20VC：美光将比Meta更有价值 | 出口管制如何助益而非伤害中国 | 电力是人工智能的瓶颈 | 为何达里奥用“劳动力替代”的言论损害了人工智能——与Perplexity创始人Aravind Srinivas对谈

我们当初从聊天转向智能代理时，以为成本会下降。

We thought when we went from chat to agent that costs would go

以为成本和令牌消耗都会降低。

down and token costs would go down.

结果反而上升了。

They've gone up.

是的，目前是这样。

Yeah, for now.

你能帮我理清一下吗？这种情况会怎么变化？

Help me understand that and how that changes.

我觉得在软件领域，你往往愿意为前沿技术买单。

I think in software, you kinda wanna pay for the frontier.

就好比，如果你知道有位工程师特别出色，比如下一个杰夫·迪恩，你会不会宁愿花同样的预算，只招他一个人，而不去雇一堆能力中等、达不到杰夫·迪恩水平的工程师？

It's kind of like, if you know some engineers awesome, if you know you have like the next Jeff Dean, would you rather hire that person and not hire people who are medium engineers, but not Jeff Dean level, with the same amount of budget you have?

对。

Yes.

假设你有一百万美元，你可以雇五个年薪二十万的人，或者雇一个杰夫·迪恩，给他一百万。

Let's say you had a million dollars, you could hire five people worth 200 ks, or you could hire one Jeff Dean and pay them a million.

你会怎么选？

What would you do?

选一个杰夫·迪恩。

One Jeff Dean.

对。

Yeah.

所以我觉得你愿意为前沿技术付费，但什么才算前沿，其实一直在变。

So I think you would pay for the frontier, but what stays frontier keeps changing?

再过十二个月，打个比方，可能会出现一个开源模型，性能跟 OPUS 4.8 一样好。

In twelve months from now, let's say, thought experiment, there is an open source model as good as OPUS 4.8.

但你还是得为推理付费，没有任何东西是真正免费的。

You still have to pay for inference, nothing is truly free.

但那时候成本可能只有现在 OPUS 4.8 的十分之一，再配上合适的智能代理框架和所有连接工具——GitHub 也好，其他开发流程也好，全都顺畅运行。

But it's gonna be like, let's say 10 times cheaper than OPUS four point And when you pair it with the right agent harness, and all the connectors, GitHub, everything, all your developer workflows work fine.

你为什么觉得token的消耗还会这么高？

Why would you assume that the token spend is gonna be still high?

对你现在做的这些事来说，消耗根本不会这么高。

It's not gonna be, for the same things you're doing today, it's not gonna be.

但你可能还没想到，未来在前沿技术上会衍生出完全不同的应用场景。

But there might be a different set of things you might do with the frontier that you're not conceiving today.

我的预测是，会出现完全自主的软件工程师型代理。

My prediction would be agents that are like completely autonomous software engineers.

现在我们都在用ClotCode或者Codecs这类工具写代码，但没人真把自己当软件工程师。

Today I think we're all using tools like ClotCode or Codecs to write code, but not as literal software engineers.

现在有很多人对你们这些前沿模型——比如OpenAI和Anthropic——变得悲观，因为他们发现，用开源模型花一小部分钱也能做很多事。

There is a large way that people that is now bearish on your frontier models who are OpenAI's and your Anthropics because they're realizing that you can actually do a lot with open models for a fraction of the price.

对，确实如此。

Yeah.

你说得没错，我们还是会为前沿技术付费，所以它们依然会创造巨大的价值。

What you're saying is actually that is true, but we will still pay for the frontier, and so they will still accrue great value.

没错。

That's right.

我觉得这个区别听起来像是个矛盾。

And I think this distinction, it feels like a contradiction.

但实际上并不是。

It's not though.

好像两件事不可能同时成立，但事实并非如此。

It feels like two things cannot be true simultaneously, but that's not quite the case.

事实上，我认为前沿技术会越来越成为少数人甚至根本不想接触的东西。

In fact, I would argue that the frontier is increasingly going to be a thing that very few individuals might even want.

比如，等到某个阶段，AI能写代码这件事就不再让人觉得特别了，你已经习以为常了，对吧？

Like you could argue that after a point, it's not even interesting that AI can write software, you've normalized it, right?

我们就假设这种情况会发生。

Let's say that's going to be the case.

未来不会再像过去那样，公司靠成千上万的工程师来构建，反而会有更多公司只保留小规模的开发团队，而每个人都会大量使用AI。

Instead of companies being built with like tens of thousands of software engineers, unlike the past, there'll be a lot more companies with smaller software teams, and each of us will be using a lot of AIs.

这对世界其实是好事，我们会看到更多元化的商业形态，软件和人力会流向过去根本不可能触及的领域。

That's actually good for the world, we'll be seeing a lot of different businesses, we'll be seeing allocation of software and labor in places that was never even possible.

未来的前沿会是AI设计芯片、研发新药、打造机器人、寻找治愈癌症的方法。

Whatever is the frontier is going to be things that AI is going and designing chips, AI is designing drugs, AI is figuring out how to build robots, AI is figuring out how to cure cancer.

癌症。

Cancer.

这些应用的用户不会有上千万人。

These are applications where you don't have like 10,000,000 users.

可能只有几家公司在做，但它们的成果却会影响无数人的生命。

It's like a few companies, but the effect of that work will touch a lot of human lives.

在我看来，前沿正朝着这个方向发展。

I think to me that's where the frontier is headed.

你也能从前沿实验室的动向看出来。

You could also see that from the moves that Frontier Labs are making.

Anthropic收购了一家湿实验室，可能是为了人才，也可能是为了搭建开展湿实验的基础设施。

Anthropic bought a wet lab could be for the talent, could be for the infrastructure to run like wet lab experiments.

但你想想，要是把所有这些数据都用在中期训练里，而不只是用GitHub上的代码。

But imagine taking all those tokens and putting it in the mid training instead of just tokens from GitHub.

那样的话，出来的结果肯定会很不一样。

So then that's gonna produce something interesting.

关掉吧。

Turn off.

解决前沿问题，有没有一个尽头呢？

Is there an asymptote to frontier problems to be solved?

我知道这听起来有点荒唐，但如果你一直追着下一个前沿问题跑，迟早会碰到癌症、气候变化，天哪，我真希望它们都能被解决，最好就在天堂里。

I know that sounds ridiculous, but if you are continuously on the chase for the next frontier problem, you get to cancer, you get to climate change, and my word, I hope they solve both in, like, heaven.

要解决的问题实在太多了。

That's a huge amount to solve.

但如果你一直陷在这种永不停歇的追赶中，这种追赶本身有没有尽头呢？

But if you're on the treadmill of continuosys, there an asymptote to that?

从数学上讲，目前没有任何证据表明AGI或ASI这类系统能创造的经济价值是有上限的。

Well, there's no mathematical argument to there being a cap on the amount of economic value one can create with AGI or ASI like systems.

埃隆有个很好的观点，他说在后AGI经济里，钱会彻底失去意义，因为你能无限生产能源，而经济本质上是建立在能源和劳动力之上的。

And Elon Elon has a good argument for this, they can where he says money loses all meaning in a post AGI economy because you'll be producing an abundance of energy and fundamentally the economy is grounded to energy and labor.

如果你能无限供应这两样东西，那钱还有什么意义呢？

If you can produce an abundance of them, well, what meaning does money have?

所以我不觉得我们在前沿领域会没东西可研究。

And so I don't think we run out of things to solve at the frontier.

我觉得人类的创造力永远不会枯竭。

I think we're always gonna be creative.

人们为什么非要弄明白宇宙的运作原理？

Like why did people even wanna understand the universe?

我们为什么要研究亚原子粒子、量子物理、黑洞理论、宇宙的起源？

Like why did we wanna understand subatomic particles, quantum physics, black hole theory, the origins of the universe?

这到底是为了什么？

Like what is the purpose?

但我们还是去做了，因为探索未知，本来就是人类一直以来的使命。

But we still went ahead and did it because that's kind of what the purpose of humanity has always been, to understand the unknown.

大卫·多伊奇曾说过，我们是唯一一种会对熟悉事物产生好奇的物种。

David Deutsch is famous for saying this, Like we are the only species capable of being curious about what is already familiar.

