Powering AI: Energy, CapEx & the Slow Takeoff

This really comes down to the difference between AI versus non-AI, and what I mean is, for regular data center, that power consumption is basically stable or flat throughout the years, versus for AI data center, that power consumption can really go from one to 100 and back to one, and all of that can happen within one second during training.

这本质上归结为 AI 与非 AI 的区别。我的意思是,普通数据中心的功耗常年基本稳定或持平;而 AI 数据中心的功耗可以从 1 冲到 100 再回到 1,在训练期间这一切可能在一秒之内发生。

If you actually talk to people inside of OpenAI, they will tell you, oh, they actually call AGI as something that will have a slow takeoff, meaning that it's actually more progressive than all of a sudden.

如果你真的去和 OpenAI 内部的人聊,他们会告诉你,他们把 AGI 视为一个会慢起飞(slow takeoff)的过程,意思是它其实是渐进式的,而不是突然发生的。

And I think Zuckerberg also said it in one of his interviews really well — he said, oh, AGI is not just this one thing, because there's no single threshold for humanity, because people can have very different skills.

我觉得 Zuckerberg 在他的一次访谈里也说得很好——他说,AGI 不是单一的一件事,因为对人类而言不存在单一的门槛,因为人与人的技能可以非常不同。

So it's really like, over time, it's more incrementally that we just add different capabilities to it.

所以更像是随着时间推移,我们以增量的方式给它添加不同的能力。

For something like translation, maybe we probably already hit AGI even today.

像翻译这类任务,可能今天我们就已经达到 AGI 了。

Welcome to Autopilot, a podcast where we discuss the past, present, and future of AI and automation.

欢迎来到 Autopilot,一档讨论 AI 与自动化的过去、现在和未来的播客。

We explore how AI is creating and disrupting large industries in conversations with top founders, investors, and historians.

我们通过与顶尖创始人、投资人和历史学家的对话,探讨 AI 如何创造和颠覆大型行业。

I'm your host, Will Summerlin.

我是主持人 Will Summerlin。

Let's dive in.

我们开始吧。

Freda, thank you for joining us on the Autopilot podcast.

Freda,感谢你来参加 Autopilot 播客。

I'm very excited for this episode — you're probably one of the most thoughtful and research-driven AI investors I know.

我非常期待这一期——你可能是我认识的最有思考深度、最以研究为导向的 AI 投资人之一。

So we're going to talk about a bunch of interesting and complex topics today, from the energy side of AI, to what AGI actually means and how far away we are, to models — open source versus closed source — and lots of other things.

今天我们要聊一系列有趣而复杂的话题,从 AI 的能源问题,到 AGI 到底意味着什么、我们离它还有多远,再到模型——开源与闭源——以及很多其他内容。

But before we jump into that, let's start with you.

但在进入正题之前,先从你开始。

You're at a firm called Altimeter — I think most people probably know Altimeter, but for those that don't, can you start by just explaining what Altimeter is and what you do there, and then spend a minute on your background?

你在一家叫 Altimeter 的公司——我想大多数人可能都知道 Altimeter,不过为了还不了解的听众,能否先解释一下 Altimeter 是什么、你在那里做什么,然后花一分钟讲讲你的背景?

Right, so Altimeter is a crossover technology-focused investment firm, so basically we do both VC and also public hedge fund.

对,Altimeter 是一家 crossover 的、专注科技的投资公司,基本上我们既做 VC,也做二级市场的对冲基金。

So I've been with Altimeter for four years, and spent a lot of time recently thinking about some of the big questions in AI.

我在 Altimeter 已经四年了,最近花了很多时间思考 AI 领域的一些大问题。

That's great — we think about AI, like, we can just go ahead and jump in here.

太好了——我们本来就是聊 AI 的,那就直接切入正题吧。

One of the big questions, speaking of big questions, that I think a lot of people are asking is really on the energy side.

说到大问题,我觉得现在很多人都在问的一个大问题,其实是在能源这一侧。

Like, we've seen, as scaling laws hold, the more compute you use to train models, the more data you use to train models, generally the better models perform across a wide range of tasks.

我们已经看到,只要 Scaling Law 成立,训练模型用的算力越多、数据越多,模型在各类任务上的表现通常就越好。

But actually training these models and also running inference on these models takes a ton of power, and it seems like that's become one of the bottlenecks.

但实际训练这些模型、以及在这些模型上跑推理,都要耗费大量电力,而这似乎已经成为瓶颈之一。

Like, people talk about GPUs being the bottleneck, and you know, we've been spending on GPUs — seems like maybe that's easy, we can talk about that a little bit.

大家常说 GPU 是瓶颈,我们也一直在 GPU 上花钱——那部分似乎相对容易解决,这个我们可以稍微聊一聊。

But it seems like energy will be the next bottleneck, and you can't just, like, spin up power grids overnight.

但看起来能源会是下一个瓶颈,毕竟电网不是一夜之间就能建起来的。

So I'm curious how you're thinking about that as a bottleneck.

所以我很好奇,你是怎么看待能源这个瓶颈的。

Right, so first of all, I want to say it's actually not just energy, but really everything that's in and around the data center that has become the bottleneck, and those names have done really well in the stock market year to date.

好,首先我想说,瓶颈其实不只是能源,而是数据中心内部及周边的一切都成了瓶颈,而这些相关公司今年以来在股市的表现都非常好。

And so as an investor, I think it's very interesting to actually see the market has been following and tracking those bottlenecks.

所以作为投资人,我觉得很有意思的一点是,市场其实一直在跟随并追踪这些瓶颈。

And last year at one point — you also mentioned it — it's all about the data and the compute being the limiting factors.

去年某个时点——你也提到了——大家都在说数据和算力是限制因素。

I think it was kind of around Q3 or Q4 last year, that's when the narrative really shifted.

我记得大概是去年三、四季度前后,叙事真正发生了转变。

And I think two things contributed to it: one is that's roughly around the time when H100s were delivered to end customers, and number two, we had some kind of breakthrough in synthetic data that gave us some runway into, call it, 2025.

我认为有两件事促成了这个转变:一是那个时间点前后 H100 开始交付到终端客户手里;二是我们在合成数据上取得了某种突破,让我们有了一段跑道,可以撑到 2025 年左右。

And so people really pivoted and very quickly shifted the focus towards infrastructure and electricity.

于是大家迅速调转方向,把注意力转向了基础设施和电力。

My mental framework on electricity is actually very simple: it's just the supply side, the demand side, and the grid that's connecting the two.

我对电力的思维框架其实非常简单:就是供给侧、需求侧,以及连接两者的电网。

And for both supply and demand, it's not just about the volume, but also really comes down to the type, or the volatility profile, of the supply and demand.

而且无论供给还是需求,关键不只在总量,更在于供给和需求的类型,或者说波动性特征。

So if we look into the demand side, the total power consumption in the US has been pretty flat at around 4,000 terawatt hours, and like 3%, or 17 gigawatts, of that is consumed by data centers.

先看需求侧,美国的总用电量一直相当平稳,大约 4,000 太瓦时,其中约 3%、也就是 17 吉瓦(GW)由数据中心消耗。

If you take Nvidia's number as the input, then in just a few years the total data center will consume closer to 100 gigawatts of power.

如果拿 Nvidia 的数字作为输入,那么只需几年,数据中心的总耗电量就会接近 100 吉瓦(GW)。

And just to put things into perspective, a mega city like New York consumes about 15 gigawatts of power, so 100 gigawatts is definitely a very big number.

给大家一个参照:像 New York 这样的超大城市耗电约 15 吉瓦(GW),所以 100 吉瓦(GW)绝对是个非常大的数字。

But I think what the public don't really appreciate is that it's not just about the volume, but also the difference between AI versus non-AI data center.

但我认为公众没有真正理解的是,关键不只在总量,还在于 AI 数据中心与非 AI 数据中心的差别。

So training itself doesn't consume a crazy amount of power, but this really comes down to the difference between AI versus non-AI.

训练本身并不消耗特别夸张的电量,但关键还是在 AI 与非 AI 的差别。

And what I mean is, for regular data center, that power consumption is basically stable or flat throughout the years, versus for AI data center, that power consumption can really go from one to 100 and back to one, and all of that can happen within one second during training.

我的意思是,对普通数据中心来说,耗电量常年基本稳定或持平;而对 AI 数据中心来说,耗电量可以从 1 冲到 100 再跌回 1,而且这一切在训练过程中可以在一秒之内发生。

And it's just the way how neural network really works.

这就是神经网络的工作方式决定的。

And I think people don't really recognize that kind of volatility that's been added to the load.

我认为人们并没有真正意识到,这种波动性已经叠加到了负载上。

Is all of the volatility on the training side, or is there also a lot of volatility on the inference side?

波动性全都在训练这一侧吗,还是推理侧也有很大的波动性?

Yeah, like, so far what I've been hearing is, like, on the inference side that is probably more controllable.

到目前为止我听到的说法是,推理侧的波动性大概更可控一些。

I guess the logical question then is, like, if the two challenges are the grids that we have today can't handle, you know, the net increase in load demands, and also aren't built for the volatility — like, how do we solve it, and how long is it going to take to solve it?

那么顺理成章的问题就是,如果这两大挑战是:现有电网既承受不了净增的负载需求,又不是为这种波动性设计的——我们该怎么解决,解决又要花多长时间?

Right, right, right.

对,对,对。

So I think given the demand profile that we see, what you really need on the supply side is probably something that's highly dispatchable energy sources, to basically match with demand on a real-time basis.

我认为,鉴于我们看到的需求特征,供给端真正需要的,大概是高度可调度的能源,来实时匹配需求。

But if you look at the supply side, we have, call it, 63 gigawatts of power that will come online in 2024, and 95% of that is solar and wind.

