第037集 - 谁在为11万亿美元的AI建设提供资金？（资本市场）| Dan Nishball, Jordan Nanos

好的，好的，好的。

Sure, sure, sure.

所以，比如说，我们看亚马逊，记得他们发行那笔债券的时候，十年期的Z利差大概在100个基点左右。

So if I look at Amazon, right, for instance, remember when they issued that bond, their ten year Z spread would have been around 100 basis points.

所以大致上，如果你有国债，以前大概是4%到4.5%，然后加上100个基点，也就是1%。

So roughly if you have treasuries, it used to be like four, 4.5%, then you add 100 basis points or 1%.

那就意味着亚马逊债务的收益率大约在5.5%到5.75%，具体看情况。

Then that means the yield on Amazon debt is about 5.5 or 5.75 as the case may be.

而五年期的利差大概在65个基点左右。

And for five year, that spread was roughly around 65 basis points.

然后紧接着发生的是，利差扩大了，大概在夏天的时候，从发行时点扩大了大约10到20个基点，这对投资级债券来说算是比较大的波动了。

And then what sort of happened right after is it blew out probably by like in the summer, it blew up by about probably 10, 20 basis points from issue, which is kind of a lot of move for IG.

我的意思是，目前情况还算相对可控，但你肯定看到甲骨文自夏天以来已经上涨了不少。

I mean, things are relatively under control for now, but you've definitely seen Oracle has moved up quite a bit since the summer.

他们的整条收益率曲线大概上移了30到50个基点。

Their entire curve probably moved up 30, 50 basis points.

所以你可以看到债务市场已经注意到了。

So you are seeing the debt market take note.

评级机构会看穿这些事情的。

Rating agencies will see through this stuff.

给我讲讲甲骨文这次上涨105个基点的情况，然后和Core Weave比较一下，Core Weave远高于投资级，高出差不多400个基点。

Walk me through this Oracle increase of 105 basis points and then compare that to Core Weave, which is well above investment grade at almost 400 basis points higher.

所以甲骨文，他们的利差大概会在……

So Oracle, their spread is gonna be around...

比如，目前来看，十年期，我记的是大约250个基点的Z利差，相对于亚马逊。

Like, looking currently, ten year, I've got them down for about 250 z spread versus an Amazon.

我记的亚马逊十年期只有100到110个基点的利差。

I've got them only at a spread of 100, 110 for about ten years.

这张图有点关于，就像你说的，相对的……

This chart is a little bit about, as you said, the relative...

我们怎么比较信用风险？

How do we compare credit risk?

所以这展示的是不同类型的结构。

So what this is illustrating is a different type of structures.

我这里要展示的是，我们来看一下CoreWeave向Meta出售承购合同的结构。

So what I'm showing you here is let's look at the structure where CoreWeave is selling an offtake contract to Meta.

问题是，我借给Core REIT这笔钱要收多少利息？

And the question is, how much am I gonna charge to lend to Core REIT for that?

所以借钱给Core的人会说，这实际上是Meta的信用风险加上执行风险。

So someone who's lending money to Core will say, Actually, this is Meta credit risk plus execution.

所以我们做这笔交易时，Meta的曲线大约是97个基点的Z利差。

So Meta curve at the time we did this for €5, around 97 basis points Z spread.

所以实际上借给Qorvi这个项目的钱，利率大约在200个基点。

So now the actual lending to Qorvi for this project was around 200.

是2.25美元，所以是加225个基点，算下来略低于200个基点的Z利差。

It was $2.25, so far plus two twenty five, which when you do a little bit of math comes in slightly under a Z spread of 200.

所以97个基点和200个基点之间的差距，其实就是执行风险。

And so the difference between that 97 basis points and the 200 is actually the execution risk.

因为即使Meta或微软在那里承购，你仍然需要把集群搭建起来，这就是为什么我们其实应该把它改名为ClusterMAX风险。

Because even though Meta is there, offtaking or Microsoft is there, you still have to stand up the cluster, Which is why we should actually rename this to the ClusterMAX risk.

因为如果他们没有一个白金评级，事实上，总有一天，乔丹，会变成这样：这是你的SSP评级，这是你的ClusterMAX评级。

Because if they don't have a platinum, in fact, there's gonna come a day, Jordan, when it's like, this is your SSP rating and this is your ClusterMAX rating.

不是吧，兄弟。

It's not, man.

每个人都想从ClusterMAX里选股票。

Everybody wants to do, make stock picks from ClusterMAX.

现在你又想从ClusterMAX里挖别人的债务。

Now you wanna raid people's debt from ClusterMAX.

好吧。

Okay.

这只是一个数据点。

It is one data point.

别告诉伊森，他会把我们停掉的。

Don't tell Ethan he's gonna shut us down.

不。

No.

不。

No.

我们不会把新的云债务产品定价叫做ClusterMAX的，丹。

We're not we're not branding new cloud debt product pricing as ClusterMAX, Dan.

抱歉。

Sorry.

抱歉，观众朋友们。

Sorry, audience.

乔丹不提供投资建议。

Jordan does not give investment advice.

有一个……

There's a...

好的。

Okay.

你看。

Look.

要建一个高质量的集群，有很多因素需要考虑。

There there's a lot of things that go into building a good quality cluster.

我觉得，按照既定时间表建立数据中心，然后维护合同，确保合同不被取消或终止，这通常属于裸金属服务，和我们评级的托管集群是完全不同的产品。

I think execution written on standing up a given data center on a given timeline and then maintaining a contract such that there isn't a cancellation or termination of the contract is typically bare metal, and it's typically a very different sort of product that is being sold to people than the managed clusters that we're rating.

但是，我仔细想了想，我对自己说的也不是完全有把握，那就是，你看，很多打造优质裸金属服务的经验，肯定也适用于集群的质量。

But, actually, I'm thinking this through, and I'm not completely convinced in what I'm saying, which is that, look, a lot of the experience that applies for building a great bare metal service certainly applies to the quality of the cluster.

比如，能建出可靠集群的人，通常也能维护好合同并按时交付。

Like people who build reliable clusters are generally able to maintain and deliver on their contracts.

所以，好吧，有道理。

So, all right, fair enough.

我们可以把它作为衡量执行风险的一个指标。

We can use this as a rating of something towards execution risk.

嗯，乔丹，你这么说，我和贷款方一起路演过。

Well, know, Jordan, you say that, I've been on a roadshow with lenders.

我去了纽约，也去了新加坡，很多人跟我说，这真的对他们的流程帮助很大。

I did New York, I did Singapore, and a lot of them have said, Yeah, this has been incredibly helpful to their process.

我觉得这就像……

I think it's like...

你看，如果一个人能上ClusterMAX并达到白金级别，那他们通常不太会在执行上出大问题。

Look, you know if someone's made it onto ClusterMAX to go to Platinum, they're less likely to mess up on execution generally.

这就像是一个认可的标志。

It's just a stamp of approval.

就像你成绩好，不代表你就是天才，但大概率还是比一般人强，对吧？

Just because you got good grades, you got good, doesn't mean you're a genius, but probably is more likely than not, right?

我觉得这是一个信号和数据点，但托管集群和大型裸金属数据中心是完全不同的产品。

I think it's a signal and a data point, but managed clusters and large bare metal data centers are very different products.

我觉得一个人在白金名单上，并不代表他所有的数据中心都能按时交付，反过来也一样。

I think just because somebody's in platinum does not necessarily mean that all their data centers are gonna be on time, and vice versa.

一个人排名靠后，或者被评为表现不佳，也不代表他的数据中心会延期，或者未来质量会差。

Just because somebody's down the list, or maybe ranked as like underperforming, or something like that, does not mean that their data centers will be late, that they will be low quality in the future.

