Latent Space
The Programmable Data Center — Evan Conrad, SF Compute
00:0036m43s
00:00
Nvidia is obviously like the 800 pound gorilla in the room. Why didn't they make SF compute?
Nvidia 显然是房间里那头八百磅的大猩猩。他们为什么不做 SF compute 呢？
00:06
So our impression is that Nvidia can't make it but benefits tremendously from it existing.
所以我们的印象是，Nvidia 做不了这件事，但它从这件事的存在中获益巨大。
00:10
The reason why they can't make it is because they would then not be independent enough
他们做不了的原因是，那样的话他们就不够独立，
00:14
for people to trade on it.
人们没法在上面交易。
00:25
Anyway, Evan, welcome back to the InSpace.
总之，Evan，欢迎回到 InSpace。
00:27
Thank you so much.
非常感谢。
00:28
Welcome back to AIE.
欢迎回到AIE。
00:29
Actually, first time in AIE.
其实是第一次来AIE。
00:30
I think so, yeah.
我想是的，对。
00:32
This is a very special one because I feel like you have been working on
这一期非常特别，因为我感觉你一直在研究
00:37
computer trading, computer markets for a while.
计算机交易、计算机市场有一段时间了。
00:39
Now it's like feeling of the moment, jump, basically announce that they are getting into
现在感觉是顺势而为，Jump基本上宣布他们要进入
00:45
a computer market which is why I'm excited to cluster people. You met Oliver from semi-analysis
一个计算机市场，这就是为什么我很兴奋能把人们聚集起来。你遇到了来自semi-analysis的Oliver
00:51
who is basically initiating coverage on computer markets as if it's my read on this.
他基本上是在启动对计算机市场的覆盖，就好像这是我对这件事的解读。
00:57
So people are going to watch your talk. I think it's a very good explanation of both your
所以人们会看你的演讲。我觉得这很好地解释了你的
01:04
origin story and why you need to exist and what the hard problems you work on.
起源故事，以及你为什么需要存在、你致力于解决的难题是什么。
01:07
Yeah. Done in a very simple, engaging manner.
是的。用非常简单、引人入胜的方式完成。
01:10
I'm very jealous of your writing skills because you showed the flyer that you showed at New
我非常羡慕你的写作能力，因为你展示了你在New上展示的那张传单
01:14
Rips. It was like, it's pretty straightforward and like simple and like, yeah, it's good.
太厉害了。就感觉，挺直接挺简单的，而且，对，挺好的。
01:19
So I mean congrats.
所以我是说恭喜啊。
01:20
Thank you.
谢谢。
01:20
I appreciate that.
很感谢。
01:22
And also congrats on AIE.
也恭喜AIE。
01:23
It seems to y'all are rolling.
感觉你们真是势如破竹啊。
01:26
We're rolling.
我们开始了。
01:26
I just came back from the site tour for next year's AIE and then for a hint.
我刚从明年AIE的场地考察回来，然后给个提示。
01:31
A Javits Center. It's the same place they do Comic-Con.
Javits中心。就是他们办Comic-Con的同一个地方。
01:34
Okay. Wow.
好的。哇。
01:34
Which, wow, wow, damn, but I think the finance focus is really hitting for folks.
哪个，哇，哇，天哪，但我觉得金融方向的聚焦真的打动了大家。
01:41
Yeah.
是的。
01:41
So we're going to try to split up.
所以我们打算试着分一下。
01:44
Like this is here, we just lump everything in finance.
就像这里，我们把金融领域的所有东西都归到一起。
01:46
Like your finance, their finance, like insurance is finance, but like they're all super different.
比如你的金融，他们的金融，比如保险也是金融，但他们全都超级不一样。
01:50
Yeah.
是的。
01:51
So we'll probably break out things more minutely in future because people want to go deep on hedge
所以我们未来可能会把东西拆得更细一些，因为人们想在对冲方面深入探讨
01:57
fund stuff or banking or personal wealth or especially in New York.
搞fund stuff或者banking或者personal wealth，尤其是在New York。
02:02
Yeah.
是啊。
02:03
Yeah.
是啊。
02:03
Anthropic was like, yeah, we're all about AIE right now.
Anthropic 当时就说，对，我们现在全力搞 AIE。
02:06
Yeah.
是啊。
02:06
AIE is like registered investment advisors.
AIE 就是注册投资顾问。
02:08
Yes.
对。
02:09
So different from you.
So different from you.
02:12
Where are this?
Where are this?
02:13
Yeah, yeah.
Yeah, yeah.
02:14
Okay, so for listeners who don't know, we did a podcast two years ago where you talked about
Okay, so for listeners who don't know, we did a podcast two years ago where you talked about
02:18
your, you know, where as a compute was.
your, you know, where as a compute was.
02:23
Today, you're much more in the sort of physical settlements.
Today, you're much more in the sort of physical settlements.
02:26
You talk, you talk about your back fabric.
你讲，你讲你的 back fabric。
02:28
What do you call it?
你管它叫什么？
02:29
The back plan.
The back plan。
02:30
Back plan, which obviously that takes a lot of hardcore engineering stuff.
Back plan，显然那需要大量硬核工程的东西。
02:34
Can you give us status update on like where is SF Compute versus listeners who last
你能不能给我们更新一下 SF Compute 现在的进展，相对于那些上次
02:38
heard you own a podcast two years ago?
两年前听你播客的听众来说？
02:40
So when we started, we were this company called Junark and then we were just trying to
所以一开始，我们这家公司叫Junark，那时候我们只是想
02:44
sublease our cluster.
转租我们的集群。
02:46
And we started there in the technical stack.
我们从技术栈那块开始做起。
02:50
And a way to think about it is it's like you're in a hotel and you're trying to sell that
你可以这样理解，就像你在一家酒店里，你想把那个
02:54
hotel room. And then somebody asks you, hey, like are you going to clean my room?
酒店房间卖出去。然后有人问你，嘿，你会帮我打扫房间吗？
02:58
And then you say, oh, I guess I probably also need the hotel room as well.
然后你说，哦，我大概也需要把酒店房间一起做了吧。
03:02
Or like I need the cleaning closet.
或者就像我需要清洁柜。
03:04
And so you go get access to the cleaning closet.
然后你去获取清洁柜的访问权限。
03:07
And then they want room service.
然后他们想要客房服务。
03:08
And so you get access to the kitchen.
于是你获得了厨房的访问权限。
03:10
And then, you know, they want security.
然后，你知道，他们想要安保。
03:12
And so you get access to the security closet.
于是你获得了安保柜的访问权限。
03:13
And you get this and you do that and you do this.
然后你得到这个，你做那个，你做这个。
03:15
And at a certain point, you're just running the whole hotel.
到了某个节点，你就是在运营整家酒店了。
03:17
And the moment you're running the whole hotel, you say, oh, wait a second.
而当你运营整家酒店的那一刻，你会说，哦，等一下。
03:21
Instead, my supplier on this market is somebody that looks closer to a fund.
相反，我在这个市场上的供应商更像是某种接近基金的东西。
03:29
Or a cloud that has access, in our case, a cloud that has access to the data center
或者一个云，它拥有访问权限，在我们的案例中，是一个能访问数据中心的云
03:34
and we're going to level up.
而我们要升级了。
03:35
And so what we've spent the last several years doing is building the entire operating capacity
所以我们过去几年一直在做的，就是构建整个云的运营能力，
03:39
of a cloud in order to offer, in order to actually deliver the compute at a reliable level.
为了提供、为了真正以可靠的水平交付算力。
03:48
We call that physical settlement.
我们称之为physical settlement。
03:49
We're taking a term from the futures markets from other markets
我们借用了期货市场和其他市场的一个术语，
03:54
where in oil or soybeans or something, physical settlement means you get delivered the oil and the
在石油或大豆之类的市场里，physical settlement意味着你实际收到石油和大豆。
04:01
soybeans. But in some markets, like in oil and also in power, there's like a delivery network.
但在某些市场，比如石油还有电力市场，存在一个交付网络。
04:12
In oil, you have to have pipelines.