但看供给端,2024 年大约有 63 吉瓦(GW)的电力会并网,其中 95% 是太阳能和风能。

And as we both know, like, solar and wind are not the most controllable type of energy sources, and that has its own set of volatilities.

而我们都知道,太阳能和风能不是最可控的能源类型,其自身就伴随一系列波动性。

So I think that's really creating a problem.

所以我认为这确实造成了一个问题。

And in terms of the solution, I'll say probably just no perfect solutions in the near term, but we can talk about some of the longer-term solutions.

至于解决方案,我想说短期内大概没有完美的方案,但我们可以谈谈一些更长期的方案。

And which I think it really comes down to: we have to see some major breakthrough in energy efficiency, either at the model level or at the chips level.

我认为归根结底是:我们必须看到能效上的重大突破,要么在模型层面,要么在芯片层面。

And the other thing we will see in the long term is we need to have more nuclear power plants coming online.

另一件长期会看到的事,是我们需要有更多核电站并网。

And in terms of the relative near-term solution, one is I think we will see more collocation deals, and the other one, just we will have more gas plants coming online — but even that would probably take three to four years.

至于相对短期的方案,一是我认为会看到更多共址部署交易,二是会有更多天然气电厂并网——但即便是后者,大概也要三到四年。

Yeah, the cycle to actually build a new plant and get it certified and get it online is not something that can happen in a matter of months, right?

是的,真正建一座新电厂、拿到认证、并网发电的周期,不是几个月就能完成的事,对吧?

Exactly.

没错。

Which is challenging, because, like, you could build a data center in six months, uh, if people are building data centers quickly.

这就很有挑战,因为数据中心六个月就能建起来,如果大家在快速建数据中心的话。

So there will be this, like, demand and supply imbalance, I think, at some point in the near future.

所以我认为,在不远的将来某个时点,会出现这种供需失衡。

I'm curious, as you think about this, like, regionally — um, it seems like a lot of the narrative has been the energy constraints, the grid, in the United States.

我很好奇,当你从区域角度思考这个问题时——似乎大部分叙事都集中在美国的能源约束和电网上。

Have you seen any of the hyperscalers or other data center companies looking at other sources of energy outside the United States, where maybe the grids can support higher loads or higher volatility?

你有没有看到 hyperscaler 或其他数据中心公司在美国之外寻找能源,那些地方的电网也许能支撑更高的负载或更高的波动性?

Yeah, I think we started to see some of that, but at the same time, like, if you want to do training or inferencing, it probably also makes sense to be closer to the end customers, so that creates some kind of limiting factors.

是的,我想我们开始看到一些这样的情况,但与此同时,如果你要做训练或推理,离终端客户更近大概也是合理的,这就构成了某种限制因素。

And the other thing is definitely also in terms of national security — whether that's a question on whether you can or should be building data centers outside of the United States.

另一点显然还有国家安全——你能不能、该不该把数据中心建在美国之外,这本身就是个问题。

But for now, we are really seeing this, like, all the hyperscalers are trying to figure out, like, where you can actually put data centers within the United States, and then try to figure out the specific states, and within the state try to find the region to actually put it.

但眼下我们真正看到的是,所有 hyperscaler 都在琢磨数据中心到底能放在美国境内的哪里,然后确定具体的州,再在州内找到实际落地的区域。

Yeah, is cooling also a bottleneck?

对了,散热也是瓶颈吗?

Volatility probably creates some challenges around cooling, um, but also just, like, the size of the training clusters now and the amount of heat that's dissipated — I would imagine cooling has also been a challenge.

波动性可能会给散热带来一些挑战,而且现在训练集群的规模这么大、散发的热量这么多——我猜散热也一直是个难题。

Right, there has been a lot of things happening on the cooling side.

没错,散热这块确实在发生很多变化。

There used to be, like, air cooling, and right now it's moving towards, like, liquid cooling, and so there are a lot of, like, new technologies in terms of that.

以前是风冷,现在正在转向液冷,所以这方面有很多新技术。

But if you really think about what that really comes down to, or the entry barrier of doing cooling, I would say it's somewhat maybe five or six out of 10 — meaning that, today, is it a bottleneck?

但如果你真正去想这件事的本质,或者说做散热的进入门槛,我会说大概是 10 分里的 5 到 6 分——意思是,今天它是瓶颈吗?

Definitely yes.

绝对是。

But if you ask me, in two years, we actually see a lot of startups and new companies getting into the field, so maybe in two years, because the supply and demand will naturally just rebalance each other — so right now it's the bottleneck, and my personal prediction is that maybe in two years that'll be solved.

但如果你问我两年后呢,我们实际看到很多创业公司和新公司正在进入这个领域,所以也许两年后,因为供给和需求会自然地重新平衡——所以现在它是瓶颈,我个人的预测是也许两年后这个问题就会被解决。

Versus for energy, I see that as more of a longer-term kind of imbalance that needs, like, tons of billions of dollars, and we may or may not eventually get there.

相比之下,能源在我看来更像是一种长期的失衡,需要几百上千亿美元的投入,而且我们最终能不能解决还不一定。

One of the other things that I've seen is companies, particularly the startups, like, try to build collocated data centers next to, like, natural gas mining operations, and using, like, the flare-off of the natural gas mines to power generators to power the data center.

我看到的另一件事是,一些公司,尤其是创业公司,尝试在天然气开采作业旁边建共址部署的数据中心,用天然气矿井的放空燃烧气来驱动发电机,再给数据中心供电。

Is that also, like, a viable path, or is that just going to be too small-scale for kind of a meaningful impact on the problem set?

这也是一条可行的路径吗,还是说规模太小,对整个问题起不到实质性的影响?

Right, right, that's nothing — one of the solutions, like, in terms of collocation, people talk about two things: one is you either collocate with the nuclear power plants, or you collocate with the gas plants.

对,对,这是方案之一——说到共址部署,大家谈的是两件事:一是你和核电站共址,二是你和燃气电厂共址。

And I almost think about the gas plants as kind of a backup solution to the nuclear power plants.

而我几乎把燃气电厂看作核电站的备选方案。

And what I mean is that there's only 22 gigawatts of nuclear power plants in the United States, and so each power plant is roughly around one to two gigawatts, so you only have, like, so much nuclear power plants that you can actually go after.

我的意思是,美国的核电站总共只有 22 吉瓦(GW),每座电站大概是 1 到 2 吉瓦(GW),所以你真正能去争取的核电站就只有那么多。

But we have a lot of gas plants — so if you ask is it more ideal than nuclear, I want to say no, but there's just more supply of gas plants.

但我们有很多燃气电厂——所以如果你问它是不是比核电更理想,我想说不是,只是燃气电厂的供给更多。

Got it, that makes sense — interesting.

明白了,有道理——很有意思。

I was talking to somebody recently who worked for a big power company that was building a nuclear plant — this was years ago — and they were building in an area that was totally flat, and they had it almost complete, and they needed, like, a cooling pond, right, so they could pump water into the reactor if something happened.

我最近和一个人聊过,他以前在一家大型电力公司工作,当时公司在建一座核电站——这是很多年前的事了——他们建在一个完全平坦的地区,工程都快完工了,而他们需要一个冷却池,以便出事时能把水泵进反应堆。

And right when it was almost complete, the regulations changed, that required the pump, or the pool, to be gravity-fed.

就在快完工的时候,法规变了,要求那个泵、或者说那个水池,必须靠重力供水。

And they were in an area that was completely flat, so once they had invested billions of dollars and got this plant pretty much complete, they had to build an artificial mountain, essentially, in the middle of, like, a flat area, to have, like, hundreds of feet of elevation to create a gravity-fed cooling system.

而他们所在的地区完全是平地,所以在投入了几十亿美元、电站基本建成之后,他们不得不在一片平地中间造出一座人工山,弄出几百英尺的高差,来建一套重力供水的冷却系统。

So it's, like, it's just such a complicated place to build.

所以说,核电就是这么一个建起来极其复杂的领域。

Um, I think you're right — I think it'll be some period of time before nuclear becomes, like, viable at scale to have, like, a meaningful impact on consumption.

我觉得你说得对——核电要在规模上变得可行、对能耗产生实质性影响,还需要一段时间。

Exactly, exactly.

没错,没错。

Maybe we can switch to kind of the buildout of GPUs and what you're seeing on that side.

也许我们可以切换到 GPU 建设这块,聊聊你在这方面看到的情况。

Like, one of the bearish arguments I think we've all heard against AI, at least in the near term, is that it's been a super capex-heavy endeavor, and, like, big tech companies, hyperscalers, even startups, have spent a ton of money, um, you know, building out data centers, lots of GPU capacity, spent a ton of money on, you know, researchers to train these great models.

我想我们都听过一种看空 AI 的论调,至少在短期内是这样:这是一项极其重 CapEx 的事业,大型科技公司、hyperscaler、甚至创业公司,已经花了大量的钱建设数据中心、大量的 GPU 算力,还花了大量的钱雇研究员来训练这些出色的模型。

Uh, but it seems like, you know, hundreds of billions of dollars in capex spend has not resulted in hundreds of billions of dollars in revenue or profit.

但看起来,数千亿美元的 CapEx 支出并没有带来数千亿美元的收入或利润。

So I'm curious how you're thinking about that, uh, both now and over the next three to five years.

所以我很好奇你怎么看这个问题,包括现在,以及未来三到五年。

Right, very interesting question.

对,非常有意思的问题。

If you look at the four big tech companies, I think combined together they will spend over $200 billion in capex this year, and that's up 40 to 45% versus 2023.

如果看四大科技公司,我想它们加在一起今年的 CapEx 会超过 $200 billion,比 2023 年增长 40% 到 45%。

If you look at it as a percentage of revenue, then we are reaching 15%, versus the long-term average has been 10%.

如果按占收入的百分比看,我们正在逼近 15%,而长期平均值一直是 10%。

So it is really like big tech, big spend.