确实如此。

So True.

我觉得

I think

如果有的话，我会建议大家去评估某个具体的项目或某个具体的地点。

that there's, if anything, I would be asking people to assess a given project, or a given site.

我的意思是，我们见过有人把Neo Cloud投资组合中的每个站点都设立为SPV。

I mean, we've seen people structure every single site within a Neo Cloud's portfolio as an SPV.

我认为这是正确的思考方式，因为正如你所说的，在某些情况下，位于某个地点、某个大陆的站点，恰好由某个Neo Cloud拥有和运营，其执行风险水平会非常不同，即使他们下周或下个月可能去完全不同的地方做另一个项目，以完全不同的方式筹集债务，由完全不同的团队来负责。

And I think that's the right way to think about it, where there is a different, in some cases, very different level of execution risk, as you're describing it, for a site in one location, on one continent, that happens to be, you know, owned and operated by a given Neo Cloud, even though they can go to a completely different continent for a different project the next week, the next month, and raise debt in a totally different way, have a totally different team working on it.

所以，你知道，情况会有些不同。

And, you know, therefore things are a little bit different.

丹，我们继续吧。

Dan, let's move on

嗯，乔丹，这是个有趣的观点，对吧？

Well, Jordan, that's an interesting point, right?

因为我们接触到的很多NeoClusts，他们之所以没空，并不是因为忙着做裸机业务。

Because a lot of the a lot of the NeoClusts we talk to that are unavailable, they're not unavailable because they're busy doing bare metal.

就像你说的，那不是他们的重点。

To your point, it's not their focus.

但他们还是拿下了大单。

So yet they've landed large deals.

所以这只是一个数据点，有些人就是不关注这个。

So it's like, it's just one data point and some people just don't focus on it.

有些做得很好的人拿了很多单子，然后就说，嘿，不好意思，我们现在没有产品给你们测试。

Some guys who are doing great have landed lots of deals, just like, Hey, sorry guys, you know, we just don't have the product for you to test right now.

整个市场都变了。

And the whole market has changed.

我们一直在说这个，但ClusterMAX刚起步的时候，买家有五家不同的供应商提供两周的256 GPU试用，让他们能做出明智的决定再签约。

I mean, we're constantly going back to this, but when ClusterMAX started, buyers had five different providers offering them two week POCs on two fifty six GPUs so that they could make their decision in an informed manner and sign up.

现在你得在GPU交付前六个月就选好合作伙伴。

Now you got to pick a dance partner six months before the GPUs get delivered.

而且你得给这些人预付一笔钱，有时候相当于你合同金额的30%到50%。

And you gotta make a prepayment to these people that's worth 30, 50% of your contract value in some cases.

你在这一集前面也说过，实际上买家就是在直接替别人承担资金成本。

I mean, you said this earlier on this episode, effectively the buyers are just paying for people's cost of capital directly.

所以你把筹来的钱——这些钱本身就有资金成本——交给别人，还得替他们承担资金成本。

And so you're taking money that you've raised, which has a cost of capital, and then you're giving to somebody else and you have to pay for their cost of capital.

这就像一层叠一层，逼得很多人想自己搞建设。

I mean, it's, you know, it's stacking on top, which driving a lot of people to try to do self build work.

我觉得这几乎就像一种震荡。

And I think is what it's almost like this oscillation.

就像，早期你会选择租用，因为你不想把它放在自己的资产负债表上。

It's like, start out early and you do rental because you don't want it on your balance sheet.

你想快速行动。

You wanna move fast.

你想要尽可能大的集群。

You want the biggest cluster possible.

然后你开始成长，你意识到，嘿，这笔交易不划算。

And then you start to grow and you realize, Hey, this is a bad deal.

感觉我有点被坑了。

Like I'm kind of getting ripped off.

我应该自己拥有这些东西。

I should own this stuff myself.

然后你开始做这些自建项目，接着你达到了资产负债表能承受的极限，以及你能筹集到的资金上限。

And then you start doing these self build sites and then you reach this limit of what your balance sheet can handle, how much money you can raise.

然后你开始去找，前沿实验室的人说，嘿，伙计们，我需要帮助。

And then you start going, the Frontier Labs go, hey guys, I need help.

我现在基本上需要抵押我的未来。

I need to basically mortgage my future right now.

我需要你们建一些东西，而且我需要保证你们在这项工作上有回报，因为，嘿，我没有足够的资产负债表来做这件事。

And I need you to build some stuff and I need to guarantee you some returns on this work because like, hey, I don't have the balance sheet to do it.

我需要把你们的资产负债表置于风险之中。

And I need to put your balance sheet to risk.

所以这是一个有趣的曲线，前沿实验室现在做交易的方式，就是想把所有东西都从资产负债表上挪走。

So it's an interesting curve where the Frontier Labs, the way they're doing deals right now and trying to get everything off their balance sheet.

甚至超大规模企业也在试图把东西从资产负债表上挪走。

And even the hyperscalers are trying to get stuff off their balance sheet.

这看起来有点像早期阶段的初创公司。

It kind of looks like an early stage startup.

他们想要现在能拿到的最多的东西。

They want the most that they can get right now.

我觉得你已经触及了原因，对吧？

And I think you've touched on the why, right?

也就是我们是怎么走到这一步的。

Which is where we kind of got here.

比如，为什么会有这么多支持？

Like, why is there all this backstop?

你描述了这些问题，对吧？

And you described the problems, right?

你得提前六个月预订。

You gotta book six months in advance.

根本拿不到你想要的算力。

Can't get the compute you want.

还得先付50%的定金。

You gotta put 50% down.

这就是他们搞这个兜底方案的原因。

And that's kinda why they've done the backstop.

让我跟你分享一下。

I mean, let me share.

我做了张幻灯片，讲的就是这些原因。

I did up this slide that describes some of the why.

跟你说的完全一样。

And it's what you said.

扩大算力供应，就是你刚才说的那个。

Broaden compute availability is what you just talked about.

现在，支持融资市场是另一回事，因为我们刚才谈到，如果没有那个投资级信用支持，什么都做不了。

Now, supporting the financing market is another thing because we just talked about how, unless you have that IG backstop, you can't do anything.

所以对新云服务商来说非常困难，从商业模式上看，没有动力去说，嘿，我要把我的订单簿填满各种推理服务的一半。

So it's very hard for a Neo Cloud, there's no incentive as a business model to say, Hey, I'm gonna fill my book with half of various inferences of service.

因为你的贷款方会说，嘿，这不是投资级信用。

Because your lenders go, Hey, it was not IG credit.

我对这个业务不够了解。

I don't understand this business well enough.

这又回到了我之前那张关于障碍的幻灯片。

Which which which kind of goes back to this earlier slide that I had on the the obstacles.

他们还在学习曲线上。

They're still on the learning curve.

他们不够了解这个业务，除了投资级信用之外，无法承销其他任何东西。

They don't understand the business well enough to underwrite anything other than IG.

而且他们也没有相应的工具。

And then they don't have the tools.

这很像乔丹和我

This is like a lot of what Jordan and

正在做的事情。

I are working on.

你知道，乔丹负责评估运营质量的工具，而我负责评估市场动态。

You know, Jordan on the tools to assess the operational quality and and myself on assessing, like, what's happening in the market.

这就是我们推出GPU指数的原因。

That's why we have the GPU index.

我们努力创建贷方需要的工具来监控风险。

That's why we're trying to create tools that lenders need to monitor risk.

你提到了残值。

And you mentioned residual value.