在石油行业，你必须得有管道。
04:14
If you would like the oil to be delivered to you, there has to be a way for it to get to you.
如果你想让石油送到你手上，就必须有办法把它运到你那里。
04:18
Yes. And for people interested in financial history during COVID, oil went negative because
是的。对于那些对新冠期间金融历史感兴趣的人，石油价格变成了负值，因为
04:26
I will pay you to take oil from me because it is sitting there in the port and I can't pay for this.
我愿意付钱让你把石油从我这里拿走，因为它就堆在港口，我付不起这个费用了。
04:32
Yes. And so I will pay you to take my oil.
是的。所以我会付钱让你把我的石油拿走。
04:34
Yes. That is the weirdness of physical settlement.
是的。这就是实物交割的诡异之处。
04:38
Yes. There's a bunch of weirdnesses like this in other markets as well.
是的。在其他市场也有很多这种奇怪的现象。
04:42
So in the power markets and the grid, there's parts of the orders that you fill out,
所以在电力市场和电网里，有些订单的部分是你要填写的，
04:47
the buyers and sellers that are there because the boiler takes a while to cool off.
买家和卖家之所以在场，是因为锅炉需要一段时间才能冷却下来。
04:55
And you can stop the boiler and then it will, you can start it back up again within some
你可以把锅炉停下来，然后它就会，你可以在某个时间内重新启动它
05:00
period of time. And so that has to get factored into the order book. This genre thing is the
一段时间。所以这必须被纳入订单簿的考量之中。这个genre的事情是
05:04
same thing that happens in compute markets. So we co-designed physical settlement and the order
同样的事情也会发生在算力市场。所以我们共同设计了实物交割和订单
05:09
book at the same time. And that meant that we get to actually handle the real constraints of compute,
同时预订。这意味着我们实际上要处理算力的真实约束，
05:15
which might include something like, oh, I'm okay with being interrupted, but only if there's
其中可能包括类似这样的情况：哦，我可以接受被中断，但前提是
05:20
like storage on the thing that I can recover from really quickly. Very similar types of problems.
设备上有存储，能让我很快恢复。非常类似的问题类型。
05:25
Yeah. So people can refer to your talk or you've done a couple other podcasts already.
是的。所以大家可以参考你的演讲，或者你已经做过几个其他播客了。
05:29
So I want to be like to sort of tool one level instead of the one-on-one level.
所以我想稍微提升一个层次，而不是一对一的那种层面。
05:33
Sure. And also because I have been a derivative
当然。而且也是因为我一直是一个衍生
05:36
straighter so I can give people a bit more detail that you normally don't get into.
说得更直接一点，这样我能给大家多讲一些通常你不会涉及的细节。
05:41
So for example, well, how mature or deep or liquid is the compute market today?
那么比如说，今天的算力市场到底有多成熟、多深入、流动性有多强？
05:46
Not very mature or deep or liquid. So what we've been doing is first we had to go get a bunch
不太成熟，也不够深入，流动性也不够。所以我们一直在做的是，首先我们得去搞一批
05:52
of supply online in the first place, which just takes a while. And then we had to build the entire
首先得让供给上线，这本身就需要一段时间。然后我们还得搭建整个
05:57
platform in order to do this. And we just now got that working. And so now we move capacity
platform才能做到这一点。而我们现在刚刚让它跑通了。所以现在我们把capacity转移过去
06:03
over onto the order book, which makes it liquid. And so 5% of our total capacity is on the order book.
转到订单簿上，这样就具有流动性了。所以我们总容量的5%放在订单簿上。
06:10
And we're consistently moving more and more. And then we're bringing on more capacity.
我们一直在不断地增加更多、更多的量。然后我们还在引入更多的容量。
06:13
That's all on the order book. The nice part is that maybe I should have said the volume on
关于订单簿就这些了。好的一点是，也许我该说的是成交量在
06:19
the actual order book is small, but the liquidity on the order book is high. So for that small amount
实际的订单簿很小，但订单簿上的流动性很高。所以对于那么小的量
06:24
of capacity that you have, you have like very concentrated liquidity. So a sort of standard metric
在你拥有的容量中，你有非常集中的流动性。所以一种标准指标
06:32
I would say for someone who is using this is a customer who buys capacity from us and they spend
我会说，对于使用这个的人来说，是从我们这里购买容量的客户，他们花费
06:37
$100 to buy that capacity. We'll make back $25 on average from selling it back to somebody else.
花100美元买那个容量。我们平均能通过把它卖给别人赚回25美元。
06:43
And that's the number that we, or this is one of the metrics that we try to track because
这就是我们，或者说这是我们试图追踪的指标之一，因为
06:48
we want the liquidity to be useful to the commercial participant. And you can measure liquidity
我们希望流动性对商业参与者有用。你可以用很多不同的方式来衡量流动性
06:52
in a bunch of different ways. You can say like, oh, I just want tons of makers or something.
。你可以说，哦，我只想要大量的做市商之类的。
06:56
And you can measure it in like high-minded ways. We try to measure it by like, is this solving the problem
你可以用很高大上的方式来衡量。我们试着用这种方式来衡量，比如，这是否解决了客户的
07:02
for the customer? Yeah. Got it. And so last time we talked as well, there was a lot of conversation
问题？是的。明白了。所以上次我们聊的时候，也有很多讨论
07:09
about standardizing the contracts, talking with like the CFTC or whatever, what's the status of
关于标准化合约，跟CFTC之类的机构沟通，现在是什么情况
07:16
those things? So I think standardizing the contracts in general is still pretty far away. And the
那些东西？所以我认为总体上标准化合同仍然相当遥远。而且那个
07:23
what we were trying to do at the time was, or basically what we learned over the last several
当时我们试图做的事情是，或者说基本上我们在过去几年里学到的是
07:29
years was we just needed to operate the whole thing. Otherwise, you cannot standardize the thing.
years 我们只需要运营整个东西。否则，你没法把这个东西标准化。
07:34
And so to get closer to something that can be standardized, you first need to make all the
所以为了更接近可以标准化的东西，你首先需要把所有
07:38
operations of the thing the same. To give an example of what can be different is a single operator can
这个事物的运作方式是一样的。举一个可能不同的例子，单个operator可以
07:45
offer two different types of virtual machines. One type of virtual machine might have a 40% performance
提供两种不同类型的虚拟机。一种类型的虚拟机可能有40%的性能
07:51
degradation versus another types of virtual machine. And 40% is a lot. When you say performance
与另一种类型的虚拟机相比的性能下降。而40%是很大的。当你说性能
07:56
degradation, are you talking like MFU type things or like a monoburn in that you've experienced or
退化，你是在说MFU之类的东西，还是你经历过的monoburn之类的，或者
08:01
what? It will affect. So in this particular case, I'm talking about GPU kernel launches and how long
什么？它会有影响的。所以在这个特定的情况下，我说的是GPU kernel的启动以及多长时间
08:05
it takes for that to happen, which is a process that happens quite frequently when you're doing sort of
这发生所需要的时间，而这个过程的频率其实挺高的，当你在做那种
08:10
either training or inference. That will basically, the end result is that you get a worse MFU, but
要么是training要么是inference。基本上，最终结果就是你会得到一个更差的MFU，但是
08:19
like is the intermediate level? Yeah, so anyone who's treated commodities before, there's different
像是中间层吗？对，所以任何以前做过大宗商品的人，都有不同的
08:26
grades of commodity. Yes, right. There's like, it's not oil, it's like sweet crude.
商品的等级。对，没错。它不像石油，它像低硫轻质原油。
08:30
Yes, versus like whatever the opposite of life sweet crude is. Or it's like, you know, corn level
对，相对于低硫轻质原油的反面之类的。或者就像，你知道的，玉米A级
08:34
A or corn B or whatever. And so like people are trying to, well, you have to establish the
或玉米B级之类的。所以人们试图，嗯，你必须建立这些
08:39
equivalence of those things. Correct. And I'm wondering if they have been established. So what we did
东西的等价关系。没错。我想知道它们是否已经建立起来了。所以我们
08:43
instead was we created an order book that had multiple different constraints. And you can say,
取而代之的做法是，我们创建了一个带有多个不同约束条件的订单簿。你可以说，
08:47
I want to have, you know, something that's between this and that, it runs from this time to this time.