比如你盯着一个水果，明明知道它是芒果，知道它什么味道、长什么样、什么季节成熟，但你还是能盯着它，提出一个以前从没问过的问题。

Like you can stare at a fruit and you know that it's a mango and you know exactly how it tastes, you know how it looks, you know the shape, you know what seasons it grows in and stuff, but you can still look at it and ask one more question about it that you haven't asked before.

其他动物物种做不到这一点。

Other animal species cannot.

一旦它们在脑子里形成了对这东西的形状、触感等认知，就会直接忽略它。

Once they have it in their mental model what it looks like and touches and feels like, they're gonna ignore it.

对它们来说，这根本没什么意思。

It's it's not very interesting to them.

我能问一下吗？

Can I ask

你刚才提到智能代理的使用，说如果是重复性任务和一次性任务，比如定时任务，有人提到我们会有一个24小时在线的AI，还说会推出一款硬件产品？

you, you mentioned about agent usage and you said if you do repetitive tasks versus one off say cron jobs, I think someone said we're gonna have a twenty four seven AI and they've talked about a hardware product that's gonna come out?

你觉得我们会拥有持续运行的智能代理吗？

Do you think we will have continuous agents running?

是的，我觉得会。

Yeah, I think so.

我觉得这正是我之前说的编排问题的核心——最大化令牌的价值。

And I think that's kind of why I believe orchestration problem I talked about, maximizing the token value.

你能帮我一下吗？

Can you just help me

解释一下？

answer that?

抱歉，你提到的编排问题是指什么？

Sorry, when you say the orchestration problem.

嗯，这么说吧，主要有四个目标：智能和准确度，还有隐私和成本。

Yeah, so okay, so there are like four objectives, intelligence and accuracy, and then privacy and cost.

这些目标之间都是相互冲突的。

These are all competing with each other.

你可以 argue 说，只要建巨大的数据中心，花大量电力运行，就能把智能和准确度做到极致。

So you could argue that you could max out on intelligence and accuracy by building giant, giant data centers and spending a lot of power to run them.

但这样会牺牲隐私和成本，因为所有东西都集中在一起，开销会非常大。

And you could miss out on privacy and costs, because everything will be centralized and you're going to be paying a lot.

也有人会说，所有东西都可以在本地运行，这样对隐私和成本更有利，但可能就达不到最前沿的智能和准确度。

You could argue that everything can run locally, and so that'll be good for privacy and costs, but may not be frontier intelligence, may not be frontier accuracy.

关键是要找到一个平衡点。

The solution is to figure out a sweet spot.

需要的时候用本地模型，需要的时候用服务器模型，在有价值的个人上下文基础上，把两者协调起来。

Use local models when necessary, use server side models when necessary, and orchestrate across local models and server side models, grounded in valuable personal context.

有时候，智能本身已经具备了，但系统就是跑不起来，因为底层工具没搭对。

Sometimes the intelligence might already be there, but the system might not work because the harness isn't grounded in the right set of tools.

所以得打造一个顶级的调度系统，哪怕用个普通的模型，也能让它表现得像顶尖的，还能根据任务和任务的每个环节，选对模型、子代理，甚至充分利用我们自己设备里的计算资源，不一定非得全靠服务器。

So build a world class harness that can even make an okay ish model appear great, and be able to use the right model for the right task, and the right part of the task, sub agents, and even like utilize the compute we all have in our own devices, doesn't need to be always on a server.

这就是个调度问题——一个智能的路由器，一个顶级的调度中枢。

That is an orchestration problem, a router, an awesome router, a master orchestrator router.

要是真能做到这一步，就能实现2047年的AI愿景，还不用让人担心破产。

Now if you do that, you can realize the vision of a 20 fourseven AI, without people freaking out about going bankrupt.

因为没人负担得起在服务器上运行的2047级前沿AI。

Because no one's gonna be able to afford a 20 fourseven AI, frontier AI running on the server.

想象一下，你一旦启动它，就再也关不掉了，除非发生什么极端情况。

Imagine you turned it on and you could never switch it off unless something crazy happened.

大多数人担心这些AI的地方，其实是它会不会做出什么疯狂的事。

The thing that most people worry about those AIs is like, oh, what if it does something crazy?

但真正让人担忧的其实是费用问题。

But the real concern actually is the cost.

没人能承受一个每几秒就运行一次、全天候不停歇的定时任务。

Nobody is gonna be able to afford a cron job at the fidelity of few seconds that runs all the time.

这里的瓶颈其实是任务调度和本地计算能力。

The bottleneck there is actually orchestration and local compute.

所以我相信，我们需要打造一个持续学习的本地模型，能帮你节省压缩开销、优化上下文窗口，尽可能把计算留在本地，只在必要时才调用服务器。

And so I believe one needs to build a continuously learning local model that can save you on like compaction, context windows, and try to preserve as much compute locally, and rely on the server side from here only when necessary.

它要不断学习、不断适应、持续进化。

And keeps learning, keeps adapting, keeps evolving.

这个模型不仅仅是个模型，它还包括配套的框架、本地芯片、计算能力，以及它所控制的整个设备生态系统。

And that model is not just a model, it's a model plus the harness, plus the local chip and the compute and the ecosystem of devices it controls.

这个系统就是你自己的智能。

That system is gonna be your own intelligence.

本质上，数据中心搬到了你的本地设备上，你可以掌控它、拥有它，再也不用担心有人窥探你，或者偷看你的所有敏感信息——那些极其重要的个人数据。

Essentially the data center moved to your local device, and you get to control it, you get to own it, you don't get to worry about somebody like spying on you, or looking at all your tokens, very valuable personal tokens.

想象一下，你手头有一些高度敏感的交易资料。

Imagine you have like very sensitive deal materials.

比如说你在谈一笔交易，结果某个前沿实验室却掌握了你用来写备忘录的所有数据。

Let's say you're doing a deal, and then a frontier lab has all your tokens that you use to like write a memo.

想象一下，要是有人黑进服务器，把你的交易内容偷走了，会怎么样？

Imagine somebody could hack into that server and steal your deal from you.

你肯定不希望这样。

You wouldn't want

对吧？

that, right?

说实话，伦敦那边值得偷的东西多得多了。

I'm gonna be honest, it is much more valuable things for people to steal from London

但说实话，我知道你不是普通的伦敦风投，你上一轮基金规模高达四亿美元，对吧？没错，你确实做到了。

But based yes, I can figure You're not just yet another London VC, you you had like $400,000,000 fund last time Yeah, you read yeah.

所以想象一下，你已经在为四十亿美元的基金布局了。

So imagine like you're already making your moves for the $4,000,000,000 fund.

每个人都有自己的敏感信息，我认为真正能实现这个24.07小时不间断代理的，不会是做模型的公司，也不会是前沿模型开发商，而是那个愿意当协调者的公司。

Everyone has certain levels of like sensitive stuff, and so I think that's where I believe that the $24.07 always on agent is gonna be realized by the company that wants to play the role of the orchestrator, not the model builder, not the frontier model builder, but the orchestrator.

而这，正是我们想做的。

And I think that's we wanna do.

我们把计算机明确地定位为智能协调者。

Computers has been positioned explicitly as the agent orchestrator.

管弦乐队里的乐手，就是那些使用不同模型的子代理。

The the musicians in the orchestra are these sub agents that utilize these different models.

可以把它们看作乐器、工具、连接器和模型——这些全都是乐器。

Think of them as the instruments, and the tools, the connectors, the models, these are all the instruments.

那些子代理就是乐手，整个工作成果是交响乐，系统是乐团，而计算机就是指挥。

The musicians are the sub agents, and the symphony is the work, and the system is the orchestra, and computer is the orchestra conductor.

这正是它的定位，所以它所协调的内容一直在变化，对吧？

That's how it's been positioned, so what it orchestrates keeps evolving, right?

它从模型变成文件、工具、芯片、设备，但其实这些都不重要，只要你能把事情协调好，为每个用户最大化token的价值就行。

It changes from models to files to tools to chips to devices, but it doesn't even matter, like you don't care as long as it orchestrates things correctly, and maximizes the token value for what per user.

如果你能解决这个问题，长期来看，你就能抓住人工智能领域最大的经济价值。

If you can solve this problem, you will capture the most economic value in AI long term.

短期内可能看起来，别的实验室收入在指数级增长，但长远来看，这才是真正关键的目标。

Short term it might look like, oh, like this other labs revenue is growing exponentially does that, but long term, this is the one objective that truly matters.

谁最有可能做到这一点？

Who is best positioned to do that?

我认为是我们。

I believe it's us.