这是我们在路演中提出的一个重大需求，因为银行考虑残值，他们贷款给资产。

This is one of the huge asks we had in our roadshow because banks think in terms of residual value, they're lending its assets.

他们习惯于贷款给房产、汽车，他们必须知道残值，而GPU的残值非常难确定。

They're used to lending its property, cars, and they have to know residual value, which is very, very hard with GPUs.

这不像直线折旧那么简单。

It's not like you straight line depreciate.

我觉得残值取决于你能通过代币赚多少，或者你能以合理的价格租出去多少。

I think the residual value is what you can earn on tokens or what you can reasonably rent it at.

现在市场非常扭曲，因为在正常市场中，供需之间会有竞争，人们根据需求生产足够的GPU，而且总能稍微多生产一点，这样价格就是真实的，结算也是真实的。

And it's so distorted right now because I think in a normal market where you have this contention between supply and demand, where people produce enough GPUs based on the demand for them, and they can always go a little bit higher, then you have real pricing, and you have real like settlement.

但现在的问题是，没人知道H100的真实理论折旧曲线是什么样的，尤其是当GB300上线，明年Verirubin也上线并且性能更强的时候。

But what's happening right now is nobody has an understanding of what the true theoretical H100 depreciation curve looks like as GB 300 comes online, as Verirubin comes online next year, and it outperforms that.

然后你会问，我怎么才能让H100盈利呢？

And then you say, how can I run an H100 profitably?

其实我们已经有Inference X的模型，显示你确实做不到。

Well, we already have models with Inference X that show that you really can't.

如果GB300的性能比你的Hoppers强七倍，那谁还会去买一个边缘的Hopper而不是边缘的GB300呢？

If a GB 300 is outperforming your hoppers by like seven times, why would anybody wanna buy a marginal hopper over a marginal GB 300?

你绝对不会想要的。

You'd never want it.

你肯定想要最新最好的。

You'd want the latest and greatest.

但这不是大家问关于Hopper的问题。

But this isn't the question people are asking about Hopper.

相反，问题是，我根本买不到B300 GPU，因为都卖光了，或者我得等六个月。

Instead, the question is, I have no option to purchase a B 300 GPU because it's all sold out or I have to wait six months.

我现在就想做研究。

I want to do research right now.

我有什么选择可以买？

What option do I have to buy?

答案就是续签我的H100合同，用四年前的GPU，价格却是我之前签的两倍。

And the answer is renew my H100 contract with four year old GPUs for double the price of what I signed it for previously.

这就在屏幕上。

And this is on screen.

Dan有我们知道的H100一年期交易的价格表。

Dan's got our pricing chart for the one year term length, H100 deals that we know about.

我是说，现在它们的价格已经接近2023年上半年签约时的价格了。

I mean, they're approaching the prices that they were signed in the first half of twenty twenty three today.

比如H100 GPU一年期合约每小时3美元。

Like $3 an hour for H100 GPUs on a one year term.

而且你看，如果你相信AGI，说实话，我和这些初创公司聊得很多，基本上我会给他们这样的建议：如果你相信你的创业公司会成功，你就得想象有人会做出OpenAI和Anthropic做不到的事情，而那个人就是你。

And look, if you believe in AG up, frankly, I talk to all of these startups a lot, and I basically give them the coaching of, if you believe in your startup being successful, you have to imagine that somebody is gonna do something OpenAI and Anthropic can't, and it's gonna be you.

所以你必须，轻率地说，你必须相信AGI。

And so you have to, flippantly, you have to believe in AGI here.

如果你不相信……

And if you don't...

乔丹，但是

Jordan, but

我有个问题，因为你说过做推理的人不会租H100。

I had a question, because you said that no one doing inference would ever rent an H100.

那么，价格还会上涨吗？

So, is the price going up then?

嗯，因为我觉得需求太大了，所以价格会涨，主要是因为很多人只有这一个选择。

Well, because I think there's so much demand for it that the price goes up strictly because of it's the only option for a lot of people to do.

其次，有些研究人员更喜欢H100，因为它比较熟悉。

Second of all, there are some researchers who prefer H100 because it's familiar.

他们更看重每天能完成多少任务，而不是每秒能处理多少token。

They are more productive by the metric of jobs per day, rather than tokens per second.

就像，我不想花时间去调试。

It's like, I don't need to spend time debugging.

我不想把代码从BF16改成FP8或者FP4。

I don't need to port stuff from BF16 to FP8 or FP4.

我在做训练，你知道的，我需要数值稳定性。

I'm doing training and, you know, I need numerical stability.

我不在乎那些用于推理的低精度数据类型。

I don't care about these low precision data types that are used for inference.

我在做一些新模型，同样，我关心的是数值稳定性。

I'm doing some new model where, again, I care about numerical stability.

我不想冒这个险。

I don't want this risk.

所以，人们想用Hoppers、H100和H200做研究是有原因的。

So, there's reasons why people want to do research with Hoppers, H100 and H200.

但我认为今天价格上涨的最大原因就是GB300卖光了。

But I think the biggest reason why the prices are increasing today is just because GB300 sold out.

所以人们就转向B200、B300，对吧？

So people go to the B200, B300, right?

那些也卖光了。

Those are sold out.

这

It

就一层层往下传。

cascades down.

没错。

Exactly.

我的意思是，如果你的选择是等六个月什么都没有，或者现在就用H100开始，那么一个刚融到钱的初创公司很容易做决定。

And I mean, if your answer is wait six months and have nothing or get started with H100s today, a startup that just raised money has a pretty easy decision.

他们会选择现在就能拿到的H100。

They go for the H100s they can get access to right now.

这太有道理了。

That makes so much sense.

这也是为什么我们做这个指数时采用调查的方式，因为每个数据点背后都有故事，对吧？

And that's kinda why when we do this index, do it survey based, because there's a story behind every data point, right?

这不仅仅是随便抓取一堆数据。

This is not just scraping a bunch.

背后有故事，这些故事很重要。

Is a story and those matter.

乔丹，我实际上做了分解。

What I did, Jordan, I actually decomposed.

哦，你想去吗？

Oh, do you wanna go?

或者我可以聊聊这个。

Or I can talk about this.

对，抱歉。

Yeah, sorry.

就最后一件事。

Just one last thing.

如果你回到之前那个，我们会有一种更学术、更理论的方法来预测H100价格曲线的贬值。

If you go back to the previous one, we would have a more academic theoretical way to forecast the depreciation in the H100 pricing curve.

如果你看这张图，去年九月或十月左右，价格就一直往下跌，跌破1.8美元，跌破1.4美元。

If you look at this chart and sometime around September or October of last year, it just kept going down below a dollar 80, below a dollar 40.

我记得那时候有些交易签的是H100每小时1.3美元。

I remember some deals being signed at that time for like 1.30 for an H100.

有人出价那么低。

People were offering stuff that low.

对，那就是最低点了。

Yeah, that was the bottom.

不过那些供应商可不怎么样。

But those were not great providers though.

你是说卖家？

For sellers?

嗯，不是。

Well, no.

我猜那些以1.30美元卖的人，供应商应该不怎么样，对吧？

I'm guessing the guys who were selling at 1.30 were not great providers, right?

有道理。

Fair enough.

对，对，没错。

Yeah, yeah, exactly.

但关键是，我知道当时有些卖家在考虑1.30美元的价格，心里盘算着：我是签一年期1.30美元，还是签两年期1.25美元之类的？

But the point is, I know some sellers at that time who were evaluating 1.30 and going, Should I do a one year at $1.30 or should I do a two year at $1.25 or something like that?

