谷歌AI基础设施负责人Amin Vahdat谈前沿AI的物理与经济

这是第一个。

That's one.

我们还有正在准备量产的芯片。

We have the chips that we're working on getting into production.

这些芯片已经从制造商那里回来了，我们正在调试，让它们正常工作。

They're already back from the manufacturer, we're debugging them, making them work.

我们还有正在实现阶段、即将流片并送往制造商的芯片。

We have the chips that are in implementation that are about to tape out and go to the manufacturer.

我们还有处于设计阶段的芯片，以及处于概念阶段的芯片。

We have the chips that are in design phase, and then we have the chips that are in concept phase.

所以这真的是一个从量产到最初想法，跨越五六个阶段、很多年的流程。

So it's really this five, six stage, many, many year pipeline from in production to in your mind, etcetera.

与DeepMind在生产上的合作意义重大，因为我们可以一起努力，最大化每瓦特带来的智能或有效吞吐量。

The collaboration with DeepMind in production is significant because we get to work together in maximizing delivered intelligence or delivered goodput per watt.

我们清楚模型里发生的一切，也清楚硬件以及两者之间所有环节的情况。

And we know exactly what's happening in the model, and we know exactly what's happening in the hardware and everything in between.

我们还能在即将流片的芯片上深入合作。

We also get to collaborate deeply on the chips that are actually just about to tape out.

为什么？

Why?

因为我们可以介入，甚至在芯片架构还在开发过程中就做出调整，如果我们是跨公司合作，这几乎是不可能的，至少非常困难。

Because we can intercept, and we can make changes to the chip literally chip architecture in flight, which would be somewhere between hard and impossible to do if we were working across company boundaries.

不是不可能，但如果我们说，天哪，离芯片完成只有几周或几个月了，现在我们要坐下来肩并肩，讨论是否要打乱计划，这会难得多。

Not impossible, but it would be much harder for us to say, oh my gosh, we're a few weeks or a few months away from getting this chip done, and now let's get in the same room shoulder to shoulder and figure out if we should disrupt the program.

有可能，但会更难，我是这么认为的。

It's possible, but harder is what I would say.

然后，当然，对于还在设计阶段的芯片，我们可以一起评估很多架构。

And then, of course, for the chips that are in design, we have many architectures we can evaluate together.

我们可以问DeepMind的同事，你们觉得两三年后模型架构会怎么发展？

We can then ask our colleagues in DeepMind, where do you see model architectures going in two, three years' time?

这是我们可以做的事情的帕累托最优方案。

Here's the pareto of things that we can do.

这是模型架构发展趋势的帕累托图。

Here's the pareto of where model architecture is going.

我们确实拥有深入且强大的模拟基础设施，可以预测工作负载如何映射到不同的硬件架构上。

We have actually deep and significant simulation infrastructures that can predict how the workloads are going to map to different hardware architectures.

迭代非常深入且快速。

Deep, deep and fast iteration.

实际上，团队并不是分开工作的。

It really isn't the teams are working separately.

很多时候，我们在同一栋楼、同一个房间里，每天都有深入的互动。

It's same building, same rooms many times, deep daily interaction.

我每周都会和他们谈好几次，等等。

I talk to whether that's or multiple times a week, etcetera.

所以这真的是工作中非常有趣的一部分。

So it really is a super fun aspect of the work.

那太棒了。

That's awesome.

最终，他们的模型也会帮助芯片设计。

And eventually their models will help with chip design.

我们也在用Gemini来设计未来Gemini的硬件。

We're using Gemini to design hardware for future Geminis as well.

那真是太酷了。

That's very cool.

我想回到刚才说的合作本身，我猜，是不是还是存在一些不匹配的地方，比如……

I want to come back to that On a little bit the collaboration itself, I guess, is there an impedance mismatch still of just...

我觉得，和硬件打交道时，周期时间可能比DeepMind的同事在软件那边要慢一些。

I figure the cycle times are probably just slower when dealing with hardware than your DeepMind folks get to deal with on the software side.

而且我猜你们的规划周期要提前很多。

And I figure your planning cycles are much further ahead.

确实如此。

For sure.

那你们实际有多少调整空间呢？

And so how much room do you actually have to adjust?

是的，这确实是个好问题。

Yeah, it's a really good question.

我们提前两三年、四五年就在规划硬件了。

We are planning hardware two, three, four, five years in advance.

这一点毫无疑问。

No question.

换句话说，我们刚才谈到的TPU 8s和8t，就是我们刚刚发布的。

In other words, we talked about TPU 8s, I and 8t, just now that we announced.

但你可以想象，第九代、第十代，可能还有其他一些产品，已经在概念执行阶段了。

But you can imagine that nine, ten, and maybe some others are in concept execution to something else.

它们可能还要好几年才能出来。

And they might be years out.

如果你在做模型架构，通常不会想那么远。

If you're working with model architecture, by default, you're not going be thinking years out.

但我觉得这也是公司一起成长起来的一个很棒的地方。

But I think this is also the great thing about how the company has grown up together.

换句话说，谷歌研究院、DeepMind 和 Transformer 的发明，所有这些都是在这些重叠的房间里发生的。

And in other words, Google Research and DeepMind and Transformers being invented, all of this was happening in these overlapping rooms as well.

所以，有一整代研究人员已经习惯了能够影响硬件，并且知道硬件是多年周期的。

So in other words, there's a whole generation of researchers who've grown accustomed to being able to influence the hardware and knowing that the hardware is operating on multiyear cycles.

他们也明白，如果他们有一个小调整，能给即将流片的芯片带来 1% 或 0.5% 的提升，他们可能不会来找我们。

And they also know that, look, if they have a small tweak that's going to deliver like 1% or 0.5% or something like that for a chip that's about to tape out, they're probably not going to come to us.

因为他们知道，这不像软件，你只要改个代码，两周后就能上线。

Because they know enough to know that actually it's not like software where you can just do a change list and it's going to ship to production in two weeks' time.

实际上，停止流片是一件大事。

Actually stopping a tape out is a big deal.

但他们也知道，如果他们有一个真正的好机会，我们绝对会一起合作，看看能不能把它加进去。

But they also know if they've got a really good opportunity, yes, we're absolutely going to work together to figure out if we can get it in there.

所以，正如我所说，收益是成倍增长的。

So as I said, it's multiplicative in terms of where the benefits come from.

硬件确实能提升所有东西。

The hardware does lift all the tides.

所以DeepMind团队里有很多人，我们对此非常感激。

And so significant portions of the DeepMind team, we're so grateful for it.

这是一个了不起的团队。

It's an amazing team.

但团队中很大一部分人都在思考如何影响技术路线图。

But significant portions of it are thinking about how do I influence the roadmap?

因为这实际上是一个流水线。

Because it's actually a pipeline.

