Today we are interviewing Satya Nadella.  

"We" being me and Dylan Patel, who is  founder of SemiAnalysis. Satya, welcome. 

Thank you. It's great. Thanks  for coming over to Atlanta. 

Thank you for giving us the  tour of the new facility. 

It's been really cool to see. Absolutely. 

Satya and Scott Guthrie, Microsoft's  EVP of Cloud and AI, give us a tour of  

their brand new Fairwater 2 data center,  the current most powerful in the world. 

We've tried to 10x the training  capacity every 18 to 24 months. 

So this would effectively be a 10x  increase from what GPT-5 was trained with. 

So to put it in perspective in the number of  optics, the network optics in this building  

is almost as much as all of Azure across all  our data centers two and a half years ago. 

It's got like five million network connections. You've got all this bandwidth between different  

sites in a region and between the two regions. So is this like a big bet on scaling in the  

future, that you anticipate in the future that  there's going to be some huge model that will  

require two whole different regions to train? The goal is to be able to aggregate these  

flops for a large training job and then  put these things together across sites. 

The reality is you'll use it for training and  then you'll use it for data gen, you'll use it  

for inference in all sorts of ways. It's not like it's going to be used  

only for one workload forever. Fairwater 4, which you're going  

to see under construction nearby, will also be on  that one petabit network so that we can actually  

link the two at a very high rate. Then we do the AI WAN connecting to  

Milwaukee where we have multiple  other Fairwaters being built. 

Literally you can see the model  parallelism and the data parallelism. 

It's kind of built for, essentially, the  training jobs, the super pods across this campus. 

And then with the WAN, you can  go to the Wisconsin data center. 

You literally run a training job  with all of them getting aggregated. 

What we're seeing right here is a cell  with no servers in it yet, no racks. 

How many racks are in a cell? We don't necessarily share that per se, but… 

That’s the reason I ask. You'll see upstairs. 

I'll start counting. You can start counting.  

We'll let you start counting. How many cells are there in this building? 

That part also I can't tell you. Well, division is easy, right? 

My God, it's kind of loud. Are you looking at this like,  

"Now I see where my money is going." It's like, "I run a software company.  

Welcome to the software company." How big is the design space once  

you've decided to use the GB200s and the NVLink? How many other decisions are there to be made? 

There is coupling from the model architecture  to what is the physical plan that's optimized. 

And it's also scary in that sense, which is,  there's going to be a new chip that'll come  

out. Take Vera Rubin Ultra. That's going to have  power density that's going to be so different,  

with cooling requirements that  are going to be so different. 

So you kind of don't want to  just build all to one spec. 

That goes back a little bit  to the dialogue we'll have,  

which is that you want to be scaling in time as  opposed to scale once and then be stuck with it. 

When you look at all the past technological  transitions—whether it be railroads or the  

Internet or replaceable parts, industrialization,  the cloud, all of these things—each revolution  

has gotten much faster in the time it  goes from technology discovered to ramp  

and pervasiveness through the economy. Many folks who have been on Dwarkesh's  

podcast believe this is the final  technological revolution or transition,  

and that this time is very, very different. At least so far in the markets, in three years  

we've already skyrocketed to hyperscalers  doing $500 billion of capex next year,  

which is a scale that's unmatched to  prior revolutions in terms of speed. 

The end state seems to be quite different. Your framing of this seems quite  

different from what I would call the  "AI bro" who's like, "AGI is coming." 

I'd like to understand that more. I start with the excitement that I  

also feel for the idea that maybe after  the Industrial Revolution this is the  

biggest thing. I start with that premise. But  at the same time, I'm a little grounded in the  

fact that this is still early innings. We've built some very useful things,  

we're seeing some great properties,  these scaling laws seem to be working. 

I'm optimistic that they'll continue to work. Some of it does require real science  

breakthroughs, but it's also a lot  of engineering and what have you. 

That said, I also sort of take the view  that even what has been happening in the  

last 70 years of computing has also  been a march that has helped us move. 

I like one of the things that Raj  Reddy has as a metaphor for what AI is. 

He's a Turing Award winner at CMU. He had  this, even pre-AGI. He had this metaphor  

for AI, it should either be a guardian angel  or a cognitive amplifier. I love that. It's a  

simple way to think about what this is. Ultimately, what is its human utility? 

It is going to be a cognitive  amplifier and a guardian angel. 

If I view it that way, I view it as a tool. But then you can also go very mystical about  

it and say this is more than a tool. It does all these things, which only  

humans did before so far. But that has been the case  

with many technologies in the past. Only humans did a lot of things,  

and then we had tools that did them. We don't have to get wrapped up in the definition  

here, but one way to think about it is, maybe  it takes five years, ten years, twenty years. 