因为你没有追求token最大化，你的目标是为用户创造最大的价值。

Because you have the incentive of not token maxing, you have the incentive of delivering the most value to the user.

只要AI技术栈的任何一部分有提升，我们的产品就会跟着变好。

Every time any part of the AI stack improves, our product improves.

今年年初以来，Anthropic的模型取得了巨大进步。

Since the beginning of the year, Anthropic's models have made tremendous progress.

但同样真实的是，我们的收入自年初以来已经增长了三倍以上。

But what's also true is that our revenue has more than tripled since the beginning of the year.

自从今年年初以来，收入翻了三倍。

Tripled since this beginning of the year.

这很大程度上要归功于Anthropic的模型进步，同时我们也因为OpenAI和他们竞争，降低了同等能力的成本，从而减少了开支。

And a lot of thanks to model progress made by Anthropic, and we also brought our burn down thanks to OpenAI competing with them and bringing down the cost of the same capability.

现在随着开源模型、本地模型和本地芯片的进步，我们会把一部分推理任务迁回本地设备，进一步降低成本。

And now with progress in open source and local models and local chips, we're gonna move some of the inference back to the local devices and bring down the cost even more.

只要AI技术栈的任何一环——不管是芯片、模型还是框架——有所提升，我们的系统就会大幅改进。

Every time any part of the AI stack, whether it's chips, models, harnesses, any of these gets better, our system improves tremendously.

一旦我们的系统显著提升，用户就会更喜欢，愿意花更多钱，我们的业务也就随之增长。

And if our system improves tremendously, our users love it and they pay more, they spend more, and so our business grows.

关于谁最有可能在那个世界中胜出，成为协调者，答案是那个能从任何一层技术进步中获益的产品或企业。

To your question of who's best positioned to win in that world for that objective of being an orchestrator, is the one whose product or business benefits from other people's progress at any layer of stack.

如果黄仁勋做出了更好的芯片，对我们来说是好事。

If Jensen produces a better chip, it's great for us.

如果达里奥推出了更强的模型，对我们来说也是好事。

If Dario produces a better model, it's great for us.

如果苹果推出了更棒的设备，同样对我们有利。

If Apple produces a better device, it's great for us.

我很欣赏的一点是，我们能在整个技术栈的每一层都扮演积极的角色，而不必依赖某一个人的成功。

And I love the fact that we are able to be a very positive some player at every layer of the stack and not have to rely on any one person to win.

当我们评估不同提供商时，会区分服务器端和设备端的方案。

When we look at the different providers that we set kind of server side versus on device.

说到服务器端，很多人在搞AI基础设施泡沫，我觉得这既可笑又愚蠢，简直荒唐。

When we look at server side, a lot of people run AI infrastructure bubble, which I think is funny, stupid and moronic.

以你所见，目前我们面临的数据中心供应问题到底有多严重？

To what extent do we have a data center supply problem today from what you see?

我觉得最大的问题其实出在电力上。

I think the biggest problem is actually in power.

那我们来拆解一下。

So what let's break down.

什么是数据中心？

What is a data center?

难道就是从戴尔或者超微买一堆芯片就行了吗？

Is it like that you just buy like a bunch of chips from Dell or Super Micro?

不是的，那只是其中一部分。

No, that's just one part of it.

你得先搞定土地，要么买要么租，还得买涡轮机发电，或者跟电力公司、电网运营商合作。

You actually have to go secure land, or you have to lease something, lease a property, you have to buy a bunch of turbines to generate power, or you have to work with like power suppliers, grid suppliers.

还得解决散热问题。

You also have to work on cooling.

还有很多其他事情，进展都慢得多。

There's a lot of other work you got to put in that is far far slower.

做这些事之前你得先拿到各种许可。

You got to get permits to do all these things.

所以通常情况是，整个过程耗时很长，而眼下正在使用的模型，都是在上一代算力条件下训练出来的。

And so usually what's happening is, there's a lot of lead time doing this, and the models that are already in use today, these have been trained in the hopper generation.

比如Blackwell世代的第一个模型MITOS，已经让人感到不安了，很多人现在就开始慌了。

So the Blackwell generation model, I think the first model that's Blackwell generation category is MITOS, And it's already scary, people are already like freaking out about it.

想象一下，如果每个人都用百万甚至数十万个黑洞的数据来预训练模型。

So imagine that everyone pre trains a model on like a million or like hundreds of thousands of black holes.

这些模型的性能将远超今天的任何现有系统。

And those models are going to be far more powerful than what exists today.

接下来明年，薇拉·鲁宾天文台将全面投入运行。

And then the Vera Rubins are coming next year in full capacity.

到时，薇拉·鲁宾产生的所有数据都会被用来训练模型。

Like all the data centers of Vera Rubins will be used next year.

那个模型的威力还会更强大。

That model will be even more powerful.

所以我认为，物理基础设施的建设周期总是会成为前沿能力发展的瓶颈。

So I think there is a certain physical build out time that always bottlenecks frontier capabilities.

因此，这一层确实有很高的价值。

That's why there's a value in that layer.

谁能掌握这套方法，把显卡、芯片、网络、电力和散热系统整合起来，再把上面的软件层协调好，最终输出前沿模型的响应，谁就掌握了关键。

Whoever knows how to do this, puts together a bunch of GPUs and chips and networking and power and cooling, and actually like orchestrating all this software layer on top, and is able to convert that into frontier output tokens.

这种端到端的整合能力价值巨大，所以市场给基础设施公司的市盈率，比像Meta这样的公司还要高。

That vertical integration has a lot of value, so that's why the markets are pricing infrastructure companies with a higher PE ratio than companies like Meta for example.

尽管Meta也建了很多基础设施，但它在资本市场被当作一家软件公司来估值。

Even though Meta builds a lot of infra, it's valued as a software company.

但当我们

But when we

看到Meta最近的资本支出大幅增加，还打算继续筹钱扩大支出时，我就明白了——像OpenAI、Anthropic这些AI公司，其实还没从AI产品里赚到钱。

see like Meta's CapEx spend and wanting to increase in the last few days and thinking about raising more and more money to increase CapEx spend, I get it with a lot of the AI providers that your OpenAI is around Anthropic, because they aren't making money from their AI products.

但对Meta来说，资本支出直接关系到广告精准度的提升，能带来6%到8%的收入增长。

For Meta, the CapEx correlates to increasing accuracy on ads, which is like a six to 8% bump in revenue.

我明白，但这么大的资本支出，实在说不通。

I get it, but for the CapEx spend, doesn't make sense.

是的，我认为他们很清楚市场在说什么，他们可不傻，不会看不见这些信号。

Yeah, well, I believe they are understanding what the market's saying, know, they don't think they're dumb to not see what's being said.

我觉得他们正在推出很多订阅类产品。

I think they're introducing a lot of subscription products.

公司不能只做一个社交平台，光靠提升用户参与度来赚广告钱，我觉得他们得推出大量智能代理、订阅服务，甚至搞一个类似马斯克旗下SpaceX那样的云服务平台。

Basically the company needs to not just be a social platform, maximizing engagement and turning that into ad revenue, And I think that requires them to launch a lot of like agents, subscription based products, and maybe even a Meta Cloud that rents out service like what Elon's doing at SpaceX.

等他们真这么做了，市场叙事或许就会改变。

And maybe once they do that, the narrative might change.

但回到我刚才的观点，未来六到十二个月，作为HBM供应商的美光，估值可能反而超过Meta。

But to go back to my point, it might not be inconceivable that Micron, the supplier of HBMs, might be more valuable than Meta in the next six to twelve months.

美光现在市值已经接近一万亿了，而Meta才一千三到一千四百亿。

It's already at like a trillion, and Meta is like 1.3 to 1,400,000,000,000.

你能帮一下吗

Can you help

你能让我明白这一点吗？

me understand that?

因为内存已经成了巨大的瓶颈。

Because memory is already a massive bottleneck.

它的生产成本已经涨了五倍。

It's increased five x in price in terms of the COGS.

对。

Right.

但大家都在说，美光现在股价已经完全反映出来了。

But people are going, wow, Micron is fully priced at this point.

那为什么还说它没被充分定价呢？

Why is it not fully priced?

因为它依然是瓶颈。

Because it's still the bottleneck.

只要是瓶颈，就能掌握定价权。

Whatever is the bottleneck will command the price.

AMD现在表现很好，因为CPU又成了瓶颈。

AMD is doing really well because CPUs became a bottleneck again.