比如我两年前的想法，现在已经投入生产了。

Like the idea I had two years ago, it's in production now.

它正在帮助谷歌的所有工作负载运行得更快。

And it's helping all the workloads at Google go faster.

这种感觉真好。

That's a good feeling.

完全同意。

Totally.

不过，要预测五年后哪些工作负载最普遍、哪些算法突破会出现，这似乎是个不可能完成的任务。

It seems like an impossible task, though, to predict in five years what workloads will be most common and what algorithmic breakthroughs will have occurred.

这看起来难道不是不可能完成的任务吗？

Doesn't it seem like an impossible task?

我明白你的意思，这看起来确实像是不可能完成的任务，但妙就妙在这里。

I hear you that it would seem like an impossible task, but here's the awesome thing.

实际上，我们正在详细地写这方面的内容，而且写起来特别有意思。

And actually, we're working on writing this in great detail, and it's a lot of fun to do it.

令人惊讶的是，TPU架构在中等细节层面上，不是超高细节，而是中等细节，从TPU v1以来基本没变过。

The stunning thing is that the TPU architecture at a medium level of detail, not at a super high level of detail, at a medium level of detail hasn't really changed since TPU v1.

可以把它看作CPU的指令集架构。

One way to look at it is the instruction set architecture for a CPU.

有加载、存储、加法、减法和分支。

You've got loads, and you've got stores, and you've got adds, and you've got subtracts and branches.

然而，在那段时间里，上面的软件又做了些什么呢？

And yet, what has the software on top of it done over that period of time?

TPU也是同样的道理。

Same thing with TPUs.

我们有一些基本的指令和基本的原语。

We have some fundamental instructions and some fundamental primitives.

当然，我们也做了扩展。

Of course, we've extended it.

并不是说指令集完全没有变化。

It's not like the instruction set hasn't gotten changed at all.

但那些原语，比如把数值计算专门化到非常大的矩阵乘法单元，还有一个处理向量操作和散聚操作的稀疏核心，等等。

But the primitives there in terms of specializing the numerics to very big, very large matrix multiply units, a sparse core that manages vector operations and scatter gather operations, etcetera.

真正定义远程加载存储的，大概有五六个关键点。

There's five or six things that really define a remote load store.

其实还有一个，就是那个。

That's another one, actually.

我们可以通过我们的ICI网络读写远程内存。

We can read and write remote memory through our ICI network.

这些基础一直都在，而且它们已经扩展到了很多代的模型，甚至很多代的深度神经网络算法和模型结构。

The fundamentals have been there, and they've extended just to many generations of models, many generations of even deep neural network algorithms and model structures.

好的，说到工作负载的变化，过去一年左右最大的变化之一，我觉得其实是从今年年初开始的，就是长周期智能体的兴起。

Okay, speaking of shifting workloads, it seems like one of the biggest changes in workloads over the last year or so, I think it really started at beginning of this calendar year, was kind of the rise of the long horizon agent.

而且

And

我猜这种工作负载的形态和过去那种快速的LLM对话很不一样。

I would guess that's a very different shape of workload than kind of the quick turn LLM conversations of years past.

这对数据中心的需求意味着什么呢？

What does that mean in terms of data center needs?

对，我觉得这有两个重要的方面。

Yeah, so I think there's two huge aspects to this.

一个是不再是人类加速交互，随便你怎么说。

One is it's no longer human to, whatever you want to say, accelerate interaction.

换句话说，当你在浏览器或手机上输入提示词时，当然，为了回应你的提示，会有很多工作要做。

In other words, when you are typing at a prompt on a web browser or on your phone or whatever, of course, in response to your prompt, there's going to be a bunch of work that happens.

但是当回复返回时，你得去读它。

But then when the response comes back, you've got to read it.

你得思考一下，然后可能还会有后续问题。

You've got to think about it, and then maybe you have a follow-up.

这会是好几秒的交互时间。

That's going to be multiple seconds of interaction time.

现在对于深度地平线代理，没有人在循环中，这自然限制了请求发送到模型的速度。

Now on Deep Horizon agents, there's no human in the loop that is going to naturally rate limits how quickly requests are going to go to the model.

所以，交互时间从几秒甚至几十秒，现在可能缩短到几毫秒。

So this, what went from seconds, maybe tens of seconds in terms of interaction time, is now going into perhaps milliseconds.

一旦我收到回复，我就能解析它。

As soon as I get a response back, I can parse it.

我可以稍微推理一下，然后确定下一个请求是什么。

I can reason about it perhaps a bit, and I can figure out what my next request is going to be.

所以这是第一个大变化。

So that's big change one.

第二个大变化是，所有这些推理和解析工作很可能都会在CPU上完成。

Big change two is all that reasoning and all that parsing is going to probably happen on a CPU.

然后那个CPU可能还得琢磨一下，在把下一个请求发给模型之前，我还需要去获取哪些其他状态信息？

And that CPU is then going to have to probably think about, okay, what other state do I need to go gather before making my next prompt back into the model?

也就是说，我从这个响应中学到了一些东西。

Like, in other words, I've learned something from this response.

我要再走一步，但实际上我需要去获取一些上下文，可能从本地的DRAM，也可能从另一台CPU上的别人的DRAM，或者从SSD、HDD，甚至其他地方。

I'm going to take another step, but I actually need to go grab some context, maybe from DRAM, local to me, maybe from someone else's DRAM, on another CPU, maybe on SSD, or maybe on HDD, somewhere else.

所以现在还需要进行大量的协调工作。

So now a huge amount of orchestration has to take place as well.

因此，设计确实发生了很大的变化，对加速器计算的需求在上升。

So the design actually is changing pretty significantly where the demand for accelerator compute is going up.

但对CPU、网络和存储，也就是传统数据中心里的CPU等资源的需求，也在急剧增长。

But the demand for CPU and networking and storage, the traditional data center, CPU, etcetera, is also going through the roof.

那这是不是意味着你会在GPU机架旁边放更多的CPU呢？

So does that mean you're putting more CPUs alongside your GPU racks then?

GPU机架？

GPU racks?

对，GPU、TPU机架。

Yeah, GPU, TPU racks.

关键问题又回到了优化和专业化这个话题上。

This is the key question is going back to this question of optimization and specialization.

如果我们开始大量部署CPU机架，并且它们紧挨着GPU和TPU，那就意味着我们实际上无法完全针对密度和网络需求进行专业化，比如说，TPU机架相对于CPU机架的需求。

If we start putting lots of CPU racks and it is a question next to the GPUs and TPUs, that means that actually we can't fully specialize to the density and network requirements, let's say, of a TPU rack relative to a CP rack.

TPU机架会比CPU机架更密集。

A TPU rack is going be more dense than a CP rack.

它可能比CPU机架需要更多的网络连接。

It's probably going need more networking than a CP rack.

