really rapidly.

要说这个故事里被忽略的关键点，我觉得就是这个。

You know, that to me is, if I had to say what the one piece that's being lost in the story, right?

大家很容易对和Anthropic的合作感到兴奋，因为这些合作看得见、摸得着。

Like it's easy for everybody to get excited about the deals with Anthropic because you could put your hands around that.

你能清楚算出它们能带来多少收入。

You know how much revenue it is.

现在大家都在争论九十天终止条款能撑多久，这些收入该用多少倍估值。

I see debate about the ninety day termination and how long they last and what multiple do you put on those revenues.

但真正被忽略的是，他们在构建前沿模型方面的能力已经取得了巨大突破。

But I think the thing that's getting lost is I think they've dramatically advanced their capability when it comes to building a frontier model.

硅谷以外的人可能对迈克尔和Cursor团队了解不多。

People outside Silicon Valley may not know Michael and the team at Cursor as well.

他直接把整个顶尖团队整个搬到了SpaceX，真是了不起。

This is an extraordinary team that he just downloaded right into SpaceX.

SpaceX本来就在做不错的模型，但他们有个独特优势——能灵活调动算力，随时把算力拉回来训练和运行模型。

SpaceX was already building good models and what they have is they have this way to monetize compute that gives you this call option that you can pull all that compute in house right, to train a model and then to run the model.

我觉得如果真有意外惊喜，那很可能就在这里。如果让大家轮流说，我敢说这个地方被关注得最少，但潜力最大。

I suspect if there's an upside surprise, if we went around the table, I'd say this is the place that's getting the least amount of attention and could have the biggest upside surprise.

克拉克，你觉得现在业界最忽略的点是什么？或者有哪些地方被误解了？

Any thoughts, Clark, on what you think is being overlooked or areas that you think are misunderstood about the business today?

过去几周已经证明了，埃隆和他的团队真的能把这么多算力快速搭建起来。

I would what the last few weeks have proven is that Elon, their team can stand up all this compute.

要是回溯一年半前，他们那时候在算力部署上其实落后的。

Actually, if you just, you know, went back one and a half years, you know, they were behind in the race to stand up compute.

那时候他们手里的H100没多少。

They were you know, they they didn't have that many H100s.

后来他们引入了Colossus，接着又上了Colossus II，规模远超其他人。

Brought in Colossus, then they brought in Colossus II at a scale much larger than anyone else.

现在，随着Verirubin的部署临近，从我跟不少人交流来看，他们似乎已经锁定了高达20%的Verirubin产能，尤其是在这些芯片刚上市、供不应求的初期阶段，这让他们在算力布局上占据了先机，因为大家都觉得他们更擅长搭建这类算力系统。

And now, you know, as we gear for Verirubin, you know, from, you know, a lot of my conversations, it looks like they've, you know, secured maybe up to 20% of Verirubin capacity, especially in the early days of, you know, when, you know, these chips are very scarce, that they're going to have a lead on all of this because, you know, people think that they can stand up this compute better.

所以我觉得，过去几周其实已经证明了，埃隆真的愿意去挑战技术前沿。

So I think they'll you know, what what the last few weeks have actually shown is that Elon, you know, Elon will take, you know, take a shot at hitting the frontier.

但如果他们因为某种原因多采购了一些算力资源，别忘了，这可是个极其稀缺的资产，而他们已经证明自己能以业内顶尖的利润率和回本周期来实现变现。

But if it, you know, if for whatever reason they have over procured some capacity, this is a very scarce asset that they have shown that they can monetize at actually, you know, best in class margins and payback periods.

说真的，咱们干这行这么久，早就该明白这一点了。

Mean, the irony is like, you know, you and I've been doing this long enough to know.

这正是贝佐斯当初要建AWS的原因。

I mean, that's why Bezos built AWS.

对吧？

Right?

他得为黑色星期五提前准备好足够的容量。

He had to build capacity for Black Friday.

没错。

Yeah.

对吧？

Right?

但接下来的一年里，他把这些必须建起来的算力资源搁置着，却想出了一个特别棒的赚钱方法。

But then the rest of the year, he sat on all this capacity they had to build and he figured out a really incredible way to monetize this.

顺便说一句，当时2009年、2010年他正在布局AWS的时候，投资者们都很反感。

And by the way, investors at the time, 2009, 2010, when he was building out the capability around AWS, hated it.

当然了。

Of course.

因为他把所有自由现金流都投进去了。

Because he was consuming all that free cash flow.

而与此同时，他正在挖一个史上最大的金矿。

My meanwhile, he was digging the biggest gold mine in the history of the world.

最大的之一。

One of the biggest.

among them.

Among them.

其中之一。

Among them.

当时可能是最大的。

At the time was probably the biggest.

是的。

Yeah.

谷歌搜索可能会有不同的看法。

Google search might wanna have a discussion.

顺便说一下，我觉得这很重要。

By the way, do think it is important.

Grok 4.3，我觉得如果他们收购了Cursor，那可能会非常关键。

Grok 4.3, I think the Cursor, if they acquire it, that may end up being very important.

但就在十到十二天前，Grok 4.3还处在帕累托前沿上，这些技术发展得太快了。

But Grok four point three was on the Pareto frontier as of ten or twelve days ago, these things move fast.