智能代理和代理框架全都在CPU上运行。

Agent loops, agent harnesses are all running on CPUs.

前沿模型用GPU生成token，但像Claw这样的程序，如果它要下载500个来自不同网站的文件，处理大量数据、做各种转换、生成图表，再部署到网站上分享给他人，这些任务全靠CPU来完成。

The tokens are produced by the frontier models on GPUs, but whatever work, let's say Claw generates a coding script that decides to download 500 files from different websites and then munges a lot of data and transforms it in certain ways and generates a plot and then host it on a website that you can share with your people.

所有这些计算都跑在CPU上。

All that compute is running on CPUs.

智能代理对CPU的使用量已经超过人类了。

Agents are using CPUs more than humans.

于是企业级CPU的需求突然上升，受益最大的就是英特尔和AMD这样的公司。

And so suddenly there's a rise in enterprise CPUs, and the beneficiaries of these are like Intel and AMD.

所以谁掌握了瓶颈，谁就能赢。

So then they get to be the bottleneck, whoever's going to be the bottleneck will win.

现在基础设施就是瓶颈，因为需求太大，而供应根本跟不上。

And so infra is the bottleneck right now because there's a lot of demand and we just don't have the supply.

所以，那些供应内存、固态硬盘和CPU计算能力的公司，突然间变得特别重要了，比那些只会建数据中心却不知道怎么把资源变成实际价值的公司强多了。

And so whoever supplies memory, SSDs for storage, CPU compute, suddenly these are all interesting, like they're more important than companies that are just building data centers and not knowing how to turn that into a valuable outputs.

你觉得你的Nebius和CoreWeave未来能持续成为一家市值超千亿的公司，还是只是在解决一个短期的供应问题？

Do you believe your nebius and your coreweaves will be a sustainable multi $100,000,000,000 company in the future, or is it solving a short term supply problem?

我确实觉得它们能持续发展下去。

I certainly think they can be sustainable.

没错。

Yeah.

我觉得其中有些公司我不太看好，老实说，我也说不准哪家最终能胜出，毕竟还有Crusoe、Firebird这些玩家，一溜儿不少公司呢。

I think there are some I don't like, look, I don't know particularly which of those is gonna win, and there's also other players like Crusoe and Firebird and there's a bunch of companies.

关键是要会灵活应对、善于利用资源。

It's all about being resourceful.

你得把数据中心建在自然资源丰富、电力成本低的地方，建设周期短，运营还得稳定可靠。

You got to take power from areas where there's a lot of natural resources and the cost to bring up the data center is pretty cheap and the time to bring up the data center is cheap and your service is reliable.

如果有人跟你签了采购十万块GPU的合同，你的服务就必须得靠得住。

Like if somebody commits to buying 100,000 GPUs from you, the service should be pretty good.

你应该提前锁定供应，做好规划。

And you should be able to secure the supply ahead of time, plan well.

我觉得有些公司甚至在能源层面上创新，自己发电是降低运营成本的一种方式。

And I think some companies are even innovating on the power layer, generating their own power is one way to bring down the margins.

所以我认为这一层确实有价值，因为这种能力很难被复制。

And so I think there's certainly like value in that layer, because it's hard to replicate work.

这就是我的看法。

That's how I see it.

你可能会说，OpenAI 能够完成 CoreWeave 做的所有事情。

You could argue that OpenAI can do all the work that CoreV was doing.

这正是他们搞 Stargate 项目的初衷。

And that's kind of what they wanted to do with Stargate.

但为什么 CoreWeave 在建设数据中心方面比 OpenAI 更成功呢？

But why is CoreVee more successful at building data centers than OpenAI?

确实如此，这事儿特别费操作。

Tots do, it's operationally intensive.

没错，运营起来特别繁琐，你得全神贯注，大部分时间都要花在申请许可、搞定电力供应、排查供应链里的各种瓶颈，还得不断提前规划，仔细测试所有系统，应对数据中心运行中随时冒出来的各种物理问题。

Yeah, operationally intensive, you got to focus, you got to like spend most of your time securing permits, like figuring out power, figuring out like bottlenecks in the supply chain here and there, and constantly plan ahead and like test all these systems carefully, deal with like random physical issues that arise in like running a data center.

有个概念叫总拥有成本，也就是运营开销。

There's something called TCO, cost of operations.

这个成本你必须算进去。

You got to factor that in.

所以这么说吧，如果你只是个服务器租赁商，没什么价值可言。

So that said, I don't think there's value if you're just like a server renter.

如果你只是把GPU服务器机架按小时租给不同公司，这种模式根本赚不到多少钱。

If you're just a GPU server rack renter, if you're just leasing it to different companies on certain hourly pricing rates, there's not a lot of value.

你得在上面加一层软件，就像AWS那样。

You have to actually build some software on top, kind of like how AWS did.

人家叫Amazon Web Services，不叫Amazon服务器，对吧？

It's called Amazon Web Services, not Amazon servers, right?

你得搞一套软件编排系统，这样才能在硬件基础上叠加软件的利润。

You have to have some software orchestration on top that allows you to get software margins on top of what you're doing.

我觉得这就是为什么像 Neebeas 这样的公司会转向 AI 模型推理，去部署开源模型或者托管自己的模型。

I think that's why you're seeing moves like Neebeas, like going for the AI model inference, taking open source models or hosting your models.

像 Fireworks 这些公司，其实就是在做这样的商业模式。

That's a business model of certain other companies like Fireworks and and all that.

但你也可以想象，Neo Clouds 完全可以走这条路。

But you could imagine Neo Clouds just going for that business.

这正是我想问的问题。

That was exactly my question.

我刚请了 Navias 的联合创始人来做节目，他明确提到一个难题：现在有大量资金只想租用算力，但他也清楚，如果想打造一个可持续的长期业务，就必须构建完整的全栈产品。

So I just had the co founder of Navias on the show and the really clear takeaway was the challenge that he has, which is there's a huge amount of money that wants just capacity and compute with the awareness that he needs to build a full stack product if he wants to have a long term sustainable business.

没错。

Yeah.

这让我有了一个关键的领悟。

That was the core realization for me.

当我看推理层的时候，比如你提到的 Fireworks 或 Base 10，你觉得这会怎么发展？

When I look at the inference layer, like you said, Fireworks or Base 10, how do you think that plays out?

单靠推理服务，能诞生一家价值一千亿美元的独立公司吗？

Do we have standalone $100,000,000,000 companies in inference alone?

还是说这种公司正在路上？

Or do we see that coming

关键是要倒着想。

It's all about working backwards.

要打造一家一千亿美元的公司，需要什么条件？

Like, what does it take to build a $100,000,000,000 company?

一百亿美元的年收入。

10,000,000,000 in revenue.

没错。

Exactly.

一百亿美元收入，毛利率三到四成，再加上不错的净利润和现金流。

10,000,000,000 revenue, 30 to 40% gross margins, good amount of net income, good cash flow.

对于一家既能做AI托管推理，又能高效运营服务器和数据中心建设的公司来说，一百亿美元的收入其实并不遥远。

Okay, 10,000,000,000 in revenue is not that inconceivable for a company that can both do AI hosted inference and server capacity and data center build outs very operationally well.

有些因素是他们无法控制的，比如开源模型一直在持续表现优异。

There are some factors beyond their control, like open source models continuing to be awesome.

如果开源模型不再优秀，跟前沿模型的差距拉大到一年半、甚至十八个月以上，那这些公司其实就没什么像样的商业模式了。

If open source models stop to actually be good, where the gap between them, the frontier is like more than twelve months, like fifteen months, eighteen months, then I don't think these companies really have a business model.

因为他们只能把算力租给 OpenAI 或 Anthropic 这样的公司。

Because they're only going to be able to rent capacity to OpenAI or Anthropic.

这正是 Roman 和 Nebula 说过的。

That's exactly what Roman and Nebula said.

他说，如果行业出现整合，只剩下 Anthropic、OpenAI 这样两三个主导玩家，那对 Nebula 来说就是最大的威胁。

He said if consolidation happens and there's Anthropic and OpenAI or two or three dominant providers, that is the biggest threat to Nebula.

没错。

That's correct.

对。

Yeah.

但你得先假设，来自中国或英伟达的模型，还有 Neematron，都在取得不错的进展。

But you gotta make a leap of faith assumption that the models from China or NVIDIA is making good progress on their models and Neematron.