所以换句话说，我们的建筑设计方案现在要改变了。

So in other words, now our building design is going to change.

另一个选择是保持统一性，把TPU或GPU都放在同一栋楼里。

Another option is maintain your uniformity, put your TPUs or your GPUs all in one building.

隔壁那栋楼，也许可以放你的CPU。

And the building next door, maybe put your CPUs.

顺便说一句，也许硬盘得放在另一边的另一栋楼里，因为它们又有另一套要求。

Maybe, by the way, the hard drives have to be in another building on the other side because they have yet another set of requirements.

现在你需要在楼与楼之间建立相当高水平的网络连接。

Now you need networking between these buildings at a pretty significant level.

所以一旦你离开一栋楼，网络连接的复杂度就会大幅上升。

So once you leave a building, the networking complexity goes up significantly.

从可靠性和成本的角度来看，延迟可能会增加，但可能还在可接受范围内，不过现在你可能会面临几百微秒的延迟，如果组件之间有排队的话，甚至可能更高。

From a reliability perspective, from a cost perspective, latency goes up probably acceptably, but still now you might go to hundreds of microseconds, potentially more with queuing between the components.

所以这些考虑因素确实会以非常有趣的方式发生变化。

So the considerations do change in pretty interesting ways.

你的工作真不容易。

Your job is hard.

但很有趣。

It's fun.

挺有意思的。

It's fun.

挺有意思的。

It's fun.

网络方面的情况怎么样？

What's happening on the networking side?

我听说谷歌在网络技术方面一直走在前沿，包括光通信。

And I've heard that Google has always been at the forefront of the newest in networking, including optical.

你能谈谈光网络目前的状况吗？

Can you say a word on what is the state of optical networking?

在谷歌，大概十五六年前，我们是最早引入波分复用技术的公司之一，这项技术可以在数据中心内的单根光纤上传输多个信号。

We at Google, this was probably fifteen, sixteen years ago, were among the first to bring essentially what's called wave division multiplexing, where you could put multiple signals on a single fiber within the data center.

实际上，我们所有的机架间通信都利用了这项技术。

And we actually leverage that for all of our communication between racks.

与此同时，我们还引入了与波分复用配套的光路交换技术。

At the same time, we introduced technology along with wave division multiplexing called optical circuit switching.

光路交换的核心就是，跟传统的电分组交换不同，它是在光域里直接传输和转发数据的。

And essentially what optical circuit switching does is, in contrast with traditional electrical packet switching, what it does is it transmits and moves data entirely in the optical domain.

这样做的好处是……

The great thing about that is...

那我来讲讲支撑这项技术背后的原理吧。

And so let me describe the technology that underpins it.

在分组交换里，你会拿到一个带包头的数据包。

In a in a packet switch, you would take a packet that has a header.

你需要在电域里查看它，判断这个包要去哪里。

You would look at it, in the electric domain, figure out where the packet is headed.

包头里可能有个IP地址，告诉你该往哪儿发。

In its header, it might have an IP address that says, okay, where do I head it?

你查一下……

You look up...

好的。

Okay.

对于那个目的地，你查一下表，看该从哪个端口转发出去。

For that destination, you look up in a table, what port do I forward it along?

所以基本上，每秒有数十亿、数百亿甚至数万亿的数据包以超高速涌入，你都得把它们转发出去。

So, basically, billions, billions and billions, perhaps trillions of packets coming through at super high speed per second, you're forwarding them along.

光路交换的意思是，我不在电域里碰这些比特。

Optical circuit switching says, I'm not touching these bits in the electrical domain.

我要做的是，对于每个输入端口，确定光应该发送到哪个输出端口。

What I'm going to do is I'm going to figure out for an input port which output port to send the light to.

明白吗？

Okay?

现在我有……

And now I have...

实现这个有多种方法。

There's multiple ways to do this.

我们最初采用的那种叫MEMS交换机，也就是微电机系统，它们基本上控制三维空间中的镜子。

The one that we started with is called MEMS switches, microelectrical motors, that basically control mirrors in three d.

所以我们现在可以通过编程来操作一个盒子，这个盒子可能有……

So we can now programmatically take a box that might have...

你选128个端口、256个端口，或者类似的数目，我们可以为每根进入的光纤配置输入端口，把它映射到输出端口，然后调整镜子的旋转角度，让光直接照在这些镜子上，反射到正确的输出端口。

You pick 128 ports, 256 ports, some number like that, and we can configure every input port for fiber that comes into it, map it to an output port, and then change the rotation of mirrors where the light literally shines down on these mirrors and gets reflected to the right output port.

最初，我们这样做的原因有两个，一个是为了在机架组之间建立局部性。

Initially, the reason we did this twofold was to create locality between groups of racks.

假设我有一个计算集群和一个存储集群，它们都在支持——还是拿搜索举例——某个功能。

Let's say I had a compute cluster and a storage cluster, and both of them were in support of, again, making it up search.

我们知道这两个集群之间会有很多通信，所以我们会配置镜子，在这两个集群之间创建纯光学的捷径，也就是机架集群之间的捷径。

We knew that these two clusters would talk to each other a lot, so we would configure the mirrors to create shortcuts purely optically between those two clusters, clusters of racks.

这是第一个原因。

That was reason number one.

第二个原因是，我们希望能够在不需要移动任何光纤的情况下，扩展或收缩网络。

Reason number two was we wanted to be able to expand the network and contract the network without actually moving any fiber.

具体细节我就不多说了，我可以在白板上画出来。

And without working into the details, I could draw this on a board.

光电路交换机可以让你重新配置网络的骨干，来扩展或缩小网络，而不需要人工干预，只需要一个控制器来管理这一切。

The optical circuit switch would allow you to actually reconfigure the spine of the network to expand it or to shrink it without a human being, having to do anything other than be literally a controller that would manage that.

现在快进到TPU。

Now fast forward to TPUs.

TPU采用环形拓扑结构，将所有TPU直接连接在一起。

A TPU has a torus topology that connects all the TPUs to one another directly.

我之前提到过吞吐量和有效吞吐量。

I talked about throughput and goodput earlier.

我们可以做的一件事是，如果某个TPU机架出现故障，我们可以用另一个TPU机架替换它，而无需移动任何光纤。

One of the things that we can do is if we have a TPU rack that fails, we can replace it with another TPU rack without moving any fiber.

同样，我得比划一下或者画个白板来解释。

Again, I'd have to wave my hands or draw out a whiteboard.

但基本上，我们可以说随时都有一个备用机架可用。

But, essentially, we can say we have a spare rack available at all times.

当一个机架出现故障时，我们就把光路重定向到那个新机架。

And when a rack fails, we redirect the light to that new rack.

而且这可以在几毫秒内完成。

And that can be done in milliseconds.

那为什么还要用光纤呢？

Why do you have the fiber at all then?

而不是完全用自由空间。

As opposed to fully free space.