At some point, eventually a machine is producing  Satya tokens, and the Microsoft board thinks that  

Satya tokens are worth a lot. How much are you wasting of  

this economic value by interviewing Satya? I could not afford the API costs of Satya tokens. 

Whatever you want to call it, are the  Satya tokens a tool or an agent, whatever. 

Right now, if you have models that cost on the  order of dollars or cents per million tokens,  

there's just an enormous room  for margin expansion there,  

where a million tokens of Satya are worth a lot. Where does that margin go and what level of that  

margin is Microsoft involved  in is the question I have. 

In some sense this goes back again to,  essentially, what's the economic growth  

picture going to really look like? What's the firm going to look like? 

What's productivity going to look like? That to me is where, again, if the Industrial  

Revolution created… After 70 years of diffusion  is when you started seeing the economic growth. 

That's the other thing to remember. Even if the tech is diffusing fast this  

time around, for true economic growth to appear  it has to diffuse to a point where the work,  

the work artifact, and the workflow has to change. So that's one place where I think the change  

management required for a corporation to truly  change is something we shouldn't discount. 

Going forward, do humans and the tokens they  produce get higher leverage, whether it's the  

Dwarkesh or the Dylan tokens of the future? Think about the amount of technology. 

Would you be able to run SemiAnalysis  or this podcast without technology? 

No chance, at the scale that you have  been able to achieve, there’s no chance. 

So the question is, what's that scale? Is it going to be 10x'ed with something  

that comes through? Absolutely. Therefore,  whether you're ramped to some revenue number  

or you're ramped to some audience number or what  have you, that I think is what's going to happen. 

The point is, what took 70 years, maybe  150 years for the Industrial Revolution,  

may happen in 20 years, 25 years. I would love to compress what happened  

in 200 years of the Industrial Revolution  into a 20-year period, if we're lucky. 

Microsoft historically has been  perhaps the greatest software company,  

the largest software-as-a-service company. You've gone through a transition in the past  

where you used to sell Windows licenses  and disks of Windows or Microsoft,  

and now you sell subscriptions to 365. As we go from that transition to where  

your business is today, there's also  a transition going on after that. 

Software-as-a-service has incredibly  low incremental cost per user. 

There's a lot of R&D, there's a  lot of customer acquisition costs. 

This is sort of why, not Microsoft, but  the SaaS companies have underperformed  

massively in the markets, because the COGS of  AI is just so high, and that just completely  

breaks how these business models work. How do you, as perhaps the greatest  

software-as-a-service company, transition  Microsoft to this new age where COGS matters a lot  

and the incremental cost per user is different? Because right now you're charging like, "Hey,  

it's 20 bucks for Copilot." It's a great question because in some  

sense with the business models themselves,  the levers are going to remain similar. 

If you look at the menu of models starting from  consumer all the way, there will be some ad unit,  

there will be some transaction,  there will be some device gross  

margin for somebody who builds an AI device. There will be subscriptions, consumer and  

enterprise, and then there'll be consumption. So I still think those are all the meters. 

To your point, what is a subscription? Up to now, people like subscriptions  

because they can budget for them. They are essentially entitlements to  

some consumption rights that come  encapsulated in a subscription. 

So I think that in some sense  becomes a pricing decision. 

How much consumption you are entitled to is,  if you look at all the coding subscriptions,  

kind of what they are, right? Then you have the pro tier,  

the standard tier, and what have you. So I think that's how the pricing  

and the margin structures will get tiered. The interesting thing is that at Microsoft,  

the good news for us is we are in  that business across all those meters. 

At a portfolio level, we pretty much  have consumption, subscriptions,  

to all of the other consumer levers as well. I think time will tell which of these  

models make sense in what categories. One thing on the SaaS side, since you  

brought it up, which I think a lot about. Take Office 365 or Microsoft 365. 

Having a low ARPU is great, because  here's an interesting thing. 

During the transition from server to cloud, one  of the questions we used to ask ourselves is,  

"Oh my God, if all we did was just basically  move the same users who were using our Office  

licenses and our Office servers at the  time to the cloud, and we had COGS,  

this is going to not only shrink our margins but  we'll be fundamentally a less profitable company." 

Except what happened was the move to the  cloud expanded the market like crazy. 

We sold a few servers in  India, we didn't sell much. 

Whereas in the cloud suddenly  everybody in India also could  

afford fractionally buying servers, the IT cost. In fact, the biggest thing I had not realized,  

for example, was the amount of money people were  spending buying storage underneath SharePoint. 

In fact, EMC's biggest segment may have  been storage servers for SharePoint. 

All that sort of dropped in the  cloud because nobody had to go buy. 

In fact, it was working capital,  meaning basically, it was cash flow out. 

So it expanded the market massively. So this AI thing will be that. 

If you take coding, what we built with  GitHub and VS Code over decades, suddenly  

the coding assistant is that big in one year. That I think is what's going to happen as well,  

which is the market expands massively. There’s a question of, the market will  

expand, but will the parts of the revenue that  touch Microsoft expand? Copilot is an example.  