所以这确实是大科技公司、大手笔支出。

And just as reference, I was looking it up — I think the oil and gas companies combined together, at the peak of their investment cycle, which happened around 2013, I think spent like $160 billion in total in capex.

作为参照,我之前查过——油气公司加在一起,在它们投资周期的顶峰,大约在 2013 年,总共的 CapEx 支出我记得是 $160 billion 左右。

So in some way, the big tech companies today are even spending more than the oil and gas companies at the peak.

所以从某种意义上说,今天的大型科技公司花的钱,甚至超过了油气公司在顶峰时期的水平。

So, um, that's a very fair question that everyone is trying to find an answer to, which is: what about the return, where's the ROI, where's the value creation?

所以,每个人都在试图回答的一个非常合理的问题就是:回报呢?ROI 在哪里?价值创造在哪里?

Um, so I think about it in terms of the application layer and the infra layer.

我是从应用层和基础设施层这两个角度来思考的。

So if you look at the infra layer — so today, before we move into AI, we are spending $200 billion in total on public cloud, and within that, call it maybe $15 billion we spend on CPU.

先看基础设施层——今天,在进入 AI 之前,我们在公有云上的总支出是 $200 billion,其中花在 CPU 上的大概是 $15 billion。

As we move to AI, the chips layer is becoming a lot more expensive, and you can actually see that from Nvidia numbers.

随着我们转向 AI,芯片层变得贵得多,这一点从 Nvidia 的数字里就能看出来。

The cloud layer is still there, right, but on top of the chips layer and the cloud layer, you now have one additional layer called the large language model.

云层还在那里,但在芯片层和云层之上,现在多了一个新的层,叫大语言模型。

So I can easily see how the infra layer spend double, if not triple, in the age of AI.

所以我很容易想象,在 AI 时代,基础设施层的支出会翻倍,甚至增至三倍。

Then the question really becomes, what's going to happen to the application layer?

那么问题真正变成了:应用层会发生什么?

There are really two possibilities: so either that layer will just be squeezed, because the infra layer is just taking up more space, or the other possibility is that the total pie is so much bigger, so the application layer can maintain the profit margin.

其实有两种可能:要么这一层被挤压,因为基础设施层占据了更多空间;另一种可能是,总的蛋糕大得多,所以应用层可以维持住利润率。

Like, we just don't have enough data points today to really see, oh, which way is going to be the future.

我们今天还没有足够的数据点,能真正看清未来会走哪条路。

But if you put a gun to my head, I want to say it's definitely somewhere in the middle, but I tend to think it might be closer to the former one, which is the application layer getting squeezed a little bit.

但非要我选的话,我想说答案肯定在中间某个位置,不过我倾向于认为它可能更接近前者,也就是应用层会被挤压一点。

I'm sure you follow the most recent, like, earnings, like the most recent quarters.

我相信你也在跟最近的财报,最近这几个季度的。

Said in another way, the Microsoft, or the hyperscalers, are really sucking the oxygen out of the room.

换个说法,Microsoft,或者说这些 hyperscaler,真的是把屋里的氧气都吸走了。

The hyperscalers are very aggressively absorbing their suite — whether it's Fabric or Data Factory, they want to provide this all-in-one solution to the customers, so customers no longer have to purchase the packaged software solutions like they used to do.

hyperscaler 正在非常激进地扩充自家套件——无论是 Fabric 还是 Data Factory,它们想给客户提供一体化的解决方案,这样客户就不再需要像过去那样购买打包的软件方案。

It seems like AI coding assistants, and eventually, like, AI coding agents, will accelerate that cycle, right?

看起来 AI 编程助手,以及最终的 AI 编程 agent,会加速这个周期,对吧?

To the extent you have, like, the fundamental infrastructure components, um, and you can just, like, piece them together with API calls — like, you can have, like, one developer who has, like, a really good AI coding agent, um, sitting next to them, and they can build, like, pretty much any custom workflow software solution you need in your company.

只要你有那些基础设施的基本组件,就可以用 API 调用把它们拼在一起——你可以让一个开发者,身边配一个非常好的 AI 编程 agent,他们几乎可以构建你公司需要的任何定制化工作流软件方案。

Like, you don't need to pay Salesforce $100,000 a year for, like, a pretty basic CRM, when you can create a highly customized CRM in a matter of a couple of days with a competent software developer and AI coding assistant.

你不需要每年付给 Salesforce $100,000 去买一个相当基础的 CRM,因为一个称职的软件开发者加上 AI 编程助手,几天之内就能做出一个高度定制化的 CRM。

Right, for sure, for sure.

对,当然,当然。

Like, I think it's actually one thing for AI to really build an entire software — we may or may not see that, but that's definitely a three-to-five-year thing.

我认为让 AI 真正构建一整套软件是一回事——我们也许能看到,也许看不到,但那肯定是三到五年之后的事。

But today, I think with all the API calls and everything, or the hyperscalers really willing to provide all the software tools to the customers, the customers no longer have to go to independent software vendors to actually do what they used to do.

但在今天,有了这些 API 调用等等,或者说 hyperscaler 非常愿意把所有软件工具都提供给客户,客户就不再需要去找独立软件供应商做他们过去做的那些事了。

And if you think about it, it's not just the cloud providers — like, even Nvidia, for example, apparently, even with all the revenue and profit margin, they are still not happy just being a GPU provider — that's their culture.

而且你想想,这不只是云服务商——比如即便是 Nvidia,显然,就算有那样的收入和利润率,他们仍然不满足于只做一个 GPU 供应商——这是他们的文化。

So they also want to sell you the cloud solutions, they want to sell you the software solutions.

所以他们还想卖给你云解决方案,想卖给你软件解决方案。

And I think they disclosed that they already have a few billions of dollars in revenue of their software solutions.

我记得他们已经披露,他们的软件解决方案已经有几十亿美元的收入了。

So I was like, well, just think about it — how many software companies there will be to actually generate, like, a few billions of revenue.

所以我当时就想,你想想看——有多少家软件公司能真正做到几十亿美元的收入。

Um, so definitely a lot of moving pieces in the market.

所以市场上确实有很多变数,而且都还在变化之中。

That's insane — I feel like people don't talk about the software side of their business, or at least haven't paid much attention to it.

这太夸张了——我觉得大家不怎么谈他们业务里软件这一块,或者至少没怎么关注过。

I'm curious, like, what does this mean for all the independent, like, infrastructure software companies?

我很好奇,这对那些独立的基础设施软件公司意味着什么?

Like, you see a lot of vector database companies, and, like, lots of — I mean, we can name every category — but there are lots of categories that kind of sit somewhere between, like, applications and the hyperscalers.

你能看到很多向量数据库公司,还有——其实每个类别都能数出一串——有很多类别大概处在应用和 hyperscaler 之间的位置。

Do you think they're going to have a tough, you know, five years ahead, as the hyperscalers continue to push up the stack?

你觉得随着 hyperscaler 继续沿技术栈向上推进,它们接下来五年会很难熬吗?

I think that's what the public market is essentially telling you.

我认为这基本上就是二级市场正在告诉你的答案。

Uh, I think it really — like, for now, we don't have a perfect answer to know, oh, is the hyperscalers going to be stronger, or those independent, like, infrastructure software companies going to have the upper hand.

我认为——目前我们还没有一个完美的答案,来判断到底是 hyperscaler 会更强,还是那些独立的基础设施软件公司会占据上风。

But like, I think we'll probably see that over time, but right now we just don't have a perfect way to predict that.

但我认为随着时间推移我们大概会看清楚,只是现在我们还没有一个完美的方法去预测。

We've also seen, like, several large startups whose core business is essentially, like, acquiring GPUs and then renting those GPUs out, which is really what the hyperscalers do.

我们也看到几家大型创业公司,核心业务本质上就是采购 GPU 然后把这些 GPU 租出去,这其实就是 hyperscaler 在做的事。

But, like, these companies have gotten, you know, access to GPUs in an environment where there's supply constraint.

但这些公司是在供给受限的环境下拿到了 GPU。

Do you think that's, like, a viable business model moving forward, or do you think we're going to get to, like, a point where the supply constraints are relieved to the point where, like, the hyperscalers have plenty of capacity and you don't need to use those third-party providers?

你觉得这往后是一个可行的商业模式吗?还是说我们会走到某个点,供给约束缓解到 hyperscaler 有充足产能,你就不再需要那些第三方供应商了?

Right, so definitely the supply-demand dynamics is really important to actually make the prediction.

对,所以供需动态确实对做出这个预测非常重要。

And the other thing, I think Nvidia is actually important enough to actually decide the fate of a lot of those companies — like, in the case of maybe, like, CoreWeave, they really have this special relationship with Nvidia, and Nvidia wants them to be successful, Nvidia wants them to be a viable cloud solution.

另外一点,我认为 Nvidia 实际上重要到足以决定其中很多公司的命运——比如 CoreWeave 这样的例子,他们和 Nvidia 有非常特殊的关系,Nvidia 希望他们成功,Nvidia 希望他们成为一个可行的云解决方案。

And just on this point, while we're here — like, do you think the hyperscalers are going to try to directly compete with the foundation model providers, like OpenAI, Anthropic, uh, xAI?

顺着这个话题,趁我们说到这儿——你觉得 hyperscaler 会去直接和基础模型提供商竞争吗?比如 OpenAI、Anthropic、xAI?

They've had, like, this basically partnership model in the past, and maybe it's similar to the relationship between the hyperscalers and application companies, where they're frenemies.

它们过去基本上是一种合作伙伴模式,也许类似于 hyperscaler 和应用公司之间的关系,亦敌亦友。

And, like, we've seen these strategic deals between, you know, Microsoft and OpenAI, and Google Cloud and Anthropic.

我们也看到了这些战略合作,比如 Microsoft 和 OpenAI,Google Cloud 和 Anthropic。

I'm curious, like, what that's going to look like in three to five years — whether or not the hyperscalers are going to look around and say, let's just, like, build these models ourselves, versus relying on a partner to do it.