天哪，那些人现在后悔死了，当初没签两年期，因为现在续约，有些人直接签到了2.50美元，比去年这个时候贵太多了。

And man, those guys are kicking themselves for not doing a two year then, because they just renewed literally some of them at $2.50, like way, way more than they were paying this time last year.

所以总之，关键是如果我们看到价格曲线像汽车或其他资产那样贬值，我们就能真正理解H100的残值。

And so anyway, the point is we would have a real understanding for H100 residual value if we had seen the pricing curve depreciate like some other asset, like a car or something.

但如果趋势继续下去，明年四年前的H100售价可能会比现在更高。

But four year old H100s going into next year are gonna be selling for likely higher prices than they are today if the trend continues.

我觉得会这样，因为我预计需求会继续增长。

And I mean, I expect it to because I expect demand to continue to rise.

我再举两个例子，这些不是基于AGI未来的假设。

And let me make two more cases for this that are not based on like an AGI future.

你不需要相信那个。

You don't need to believe in that.

你需要相信的是，由于政治姿态、暂停令和执行风险，数据中心会出现延误或取消，就像你描述的那样，我们经常看到这种情况。

You need to believe in data center delays or cancellations due to political posturing and moratoriums and execution risk, as you were describing, which we see a lot of.

你还得相信消费者对Anthropic和OpenAI的API端点的需求会增加，我认为即使你不相信模型在变好，你也可以相信这一点。

And you also have to believe in more demand from consumers on Anthropic and OpenAI's API endpoints, which I think you can believe in, even if you don't believe the models are getting better.

你只需要看看像Instinct的增长或者Meta的Muse，想象一下一个更多使用AI驱动产品的世界，即使模型不再进步。

You just need to look at things like instinct growing or muse from meta and just imagine a world where more people use AI driven products, even if the models stop getting better.

所以，数据中心可能建得更慢，政客们可能会插手阻止，这会限制供应，而现有模型的需求增加，模型也在变好，这些都是价格随时间上涨的因素。

So the combination of like data centers might take longer to build, politicians may get in the way and stop them, that will constrain supply, more demand for the existing models and the models get better are all a recipe for increased prices over time.

这也是我一直在跟很多人讲的观点。

And that's generally the case that I've been making to a lot of people.

而且问题是，对吧？

And the thing is, right?

如果你用科学的方法来看，我们从2024年开始做这个预测，但我们的预测实际上是向下倾斜的。

If you did the scientific version, we started this in 2024, but we actually have a forecast, and that actually looked like downward sloping.

它开始偏离了。

It actually started detaching.

我们在大约一年半的时间里，误差都在2%以内。

We were like within 2% for a good year and a half, a good eighteen months.

但从九月份开始，它开始脱节了，因为我们计算理论折旧曲线的方式是基于新GPU的相对计算能力。

But then it started dislocating for September because the way we calculate the theoretical depreciation curve is the relative compute power of new GPUs.

比如，GB3的推理速度提高了七倍，FLOPS提高了三倍。

So for instance, a GB3 added had seven times more inference speed and three times more FLOPS.

所以基本上，那些新增的算力会把旧算力的价格压下来。

So basically, those new capacity would push down the price of your old capacity.

理论上，应该是这样。

In theory, that's what should happen.

但我们一直说的是，我们了解供给，也追踪价格，但实时追踪需求很难。

But what we always said is, we understand supply, we track price, but it is hard to track in real time demand.

所以我们反其道而行之，我们观察价格，然后从价格推导出需求。

So what we do is we actually flip it on its head and we say, Let's observe price and let it derive demand.

所以在这张幻灯片上，我就是这么做的。

And so in this slide, this is exactly what I did.

我做的就是，如果你有一个供需平衡的中性状态，那么你会看到价值在科学地贬值，或者租金价格随着新增算力而下降。

What I did is I said, If you have a balanced neutral supply demand, then what you see is it depreciating in value or rental price going down in a scientific way, in line with new capacity.

但实际情况是，到了十月份，这个规律被打破了。

But what happened is in October, that dislocated.

而这意味着，如果你的需求高于供给，那么价格就会超出预期地上涨。

And what that implies, because if your demand is higher than supply, then your price surprised the upside.

所以我就反过来想，嘿，这从价格上能看出供需平衡是什么情况呢？

And so what I did is I reversed it and I said, Hey, what does this imply for demand supply balance based on price?

然后我们看到，在2025年底，这个情况彻底失控了，而且再也没有恢复过来。

And what we saw is it blew out completely in late twenty twenty five, and it's never looked back.

Jordan，这是不是跟你看到的情况有点吻合？

Does this kinda tie in with what you've seen, Jordan?

我觉得应该是吧。

I assume so.

可能不是那些理论上的东西，但我们在市场上应该关注什么呢？

Maybe not the theoretical stuff, but what should we serve in the market?

嗯，去年差不多这个时候发生了几件事。

Well, a couple of things happened around that time last year.

我觉得基本上模型不再变大了。

I think basically models stopped getting bigger.

中国模型花了一段时间才赶上，但直到去年这个时候左右，当推理模型开始起飞时，大家发现可以用算力换取参数数量，并通过后训练中的强化学习创新来继续提升模型质量。

It took a while for the Chinese models to catch up, but up until around this time last year, when reasoning models were taking off, people kind of discovered we can trade flops for parameter count and continue to improve the quality of the model with a lot of innovations around RL in post training.

他们一直在利用这一点，而我们可能已经做出了更大的模型。

And they've just continued to exploit that where we may have gotten bigger models.

我觉得我们确实做出了更大的模型，比如Astra、Fable或者Mythos那一代，但模型并没有大到能超过硬件改进的速度，尤其是GB300和VL72上线之后，内存、内存带宽和FP4性能都大幅提升，足以跟上模型规模的扩大。

I think we likely did get bigger models, you know, with the Astra and the Fable generation or the Mythos generation, but we didn't get significantly bigger models that outpaced the improvement in hardware that came when the GB 300 and VL 72 came online, which is to say significantly more memory, memory bandwidth and FP4 performance were able to keep up with the increased size of the models.

所以这意味着这些公司可以在API上继续提高每token的价格，而不必大幅增加服务模型的成本。

And so, what that meant is that these companies could continue increasing price per token on their APIs without necessarily increasing their costs to serve the model significantly.

结果就是，他们的利润高多了。

And, you know, that the result is just they're way more profitable.

简直是高得离谱。

Like way, way more profitable.

而且你根本不需要了解Anthropic或OpenAI的模型就能得出这个结论。

And you don't even need to know anything about the Anthropic or OpenAI models to make this case.

你只需要看看中国的模型，看看去年H100、H200甚至GB200上线时DeepSeek v3的表现。

You just need to look at the Chinese models and you need to look at what the performance of a DeepSeek v3 was on like an H100, H200, or even a GB200 when, you know, those were being brought online last year.

然后你再对比一下DeepSeek v4、4.1 Pro、4.1 Flash、GLM 5.3或者Kimi K3，看看它们在稀疏注意力上的各种创新。

And then you need to compare it to DeepSeek v4 or 4.1 pro or 4.1 flash or GLM 5.3 or Kimi K three, where you just look at all the innovations on sparse attention.

你看参数数量增长了一些，有些情况下大概翻了两倍或三倍。

You look at the parameter count growing a bit, like around two times or three times in some cases.

但是，你知道，算力并没有大幅增加，而性能却提升了很多。

But, you know, flops not increasing significantly and performance increasing so much.

所以他们能比之前的令牌收取更高的价格，因为这些令牌更有价值，但服务成本更低。

And so they can command a pricing premium over the previous tokens because the tokens are more valuable, but it's cheaper to serve.