全球最智能的5000亿参数模型。

Most intelligent 500,000,000,000 parameter model in the world.

当时有四家公司处于领先地位。

And they were on the frontier, there are four companies on the frontier.

xAI、SpaceX的AI、谷歌凭借Gemini 3.1 Pro胜出，其余的则由Anthropic和OpenAI主导。

XAI, SpaceX AI, Google won with Gemini 3.1 Pro, and then the rest of it was dominated by Anthropic and OpenAI.

它们都在帕累托前沿上，接下来我们看看它们在Cursor上会有什么表现。

But they were on the Pareto frontier, and now we'll see what they do with Cursor.

对。

Yeah.

我待会儿再回到那个话题。

I want to come back to that in a second.

对了，我想问

By the man, I want ask

你几个问题。

you some question.

问吧。

Go.

你觉得呢？

What do you think?

所以你觉得最大的潜在增长点来自模型本身？

So you think the biggest source of potential upside is the model?

对。

Yes.

你觉得呢？

What do you think?

我觉得最没人提到的其实是这一点。

I the thing think that's least talked about.

最没人提到的。

Least talked about.

对吧？

Right?

你听我说。

And so, listen.

我怎么看这个IPO的乐观和悲观情况呢？悲观的一方盯着去年的营收。

When I look at the bull bear case on the IPO, right, the bears are looking at last year's revenue.

他们说去年是180亿美元，再看银行预测三年后能达到1600亿，心里就犯嘀咕了——世界上没几家公司在三四年里能把收入翻八倍。

Say it was $18,000,000,000 and they're looking at the forecast from the banks of $160,000,000,000, you know, three years from now and they're saying, not many companies in the history of the world have basically 8x their revenue over three to four years.

对。

Right.

所以我觉得大家对估值感到紧张，原因就在这儿。

So that's where I think people get nervous about the valuation.

但我再仔细想想，如果从分析师的角度，一码一码拆开来看——这正是我刚才做的，对吧？

When I look at this again, when you break it down as an analyst first principles part by part, which is what I tried to do here, right?

看看星链项目，完全可行。

When you look at Starlink, it looks totally doable.

再看他们在地面AI算力上的布局，未来三年内实现也完全没问题。

When I look at what they're building in AI compute terrestrially, looks totally doable over the course of the next three years.

至于他们那个模型，尤其是收购了Cursor之后，把算力和这些技术整合起来，我觉得很可能成为意外的亮点。

When I look at the model itself after the acquisition of Cursor, combining those things around the compute they have, that looks to me like it could be an upside surprise.

所以我觉得，虽然现在是IPO阶段，但三年后回头看，很可能大家都会说：天啊，这明明很明显啊。

So I would say that I think that in the IPO, but I think when you look back three years from now, there's a decent chance that everybody's like, oh my god, that was super obvious.

对吧？

Right?

尽管今天这些事情都伴随着风险。

Even though today all of these things have risk associated.

回到我们一开始说的，我们没人是来鼓吹这个17700亿美元的IPO的。

And back to where we started, I'm not, know, none of us are here to pump the IPO at 1,770,000,000,000.

我们只是想像我们内部那样，一步步拆解，看看未来各种可能性的分布是怎样的。

It's really to just break it down as we do inside our shop and to say, what is that distribution of future probabilities?

从现在开始，哪种情况发生的概率更高？

What's the probability that is higher from here?

我觉得我们都已经被AI洗脑了。

What's the and I think we're all pretty AI pilled.

如果你被AI洗脑了，那就意味着我们必须建造远超世人想象的算力，这些模型的价值也会比大家预期的高得多。

And if you're AI pilled, that means we gotta build a lot more compute than the world thinks and that these models are gonna be a lot more valuable than people think.

把这一点和他们的主营业务结合起来，我真的想不出还有哪个创业者或哪家公司，比SpaceX更值得押注未来。

You combine that with their core business, I don't know another entrepreneur or another business that's a better bet on the future than SpaceX.

所以我觉得，对大多数机构投资者来说，这绝对是必须买入、必须持有的标的，买了就别管它，这样才能真正押注太空和人工智能的未来。

And so I think for most institutional investors it's a must buy, a must own is set it and forget it in order to have a real bet on both the space and the AI future.

但愿如此吧。

From your lips to God's ears.

我的意思是，你还是得等一等。

I mean, again, I think that you're going to have to wait.

但上周我们看到过一张图表。

But we had this chart last week that came out.

大家纷纷在推特上转发，时间点也太巧了。

Everybody was sending around Twitter conveniently timed.

数据显示，像Facebook、Twitter、阿里巴巴、Shopify这些公司上市后，平均最大回撤超过了50%。

And it shows the average max drawdown post IPO for like 20 companies from Facebook, Twitter, Alibaba, Shopify is over 50%.

所以也许我们就在这里先告一段落吧。

And so maybe that again will end this section here.

加文，你我在这行干了这么多年了。

Gavin, you and I have been doing this a long time.

我们都知道，上市前后股价肯定会大幅波动。

We know it's going to be bouncy around the IPO.

作为基金经理，你是怎么应对这种情况的？

How do you as a manager try to manage that?