我想要，你知道的，某个介于这个和那个之间的东西，它从这个时候运行到那个时候。
08:53
It's like X to Y. It's a bunch of different properties that you can attach your order.
这就像X到Y一样。它是一堆不同的属性，你可以把你的订单附加到上面。
08:58
And then we match that against cell orders that have those properties in a big matching engine.
然后我们在大匹配引擎里把它和具有这些属性的细胞订单进行匹配。
09:02
And the reason why we did it this way is it allows us to generate liquidity without having to
我们之所以这样做，是因为它让我们能够在不需
09:07
standardize before you know what the technology can do. And because the technology was shifting so
在你还不知道这项技术能做什么之前就标准化。而且因为这项技术在不断变化，所以
09:12
quickly, we wanted to basically be able to do that. The trade off of this is that you have less
很快地，我们基本上希望能够做到这一点。这样做的代价是你拥有更少的
09:17
concentrated liquidity. And so the prices are going to have, there's going to be more spreads between
集中流动性。所以价格会有，之间会有更多的价差
09:22
the buy and the sell. Yeah. This sucks. Yeah. But on the other hand, it means there's lots of
买和卖。是的。这很糟糕。是的。但另一方面，这意味着有很多
09:27
arbitrage to make if you can do that correctly. So if you have a product that can use lots of
如果你能正确做到这一点，就能套利。所以如果你有一个产品，可以使用大量
09:32
different types of capacity, then you might really like SF compute because you will be able to use
你可能会很喜欢SF compute，因为你可以使用不同类型的能力，然后你就能用
09:37
lots of different types of capacity and you'll make more money. You can take the spread.
很多不同类型的capacity，你就能赚更多钱。你可以赚那个差价。
09:41
Okay. What dimensions matter in this grid of things? So a bunch of different things. One is
好。在这个网格里，哪些维度是重要的？所以是一堆不同的东西。一个是
09:50
dimension categories that I will say is there a workload running on it right now? Or is it
我将会说的维度类别是，现在上面有工作负载在运行吗？还是说
09:54
capacity in the future? So is it capacity right now? Capacity right now in order for you to sell that,
未来的capacity呢？那现在是capacity吗？现在的capacity，你要卖出去的话，
10:01
you have to interrupt your current workload. So you have to think, do I care if my current workload
你必须中断你当前的工作负载。所以你得想，我到底在不在乎我当前的工作负载
10:05
is interrupted or not? Capacity that's sold that's starting a month from now, you can buy and sell
被中断？而一个月后开始的capacity，你可以随时买卖。
10:10
that all the time. And so you get liquidity on that really easily. There is some concern like, oh no,
所以那个流动性很容易就有了。确实有一些担忧，比如，哎呀，
10:15
well I have capacity next month, but it's very different than if you're interrupting a workload
我下个月有capacity，但这跟你现在中断一个工作负载是非常不同的。
10:18
right now. The way that the interruption happens matters. So an example is if you're on a cluster and
中断发生的方式很重要。举个例子，如果你在一个cluster上，
10:24
the cluster doesn't have as amazing storage as you want it to have. So an example is like, do you
这个集群的存储没有你期望的那么出色。举个例子，比如你
10:29
have a weko or vast system under the hood? And do you offer that? And like what's the size of that?
底层是不是有一个Weko或者Vast系统？你们提供这个吗？大概是什么规模？
10:35
This changes how you respond to being preempted or removed off the cluster.
这改变了你在被抢占或被移出集群时的应对方式。
10:41
Meanwhile, the location of that storage matters. So if you're able to move that data from one cluster
与此同时，那个存储的位置也很重要。所以如果你能把那些数据从一个集群迁移走
10:48
to another quickly, then maybe you might be able to spread across multiple different clusters.
快速切换到另一个，然后也许你就能分散到多个不同的cluster上。
10:53
And you can be interrupted over here and then pop back up over there. To make that work,
你可以在这边被打断，然后又从那边冒出来。为了实现这一点，
10:58
you also need like a fast transit path between one cluster and another. Typically,
你还需要一个集群之间的快速传输路径。通常，
11:03
this is implemented with a WAN, but most of these clusters don't necessarily have
这是用WAN实现的，但这些集群中的大多数不一定具备
11:07
like a wide area network. Yeah, a wide area network. So basically cables that connect the clusters
就像一个广域网。对，一个广域网。所以基本上就是连接各个集群的线缆
11:11
together. A private internet is another way to think about it. A lot of clusters that are being
一起。私有互联网是另一种思考方式。很多集群正在
11:16
spun up today are like, they're not attached to traditional CPU clusters were next to the customer
今天启动的那些，跟传统的CPU集群没有关联，它们就在客户旁边
11:25
for latency reasons. In GPUs, they're next to the power wherever the like, you know, you put the
出于延迟方面的考虑。在GPU里，它们紧挨着电源，随便你放在哪儿，你知道的，你把那个
11:32
data center next to the power rather than next to the customer. This means that a lot of stuff is
data center next to the power rather than next to the customer。这意味着很多东西
11:37
getting put in places that previously they weren't expecting data centers there. And so you may not
正在被放到以前他们不预期会有data center的地方。所以你可能不会
11:42
have as excellent of a transit path from one to the other. So this starts to matter. Other examples
拥有一条从一个到另一个那么优秀的transit path。所以这开始变得重要了。其他例子
11:49
are weird shit where we theoretically support but don't actually like this is a thing that we're
是一些奇怪的东西，我们理论上支持但实际上并不喜欢，这是我们在
11:59
thinking about but haven't done basically. The early GB 200s broke the early H 100s broke.
考虑但基本上还没做的事情。早期的GB 200坏了，早期的H 100坏了。
12:07
And so if you were buying H 100s or you're buying GB 200s, you might really want the ones that
所以如果你在买H 100或者在买GB 200，你可能真的想要那些
12:11
were deployed after a certain time because you're not getting the first JD ones off the rack. And so
它们是在某个时间之后才部署的，因为你拿不到第一批JD的产品。所以
12:16
if you're a super sophisticated customer, you may be willing to pay more because for that. Oh, so it's
如果你是一个非常精明的客户，你可能愿意为此多付一些钱。哦，所以这
12:21
about like what cohort they were they were produced and they get better over time. So like the first
跟它们是哪一批生产的有关系，而且它们会随着时间变得更好。所以就像第一批
12:28
ones off the rack maybe had some faults or some problem. And so if you were the ones who got the
产品可能有一些缺陷或一些问题。所以如果你是那些拿到
12:32
first ones, you know, first ones from the first shipment of them, then yeah, they would they would
第一批的人，你知道，第一批从第一批货里拿到的，那么是的，它们会
12:39
be broken or in some way. I assume the sort of process is like look, there's defects everywhere.