所以市场上会有足够的因素，阻止行业进一步整合。

So there's gonna be enough factors in the market to keep consolidation as an outcome from stopping from happening.

但如果你是这些公司，你就无法掌控自己的命运。

But you don't control your own destiny if you're those companies.

这基本上就是问题所在。

That that that's basically the problem.

明白了。

Okay.

我们完全可以有独立的公司，单靠推理服务就能值一千亿美元。

So we can have standalone companies that are a $100,000,000,000 in inference alone.

抱歉。

Sorry.

我只是在向你请教，汲取你的见解。

I'm just pillaging you for your knowledge.

当我们看像 OpenRooter 或 Factory AI 这样的模型选择公司时，他们刚推出了第二个模型选择或模型路由产品。

When we look at the model selection companies like an OpenRooter or, like, Factory AI just released second model selection or model routing product Yep.

发布时表现非常出色。

Which did very well on launch.

在模型选择和路由这个领域，能值一千亿美元吗？

Is that a $100,000,000,000 companies in the model selection and routing business?

大概不行。

Probably not.

你完全可以只做路由器的供应商，但光有路由器本身，没法产生实际价值。

I think you can just be a provider provider of of a a router, you have to use the router to produce something meaningful.

其实OpenRouter的商业价值主要不在路由功能上，尽管产品叫OpenRouter，它根本不是在不同模型之间做路由。

Actually most of the business value of OpenRouter is less than the router, even though the product is called OpenRouter, it's not routing across models there.

它只是在同一个模型的不同接口之间做调度。

It's actually just routing across different endpoints of the same model.

那我们不妨换个问题。

Okay, so maybe let's ask this question.

如果你是个开发者，想用Claude Opus或者GPT55，为什么不用自己的API密钥，非得通过OpenRouter呢？

If you wanted to use Claude, Opus, or I don't know, like GPT55 as developer, why would you not want to just use it with your own API key versus using it inside OpenRouter?

第一个理由。

Number one argument.

你之所以想这么做的最简单原因，就是模型的备用机制。

The single simplest argument as to why you would want to do that is model fallbacks.

有时候你的API密钥可能没有足够的速率限制，或者即使有，OpenAI的服务器也可能出现故障，无法保证你应用所需的响应时间。

Sometimes your API keys might not have the rate limits, or even if you have the rate limits, there might be an error on OpenAI servers that don't guarantee you the response time you need to run your application.

而OpenRouter会提前用融资资金支付一年的容量费用，确保从多个提供商——比如Bedrock、Azure或OpenAI自身——获取多个端点的速率限制。

And OpenRouter would go and they would pay for capacity for like one year ahead with the funding they have, and secure the rate limits and multiple endpoints across multiple different providers of OpenAI models, be it Bedrock or Azure or OpenAI themselves.

所以这种路由功能很有价值。

And so that routing is valuable.

他们实际上是在解决一个基础设施问题：稳定的令牌供应。

It's essentially an infra problem they're solving, which is reliable token supply.

他们并不是通过判断某个提示该发给GPT、Claude还是别的模型来降低成本。

It's not actually, oh, like they're lowering the cost by deciding if this prompt should go to like GPT or clot or something like that.

这并不是他们真正卖给开发者的卖点。

That's not what they're actually selling to the developer.

这根本不是他们的商业模式。

That's not actually the business model.

对于很多中国开源模型，你可能不希望自己的API密钥流向中国。

And then for a lot of these Chinese open source models, you probably don't want your API tokens from going to China.

假设你没有足够资源去评估和选择不同的推理服务商，分辨谁更可靠。

And let's say you don't have the bandwidth to work with different inference providers that verify who's good and who's not.

你只是把这一切交给OpenRouter处理，由他们为你提供所需的令牌。

You're just trusting OpenRouter to take care of all that, and then they're supply the tokens to you.

这种路由不是在比较哪个模型更便宜、适合什么任务，而是确保令牌供应的稳定性。

So it's routing not at the level of like, oh, deciding which model is cheaper, what task, it's more like a reliable token supply.

我认为这一层确实有价值，否则他们不会有这么多用户，也不会每个月路由数万亿个令牌。

And I think there's some value in that layer, definitely, otherwise they wouldn't have these many users and these many trillions of tokens being routed a month.

但这并不是高利润率的生意。

But it's not like high gross margins business.

他们的商业模式实际上是：通过承诺大量采购，从模型提供商那里拿到折扣，但对用户仍按官方标价收费。

The way the business model works for them is actually, they would secure a discount from the model providers by guaranteeing a lot of supply, but they would still charge the user listing price on the API.

这就导致了他们利润率的差异。

And that differences their margins.

我们之前聊过瓶颈问题，你提到了HBM高带宽内存，还有美光这家公司，以及它们所代表的价值和潜力。

We spoke about bottlenecks and you said about HBM, high bandwidth memory and Micron and the value that they have to say and what it can be.

那三年后会出现什么我们现在还没讨论到的瓶颈呢？

What bottleneck will we have in three years that we're not discussing today?

我觉得电力供应仍然会是主要的瓶颈。

I think power will remain the bottleneck.

我感觉

It feels like

我也是这么觉得。

that to me.

除非数据中心建设的方式发生巨大的变化。

Unless something dramatically changes in the way data center build outs happen.

我确实认为，在建造数据中心方面会遇到很多阻力。

I actually believe that there'll be a lot of resistance to building data centers.

因为人们误以为数据中心耗水量大、耗电多，其实这两点都不对。

It's because people incorrectly think that data centers consume a lot of water or eat up a lot of power, which is both are untrue.

萨提亚甚至说过，这些公司的能效高得就像一罐水一样。

Satya even made the statement that it's like a can of water or something in terms of how efficient these companies are.

你觉得这是他们抵制的原因吗？

Do you think that's why they're putting up resistance in them?

我不这么认为。

I don't.

我觉得是因为它象征着失业、财富差距扩大，还有

I think it's because it's a symbol of job losses, increasing wealth and

原因有很多。

It's a lot of things.

原因挺多的。

It's a of things.

各种担忧和恐惧交织在一起，以多种方式表现出来。

It's a lot of apprehensions, fear about what's gonna happen, channelizing in so many different ways.

有时候，这种情绪通过憎恨财富不平等、想要对富人征税来表达。

Sometimes the channelizing through hatred for wealth inequality and wanting to tax people.

有时候，则是通过担忧环境和气候变化来体现。

Sometimes it's channeling through concerns with environment and climate change.

有时候，人们会觉得：电网涨价是因为你们建了这么多数据中心，或者我的手机、电脑现在更贵了，因为你们把内存都买光了，导致内存价格上涨。

Sometimes it's channelizing in a way where you're all like, oh, the price of the grid is going up because you guys are building all these data centers and then, or like I'm paying more for my phones and laptops now because the RAM prices have gone up because you guys went and bought all of it.

所以我觉得，这些情绪有各种各样的表达方式，但共同点是，大家对人工智能普遍抱有负面看法。

So I think there's a lot of different ways in which it's getting channelized, but the common sentiment is like a pretty bad sentiment about AI.

你觉得呢？

Do you

这会对这些数据中心的建设产生实质性影响吗？

think it will be meaningful to the development of those data centers?

我觉得现在有一半左右的数据中心项目，都因为公众的抵制而搁置了。

I think right now 40 out of a 100 are not being developed because of public resistance.

是的，所以电力供应现在成了瓶颈。

Yeah, so that's where the power bottleneck is.

你可能会看到一些国家抓住这个机会，允许这些模型开发者在那里建设数据中心。

You could see maybe certain countries seize the opportunity for this, and allow these model builders go build data centers there.

埃隆正打算去太空做这件事。

Elon's going to space to do that.

这会是个有趣的实验，因为那里有大量可以利用的太阳能。

So that's gonna be an interesting experiment, because there's a lot of energy from the sun that can be harnessed there.

其他国家拥有丰富的自然资源，监管环境也可能更宽松。

There's a lot of natural resources in other countries, regulations might be more friendly.

数据中心的建设还是会继续，只是可能不会发生在美国了。

So we're still gonna see data center build out, it might not happen in The US.

但你得解决实实在在的物理问题——供应链、审批许可、电力保障，确保一切顺利运转，还要不断缩短交货周期。

But the fact that you have to solve physical problems, like you actually have to deal with the supply chain, the permits, securing power, like making sure like things work and getting the lead times lower and lower.