你会不会在上市前后做短线交易？

Do you try to trade around the IPO?

还是干脆买进后就一直拿着？

Do you set it kind of and forget it?

从Altimeter的角度来说，我们通常会先建一个基础仓位，然后长期持有，对吧？

I would say from an Altimeter perspective what we tend to do is we take a base position that we set and forget, right?

之后再根据市场在特定时刻的反应，适当加仓或减仓。

And then we may size up or size down depending upon how the market reacts in, you know, in a particular moment.

关于这张图，你有什么看法？你们平时是怎么看待这个问题的？毕竟你们之前就持有了大量股份。

But any thoughts on on this chart or, you know, how how people you guys are thinking about it in particular, you obviously own a lot going into it.

我完全同意你说的每一句话。

First, agree with absolutely everything you said.

我也是这么想的。

I actually think about it the same way.

定好之后就别管了。

Set it and forget it.

你提到过，有稳定仓位，也会根据情况调整，就像船航行时，想迎风快一点就把压舱物移到一侧，想避风就移到另一侧。

You've talked about you have ballast, you move around, and you move the ballast to one side of the ship when you want a ship to lead into the wind to go faster and you move it to the other side.

但我们不能让船翻了。

We don't want the ship to tip over.

这个比喻太贴切了。

Think that's a great analogy.

对投资组合里所有重要的公司，我们都这么看待，百分之百认同。

Think about all important companies in the portfolio the same way, so 100% agree.

这张图真是让人有点泄气。

I mean, this chart is a bummer.

我想说的是，虽然有关于IPO的数据，但这种情况真的前所未见。

What I would say is, you know, there's data on IPOs, but what I would just say is this is a really unprecedented situation.

没错。

Yes.

我们从来没经历过这么大规模的IPO。

We've never had an IPO this big.

我们从来没看过哪家公司上市后这么快就进指数。

We've never had an IPO that's going go into an index this quickly.

投资者到底会抛多少，我们完全没法预估。

We simply do not know how much selling there will be from investors.

我只能猜一猜。

I would hazard a guess.

说实话，我真的说不准。

I mean, I'm I don't know.

但我觉得，埃隆并不需要套现。

But Elon, I don't think he needs liquidity.

我觉得他持有多少股份来着，Foxy？

And I think he owns what does he own, Foxy?

百分之五十。

-fifty percent

公司百分之五十的股份。

-fifty percent of the company.

顺便说一下，他锁仓了365天，还有366天。

By the way, he's locked up for three sixty five days and three sixty six days.

所以我们知道他不会卖，所以——

So we know he's not selling, -So

我只是觉得这情况前所未有，真正准确的回答就是，短期会发生什么，我真的说不准。

I just think it's an unprecedented situation, and the right answer is I don't know what's going to happen in the short term.

对于每个自己做决定的投资者，我只想鼓励你们，就按你刚才说的那样去思考。

And the right answer that I would just encourage every investor making their own decision is to just think exactly the way you articulated it.

我们有这些不同的因素。

We have these different levers.

我们有这些不同的变量。

We have these different variables.

从基本原理出发，逐个分析每一个变量。

Think about each one of them from first principles.

自己做决定。

Make your own decision.

自己做充分的调查研究。

Do your own due diligence.

要认真思考。

Be thoughtful.

但这里涉及的变量实在太多了。

But there are a lot of variables here.

说真的，我觉得用100倍的 trailing TTM 收入来估值，有点奇怪。

Then it is a little funny to me that it was 100 times trailing TTM revenue.

在他们签下那些交易之后，我觉得现在大概是39倍。

Well, after the deals they signed, That I think it's at 39

变化可能很快。

can change fast.

他们一个月内增加了两千九百亿美元。

So they added $29,000,000,000 in a month.

顺便问一下，你见过这种情况吗？

By the way, have you ever seen that happen?

从来没见过。

Never.

从来没见过。

Never.

这说明他不仅是个出色的工程师，他和格温还有整个团队在商业运作上也非常出色。

And it just goes to show, first, is not only a great engineer, he and Gwen and the team are great at business.

那布拉德呢？

And Brad?

他们很清楚该如何筹集资金，推动公司进入下一阶段。

They understand what needs to be done to raise the capital, to get to the next phase.

他们在业务上有着长远的愿景。

They have a long term mission in the business.

所以对我来说，过去几周我们看到Cursor的所作所为，还有他们谈成的这些交易，我不觉得那七大巨头中的任何一家能这么快地调整自己的业务。

And so to me, again, what we saw over the course of the last few weeks with Cursor, what we saw with these deals that they cut, I don't know that any of the Mag-seven could have moved that quickly to adjust the business that they did.

这种大规模的创业精神，我们平时在企业里几乎看不到。

It's exceptionally entrepreneurial at scale, which we very rarely see in businesses.

我还想提另外两点。

Two other things I would just say Can

布拉德，我抱你一下？

I give you a hug, Brad?

我还想再说两点。

Two other things I would just say.

第一，大家总在谈论融资的总额。

Number one is people talk a lot about the total amount of capital being raised.