这是个很好的问题。

It's a very good question.

衰减损耗和带宽会大幅下降。

Attenuation loss and the bandwidth would drop dramatically.

而且，在一个非常大的建筑范围内，如果没有光纤把所有东西在三维空间里连接起来，要精确对准所有设备会非常困难，甚至可能不可能。

Furthermore, across the range of a very large building, aiming everything without the benefit of fiber to connect it all together in three d would be challenging to possibly impossible.

我们讨论过这个。

We've talked about it.

我们确实讨论过，而且有一些非常有趣的讨论。

We've talked about it, actually, and there have been some really interesting discussions.

是的，大部分是在光纤里，但当它们到达光电路交换机时，光基本上就直接照射在这些微小的芯片上。

Yes, it's mostly in fiber, but then when they hit the optical circuit switch, essentially the light then literally shines down on these tiny chips.

所以这是网络发展的一个大方向，但还有很多其他方面。

So that's one big direction of networking, but there's a lot.

我觉得在数据中心，网络的能力和需求可以说正在飞速增长。

I would say networking is exploding in terms of its capabilities and need, frankly, in the data center.

太酷了。

So cool.

我们刚才一直在聊

We're have an entire conversation on

是的，这确实非常非常酷。

Yes, it is very, very cool.

甚至回到那幅不可言喻的画面，都是那些被切断的电缆。没错。

Even going back to the picture from ineffable, it was all the cables that cut out Exactly.

这些

Of these

对，那些电缆，最后有些会接到我们数据中心的电路交换机上。

Yeah, and those cables, eventually, some of them wind up at optical circuit switches in our data centers.

有道理。

Makes sense.

好，我想聊聊电力这个话题。

Okay, I want to flip to talking about power.

你一直在讲每瓦特的有效吞吐量，还有其他每瓦特的单位。

You've been talking about goodput per watt and other units per watt.

这让我觉得，每瓦特意味着电力在某种程度上是限制因素，或者说是稀缺资源，或者说是成本高昂的因素。

Makes me think per watt means power is kind of the binding constraints or the scarce constraints or the expensive constraints in some way.

所以经常有人问我，我们面临的最大限制是什么？

So I get asked this question, what is the biggest constraint that we face?

实际上，我们并没有一个单一的最大限制。

And the reality is there is no single biggest constraint that we face.

它们全都是限制。

They're all constraints.

它们都非常难。

They're all super hard.

而且它们还在不断变化。

And they shift continuously.

但如果非要回答根本问题，我会说电力是我们面临的最根本的约束。

But if I had to answer fundamentally, I would say that power is the single most fundamental constraint that we face.

其他问题似乎我们都有办法解决，只是时间问题。

Everything else seems like we know how to solve them, and it's a question of solving them over some period of time.

电力，我觉得你刚才说的那个说法特别好。

Power, I think the way that you put it is really nice.

这是一个长期的、有约束性的问题，我们目前还没有解决办法。我是说，也许核能会带来丰富的清洁能源，那肯定会解决很多问题。

It's a binding long term issue that we do not have a I mean, Perhaps nuclear is going to abundant clean energy, which would solve a lot of problems for sure.

但什么时候能实现，以及能否大规模实现，还是未知数。

When that happens and when it happens at scale, still unknown.

那实际操作中是怎么运作的呢？

And so how does it work in practice?

你要新建一个数据中心。

You're standing up a new data center.

它需要一千兆瓦的电力。

It needs a gigawatt of power.

我想你不可能直接打电话给太平洋煤气电力公司说，嘿，请送一千兆瓦的电过来。

I imagine you can't just call PG and E and say, Hey, please send a gigawatt of power.

那么，电力供应实际上是怎么操作的呢？

So what does provisioning power actually look like?

你是不是真的需要一路垂直整合，连自己的涡轮机都自己做？

And are you having to actually vertically integrate all the way down doing your own turbines?

你怎么解决电力瓶颈的问题？

How do you solve the power bottleneck?

这又是一个大问题，也是重要的问题。

Again, it's a big question, important question.

在谷歌，我们偏好的模式始终是保持连接，保持良好的连接。

Our preferred model at Google is always to be connected, to be good connected.

所以这就相当于给你的……

So it's the equivalent of calling your...

哦，所以你不能直接给太平洋煤电打电话。

Oh, so you can't just call PG and E.

对，没错。

Exactly.

你得非常客气地联系你数据中心所在地的那家公用事业公司。

You call very politely your favorite utility, wherever it is that your data center is.

当然，你得提前很多年通知他们。

And of course, you give them many, many years of notice.

换句话说，如果我们要的是千兆瓦级别的电力，那可不是你说“嘿，我明天就需要一千兆瓦”就能解决的。

In So other words, if we're talking about a gigawatt scale, it's not something that you can say, hey, I need a gigawatt tomorrow.

你能开始给我计费吗？

Can you start billing me?

这是我们需要一起规划的事情。

It's something that we co plan together.

对我们来说，这也是我们非常重视的一点，就是确保在和这些公用事业公司合作时，建设这些基础设施的成本——这个话题我们也能聊很久——我们都会承担，因为按照现在的计费方式，如果公用事业公司为我们扩容，比如说，其他人的费率可能会上涨。

For us, it's something that we also take very seriously from the perspective of ensuring that when we work with these utilities, the costs of putting that infrastructure in place, we could get into long conversations just on this topic as well, that we cover those costs because with the way that billing works, it actually could be that by the act of the utility building up capacity for us, let's say, other people's rates could go up.

理论上，我们确保的是，比如说需要新建或升级的输电线路、额外的公用事业基站等等，这些费用我们也会支付。

In theory, what we ensure is that actually, let's say the transmission lines that have to be built, upgraded, additional utility base stations, etcetera, that we pay for those as well.

所以这是一个很长的规划过程。

So it is a long planning process.

完全有可能出现这种情况，比如说，我们需要一千兆瓦，我随便说个日期，2028年。

It can absolutely be the case that, let's say, that we need a gigawatt in I'll make up a date 2028.

而公用事业公司能在2029年给我们一千兆瓦。

And the utility can get us a gigawatt in 2029.

他们可能在2028年给我们，比如说，七百兆瓦。

That can get us, let's say, 700 megawatts in 2028.

那我们就得考虑，剩下的三百兆瓦怎么解决？

So now we might be left with a question of, how do we cover those 300 megawatts?

嗯，一个答案就是等。

Well, one answer is just wait.

另一个答案是，好吧，我们能不能自己想办法发一部分电？

Another answer is to say, Okay, well, would we figure out how we generate some subset of that power ourselves?

也许是用太阳能电池板，或者用电池做备用，等等。

And maybe that would be with solar cells or with batteries as a backup, etcetera.

会不会有其他来源？

Would it be for other sources?

然后我们又和电力公司合作。

So then we again work with the utility.