我很好奇三到五年后这会是什么样——hyperscaler 会不会环顾四周然后说,我们干脆自己造这些模型,而不是依赖合作伙伴去做。

I think in terms of building the models, like, right now, in terms of the training cost, we are already at roughly $1 billion, so the next version might be like $10 billion total training run.

我认为在造模型这件事上,现在训练成本已经大约在 $1 billion,所以下一个版本可能是总计 $10 billion 的训练任务。

So very soon, I think we'll just run out of the long tail of model companies.

所以很快,我认为长尾的模型公司就会所剩无几。

So in terms of the relationship, it's like, they definitely do not want one single model company to be, like, too powerful.

所以在这层关系上,它们肯定不希望任何一家模型公司变得过于强大。

They also want to keep their internal teams, so they have some kind of lever when they are talking to those model companies.

它们也想保留自己的内部团队,这样在和那些模型公司谈判时手里有筹码。

But in terms of whether the cloud companies will really be the ones to have the best models internally, I think increasingly that's probably unlikely, given how capital-intensive it is to actually train those models.

但至于云公司会不会真的成为内部拥有最佳模型的那一方,我认为这越来越不可能,因为实际训练这些模型的资本密集度太高了。

And since we're on this point, like, how do you think about open source versus closed source here on the model layer?

既然聊到这一点,你怎么看模型层的开源与闭源之争?

Right, so, like, first of all, I want to clarify — I don't think people are calling it the right name.

好,首先我想澄清一点——我觉得大家的叫法并不准确。

Like, Llama or Mistral, those companies, they are open weight, but not really open source.

比如 Llama 或 Mistral,这些公司做的是开放权重(open weight),而不是真正的开源。

So you are given the parameters of the model, but you are not really given the pre-training dataset or the training secrets behind it.

你拿到的是模型的参数,但拿不到预训练数据集,也拿不到背后的训练秘诀。

If you think about it, Android or Linux, those are true open source.

仔细想想,Android 或 Linux 才是真正的开源。

In terms of open versus closed, I think the question people are really trying to get to is: is the model layer really getting commoditized, because Llama 3 was getting really powerful.

关于开源与闭源,我认为大家真正想问的问题是:模型层是不是真的在被商品化,因为 Llama 3 当时变得非常强。

And, um, I think about it in two ways.

我从两个方面来看这个问题。

One is in terms of the model capability — uh, I keep reminding myself that GPT-4 was actually released in March 2023, and it's been 16 months, and so far we don't have any model, maybe other than Claude 3, that really surpassed GPT-4's level.

一是模型能力——我一直提醒自己,GPT-4 其实是 2023 年 3 月发布的,至今已经 16 个月,到目前为止,或许除了 Claude 3,还没有任何模型真正超越 GPT-4 的水平。

So from a model capability, is it getting commoditized?

那么从模型能力看,它在被商品化吗?

I want to say no.

我想说没有。

And the other point is, if you go back to the technology, or how the models are really built — um, I mean, Llama 3 is very impressive; it basically continued to train the model beyond an optimal point, to basically change the efficient frontier of model training.

另一点是,如果你回到技术本身,也就是模型究竟是怎么构建的——Llama 3 确实非常令人印象深刻,它基本上是在超过最优点之后继续训练模型,从而改变了模型训练的效率前沿。

But I always think about it in this way: if you give OpenAI or any other model company the same amount of GPU and resources, they can probably do the same, if not even better, versus Llama.

但我总是这样想:如果给 OpenAI 或任何其他模型公司同样数量的 GPU 和资源,他们大概率能做到和 Llama 一样,甚至更好。

So what that really tells us is that nothing today has really deviated from the scaling law.

这真正告诉我们的是,今天没有任何东西真正偏离 Scaling Law。

And what does the scaling law tell us?

那 Scaling Law 告诉我们什么?

It's just like, bigger models are just better models.

很简单:模型越大,就越好。

So so far, with everything we've seen, um, I don't really think there's, like, a true risk of the model layer getting completely commoditized.

所以就目前我们看到的一切而言,我并不认为模型层存在被彻底商品化的真实风险。

It seems like, you know, as long as scaling laws continue to hold, um, these training cycles are going to get more and more expensive, and you really will need, like, massive economies of scale in order to make the underlying economics make sense, right?

看起来,只要 Scaling Law 继续成立,这些训练周期就会越来越贵,你真的需要巨大的规模经济,才能让底层的经济账算得过来,对吧?

If you're spending a billion dollars on a training cycle, right, like, you probably need to produce, you know, at least two, three, four, five billion dollars in profit off of that model in order for it to make any sense.

如果你在一个训练周期上花 $1 billion,你大概需要从那个模型上产出至少 $2、$3、$4、$5 billion 的利润,这才说得通。

And you're only going to do that if you have a large number of customers, which you're only going to have if you have one of the best models.

而只有拥有大量客户你才能做到这一点,而只有拥有最好的模型之一,你才会有大量客户。

And so the sort of economies-of-scale dynamic will likely, to your point earlier, like, result in kind of a winner-take-most, like, outcome here, where you have a couple of companies that are able to acquire the resources to train best-in-world models.

所以这种规模经济的动态,很可能像你之前说的那样,导致一个赢家拿走大部分的结局——只有少数几家公司能获得资源,去训练世界最强的模型。

For sure — the state-of-the-art models acquire more customers, generate more revenue, which gives them the resources to train the next generation of models.

当然——最前沿的模型获得更多客户、产生更多收入,这又给了它们训练下一代模型的资源。

So it's like a very clear logical path forward, where you end up with, you know, a couple of winners.

所以这是一条非常清晰的逻辑路径,最终你会得到少数几个赢家。

Yeah, exactly.

对,正是如此。

And just to give you some numbers on that: so I think GPT-4 was trained on 8,000 H100 equivalents for about 100 days — the total training cost, if you do the math, is around like $300 million.

给你几个数字:我记得 GPT-4 是用相当于 8,000 张 H100 训练了大约 100 天——算下来,总训练成本约 $300 million。

And for GPT-5, I think the rumor was that that was trained on 30,000 H100s for about a few months, and the total training cost will be closer to $1 billion.

至于 GPT-5,传闻是用 30,000 张 H100 训练了几个月,总训练成本会接近 $1 billion。

And of course, you can say, oh, I don't want to hold it on my balance sheet — how about I rent it from the hyperscalers?

当然,你可以说,我不想把它放在自己的资产负债表上——那从 hyperscaler 那里租怎么样?

And the cost for that is $10,000 on a per-GPU-per-year basis, and so that will be $500 million on a per-year basis.

租的成本是每张 GPU 每年 $10,000,那一年下来就是 $500 million。

And as we move to the version of the model that will be in training in 2025 or 2026, uh, I think that training cost is going to be $5 to $10 billion.

而当我们进入 2025 或 2026 年将要训练的那一代模型,我认为训练成本会到 $5 到 $10 billion。

And you are already seeing xAI building this 100,000-H100 cluster, and that in itself is a $4 billion total investment — and again, if you rent it from the cloud, that will be $2 billion per year.

你已经能看到 xAI 在建那个 100,000 张 H100 的集群,仅这一项就是 $4 billion 的总投资——同样,如果从云上租,那就是每年 $2 billion。

And I think, like, Jensen was already talking about 1 million GPUs connected together for training, and that's easily, like, a $50 billion, if not even more, of training cost that we are talking about.

而且 Jensen 已经在谈把 100 万张 GPU 连在一起做训练,那我们谈论的训练成本轻松就是 $50 billion,甚至更高。

So those are very big numbers.

这些都是非常大的数字。

And, um, I think, to your point, the implication, or the good news from it, is just, very soon the long tail of model companies will just, like, disappear.

我认为,如你所说,这背后的含义,或者说好消息是:很快,模型公司的长尾就会消失。

And, um, I think the end state of the model companies will look very similar to something like the operating systems, where you have one or two oligopolies, and at the same time you have something like Linux, which is open source.

我认为模型公司的终局会非常像操作系统的格局:有一两家寡头,同时存在一个像 Linux 这样的开源方案。

Yeah, it seems like we're already sort of seeing that play out, right?

是啊,看起来这个局面已经在上演了。

Like, OpenAI's at a couple billion ARR, Anthropic's at a couple hundred million, Cohere is at like 15 million ARR, as of a month or two ago, and Cohere is, like, arguably the number three.

OpenAI 的 ARR 有二十来亿美元,Anthropic 是几亿美元,Cohere 大概是 1500 万美元 ARR——这是一两个月前的数据——而 Cohere 可以说是第三名。

Now you have xAI, which totally changes the dynamic — but it's like, the top two were starting to run away with the lead already; now xAI, given the amount of compute they have, will probably be in an interesting position.

现在又有了 xAI,这完全改变了格局——但前两名本来就已经开始甩开身位领跑了;现在 xAI 凭借他们手里的算力,可能会处在一个很有意思的位置。

I'm curious, though, on the open source, or open weight, side — there's, like, a counterargument to all of this, which is, for a lot of applications, like, you don't need a huge model.

不过我好奇的是开源、或者说开放权重(open weight)这一边——针对上面这一切,有一个反方论点:对很多应用来说,你并不需要一个巨大的模型。

Like, if you have an application that has, like, kind of a constrained, narrow set of use cases — maybe it's an application that only answers, you know, a small subset of questions, doesn't have to have open-ended conversations, like, doesn't need to be multimodal — you can get away with a much smaller model, um, whether you take an open source model and fine-tune that, or train your own model on some proprietary dataset.

如果你的应用只有一组受限的、狭窄的使用场景——比如它只回答一小部分问题,不需要开放式对话,也不需要多模态——你完全可以用一个小得多的模型,不管是拿一个开源模型来微调,还是在某个专有数据集上训练自己的模型。

And the inference cost of these smaller models is likely going to be a lot lower than, like, the inference cost of GPT-5.