你把这里所有公司的融资加起来——Anthropic、OpenAI、SpaceX这些，算上他们可能融到的钱，大概2500亿美元，还不到那七大巨头总市值的1%。

If you add up the capital here, right, for Anthropic, what they may raise, what OpenAI may raise, what SpaceX may raise, let's call it $250,000,000,000 That's 1% of the Mag seven.

明白吗？

Okay?

这只是Mag七的1%。

It's 1% of the Mag seven.

对。

Yeah.

我们也会这样做的。

And we will as well.

在我看来，这其实是我们共同相信的、对未来的押注。

You know, like, that to me is like a bet on the future that we all believe in.

如果我要问，我们在哪里和主流看法不同？

And so if I said, where are we out of consensus?

我们的独特观点是什么？

What is our variant perception?

我们其实认为，它会变得更大、更快。

We actually think it's going to be bigger, faster.

我们几年前就开始这么想了。

And we've thought that for a couple of years.

所以一开始，它只占魔七市值的1%。

So first, it's only 1% of the Mag seven market cap.

然后你提到了这一点。

And then you referenced it.

抛售的规模。

The amount of selling.

我这儿有个图表，待会儿放上来。

I've got a chart we'll post here.

这是SpaceX股东的Dribbles股份解禁情况。

This is the Dribbles share release for SpaceX shareholders.

在首次财报发布之前，能解禁的股份其实不多。

So there's not a lot that can be released up until after the first earnings.

我们在Cerebras上市时就见过这种情况。

We saw this in the Cerebras IPO.

这次IPO里也有类似的情况。

There's a version of it here in this IPO.

所以我觉得，银行这边很谨慎，毕竟这是个规模巨大的IPO。

And so again, I think the banks have been thoughtful here knowing that this is a very large IPO.

我不是说股价不会下跌。

And I'm not saying that won't trade down.

毕竟这种事确实有可能跌，你知道的。

Like there's possibility, you know, these things trade down.

但对我而言，咱们眼光放长远一点。

But again, for me, telescope out.

还有哪家公司比它更适合作为未来的投资标的？

Is there any company better positioned as a bet on the future?

我觉得过去五周里，他们展现出来的实力，大概已经排第一了。

I think what they've shown over the course of the last five weeks, they're probably number one.

咱们接着往下说。

But let's move on.

没有。

No.

不会。

No.

往前看。

Ahead.

我想说说这些员工的事。

I'll say one thing about the employees.

这里另一件前所未有的事是，员工和很大程度上的投资者，每六个月就能实现一次流动性，确实如此。

I think another thing that's unprecedented here is the employees and to a large degree the investors here, have had liquidity every six Exactly.

没错。

Exactly.

过去十年都是这样。

Like the last ten years.

对。

Yes.

所以如果你是SpaceX的员工，或者前员工，想卖股票的话，差不多已经有二十次机会了。

So if you're a SpaceX employee, our former employee, and you wanted to sell, you've had, whatever that is, close to 20 chances.

历史记录很清楚，很多大投资者都能做到，我想大多数人还是选择继续持有。

It is a matter of historical record that large investors have been able to I would think a lot of the people, they've chosen to own it.

现在有了新的估值，我们看看他们会怎么选择。

Now there's a new valuation, and we'll see what they do.

但这种情况绝对是前所未有的，咱们走着瞧吧。

But just this is utterly unprecedented, and we'll see.

是的。

Yeah.

不对。

No.

这个观点非常到位。

It's it's it's a great point.

我们实际上已经把这些公司称为准上市公司了。

We've, in fact, called these companies quasi public.

你我都清楚，SpaceX，我也把Anthropic和Databricks归到这一类里。

You and I both know that SpaceX, and I put Anthropic in in in this category as well, Databricks in this category.

在过去三年里，这些公司的股权流动性，其实比一些我们熟知的公开上市生物技术公司还要强。

These things in many ways have been more liquid over the course of the past three years than some public biotech companies we know.

没错。

Absolutely.

对吧？

Right?

所以这里其实存在一个流动性梯度。

And so there's a continuum of liquidity here.

我们总把情况简单分成私有和公开两种，但本质上它是一个连续谱。

We treat it as a binary private versus public, but it's really about this continuum.

咱们继续聊聊模型吧。

You know, let's keep going on models.

Anthropic最近推出了Fable 5，你昨天提到过，其实就是Mythos加上一些针对网络、生物、化学领域的分类器和安全防护机制。

You know, Anthropic launched Fable five, which you referenced yesterday, which is basically Mythos with some classifiers and safeguards around cyber and biology, chemistry and distillation.

一旦触发这些条件，系统就会回退到 Opus 4.8。

When those things get triggered, it fails back to Opus 4.8.

昨天麦卡锡还发过一条推文谈到这个。

You know, there's a McCarthy tweet about this yesterday.

他说，这模型在所有基准测试上表现都很出色，但真正厉害的是处理长时间任务的能力。

He said, you know, it's soda on all the benchmarks, but what really makes it special is long running tasks.

明白了。

Okay.

你转推了我们那位好朋友诺姆·布朗的内容。

You retweeted our good friend, you know, Noam Brown.

ChatGPT 5.5 也展现了类似的能力。

You know, ChatGPT 5.5 also exhibited these capabilities.