这可不是简单地复制一个SaaS应用那么简单，对吧？

You're not solving problems like cloning some SaaS apps here, right?

也不是组建一个销售团队，或者用更好的营销手段去对抗竞品。

Or like you're building a go to market team or like doing better marketing against the competitors products.

是的，那些也是难题，但这些问题要难得多，因为你没法完全掌控自己的命运，需要大量资金、人脉，甚至有时候还得靠政治上的助力才能推动进展。

Yes, those are also hard problems, but these are like much harder problems where like you're not in full control of your destiny and you need a lot of capital and connections and like the right people, sometimes even like political help to unlock progress.

所以在我看来，这会继续成为瓶颈。

And so that's why this will continue to remain the bottleneck in my opinion.

而且风险也不小，万一哪天又出现类似DeepSeek那样的时刻——有人用一种完全不同的垂直整合架构，造出了效率高得多的模型。

And there's a lot of risk as well because if you do encounter another deep seek moment here, where there's a vastly more efficient model that's been built with a very different vertically integrated architecture.

你辛辛苦苦建了这么多算力，结果一拍大腿：糟了，我建过头了。

And you built out all this capacity and you're like, damn, that's I overbuilt.

因为现在有更高效的方式，能在普通人的设备上跑，比如MacBook、Windows电脑。

There's something far more efficient that can run on people's local devices, their Macbooks, their Windows PCs.

那时候你肯定慌得不行。

Yeah, like you're probably freaking out then.

不过你觉得这种可能性有多大？

How likely do you think that is though?

大概有两到三成的概率吧。

It's probably like 20%, 30% chance.

我觉得之所以还有这种可能性，是因为出口管制。

The reason I think there are some possibilities that because of the export controls.

DeepSeek 并没有用视频栈，而是用了华为的架构。

So DeepSea is not building within video stack, they're building the Huawei stack.

出口管制不仅针对 NVIDIA 的 GPU，还涵盖了 HPM。

And because there are export controls on not just NVIDIA GPUs, but also on HPMs.

DeepSeek 所构建的这些架构在内存效率上要高得多。

These architectures that DeepSeek's building are far more like memory efficient.

他们对 KV 缓存做了优化，让它小到可以放在 SSD 上，推理时也不需要高带宽内存。

They've made innovations on KV cache to be really small enough that you can host it on the SSDs, and they don't need high bandwidth memory for inference time.

由于不能使用 3D NAND，他们的推理架构和存储架构都得完全重新设计。

And they're gonna have a completely different architecture for inference, completely different architecture for storage, because they're not allowed to use the three d NANDs.

他们的架构不只是模型结构不同，模型本身就已经很不一样了。

So their architecture, it's not just a model architecture, the model architecture is already pretty different.

他们在注意力层做了创新，训练算法也做了调整，这样就不需要大量互联带宽。

They've made innovations on the attention layer, they've made innovations on like the training algorithm, so that it doesn't consume a lot of interconnect capacity.

所以他们的整个技术栈都在向硬件、芯片和晶圆厂等环节深度整合。

So basically their whole stack is getting vertically integrated to their hardware and their chips and their fabs and so on.

这跟美国的做法完全是两种不同的策略。

And so that's a very different bet from what America's making.

你觉得出口管制对我们是有帮助还是有损害？

Do you think the export controls have helped or hurt us?

现在还不好说。

Jury's still out.

短期来看确实有帮助，我认为开源模型和前沿模型之间之所以能拉开十二个月的差距，全靠出口管制。

Short term is helping because my belief, the only reason why there's even like a twelve month gap between open source and frontier is export controls.

所以它确实起到了积极作用。

And so it's definitely helped.

像Anthropic这样的公司，还大力游说推动了这些限制。

And and and definitely like companies like Anthropic lobbied very hard for it.

但也有这种可能：正因为这些限制，他们反而在物理层面上变得特别厉害。

But there is a chance that because of that, they now get really good at the physical layer.

他们还有一个优势，就是能快得多地建成数据中心。

And one advantage they have is they can actually build data centers a lot, lot faster.

电力不是问题，审批不是问题，人手也不是问题。

Power is not a problem, permits are not a problem, people are not a problem.

劳动力不是问题。

Labor is not a problem.

专业能力也不是问题。

Expertise is not a problem.

所以，逼着他们去建这些设施，反而让他们成了更强大的对手。

And so by forcing them to go out there and build all this, you're converting them into a far more potent competitor.

你觉得呢

Do you

我们是不是还是严重低估了中国的实力？

think we still dramatically underestimate China's capabilities?

我觉得是的。

I think so.

如果人工智能不只是数字层面的，而是涉及物理实体的，那就得建晶圆厂、造机器人、生产芯片，还得高效利用能源，把它们集成到本地设备里。

If AI is like not just digital, that's also physical AI, you gotta build fabs, robots, chips, and harness the energy really well, package it into local devices.

我觉得他们相比美国拥有更多优势。

I think they have a lot more advantages than America.

我们美国自己拥有像台积电这样的企业有多重要？

How important is it that we have our own TSMC in The US?

台积电其实已经在那儿了。

TSMC is actually, there

台积电在亚利桑那州建了一个晶圆厂。

is a fab of TSMC in Arizona.

没多少人提到这点，但台积电正投入高达一千五百亿美元在美国建设晶圆厂。

Like not a lot of people talk about this, but TSMC is investing like $150,000,000,000 into building American fabs.

他们已经投了四百亿，或者六百亿，我上次查的时候是这个数。

They've already invested $40,000,000,000 or something like that, 60,000,000,000 last time I checked.

所以，亚利桑那州的台积电工厂马上就要建成了。

So there is a TSMC in Arizona that's coming up.

还有英特尔，所以美国政府持有英特尔10%的股份，英伟达和软银各自持有5%。

There's also Intel, and that's why American government owns 10% of Intel, Nvidia and SoftBank own 5% each.

因此，美国本土的晶圆厂也在大量投入，同时台积电也在加码在美国的晶圆厂建设。

So there is a lot of investment going into an American fab as well as TSMC is investing into its American fabs.

埃隆正在打造TerraFab，看来大家终于意识到建晶圆厂有多重要了，但这也正是中国特别厉害的地方。

Elon's building TerraFab, like I think people have woken up to the importance of building fabs, but this is also why China is particularly very, very competent.

考虑到我们刚才提到的中国那些出色的能力，我知道这问题有点荒谬，但Sota，假如我让你……

Given the capabilities of China that we just mentioned that really articulately, I know it's a ridiculous question, but, Sota, if I were

你的任务是确保美国保持竞争力，面对越来越强大的中国，你会怎么做？

to say to you, your job is to make sure America stays competitive, what would you do to ensure that you retained competitiveness in an increasingly strong China?

我觉得得认真对待实体基础设施，持续投入资金，别再到处传播那些关于数据中心污染水源、耗尽水资源的不实言论了，要基于事实说话。

I think take physical infrastructure a lot more seriously and continue funding it and not, like, have all these I wouldn't say meaningless, it's more like not propagate fake news around data centers about how data centers are polluting and contaminating water or like they're sucking up all the water and actually be fact driven.

我希望我们的产品能帮上忙，你上Perplexity问任何问题，都能验证自己的假设是否靠谱。

And so I hope our product helps there, you can go to Perplexity and ask any question and get fact checked on your assumptions.

但确实很重要的是，我们要用大众听得懂的语言，把真实情况讲清楚，别搞恐吓式宣传。

But yeah, it's very important that we educate the public about what's actually going on in a language they easily understand and not fear monger.

别动不动就说所有工作都要没了，实际上会有大量了不起的公司涌现，只用二三十个人，就能打造出估值上百亿美元的企业，推动数万亿美元的新增GDP。

Okay, like not be like, oh, all their jobs are going to go away, this, like there's going to be lots of amazing companies that are going to get built with far fewer people getting multi billion dollar, multi $100,000,000 evaluations with like twenty, thirty people and propelling like trillions of dollars of new GDP.

咱们该聊聊怎么促成这种局面，怎么去建设它，一起创造一个更积极的未来，而不是老说90%的工作都要消失，说我们的模型会让你们全被坑了。

Like let's talk about how to enable that, let's talk about how to build that, and create a more positive future together, instead of, oh, like 90% of the jobs are going be gone, like you're all going to get screwed over by our models.

我有责任把这一切如实告诉你们。

Like, it's moral duty to tell you all this.