坏掉或者以某种方式出问题。我猜这个过程大概是这样的：你看，到处都是缺陷。
12:44
Like it just happens, right? Totally. And so like you get it, you get a new shipment off
就像它就是自然而然发生的，对吧？完全是这样。所以就像你懂的，你收到一批新货
12:48
from TSMC or whatever. And you just kind of run your tests and like you figure out which ones,
从TSMC或者别的什么地方。然后你就跑你的测试，看看哪些，
12:53
which was a bad and you throw them out. And isn't that how it works? And I didn't know about this
这很糟糕，你就把它们扔掉了。不就是这么运作的吗？而且我之前不知道这件事
12:58
element of cohorting where like let's say Q for 2026 B 300s are way worse than 2027 B 300s.
cohorting 的一个要素是，比如说2026年Q的B 300比2027年的B 300差多了。
13:07
Yeah. So I'll give an example. You're describing Burnin. Yes. So when you get a cluster,
对。那我举个例子。你在描述Burnin。是的。所以当你拿到一个集群的时候，
13:14
a good cloud, what it does is it tries to kill all the GPUs before they ever get in the hands of
一个好的云服务，它做的事情是试图在所有GPU落到别人手里之前就把它们干掉
13:19
the customer. And then the GPUs that are broken or the nodes that are broken, you get RMAID. You
客户。然后那些坏掉的GPU或者坏掉的节点，你就走RMA流程。你
13:25
basically go and you return it. Yeah. You return it and you get a new one and there's a lead time
基本上就是去，然后你把它还回去。对。你把它还回去，然后拿到一个新的，而且有一个等待时间。
13:29
to that. But whatever you'll get, yeah, you'll get some back. And then the customer never has to
对此。但不管你会得到什么，是的，你会得到一些回报。然后客户就永远不必
13:34
experience this. The thing I'm referring to is that like in early age 100s, there were like firmware
体验过这个。我指的是，在早期100年代，有像firmware这样的东西
13:44
bugs that may not even be related to the specific like they may not be in video. They might be
可能甚至跟具体的那个东西都没关系的bug，它们可能不在视频里。它们可能是
13:49
like super micro has a firmware bug in the in the box that you bought. And it applies to all of
比如超微在你买的那台机器里有个固件bug。而且它适用于所有
13:56
them. And it doesn't matter like the solution to that was you need to get them to update their
他们。而且这跟解决方案是什么没关系，比如解决方案就是你需要让他们更新
14:03
firmware. Or they need to ship a new thing in every new box. Yeah. The condition has is for
firmware。或者他们需要在每个新盒子里装一个新东西。对。这个情况是针对
14:08
like the the B 200s that we got. Yeah. And the IQ had to use them and say go fix it. Yes.
比如我们拿到的那些 B 200。对。然后 IQ 不得不使用它们并说去修好它。是的。
14:13
So like different types of problems like this can happen. Interesting. Okay. So I'm like,
所以像这类不同类型的问题都可能发生。有意思。好吧。所以我就想，
14:18
okay. That seems small compared to the overall. Like I just want to like I want people to understand
好吧。跟整体相比这似乎很小。就像我只是想让大家理解
14:22
the shape of this market. You mentioned like preemption, you know, and we talked about spot instances
这个市场的形态。你提到了 preemption，你知道，我们还聊了 spot instances
14:28
and how that the analogy is. Yeah. And then there's also booking a month out. How far out do
那这个类比是怎么回事。对。然后还有提前一个月预订。提前多久
14:34
those do these markets go? And is real time or sort of preemption more interesting to you versus
这些市场会走向何方？对你来说，实时还是某种抢占更有意思，相比
14:42
far be further out? To answer your first question. Yeah. Three to five years. Wow. We're starting in
远得多的多？来回答你的第一个问题。是的。三到五年。哇。我们开始吧
14:48
the month by month or quarter by quarter or the houses. Are you asking how the cashflow works?
一个月一个月地看，或者一个季度一个季度地看，还是那些房子。你是在问现金流是怎么运作的吗？
14:52
The contracts, like the futures contracts, right? Like again, like I'm a former contracts
那些合同，就像期货合约，对吧？就像，再说一次，我以前是做合约的
14:57
derivatives trader in the in the market itself. Yeah. Like is in the broader market. Yeah. Clusters are
市场本身的衍生品交易员。是的。就像在更广泛的市场中。是的。集群是
15:02
financed three to five years. Okay. So the payments are totally getting that period of time. Yeah.
资助了三到五年。好的。所以付款完全是在那段时间内完成的。是的。
15:08
There is also like blocks of capacity in our order book, which can extend any arbitrary time.
我们的订单簿里也有容量块，可以延长任意时间。
15:14
Yeah. There's some stuff that we have to do for the cashflow on that. Right now,
对。那个上面有些事我们得为了现金流去做。现在，
15:18
what we do is we just ask people to pay entirely upfront for a block. We don't ask them to do that
我们做的就是直接让人们为一个区块全额预付。我们并不要求他们那样做
15:23
for a lease for a longer term contract. This gets into something that I definitely need a whiteboard
对于长期合同的租约来说。这就涉及到一些我肯定需要白板来讲的东西了
15:27
to explain. But you had another question I was about to answer after. Do you want real time or
来解释一下。不过你还有一个问题我正准备之后回答。你是想要实时的还是
15:34
do you want more what's more interesting to you real time or longer out? So we care more about
你是想要更多，什么对你来说更有意思，实时还是更长远的输出？所以我们更关心的是
15:39
longer out because we're trying to build a forwards curve that has a lot of liquidity on it.
更长的时间，因为我们要构建一条流动性很高的远期曲线。
15:47
These are kind of the same thing and kind of not. So what we do is we since say you can buy
这些基本上是同一回事，但也不完全是。所以我们做的是，既然说你可以在
15:55
an hour out. You can buy right now to the next hour or you can buy right now to the next three
一小时后的。你可以现在买下一小时的，或者现在买接下来三小时的
16:00
years and anywhere in between. That's wild. I know. And so the, you know, or five years,
几年到五年之间都有可能。太疯狂了。我知道。所以那个，你知道，或者五年，
16:07
technically, there's a certain difference that happens that's dependent for each customer.
从技术上讲，每个客户所发生的变化是有一定差异的。
16:14
Basically, like if I have a workload that is that's a batch inference, basically, where I can
基本上，就像如果我有一个工作负载，是那种批量inference，基本上，就是那种我可以
16:22
be interrupted at any point, I don't really care. Then my ability to be cut off is probably like,
随时被打断都行，我真的不在乎。那我被切断的能力大概就像是，
16:32
I need an hour of time to know that I'm going to be cut off and then I'm good, or I can even
我需要一个小时的时间来知道我会被打断，然后我就没问题了，或者我甚至可以
16:38
be notified in a minute or two. But if I have live inference customers who are expecting me to be
一两分钟内就能收到通知。但如果我有实时inference的客户，他们期望我
16:45
online, then you can interrupt me, but in a month, maybe two months, maybe three months. And
在线的时候，你可以打断我，但一个月后，也许两个月，也许三个月。而且
16:51
then I can adapt to that capacity over that period of time. But you cannot interrupt me in a minute.
那我可以在这段时间内适应那个容量。但你不能在一分钟内打断我。
16:58
And so those two things are priced differently. And the reason why they're priced differently
所以这两件事的定价方式是不同的。而它们定价不同的原因
17:03
is because what is the price you would pay in order to interrupt your batch job? It's very
在于，你愿意为中断你的 batch job 支付什么价格？这跟
17:09
different than the price you would pay to interrupt your live job. If you have a live job, you're,
你愿意为中断你的 live job 支付的价格非常不一样。如果你有一个 live job，你
17:14
you are probably paying an SLA when you break for your customer. Does that make sense?
很可能是在为你的客户提供 SLA 保障。这样说有道理吗？
17:20
Yeah, I see. So this is actually the market for breaking leases.
嗯，我明白了。所以这实际上是打破租约的市场。
17:23
These are the market for breaking leases, basically. Oh my god. And then there's also the market for
这些基本上就是打破租约的市场。我的天。然后还有那个市场，关于
17:26
future leases, or like for leases that are, oh, I thought this was like a subletting situation.
未来的租约，或者说那些租约，哦，我以为这是类似转租的情况。
17:31
This is like, oh, these break. So one way to think about it is in power markets, there's actually
这就像是，哦，这些会中断。所以一种理解方式是，在电力市场里，实际上有
17:37
two markets. There's like the day ahead market and there's the live market. Compute kind of has the
两个市场。有日前市场，还有实时市场。算力大概也有
17:42
same thing where it has capacity that is coming online or capacity that's about to arrive to you.
同样的东西，它有即将上线或即将到达你手中的容量。
17:50
And then capacity that is currently in use. And these have different price dynamics.
然后是当前正在使用中的容量。而这些有不同的价格动态。
17:55
Okay. Interesting. Backward nation or contango?
好。有意思。是 backwardation 还是 contango？
18:00
Backward nation or contango? So is compute further out priced hired in spot or lower than spot?
落后国家还是contango？那么compute远期的定价是比现货高还是比现货低？
18:08
Is compute further out priced? Does it go like this? Does it go like that?
compute远期的定价是这样吗？是那样吗？
18:13
Over like one year out two year out, do, oh, this is both.