根本没人再觉得继续做快照式基准测试还有意义了。

You know, it led no one right to suggest that it's not very relevant to do the snapshot benchmarks anymore.

因为现在的问题是，x 轴得换成时间、令牌数或者算力，毕竟只要给这些前沿模型足够长的时间，大多数问题都能解决。

Like the x axis has to be time or tokens or compute because we can solve most problems now if we just let these frontier models for a very long point in time.

所以加文，这个新型模型到底是什么？

So Gavin, what is this new class of model, right?

Fable 5.5.5。

Fable 5.5.5.

这对超级智能的竞争意味着什么？

What does it mean for the race in super intelligence?

谁现在领先？

Who's up?

谁被甩在后面了？

Who's down?

谁还站在前沿？

Who's still on the frontier?

说说你的看法吧。

Give us your thoughts.

很难说Anthropic现在

It's hard to say that Anthropic's

不算领先。

not up.

对。

Yeah.

你看他们公布营收数据之后，Fable 5发布，还有Mythos明显更厉害。

Like after the revenue numbers they put up, after the Fable 5 release, and Mythos is evidently even better.

但我总觉得，昨天那个Gnome Brown主持的Polynomial节目，观点太深刻了。

But I just think that Gnome Brown host from yesterday, Polynomial, is so profound.

关键是，我们根本不知道这些模型到底有多聪明。

And just the idea that we do not know how smart these models are.

我们得再多说说这个。

We made Say more about that.

那为什么我们不知道它们有多聪明呢？

Why don't we know how smart they are?

因为没人让Mythos跑满一年，我们可能永远没法真正弄清楚每一代模型到底有多强，毕竟还没来得及好好评估，下一个模型就出来了。

Because nobody has run Mythos for a year And we may never know how smart each generation of models actually is or was, but because we don't have time to appropriately evaluate their intelligence before the next model comes out.

这话说得真深刻。

This is a profound statement.

你想想看。

Just imagine.

我老说，一想到全自动驾驶，就想象一个人从不分心、从不疲倦、开车时绝不打电话、绝不酒驾、绝不对孩子大吼大叫，也不用回头给婴儿喂奶。

I always say, when you think about FSD, just imagine a human being who never gets distracted, never gets tired, never talks on the phone in the car, never drinks and drives, never yells at their kids, never has to go to the back seat to give their baby a bottle.

当然，时间一长，这种表现肯定比容易分心的人类强多了。

And of course you would think that over time that is superior to humans who are distracted.

我不清楚你能连续专注思考一个话题多久，布拉德？

I don't know how long How long can you think deeply about one topic, Brad?

大概一个小时吧。

It be an hour.

差不多一个小时。

Be like an hour.

稍微久一点。

A bit.

这让我感觉很难受。

That makes me feel terrible.

因为我觉得自己能连续专注一个话题长达五分钟，才会有杂念冒出来。

Because I think I could think deeply about one topic continuously before having a stray thought enter my mind for maybe five minutes.

然后我就能再回到那个状态。

Then I can come back to that.

想象一下，如果爱因斯坦能一直思考，也许他能连续思考三个小时。

Imagine if Albert Einstein had been able, instead of Maybe he could think for three hours

显然，

at Clearly a

他拥有非凡的头脑。

an exceptional intellect.

但想象一下，如果爱因斯坦每天二十四小时都在思考基础物理。

But imagine Albert Einstein had just thought about fundamental physics twenty four hours a day.

他不需要吃饭，不需要睡觉，不需要休息，也不需要喝酒。

He doesn't have to eat, he doesn't have to sleep, he doesn't have to relax, he doesn't drink.

而且永远不会衰老。

And never gets old.

永远不会

Never gets

衰老。

old.

也不会衰退，没错。

Never has diminution Yes.

这种能力的

Of

他整整思考了一年。

And he thought for one year.

我的意思是，也许我们早就

I mean, we might already, you know

解决掉很多这些棘手的问题了。

Have solved a lot of these intractable problems.

我觉得这个想法真的很了不起。

I just think that's an extraordinary thought.

我之前的看法已经很乐观了，但现在我觉得更乐观了。

My takeaway was, however bullish I was on compute before then, I'm just a lot more bullish.

对，对，没错。

Right, right, right.

我们当时看到的情况，很可能才是真正让 Opus 4.6 突破的关键。

We saw when, that was probably what really unlocked Opus 4.6.

这是第一个能长期保持上下文和记忆、解决复杂长任务的模型。

It was the first really long running model that could maintain that context, maintain that memory, solve some of these longer running problems.

对我们来说，关键信号出现在一月份。

For us, the signal was in January.

我们当时就感觉这是个重要时刻，但后来看到收入开始增长，才明白很多人其实都在默默认可，这个模型真的变得有用多了。

We knew, we felt like that was a big moment, but then when you started to see the revenue go up, we knew that lots of people were voting independently, that that was a profound moment that they became much, much more useful.

但今年年初，大家普遍都有个共识，对吧？

But one of the things that the consensus going into this year, right?

所以今年一开始，大家最关心的问题就是：AI的收入到底会不会出现？

So the big question going into this year was, was the AI revenue going to show up?