所以也可能是组合方案，我们保留一些本地发电能力，即使电力公司完全上线，比如达到吉瓦规模，我们也能把电送回电网，这样他们需要的时候我们也能提供。

So it might also be a combination where we might maintain some power generation local and have the capability even when the utility comes online fully at, let's say, the gigawatt scale, where we could actually provide power back to the grid, so having that when they need it as well.

换句话说，可以这样看：一年里可能有两周，也许是最热的两周，居民用电需求特别大。

So in other words, one way to look at it is there might be the two weeks of the year, maybe it's the hottest two weeks of the year, where there's huge amount of residential demand.

如果我们有一些本地发电，我们也能把电送回电网。

If we have some local generation, we can then provide that back to the grid as well.

所以对我们来说，这真的是和电力公司多年合作的过程。

So it really is working over multiple years with the utilities for us.

为什么你倾向于这样做，而不是去……

Why is your preference to do that versus to of go...

就是跟公用事业公司合作，而不是自己垂直整合？

Like to work with the utilities as opposed to kind of vertically integrate yourselves?

对，主要原因就是灵活性和双方都能受益。

Yeah, so the main reason is flexibility and uplift on both sides.

一种看法是统计复用，或者你也可以说是大数定律。

So one way to look at it is statistical multiplexing, or if you want, the law of large numbers.

如果我们想要一千兆瓦的电力，假设我们希望可靠性达到99.99%以上，那可能就得建两千兆瓦的发电能力。

If we need to have a gigawatt of power, let's say we wanted to have that with 99.99% plus reliability, Probably means we have to build two gigawatts of power.

要达到99.99或99.999的可靠性，就必须有一加一的冗余。

At that level of 99.99 or 99.999, you have to have one plus one redundancy.

那成本就高了。

That gets expensive.

而且还要确保这最好是清洁能源，最好就在我们数据中心旁边，这也有挑战性。

And then making sure that that's going to be ideally a clean energy source, probably right next to our data center, that can also be challenging.

现在我们和数据中心合作，我们自己可能也能提供一部分电力，当电网不需要那么多电的时候，我们可以把电给电网，需要的时候再从电网取电。

Now, we partner with the data center, again, maybe we have some amount of that that we can bring ourselves, that we can give to the grid when they need it less, that we can take their power.

所以换句话说，通过利用这种统计复用，在更大的基础上，其实大家都能受益。

So in other words, by leveraging that statistical multiplexing over a much larger base, actually everybody wins.

我们受益，电网受益，居民也受益，等等。

We win, the grid wins, residences win, etcetera.

我们更倾向于这种方式。

We prefer that.

在极少数情况下，我们也会在电表后面做，比如自己发电自己用。

And in rare, rare cases, we will do it, let's say, behind the meter.

但即便如此，我们也是假设未来会和电力公司合作，可能一年以后。

But even then, we're doing it under the assumption that we're going work with the utility where it might be a year later.

我们想要连接到电网上。

We want to be connected to the grid.

再说一次，这对我们有利，对电网也有利。

Again, it's an uplift for us, it's an uplift for the grid.

有道理。

Makes sense.

你们是怎么决定数据中心建多大的？

How do you decide how big to make a data center?

对，这其实是一门艺术，而且也是争论的焦点。

Yeah, that's an art, and it's a source of significant debate.

我们曾经有过这样的争论。

At some point, we had a debate.

我记得十多年前，甚至十五年前，就有过一场大争论：要不要把所有东西都放在一个数据中心里？

I remember even ten plus years ago, fifteen years ago, big debate at Should we just put everything into one data center?

那时候，一个数据中心就要达到千兆瓦级别，那可是……

Back then, it was going to be a gigawatt, which was

那时候简单多了。

Simpler times.

对，时代不同了，那是十到十五年前的事了。

Yeah, different times, like ten, fifteen years ago.

我们要建一个千兆瓦的数据中心。

We're going have a gigawatt data center.

所以一个明显的担忧就是单点故障。

So one obvious concern there is single point of failure.

如果你从三十年的角度来看，那是个大问题。

If you're looking at things from a thirty year perspective, that's a huge concern.

但另一方面，如果谈的是今天的训练负载，越大越好。

On the other hand, if you're talking about training workloads today, bigger is better.

因为从网络的角度来看，实际上你希望把东西集中在尽可能小的距离内。

Because from a networking perspective, actually, you want to have things concentrated in as small a distance as possible.

但话说回来，有两个问题：单点故障，还有现在的电力供应。

But then again, two issues, single point of failure, but then also now power availability.

虽然千兆瓦在十年前或十五年前是巨大的，但可以想象，但在全国或全球任何地方，都不可能把谷歌的总需求建在一个地方。

While a gigawatt was huge ten or fifteen years ago, but imaginable, there's no way anywhere in the country or the world we're going to get the total requirements of Google built in one place.

那根本不可能。

It's just not going to be possible.

那现在最优的规模是多大呢？

So now what's the optimum size?

同样，我们有模型来模拟这些，还有模拟器等等。

Again, we have models for this, simulators, etcetera.

这也取决于……但你看，有些地方我们会部署在网络边缘，或者会在某个国家内部。

It also depends on But look, in some places, we're going to be at the edge of the network or we're going to be in country.

那可能是几十兆瓦。

That might be tens of megawatts.

我们实际上和互联网服务提供商合作。

We actually partner with ISPs.

那可能是伊拉克。

That might be Iraq.

真的，可能就是伊拉克。

Literally, it could be Iraq.

这是一个训练集群。

This is a training cluster.

可能那会更接近一个吉瓦。

Maybe that's going to be closer to a gigawatt.

其他一些站点可能是几百兆瓦，等等。

Some other sites might be hundreds of megawatts, etcetera.

有意思。

Interesting.

我很想了解你是怎么看待投资组合生命周期管理的。

I'd love to understand how you think about the portfolio lifecycle management.

所以我猜，最大的、最新的集群是用来训练最新的前沿模型，然后你把其他的东西回收利用，用来跑推理。这个框架合理吗？

So I guess my guess would be that the biggest, newest clusters are used for training the latest frontier model, and then you kind of recycle the other stuff and run inference on Is that a fair framework?

你们是不是也在搭建专门用于推理的集群？

Are you also standing up inference specific clusters?

这一切是怎么运作的？

How does that all work?

你考虑的这个框架非常合理，我觉得很有道理。

Very reasonable framework that you're thinking about, and I think it makes a lot of sense.

但我要说的是，推理需求非常大，我们不能仅仅依赖那些不再完全用于训练的老训练集群来支撑推理。

But I would say that the demand for inference is such that we can't just rely on older training clusters that are no longer being fully used just for training as the basis for inference.

而且，你想啊，某一年我们可能会再次集中到少数几个大型站点或训练集群。

And also, if you think about it, we might centralize again in a particular year into a small number of large sites or training clusters.