大概一年后、两年后，会，哦，这两个都是。
18:18
I find it fascinating. You don't know. Yeah. I mean, both it depends on who you are.
我觉得这很有意思。你不知道。是啊。我是说，两者都取决于你是谁。
18:22
Okay. So I'll make the case. Wait, so you're asking is compute right now?
好。那我来论证一下。等等，所以你问的是现在的compute？
18:28
The same unit of compute today, maybe not today, maybe a month out versus one year out versus two
同样一单位的compute今天，也许不是今天，也许一个月后，对比一年后、两年后
18:35
year out is that curve going up or going down? We see like a hump. Oh my god. Okay. So explain.
一年之后这条曲线是往上还是往下？我们看到像是一个驼峰。我的天。好吧。那解释一下。
18:42
I mean, like I think this will make interior sense is basically like if I have milk and it is about
我是说，我觉得这从直觉上很好理解，基本上就是如果我有牛奶而且它快要
18:46
to expire, I will lower the price on it. Yes. And so that's basically what we see.
过期了，我就会降价。对。所以这基本上就是我们看到的情况。
18:51
So it's the Nvidia release cycle basically, right?
所以这基本上就是Nvidia的发布周期，对吧？
18:54
Well, we're more saying like if there's a block of compute and it's approaching now,
嗯，我们更多是在说，如果有一块compute而且它临近了某个时间点，
18:58
it can't be sold in the past. And so it has to drop its price in order to clear at some price.
它没法在过去被卖掉。所以它必须降价才能以某个价格出清。
19:03
And so initially you see this drop where stuff that's sold right now, anything that's on the market
所以一开始你会看到这个下降，就是现在正在卖的东西，市场上现有的任何东西
19:08
right now is something that hasn't been sold. And so therefore the price will drop.
现在这个东西还没被卖出去。所以因此价格会下降。
19:13
And then it goes up to where it wants to be. And then maybe compute in the future is worth less
然后它会上升到它想要达到的位置。然后也许未来的compute会变得更不值钱
19:19
because people don't want like people would like the compute now. And so they're sort of paying
因为人们不像……人们现在想要算力。所以他们某种程度上是在付钱
19:23
for the short of lead time. And they pay less for the compute that's off into the future.
因为交付周期短。而且他们对未来算力的付费更少。
19:28
This is what I think a normal healthy curve looks like. Yeah. The place where I'm like, oh,
这就是我认为一条正常健康的曲线应该有的样子。是的。就是那种让我觉得，哦，
19:32
yeah, no shit, is that right now if you took a B300 and you tried to run inference on it,
是啊，废话，现在如果你拿一块B300去跑inference，
19:40
you can make $40 an hour or so selling that any model basically. And so that's a few
你基本上卖任何模型每小时能赚40美元左右。所以那是几个
19:49
that's a hundred percent. And maybe if you only have like 50% utilization or something,
那是百分之百。也许如果你只有50%的利用率之类的，
19:52
so like $25 an hour. So $25 an hour. Meanwhile, you can buy B300s at like $5 to $6 an hour.
那大概就是每小时25美元。所以每小时25美元。与此同时，你买B300大概每小时5到6美元。
19:59
One of these numbers has to change. Yeah. It's a huge arbitrage. It's a huge arbitrage.
这两个数字里总有一个得变。是啊。这是个巨大的套利。这是个巨大的套利。
20:03
This and your talks a lot. Yeah. So a bunch of people are basically looking at this and thinking,
这个和你聊了很多。是啊。所以一堆人基本上看着这个就在想，
20:09
if I think prices are going to go up, I might be willing to pay more than what I previously thought
如果我觉得价格会涨，我可能愿意比之前想的付更多
20:13
in the future because I might think, okay, I can buy a five right now. But if I can't buy anymore,
未来因为我也许会想，好吧，我现在可以买一个五。但如果我再也买不到了，
20:19
then maybe I will be willing to buy even more in the future just to lock it in if it's going to
那也许将来我会愿意买更多，就为了趁还能锁定的时候赶紧锁定
20:22
come online later. So we see the other side as well. Kind of depends on who you are.
之后才上线。所以我们也会看到另一面。这有点取决于你是谁。
20:28
Cool. So is it okay? And just I just really want to make super clear on this.
好。那可以吗？我只是想在这个问题上说得非常清楚。
20:33
Of course. Do you care about the different generations of chips or do you just care about
当然。你关心的是不同代际的芯片，还是你只关心
20:37
overall flops and you spread it among whatever chips you want? We care about different generations
总的flops，然后你想分配到哪些芯片上都行？我们关心的是不同代际
20:42
of chips because you care. I guess do your customers care? Yes. Right. Oh, wait. When you say
因为你在意芯片。我猜你的客户在意吗？是的。对。哦，等等。当你说
20:47
different generations of chips, do you mean like? So like? Yeah. Exactly. Yeah. Or like even AMD or
不同代的芯片，你是指像？所以像？对。没错。对。或者甚至像AMD或者
20:53
whatever. Right. Yes. I think people totally care about the generation of chips. I think part of
随便吧。对。是的。我觉得人们完全在乎芯片的代际。我觉得部分
20:57
this is because of things like VRAM. And part of this is because of like a larger ND length
这是因为像VRAM这样的东西。而这部分是因为更大的ND长度。
21:03
domains and things like that. So there's just no way to abstract all this to like some
domains 之类的东西。所以根本没办法把这一切都抽象成某种
21:07
generalized unit of compute that you can just mush around there and whatever you like.
一种泛化的计算单元，你可以随便在里面搅和，想怎么弄都行。
21:11
I think like when you were saying oil, even oil is the oil, oil is actually just like a bunch of
我觉得就像你说石油的时候，其实石油本身也就是一堆
21:15
different types of oil. Yeah. Like jet fuel or car fuel or whatever. I think it's the same thing.
不同类型的油。对。比如航空燃油、汽车燃油之类的。我觉得是一回事。
21:19
You're going to get a bunch of different types. It is, but you have oil refineries. Yeah. You
你会得到一堆不同的类型。确实是，但你有炼油厂。对。你
21:22
can effectively have black boxes. You can put oil in and get it like whatever out. If compute
实际上可以有黑箱。你把石油放进去，就能得到各种产出。如果计算
21:28
it is so heterogeneous, I guess the word is that you just have to expose the underlying complexity
它如此异构，我猜这个词就是说你只能把底层的复杂性暴露出来
21:34
to your end customer, which is bad for you as a market maker. Maybe. Like you want to identify
对你的终端客户来说，这对你作为做市商是不利的。也许吧。就像你想识别
21:41
the one like choke point or like I don't know what the better word for choke point is like
那个像瓶颈一样的，或者我不知道瓶颈更好的词是什么，就像
21:46
shelling point where everyone agrees like, okay, like this is kind of the standard and we'll
这是一个焦点，大家都同意的，比如，好吧，这算是某种标准，我们会
21:51
all agree to this standard. And you know, it'll be somewhat inefficient. Yes.
大家都同意这个标准。而且你知道，这会在某种程度上效率不高。是的。
21:56
So we've been able to reduce the problems to a certain level. Yeah. Your job is like
所以我们能够把问题降到一定程度。是的。你的工作就像是
22:03
dimensionality reduction. Yes. We reduce a lot of the dimensions, but nonetheless there are still
降维。是的。我们减少了很多维度，但尽管如此仍然有
22:07
dimensions. Yeah. It's still going to exist. Do you know how many skews you have?