而这些小模型的推理成本,很可能会比 GPT-5 的推理成本低得多。

And, uh, it seems like that's a case where the open weight models could end up, like, uh, creating some dent in the opportunity set for the giant foundation models.

看起来在这种情况下,开放权重模型最终可能会在巨型基础模型的机会盘子里啃掉一块。

I'm curious if you think that's a possibility, or the capability set of the large models will just be so incredibly good that people will abandon using smaller models and trying to optimize for inference cost.

我好奇你觉得这是一种可能性,还是说大模型的能力会好到令人难以置信,以至于人们会放弃使用小模型、放弃为推理成本做优化。

Yeah, that's a fair point.

是,这个观点有道理。

I think definitely there will be some use cases that you simply do not need the best models to actually do that, and in that case, I think open source, or open weight, definitely has a place to go.

我认为肯定会有一些使用场景,你根本不需要最好的模型也能做到,在那种情况下,开源或开放权重(open weight)肯定有它的空间。

And — but the other thing is, I never really think about the model companies as just having one model.

但另一点是,我从来不把模型公司看成只有一个模型的公司。

I almost think about them as a model factory, or like a model store.

我几乎把它们看成一个模型工厂,或者说一个模型商店。

And of course, today they are going towards AGI, right, so they want to spend all the efforts into building the best and the largest models.

当然,今天它们都在奔向 AGI,所以它们想把所有精力都投入到构建最好、最大的模型上。

But if you really think about it, it's probably very simple for them to actually create a 70-billion-parameter model, or even like a 1-billion-parameter model.

但你仔细想想,对它们来说,做一个 700 亿参数的模型,甚至一个 10 亿参数的模型,大概是非常简单的事。

And I think it's more likely that they can do it better versus some of the small model companies, or the open source or open weight companies.

而且我认为,比起一些小模型公司、或者开源和开放权重公司,它们更有可能把这件事做得更好。

And I think that's kind of the potential competition that I see from the larger model companies, even when they have the time and energy to focus on some of the smaller side.

我认为这就是我看到的来自大模型公司的潜在竞争——一旦它们腾出时间和精力来顾及小模型这一侧。

That makes total sense — and we'll flip back to the capex question in a second.

这完全说得通——我们稍后再绕回 CapEx 的问题。

One of the concerns I think I have about the foundation model companies, and also the hyperscalers, is they're investing a ton of capex to create models that are really useful, and that will likely push forward, um, you know, the rate of new technology adoption, the rate of AI adoption — you have more capabilities, you can solve more problems, it will catalyze growth of the entire ecosystem, increase the TAM for the technology industry at large.

关于基础模型公司,也包括 hyperscaler,我有一个担忧:它们投入了海量 CapEx 去打造真正有用的模型,这大概率会推动新技术采用的速度、AI 采用的速度——能力更强了,能解决的问题更多了,会催化整个生态的增长,扩大整个科技行业的 TAM。

Like, all these great things will happen, but I think there is a risk that, like, they're unable to capture a lot of the economic value that's created.

所有这些好事都会发生,但我认为存在一个风险:它们可能无法捕获所创造的经济价值中的很大一部分。

And I think, like, yeah, an analogy that I like, or have heard — you know, history doesn't repeat, but it rhymes — and if we go back to the Industrial Revolution, which I think has some parallels to where we are today.

我喜欢、或者说听过的一个类比是——历史不会重演,但总押着同样的韵脚——如果我们回到工业革命,我认为它和我们今天的处境有一些相似之处。

And you go back to the Industrial Revolution England — there's an index of railroad stocks, and if you would have invested in the index at the beginning of the Industrial Revolution and sold at the end of the Industrial Revolution, you would have actually lost money, which is, like, insane, given how much value is created in England in the Industrial Revolution.

回到工业革命时期的英国——有一个铁路股指数,如果你在工业革命开始时买入这个指数,在工业革命结束时卖出,你实际上是亏钱的,这太离谱了,毕竟工业革命在英国创造了那么多价值。

And the reason was twofold: first, these railroads were, like, very capex-heavy projects, where you had to raise a bunch of capital and build out the railroads.

原因有两个:第一,这些铁路是非常重 CapEx 的项目,你得募集一大笔资本去把铁路修出来。

The second was they became super competitive, and so you'd have, like, three railroads running from, you know, point A to point B, and, uh, in order to compete, the operators were basically running them at a loss.

第二,它们变得竞争极其激烈,于是从 A 点到 B 点会有三条铁路在跑,为了竞争,运营商基本上是在亏本运营。

So if you spend, you know, $100 million, like, building a railroad, and then you have to run it at a loss, like, obviously it's not going to end well.

所以如果你花了 $100 million 修一条铁路,然后还得亏本运营,这显然不会有好结果。

Um, and then you had, like, consolidation, and that started to fix things.

后来出现了整合,情况才开始好转。

But I guess the argument that that won't happen here is that there are really only, like, three hyperscalers that matter — you could say there's a fourth in there as well — and, um, like, given the sort of moat they've developed around, um, the economies of scale with the data centers, and all the capex, but also, like, the moving up the stack, and all the advantages of having the software on top of just the pure hardware, um, put them in a good competitive position.

但我想,认为这次不会重演的论据是:真正重要的 hyperscaler 其实只有三家——你也可以说还有第四家——考虑到它们围绕数据中心的规模经济、所有这些 CapEx 建立起来的护城河,还有向上层堆栈延伸、在纯硬件之上叠加软件带来的种种优势,它们处在一个很好的竞争位置。

And now that it's been consolidated into a few players, like, they'll have some level of pricing power.

而且现在已经整合到少数几家玩家手里,它们会拥有一定程度的定价权。

But I'm curious if that's a scenario you've thought through, and, uh, whether there are other reasons why that's not going to play out.

但我很好奇你有没有推演过这种情景,以及有没有其他理由说明这一幕不会上演。

So, um, I think this is really a multi-year investment cycle that we are actually going through.

我认为我们正在经历的,其实是一个多年期的投资周期。

So I'm a financial analyst — I also tend to think about, oh, what is the unit economics, and how much ROI can we generate in 12 months or 18 months.

我是金融分析师——我也习惯去想,单位经济模型是什么样的,12 个月或 18 个月内能产生多少 ROI。

But I think that's probably not the right framework to think about this.

但我认为那可能不是思考这个问题的正确框架。

Um, so to your point, every time I receive a package from Amazon these days, I try to remind myself that that's only made possible with years of investment in, say, freeway, railway, and last-mile deliveries.

呼应你的观点:如今每次收到 Amazon 的包裹,我都会提醒自己,这背后是多年来对高速公路、铁路和最后一公里配送的投资。

And the same thing goes for YouTube videos — that's only made possible with years of investment in high-speed internet, or 5G.

YouTube 视频也是同理——那是多年来对高速互联网或 5G 投资的结果。

Uh, so in some way, I think about all the investment in AI today similar to the investment in the information superhighway — I think that happened 25 to 30 years ago in America.

所以在某种意义上,我把今天对 AI 的所有投资,类比成当年对信息高速公路的投资——那大概发生在 25 到 30 年前的美国。

And the total GDP back 25 years ago was $10 trillion, and I think the country spent five to eight percent of the total GDP into the information superhighway.

25 年前美国的 GDP 总量是 $10 trillion,我想这个国家把 GDP 总量的 5% 到 8% 投进了信息高速公路。

And versus today, total GDP is $25 trillion, and even 1% of that is $250 billion.

而今天,GDP 总量是 $25 trillion,哪怕只拿出 1%,也是 $250 billion。

So I'm just saying, in the grand scheme of things, I can sort of make sense of the hundreds of billions of dollars of GPU revenue that we're seeing from Nvidia.

我想说的是,放在大格局里看,我大致能理解我们在 Nvidia 那里看到的数千亿美元的 GPU 收入。

But in terms of the return, I think too early to tell.

但就回报而言,我认为现在下结论还太早。

But, um, for example, like, customer service is, say, a $500 billion industry, and even if, say, AI can save 10%, or even 20%, of the cost, that's probably enough to justify all the capex that we already put into AI.

举个例子,客户服务是一个大约 $500 billion 的行业,哪怕 AI 只能节省 10%、甚至 20% 的成本,大概也足以证明我们已经投入 AI 的全部 CapEx 是合理的。

Comparing maybe the investment cycle today versus, call it, 20, 30 years ago, um, I want to say I think the environment we are in today is a lot healthier.

拿今天的投资周期和 20、30 年前的相比,我想说,今天我们所处的环境要健康得多。

So in the sense that, number one, in terms of the parties who are actually funding the capex — back 25, 30 years ago, I think those were the telcos who were the most aggressive, and those companies tend to be the most highly levered, and just a lot of those companies simply did not really make it in a higher interest rate environment.

首先,从真正出资承担 CapEx 的主体看——25、30 年前,最激进的是电信公司,而那些公司往往是杠杆最高的,其中很多在利率走高的环境里就是没能撑下来。

And versus today, we have the highest-quality technology companies who are funding the capex, and they can afford to be more patient with their capital invested.

而今天,出资承担 CapEx 的是最优质的科技公司,它们有本钱对投入的资本更有耐心。

And, um, the other thing is, 20, 30 years ago, you actually had companies called maybe Pets.com or eToys.com — those were valued at, like, 500 times revenue multiples.

另外,20、30 年前,市场上确实有 Pets.com 或 eToys.com 这样的公司——估值达到 500 倍收入的水平。

And we definitely do not have that today, or at least not yet, in the public market.

而今天的二级市场里绝对没有这种情况,至少目前还没有。

So that relatively healthier capital market, I think, is also very important to make sure that we don't really run into a bubble.

