微软CEO萨提亚·纳德拉：微软如何为通用人工智能（AGI）做准备

我需要能够看到'哦，我把那个公式弄错了'。

I need to be able to see, oh, I got that formula wrong.

而且我需要理解这一点。

And I need to understand that.

所以在某种程度上，这些都不是通过用户界面封装层的提示完成的，而是通过在中间层教会它Excel的所有工具来实现的，对吧？

And then so, to some degree, that's all being done not at the UI wrapper level with some prompt, but it's being done in the middle tier by teaching it all the tools of Excel, right?

所以我实际上是在给它一个标记语言，教它成为一个高级Excel用户所需的技能。

So, I'm giving it even essentially a markdown to teach it the skills of what it means to be a sophisticated Excel user.

所以，这有点奇怪，它又回到了AI大脑的概念，对吧？

So, it's a weird thing that it goes back a little bit to AI brain, right?

你构建的不仅仅是Excel。

Which is you're building not just Excel.

你现在是传统意义上的业务逻辑。

You are now business logic in its traditional sense.

你正在将传统意义上的Excel业务逻辑包裹上一层认知层，通过这个知道如何使用工具的模型来实现。

You're taking the Excel business logic in the traditional sense and wrapping essentially a cognitive layer to it using this model, which knows how to use the tool.

所以，在某种意义上，Excel将自带分析师功能，并配备所有使用工具。

So, in some sense, Excel will come with an analyst bundled in and with all the tools used.

这类东西将会被所有人构建出来。

That's the type of stuff that'll get built by everybody.

因此，即使是模型公司也不得不参与竞争，对吧？

So, even for the model companies, they'll have to compete, right?

所以，如果他们定价过高，猜猜会怎样？

So, if they price stuff high, guess what?

如果我是这类工具的构建者，我会替换掉你。

If I'm a builder of a tool like this, I'll substitute you.

我可能会暂时使用你。

I may use you for a while.

因此，只要有竞争存在，就总会有赢家通吃的局面。

And so, as long as there's competition, there's always a winner take all thing.

如果不会出现一个遥遥领先的模型，是的，那就是赢家通吃。

If there's won't be one model that is better than everybody else with massive distance, yes, that's a winner take all.

只要存在多模型竞争格局，就像超大规模竞争那样，再加上开源机制的制衡，我们就有充分空间在这些模型基础上创造价值。

As long as there's going to be competition where there's multiple models, just like hyperscale competition, and there's an open source check, there is enough room here to go build value on top of models.

但在微软看来，我们的战略是深耕超大规模计算业务，该业务将支持多种模型并行。

But at Microsoft, the way I look at it and say is, we are going to be in the hyperscale business, which will support multiple models.

未来七年我们将持续获得OpenAI模型的访问权，并在此基础上进行创新。

We will have access to OpenAI models for seven more years, which we will innovate on top of.

因此从本质上说，我们拥有一个前沿级模型，可以完全灵活地使用和创新。

So, royalty And essentially, I think of ourselves as having a frontier class model that we can use and innovate on with full flexibility.

同时我们将通过MAI构建自己的模型。

And we'll build our own models with MAI.

这样我们将始终保持在模型层面的竞争力。

And so, we will always have a model level.

无论是安全领域、知识工作、编程还是科学研究，我们都会构建自己的应用框架——这些框架都将以模型为核心，对吧？

And then, we'll build these, whether it's in security, whether it's in knowledge work, whether it's in coding or in science, we will build our own application scaffolding, which will be model forward, right?

这不是在模型外包裹一层，而是将模型深度集成到应用中。

It won't be a wrapper on a model, but the model will be wrapped into the application.

关于你提到的其他内容，我有很多问题想问。

I have so many questions about the other things you mentioned.

但在深入讨论前，我仍怀疑这是否高估了当前AI能力——你设想的模型现在只能截取屏幕画面，却无法查看单元格内的具体公式。

But before we move on to those topics, I still wonder whether this is not forward looking on AI capabilities where you're imagining models like they exist today where, yeah, it takes a screenshot of your screen but it can't look inside each cell and what the formula is.

我认为更准确的类比是人类：想象这些模型最终能像人类知识工作者一样操作电脑，既能查看Excel公式，又能使用其他软件，还能在Office365和其他软件间迁移数据等等。

I And think the better mental model here is like, a human, just imagine that these models actually will be able to actually use a computer as well A as a human knowledge worker who is using Excel can look into the formulas, can use alternative software, can migrate data between Office three sixty five and another piece of software if the migration is necessary, etcetera.

这正是我想表达的意思。

That's kind of what I'm saying.

但如果真是这样，那么与Excel的集成就不那么重要了。

But if that's the case, then the integration with Excel doesn't matter that much.

不。

No.

因为不用担心Excel集成的问题。

Because don't worry about the Excel integration.

毕竟，Excel本就是为分析师打造的工具。

After all, Excel was built as a tool for analysts.

很好。

Great.

所以，任何作为分析师的AI都应该配备工具，它们可以使用...

So, whoever is this AI that is an analyst should have tools that they can They can use a

电脑，对吧？

computer, right?

就像人类能使用电脑一样，电脑就是它们的工具。

Just the way a human can use a computer, that's their tool.

工具就是电脑本身。

The tool is the computer.

对。

Right.

所以我的意思是，我正在构建的本质上是一个AI分析师，它天生就具备使用所有这些分析工具的知识。

So, all I'm saying is I'm building an analyst as essentially an AI agent, which happens to come with an a priori knowledge of how to use all of these analytical tools.

但或许需要确认我们说的是同一件事。

But is it something maybe just to make sure we're talking about the same thing.

像我这样用Excel做播客的人常见吗

Is it a thing that like me using Excel as a podcaster

完全精通某种自主性

Completely and a proficient at some autonomous.

想象一下我的工作方式，我们现在应该规划我对公司未来的构想，对吧？

So just imagine I work so we should now maybe lay out how I think the future of the company is, right?

公司的未来将是工具业务，我有电脑，我用Excel

The future of the company would be the tools business, which I have a computer, I use Excel.

事实上，未来我甚至还会配备一个副驾驶

And in fact, in the future, I'll even have a copilot.

那个副驾驶也会拥有多个智能体

And that copilot will also have agents.

仍然由我掌控全局，一切最终都会回归到我这里

It's still me steering everything and everything is coming back.

这算是其中一种发展模式

So that's kind of one world.

第二种模式是公司直接为AI智能体提供计算资源

Then the second world is the company just literally provisions a computing resource for an AI agent.

它将完全自主运行

And that is working fully autonomously.

这个全自主智能体将实质性地配备同样的工具套件

That fully autonomous agent will have essentially embodied set of those same tools.

对

Right.

这些对它来说是可用的，对吧？

That are available to it, right?

所以，这个AI工具不仅拥有原始计算能力，因为使用工具来完成工作会更节省令牌。

So, this AI tool that comes in also has not just a raw computer because it's going to be more token efficient to use tools to get stuff done.

事实上，我这样看待它：我们目前的终端用户工具业务，本质上将转变为支持智能体工作的基础设施业务。

In fact, I kind of look at it and say our business, which today is an end user tools business, will become essentially an infrastructure business in support of agents doing work.

这是另一种思考方式，对吧？

Is another way to think about it, right?

所以，你会看到我们实际做的一件事是，我们在M365下构建的所有内容仍然非常重要。

So, if one of the things that you'll see us do in fact, like all the stuff we built underneath M365 still is going to be very relevant.

你需要地方存储、归档、发现和管理所有这些活动，即使你是一个AI智能体。

You need some place to store it, some place to do archival, some place to do discovery, some place to manage all of these activities even if you're an AI agent.

所以，这有点像

So, it's kind

一种新的基础设施。

of a new infrastructure.

那么，为了确认我理解正确，你是说理论上，未来具有实际计算机使用能力的AI，所有这些公司正在研发的模型公司，即使不与微软合作或不在我们旗下，也可以使用微软软件。

So, just to make sure I understand, you're saying like, look, theoretically a future AI that has actual computer use, is all these companies are working on, model companies are working right now, could use, even if it's not partnered with Microsoft or under our umbrella, could use Microsoft software.