维度。对。它还是会存在的。你知道你有多少个SKU吗？
22:12
So like five times or 100 times like 5,000 equals, I don't know, like what's
所以就像5倍或100倍，比如5,000等于，我不知道，大概是什么
22:20
we actually made skews like a concept that and the goal is to make it so that somebody else can make
我们实际上把SKU做成了一个概念，目标是让其他人也能做
22:27
their own skew. So if I want to buy black wells, I can say I'm willing to buy black wells at
他们自己的SKU。所以如果我想买Blackwell，我可以说我愿意以
22:33
$6 an hour or less. And you can match me against GB 300s or you can match me against B 300s. And I
每小时6美元或更低的价格买Blackwell。你可以把我匹配到GB 300，或者你可以把我匹配到B 300。我
22:39
don't care which black well as long as it's less than $6 an hour. But you just say you care about memory
不在乎是哪个Blackwell，只要低于每小时6美元就行。但你只是说你在乎内存
22:44
and other stuff. Totally. But not every customer does. So some customers like just can work on black
还有其他东西。完全对。但不是每个客户都这样。所以有些客户就只能在black上工作
22:49
wells, whatever it is. Some customers are super hyper specific and they need this exact thing and
wells，不管是什么。有些客户特别特别具体，他们就需要这个确切的东西，而且
22:52
they're going to pay more for that because the market is going to charge a more. Okay. And that's
他们会为此付更多钱，因为市场会收取更高的费用。好吧。那就是
22:56
what we do. So people use your API to programmatically define what they want and then you match that
我们做的事情是这样的。人们用你的API以编程方式定义他们想要什么，然后你来匹配这个需求
23:02
with whatever composition of sellers they might be able to fulfill. Yes. With the caveat that
with whatever composition of sellers they might be able to fulfill. Yes. With the caveat that
23:10
having a named query is a new thing we're doing. But the query system as a concept is how the
拥有一个命名查询是我们正在做的一件新事情。但查询系统作为一个概念，是那个
23:15
order book works right now. Okay. Interesting. All right. So you know, I think that's like a lot of
订单簿现在是怎么运作的。好。有意思。行吧。所以你知道，我觉得那就像是很多
23:22
I guess nitty-gritty. Have you ever any anything about like I guess that you want to educate people on
我猜是些细节问题。你有没有讲过什么，我猜是你想教育大家的内容
23:27
whether it's financial people or engineers about the state of the compute trading market as you
无论是金融界人士还是工程师，关于算力交易市场的现状，正如你
23:33
have developed it because effectively you guys are defining it. Mainly just the requirement for
已经开发出来了，因为实际上你们在定义它。主要就是对于
23:39
physical settlement. Yeah. We think that if you we spent three years learning why you need physical
物理交割。对。我们认为，如果我们花了三年时间才搞明白为什么你需要物理
23:44
settlement. And our entire conversation today is that it is these things are not as fungible as you
结算。而我们今天整个对话的核心就是，这些东西并不像你
23:52
think. We're going to try and make them fungible. That's the industry's job to do that. But without
思考。我们要试着让它们变得可互换。这是行业的工作。但没有
23:58
that you run a lot of risks. If you try to build too much financialization before you the industry
你会冒很大的风险。如果你在行业发展起来之前就试图建立过多的金融化
24:05
has had time to build out the physical settlement then I think what you have is you have
如果已经有时间把实体结算体系搭建起来，那我觉得你拥有的就是
24:14
a market that's very easy to manipulate and a market that has a lot of basis risk where the price you
一个非常容易被操纵的市场，以及一个存在大量基差风险的市场，在这个市场里你
24:19
may be hedging against actually is nothing to do with the price that the market. Yeah. Yeah. You can
可能是在对冲，实际上跟市场价格没什么关系。对。对。你可以
24:25
actually buy the capacity for. Yeah. So for listeners, we talked about this briefly last time. But
实际上要为此购买产能。是的。所以对听众来说，我们上次简要谈过这个。但是
24:31
people are often shocked to find out that let's say the market for credit defaults is 12 times
人们常常很震惊地发现，比如说信用违约的市场规模是12倍
24:35
the size of the bonds that they ensure. Something like that or like the market for mortgage sort of
他们担保的债券规模。类似这样的，或者说像抵押贷款市场那种
24:43
reinsurance or all that. It's larger than the underlying physical settlement. So that's why
再保险之类的。它比标的实物结算还要大。所以这就是为什么
24:48
physical versus cash settled. Yeah. And like basically cash settlements are like exactly the kind
实物交割还是现金结算。是的。基本上现金结算就完全是那种
24:54
of the side bet that you see in the big short where they may or may not have anything to do with
就像你在《大空头》里看到的那些边注赌局，他们可能跟这些有关系，也可能完全没关系
25:01
the underlying thing itself. And it may be used as a proxy for overall other stuff. Yes. And you
底层的东西本身。它可能被用作其他整体事物的代理。是的。还有你
25:06
don't want it to that to go out of control? Yeah. So to be clear, we think cash settlement as a
不想让它失控？对。所以明确一点，我们认为现金结算作为一种
25:12
future concept is a great idea. Yeah. We just think it should be derived from physical settlement.
未来概念是个好主意。是的。我们只是认为它应该从物理结算中衍生出来。
25:17
Because as you said, it's a side bet. And it would be kind of like if you kind of like if you made
因为就像你说的，这是一个side bet。而且这有点像，如果你有点像，如果你做了
25:25
a cash settlement a future before you made the physical settlement. Which I think people are
在你进行实物结算之前，先进行现金结算，一个未来。我认为人们是
25:29
kind of doing already? Yeah. Totally. People are trying to make cash settlement before you make
已经在做了吗？是的。完全是这样。人们正试图在你做之前先拿到现金结算。
25:32
the physical settlement. And people do that for all sorts of things, including startup stock.
物理交割。人们为了各种各样的事情这么做，包括初创公司的股票。
25:36
Yes. Right. Like you can buy and sell a made up share opening eye, but it's nothing to do with
是的。对。就像你可以买卖一个凭空捏造的OpenAI股份，但这跟
25:40
the actual share. Yes. And lots of those go very poorly. Like it would be kind of bad if you made a
实际的份额。是的。而且其中很多都进行得非常糟糕。就像如果你做了一个
25:49
it would be bad if you started selling the the derivative on a stock that doesn't exist yet.
如果你开始卖一个还不存在的股票的衍生品，那就糟糕了。
25:57
Yeah. Yeah. So I mean, one thing that we do, so I used to be currency trader and interest rates
嗯。嗯。所以我是说，我们做的一件事，我以前是货币交易员和利率交易员
26:02
as well. And what one thing we do is you have a board that gets together and says like this is the
同样地。我们做的一件事是，你有一个董事会聚在一起说，这就是那个
26:07
London interbank offered rate, Libor. Yes. And we all agree like this is the monthly fixing
伦敦银行间同业拆借利率，Libor。是的。而且我们都同意，这就像是每月的定盘价
26:11
and we all trade to that fixing. Yes. Is that even meaningful at some point in three years from now?
而我们都按照那个定价来交易。是的。三年后这还有意义吗？
26:18
So critically, Libor was manipulated. Yes. Our concern at the moment is this toll banks.
所以关键是，Libor是被操纵的。是的。我们目前担心的是这个收费银行。
26:24
Yes. Well, our concern is there's a very similar problem where you can manipulate an index.
是的。嗯，我们担心的是存在一个非常类似的问题，就是你可以操纵一个指数。
26:30
Our impression is why you can manipulate that index is because a lot of the indexes are based on
我们的印象是，为什么你能操纵那个指数，是因为很多指数是基于
26:34
a very small amount of data that has not a lot of correlation with the actual price of buying a
非常少量的数据，而这些数据与购买新储备合约的实际价格没有太大相关性。
26:42
new reserve contract. That's what we're concerned about. Okay. Interesting. All right. I think that's
这就是我们所担心的。好的。有意思。好。我觉得这
26:49
the the rough state of things. I think the other part of just today is you talk to Oliver and
大致的情况就是这样。我觉得今天另一部分就是你和Oliver聊了聊，然后
26:59
you're doing an event with Jump and your friends with Silicon Data. What are you learning from all
你在和Jump以及你的朋友们一起做一个活动，跟Silicon Data合作。你从所有这些中学到了什么
27:03
these other players? Who are the major players that you would want to call out as like doing
其他那些参与者呢？你会想点名说哪些主要参与者在做
27:07
interesting things? So we're big fans of Silicon Data and Computer Change. I think they were also
有趣的事情？所以我们都是Silicon Data和Computer Change的忠实粉丝。我觉得他们也是
27:14
very early to this. And I think they are doing like a reasonable shot at the A path that we think
很早就开始做这个了。而且我觉得他们在我们设想的A路径上做得还挺靠谱的
27:21
that will likely end up being partners with. Can you distinguish that you're different paths?