保持它们之间的网络距离短是有好处的。

And there is benefits to keeping the network distance between them small.

所以现在，无论它们位于世界何处，很可能在同一大陆，甚至可能在同一年份位于同一大陆的同一区域。

So now wherever they're located in the world, they're probably going to be, let's say, on the same continent, or they might be even in the same portion of a continent in a particular year.

这样一来，其他大陆可能就没有足够的服务能力了，等等。

So now you might be left with not enough serving capacity on the other continents, etcetera.

所以，实际上，我们不得不在世界其他地方专门建设推理设施。

So then, actually, then have to go build this specialized inference elsewhere across the world.

所以我觉得你的直觉很准，但不完全对。

So I think your intuition is spot on, but not wholly.

换句话说，确实必须这样：好，这里是训练集群。

In other words, it really does have to be, Okay, here are the training clusters.

是的，可能几年后它们会被用于服务。

Yes, probably they're going be used for serving in some number of years.

但我们也得同时建设服务集群。

But then we're also having to build the serving clusters as well.

那你的服务集群和训练集群不一样吗？

And your serving clusters, are they different than the training clusters?

它们更小吗？

Are they smaller?

每兆瓦的成本更低吗？

Are they cheaper per megawatt?

其实不一定更便宜，因为服务需要更紧密地整合计算、网络和存储。

Not necessarily cheaper, actually, because for serving, there is more need to co locate compute, networking, and storage.

这样一来，我们就没法做到专业化。

And so then we get to the inability to specialize.

训练时，你可以进行高密度、均匀的部署，但服务时，你得混合搭配存储、计算和加速器。

So for training, you actually have this big density uniform deployment, etcetera, whereas for serving, you're going to have to have the mix of storage, compute, and accelerators.

而且，这一点其实很有意思，我们还可以再深入聊聊。

Furthermore, it's a very interesting aspect of this that we could go into more detail on as well.

但对于推理服务来说，你其实不希望把所有东西都集中在一个地方。

But for serving, you actually don't want to have too much in one place.

你希望从全球各地来提供你的工作负载服务。

You want to be serving your workloads from all over the planet.

但现在我们有了各自的模型端点，而且模型之间实际上可能存在差异。

But now we wind up having individual model endpoints, and there actually can be variations of models.

所以现在我们必须把这些模型分布到全球，同时还要考虑本地化的问题。

So now we have to spread these models out across the world, also accounting for locality.

所以它们会更小一些。

So they will be smaller.

在推理端，它们可能不会那么垂直整合。

They'll actually be less vertically integrated probably on the inference side.

非常有意思。

Super interesting.

所以如果你五年前建了一个数据中心，那时候最先进的加速器和现在造的相比，效率差远了。

So if you made a data center five years ago, state of the art accelerator five years ago was very different, vastly less efficient than the ones that are being made today.

你们真的会回去把那些旧数据中心里的芯片换掉吗？

Do you actually go back and swap out the chips in those old data centers?

而且我知道这跟一个一直在讨论的话题有关，就是芯片的实际使用寿命到底有多长。

And I know this kind of relates to there's been an ongoing debate, I think, of what is the actual useful life of a chip?

我公开说过这个观点，在我比较……

So I've been on record of saying this in one of my more.

我没想到这个话题会引起这么大的关注。

I was surprised by how much pickup this had.

我说过我们用了七八年的TPU，利用率还是百分之百。

I said that our seven and eight year old TPUs are still at 100% utilization.

哇。

Wow.

我想我见过这个。

Think I saw this.

我没

I didn't

想到这居然成了个话题。

know if this was mistake.

我本来没想把它当成一个声明，但显然它成了一个很有分量的声明。

I didn't mean for it to be statement, but it was a meaningful statement apparently.

所以我们的老款TPU，它们的使用率其实相当高。

So our older TPUs, but our older TPUs are seeing significant utilization.

我们确实会更换它们。

We do replace them.

说到底，问题在于，折旧年限大概是六年。

In the end, there's a question of, okay, depreciation lifetime is approximately six years.

所以一旦它们完全折旧，再加上新一代芯片的能效优势等等，更换和升级确实是合理的。

So once they're fully depreciated and given the power efficiency of newer generations, etcetera, it does make sense to replace them and upgrade them.

我们确实会更换芯片，但真正换的是整个系统。

We do replace the chips, but it's really the systems.

所以换句话说，我们是以pod为单位来考虑的。

So in other words, we think in terms of pods.

比如说，一个TPU 8t pod可能有9600块芯片。

So a TPU, let's say, 8t pod might be 9,600 chips.

它可能有140多个机架，大概是152个机架，等等。

And it might be 140 something, 152 racks, etcetera.

然后我们会说，好，我们要把那个pod撤掉，换成一个新的，比如TPU 12、13或14的pod。

So then we would say, Okay, we're going to pull out that pod, and we're going to replace it with a, whatever it is, TPU 12 or 13 or 14 pod.

可能不会完全匹配。

It might not be a perfect fit.

也就是说，新pod的占地面积可能跟旧pod留下的空间不完全吻合。

Then we have to in other words, the new pod's footprint might not be a perfect fit for the old pod's vacancy.

那我们就得想办法处理这个差异。

We then just have to account for that.

我们得考虑怎么进行改造。

And we have to figure out how we would retrofit.

这个我们没法提前规划，因为不知道那么多代TPU会是什么样。

We can't plan that because we don't know what that many generations of TPUs are going to be.

所以这又有点像一门艺术，得想办法——而且这活儿其实挺难的——怎么尽快把旧的拆掉，换上新的TPU。

So then it's, again, a bit of an art to figure out how we would, and it's actually hard work, figure out how we would decom and then replace as quickly as possible with new TPUs.

你公开写过关于开放标准的事。

You've written publicly about open standards.

你能简单说说吗？

Can you say a word on that?

我觉得这就是我们之前聊到的互操作性问题。

So I think that's the we talked about this interoperability question.

对我们来说，虽然我们支持垂直整合，也让你能榨取尽可能多的性能，但更重要的是，我们不想强制锁定，也不想在系统端到端的运作方式上搞一个封闭的花园。

And so for us, while we support vertical integration and we make it possible to extract just as much performance as you want, for us, it's really important to not force lock in and not force a sort of walled garden in terms of how the system works end to end.

我举个例子。

So I'll give one example.

在谷歌，我们开发了一个叫JAX的模型开发框架。

At Google, we've developed a model development framework called JAX.

我们非常喜欢它。

We like it a lot.

我们认为它真的非常非常好。

We think it's really, really good.

而且我们在内部广泛使用它。

And we use it extensively internally.

我们的很多客户喜欢用PyTorch。

Many of our customers like PyTorch.