但你说的是，如果他们使用我们的基础设施，我们会提供更低层级的访问权限，让他们更高效地完成原本可以完成的工作

But you're saying we're going to give them if you're working with our infrastructure, we're going give you like lower level access that makes it more efficient for you to do the same things you could have otherwise A done

100%正确。

100%.

我的意思是，整个事情实际上是这样的：我们有了服务器，然后是虚拟化，接着我们有了更多的服务器。

I mean, so the entire thing and in fact, the way like what happened is we had servers, then there was virtualization, and then we had many more servers.

这是另一种思考方式，也就是说，不要把工具本身当作最终目的。

So that's another way to think about this, which is, hey, don't think of the tool as the end thing.

人类使用该工具时，其背后完整的底层基础是什么？

What is the entire substrate underneath that tool that humans use?

而这整个底层基础也是AI代理的启动平台，因为AI代理需要计算机。

And that entire substrate is the bootstrap for the AI agent as well because the AI agent needs a computer.

这算是其中一点。

That's kind of one.

事实上，我们看到一个显著增长的有趣现象是，那些开发办公自动化工具和自主代理等产品的团队，都想要部署Windows 365。

So, in fact, one of the fascinating things we are seeing a significant amount of growth is all these guys who are doing these office artifacts and what have you as autonomous agents and so on want to provision Windows three sixty five.

他们非常希望能为这些代理配置计算机。

They really wanna be able to provision a computer for these agents.

所以，确实如此。

And so, absolutely.

因此我认为，我们终将发展出一个终端用户计算基础设施业务，而且这个业务会持续增长——你猜为什么？

And that's where I think we're going to have essentially an end user computing infrastructure business, which I think is going to just keep growing because guess what?

它的增长速度将超过用户数量的增长。

It's going to grow faster than the number of users.

事实上，这也是人们常问我的另一个问题：按用户收费的业务会怎样？

So, in fact, that's one of the other questions people ask me is, hey, what happens to the per user business?

至少从早期迹象来看，或许应该这样理解：按用户收费的业务不仅是按用户，更是按代理收费。

At least the early signs maybe, the way to think about the per user business is not just per user, it's per agent.

如果你将其视为按用户和代理收费，关键在于每个代理需要配置什么资源？

And if you sort of say it's per user and per agent, the key is what's the stuff to provision for every agent?

一台计算机，周围有一系列安全措施，还有身份验证机制。

A computer, a set of security things around it, an identity around it.

所有这些可观测性等功能构成了管理层。

And all those things are observability and so on are the management layers.

我认为所有这些都将被整合进去。

And that's, I think all going to get baked into that.

至少就我目前的思考方式而言，我想听听你的看法——这些模型公司都在构建环境来训练它们的模型使用Excel或亚马逊购物等功能，预订航班等等。

The way to frame it, at least the way I currently think about it, I'd like to hear your view is that these model companies are all building environments to train their models to use Excel or Amazon shopping or whatever it is, book flights.

但与此同时，它们也在训练这些模型进行迁移工作，因为这可能是最直接产生价值的事情，对吧？

But at the same time, they're also training these models to do migration from because that is probably the most immediate valuable thing, right?

将基于大型机的系统转换为标准的云系统。

Converting mainframe based systems to standard cloud systems.

将Excel数据库转换为使用SQL的真实数据库。

Converting Excel databases into real databases with SQL.

对吧？

Right?

或者将Word和Excel中的内容转换为更具程序化、更高效且人类也能完成的传统形式。

Or converting, you know, what is done in Word and Excel to something that is more programmatic and more efficient in a classical sense that can actually be done by humans as well.

只是对软件开发人员来说这样做成本效益不高。

It's just not cost effective for the software developer to do that.

这似乎是未来至少几年内所有人利用AI大规模创造价值的方向。

That seems to be what everyone is going to do with AI for the next, you know, few years at least to massively drive value.

如果模型能自行利用工具进行迁移，微软将如何适应这种趋势？

How does Microsoft fit into that if the models can utilize the tools themselves to migrate to something?

确实，微软在数据库、存储以及其他诸多领域都占据领导地位，但办公生态系统的应用将会非常重要。

And yes, Microsoft has, you know, a leadership position in databases and in storage and in all these other categories, but the use of say, a office ecosystem is going to be significant.

这就好比大型机生态系统的应用可能会相对较少。

That's just like potentially the use of a mainframe ecosystem could be potentially less.

实际上，尽管现在没人再谈论大型机，但过去二十年间它们仍在持续增长。

Now, mainframes have grown for the last two decades actually, even though no one talks about them anymore, they've still grown.

是的。

Yeah.

百分之百同意。

A 100%.

我同意这个观点。

Agree with that.

未来会如何发展？

How that flow forward?

对。

Yeah.

意思是，在

Mean, at

归根结底，这并不是说我们将长期处于一个混合世界，对吧？

the end of the day, this is not about sort of, hey, there is going to be a significant amount of time where there's gonna be a hybrid world, right?

因为人们会使用工具，会与智能体协作，他们必须使用这些工具。

Because people are gonna be using the tools, they are gonna be working with agents, they have to use tools.

而且，他们还需要彼此沟通。

And by the way, they have to communicate with each other.

我生成的、需要人类查看的产物是什么？

What's the artifact I generate that then a human needs to see?

因此，所有这些因素在任何地方都是实际需要考虑的问题。

So all of these things will be real considerations in any place.

所以这些是输出和输入。

So the outputs, inputs.

所以我认为这不仅仅是关于'我已经迁移完成了'这么简单。

So I don't think it'll just be about, oh, I migrated off.

但归根结底，我必须生活在这个混合的世界里。

But the bottom line is I have to live in this hybrid world.

但这并没有完全回答你的问题，因为它们可能处于真正高效的新前沿领域，那里只有智能体与智能体协作，进行完全优化。

But that doesn't fully answer your question because they can be in real new efficient frontier where it's just agents working with agents and completely optimizing.

即使当智能体与智能体协作时，需要哪些基本要素？

And even when agents are working with agents, what are the primitives that are needed?

你需要存储系统吗？

Do you need a storage system?

这个存储系统需要具备电子取证功能吗？

Does that storage system need to have e discovery?

你需要具备可观测性吗？

Do you need to have observability?

你需要一个能让多个模型共用的身份识别系统吗？

Do you need to have an identity system that is going to use multiple models with all having one identity system?

所以，这些就是我们今天办公系统等所依赖的核心基础架构。

So, these are all the core underlying rails we have today for what are office systems or what have you.

我认为这也是我们未来将会拥有的。

And that's what I think we will have in the future as well.

你谈到了数据库。

You talked about databases.

老兄，我真希望整个Excel都能有个数据库后端。

Man, I would love all of Excel to have a database back end.

事实上，我希望这一切能立即实现。

In fact, I would love for all that to happen immediately.

而且那个数据库会是个好数据库。

And that database is a good database.

事实上，数据库将成为重要的发展方向。

Databases in fact will be a big thing that'll grow.

事实上，如果考虑到所有Office组件都能更好地结构化，借助智能体技术，结构化和非结构化数据之间的连接能力将得到提升，这将推动底层基础设施业务的发展。

In fact, if I think about all of the Office artifacts being structured better, the ability to do the joins between structured and unstructured better because of the agenting, that'll grow the underlying what is infrastructure business.

这一切的消费需求都是由智能体驱动的。

It happens the consumption of that is all being driven by agents.

你可以说这些都是模型公司实时生成的软件。

You could say all that is just in time generated software by a model company.

这种说法也可能是对的。

That could also be true.

我们也将成为这样一家模型公司。

We will be one such model company too.

因此，我们将持续构建。

And so, we will build in.

那么，竞赛内容可以是我们要构建一个模型加上所有基础设施并进行部署。

So, the competition could be that we will build a model plus all the infrastructure and provision it.

然后会有一批能做到这点的团队之间展开竞争。

And then there will be competition between a bunch of those folks who can do that.

我想说到模型公司，你说，好的，我们也将成为其中之一。不仅拥有基础设施，我们还将拥有模型本身。

I guess speaking of model companies, you say, okay, we will also be one of the Not only will have the infrastructure, we'll have the model itself.