那很可能最终会成为合作伙伴。你能区分出你们走的是不同的路径吗？
27:26
So the general justice they're doing cash settlement. We're doing physical settlement. We think both
所以他们一般做的是现金结算。我们做的是实物结算。我们认为两者
27:31
can exist and should exist. And that they're sort of mutually beneficial to each other. And so
可以并存，也应该并存。而且它们某种程度上是互利互惠的。所以
27:36
totally fine. So I think they are one of the earliest in space. And I think they are very clearly
完全没问题。我觉得他们是这个领域最早的参与者之一。而且我觉得他们非常明显地
27:41
took the bat well before anybody else did. And I think that's like very honorable to do.
在其他人之前就率先行动了。我觉得这非常值得尊敬。
27:47
You mentioned jump analysis, jump. These are the players emerging. It's like you were my first
你提到了jump analysis，jump。这些是涌现出来的参与者。就像你是
27:54
introduction to this. And then I see all these other guys coming in and I'm like okay it's
我第一次接触这个的引路人。然后我看到所有这些其他人进来，我就想好吧这
27:57
something's happening here. I don't know the landscape as clearly as you do. Yeah. So the categories
这里正在发生一些事情。我不像你那样清楚地了解这个格局。是的。所以我们看到的几类人
28:03
of folks that we've seen is there's folks that are trying to set better standards. So semi-analysis
是，有些人在试图制定更好的标准。所以semi-analysis
28:07
I think is working to try and create better standards. There's folks that are trying to do residual
我觉得正在努力尝试建立更好的标准。有些人在尝试做残值
28:12
value insurance and so on. There's a number of different companies in the space. I don't know
保险等等。这个领域有很多不同的公司。我不太
28:16
any of them super well. But I know this is a concept and I think that's important. There are folks that
深入了解他们中的任何一个。但我知道这是一个概念，我觉得这很重要。有些人
28:21
are trying to do deposit lending where they'll front your credit risk for the cloud. There's
在尝试做存款借贷，他们会为你的云端信用风险垫付。还有
28:29
an ecosystem of companies that are forming around how do we resolve the financial bottlenecks
一个围绕如何解决财务瓶颈而形成的公司生态系统
28:33
of compute. And it is cool to see. I think effectively what you want in this ecosystem is
算力的。看到这个真的很酷。我觉得在这个生态里，你真正想要的效果其实是
28:39
buyers and sellers. You want an analyst. You want we said insurance or what's the residual value
买家和卖家。你想要一个分析师。你想要我们说保险还是什么残值
28:46
insurance. Yeah. And we want this like healthy ecosystem out there. But obviously your
insurance. 对。我们希望外面有一个健康的生态系统。但显然你的
28:53
guys are owning a problem of physical settlement which is really interesting. Okay so I mean we can
guys正在own一个physical settlement的问题，这真的很有意思。Okay，所以我的意思是我们可以
28:58
go into like the hard tech site because obviously you are also hiring for extremely cracked engineers.
去硬核技术那边招人吧，因为很明显你们也在招那种特别厉害的工程师。
29:04
Describe your stack. We do all the things the cloud does without being the cloud. As in we don't
描述一下你们的技术栈。我们做云做的所有事情，但不是云。也就是说我们不
29:12
own the hardware. The reason why we don't own the hardware is because if we did we would be able
own the hardware。我们之所以不拥有硬件，是因为如果我们拥有的话，我们就能够
29:17
to manipulate the forward market. So we have to be Switzerland. We have to be independent
去操纵远期市场。所以我们必须保持中立。我们必须保持独立
29:22
which creates a lot of sort of strange things about the company. But in order to do physical settlement
这就造成了很多关于这家公司的奇怪现象。但为了进行实物结算
29:27
you have to do site visits. You have to visit the data center. You have to you have to design
你得去实地考察。你得去数据中心看看。你得去设计
29:35
the bill of materials and ensure that you have to basically do like supply chain consulting
物料清单，并且确保你基本上得做类似供应链咨询的事情
29:41
where you have relationships with different OEMs that ensure that things get delivered on time.
你与不同的 OEM 建立了关系，确保东西能按时交付。
29:48
Then you have to do cluster architecture. So you're sort of designing the cluster itself.
然后你得做集群架构。所以你在某种程度上是在设计集群本身。
29:55
Then you're rolling out racks or you're working with partners to roll out racks and supervising
然后你在部署机架，或者你在跟合作伙伴一起部署机架并监督
29:59
that. Then you're doing backplane management. So you're taking over the PDUs and the switches and
这件事。然后你在做背板管理。所以你接管了 PDU、交换机、
30:05
the DPUs and UFM and BMC and you're operating the whole harness around the nodes that you're selling.
DPU、UFM 和 BMC，你在运营你所销售的节点周围的整个线束。
30:13
You say you as an SF compute employees. Yeah. Wow. So when you buy a node from a cloud
你说你作为 SF compute 的员工。是的。哇。所以当你从云厂商那里买一个节点时
30:20
there's the bare metal. Yeah. But there's all this shit around the bare metal. Yeah.
有裸金属。对。但裸金属周围还有一大堆乱七八糟的东西。对。
30:24
And in order to make that bare metal node work you have to operate everything else.
而为了让那个裸金属节点正常工作，你得运维其他所有东西。
30:30
Then we because I just think traditionally the job of the data center is to extract some of this
然后我们，因为我觉得传统上数据中心的工作就是帮你把其中一部分
30:36
away from you. Yeah. We're the ones that do that. Yeah. But you're not you're not in the data center
抽离出去。对。我们就是做这个的。对。但你们不是，你们不在数据中心
30:42
business. You're not the sort of equinexes of the world. We're not in the data center business
这个行业里。你们不是像Equinix那样的公司。我们严格来说不算在数据中心行业里。
30:47
sort of. What we do work with is so we offer super computing as a service. And what this means
我们所做的是，我们提供超级计算即服务。而这意味着
30:55
is folks come to us and say, I've got x where x is I've got x megawatts here. And maybe sometimes
有人来找我们说，我有x，x就是我有x兆瓦的电力。有时候
31:04
they mean I've got a data center that's like I would like to turn into a GPU cluster.
他们的意思是，我有一个数据中心，我想把它变成GPU集群。
31:10
Like my options are I could sell the GPU or I could sell the data center to somebody or I could
我的选择是，我可以卖GPU，或者我可以把数据中心卖给某人，或者我可以
31:14
become the cloud. And then we go to them and say, great, we will turn you into the cloud. We're the
成为云服务商。然后我们去找他们说，太好了，我们会把你变成云。我们是
31:19
super computing part. You're the data center. Let's partner together and then we will sell and
超算部分，你是数据中心。我们合作，然后我们一起销售，
31:23
all make money. Sometimes it's people who have something that could turn into a data center in
大家一起赚钱。有时候是那些拥有可以变成数据中心的东西的人
31:29
which case we're going to basically project manage that up to the layer. And then we will work
在哪种情况下，我们基本上会把这个项目管理到那一层。然后我们会去做
31:34
with those far we'll work with those folks to turn their powered land or whatever they have
有了那些，我们会和那些人合作，把他们的powered land或者他们拥有的东西转化过来
31:42
into a data center by working with partners. And then we take over at the super computing part.
通过与合作伙伴一起进入数据中心。然后我们在超级计算部分接手。
31:47
This requires us to have some sort of data center expertise in house, but we don't self-perform
这要求我们在内部具备一定的data center专业能力，但我们并不自己执行
31:52
builds. So yeah. Yeah. So it's someone modular because you have different scenarios you need to
构建的。所以是的。是的。所以它是某种模块化的东西，因为你有不同的场景你需要去
31:59
plan. Yeah. Someone wants to describe this as like, like, volunteer for compute. Because it's like
计划。对。有人想把这描述成，像是，像是，自愿贡献算力。因为这就好比
32:07
these each time it's this new rodeo. Yeah. But and we like we deeply partner with folks on both sides.