这些话都说烂了。

Like, blah, blah, blah.

这在我听来完全说不通。

Like that doesn't make any sense to me.

你一边抱怨建不了数据中心，一边又说这种话，怎么可能赢呢？

Like you can't win by saying that and also like complaining about not being able to build data centers fast.

你觉得我们是不是因为达里奥一直宣传‘所有工作都要消失’‘末日将至’这种说法，反而造成了严重的误导？

Do you think we've done a complete disservice by having the marketing message that Dario has had that all jobs are going and it's all doom and gloom?

是的，我觉得是这样。

Yeah, I think so.

我的意思是，到目前为止，他们在不同场合说的话其实自相矛盾，我最近听到的说法是，根本没有证据表明AI正在取代工作岗位。

I mean, I think they have contradictory messages in their own different social engagements so far, where the most recent one I heard was there is no evidence that AI is taking over jobs.

这方面需要有一致的沟通口径。

There needs to be a consistent communication around this.

而且我觉得，很少有人谈到AI能以一种完全不同的方式帮你创办公司。

And I also think that very little is being spoken about how AIs can help you build companies in a very, very different way.

就是现在的AI，或者说是智能代理。

Like the current AIs, or Agents AI.

很多原本要雇人做的事，现在用智能代理就能搞定。

So many things you would hire people for, you can do it with agents.

但有人一想就担心：那这些人的工作怎么办？

But one way of looking at it is like, oh, what happens to all the jobs?

但换个角度看，我其实一直有个想法，却从来没机会去实现，也许我和几个朋友能一起把它做起来，你们能不能给我们一些算力支持？

But the other way of looking at it is like, hey, I never had the chance to go build out a company on this idea that I've been having all this while, and maybe me and a group of friends can come together and build this, and can you guys figure out a way to give us us compute credits?

亚马逊就给很多初创公司提供了大量算力补贴。

Amazon gave a lot of compute credits to a lot of startups.

我们刚开始做Perplexity的时候，手头有大约二十万美元的亚马逊、谷歌云和微软Azure的算力额度，加起来总共值一百万美元左右。

When we started Perplexity, we had around $200,000 worth of Amazon credits and GCP credits and Azure credits that together cumulatively this was worth like a million dollars in compute credits.

现在这个年代，一百万美元的算力额度已经不算什么了。

Now in today's world, going be like a million dollars of computer credits.

我们现在正在推动一个叫‘十亿美元建造计划’的项目，只要是一群有潜力打造十亿美元公司的团队，我们就给他们一百万美元的算力支持。

We're doing that, we're funding this thing called a billion dollar build, where we're giving a million dollars of computer credits to any group of people who have a credible path to building a billion dollar company.

我希望未来能有成千上万个这样的公司冒出来。

And I want like thousands such companies to be built later.

你当时没考虑到什么？

What didn't you

萨姆·阿尔特曼不是用两百万美元的代币换取YC公司的股权吗？

think of Sam Altman giving $2,000,000 of tokens to YC companies in exchange?

我觉得我们该多做点类似的事。

I think we should do more of that.

对。

Yeah.

这事儿就该这么干。

That's the right thing to do.

咱们得再多做点这样的事。

We should do a lot more of this.

因为咱们得让新公司冒出来。

Because you want new companies to be built.

就算它们值几亿美元，也是好事。

And even if they're worth multi $100,000,000, it's good.

如果有成千上万家，那可是相当可观的新经济增长。

If there are thousands of them, like, that's a lot of new GDP.

我节目前跟阿莫迪谈过，她说你们团队在AI领域干得真漂亮。

I I spoke to Amodei before the show, and she said how AI pilled the team is for you.

现在团队有多少人？

How big is the team today?

大概四百人左右。

It's like 400 people.

四百人。

400 people.

那两年后呢，

How big will it be in

团队规模会扩大到多少？

two years' time?

我也不清楚。

I don't know.

这个不好说。

It's hard to say.

可能八百到一千人左右。

Maybe 800 or a thousand.

那公司会不会还是按照以前的人数增长模式来走，去解决新的问题？

So will companies follow the same headcount trajectory that they have always followed, and we will just solve new problems?

还是会大幅提高效率，用少得多的人就能搞定？

Or will they be dramatically more efficient with a much fewer number of people?

他们肯定会高效得多。

Definitely they'll be dramatically more efficient.

所以我相信，我们应该打造更高效的公司，自己先做出榜样。

And that's why I am a believer in building a lot more efficient companies, and being an example for all these companies ourselves.

人们应该看看Perplexity，心想：有400人，就能做出一家估值两万亿美元的公司。

People should look at Perplexity and be like, oh, with 400 people, you can build a multi, I don't know, $20,000,000,000 company.

那意味着，只要有40个人，我大概就能打造一家百亿甚至两千亿美元的公司。

And so that means with 40 people, I could probably build a billion dollar or $2,000,000,000 company.

这完全做得到，完全做得到。

And that's totally doable, totally doable.

对我们来说，也许有4000人，公司估值就能达到两万亿美元。

And so for us maybe that means is with 4,000 people, we could be worth 200,000,000,000.

如果有1万人，我们甚至可能值20万亿美元。

We could be worth $2,000,000,000,000 with like 10,000 people.

但这并不意味着，没招那10万个在传统公司里该有的员工就是坏事。

That doesn't mean it's bad for all the 100,000 people we did not hire for a typical $2,000,000,000,000 company.

我宁愿让这一十万个人分成十万个小团队，每个团队一百人左右。

I would rather have those 100,000 people be split into groups of like 100,000 groups like that.

每个这样的团队都能值上几十亿美元。

And each of those thousand groups are worth a few billion dollars.

这太棒了。

That's awesome.

我觉得更多人应该去创业。

And I think a lot more people need to be entrepreneurial.

有些人不管在哪家公司都当不好员工，因为他们难相处、不听指挥、不按计划做事，也不好合作。

There are people who would be bad employees in any company because they're just like difficult to work with, they don't listen to like instructions, so like they don't follow like roadmaps, they're not easy to collaborate with.

但换个角度看，这些特质恰恰是创业者通常具备的。

But maybe the flip side of that is those are the kind of qualities that founders typically have.

阿拉文，有一大群人根本不是数字原住民，也不用AI来提升工作效率，你对他们有什么建议？

Aravind, there is a population and a very large population that are not AI native people, that are not using AI to improve workflows, improve efficiency, what would you advise them?

先行动起来。

Get started.

先行动起来，迈出第一步。

First steps, get started.

把你的好奇心引导到正确的方向上。

And channelize your curiosity.

你并不需要用AI来完成你现有的工作。

You don't need to use AIs to do your existing work.

如果你本来就不喜欢手头的工作，就算用AI来做，你也照样不会觉得有趣。

If your existing work is boring to you, you probably won't enjoy it even if you use AIs to do it.

你有

You got a

因为你说了很多人不喜欢自己的工作，所以你遭到了不少批评，

lot of heat for saying people don't like their jobs, so

如果你真听了我的访谈，就知道我根本没这么说。

I didn't say if you actually listened to my interview, I did not say that.

人们就爱看吸引眼球的标题，把我一句话单独拎出来，断章取义，硬生生搞成新闻标题。

So people want clickbait articles and they take something I said in one sentence and out of context and make it into a headline.

你刚才说什么？

What did you say?

我明明说了，有很多人其实并不喜欢自己的工作。

I specifically said this, hey, like, there are a lot of people who don't enjoy their jobs.

顺便说一句，这段话之所以火起来，不是因为我完全说错了。

By the way, the fact that thing went viral is not because I was completely wrong.

我觉得很多人之所以有共鸣，是因为我坦率地说出了一个事实：很多人确实不喜欢自己的工作。

I think a lot of people resonated with the fact that I was actually honest in saying a lot of people don't enjoy their jobs.

这跟你的经济状况或社会地位一点关系都没有。

That has nothing to do with your economic position or standing in society.

你可能很有钱，但做的工作却让你完全提不起劲，眼睁睁看着人生最黄金的几年，耗在一份糟糕又压抑的工作上。

You might even be like really wealthy, but doing a job that you completely don't enjoy and like destroying like the peak years of your adult life, working on something that is horrible or like depressing.

我的意思是，如果你就是这样，而你一直不敢辞职，是因为总在担心：我怎么从零开始创业？

My point is that if that's you and if the reason you could never leave your job is because you were always worried if you how would you build a company from scratch?