用这些模型，每块GPU的吞吐量更高。

You get more throughput per GPU with these.

所以现在在GV 300和VL 72上，服务中国模型的人毛利率能达到80%。

And so people serving Chinese models are enjoying gross margins of 80% right now on a GV 300 and VL 72.

所以很明显，对这种计算资源、这种资产的需求会更大。

And so, it's just like, it's like, look, obvious that there's going to be more demand for this compute, this asset.

现在，让我困惑的是，在某些情况下，也许你有答案，就是为什么没有人预测到这一点。

Now, the question or what gets confusing to me here in some cases, and maybe you have an answer for this, is how nobody forecast this.

比如，我们已经尽力了。

Like, we did our best.

我们明白了。

We understood.

而且我觉得，即使是那些非常深入了解这些模型运作方式的前沿实验室，也没有预料到自己会这么赚钱，能服务这么多用户，还能增长得这么快。

And I think even the Frontier Labs who really deeply understand how these models work, did not expect themselves to be this profitable and to be able to serve this many users and to grow this fast.

所以，连他们自己都措手不及，不得不做一些疯狂的事情，签下高溢价的大合同，就为了第二天能拿到GPU来服务用户，以及进行更大的研究项目。

And so, even they were caught by surprise and had to do crazy things and sign big contracts at very high premiums, just to get access to GPUs tomorrow to serve their users and to go for the bigger research jobs.

我不认为这是AI的快速起飞，但看起来很多实验室的收入和盈利能力确实在快速起飞。

I don't view this as like an AI fast takeoff, but it seems to be like a revenue and profitability fast takeoff for a lot of these labs.

对，对，我觉得在我们的讨论中，我们确实触及了未来的争论焦点，那就是意愿和能力，而我认为能力就是盈利能力。

Yeah, yeah, and I think, in our discussion, we've really hit on what is the debate going forward, it's the willingness and the ability, and I think the ability is the profitability.

很多刚接触这个领域的人没有意识到这有多赚钱。

A lot of people who are new to this don't realize just how profitable it is.

而这正是决定我们能否克服这个障碍，也就是资金瓶颈的关键因素之一。

And that's one of the things that will determine whether we overcome this obstacle, this kind of funding wall.

嗯，我觉得让人困惑的还有，这个蛋糕的哪一层是赚钱的，也就是说，对。

Well, I think it's also confusing which layers of the cake are profitable, which is to say Right.

看看这个。

Look at that.

英伟达显然在很长一段时间里都保持了很高的利润率。

Nvidia clearly had durable profit margins for a long, long time.

不过，你的那个分层蛋糕图解释得特别好。

But your, yeah, your layer cake chart is great to explain.

比如说，新云服务的盈利能力明显提升了，因为现在他们能把GPU租给别人的价格涨了，但GPU本身没什么变化，运营成本也没变，对吧？

Like, clearly there has been an increase in the profitability of the Neo Clouds, because the amount that they can rent the GPUs for to people has increased recently, when the GPU has not, there's nothing changed on the operational costs or anything, right?

他们用更高效的模型或者更高的价格卖token，所以他们的利润在增加。

The people are selling tokens with more efficient models or higher prices, and so therefore, their profitability is increasing.

然后从我个人的经验来看，我觉得很多人也有同样的感受，就是使用这些AI应用能获得的价值或效用也在随着质量的提高而不断增加。

And then my personal experience, and I think a lot of other people's experiences, the value or utility that you're able to derive from using these AI applications is also increasing over time as they get higher quality.

所以我不觉得这是在拆东墙补西墙。

And so I don't look at this and say, we're taking from one to give to another.

看起来每一层都在某种程度上提高了盈利能力，除了服务器OEM厂商，他们还在打价格战，利润率只有9%。

It seems like every layer of the stack in some ways is increasing profitability, except for the server OEMs where it's still a pricing war and they get 9% margins.

不不不。

No, no, no.

每个人还是能分到一点的，但是……

There's still a bit for everyone, but

这是种棘手的商品。

It's tricky commodity.

他们就是铁皮盒子，加上插电的人。

They're tinned box and the people to plug it in.

嗯，我可以深入讲这个，但我想，为了让大家思路清晰，我觉得，到目前为止我们聊的所有内容，乔丹和我聊的，都可以归为两大类：意愿和能力。

Well, I can get into that, but I wanna, just so everyone stays organized, I think, again, everything we've talked, Jordan and I have talked to so far, can fit into two buckets, the willingness and the ability.

我们刚才聊的就是意愿。

What we're just talking about now is the willingness.

回报是什么？

What is the return?

实验室的回报能保持高位吗？

Will the labs remain high?

利润率有多高？

Like how high are the margins?

最终用户能不能……

Can end users...

之前我们聊的是能力方面。

And earlier on, we were talking about the ability.

我们谈到了资产负债表即服务。

We talked about the Balance Sheet as a Service.

我们聊了英伟达的兜底方案，那是为了解决能力问题，以及谁来资助他们。

We talked about Nvidia Backstop, which is trying to address the ability and who's gonna fund them.

那个聊了很多，但我们不如深入聊聊意愿方面？

Talked a lot about that, but why don't we dip into the willingness?

因为那正是我们刚才在说的。

Because that's what we were just talking about.

我们做过分析，实际上每兆瓦通过GB 300生成前沿代币，可以赚到高达一亿。

And we've done analysis where you can actually earn up to a 100,000,000 per megawatt generating frontier tokens off a GB 300.

利用率方面可能有点乐观，但这个东西建起来每兆瓦要花大约4000万，如果你能提前锁定自用，半年就能回本。

It may be a little optimistic in terms of utilization, but when this thing costs about 40,000,000 per megawatt to stand up, if you're foreshadowed to self fable, you make this back in half a year.

所以利润非常可观。

So it's incredibly profitable.

我们再来对比一下，新云租用GB3的典型价格，至少几个月前是这样的。

And let's compare that against what neo clouds are typically renting for a GB3 at least a few months ago.

大约每兆瓦1200万，或者说每吉瓦120亿，这是新云签订光纤合同的价格。

About 12,000,000 per megawatt or 12,000,000,000 per gigawatt was how much neo clouds run a fiber contract.

但如果你做前沿推理，就像我说的，每兆瓦可以赚到1亿。

But if you're doing frontier inference, like I said, you can do 100,000,000 per megawatt.

当人们看到SpaceX和谷歌以每兆瓦4800万的价格租用时，都感到震惊。

And people were shocked when they saw SpaceX and Google renting at 48,000,000 per megawatt.

但这很合理，对吧？

But that makes sense, right?

因为他们低价买入，高价卖出。

Because they're buying here, selling here.

所以他们当然会这么做，但人们已经习惯了看到这种情况。

So of course they would, but people were used to seeing this.

所以利润是存在的。

So the profit is there.

乔丹，我记得你提到过，每一层的利润都留在哪里？

I think, Jordan, you mentioned where does a profit stay in each layer?

所以我们在这里做的是，把整个业务组织成一张损益表。

So what we did here is we actually organized the entire business as an income statement.

你可以这样想。

So you can think about this.

想象一下你是Anthropic。

You can imagine you're an anthropic.

这是你的收入，我们用几种不同的方式来表示。

So this is your revenue, and we put it in a couple different ways.

每百万token大约1.85美元，这是Sonnet和Opus加上现金支付的合理组合。

So million token, about $1.85 per million token, which is a good blend of SONNET, Opus with cash hits.

这其实是我们自己的混合支出。

This is actually our own blended spending.