目前微软AI最新发布的模型是两个月前的36号和Shotbot Arena。

Right now Microsoft AI's most recent model that was released two months ago is 36 and Shotbot Arena.

我的意思是，你们显然拥有OpenAI的知识产权。

And there's a I mean, you obviously have the IP rights to OpenAI.

所以首先的问题是，在你们认同这一点的前提下，看起来落后了，为什么会这样？

So there's a question of first, to the extent you agree with that, it seems to be behind, why is that the case?

特别是考虑到理论上你们有权直接分叉OpenAI的单一代码库或蒸馏他们的模型。

Especially given the fact that you theoretically have the right to just fork OpenAI's monorepo or distill on their models.

特别是如果'我们需要拥有一家领先的模型公司'是你们战略的重要组成部分。

Especially if it's a big part of your strategy that we need to have a leading model company.

是的。

Yeah.

首先，我们绝对会在所有产品中最大限度地使用OpenAI的模型。

So first of all, we are absolutely going to use the OpenAI models to the maximum across all of our products.

我认为，这是我们未来七年将持续做的核心工作。

Mean, that's, I think, the core thing that we're going to continue to do all the way for the next seven years.

不仅仅是使用它，还要为其增值。

And not just use it, but then add value to it.

这大概就是分析师和Excel代理的定位。

That's kind of where the analyst and this Excel agent.

这些都是我们将要实施的事项，包括进行RL（强化学习）微调。

And these are all things that we will do where we'll do RL fine tuning.

我们将在GPT系列模型基础上进行中期训练，利用独特的数据资产构建能力。

We'll do some mid training runs on top of a GPT family where we have unique data assets and build capability.

关于MAI模型，我认为好消息是——事实上根据新协议，我们可以非常明确地组建世界级的超级智能团队，并以高度雄心去实现它。

The MAI model, the way I think we're going to think about it is the good news here, in fact, with the new agreement is even we can be very, very clear that we're going to build a world class superintelligence team and go after it with a high ambition.

但同时，我们也会明智地利用这段时间来统筹运用这两方面。

But at the same time, we're also going to use this time to be smart about how to use both these things.

这意味着我们将一端高度聚焦产品，另一端则全力投入研究。

So that means we will, on one end, be very product focused and on the other end, be very research focused.

换句话说，因为我们能使用GPT系列模型。

In other words, because we have access to the GPT family.

我最不希望的就是把计算资源浪费在重复性且低价值的事情上。

The last thing I don't want to do is use my flops in a way that is just duplicative and doesn't add much value.

因此我希望优化用于生成GPT系列的计算资源，最大化其价值——同时MAI的计算资源正用于...比如我们刚发布的图像模型，目前在图像领域排名第九。

So, I want to be able to take the FLOPs that we use to generate a GPT family and maximize its value while my MAI flops are being used for Let's take the image model that we launched, which I think just launched is a number nine in the image arena.

我们同时将其用于成本优化。

We're using it both for cost optimization.

它已集成到Copilot中。

It's on Copilot.

它也在Bing里，我们将持续利用这个优势。

It's in Bing and we're going to use that.

我们在CoPilot中有一个音频模型，它具备个性和各种特性。

We have an audio model in CoPilot, which has got personality and what have you.

我们针对自家产品对其进行了优化。

We optimized it for our product.

所以我们会做这些工作。

So we will do those.

即使在Ella Marina项目上，我们也从文本模型开始着手。

Even on the Ella Marina, we started on the text one.

我记得它是在第13个晚上首次亮相的。

I think it debuted at night thirteen.

顺便说一句，当时仅用了大约15,000个H100芯片就完成了。

And by the way, it was done only on whatever, 15,000 H100s.

因此那是个非常小的模型。

And so it was a very small model.

这再次验证了核心能力——指令遵循等功能，我们要确保能达到业界最先进水平。

And so it was again to prove out the core capability, the instruction following, and everything else, we wanted to make sure we can match what was state of the art.

这向我们展示了，根据扩展定律，如果提供更多计算资源我们能实现什么。

And so that shows us, given scaling laws, what we are capable of doing if it gave more flops to it.

接下来我们要做的是全模态模型，整合我们在音频、图像和文本领域的工作成果。

So the next thing we will do is an omni model where we will take sort of the work we have done in audio, what we have done in image, and what we have done in text.

这将是MAI路线上的下一个里程碑。

That'll be the next pit stop on the MAI side.

当我规划MAI发展路线时，我们要组建世界一流的超级智能团队。

So when I think about the MAI roadmap, we're going to build a first class super intelligence team.

我们将继续公开推出并实践这些模型。

We're going to continue to drop and do in the open some of these models.

它们要么会被集成到我们的产品中，因其低延迟、低成本等优势而被使用，要么将具备某些特殊能力。

They will either be in our products being used because they're going to be latency friendly, cogs friendly, or what have you, or they'll have some special capability.

我们将开展实质性研究，为未来五到八项突破性进展做好准备——这些突破都是通往超级智能之路所必需的。

And we will do real research in order to be ready for some next five, six, seven, eight breakthroughs that are all needed on this march towards superintelligence.

因此我认为，在充分发挥我们现有GPT系列模型优势的基础上，我们还能实现更多突破。

So I think that's And while exploiting the advantage we have of having the GPT family that we can work on top of as well.

假设七年之后，你们无法再使用OpenAI的模型。

Say we roll forward seven years, you no longer have access to OpenAI models.

人们如何建立信心？或者说微软要如何确保自己拥有领先的AI实验室？

What does one get confidence or what does Microsoft do to make sure they are leading or have a leading AI lab, right?

如今OpenAI已实现诸多突破，无论是规模扩展还是逻辑推理；谷歌也研发了如Transformer等突破性技术——但人才争夺同样关键，对吧？

Today, all OpenAI has developed many of the breakthroughs, whether it be scaling or reasoning or Google's developed all the breakthroughs like transformers, but it is also a big talent game, right?

你知道Meta在人才引进上花费超过200亿美元对吧？

You know, you've seen Meta spend, you know, north of $20,000,000,000 on talent, right?

去年你看到Anthropic从谷歌挖走了整个蓝移推理团队。

You've seen Anthropic poach the entire blue shift reasoning team from Google last year.

更近些时候，Meta又从谷歌挖走了一个大型推理与后期训练团队。

You've seen Meta poach a large reasoning and post training team from Google more recently.

这类人才争夺战需要巨额资金支持。

These sorts of talent wars are very capital intensive.

可以说——如果你在基础设施上投入1000亿美元，那么也应该投入相应资金用于培养使用这些基础设施的人才，这样才能更高效地实现新突破。

They're the ones that, you know, arguably, you know, if you're spending $100,000,000,000 on infrastructure, you should also spend, you know, X amount of money on the people using the infrastructure so that they're more efficiently making these new breakthroughs.

人们能获得怎样的信心呢？你看，微软将拥有一支世界一流的团队来实现这些突破，一旦决定打开资金阀门，现在你们在资金使用上相当高效，这很明智，似乎避免了重复工作的浪费。

What confidence can one get that, you know, hey Microsoft will have a team that's world class that can make these breakthroughs and, you know, once you decide to turn on the money faucet, you know, you're being a bit capital efficient right now, which is just smart it seems to not waste money to doing duplicative work.

但一旦你们决定需要行动，怎么说呢，谁能断言‘没错，现在你们就能跃升至顶尖五大模型之列’呢？

But once you decide you need to, you know, how can one say, Oh yeah, now you can shoot up to where the top five models are.

哦，看。

Oh, look.

我的意思是，归根结底，我们将打造一支世界级团队，而且我们已经有了一个初具雏形的世界级团队样本，对吧？

I mean, at the end of the day, we're going to build a world class team and we already have a world class team that's beginning to be sort of a sample, right?

随着Mustafa的加入。

With Mustafa coming in.

我们有Karen，还有Amar Subramanian——他在Gemini 2.5中负责大量后期训练工作，现在在微软；还有Nando，他在DeepMind从事过多媒体相关工作，也在这里。

We have Karen, we have Amar Subramanian, who did a lot of the post training at Gemini, two point five, who is at Microsoft, Nando, who did a lot of the multimedia work at DeepMind is there.