所以我们可以说，嘿，如果你想在TPU上运行，你就得用JAX。

And so one thing we could say is, hey, if you want to run on TPUs, you've to use JAX.

它是最好的。

It's the best.

我有点开玩笑，但也许是真的，也许不是。

I'm being facetious a little bit, but maybe it is, maybe it isn't.

我们认为它是最好的，而且这是你唯一的选择，因为它太棒了。

We think it's the best, and that's your only choice because it's so great.

或者我们可以说，看，如果你喜欢JAX，我们也喜欢JAX。

Or we could say, look, if you like JAKs, we like JAX.

如果你喜欢JAX，你可以用它。

If you like JAX, you can use it.

但如果你喜欢PyTorch，我们有Torch TPU，你的未修改模型等等都可以在上面运行。

But if you like PyTorch, we have Torch TPU where your unmodified models, etcetera, can run.

实际上，我们在历史上已经看到过很多这样的例子。

And we've seen this throughout history, actually, with many, many examples.

我举过IP的例子。

I've used the example of IP.

为什么互联网协议会胜出？

Why did the internet protocol win?

当时其实有很多与IP竞争的协议。

There were actually many competing protocols to IP.

这是上世纪70年代和80年代初的古老历史了。

This is ancient history in the '70s and early '80s.

为什么IP会赢呢？

Why did IP win?

因为它是一个开放标准，就像沙漏的细腰，上面什么软件都能跑，下面什么硬件都能接。

It was because it was open standard, it was this narrow waist to the hourglass where anything could run on top in terms of software and anything could run underneath it in terms of hardware.

所以就是开放标准，互操作性强。

So open standard, interoperable.

任何人只要用IP，插到路由器端口上，就能成为互联网的一部分。

Anyone who brought IP could plug into a router port and become part of the internet.

这太棒了。

It was beautiful.

正是这样，互联网才能爆炸式增长，遍布全球。

And that's what allowed the internet to explode in its growth and reach across the world.

所以我们真的相信这些开放标准，尤其是在接入点上。

So we really do believe in those open standards in terms of plug in points.

现在，如果你想接入一些高度专业化的东西，你也可以做到。

Now, if you want to plug in something highly specialized, you can.

如果你觉得你有更好的东西可以接入我们的框架，那完全可以。

If think that you've got a better thing to plug into our framework, you absolutely can.

但我们真的想支持开放标准，最好周围也有开源的东西。

But we want to really support open standards, ideally open source around it as well.

这个建设太庞大、太重要了，不能是任何一家厂商的封闭专有标准。

It's too big and too important of a build out to be any one vendor's closed proprietary standard.

对，我们确实这么认为。

Yeah, we really believe that.

而且它必须是一个可选择的标准。

And it's got to be one of choice.

这也是为什么，比如说，我们完全支持TPU、GPU和其他加速器等。

And that's also why, for example, we fully support and have TPUs, GPUs, other accelerators, etcetera.

AI给你们团队的日常生活带来了什么变化？

How has day to day life for your team changed with AI?

它在哪些方面对你们的工作影响最大？

Where is it changing your function the most?

所以我觉得，最简单的回答就是软件工程方面，这个外界已经有很多报道了。

So I think the, whatever you want to say, the easiest answer is on the software engineering side, where it's been documented externally quite a bit.

所以我觉得在谷歌，在我的团队里，我们已经在软件开发能力上充分利用了AI，而且说实话，在测试、发布，甚至设计辅助等方面也都有用到。

So I think that we are, at Google and my team, using it to great effect in terms of our software development capabilities, but also, frankly, tests, rollouts, even helping with design, etcetera.

但在硬件方面，虽然外界报道可能少一些，变化也非常大。

But there have been, on the hardware side, which has maybe gotten less external coverage, also significant change.

换句话说，我的硬件工程师们——我今天早些时候刚看过数据——他们也在大量使用AI，如果你愿意的话，可以用token数量来衡量。

In other words, my hardware engineers I was just looking at the numbers earlier today, in fact are using AI, using whatever the token counts, if you want.

虽然不是最好的指标，但至少是个指标。

Not the best metric, but it's still a metric.

我们的硬件工程师使用AI的频率和软件工程师一样高。

Our hardware engineers are using AI as much as the software engineers are.

而且生产力确实提升了。

And productivity has gone up.

从设计启动到流片的时间在缩短。

Time from design kickoff to tape out is shrinking.

启动时间也在缩短。

Time for bring up is shrinking.

所以换句话说，生产力正在显著提升。

So in other words, productivity is going up significantly.

但还有其他可能不太被预料到的变化，比如我们设计数据中心的方式已经发生了重大改变。

But then other maybe less expected changes, how we do data center design has changed significantly.

也就是说，在评估时，你提到过，我们是建一个千兆瓦的大楼，还是建一个100兆瓦的园区，或者200兆瓦的园区？

In other words, and how we evaluate you mentioned, hey, do we do a gigawatt building or 100 megawatt campus or a 200 megawatt campus?

过去，这些决策过程非常细致，主要靠电子表格和人工驱动。

In the past, this would be very detailed, very spreadsheet driven, human driven processes.

现在虽然在一定程度上还是这样，但已经有很多AI参与其中，真正简化了规划和开发流程。

And it still is to some extent, but there's a lot of AI now involved that really streamlines the process in terms of planning and development as well.

有意思。

Interesting.

是推理模型吗？

Is it reasoning models?

我还不至于说已经到了推理模型那一步。

I wouldn't say quite yet reasoning models.

它不是在取代人的判断，而是让把必要的信息集中到一起变得容易得多，基本上就是把正确的信息放到做决策的人面前。

It's replacing human judgment, but it's making it much easier to bring the necessary information together in one place and basically put the right information in front of the humans making the decisions.

好的，最后我想用两个有趣的问题来收尾。

Okay, I'm going to bring us home with two out there fun questions.

第一个问题，轨道数据中心。

Question number one, orbital data centers.

我看到谷歌在认真考虑这个，你想让我评论一下。

I've seen that Google is thinking about this quite seriously, and you want me to comment on that.

但人们认真计算轨道计算的成本，是不是意味着地球上真的有很严重的限制？

But does the fact that people are seriously running the numbers on orbital computes mean that there are serious binding constraints on Earth?

你对轨道数据中心怎么看？

And what do you think of orbital data centers?

是的，这是个令人兴奋的方向。

Yeah, it's an exciting direction.

我们确实在推进这个项目，而且我们毫不开玩笑地把它称为‘登月计划’，这是我们非常期待投入的大项目之一。

We are actually pursuing it, and we've referred to it with no tongue in cheek as a moonshot, as one of our big efforts that we're excited about investing in.

这又回到了我们之前聊到的，你提得很对，从根本限制来看，能源和能源生产是一个非常非常关键的挑战。

It goes back to the earlier part of the conversation where you raised, I think correctly, that in terms of fundamental constraints, energy and energy production is a key, key challenge.