即使Opus的输入价格是55美元，现金交易也会把价格拉下来。

Even if Opus is is $55 input, the the cash hits will take it down.

所以大约是每百万token 1.85美元，这相当于如果你生成，我记得我们大概每秒生成9000个token。

So about $1.85 per million token, which is equivalent if you're generating, I think we put around, I wanna say 9,000 tokens per second.

我不记得我们用的具体数字了，但相当于每小时38美元。

I can't remember what we used, but it's equivalent to $38 per hour.

然后我们已经讨论过每小时2.80美元的成本。

And then we already talked about the cost per hour of $2.80.

然后还有Neo Cloud……

And then a Neo Cloud...

所以拥有这个系统的成本是2.80美元，但从Neo Cloud租用的成本大概在4美元以上，现在可能4.50美元，接近5美元。所以如果我们简单看一下，毛利率是多少？

So the cost to own the system is $2.80, but the cost to rent from a Neo Cloud is probably around $4 north of $4 by now, $4.50, almost $5 So if we just simplicity look at that, what is a gross margin?

是90%，对吧？

It's 90%, isn't it?

因为你卖38，成本是3.8，这就是你的利润。

Because you're selling at 38, your cost is $3.80, and this is your profit.

所以这非常赚钱，

So it's incredibly profitable,

至少对前沿模型来说是这样。

At least for frontiers.

顺便说一句，这真的很酷。

By the way, this was really cool.

每个人都应该读一读这个。

Everyone should read this.

放下手头的事，现在就去看，因为Max和团队……

Drop everything and read this now because Max and team

我们今天早些时候录了一期播客。

We recorded a pod earlier today.

你还不知道呢。

You don't know.

这是我今天的第三个播客了。

This this is my third pod of the day.

丹，我一直在背着你录节目。

I've been cheating on you, Dan.

不过，麦克斯和安德鲁确实聊过这个，天哪。

But yeah, Max and Andrew went through Oh this my gosh.

播客。

Pod.

我宁愿舔一下雾。

Get I'd lick a mist.

拜托，老兄。

Come on, man.

好吧。

Okay.

总之，麦克斯和安德鲁在上一集里聊过这个。

Well, anyway, Max and Andrew went through this on the previous episode.

可能有些朋友错过了，那就再讲一遍吧。就是说，对于这些全包消费套餐，在某些业务上，你真的能获得很多价值。

Maybe for people who who missed it, do this one again, which is like for the all in consumption plans on some of these on some of these these, yeah, these these businesses, you can get a lot of value.

但这得看情况……

But it kinda depend...

他们在这个层面能不能盈利，主要取决于人们实际使用这些套餐的程度。

If they're gonna be profitable at this layer or not, kinda depends on how much people actually use the plans that they let people have access to.

我们预估API的利润率是90%。

We peg 90% API margin.

然后我觉得在商城那边，大概用50到60。

And then I think in the mall, use like 50, 60.

但我的问题是，这些人里有多少真的用了10%？

But the question I have is how much of these people are really using 10%?

在中国，有一整个小作坊产业，专门把苹果加工成苹果酱。

And there's a whole cottage industry in China that takes apples and makes applesauce.

他们把这些20美元的套餐拿过来，然后在灰色市场上当API卖。

They take these $20 plans and then just sell them as API on the gray market.

然后我，我用的就是20美元的套餐。

And then me, I'm on a $20 plan.

我其实会特意安排我的使用时间来充分利用它。

I actually plan my sessions around utilizing.

我设置了循环，这样即使我不在也能用。

I have loops so that I use it even when I'm not there.

所以我肯定属于这个群体，但我不清楚其他人怎么样。

So, I'm definitely on this bucket, but I don't know where everyone else is.

我是说，Max之前说过，我觉得很多人可能是通过公司或者雇主拿到这些套餐的，他们其实不太在意价值是否对等，你只要把成本摊到大量用户身上，就能平衡过来。

Mean, Max made the case earlier that I think there's a lot of people who maybe get these through their business, or, you know, who they work for, and they're just not really worried about the value trade exactly, and you just kinda blend it over a bunch of users, and it works out.

那些嚷嚷着要更多额度的人，才是让你亏钱的人，但你留着他们是因为你不知道怎么区分优质用户和劣质用户，只把全包套餐卖给最赚钱的那部分用户，也就是那些不怎么用的人。

The loud people are the ones pushing for more limits, are the ones you're losing money on, but while you keep them around because you don't know how to separate the wheat from the chaff and only sell the all in plans to the most profitable segment of users, which is the ones that don't use it.

就像Planet Fitness，他们的营销策略就是卖给那些新年一月来一次就再也不来的人，赚钱的健身房就是空荡荡的。

Like, you know, the whole thing about Planet Fitness is marketing it to people, marketing gym memberships to people who come once in January for the new year and then never come back, that a profitable gym is the one that's empty.

他们根本不想有人来。

They don't want anybody to show up.

我这么说可能有点刻薄，但这就是他们推出全包计划的激励机制，而按token计费则完全相反，他们就是想让用户尽可能多地消费。

I mean, of cynical there, but that's the incentive structure for these guys to do all in plans, whereas the exact opposite is true for the per token pricing, where they're just trying to get people to consume as much as possible.

那接下来的问题就是，好吧，现在这样是赚钱的，但大家问的是，将来还会这样吗？

And then the next question becomes like, okay, it's profitable now, but like what people ask is, is that gonna be the case in the future?

让我找一下我们的token定价。

Let me find our token pricing.

天哪，我把token定价放哪儿了？

Oh gosh, where'd I put the token pricing?

然后人们会说，token价格是不是在快速下降？

And people say, fast is token price dropping?

抱歉？

Sorry?

我是说，

I mean,

不是想打断你，但让我...

not to cut you off, but let me...

好的。

Okay.

做这个。

Do this one.

因为我的问题是，乔丹，人们问我代币价格下降的速度有多快？

Because my question for you, Jordan, is people ask me how fast is token price falling?

你说下降是什么意思？

What do you mean it's falling?

它是在上涨的。

It's going up.

实际上，即使价格持平，毛利率也在上升，因为虽然代币价格保持不变，但他们生成了更多的代币。

And in fact, even if it's going flat, that's gross margin going up because while token price is staying constant, they've been generating more tokens.

然后他们就说，好吧，那很好。

And then they're like, okay, well, great.

所以我知道前沿模型一直保持稳定，但开源模型下降的速度有多快？

So I see frontier models been holding up, but how fast are open models falling?

我就说，它们其实在涨。

I'm like, well, they're going up.

另外，你看token支出的时候，它就像一个不断变化的篮子。

The other thing is when you look at token expenditure, it's like a changing basket.

所以那些追踪token支出的指数，在我看来，它们会调整这个因素。

So these indexes that track token expenditure, they adjust for that, in my opinion.

我不知道它们具体怎么运作，但我觉得它们没有调整这个。

I don't know how they work, but they don't adjust for that.

它们没有针对增长进行调整。

They don't adjust for increasing.

对于AgenTic来说，关键一直是有效价格和现金消耗。

With AgenTic, it's been all about effective price and cash hits.

我不太确定。

I don't know.

Jordan，未来token定价会怎么走？

Token pricing is going in the future, Jordan?

是的，我之前发过一条推文，点名了All In播客的大卫·弗里德伯格，他当然没回应。

Yeah, I mean, I had a little tweet thread where I went at David Friedberg from the All In podcast, he didn't respond, of course.

但很多人都在说，开源模型正在超越闭源模型。

But a lot of people are saying that the, there's like this overtaking from open models overtaking closed models.