因此，我们将组建一支世界级团队。

And so, we're going to build a world class team.

事实上，我想本周晚些时候，Mustafa还会更清晰地阐述我们实验室的发展方向。

And in fact, I think later this week, even Mustafa published a little more clarity on what our lab is going to go do.

我想让全世界知道的是：我们将构建支持多模型的基础设施。

I think the thing that I want the world to know perhaps is we are going to build the infrastructure that'll support multiple models.

从超大规模视角出发，我们要打造最具规模的基础设施集群，能够支撑全世界所需的所有模型——无论是开源模型，还是来自OpenAI等机构的模型。

Because from a hyperscale perspective, we want to build the most scaled infrastructure fleet that's capable of supporting all the models the world needs, whether it's from open source or obviously from OpenAI and others.

所以，这算是首要任务。

And so, that's kind of one job.

其次，在自有模型能力方面，我们绝对会在产品中使用OpenAI模型，同时也会开始构建自己的模型。

Second is in our own model capability, we will absolutely use the OpenAI model in our products, and we will start building our own models.

就像GitHub Copilot那样，我们可能会使用Anthropic的技术。

And we may, like in GitHub Copilot, Anthropic is used.

因此我们还将整合其他前沿模型到我们的产品中。

So we will even have other frontier models that are gonna be wrapped into our products as well.

我认为至少在每天结束时，产品在完成特定任务或工作时的评估才是最重要的。

So I think that that's kind of how, at least each time, at the end of the day, the eval of the product as it meets a particular task or a job is what matters.

我们将从那里反推所需的垂直整合，只要产品能很好地服务市场，成本总是可以优化的。

And we'll sort of back from there into the vertical integration needed, knowing that as long as you're serving the market well with the product, you can always cost optimize.

这里有个前瞻性的问题。

There's a question going forward.

目前我们的模型存在训练和推理的区别，但可以说不同模型之间的差异越来越小。

So right now we have models that have this distinction between training and inference and one could argue that there's like a smaller and smaller difference between the different models.

展望未来，如果真要实现类人智能，人类是在工作中学习的。

Going forward if you're really expecting something like human level intelligence, humans learn on the job.

想想你过去三十年的经历，是什么让Saita代币如此有价值？

If you think about your last thirty years, what makes Saita token so valuable?

是你在微软三十年来积累的智慧和经验。

It's the last thirty years of wisdom and experience you've gained in Microsoft.

最终如果我们实现人类水平的模型，它们将具备在工作中持续学习的能力。

And we will eventually have models if they get to human level which will have this ability to continuously learn on the job.

这将为领先的模型公司创造巨大价值，至少在我看来如此，因为你可以让一个模型广泛部署于整个经济体系，学习如何完成每一项工作。

And that will drive so much value to the model company that is ahead, at least in my view, because you have copies of one model broadly deployed through the economy, learning how to do every single job.

而且与人类不同，它们可以将所学知识融合到那个模型中。

And unlike humans, they can amalgamate their learnings to that model.

这种持续学习形成的指数级反馈循环，几乎像是一种智能爆炸。

So there's this sort of continuous learning sort of exponential feedback loop which almost looks like a sort of intelligence explosion.

如果这种情况发生，而微软届时还不是领先的模型公司，那么你说的这种用一个模型替代另一个模型等等的事情就不那么重要了，因为就像这个单一模型已经掌握了经济中所有工作的运作方式。

If that happens and Microsoft isn't the leading model company by that time doesn't then this you're saying, well, we substitute one model for another, etcetera, matter less because they're just like, this one model knows how to do every single job of the economy.

其他长尾部分则不具备这种能力。

The other long tail don't.

是的。

Yeah.

不，我认为你的观点是：如果存在一个唯一被广泛部署的模型，它能获取所有数据并持续学习，那就直接定胜负关店大吉了，对吧？

No, I think your point about if there's one model that is the only model that's most broadly deployed in the world and it sees all the data and it has continuous learning, that's game set match and is shut shop, right?

我的意思是，至少我看到的事实是，即使在今天，尽管存在某个模型的主导地位，但实际情况并非如此。

I mean, the reality, at least I see, is the world, even today, for all the dominance of any one model, it's not the case.

以编程为例。

Take coding.

存在多个模型。

There's multiple models.

事实上，每天的情况都更偏离'只有一个模型被广泛部署'的状态。

In fact, every day, it's less the case where there is not one model that is getting deployed broadly.

实际上，是有多个模型正在被部署。

In fact, there's multiple models that are getting deployed.

这有点像数据库的情况，对吧？

It's kind of like databases, right?

总是会出现这样的疑问：是否真的能有一个数据库可以通用于所有场景？

It's always the thing is like, Hey, can one database be the one that just is used everywhere?

但事实并非如此。

Except it's not.

目前有多种类型的数据库被部署于不同的应用场景。

There are multiple types of databases that are getting deployed different use cases.

因此我认为，持续学习或数据——我称之为数据流动性——会带来网络效应，这是任何单一模型都具备的特性。

So, I think that there is going to be some network effects of continual learning or data, I'll call liquidity, that any one model has.

这会发生在所有领域吗？

Is it going to happen in all domains?

我不这么认为。

I don't think so.

这会发生在所有地区吗？

Is it going to happen in all geos?

我不这么认为。

I don't think so.

这会发生在所有细分市场吗？

Is it going to happen in all segments?

我不这么认为。

I don't think so.

所有类别会同时发生这种情况吗？

It'll happen in all categories at the same time?

我不这么认为。

I don't think so.

因此，我认为设计空间如此广阔，存在着大量机遇。

So, therefore, I feel like the design space is so large that there's plenty of opportunity.

但你核心观点是要具备基础设施层、模型层和框架层的能力。

But your fundamental point is having a capability which is at the infrastructure layer, model layer, and at the scaffolding layer.

然后要能够组合这些要素，不仅是垂直堆叠，更要按各自用途灵活组合。

And then to be able to compose these things, not just as a vertical stack, but to be able to compose each thing for what its purpose is.

你不能构建只为单一模型优化的基础设施。

You can't build an infrastructure that's optimized for one model.

如果那样做，万一技术落伍了怎么办？

If you do that, what if you go fall behind?

事实上，你搭建的所有基础设施都将报废，对吧？

In fact, all the infrastructure you built will be a waste, right?

你需要构建能支持多种模型家族和谱系的基础设施。

You kind of need to build an infrastructure that's capable of supporting multiple sort of families and lineages of models.

否则，你投入的资本若只为单一模型架构优化，意味着只要别人在MOE等架构上取得突破，你的整个网络拓扑就会瞬间过时，这很可怕。

Otherwise, the capital you put in, which is optimized for one model architecture, that means you're one tweak away from some MOE like breakthrough that happens for somebody else and your entire network topology goes out of the window, then that's a scary thing.

因此，你希望基础设施能支持未来可能出现的任何模型，包括自家和其他家族的模型。

So therefore, you kind of want the infrastructure to support whatever may come, in fact, in your own model family and other model families.

你必须保持开放态度。

And you've got be open.

如果你真想做好超大规模业务，就必须认真对待这一点，对吧？

If you're serious about the hyperscale business, you've got to be serious about that, right?

如果想成为真正的模型公司，就必须思考：人们能在模型基础上开发哪些应用？这样才能建立ISV生态系统，除非你认为自己能垄断所有领域。

If you're serious about being a model company, you've got to basically say, Hey, what are the ways people can actually do things on top of the model so that I can have an ISV ecosystem, unless I'm thinking I'll own every category.

这根本不可能。

That just can't be.

那么你将无法开展API业务。

Then you won't have an API business.

按照定义，这意味着你永远无法成为一家能成功部署到各处的平台公司。

And that by definition will mean you'll never be a platform company that's going to be successfully deployed everywhere.

因此，行业结构将迫使人们走向专业化。

So therefore, the industry structure is such that it will really force people to specialize.

在这种专业化进程中，像微软这样的公司应当凭借自身优势在每个层级展开竞争，而非认为这是一场决胜局——只需垂直整合所有层级就能获胜。

And in that specialization, a company like Microsoft should compete in each layer by its merits, but not think that this is all about all a road to game set match where I just compose vertically all these layers.