所以这个相对更健康的资本市场,我认为也非常重要,能确保我们不会真的走进泡沫。

That's a great point, and I totally agree about thinking about this on a longer investment cycle.

这点说得很好,我完全同意要放在更长的投资周期里来看这件事。

I think people are probably overestimating — some people are probably overestimating the impact of AI on, like, a one-to-two-year time horizon, but massively underestimating the impact on, like, a 5-to-10-year time horizon.

我觉得有些人可能高估了 AI 在一到两年时间维度上的影响,但严重低估了它在 5 到 10 年时间维度上的影响。

Like, one way I like to dimension this: depending on how you cut it, the global enterprise software market today is roughly a trillion dollars, so companies spend about a trillion a year on enterprise software.

我喜欢用一个维度来衡量这件事:取决于你怎么切分,今天全球企业软件市场大约是 $1 trillion,也就是说企业每年在企业软件上的支出大约是 $1 trillion。

But that same cohort of companies pays humans $32 trillion a year to do knowledge work, right — to write software code, to review contracts, to design circuit boards, to do things that AI can increasingly do not only a lot cheaper and a lot faster, but in many cases a lot better than humans — at least better than, you know, the bottom 90% of humans.

但同一批公司每年要付给人类 $32 trillion 去做知识工作——写软件代码、审合同、设计电路板,这些事 AI 越来越能做到不仅便宜得多、快得多,而且在很多情况下比人做得更好——至少比后 90% 的人做得好。

And so if you get to a point where AI can do a task, you know, five times better, and 10 times cheaper and faster, than a human, like, of course you're going to use AI over the human.

所以如果到了某个点,AI 做一项任务能比人好 5 倍、便宜 10 倍、快 10 倍,你当然会选 AI 而不是人。

And so it seems logical that, as long as, you know, AI continues to progress — even if it progresses at a slower rate than it has historically — but as long as it continues to progress at some level, like, in the next 5 to 10 years, AI is really going to start aggressively eating into that $32 trillion in labor spend.

所以逻辑上讲,只要 AI 继续进步——即便进步速度比历史上慢——只要它保持某种程度的进步,未来 5 到 10 年,AI 就会开始大举蚕食那 $32 trillion 的人力支出。

Um, which means it's probably, in my opinion at least, going to create more net-new enterprise value than the internet, or mobile, or cloud did in their super cycles.

这意味着,至少在我看来,它创造的净新增企业价值可能会超过互联网、移动或云在各自超级周期里创造的价值。

So I totally agree — given the opportunity set in front of us, like, $200 billion in capex is not as significant as I think people are making it out to be.

所以我完全同意——鉴于摆在我们面前的机会,$200 billion 的 CapEx 并没有人们渲染得那么夸张。

Yeah, exactly.

对,正是如此。

I think people do the same as they look at the total TAM, and at the same time they look at the scaling law — and then the scaling continues, so the improvement continues.

我觉得人们的思路也是一样:一边看总的 TAM,一边看 Scaling Law——scaling 在继续,所以改进也在继续。

So they're like, oh, even though this is, like, $10 billion, or even like $200 billion, total capex, we just have to continue investing into it, because we have to capture the value.

所以他们会想,即便总 CapEx 是 $10 billion,甚至 $200 billion,我们也必须继续往里投,因为必须把这块价值拿下。

Yeah, and I think, on the railroad point that I made earlier — yeah, I think the fundamental difference is, like, the economies-of-scale component.

对,回到我之前讲的铁路那个点——我认为根本差别在于规模经济这个因素。

Which is both with the foundation model providers, where you'll have a few that are able to, you know, invest billions of dollars in a training cycle and then generate, you know, tens of billions of dollars in profit on top of that, um, where it'll be impossible for new entrants to compete.

这一点在基础模型提供商身上成立:会有少数几家能够在一个训练周期里投入数十亿美元,然后在此之上赚到数百亿美元的利润,新进入者将无法与之竞争。

But it's also true on the hyperscaler side, I think, as well, right — given the amount of investment they've made in their infrastructure, building out data centers, all the software on top of the hardware stack, that creates more value on their platform.

但我认为在 hyperscaler 这边同样成立——考虑到它们在基础设施上的投入规模,建设数据中心,以及硬件栈之上的全部软件,这些都在它们的平台上创造出更多价值。

So it seems like, given that there are, you know, two, three, four players in both of those areas that really matter, it's unlikely that it'll be totally commoditized.

所以看起来,既然这两个领域里真正重要的玩家各只有两家、三家、四家,它就不太可能被完全商品化。

Right, for sure, for sure, for sure.

对,确实,确实。

I think also it's also up to the model companies — do they just want to be an infrastructure company, or do they also want to be the product company?

我觉得这也取决于模型公司自己——它们是只想做一家基础设施公司,还是也想做产品公司?

And I think increasingly it's very clear that they themselves want to be the product, or the application, themselves.

而且我认为越来越清楚的是,它们自己想成为产品本身,或者说应用本身。

Uh, so it's actually very similar to the operating system analogy that I think Karpathy first came out with.

所以这其实和 Karpathy 最早提出的操作系统类比非常相似。

So he was saying, oh, you know, the model companies, they are just like the new operating systems.

他当时说,模型公司就像新的操作系统。

And if you think about it, each new operating system actually comes with some default apps — like, in the case of Microsoft, for example, like, the Windows comes with the browser as a default app.

仔细想想,每个新操作系统其实都自带一些默认应用——比如 Microsoft 的例子,Windows 就自带浏览器作为默认应用。

It doesn't mean that you cannot use the browser that you prefer to use, but I think it does mean, if you're a browser startup, it's probably going to be relatively more difficult for you to be successful, given the bundling nature of Windows' go-to-market strategy when it comes to their default apps.

这不意味着你不能用自己偏好的浏览器,但我认为这确实意味着,如果你是一家浏览器创业公司,考虑到 Windows 在默认应用上捆绑式的市场推广策略,你要成功大概会相对更难。

Absolutely, yeah.

完全同意。

And now that we're kind of shifting here anyways to the application layer — uh, most people's experience with AI today is, like, ChatGPT, right, where you're interacting with a model in sort of an open-ended chat conversation.

既然我们正好要转到应用层——今天大多数人对 AI 的体验就是 ChatGPT,也就是在开放式的聊天对话里和一个模型交互。

Um, and it seems like what comes very shortly after that, which is already here, is multimodal, where now the interaction can be through audio, through video, etc. — like, we saw this in, you know, Google I/O's demo, and we saw this in the OpenAI demo recently.

而紧随其后、其实已经到来的,是多模态,现在交互可以通过音频、视频等方式进行——我们在 Google I/O 的演示里看到了这一点,最近也在 OpenAI 的演示里看到了。

Um, and after that, we're going to start to see AI agents.

再之后,我们会开始看到 AI agent。

Curious if you could walk through, like, how are you thinking about the agent opportunity — like, what is an AI agent, and will that ultimately be owned by a small number of model companies, as we just suggested?

想请你讲讲,你是怎么看 agent 这个机会的——AI agent 到底是什么,它最终会不会像我们刚才说的那样,被少数几家模型公司拿下?

So, like, I think everyone is trying to make the agent work.

我认为所有人都在设法让 agent 真正可用。

Uh, and if you really ask me, I think the most likely winners are still going to be the large language model companies.

但如果你真要问我,我认为最有可能的赢家仍然会是大语言模型公司。

Because, like, what exactly is an agent?

因为,agent 到底是什么?

It's the coding plus the reasoning capability that give you some kind of agentic behavior.

是代码能力加推理能力,共同带来某种 agentic 行为。

So behind the agents, each move is basically a bunch of code, and they need to understand your intention, they need to be able to execute.

所以在 agent 背后,每一步动作本质上是一堆代码,它们需要理解你的意图,还需要能够执行。

Uh, and who has the most capability in terms of controlling that?

那么在掌控这些方面,谁的能力最强?

It actually goes back to some of the large language model companies, and I think they are also really pushing towards that.

这其实又回到了一些大语言模型公司身上,而且我认为它们也在全力朝这个方向推进。

I think that's going to be, like, the aha moment for a lot of people.

我觉得这会成为很多人的顿悟时刻。

I think a lot of people who are, like, maybe on the boundaries of the tech ecosystem, who, like, use ChatGPT — um, like, you know, my parents, or people who aren't in the industry — like, they understand what's happening, but also their interaction is kind of limited right now.

很多处在科技生态圈边缘、用 ChatGPT 的人——比如我父母,或者不在这个行业里的人——他们理解正在发生什么,但目前他们的交互方式还比较有限。

It's like, oh, I have this thing that I can chat with, and it answers questions.

就是,哦,我有个可以聊天的东西,它会回答问题。

I think when they start to see, like, agents operating, and the agentic behavior — like, that's going to be, like, the aha moment for people, where they go, oh my god, this can do a lot more than I previously expected or understood.

我觉得当他们开始看到 agent 实际运作,看到 agentic 行为——那才会是人们的顿悟时刻,他们会想,天哪,这东西能做的远比我之前预期或理解的多得多。

Yeah, it's like, not just talking to me, but also doing things for me — which is definitely a lot more valuable than that, for sure.

对,不只是跟我说话,还能替我做事——这肯定比前者有价值得多。

And obviously, what comes after that is AGI, um, which is, you know, one of the most talked-about things on Twitter right now.

显然,再往后就是 AGI 了,这是眼下 Twitter 上讨论最多的话题之一。

I'm curious — one, how are you defining AGI, and do you buy into the whole, you know, doom thesis, that AI — AGI — is going to overtake the world and destroy the economy and destroy humanity?

我很好奇——第一,你怎么定义 AGI?第二,你买不买账那套末日论——AI,或者说 AGI,会接管世界、摧毁经济、毁灭人类?

Right, so I think everybody talks about AGI.

好,我觉得大家都在谈 AGI。

I was even asking ChatGPT — I was like, what is the definition of AGI?