我觉得实际情况就是，大家消费得更多了，而且价格也更高了。

And I think the dynamic is just like, people are consuming more of everything, and they're consuming it at higher prices.

所以你看屏幕上的输入价格，我觉得这是个不错的指标，但中间那个下拉菜单

So you've got the price of input on screen here, which I think is a decent metric, but that little drop down in the middle

你用的是有效价格。

You do the effective.

嗯，你也可以这么做。

Well, you could.

只是方法不同而已。

It's just a different approach.

你可以按现金读取来算，我觉得我们大部分支出其实都在现金读取这一档，因为你知道，智能体编程真的太烧钱了。

You can do it for cash read, which I think is the bulk of our spend is actually on the cash read tier, because, know, agentic coding is just so much cash it.

然后输出是最贵的层级，尽管你实际上不会生成很多输出。

And then output is like the most expensive tier, even though, you know, you don't actually generate a bunch of output.

当你的内存带宽受限时，那才是占用GPU所有时间的东西。

That is what takes up all the time on the GPU when your memory bandwidth is constrained.

总之，我的观点是，我完全同意模型定价在每百万token的混合美元基础上会随时间上涨。

Anyway, the point is like, I totally agree that model pricing is going up over time on a blended dollar per million token basis.

我认为token效率是个问号，因为你必须在解决简单问题的效率和人们尝试解决更复杂问题时的更宏大期望之间做出权衡，这些问题以前用无限token也无法解决，现在用大量token就能解决。

I think that token efficiency is a question mark because you have to trade off efficient ability to solve simple problems with people's more grandiose expectations as they try to solve more complicated problems that weren't possible before with unlimited tokens and now are possible with a lot of tokens.

我不知道怎么衡量这种权衡。

I don't know how you measure that trade off.

看，这些动态中没有一个表明有人会决定‘我对AI受够了’。”

Look, none of these dynamics to me point to anybody deciding I've had enough of AI.

我要少用点AI。”

I'm going to use less of it.

我对这些模型受够了。”

I've had enough of these models.

让我们把它们做得更差、更低效。

Let's make them worse and less efficient.

我已经受够了这些新GPU。

I've had enough of these new GPUs.

让我们把下一代做得更差、性能更低。

Let's make the next one worse and less performing.

一切似乎都在朝相反的方向发展。

It all seems to be going the opposite direction.

他们最终会提高价格，需求也会更大。

They're gonna raise prices over time, more demand for it.

GPU会变得更高效，模型性能也会更高。

The GPUs are gonna be more efficient and the models are gonna get higher performance.

所以，走着瞧吧。

So, we'll see.

有很多地方可能会出问题，如果人们决定用新的NVL 72去搞一个20万亿参数的大模型，并且真的去用它，如果他们开始引入超快模式和一些低批处理大小的优化之类的东西，那可能会把价格推高。

There's a lot of ways in which this can break if people go and decide to YOLO for the 20,000,000,000,000 parameter model with the new NVL 72 and really use it, If they start introducing, but you know, ultra fast mode and some of these low back size optimizations and stuff that can jack prices up.

但我觉得所有这些都只是人为的限制，在我看来，这些产品的需求依然非常旺盛，正是这种需求在推动底层GPU的使用。

But I think all of that is just like artificial constraints on what to me feels like continued overwhelming demand for these products that drive the usage of the GPUs underlying it.

我们之前聊了一下利润表，就是那个显式的，但上面的利润表呢？

And we talked a bit about the income statement, the explicit one, but what about the income statement above?

用户那边怎么样？

What about the users?

那也是等式的一部分。

That's part of the equation too.

他们得到了什么价值？

What value are they getting?

所以，当他们能获得这样的效率时，我不明白为什么有人会退缩。

So I don't see why anyone would step back when these are the kind of efficiencies they're getting.

我们建立了一个包含数千个债务融资工具和所有这些关系的完整数据库，完全来自公共领域。

We built an entire database of thousands, thousands of debt facilities and all these relationships, purely from public domain.

这件事一直是我、奥利弗和特伦斯在做。

It's been myself, Oliver, Terrence have been doing this.

而且我们都是当前团队的一部分，因为我们在招人，就像我提到的，PSA。

And we're all part of current groups because we're hiring, like I mentioned, PSA.

但如果是以前，我可能需要四个月时间和四个人才能做到我们现在建立这个数据库和关系表的一半好。

But it would've taken me four months and a team of four to do half as good of a job as we've done so far in building up this database and relationship table so far.

所以这些，我们经常展示这张幻灯片。

And so these are, we've showed this slide a lot.

我们对此有过讨论，但再说一次，我不明白为什么有人会花6美元去创办一家公司，然后还会回头。

We had a pot on this, but again, I don't see why anyone paying $6 to initiate a company would ever go back.

乔丹，你会回头吗？

Would you ever go back, Jordan?

还是你会直接不用快速模式？

Or would you just stop using fast mode?

比如，你会怎么做？

Like, what would you?

我不用快速模式。

I don't use fast mode.

我是说，两者都行。

I mean, either.

你得从我的冰冷、死掉的手里才能抢走这些东西，兄弟。

You'd have to pry this stuff out of my cold, dead hands, man.

现在即使在个人生活中，和一些客服代理打交道时也是这样。

Even in my personal life now, with some of these like support agents.

就像

It's like

挑一瓶酒，挑一瓶带去和朋友吃饭的酒。

pick a wine, pick a wine to bring to dinner with my friends.

比如，好吧，妈妈们，三美元我就

Like, okay, moms for $3 I

还是保留一点人情味，丹。

still still have a personal touch, Dan.

好吗？

Okay?

比如，我不会去写生日贺卡。

Like, I'm not writing birthday cards.

那是你自己的想法。

Speak for yourself.

我还没那么夸张。

I'm not that far off

但确实，我刚刚拿到了一笔航班延误四小时的退款。

the deep end, but yeah, like I just got this refund on a flight where I was delayed on for four hours.

你知道，我绝不会再打电话去要退款了，对吧？

Few, you know, There's no way I would sit on the phone trying to get this refund ever again, right?

我就像是在查时间。

I'm like checking times

在

at

健身房，或者查航班、找酒店、看购物清单，或者查邮件，老兄。

the gym, or I'm checking booking a flight, or looking for hotels, or checking on the grocery list or just check my email, man.

比如昨晚收到的那些东西。

Like what came in last night.

我是说，这有点奇怪，但我跟一些朋友聊过，他们还没意识到有个私人助理能带来多少好处。

I mean, it's a little weird, but I talked to some friends who are not enlightened to the possibilities that having a personal assistant can provide.

而且感觉一旦体验过节省下来的时间和那种浪费了多少时间的感觉，就再也回不去了。

And it does feel like everybody can never go back once they get the experience of the amount of time that you save and the feeling of like how much wasted time.

对。

Yes.

对。

Yes.

我觉得Muse会让人大开眼界。

And I think muse is gonna be eye opening.

就像，我干嘛要花两个小时在那些订机票的网站上翻来翻去？

It's like, why did I ever sit there for two hours and browse all these airline booking sites?

我干嘛要……

Why did I ever...

我觉得会有那么一刻，

I think there's gonna be that moment,

我觉得这么说有点自以为是了，因为你知道，你有点把AI拟人化了，然后你会觉得今天有个人助理的人像是在利用这个东西。

It gets a little egotistical when you talk about it this way, I think, because, you know, you kinda anthropomorphize the Asian, then you view people who have personal assistance today as kind of exploiting like this thing.

但不管怎样，我个人看法是，我们浪费在那些单调事情上的时间多得惊人。

But, anyway, my personal view is kinda like, it's shocking how much time we waste on this monotonous stuff.