这种情况根本不会发生。

That just doesn't happen.

根据迪伦的数据，仅明年AI资本支出就将达五千亿美元，实验室已投入数十亿美元争夺顶尖研究人才。

So according to Dylan's numbers, there's gonna be half a trillion in AI CapEx next year alone, and labs are already spending billions of dollars to snag top researcher talent.

但如果没有足够的高质量训练数据，这一切都毫无意义。

But none of that matters if there's not enough high quality data to train on.

缺乏正确数据，即使最先进的基础设施和世界级人才也无法为用户创造最终价值。

Without the right data, even the most advanced infrastructure and world class talent won't translate into end value for the user.

这正是Labelbox的用武之地。

That's where Labelbox comes in.

Labelbox能大规模生产高质量数据，为你想要的任何模型能力提供支持。

Labelbox produces high quality data at massive scale, powering any capability that you want your model to have.

无论你需要一个需要多小时轨迹详细反馈的编程代理，还是需要日常任务数千样本的机器人模型，亦或是能为用户执行预订航班等现实操作的语音代理。

It doesn't matter whether you need a coding agent that needs detailed feedback on multi hour trajectories or a robotics model that needs thousands of samples on everyday tasks or a voice agent that can also perform real world actions for the user, like booking them a flight.

需要明确的是，这绝非现成的通用数据。

To be clear, this isn't just off the shelf data.

Labelbox能在48小时内设计并启动一个定制化的生产级数据管道，几周内就能为您提供数万个目标样本。

Labelbox can design and launch a custom production scale data pipeline in forty eight hours, and they can get you tens of thousands of targeted examples in weeks.

请访问libobox.com联系dwarkesh。

Reach out at libobox dot com dwarkesh.

好的，我们回到Satya的话题。

All right, back to Satya.

去年微软原本即将成为遥遥领先的最大基础设施供应商。

So last year Microsoft was on path to be the largest infrastructure provider by far.

你们在2023年最早行动，抢占先机，通过租赁数据中心、启动建设、确保电力供应等各种方式获取了所有资源。

You were the earliest in '23, so you went out there, you acquired all the resources in terms of leasing data centers, starting construction, securing power, everything.

按照这个速度，你们本可以在2026或2027年超越亚马逊。

You guys were on pace to beat Amazon in '26 or '27.

最迟到2028年，你们必将超越他们。

But certainly by '28, you're going to beat them.

但从去年下半年开始，微软突然叫停了大规模扩张计划对吧？

Since then, know, in let's call it the second half of last year, Microsoft did this big pause, right?

他们放弃了许多原计划租赁的场地，这些场地随后被谷歌、Meta、亚马逊甚至甲骨文接手。

Where they let go of a bunch of leasing sites that they were going to take, which then Google, Meta, Amazon in some cases, Oracle took these sites.

我们现在身处全球最大的数据中心之一，显然这还不是全部。

We're sitting in one of the largest data centers in the world, so obviously it's not everything.

你们的扩张速度依然惊人。

You guys are expanding like crazy.

但确实有些项目站点你们直接停止了推进。

But there are sites that you just stopped working on.

你为什么要这么做，对吧？

Why did you do this, right?

是的。

Yeah.

我是说，最根本的问题，这稍微回到超大规模业务的核心是什么，对吧？

I mean, the fundamental thing, this goes back a little bit to what is the hyperscale business all about, right?

我们做出的关键决策之一就是，如果你要把Azure打造成从训练到中期训练、数据生成到推理等AI各阶段都出色的平台，我们就需要整个机队的可互换性。

Which is one of the key decisions we made was that if you're gonna build out Azure to be fantastic for all sort of stages of AI from training to mid training to data gen to inference, we just need fungibility of the fleet.

因此这整个情况导致我们基本上没有针对特定几代技术去大量建设产能。

And so that entire thing caused us not to basically go build a whole lot of capacity with a particular set of generations.

因为你必须意识到另一个事实是，迄今为止我们每18个月就要为各种OpenAI模型准备10倍于前的训练容量。

Because the other thing that you've got to realize is having actually up to now 10x every eighteen months enough training capacity for the various OpenAI models.

我们意识到关键在于坚持这条道路。

We realized that the key is to stay on that path.

但更重要的是要保持平衡，不仅要训练，还要能在全球范围部署这些模型。

But the more important thing is to actually have a balance to not just train, but to be able to serve these models all around the world.

因为归根结底，货币化速度才能让我们持续获得资金支持。

Because at the end of the day, the rate of monetization is what then will allow us to even keep funding.

正如我所说，基础设施需要支持多种模型等等。

And then the infrastructure was going to need us to support, as I said, multiple models and what have you.

所以一旦我们确认了这一点，之后就沿着当前路径调整方向。在我看来，我们现在正启动更多项目。

So once we said that that's the case, since then we just course corrected to the path we're on, If I look at the path we're on is we are doing a lot more starts now.

我们也在尽可能多地采购托管容量，无论是自建、租赁还是GPU即服务。

We are also buying up as managed capacity as we can, whether it's to build, whether it's to lease, or even GPUs as a service.

但我们正在根据我们看到的服务需求、培训需求以及市场需求来构建它。

But we are building it for where we see the demand and the serving needs and our training needs.

而且我们不想仅仅成为一家公司的托管方，只拥有单一客户的大规模业务。

And we didn't want to just be a hoster for one company and have just a massive book of business with one customer.

那不算真正的业务，对吧？

That's not a business, right?

你应该与那家公司进行垂直整合。

You should be vertically integrated with that company.

因此，OpenAI将成为一家成功的独立公司，这很棒，我认为这是合理的。

And so, the thing that OpenAI was going to be a successful independent company, which is fantastic, I think it makes sense.

甚至Meta也可能使用第三方资源，但最终它们都会成为第一方。

And even Meta may use third party capacity, but ultimately, they're all going to be first party.

任何拥有大规模需求的企业，最终都会成为自己的超大规模服务商。

For anyone who has large scale, they'll be a hyperscaler on their own.

所以对我来说，就是要建立超大规模集群和我们自己的研究算力。

And so, to me, it was to build out a hyperscale fleet and our own research compute.

这就是调整的方向。

And that's what the adjustment was.

因此，我感到非常非常满意。

And so, I feel very, very good.

哦，顺便说一句，另一个原因是我不想被单一世代的超大规模所束缚。

Oh, by the way, the other thing is I didn't want to get stuck with massive scale of one generation.

我是说，我们刚刚看到了GB200系列。

I mean, we just saw the GB200s.

我是说，GB300系列就要来了，对吧？

I mean, the GB300s are coming, right?

等到Vera Rubin和Vera Rubin Ultra问世时，你猜怎么着？

And by the time I get to Vera Rubin, Vera Rubin Ultra, guess what?

数据中心的面貌将大不相同，因为每个机架、每排机柜的功耗都会发生巨大变化。

The data center is gonna look very different because the power per rack, power per row is gonna be so different.

冷却需求也将截然不同。

The cooling requirements are going to be so different.

这意味着我不想投入数十亿瓦特的基础设施只为支撑单一世代、单一产品线。

And that means I don't want to just go build out like a whole number of gigawatts that are only for a one generation, one family.

因此，我认为推进节奏、资源通用性和选址布局都至关重要。

And so, I think the pacing matters and the fungibility and the location matters.

工作负载的多样性很关键。

The workload diversity matters.

客户群体的多样性也很重要。

Customer diversity matters.

而这正是我们努力构建的方向。

And that's what we're building towards.

我们深刻认识到，每个AI工作负载不仅需要AI加速器，还需要大量配套支持。

The other thing that we've learned a lot is every AI workload does require not only the AI accelerator, but it requires a whole lot of other things.

事实上，我们的利润结构将主要来自这些配套环节。

And in fact, a lot of the margin structure for us will be in those other things.

因此我们希望将Azure打造成能完美支持长尾工作负载的平台——这才是超大规模业务的本质，同时确保我们在高端训练场景的底层硬件上保持绝对竞争力。

And so, therefore, we want to build out Azure as being fantastic for the long tail of the work workloads, because that's the hyperscale business, while knowing that we've got to be super competitive, starting with the bare metal for the highest end training.