我们会不会达到那些让企业和消费者更广泛使用AI的智能门槛？

Were we going to get to these thresholds of intelligence that caused enterprises and consumers to use them more?

当时，至少在这档播客里，和比尔讨论的时候，大家普遍觉得开源模型和低价token正在快速追赶前沿技术。

And I think the consensus at the time, at least on this podcast, debate with Bill, was that open source models, cheap tokens, were catching up on the frontier.

有人认为这些模型可能已经接近瓶颈了，人们不会愿意为高端token付费。

That perhaps these models were beginning to asymptote, that people wouldn't really pay for premium tokens.

但在我看来，今年上半年的实际情况恰恰相反，对吧？

And it seems to me that the evidence on the field six months into the year is just the opposite, right?

前沿模型拿走了绝大部分收入，而且如果你相信长期运行能力和更多算力能带来优势，它们可能反而在扩大对那些基于蒸馏技术的模型的领先优势，对吧？

That frontier tokens are capturing the vast majority of all the revenues and that in fact if you believe in the long running capabilities and more compute allows you to do that, they may actually be extending their lead, right, on some of these models that were built on distillation.

所以我想听听在座各位的看法。

So I just open it up to anyone around the table.

你们觉得，我们有没有推翻之前那个观点，就是说

What are your thoughts on whether or not, you know, have we challenged this thesis that

廉价的开源模型真能一直追上前沿模型，还是说它们的领先优势反而在扩大？

cheap open source tokens are going to always, you know, close the gap on these frontier models or are they extending their leads?

我觉得这个争论从我们刚开始训练这些模型时就一直存在——总说我们落后前沿三到六个月，但事实上，所有收入都集中在前沿模型上，这一点清清楚楚。

I think this debate, like this same debate has existed since the beginning of since we started training these models to begin with which was hey we're always kind of three, six months behind the frontier but empirically like you can just see all of the revenue has actually just accrued at the frontier.

我认为这是因为每次前沿模型发布，都会带来一大批以前根本做不到的新应用场景，比如编程。

And I think that's because every time we release the frontier, a whole new like slew of use cases that previously we could have never tackled like coding.

而且说实话，我们昨天一整天都坐在桌前疯狂用Claude，因为Fable 5.5现在能做的事，跟前一天的Opus 4.8比起来，简直太让人惊讶了。

Also just, you know, we've just been locked at our desks for the last day just, you know, hammering Claude because, you know, it's just fascinating the things that now we can do with Fable 5.5 that we just couldn't do with Opus 4.8just a day before.

那具体是哪些方面呢？

So what are some of those

说来听听，我挺好奇的。

things, man?

我也很好奇。

I'm curious.

我觉得它现在在多智能体协同调度方面特别厉害。

So I think it's really, really good at multi agent orchestration now.

Anthropic 发布了一篇博客，讲了六种不同的智能体编排模式。

So Anthropic released a blog post about like different agent six different agent like orchestration patterns that they've talked about.

但一旦你能掌控这些智能体，它们和模型之间就开始紧密协作了。

But really like once you start being able to manage all these agents, the harness and the model itself is being rrled with one another.

它们正变得越来越融合，而模型还能理解你工作的全貌。比如，我最近把七个模型都扔进去，说：‘帮我整合一下这些公司的假设，像台积电的产能之类的，然后给我出一份综合报告。’

They're actually being fused closer and closer together but the model can understand the extent of your work so you know one of the things for instance is I just threw in like seven of our models and just said, Okay, like I want to create a master view of like my beliefs, given all of these assumptions of all these companies, TSMC capacity, like and then produce me a report on all this stuff.

模型真的能推理出我们所有假设之间的关系。

And you know, the model is able to reason through all of our assumptions.

如果你接受这些前提，那结果就会是……

Actually if you believe this, this thing is

那哪些是矛盾点呢？

What are the contradictors?

没错，正是这样。

Exactly.

是啊，太有意思了。

Yeah, was fascinating.

以前我们可从没这么做过，但现在我觉得，我们才刚迈入多智能体协同的第一步。

And you know, before we never do that, but now, you know, I think we're just step one into multi agent orchestration.

接下来我们会做得更深入。

We're going to do this even further.

我还把所有笔记都丢进去了，它能跨我过去三年的全部笔记进行推理，然后告诉我：这些是你一直坚持的观点。

That's one example I've also dumped all my notes into it, and it's reason across all my notes from the last three years and said, here are some of your ideas that were consistent.

还有那些真正高价值的信源，哪些预测最终成真了，等等。整个过程太让人惊叹了，我们已经远远突破了原来的

Here are like you know the sources that were actually the highest signal to what actually played out you know and then it was actually just super fascinating what you could do and we've just blown through our

限制。

limits.

它正在释放出这么多可能性，我是说，Anthropic 昨天发布时就举过例子。

It's unlocking all this I mean like they gave examples yesterday in the release Anthropic did.

比如 Stripe 那个五千万行的 Ruby 代码库，一天之内就重构完了，以前可得花上几周、好几个人才行。

Know 50,000,000 line Ruby code base at Stripe that was you know refactored in a day versus many weeks with many people.

你想想，这种技术对生物学和生命科学各个领域会带来多大的影响。

You think about where this is impacting biology and life sciences just across the spectrum.