所以从根本上看，在太空中，因为大气衰减等因素，你能获得大约多40%的电力。

And so really, line is, from a fundamentals perspective, in space, you have something like 40% more power available because of lack of attenuation in the atmosphere, etcetera.

换句话说，就是太阳能容量更大，是地球的1.4倍。

In other words, there's just more solar capacity, 1.4 x.

这是其中一方面。

So that's one part of it.

但在太阳同步轨道上，太阳能电池板能接收到90%到100%的阳光照射，而在地面上只有28%到30%，可能最多35%。

But in a, sun synchronous orbit, you have 90% to 100% coverage of sunlight on your solar cells relative to 2830%, maybe 35% on land.

所以，简单说，就是巨大的能量，当然来自太阳。

So in other words, just this enormous amount of power, obviously the sun.

你把1.4倍和每天日照时间多3到4倍结合起来。

You take the 1.4x and you take the 3x to 4x in terms of number of hours per day.

基本上就可以把电池排除在外了。

You remove batteries more or less from the equation.

现在你就有潜力做出真正能……的东西。

And now you have the potential for something that can really...

而且当然，还是零碳的。

And of course, carbon free.

好处很多，挑战也很多。

Lots of benefits, lots of challenges.

挑战特别特别多。

Lots and lots of challenges.

所以换句话说，好，那冷却怎么办？

So in other words, okay, now what about cooling?

你可能天真地以为在太空里冷却更容易，但实际上更难。

Naively, you might think cooling in space is easier, but it's actually harder.

可靠性。

Reliability.

我们之前聊过这些东西有时候会出故障。

We talked about how these things fail sometimes.

维修会变得更难，但不是不可能，只是在太空里更难了。

Repairs becomes harder, not impossible, but it becomes harder in space.

我们能不能

Can we

把机器人拉上来，用

pull the robot up with

那件宇航服？

the costume?

对，就是这样。

Exactly.

所以也许这个模式是个有前景的方向。

So perhaps this model is a promising direction.

冗余设计可能会帮上大忙。

Redundancy will probably be your friend.

关于自由空间光学的那个前瞻性问题，现在真的要变成现实了。

Prescient question on free space optics is now going to become reality.

我们可能不会在这些组件之间拉光纤，所以实际上会采用自由空间。

We're probably not going to be stringing fiber between these components, so it will actually be free space.

我们会让激光对准接收器，并实时校准。

And we're going to have the lasers pointing at the receivers and calibrating in real time.

这里没有障碍，没有根本性的障碍。

There are no showstoppers here, no fundamental showstoppers.

最后一个问题。

Last question.

我脑子里一直有个画面，就是你建造的那台难以形容的巨型超级计算机。

I've had this image of the ineffable big supercomputer that you built them in my head this whole time.

所以问题是这样的。

And so here's the question.

十年后，最前沿的超级计算机长什么样？

Ten years from now, what will the most frontier supercomputer look like?

天哪。

Oh my gosh.

对，十年正好处在一个临界点上。说实话，可能已经有人在考虑2036年的计算机会是什么样子了。

Yeah, so ten years is right at that edge where it's, I would say, in all sincerity, probably some people are thinking about what the 2036 computer looks like.

可能有一小部分人会想得那么远。

We might have a few thinking that far ahead.

但哇，那里的不确定性范围实在太大了。

But wow, the cone of uncertainty there is way, way too wide.

换句话说，如果你看看趋势，集成度将会非常高。

In other words, if you look at the trends, the level of integration is going to be tremendous.

我们看到了为Ineffable组装的那些漂亮的光纤和VR200机架。

So we saw the beautiful fiber and the VR200 racks that we put together for ineffable.

我的猜测是，它会更加集成，光纤会更少——我不会说2036年没有光纤，但我觉得从机架的角度来看，机架看起来会更加紧密集成。

My guess is it's going to be more integrated, less fiber to I won't say no fiber in 2036, but I think that probably from a rack perspective, the rack will look much more tightly integrated.

光纤的数量可能会减少，也许机架外面只有一小束光纤。

The amount of fiber there'll be maybe a small fiber bundle coming out of the rack.

所以换句话说，从模块化制造的角度来看，这些机架到2036年很可能会在中央工厂生产。

And then, so in other words, one view you might have from a modular manufacturing perspective is these racks, likely in 2036, are going to be manufactured centrally.

不管是72个、144个、288个，还是更多，576个或1152个GPU，随便你选个喜欢的倍数，因为我们说的是ineffable的情况，但TPU也是一样的。

And so whether they have 72 or 144 or two eighty eight or probably more, five seventy six or eleven fifty two, pick your favorite multiple of GPUs since we're talking about the ineffable case, but then be the same for the TPUs.

深度集成到一个机架里。

Deeply integrated into a rack.

一个机架会不会有好几兆瓦的功率？

Would a rack be multiple megawatts?

我们不知道具体怎么实现，但2036年，我们可以尽情想象。

We don't know exactly how, but 2036, we get to imagine things.

单个机架可能就有好几兆瓦。

Could be multiple megawatts in a single rack.

然后你接上水、电和光纤。

And now you bring water, and you bring power, and you bring fiber.

把机架推进去，插上这三样东西，就可以直接开工了。

You wheel that rack in, you plug those three things in, and you're off to the races as well.

好的，那它不会是一个巨大的外星球体吧？

Okay, so it's not going be like a big alien orb then?

在太空中。

Up in space.

到2036年，我不能说它不可能直接被发射到太空，然后被空间站的机械臂抓住，再插到合适的模块上。

By 2036, I wouldn't say that it's not possible for 2036 for it to be then directly launched into space and then maybe taken by a robotic arm at some space station and then plugged into the right module.

也许那就是2036年的事。

Maybe that's for 2036.

有趣。

Fun.

想想就觉得有趣。

Fun to think about.

我真的很喜欢这次对话。

I really enjoyed this conversation.

是的，内容很丰富。

Yes, it's ton

这是历史上最庞大、最激烈的资本支出建设，但我也认为这是一场技术

of This is the most intense, biggest CapEx build out in history, but I also think it's a technology

革命。

Revolution.

这场革命美得令人惊叹。

Revolution that is a thing of beauty.

我觉得你对技术之美、制约因素以及如何平衡这一切，把握得非常透彻。

And I think you have such a grasp over the beauty of the technology and the binding constraints and how to balance this all.

谷歌交到你手上，让人放心。

Google is in good hands.

感谢你抽出时间分享

Thank you for taking the time to share

你所做的事情。

what you do.

这真的非常有趣。

It's a ton of fun.

我们正亲身经历这一切，也在定义这一切。

We get to live through this and we get to define it.

非常感谢。

Thank you very much.

这是一次很棒的对话。

It's a great conversation.

谢谢。

Thank you.

多谢。

Thanks.