但这不会挤占其他业务，对吧？

But that can't crowd out the rest of the business, right?

因为我们不是只做五份合同、为五个客户提供裸机服务的业务。

Because we're not in the business of just doing five contracts with five customers being their bare metal service.

这不是微软的业务模式。

That's not a Microsoft business.

这可能是别人的业务模式，这很好。

That may be a business for someone else, and that's a good thing.

我们说过，我们从事的是超大规模业务，归根结底是为AI工作负载提供长尾服务。

What we have said is we are in the hyperscale business, which is at the end of the day, a long tail business for AI workloads.

为此，我们将为包括自研模型在内的一系列模型提供领先的裸机即服务能力。

And in order to do that, we will have some leading bare metal as a service capabilities for a set of models, including our own.

我认为这就是你们看到的平衡点。

And that I think is the balance you see.

围绕这个可替代性话题的另一个问题是：现状并不理想，对吧？

Another sort of question that comes around this whole fungibility topic is, okay, it's not where you want it, right?

你们更希望把资源放在像亚特兰大这样优质的人口中心，就像我们现在这里。

You would rather have it in a good population center like Atlanta's, we're here.

还有个问题是：随着AI任务范围的扩大，这种地域差异还重要吗？

There's also the question of like, well, how much does that matter if as the horizon of AI tasks grows?

实际上，你知道的——推理提示可能只需30秒，深度研究可能需要30分钟，而软件代理未来可能需要数小时甚至数天才能完成人机交互。

Well, actually, you know, thirty seconds for a reasoning prompt or, you know, thirty minutes for a deep research or it's going to be hours for software agents at some point and days and so on and so forth, the time to human interaction.

如果延迟是——确实如此。

Why does it matter if it's- Yeah.

A

A

好问题。

great question.

位置A、B或

Location A, B or

C。

C.

完全正确。

That's exactly right.

事实上，这正是我们想要思考的另一个原因：Azure区域应该是什么样子的？

So, fact, that's one of other reasons why we wanna think about, Hey, what does an Azure region look like?

实际上，Azure区域之间的网络连接是怎样的？

And what is, in fact, the networking between Azure regions?

因此，我认为随着模型能力的演进，这些令牌的使用方式（无论是同步还是异步）也会发展。

So, this is where I think as the model capabilities evolve and I think the usage of these tokens, whether it's synchronously or asynchronously, evolves.

事实上，你不想处于不利位置。

And in fact, you don't want to be out of position.

除此之外，顺便问一下，数据驻留法律有哪些？

Then on top of that, by the way, what are the data residency laws?

对我们来说整个欧盟数据边界问题，意味着你不能随意将调用往返发送到任何地方，即使是异步的。

The entire EU thing for us where we literally had to create an EU data boundary basically meant that you can't just round trip a call to wherever, even if it's asynchronous.

因此，你可能需要高密度的区域性设施，以及考虑电力成本等因素。

And so, therefore, you need to have maybe regional things that are high density and then the power costs and so on.

但你完全正确，提出我们在建设过程中拓扑结构必须不断演进这一点。

But you're 100% right in bringing up that the topology as we build out will have to evolve.

第一点，关于每美元每瓦特的代币效率。

One, for tokens per dollar per watt.

经济效益如何？

What are the economics?

在此基础上叠加使用模式分析——同步、异步场景，同时考虑计算存储需求，因为延迟可能对某些场景至关重要。

Overlay that with what is the usage pattern in terms of synchronous, asynchronous, but also what is the compute storage because the latencies may matter for certain things.

存储资源必须到位。

The storage better be there.

如果我在附近部署Cosmos DB来处理会话数据甚至自动驾驶数据，那么这些资源也必须就近部署，以此类推。

If I have a Cosmos DB close to this for session data or even for an autonomous thing, then that also has to be somewhere close to it and so on.

因此我认为，所有这些考量因素将塑造超大规模业务的形态。

So, I think that all of those considerations is what will shape the hyperscale business.

在暂停之前，按我们为你们预测的28年目标，本该达到12-13吉瓦规模，但现在只有9.5吉瓦左右对吧？

Prior to the pause, versus what we had forecasted for you by '28, you're going to be like twelve, thirteen gigawatts and now we're at 9.5 or so, right?

但有个更关键的问题对吧？

But something that's even more relevant, right?

我希望你能更明确地声明：这就是你们不想涉足的领域。

And it's, I just want you to like more concretely state that this is the business you don't want be in.

比如甲骨文将从只有你们五分之一的规模发展到2027年超越你们。

Like Oracle is going from like one fifth your size to bigger than you by 2027.

虽然其资本回报率还达不到微软的水平对吧？

And while it's not a Microsoft level quality of return on invested capital, right?

他们仍然保持着35%的毛利率，对吧？

They're still making 35% gross margins, right?

问题的关键在于，这或许不是微软的业务范畴，但你们通过拒绝这项业务、放弃优先权，实际上创造了一个超大规模服务商

Sort of the question is like, hey, it's not Microsoft's business to maybe do this, but you've created a hyperscaler now by refusing this business, by giving away the right

首先声明，我并非要否定甲骨文在构建业务方面取得的成功

of I'm first refusal, not, first of all, I don't want to take away anything from the success Oracle has had in building their business.

我衷心祝愿他们发展顺利

And I wish them well.

因此，我认为已经回答您的是：成为一家单一模型公司的托管商对我们没有意义，何况其时间范围和RPO都有限？

And so, the thing that I think I've answered for you is it didn't make sense for us to go be a hoster for one model company with limited time horizon RPO?

就这么说吧，对吧？

Let's just put it that way, right?

真正需要深思的不是未来五年做什么，而是未来五十年做什么

Thing that you have to think through is not what you do in the next five years, but what you'll do for the next fifty.

因为这正是我们做决策时的考量依据

Because that's kind of what I We made our set of decisions.

我对与OpenAI的合作及我们的工作方向感到非常满意

I feel very good about our OpenAI partnership and what we're doing.

我们拥有可观的业务量

We have a decent book of business.

我们祝愿他们取得巨大成功

We wish them a lot of success.

事实上，我们甚至采购甲骨文的服务器容量

In fact, we are buyers even of Oracle capacity.

我们祝愿他们成功。

We wish them success.

但在这一点上，我认为我们正在尝试的工业逻辑相当清晰，这不是关于追逐。

But at this point, I think the industrial logic for what we are trying to do is pretty clear, which is it's not about chasing.

首先，我顺便追踪你们的东西，无论是AWS、谷歌还是我们的，我觉得这非常有用。

First of all, I track, by the way, your things, whether it's the AWS or the Google and ours, which I think is super useful.

但这并不意味着我必须追逐它们。

But doesn't mean I got to chase those.

我必须追逐它们，不仅仅是因为它们可能在一段时间内代表的毛利率。

I have to chase them for not just the gross margin that they may represent in a period of time.

微软独特能开拓且对我们有意义的业务领域是什么？

What is this book of business that Microsoft uniquely can go clear which makes sense for us to clear?

这就是我们要做的。

And that's what we'll do.

我想我有个问题，甚至退一步说，好吧，我同意你的观点，在所有条件相同的情况下，拥有一长串能带来更高利润的客户，比仅为少数实验室提供裸机服务是更好的业务。

I guess I have a question, even stepping back from this of, okay, I take your point that it's a better business to be in all else equal to have a long tail of customers who can have higher margin from rather than just serving bare metal to a few labs.

但接下来的问题是，好吧，行业正在向哪个方向发展？

But then there's a question of, okay, which way is the industry evolving?

因此，如果我们相信我们正走在通往越来越智能AI的道路上，那么为什么行业形态不是由OpenAI、Anthropic和DeepMind作为平台，而企业长尾实际与之开展业务，它们需要裸机但更像是平台呢？

And so if we believe we're on the path to smarter and smarter AIs, then why isn't the shape of the industry that the OpenAI's and Anthropix and DeepMinds are the platform which the long tail of enterprises are actually doing business with where they need bare metal but like they are platform.

直接使用Azure的长尾是什么？

What is the long tail that is directly using Azure?

因为你想使用通用的云，但

Because you want to use the general But cloud what

Azure上这个功能要多久才能用？

long is that's to be available on Azure?