比如要招人、要租办公室，诸如此类的问题——现在这些都已经不一样了。

Like there are all these things to figure out, have to hire a lot of people, you have to like set up an office, this, that, like that's changed.

历史上第一次，你和一两个朋友就能从一个想法开始，真正有机会打造出一家十亿美元的公司。

For the first time in history, you can get started on an idea with like one or two other friends and maybe have a real genuine shot at building a billion dollar company.

我完全理解。

I totally get that.

今天我们聊的所有内容，都是建立在前所未有的持续增长需求基础上的。

Everything that we've discussed today has been on the back of unprecedented demand up and to the right.

我们需要更多内存，更多数据中心的供给，还有按需服务。

We need more memory, we need more data center supply, we need on demand and service.

所有东西都在一路向上增长。

Everything is like up and to the right.

看到优步那边有点松动，他们说：我不确定自己真拿到了预期的效率提升。

Seeing some cracks in an Uber saying, I'm not sure I'm getting the productivity gains that I thought.

微软和他们合作，提供了1500美元的算力预算。

Microsoft aligning with them, putting a $1,500 token budget.

你觉得这种效率持续提升、必须这么做的趋势会一直持续下去，还是会中途遇到波折？

Do you think we will have a continuous up and to the right acceptance that productivity gains are unwavering, we have to do this, or will there be falterings along the way?

我的意思是，我肯定会出现一些波折。

I mean, I'm sure there's gonna

一路上会有些反复。

be falterings along the way.

大家担心令牌用量超标是完全合理的，这也是我觉得你需要某种混合式智能体推理的原因。

And people are rightfully freaking out about token maxing, which is why I think you need some form of hybrid agentic inference.

你需要一部分在本地运行的推理算力，不用按令牌计费，基本上就是按量计费的智能服务。

You need some amount of inference computer run locally that you're not paying for tokens on, on metered intelligence essentially.

未来的顶尖公司会如何安排代币预算？

How will the best companies of the future structure token budgets?

我希望他们根本不用操心这个。

My hope is that they don't have to understand that.

他们可以直接跟一个协调方合作，让对方来搞定。

They will be able to work with an orchestrator who does it for them.

想一直盯着哪个模型擅长什么、怎么分配预算——比如编程用多少、财务用多少——这活儿可不好干。

It's not gonna be easy for you to constantly keep track of like which models are the best at what things and how do you allocate, oh, this is the budget for coding, this is the budget for finance.

你怎么可能搞得清楚，哪个模型在哪些方面更擅长，又该在每个领域花多少钱呢？

Like, how do you even understand like which models are good at each of those things and like how much do you spend on each of these divisions?

你根本没法一直盯着这些细节。

You're not gonna be able to keep track.

我曾经

I had

前几天有个朋友来节目上说，谷歌会成为令牌之王。

a friend on the show the other day say that Google will be the token king.

他们能以最低的成本生产令牌。

They can produce the lowest cost tokens out of anyone.

他们拥有完整的TPU、数据中心、网络和电力采购体系。

Own full stack TPUs, data centers, networking, power procurement.

你觉得他们真的能成为成本最低的令牌生产者吗？

Do you think that's true that they will be the lowest cost token producer?

他们确实具备了成为这样的生产者所需的所有优势。

They have advantages, all advantages one needs to have to be that.

但他们低估了编码模型的重要性。

But they underestimated the importance of coding models.

所以现在他们远远落后于前沿水平。

And so they're far behind the frontier right now.

所以再说一次，他们

So again, they

其实完全有能力追上来，团队非常出色。

could catch up, totally capable, totally competent team.

但目前还没达到前沿水平。

But today they're not quite at a frontier.

前几天我真是吓了一跳。

I was shocked the other day.

我看到Cloudflare宣布，他们的代理流量已经超过了人工流量。

I saw the Cloudflare announcement that now agent traffic has overtaken human traffic for them.

你为什么这么惊讶？

Why why are you shocked?

比我想象的还要快，我个人觉得。

It was quicker than I thought, personally.

我觉得这事迟早会发生，但没想到这么快，可能两年后吧，现在还没到。

I thought that would happen, but in two years, maybe not now.

当智能代理的流量远远超过人类流量时，世界会变成什么样？

How does the world change when agent traffic far exceeds human traffic?

我觉得人们会拥有更多的自主权。

I think people are just gonna have a lot more agency.

那网站会消失吗？

But do websites go away?

设计还重要吗？

Does design not matter?

互联网的广告模式会彻底消亡吗？

Does the advertising model of the internet die completely?

我认为，旅游、购物和时尚领域的广告模式，并不会被智能代理颠覆。

My belief is that the advertising model around travel or shopping or fashion are not getting disrupted by agents.

因为判断不是客观的。

Because the judgment is not objective.

凡是判断客观、交易基于客观标准的，都会被智能代理颠覆。

Anything where the judgment is objective, the transaction is based on objective judgment, that's gonna get disrupted by agents.

凡是交易更主观、决策更依赖个人感受的，

Anything where the transaction is more subjective, like the decisions are more subjective.

比如，这个房间里最好的家具是哪一件？

Like what is the best piece of furniture inside this part?

为什么偏偏选这张桌子？

Like why this particular table?

类似这样的问题。

Or like those kind of things.

至于麦克风这种东西，你大概会做客观判断来买。

Probably for the mic you would buy an objective decision.

对。

Mhmm.

这张桌子，你更在意的是它对房间整体美感的影响。

The table, you probably are caring about the aesthetics of the room.

我觉得世界未来就会这样分化：主观的事还是靠广告驱动，客观的事则由智能代理主导。

I think that's kinda how I feel the world will split, and subjective things will still be ad based, objective things will be agent based.

我看了你在Excel跟Sam对话后做的毕业演讲，他说我必须看，所以我当然就看了。

I watched your commencement speech on the back of speaking to Sam here at Excel, he said I had to watch it, so obviously I watched it.

你提到过，ARR的核心能力是提出更好的问题。

And one of the points you made was the defining skill of the ARR is asking better questions.

对。

Yeah.

现在有什么问题，是根本没人问，但其实每个人都该问的？

What question is no one asking today that maybe everyone should be asking?

我觉得大家该多想想：假设我手上有足够的自主权，我到底该做什么？

I think people need to ask more about like, okay, assuming I have a lot of agency available to me, what do I do?

想象一下，我给你一百万人或者十万人的编制，再配上足够运行这些智能代理的算力。

Imagine like I gave you a headcount of like 100,000 people or 10,000 people, and enough compute credits to run those agents.

你会怎么做？

What would you do?

假设我问你，哈利，假设你现在突然有了上万个智能代理可用。

Let's say I ask you Harry, let's say you have suddenly like 10,000 agents at your disposal.

你会怎么做？

What would you do?

我记得你曾经说过，不是对我，而是在你某个节目里提到过，你做这个播客是因为觉得没别的途径去拿下项目。

I I remember you telling me, or not me, but but in some episode of yours where you said you're only you only did this podcasting because you felt like you didn't have an arbitrage to go win deals.

完全没错。

A 100%.

这就是我为什么还在坚持。

Why I still do it.

我真的很喜欢我做的事，当然了。

I I love what I do, but yeah.

好吧，你现在已经获得了一定的影响力。

Okay, so you've gotten some amount of distribution.

所以现在假设你有一亿美元，可以用来做智能推理，连接所有系统，全部运行顺畅。

So now assuming that you have, let's say you could spend a $100,000,000 on Ingenic inference and grounded with all the connectors and stuff, and it's all working.

有了这种能力，你会怎么用它来推进你的目标？

What would you do with that capability to further your goals?

那你那时的目标又该是什么呢？

What should your goals even be then?

我觉得这才是我真正想问的问题。

I think that's the question I would ask.

假设在未来三到五年，你可以把任何数字任务都交出去，有合适的工具和代理来帮你处理。

Assuming that in the next three to five years, you're gonna be able to delegate whatever digital you want and with the right harness and agents and be able to delegate that.

从根本上说，我会打造一套智能代理系统，用来发现、识别、接触、促成并拿下优质投资，再让媒体内容为它赋能。

Fundamentally, it would be to build an agentic infrastructure to be able to find, identify, outreach, set up, win great investments and have the media sit on top and power that.

这事儿难度极高，但却是投资领域的终极目标。

That is intensely difficult to do and would be the holy grail to investing.

而这一切，正是我最终想要实现的愿景。

But that would power what my end goal ambition is.