对我来说，这其实回到了一个根本问题。

And to me it really gets back to this fundamental point.

第一，如果你相信长期运行的智能体真的可行，那么未来我们生成和消耗的令牌量将会持续增长，前景无限。

Number one, if you believe this to be true about long running agents then we're going to produce and consume more tokens in the future as far as the eye can see.

所以这个世界，这让我想到TeraFab、Space Orbital这些项目，因为我们可能真的会解锁智能的全新门槛，但要达到那里，我们必须让这些系统长时间运行下去。

So the world, this gets me back you know, TeraFab and Space Orbital and all this because we we we may in fact unlock real thresholds of intelligence, but we're going to have to let these horses run for a long time in order to get there.

是的。

Yeah.

有两件事可以同时成立。

Would just say two thing two things can be true.

大部分经济价值可能依然会集中在前沿技术公司手中。

The majority of economic value may continue to accrue to the frontier.

过去确实如此，今年头六个月更是如此。

Man, has it ever accrued to the frontier thus far, and for sure the first six months this year.

但全球消耗的大部分令牌，可能反而来自开源项目。

But the majority of tokens consumed in the world may be open source.

而它们现在就是这样。

And they are today.

没错。

Yes.

我认为当前这种状态很可能会持续下去。

And I think that this current state is likely to persist.

哈维在X上发了一篇很棒的博客，他们用的是自家的专有法律数据，结合Fireworks对开源模型做强化学习和监督微调，真是让人惊叹，东西过五天就全过时了。

Harvey had a great blog post that they put out on X, and they used, and it's just amazing how everything gets out of date, like, five days, you know?

他们用自家的专有法律数据，通过Fireworks对开源模型进行强化学习和监督微调。

But they used their own proprietary legal data to do reinforcement learning and supervised fine tuning with fireworks on an open source model.

然后他们用了一个路由系统，就是根据问题类型决定把请求发给哪个模型，再用另一个模型来验证结果。

And then they used a router, and a router being something that picks which model you send which query to, and which model you use to check which model.

他们的效果比OPUS 4.7或4.8更好，成本却更低。

And they got better outcomes than an OPUS either 4.7 or 4.8 at a lower cost.

我觉得这才是未来。

And I think that is the future.

但实际上，他们还是在大量使用OPUS，不过他们处理的大部分令牌很可能都来自他们自己的开源模型。

And the reality is they were still consuming a lot of OPUS, but a majority of the tokens they were processing probably were in their own open source

我们听到的情况也一样。

We hear the same thing.

我们做了一项针对300家企业的调查，很快就会发布。

An enterprise survey that we'll post of 300 companies.

哪些公司在做优化？

Which ones were optimizing?

这些公司正在研究模型路由，决定把某些令牌发到特定的模型上。

So these are folks who are kind of looking at model routing and saying we're gonna send certain tokens over here.

哪些公司在考虑优化？

Which ones are thinking about optimizing?

哪些公司还没开始优化？

Which ones aren't optimizing yet?

他们对前沿模型令牌的预期使用量又是怎样的？

And then what is their expected use of frontier model tokens?

没错。

Right.

尽管他们已经在优化了，但还是预计会消耗更多资源。

And they're all expecting to consume a lot more even though they're already in the process of optimizing.

想想摩根大通的情况。

Think of it in the context of JP Morgan.

如果他们做一些后台工作，比如客户服务之类的，很可能会用开源模型。

If they're doing some back of the house stuff, right, on customer service or whatever, they may very well use an open source model.

不过我觉得他们不太愿意用中国开源模型，更倾向于等美国的开源模型真正达到他们需要的性能。

Now I think they're loath to use Chinese open source models, so they're waiting on kind of US open source models to, you know, be able to really deliver the bang that they need.

但我的猜测是，对这些企业来说，大部分后台工作最终都会落在这些模型上。

But my hunch is for these enterprises, a lot of that back of the house stuff will get rooted there.

这些后台任务很可能占大多数的token处理量。

That will probably be a majority of the tokens.

但我觉得那些高价值的任务，比如编程，他们不想用次一等的代码。

But I think the really high value stuff, you know, coding as an example, they don't wanna write second tier code.

我认为绝大部分还是会集中在前沿领域。

I think the vast majority of that will continue to be on the frontier.

订机票根本用不着爱因斯坦。

You don't need Albert Einstein to book you a trip.

做客户身份验证也用不着爱因斯坦。

You don't need Albert Einstein to do KYC.

但这个争论，我们两年前就在这张桌子旁真真切切地吵过。

But but but this is the debate we had it literally at this table two years ago.

不过，如果你单看收入曲线，当时大家下结论时说，因此前沿模型不会获得大部分收入。

However, if you just look at the revenue curves, what bill what what folks concluded they when they said that, they said, therefore, the frontier models will not accrue most of the revenue.

可我们现在看到的情况是，90%的收入都来自这里。

And what we're seeing right now, it's 90% of

他们的判断完全错了。

been decisively wrong.

可能远超90%，而且这种状况还会持续下去。

Probably more than 90%, and it may continue to be decisively wrong.

前沿模型可能占了九成的经济价值。

Frontier might be 90% of the economic value.