而且让我震惊的是，你知道，我身边有些亲近的人根本不珍惜自己的时间。

And it's shocking how, you know, some people that I'm close with do not value their own time.

你知道，我就不细说了，但你可以想想我们被说服了

You know, I'm not gonna go into all the details there, but you can think we're sold

被说服了，对吧？

on willingness, right?

我们被说服AI很有用，对吧？

We're sold on the AI is useful, right?

感觉，

Felt,

自从在SemiAnalysis工作以来，我觉得我们一直处于采用很多新技术的尖端，并且感受到了使用它的能力，这不仅让我们的研究更好，也为我们看到的一些数据背后的研究提供了信息。

since working at SemiAnalysis, I felt like we've been at the tip of the spear for adopting a lot of and this just like feeling the ability to use it, which both makes our research better and informs the underlying research that motivates some of the data that we see.

看到我生活中和工作中的其他人也开始采用这些东西，并重新掌控生活中的很多事情，真的很好。

And it's really nice to see other people that we work with in my personal life adopt these things and then start to like regain control over a lot of things in their life.

真的很酷。

It's really cool.

看到这些真的很酷。

Really cool to see.

而且，我就是，是的，当我看到人们有这些好的使用体验，并开始受到激励去共同建设一个更积极、高质量的未来时，我真的非常开心。

And, just, I, yeah, I just, I get really happy when I see people have these good experiences using it, and start to get motivated to like build a more positive and, high quality future together.

我不知道怎么总结，但好吧，跟我聊聊融资意愿吧。

I don't know how to wrap But yeah, talk to me about the willingness to finance.

你觉得

Do you

是的，是的。

think Yeah, yeah.

我们刚才已经聊过这个了。

Well, we just talked about that.

我觉得我们已经信了。

I think we're sold.

我需要说服你的，还是意愿和能力。

What I gotta convince you on is, again, willingness and ability.

而且我得让你相信能力这部分。

And I gotta sell you on ability.

我可能刚才吓到你了，但是……

I might've scared you with this, but...

跟我说说

Talk to me

这里的人的因素。

about the human element here.

因为我觉得在SemiAnalysis工作，我也意识到一件事，就是这些看起来特别数据驱动的事情，最后往往取决于某个人的决定。

Because I think one thing I've also realized working on SemiAnalysis is how much of the, this stuff that seems so data driven comes down to some human being making a decision.

就像萨蒂亚说暂停，那就暂停。

Like Satya said pause, therefore pause.

或者黄仁勋说建，那就建，或者兜底什么的。

Or Jensen said build, therefore builds or backstop or whatever.

所以，这到底是怎么落到实处的？这些公司的领导者，他们真正体验过用个人助理或AI，或者有过那种个人时刻，让你觉得：天哪，我能在数学发现发表之前就看到它。

So where does this break down to genuinely what are the leaders of these companies having this experience using a personal assistant or an AI, or just like having that personal moment where you're like, oh my God, I can see the math discovery before it comes out and gets published.

或者我能看到我们在癌症药物上取得的进展，或者别的什么。

Or I can see this progress we're making on this cancer drug or whatever.

然后我就真的想把整个公司押在这上面，我想做这项工作，因为它对我意义重大，这就是我毕生的事业。

And like, I just really want to bet the company on this, and I wanna do this work because it means something to me, and this is like my life's work.

而不是去合理化所有财务细节来让它运转。

As opposed to rationalizing all of the financial details to make it work.

你能解释清楚这两者之间的平衡吗？

Can you even contend with explaining that balance between those two things?

但这是个大问题，因为刚才说到……

But that is one of the big questions because covered...

这还是在意愿层面，对吧？

This is still under willingness, right?

因为我们已经信了，对吧？

Because we're sold, right?

我觉得你的意思是，其他人也信了吗？

And I think what you're saying is, Are other people sold too?

这张图其实讲的就是这个，对吧？

And that's kind of what this chart is a little bit about, right?

我想提一个稍微不同的问题，因为我会把用户分成几类：消费者、API、最前沿的人，比如AI原生用户，然后是其他企业用户，对吧？

And I think the question I would make one that's slightly different, because what I would do is I stratify to consumer, API, tip of the spear, people like AI natives, and then the rest of the enterprise, right?

所以我觉得我们在最前沿这部分已经确信了。

So I think we're convinced on the spear, tip of the spear.

我觉得音乐会在消费者领域证明这一点，但企业领域我不确定，因为我经常参加这些活动。

I feel like music is going to prove the case in consumer, but I don't know an enterprise because I sit at these...

我参加很多这类活动，也见到很多人。

Do a lot of these events and I meet a lot of folks.

我们会办一些关于AI的讲座和晚宴。

We do talks and dinners on AI.

然后就是财富500强的那些人。

And then kind of Fortune 500.

他们就会说，哦，我拿它来干嘛呢？

They're like, Oh, what would I use it for?

就用来做做摘要什么的？

Just summarizing stuff?

不，它的用处可远不止这些……

No, it's so much so...

我觉得财富500强里就是缺乏发现，也缺乏想象力。

I feel like there's just a lack of discovery in Fortune five hundred and lack of imagination.

很多人其实并不是带着开放的心态来的。

A lot of people, they actually kind of don't come out this with an open mind.

他们来的时候就是，哦，挺好的。

They come out this like, Oh, great.

人工智能。

AI.

我怎么裁员？

How do I fire people?

他们满脑子想的都是这个。

That's all they think about.

就像……

It's like...

所以我觉得企业这块，三者之中，我最不了解的就是它。

So I just think the enterprise is, of the three, it's the one I don't know the most.

我不了解是因为这里面牵扯太多，比如CEO会不会说，嘿，咱们拥抱它吧。

I don't know it because it's so much to like, does the CEO say, Hey, let's embrace it.

咱们相信亚马逊。

Let's trust Amazon.

咱们全力以赴。

Let's go full blast.

或者他们会说，工作中不要用AI。

Or they go like, No AI in your work.

你觉得呢，你和《财富》杂志聊过吗？

When do you think, have you talked to Fortune?

你觉得他们在这件事上是什么态度？

Like, where do you think they are on this?

我觉得《财富》500强的情况可能是，它们会慢慢被那些全面拥抱AI的公司取代。

Well, I think the dynamic that I imagine happens with the Fortune five hundred is that the Fortune five hundred just slowly gets replaced with companies that are all in on AI.

看看排名靠前的公司。

Well, look at the top

可能前十名确实是这样。

Possibly 10 companies true.

看看全球市值前十的公司。

Look at the top 10 companies in the world by market cap.

我的意思是，这些公司里，十有八九现在都跟AI、半导体或者数据中心有关。

I mean, these are all, like nine out of 10 of them are AI or semiconductor related or data center related at this point.

比如我们说超大规模，那五家都在里面。

Like when we say hyperscale, all five of them are in there.

然后还有英伟达、博通、台积电。

Then you've got Nvidia, Broadcom, TSMC.

内存公司也快进前十了。

The memory guys are getting close to the top 10.

所以，财富500强，我觉得我们用这个词来形容，如果你指的是按收入排名的全球最大的500家企业，我觉得很多公司可能就……

So, Fortune 500, I think we use that term to describe, if you actually mean the 500 biggest businesses in the world measured by revenue, I think a lot of them are just going to Maybe

我应该说是美国企业界。

I should say corporate America.

传统企业，美国企业界。

Legacy businesses, corporate America.

传统企业，对。

Legacy business, yeah.

政府，就是那些对采用AI比较慢的群体。

Government, just like people that are slow to adopt AI.