我甚至去问过 ChatGPT:AGI 的定义是什么?

And I think ChatGPT versus Claude versus Perplexity are just giving me very different answers, so there is actually no definition of AGI, which is kind of interesting.

ChatGPT、Claude、Perplexity 给我的答案各不相同,所以 AGI 其实并没有一个定义,这一点挺有意思。

So I personally probably think about AGI as something like: when the model, or the AI, can do 90% of the jobs better than, call it, 90% of the people.

我个人大概把 AGI 理解为:当模型或者说 AI,能在 90% 的工作上做得比大约 90% 的人更好。

And, uh, I think how the market really thinks about AGI is, like, people really think that one day they wake up — let's say GPT-10 just finished training, and the model is so awesome, and they can unlock all the capabilities all at once, and that's the day we call AGI.

而市场对 AGI 的真实想法是:人们真的以为某天一觉醒来——比如 GPT-10 刚训练完,模型强大无比,一次性解锁所有能力,那一天我们就称之为 AGI。

But if you actually talk to people inside of OpenAI, they will tell you, oh, they actually call AGI as something that would have a slow takeoff — meaning that it's actually more progressive than all of a sudden.

但如果你真去和 OpenAI 内部的人聊,他们会告诉你,他们其实把 AGI 看作一个慢起飞(slow takeoff)的过程——意思是它更多是渐进式的,而不是突然发生。

And I think Zuckerberg also said it in one of his interviews really well — he said, oh, AGI is not just this one thing, uh, because there's no single threshold for humanity, because people can have very different skills.

Zuckerberg 在一次访谈里也说得很好——他说,AGI 不是单一的一件事,因为人类没有单一的门槛,人和人的技能可以非常不同。

So it's really like, over time, it's more incrementally that we just add different capabilities to it.

所以更像是随着时间推移,我们以增量方式往里面加入不同的能力。

Uh, so how I think about it is very simple: like, for example, for translation, we probably already hit AGI even today.

我的思考方式很简单:比如翻译,我们可能今天就已经达到 AGI 了。

Or for something like coding, as we just talked about, which is something that I can maybe see we get to AGI in one to two years.

又比如刚才谈到的编程,我大概能看到一到两年内达到 AGI。

And there is a really long tail of different other skills that can be more complex and more nuanced, and that might take five, if not 10, years for us to get there — and that's totally okay.

此外还有一条很长的长尾,是各种更复杂、更微妙的技能,可能要五年甚至十年才能达到——这完全没问题。

So this is roughly how I think about the whole AGI thing.

这大致就是我对整个 AGI 问题的看法。

I totally agree — I think the concept is kind of stupid.

我完全同意——我觉得这个概念有点蠢。

Like, we're going to have, like, AGI within domains — to your point, we already have it in some domains.

我们会在具体领域内实现 AGI——正如你说的,有些领域已经实现了。

Um, and like, you don't need one model that's able to do, you know, everything across every domain to have, like, a really meaningful impact on the economy.

而且你不需要一个在所有领域什么都能干的模型,才能对经济产生真正有意义的影响。

You just need models that are good at specific, you know, domain-specific tasks.

你只需要在特定领域任务上表现出色的模型。

Like, right, if you're good at coding, and you're good at customer support, and you're good at reviewing contracts, doing legal work — like, you're going to have an incredible impact on the economy.

比如,如果你擅长编程、擅长客服、擅长审合同、做法律工作——你就会对经济产生惊人的影响。

Right, so, like, the whole premise of AGI is, like, maybe a little bit overdone.

所以 AGI 这个前提本身,可能有点被过度渲染了。

What about the second part of the question — do you buy into the whole thesis that it's going to destroy humanity and destroy the economy?

那问题的第二部分呢——你买不买账它会毁灭人类、摧毁经济这套论调?

I actually believe that maybe in the future, you and I probably don't need to work, unless we choose to do the work.

我其实相信,也许未来你我都不需要工作了,除非我们自己选择去工作。

Uh, because there will be one day that AI can do a lot of things for humans.

因为总有一天,AI 能替人类做很多事情。

And I'm also personally very excited about the robotics opportunities, so which means the AI can not only do some of the mental work for us, but can also do the physical part of the world.

我个人也对机器人领域的机会非常兴奋,这意味着 AI 不仅能替我们做一部分脑力工作,还能承担物理世界的那部分。

Uh, so yeah, I think it's very exciting to actually think that we can potentially one day have all the freedom to actually do what we really want to do.

所以,想到有一天我们可能拥有完全的自由,去做真正想做的事,这非常令人兴奋。

I agree — I think it'll be a net positive for society, and, like, there will be bumps in the road, obviously, as there is with any disruptive cycle.

我同意——我认为这对社会是净正面的,当然路上会有颠簸,任何颠覆性周期都是如此。

But, um, I think, you know, if you look at the history of technology innovation — innovation typically led to, you know, productivity growth, GDP growth, like, higher quality of living.

但如果你回顾技术创新的历史——创新通常带来生产率增长、GDP 增长和更高的生活质量。

Like, GDP growth was flat for, like, a thousand years in England — basically flat leading up to the Industrial Revolution — then it massively accelerated, um, and quality of life across the board improved for most people.

英国的 GDP 增长有一千年基本是平的——直到工业革命前都几乎没动——之后大幅加速,大多数人的生活质量全面提升。

So it'll probably be the same thing here, um, perhaps at an even more meaningful scale.

这次可能也是一样,规模也许还要更大。

And I think we also shouldn't really underestimate how creative humans can be.

我觉得我们也不该低估人类的创造力。

So there are probably a lot of things that, today, because we are really constrained, so we don't really have the time, or the luxury if you will, to actually think about it.

今天可能有很多事情,因为我们受到很大约束,没有时间——或者说没有那份余裕——去真正思考。

But if, say, one day AI can do a lot of things for us, then I think humans can be very creative in creating new domains.

但假如有一天 AI 能替我们做很多事,我认为人类会在开创新领域上表现出极大的创造力。

Uh, and, uh, yeah, it's so interesting — if you go back to, like, the invention of the plow: um, you know, you had hunter-gatherer societies, and you look at, say, fast-forward 500 years, or even a couple hundred years, from the invention of the plow — the societies who adopted the plow were the ones who ended up dominating, right, both economically, militaristically, etc.

这很有意思——回到犁的发明:当时是狩猎采集社会,你把时间快进到犁发明后的 500 年,哪怕只是两三百年——采用犁的社会最终占据了主导地位,无论在经济上还是军事上。

And the hunter-gatherer tribes basically just got wiped off the face of the earth.

而狩猎采集部落基本上就被从地球上抹去了。

And the reason that was true is because, if you're a hunter-gatherer tribe or society, you spend almost 100% of your energy and time focused on survival, right — focused on, like, finding food to eat, finding water to drink, etc.

之所以如此,是因为如果你是一个狩猎采集部落或社会,你几乎 100% 的精力和时间都花在生存上——找吃的、找水喝等等。

But if you have an agrarian society, you can have one or two or five people in your village farming, and the rest of the people can go build tools, right — they can go explore.

但在农耕社会,村里只需要一两个或五个人种地,其余的人可以去造工具、去探索。

Exactly.

正是。

They can build weapons, they can do all these things beyond just, like, feeding themselves — and that's how you get productivity and leverage in a society, right?

他们可以造武器,可以做各种超出填饱肚子之外的事情——一个社会的生产率和杠杆就是这么来的。

Right.

对。

And it's kind of the same thing here, to your point — it's like, if we have models that can automate a lot of the mundane work we do now, or instead of having a hundred people do a mundane task, you could have one person, you know, with an AI copilot, it frees those people to do other things that will create more productivity and value in society.

这里也是类似的道理,正如你所说——如果模型能把我们现在做的大量琐碎工作自动化,或者说原本一百个人做的琐碎任务,现在一个人配一个 AI copilot 就能完成,这就把那些人解放出来去做别的事,为社会创造更多生产率和价值。

So it seems like there will be bumps in the road — like, there will be displacement in certain job categories — but net-net, on, like, a 15-, 20-year basis, like, it'll probably be a very positive thing for society.

所以路上会有颠簸——某些职业类别会出现岗位替代——但综合来看,放在 15 年、20 年的尺度上,这对社会大概率是非常正面的事。

For sure, for sure.

当然,当然。

Like, even at one point, I think 70% of the population was actually working in the farming industry.

甚至在某个时期,我想有 70% 的人口其实都在从事农业。

If you think about how that has really shifted over time, and how much productivity gain, or GDP growth, there has been in the meantime, I think that's very significant.

想想这个比例随时间发生了多大变化,以及这期间有多少生产率提升或 GDP 增长,我认为意义非常重大。

A hundred percent, agree.

百分之百同意。

Cool — so we've covered the energy bottlenecks, we've covered the position of the hyperscalers and the big model developers, uh, we've covered the capex investments that they're making, some predictions about the future with AGI.

好——我们已经聊了能源瓶颈,聊了 hyperscaler 和大型模型开发商的处境,聊了他们正在做的 CapEx 投入,以及一些关于 AGI 未来的预测。

Let's talk about, like, the next five years, and the opportunities you're excited about as an investor — where do you think there's going to be a lot of value creation that's going to be kind of net-new and investable?

我们来聊聊未来五年,以及作为投资人你感到兴奋的机会——你认为哪些地方会出现大量全新的、可投资的价值创造?

Right, um, I think there are definitely a lot of very exciting opportunities, but maybe we talk about a few more debatable ones today.

好,我认为确实有很多非常令人兴奋的机会,但今天我们不妨聊几个更有争议的。

And, uh, I think one of those actually happens to be the model companies themselves.

我认为其中一个恰好就是模型公司本身。

Uh, I've been very bullish on just investing in the model companies, because I truly think the model itself is the application, and so far there just hasn't been any boundaries that we see from the model companies.