每一次都是这种新的rodeo。是的。但我们喜欢，我们和双方的人都深度合作。
32:14
Yeah. And primarily the US. Primarily the US, but we are operating around the world. Yeah.
是的。而且主要是在美国。主要是在美国，但我们在全球各地都有运营。是的。
32:21
At some point as SF compute become a misnomer. So when we're deploying clusters all around the
在某个时候，SF compute 变成了一个不太准确的叫法。所以当我们在各地部署集群的时候
32:27
world. Yeah. And when we do that, we're working with people who themselves are trying to be clouds
world。是的。当我们这么做的时候，我们是在和那些本身就想成为云的人合作
32:31
all around the world. They look like sovereign compute players. And our pitch to sovereign compute
在世界各地。他们看起来像是主权算力参与者。而我们对主权算力的推介
32:36
is amazing. Let's help you set up a cloud in your own country. And we will bring the San Francisco
太棒了。我们来帮你在你自己的国家搭建一个云。然后我们会把旧金山
32:42
off-take. We'll bring the San Francisco expertise. We'll set up the clusters. That's our whole business.
off-take。我们会带来 San Francisco 的专业能力。我们会搭建集群。这就是我们的整个业务。
32:48
Yeah. We're going to make it for who you mean your biggest announced partners that you can name.
是的。我们会为你所说的你能点名的最大的已公布合作伙伴来做这个。
32:54
Nobody I can name. Really? Wow. We will have some announcements in week or two that we expect to come
没有我能点名的。真的吗？哇。我们在一两周内会有一些公告，我们预计会
33:02
on. Yeah. All things sort of. Sure. Sure. Good luck. Yeah. No, I mean, I think this is really
出来。是的。各种事情吧。当然。当然。祝好运。是的。不，我是说，我觉得这对每个
33:08
interesting for every country to like who's like investing in sovereign AI to figure this out.
投资主权 AI 的国家来说都很有意思，去搞清楚这件事。
33:14
And it's interesting that like SF is still sort of like that brand name and that is that it's
而且有意思的是，SF 仍然算是那个品牌名字，就是那个，它是
33:20
a very marketable. So I think one of the interesting things is just thinking through this market.
非常有市场。所以我觉得有意思的一点就是去思考这个市场。
33:26
Nvidia is obviously like the 800 pound gorilla in the room. Why didn't they make SF compute?
Nvidia显然是房间里那个800磅的大猩猩。他们为什么不做SF compute？
33:32
And again, the answer is going to be market manipulation but like let's just really get
答案 again 还是市场操纵，但我们就真的深入
33:36
into the incentives of the multiple players. Yeah. So our impression is that Nvidia can't make it but
聊聊多方参与者的激励机制。对。所以我们的印象是Nvidia自己做不了，但
33:42
benefits tremendously from it existing. The reason why they can't make it is because they would
从它的存在中获益巨大。他们做不了的原因是，如果做了，他们就不够
33:46
then not be independent enough for people to trade on it. If you don't trade on it, then you don't
独立，人们就不会在上面交易。如果没人交易，那你就不
33:51
have enough liquidity for lenders to finance against the forward market. Yeah. This is also why
有足够的流动性让贷款方能够基于远期市场进行融资。是的。这也是为什么
33:56
Crusoe and Corrieve and Lambda, why the hyperscalers can't make it, why base 10 and fireworks and
Crusoe和Corrieve和Lambda，为什么超大规模云厂商做不了，为什么base 10和fireworks以及
34:01
modal can't make it because they have too much demand on that side. This is a thing that benefits
modal那边来不了了，因为他们那边需求太大了。这是一件有好处的事
34:07
the entire industry because it makes the financing costs go down. But nobody can make it unless they
整个行业都会受益，因为融资成本降低了。但没人能做成，除非他们
34:13
give up the ability to do all the big inference money and all the big cloud deals. You have to kill
放弃所有大型inference收入和所有大型云交易的能力。你必须杀掉
34:20
your business in order to make this. That's why nobody does it. It's like structurally they're not
你的业务才能做到这一点。这就是为什么没人这么做。就像结构上他们并不是
34:24
set up to do that. Yeah. And you're basically reducing friction and that's all your revenue.
就是为了做这个而设置的。对。你基本上是在减少摩擦，而这就是你所有的收入来源。
34:30
You've announced some milestones. Actually, I forgot to say that upfront. Can you repeat what they are
你们公布了一些里程碑。其实我一开始忘了说这个。你能重复一下是什么吗，
34:36
and just take a victory lap a little bit? Close $700 million of volume. We have a significant number
然后稍微来个胜利巡礼？完成了7亿美元的交易量。我们有很多新的集群
34:46
of new clusters on the way. Yeah. And then the things that I think I'm most excited about is we've
即将上线。对。然后我觉得最让我兴奋的是，我们推出了
34:55
shipped a whole bunch of new features that make SF compute very viable for platform as a service
一大堆新功能，让SF compute对于平台即服务公司来说非常可行。
35:01
company. An example is we allow customers to subdivide their infinite band fabric without having
一个例子是我们允许客户细分他们的infinite band fabric，而无需
35:09
to ask us. So many clouds when you go to buy them, they will give you one infinite band fabric.
来问我们。你去买的时候有那么多云，他们会给你一个无限带宽的fabric。
35:14
And then if you want to redraw the lines of that infinite band fabric, you need to contact them and
然后如果你想重新划定那个无限带宽fabric的线路，你需要联系他们，
35:21
they will then do something for you. We now can do that on the fly for customers, which I think
然后他们会为你做一些事情。我们现在可以为客户实时做到这一点，我觉得
35:26
is really great because it interfaces with the market really well, where you want to be able to
这真的很棒，因为它与市场对接得非常好，你希望能够
35:32
redraw things in order to scale up and down your clusters. Would you describe this as virtualizing
重新划定这些东西，以便对你的集群进行扩容和缩容。你会把这描述为将
35:37
the data center and making it programmable? Yeah. We're virtualizing the data center and making it
数据中心虚拟化并使其可编程吗？是的。我们正在将数据中心虚拟化并使其
35:40
programmable. So all the stuff that you would have gotten if you had owned the hardware, we're trying
可编程的。所以所有那些如果你拥有硬件就能得到的东西，我们正在尝试
35:44
to virtualize that, which I think is exciting to us. It's funny like the last time this happened
去虚拟化那个，我觉得这对我们来说很令人兴奋。有意思的是，就像上次发生这种情况一样
35:50
was like the cloud, which is really cool. And then finally, I think what is your call to action
就像是cloud，这真的很酷。然后最后，我想你的call to action是什么
36:00
for people? Is it engineers? Give me your supply. What is it? If you would like to own a GPU
对人来说？是工程师吗？给我你的供给。是什么？如果你想拥有一块GPU
36:06
cluster and you want the residual value from it, you want to have it for your portfolio, we would
你想从集群中获得残值，你想把它纳入你的投资组合，我们会
36:11
love to work with you in order to build that GPU cluster. We will take over and we will do the
很乐意与你合作，一起搭建那个GPU集群。我们会接手，然后我们会做那个
36:16
supercomputing portions of that and ensure the things go right, we own the SLA. And then if it turns
超级计算的那部分，确保一切正常运行，我们负责SLA。然后如果它变成
36:21
out that that was a bad idea that you need out of that cluster, we can immediately sell that cluster
发现那是个坏主意，你需要脱离那个集群，我们可以立刻卖掉那个集群
36:24
to somebody else. So it doesn't hit your balance sheet. Very simple pitch. Well, thank you so much for
给别人。所以它不会影响你的资产负债表。非常简单的推销。好吧，非常感谢你
36:29
your time. You're doing a great service to the industry and I just am happy to be your friend.
你的时间。你为这个行业做了很大的贡献，我很高兴能成为你的朋友。
36:33
I appreciate it. I think it's it.
我很感激。我觉得就是这样。