所以任何工作负载如果说，嘿，我想用些开源模型和OpenAI模型。

So, any workload that says, Hey, I want to use some open source model and an OpenAI model.

我是说，你现在去Azure Foundry，就能通过PTU配置所有这些模型，获取Cosmos DB、SQL数据库、存储资源和计算资源。

I mean, you go to Azure Foundry today, you have all these models that you can provision by PTUs, get a Cosmos DB, get a SQL DB, get some storage, get some compute.

这才是真正的工作负载形态。

That's what a real workload looks like.

真正的工作负载不只是简单调用模型API。

A real workload is not just Hey, I did an API call to a model.

真正的工作负载需要整合所有这些资源来构建应用或实例化应用程序。

A real workload needs all of these things to go build an app or instantiate an application.

事实上，模型公司也需要这些来构建任何东西。

In fact, the model companies need that to build anything.

这不仅仅是拥有一个访问令牌那么简单，而是需要整套资源。

It's just not like I have a token I have to have all of these things.

这就是超大规模业务的本质。

That's the hyperscale business.

而且不只是一个模型，而是所有这些模型。

And it's not only one model, but all these models.

所以，如果你想要Grok+，比方说OpenAI加开源模型，就来Azure Foundry配置它们，构建你的应用。

And so, if you want Grok plus, let's say, OpenAI plus an open source model, come to Azure Foundry, provision them, build your application.

这里有个现成的数据库。

Here is a database.

这基本上就是业务的本质。

That's kind of what the business is.

有一个独立的业务就是单纯销售裸金属服务的公司。

There is a separate business called just selling raw bare metal services companies.

这就是关于你想涉足多少这类业务、不想涉足多少以及它究竟是什么的争论。

And that's the argument about how much of that business you want to be in and not be in and what is that.

这是我们参与的、非常不同的业务领域。

It's a very different segment of the business, which we are in.

而且我们对它挤占其他业务的程度也有限制。

And we also have limits to how much of it is going to crowd out the rest of it.

但至少我是这么看待这个问题的。

But that's kind of at least the way I look at it.

所以这里其实有两个问题，对吧？

So, there's sort of two questions here, right?

比如，你们难道不能两者兼顾吗？这是第一个问题。

Like, couldn't you just do both is one.

另一个问题是根据我们对你们2028年产能的预估，3.5吉瓦的容量是否偏低。

And then the other one is given our estimates on what your capacity is in 2028, is 3.5 gigawatts lower.

当然，你们本可以把这部分专门用于OpenAI的训练和推理算力。

Sure, you could have dedicated that to OpenAI training and inference capacity.

但你们也可以把这3.5吉瓦用于运行Azure、运行Microsoft 365、运行GitHub Copilot。

But you could have also dedicated that to, hey, this 3.5 gigawatts is actually just running Azure, is running Microsoft three sixty five, it's running GitHub Copilot.

我本可以建设这些产能而不分配给OpenAI。

I could have built it and not given it to OpenAI.

或者

Or

我可能想在其他地方建造它。

I may want to build it in a different location.

可能想在阿联酋建造。

May want to build it in UAE.

我可能想在印度建造。

I may want to build it in India.

我可能想在欧洲建造。

I may want to build it in Europe.

所以，正如我所说，另一个问题是，考虑到监管要求和数据主权需求，我们目前确实面临产能限制。

So, one of the other things is, as I said, where we have real capacity constraints right now are given the regulatory needs and the data sovereignty needs.

我们必须在全球范围内建设。

We've to build all over the world.

首先，美国本土的产能极其重要，我们将建设所有设施。

First of all, stateside capacity is super important and we're going to build everything.

但当我展望2030年时，我对微软业务形态有一个全球视角，包括第一方和第三方业务。

But one of the things is when I look out to 2030, I have a sort of a global view of what does Microsoft shape of business by first party and third party.

第三方业务按前沿实验室划分，他们所需与我们为多种模型构建的推理能力及自身研究计算需求之间的对比，对吧？

Third party segmented by the Frontier Labs and how much they want versus the inference capacity we want to build for multiple models and our own research compute needs, right?

所以，这些都是我考量中的因素，而不是简单地说，嘿，你正确地指出了暂停。

So, that's all what's going into my calculus versus saying, Hey, I think you're rightfully pointing out the pause.

但暂停并非因为我们说'天啊，我们不想建那个'而做出的决定。

But the pause was not done because we said, Oh my god, we don't want to build that.

我们意识到，哦，我们想根据工作负载类型、地域类型以及时间因素，以略有不同的方式构建我们想要的东西。

We realized that, Oh, we want to build what we want to build slightly differently by both workload type as well as geo type and timing as well.

我们将继续增加我们的千兆瓦产能。

We'll keep ramping up our gigawatts.

问题在于，以何种速度以及在何地推进？

And the question is, at what pace and in what location?

我该如何在这上面应用摩尔定律呢？

How do I write even the Moore's Law on it?

也就是说，我是否真的想在2027年过度建设3.5倍产能？

Which is, do I really want to overbuild 3.5 in 'twenty seven?

还是说考虑到已知因素，我更应该把建设周期分摊到2027和2028年？

Or do I want to spread that in 'twenty seven, 'twenty eight knowing even?

我们从英伟达身上学到的重要一课就是：他们的技术迭代速度正在加快。

One of the biggest learnings we had even with NVIDIA is their pace increased in terms of their migrations.

所以这是个关键因素。

So, that was a big factor.

我不想因为一代设备就被困在四五年折旧周期里动弹不得。

I didn't want to go get stuck for four years, five years depreciation on one generation.

我只想简单直接地采购。

And I wanted to just basically buy.

实际上，黄仁勋给了我两条建议。

In fact, Jensen's advice to me was two things.

第一条是：要像光速执行那样迅速行动。

One is, hey, get on the speed of light execution.

这就是为什么我认为即使在亚特兰大数据中心的实施，我是说，大约九十天内，从我们接手到交付实际工作负载的这段时间。

That's why I think even the execution in this Atlanta data center, I mean, like in ninety days, right, between when we get it to hand off to a real workload.

这算是他们那边真正的光速执行了。

That's sort of real speed of light execution on their front.

所以我想在这方面做好。

And so I wanted to get good on that.

这样一来，我就能在每一代产品上逐步构建和扩展。

And then that way, then I'm building this each generation and scaling.

然后每五年，你就能获得更均衡的发展。

And then every five years, then you have a much more balanced.

所以，这真的就像大型工业操作的流程，突然之间你不会失衡，不会在某段时间大量建设后又长期停滞——就像你说的，集中在一个地点可能对训练很有利，但对推理就不太理想，因为即使全是异步处理我也无法提供服务。

So, it becomes really literally like a flow for a large scale industrial operation like this, where you suddenly are not lopsided, where you built up a lot in one time and then you take a massive hiatus because you're stuck with all this, to your point, in one location, which may be great for training, may not be great for inference because I can't serve even if it's all asynchronous.

但欧洲不会允许我往返德克萨斯州。

But Europe ain't going to let me round trip to Texas.

所以，这些就是全部情况。

So, that's all of the things.

我该如何将这番声明与你过去几周的行为协调起来？

How do I rationalize this statement with what you've done over the last few weeks?

你已经宣布与Iris Energy、Nebius和Lambda Labs达成协议，还有几个即将公布。

You've announced deals with Iris Energy, with Nebius and Lambda Labs, and there's a few more coming as well.

你正在通过租用Neo Clouds的算力来获取容量，而不是自己建设。

You're going out there and securing capacity that you're renting from the Neo Clouds rather than having built it yourself.

我觉得这没问题

What was the I think it's fine

这对我们来说是个好消息，因为我们现在拥有了它。

for us because we now have it.

当你能预见需求，并且这些需求能在人们建设的地方得到满足时，那就太棒了。

When you have line of sight to demand, which can be served where people are building it, great.

事实上，我甚至可以说，我们会接受租赁。

In fact, we'll even have, I would say, we will take leases.

我们会接受定制化建设。

We will take build to suite.

我们甚至会采用GPU即服务模式，在我们自身容量不足但需要时，利用他人提供的资源。

We'll take even GPUs as a service where we don't have capacity but we need capacity and someone else has that.