我一直非常看好直接投资模型公司,因为我真心认为模型本身就是应用,而且到目前为止,我们还没有看到模型公司遇到任何边界。

And so I talked about the operating system analogy, which I personally really like.

我之前讲过操作系统的类比,我个人非常喜欢这个类比。

So I think what we do as investors, or how I think about it, is really to think about what will be the default apps for the new operating system, which is the model companies.

所以我认为我们作为投资人要做的,或者说我的思考方式,是去想新操作系统——也就是模型公司——的默认应用会是什么。

And I think that leads to one of the more debatable investment opportunities, which is coding, for example.

我认为这就引出了一个更有争议的投资机会,比如说编程。

Uh, if you really ask people who are working at the model companies, they will tell you coding is everything they do — it's core to the capability of the model.

如果你真去问在模型公司工作的人,他们会告诉你编程就是他们做的一切——它是模型能力的核心。

And, uh, I think Llama 2 didn't really have coding capabilities in it, because Zuckerberg was saying, oh, I don't think people will be asking coding questions inside of Instagram or WhatsApp — and he's 100% right.

我认为 Llama 2 当初没有真正加入编程能力,因为 Zuckerberg 说,我觉得没人会在 Instagram 或 WhatsApp 里问编程问题——他说得 100% 对。

But after they included coding in the model layer, uh, I think they just realized the model became so much more rigorous, and just so much more powerful.

但在他们把编程加进模型层之后,我认为他们才发现模型变得严谨得多,也强大得多。

I think the mental framework is, like, coding can just lead to better reasoning, and with coding and reasoning combined together, you can get to some kind of agentic behavior, which is what everybody is trying to push towards.

我认为思维框架是:编程能带来更好的推理,而编程和推理结合在一起,就能得到某种 agentic 行为,这正是所有人都在努力推进的方向。

So if you really ask me today, um, I would say, if I'm the model company, I'll probably have coding as one of the, quote-unquote, default apps.

所以如果你今天真要问我,我会说,如果我是模型公司,我大概会把编程作为所谓的默认应用之一。

And I'll say the same thing goes for maybe the independent agent companies.

对于独立的 agent 公司,我想说同样的道理也适用。

Uh, and there's another one, which is, uh, what will happen to some of the vertical models — whether this is voice models, or image models, or video models.

还有一个问题是,一些垂直模型会怎样——无论是语音模型、图像模型还是视频模型。

I think that's a true debate that we have in the market, which is: is single modality good enough, or would the multimodality totally overshadow all the single-modality companies?

我认为这是市场上一个真正的争论:单一模态够不够好,还是说多模态会彻底盖过所有单一模态公司?

Um, I think in terms of looking at the startup landscape, uh, it's probably really easy and also extremely difficult today to be an AI startup.

从创业公司格局来看,今天做一家 AI 创业公司可能既非常容易,又极其困难。

It is very easy because literally everything is just one API call away, right?

说容易,是因为几乎一切都只需一次 API 调用。

But it's also extremely difficult because you don't really control what will actually happen in the future — you don't know how much better the next version of the model will become versus today's version.

但说极其困难,是因为你并不真正掌控未来会发生什么——你不知道下一版模型会比今天的版本好多少。

So you have to be pretty careful with your relationship with the model companies.

所以你必须非常谨慎地处理你和模型公司的关系。

It could be, today you are a standalone AI startup, and tomorrow you may just become a feature in a model company.

可能今天你还是一家独立的 AI 创业公司,明天你就变成了模型公司里的一个功能。

Yeah, I totally agree.

是,我完全同意。

And, uh, I think Sam Altman said this recently, which was, like, you can determine the fate of your startup — whether, uh, you know, if OpenAI, or name any other foundation model, announces that their model has a great new capability, like, do you get terrified or excited?

我记得 Sam Altman 最近说过:你可以据此判断你创业公司的命运——如果 OpenAI 或随便哪家基础模型公司宣布他们的模型有了一项强大的新能力,你是感到恐惧还是兴奋?

If you get excited, you're probably going to be okay.

如果你感到兴奋,你大概率会没事。

Totally.

完全同意。

One of the areas that I think is interesting — and it's, like, sort of this, like, undefined, uh, place in the spectrum between, like, private equity and venture — are companies that build, like, AI-native — let me rephrase that, I'm going to note to the editors to cut that out.

有一个我觉得有意思的领域——它处在 PE 和 VC 之间某个没有明确定义的位置——是那些做 AI 原生的公司——让我重新说一下,我会提醒剪辑把这段剪掉。

Um, one of the areas that I think is really interesting is probably somewhere between, like, venture and private equity.

有一个我觉得非常有意思的领域,大概处在 VC 和 PE 之间。

But I think there's an opportunity in a lot of industries, like accounting or manufacturing, to build vertically integrated, AI-native companies, to compete with the incumbents in that industry.

我认为在很多行业,比如会计或制造业,存在一个机会:建立垂直整合的 AI 原生公司,去和行业里的在位者竞争。

So for example, instead of trying to sell AI software to accounting firms — who are probably, like, you know, owned by someone in their 60s who's been doing it for 30 years, and, like, has 10 employees, and it's not going to change their ways — instead of doing that, like, why don't you just create an accounting firm that's AI-native, and, like, leverage the cutting-edge technology, and you'll get the sort of tailwinds of, as the technology improves, like, your efficiency improves.

比如说,与其把 AI 软件卖给会计师事务所——它们的老板很可能是干了 30 年、六十多岁的人,手下 10 个员工,根本不会改变做事方式——与其那样,为什么不直接创办一家 AI 原生的会计师事务所,利用最前沿的技术,这样随着技术进步,你的效率也跟着提升,你能吃到这样的顺风。

And it's not a software business — it's probably not a venture-scale business — but, like, you can probably run an accounting firm at, you know, 50, 60% EBIT margins, if you're super efficient, like, leveraging the frontier of what's possible in AI.

它不是软件生意——可能也不是 VC 规模的生意——但如果你足够高效,充分利用 AI 能力的前沿,你大概能把一家会计师事务所做到 50%、60% 的 EBIT 利润率。

I think it's true across a lot of industries, like manufacturing as well.

我认为这在很多行业都成立,比如制造业也是。

Totally agree, yeah.

完全同意。

Yeah, a lot of manufacturers are, like, these fragmented mom-and-pop shops that don't know what the cloud is, and if you can build, like, a highly integrated, AI-native manufacturing company, you can probably dominate in a given category.

很多制造商是分散的夫妻店,连云计算是什么都不知道,如果你能建立一家高度整合的 AI 原生制造公司,你大概能在某个品类里占据主导。

So I think we might see the emergence of these types of companies in the future, but I just don't know where they're going to fall in the spectrum — seems like probably too tech-oriented for a lot of private equity, and probably the upper bound is too limited for most venture.

所以我认为未来我们可能会看到这类公司涌现,但我不确定它们会落在这个光谱的哪一段——对很多 PE 来说可能太偏科技,而对大多数 VC 来说上限又太有限。

But I think for founders, it's kind of an interesting place to build.

但我认为对创始人来说,这是个挺有意思的创业方向。

Oh, definitely it is.

那是当然。

If you actually think about, in the accounting example you were talking about, it's probably easier to just build a new startup, versus for an existing accounting firm to really pivot and go through a ton of organizational changes to get to AI.

拿你刚才说的会计的例子来讲,直接建一家新创业公司,可能比让一家现有的会计师事务所真正转型、经历大量组织变革去拥抱 AI 要容易。

So yeah, in that sense, yeah, it's very interesting and very exciting to see how that will play out vertically.

所以从这个意义上说,看这件事在各个垂直领域如何展开,非常有意思也非常令人兴奋。

Yeah, agree.

是,同意。

I think we're starting to see some of these companies pop up, so we'll see where they land in, you know, three to five years.

我认为我们已经开始看到一些这类公司冒出来,三到五年后再看它们落在哪。

But, uh, that's an example where you're very excited when models get better — not terrified — because you're capturing the benefit.

而且这正是一个模型变强时你会非常兴奋——而不是恐惧——的例子,因为受益的是你。

Yeah, it's the verticals, yeah.

对,就是垂直领域。

I think the models, like, they won't be getting into a lot of the vertical industries, so you definitely have that kind of benefit.

我认为模型公司不会进入很多垂直行业,所以你确实能拿到这种好处。

Absolutely.

没错。

And if people want to follow you — you have a great Twitter account — what's your Twitter handle, and where else can they go to find you?

如果大家想关注你——你的 Twitter 账号很不错——你的 Twitter handle 是什么,还有哪里可以找到你?

My name, Freda Duan, uh, and then you can just find me on Twitter.

就是我的名字,Freda Duan,你直接在 Twitter 上就能找到我。

Awesome, and we'll link you to it.

太好了,我们会附上链接。

Awesome, yeah — thank you so much, thank you so much for joining.

太好了——非常感谢,非常感谢邀请我。

Cool, yeah, thank you.

好,谢谢你。

Thanks for tuning in to the Autopilot podcast.

感谢收听 Autopilot 播客。

If you're using AI to create or disrupt a large industry, I would love to learn more about your story — drop me a line on Twitter, or go to our website.

如果你正在用 AI 创造或颠覆一个大行业,我很想了解你的故事——在 Twitter 上给我留言,或者访问我们的网站。

If you want to support the show, the best ways are to leave a review wherever you're listening, and subscribe.

如果你想支持这个节目,最好的方式是在你收听的平台留下评价,并订阅。

The Autopilot podcast is part of Turpentine, the podcast network behind Moment of Zen, Turpentine VC, Age of Miracles, and more shows for experts, by experts, in tech.

Autopilot 播客隶属于 Turpentine,这个播客网络旗下还有 Moment of Zen、Turpentine VC、Age of Miracles 等更多科技领域由专家做给专家听的节目。