开源模型可能占了八成的令牌量。

Open source might be 80% of tokens.

我觉得开源很重要的一点是，很多人误以为这对人工智能是利空的。

That I think is very important on open source is that I think there's this belief that it's bearish for AI.

实际上，这可能对前沿模型才是利空。

It's actually It may be bearish for the frontier models.

就是你提到的那个看空的观点。

There's that bear case you talked about.

但对算力和硬件来说，这其实是利好，因为如果前沿模型赚的利润变少了，大家就会在算力上花更多钱。

It's actually really bullish for compute and hardware because if the frontier models are capturing less of the margin, then you're going spend more on compute.

所以开源做得越好，对算力提供商就越有利。

So the better open source does, better it is for compute providers.

我觉得吧，在西海岸硅谷待久了，和在亚洲待久了，会有两种截然不同的看法。在硅谷待久了，就会觉得所有东西都该是闭源的、上云的，流量都得走这条路。

Will say, there is a very I would say between spending time in the heart of the West, Silicon Valley, and also spending time in Asia, there is a very big, like a deep seated belief in one versus the other, which is like if you spend a lot of time here, it's like all closed source, cloud, every all traffic is gonna go, you know, by way of this direction.

但如果你在亚洲待久了，就会发现大家普遍相信，我们会找到最适合特定任务的模型，不会浪费资源。

And then you spend time in Asia, you know, the overwhelming belief is that we're going to find the right model to the right workload, and we're not going to overspend.

我觉得，接下来这一年最能说明问题了，因为封闭模型之所以能占据这么大价值，是因为它们真能理解意图并把事情做完。

I think, you know, I would say I would say the next year is probably going to be the most indicative of which way this falls because I think the reason why closed source models have captured so much of the value is because the models actually get the intention and actually carry through the work.

今年是第一次，我们有了真正能执行用户意图的智能体——不只是回答聊天请求，而是能产出实际成果。

And this was the first year where we actually had agents that actually carried out user intention from just answering a chatbot request to actually producing useful work.

现在，这种智能水平飞速提升，我们正不断向最具经济价值的任务发起冲击，比如编程、金融，还有各种知识型工作。

Now the level of this intelligence has scaled so rapidly and we continue to push against like the most economically valuable tasks, which are coding and finance and all these like knowledge work tasks.

但对于那些长尾任务，如果开源模型继续保持六个月的滞后，我们可能会看到更多开源工具被用在日常事务上，毕竟——

But like for the long tail of tasks, if open source continues to maintain a six month lag, we might actually see a lot more open source used for, you know, our everyday tasks that we might actually And

这基本上就是詹森的观点。

that's basically Jensen's argument.

对。

Yes.

没错吧？

Right?

詹森的观点是，未来会出现模型路由，现在正好是前沿模型占据优势、能处理长流程任务的时刻。

Jensen's argument is you're gonna have model routing, and we're just in a moment in time where the frontier models gained the advantage, can do long running tasks.

开源模型之前还做不好这些任务。

The open source models couldn't do it very well.

所以所有的价值都被它们拿走了。

And so they're accruing all of the value.

但一旦开源模型也能做好长流程任务——这不会太远了——它们也会分走一大块收入。

But as soon as the open source models can do the long running tasks as well, which is not far away, that they too will grab a bunch a bunch of this revenue.

你们投资了Reflection吗？

Are you investors in reflection?

我们没有。

I'm not.

明白了。

Okay.

我们也没投，但我对米沙和他们的团队所做的事情非常佩服。

Nor nor nor are we, but I I'm very impressed by Misha and and the team and what they're doing.

我真心希望有一个美国的前沿开源实验室能胜出。

I very much want a frontier open source US lab to win.

我们知道，你最近说过，我也觉得没错，英伟达要是真想做，其实早就有了不错的开源模型。

We know that, you know, I heard you say recently and I believe it to be true, Nvidia any day that they really wanted to, right, they already have some great open source models.

他们随时都能打造出一个前沿的开源模型，这根本不是问题。

They could absolutely build a frontier open source model whenever they chose to do it.

所以在我眼里，美国会不会出一个前沿开源模型，根本不用怀疑。

And so it's not a question in my mind as to whether or not The US is going to have a frontier open source model.

唯一的问题只是时间早晚。

It's just a question about timing.

等到那时候，假设他们真的实现了长任务处理能力，

And then, like, at that point in time, is the you know, let's assume they get these long running capabilities.

前沿实验室是不是又能再次抓住优势，继续垄断收入？

Have the Frontier Labs now achieve something yet again that allows them to keep the stranglehold on the revenues?

是的，我觉得如果你说，‘哇，你这ASIC做得真不错。’

Yeah, and I just think it's, if you're, Wow, that's a cute ASIC you've built there.

这太可爱了。

That is so cute.

没错。

Right.

你希望开源怎么参与到前沿领域呢？

How would you like open source to join the frontier?

对啊。

Right.

你希望它怎么实现？

How would you like that?

你喜欢吃新鲜苹果吗？

How do you like live apples?

是的。

Yeah.

我的意思是，我不确定这是他们明确的算计，但我认为詹森

I mean, I'm not sure that's the explicit calculation, but I do think Jensen