顺便说一句，我甚至欢迎所有Neo Cloud加入我们的市场平台。

And by the way, I would even sort of welcome every Neo Cloud to just be part of our marketplace.

因为，再猜怎么着？

Because again, guess what?

如果他们将其容量引入我们的市场，通过Azure来的客户就会使用Neo Cloud——这对他们来说是重大胜利，同时也会使用Azure的计算、存储、数据库等其他服务。

If they go bring their capacity into our marketplace, that customer who comes through Azure will use the Neo Cloud, which is a great win for them, and will use compute storage, databases, all the rest from Azure.

所以我完全不认为这只是‘我应该独吞所有’的事情。

So I'm not at all thinking of this as just, Hey, I should just go gobble up all of that myself.

所以

So

你提到如何对这项五六年寿命的资产进行折旧，这占据了数据中心总拥有成本75%的大部分。

you mentioned how you're depreciating this asset that's five, six years and this is the majority of the 75% of the TCO data center.

而Jensen从中获得了75%的利润率。

And Jensen is taking a 75% margin on that.

所有超大规模厂商都在尝试开发自己的加速器，以降低设备带来的巨额成本，从而提高利润率。

So what all the hyperscalers are trying to do is develop their own accelerator so that they can reduce this overwhelming cost for equipment to increase their margins.

是的。

Yeah.

然后你看他们所处的位置，对吧？

And then like you know, when you look at where they are, right?

谷歌遥遥领先于其他公司，对吧？

Google's way ahead of everyone else, right?

他们从事这方面的时间最久。

They've been doing it for the longest.

他们计划生产约500到700万枚芯片，对吧？

They're going to make something like five to 7,000,000 chips, right?

都是他们自研的TPU。

Of their own TPUs.

看看亚马逊，他们正试图生产300到500万枚。

You look at Amazon, they're trying to make three to 5,000,000.

但当我们观察微软订购的自研芯片数量时，远低于这个数字。

But when we look at what, Microsoft is ordering of their own chips, it's way below that number.

你们开展这个项目的时间同样很长。

You've had a program for just as long.

你们内部情况如何？嗯，问得好。

What's going on with your internal Yeah, good question.

有几件事要说。

So the couple of things.

对于任何新加速器来说，最大的竞争对手其实是英伟达的上一代产品，对吧？

One is the thing that is the biggest competitor for any new accelerator is kind of even the previous generation of NVIDIA, right?

我的意思是，在整体部署中，我关注的是总拥有成本(TCO)。

I mean, in a fleet, what I'm going to look at is the overall TCO.

所以，我甚至对我们自己的产品也设定了标准——顺便说下，我刚看了Maya 200的数据表现很好，但我们在计算领域也发现：最初主要用英特尔，后来引入AMD，再然后是钴架构。

So, the bar I have even for our own, and which, by the way, I was just looking at the data for Maya 200, which looks great, except that one of the things that we learned even on the compute side, which is we had a lot of Intel, then we introduced AMD, and then we introduced Cobalt.

我们就是这样实现规模扩展的。

And so, that's kind of how we scaled it.

因此至少在核心计算领域，我们拥有充分的实证：如何自研芯片并管理同时运用这三种架构的混合集群。

And so, we have good sort of existence proof of, at least in core compute, on how to build your own silicon and then manage a fleet where all three are at play in some balance.

毕竟连谷歌和亚马逊也都在采购英伟达。

Because, by the way, even Google's buying NVIDIA and so is Amazon.

这很合理，因为英伟达持续创新且具备通用性。

And it makes sense because NVIDIA is innovating and it's the general purpose thing.

所有模型都能在其上运行。

All models run on it.

而且客户需求确实存在。

And customer demand is there.

如果你要打造垂直解决方案，就必须配套自有模型——无论是用于训练还是推理，同时还要自主创造需求或补贴需求。

Because if you build your own vertical thing, you better have your own model, which is either gonna use it for training or inference, and you have to generate your own demand for it or subsidize the demand for it.

因此你需要确保规模扩展的合理性。

So therefore, you wanna make sure you scale it appropriately.

所以我们的做法是建立MAI模型与自研芯片的闭环，我认为这才是获得芯片自主权的根本——即根据实际需求设计微架构，并与自有模型保持同步发展。

So the way we are gonna go do it is have a closed loop between our own MAI models and our silicon because I feel like that's what gives you the birthright to really do your own silicon, right, where you literally have designed the microarchitecture with what you're doing, and then you keep pace with your own models.

对我们来说，好消息是OpenAI有一个我们可以使用的项目。

In our case, the good news here is OpenAI has a program which we have access to.

所以，认为微软不会拥有某种程度的东西——什么

And so therefore, to think that Microsoft is not going to have something that's- What

你们能接触到什么程度的权限？

level of access do have to that?

全部权限。

All of it.

你只是

You just

能获得所有这些的知识产权吗？

get the IP for all of that?

那么，你们唯一没有的是消费级硬件？

So, the only IP you don't have is a consumer hardware?

就是这样。

That's it.

哦，哇。

Oh, wow.

好的。

Okay.

是的。

Yeah.

有意思。

Interesting.

是的。

Yeah.

顺便说一下，我们还给了他们大量知识产权来帮助他们起步，对吧？

And by the way, we gave them a bunch of IP as well to bootstrap them, right?

所以，这就是他们拥有大量资源的原因之一，因为我们共同建造了所有这些超级计算机。

So, this is one of the reasons why they had a mass because we built all these supercomputers together.

我们为他们建造了这些，他们理应从中受益。

We built it for them, and they benefited from it, rightfully so.

现在，即使他们在系统层面进行创新，我们也能获取所有成果。

And now, as they innovate even at the system level, we get access to all of it.

我们首先想要实例化他们为自己构建的内容。

And we first want to instantiate what they build for them.

但之后我们会进行扩展。

But then we'll extend it.

如果要我说，我对你问题的理解是：微软希望成为NVIDIA极速执行方面的绝佳合作伙伴，因为坦率地说，那支舰队就是生命本身。

And if anything, the way I think about your question is Microsoft wants to be a fantastic, I'll call it, speed of light execution partner for NVIDIA because quite frankly, that fleet is life itself.

显然，Jensen的利润率表现非常出色，但总拥有成本（TCO）涉及多个维度。

Mean, obviously Jensen is doing super well with his margins, but the TCO has many dimensions to it.

我希望在降低总拥有成本方面做得非常出色。

And I want to be great at that TCO.

除此之外，我还希望能够真正与OpenAI系和MAI系以及系统设计合作，因为我们在这两端都拥有知识产权。

On top of that, I want to be able to sort of really work with the OpenAI lineage and the MAI lineage and the system design, knowing that we have the IP rights on both ends.

说到权利，有件事——你前几天在采访中提到，根据你们与OpenAI达成的新协议，我们拥有相关权利。

Speaking of rights, one thing, you had an interview a couple of days ago where you said that we have rights to the new agreement you made with OpenAI.

你们拥有权利，即对OpenAI进行的无状态API调用的专有权。

You have rights, the exclusivity to the stateless API calls that OpenAI makes.

我们之前对于是否存在任何状态感到有些困惑。

And we were sort of confused about if there's any state whatsoever.

我的意思是，你刚才提到所有这些即将出现的复杂工作负载将需要内存、数据库、存储等等。

I mean you were just mentioning a second ago that all these complicated workloads that are coming up are going to require memory and databases and storage and so forth.

那么现在ChatGPT在会话中存储内容，这还是无状态的吗？

And is that now not stateless of ChatGPT storing stuff on session?

所以，这就是原因所在。

So, that's the reason why.

因此，这个决定，我们在业务和战略上做出的决策，同时也是为了适应OpenAI在获取计算资源时所需的灵活性。本质上，OpenAI既有PaaS业务也有SaaS业务。

So, the thing, the business, the strategic decision we made and also accommodating for the flexibility OpenAI needed in order to be able to procure compute for Essentially, of OpenAI having a PaaS business and a SaaS business.

SaaS业务就是ChatGPT。

SaaS business is ChatGPT.

他们的PaaS业务是其API。

Their PaaS business is their API.