If you look earlier this year, according  to Dylan's numbers, GitHub Copilot revenue  

was like $500 million or something like  that and there were no close competitors. 

Whereas now you have Claude  Code, Cursor, and Copilot  

with around similar revenue, around a billion. Codex is catching up around $700–800 million. 

So the question is, across all the surfaces that  Microsoft has access to, what is the advantage  

that Microsoft's equivalents of Copilot have? By the way, I love this chart. 

I love this chart for so many reasons. One is we're still on the top. 

Second is all these companies that are listed  here are all companies that have been born in  

the last four or five years. That to me is the best sign. 

You have new competitors,  new existential problems. 

When you say, who's it now? Claude's going to kill you,  

Cursor is going to kill you, it's not boreland.  Thank God. That means we are in the right  

direction. This is it. The fact that we went from  nothing to this scale is the market expansion. 

This is like the cloud-like stuff. Fundamentally, this category of  

coding and AI is probably going to  be one of the biggest categories. 

It is the software factory category. In fact, it may be bigger than knowledge work. 

I want to keep myself open-minded about it. We're going to have tough competition. 

That's your point, which is a great one. But I'm glad we have parlayed  

what we had into this and now we have to compete. On the competing side, even in the last quarter we  

just finished, we did our quarterly announcement  and I think we grew from 20 to 26 million subs. 

I feel good about our sub growth and  where the direction of travel on that is. 

But the more interesting thing that has happened  is, guess where all the repos of all these other  

guys who are generating lots and lots of code go?  They go to GitHub. GitHub is at an all-time high  

in terms of repo creation, PRs, everything. In some sense we want to keep that open,  

by the way. That means we want to have that. 

We don't want to conflate  that with our own growth. 

Interestingly enough, we are getting  one developer joining GitHub a second  

or something, that is the stat, I think. And 80% of them just fall into some GitHub  

Copilot workflow, just because there are. By the way, many of these things will even  

use some of our coding code review agents, which  are by default on, just because you can use it. 

We'll have many, many structural shots at this. The thing that we're also going  

to do is what we did with Git. The primitives of GitHub, starting with Git,  

to issues, to actions, these are powerful, lovely  things because they kind of are all built around  

your repo. We want to extend that. Last week  at GitHub Universe, that's kind of what we did. 

We said Agent HQ was the conceptual thing  that we said we're going to build out. 

This is where, for example, you  have a thing called Mission Control. 

You go to Mission Control, and now I can fire off. Sometimes I describe it as the cable TV of all  

these AI agents because I'll have, essentially  packaged into one subscription, Codex, Claude,  

Cognition stuff, anyone's agents,  Grok, all of them will be there. 

So I get one package and then I can literally  go issue a task and steer them so they'll all  

be working in their independent branches. I can  monitor them. I think that's going to be one of  

the biggest places of innovation, because right  now I want to be able to use multiple agents. 

I want to be able to then digest  the output of the multiple agents. 

I want to be able to then  keep a handle on my repo. 

If there's some kind of a heads-up display  that needs to be built and then for me to  

quickly steer and triage what the coding agents  have generated, that to me, between VS Code,  

GitHub, and all of these new primitives we'll  build as Mission Control with a control plane. 

Observability… Just think about everyone  who is going to deploy all this. 

It will require a whole host of observability  of what agent did what at what time to what  

code base. I feel that's the opportunity. At  the end of the day your point is well taken,  

which is we better be competitive and innovate. If we don't, we will get toppled. 

But I like the chart, at least as long as  we're on the top, even with competition. 

The key point here is sort of  that GitHub will keep growing  

regardless of whose coding agent wins. But that market only grows at say 10,  

15, 20%, which is way above GDP. It's a great  compounder. But these AI coding agents have  

grown from say $500 million run rate at the end  of last year—which was just GitHub Copilot—to  

now where the current run rate across GitHub  Copilot, Claude Code, Cursor, Cognition, Windsurf,  

Replit, OpenAI Codex… That’s run rating at  $5–6 billion now for the Q4 of this year.  

That's 10x. When you look at the TAM of software  agents, is it the $2 trillion of wages you pay  

people, or is it something beyond that? Because every company in the world will  

now be able to develop software more? No question Microsoft takes a slice of that. 

But you've gone from near 100%,  or certainly way above 50%,  

to sub-25% market share in just one year. What is the confidence that people can get  

that Microsoft will keep winning? It goes back a little bit, Dylan,  

to that there's no birthright here,  that we should have any confidence  

other than to say we should go innovate. Knowing the lucky break we have, in some sense,  

is that this category is going to be a lot  bigger than anything we had high share in. 

Let me say it that way. You could say we had high  

share in VS Code, we had high share in the  repos with GitHub, and that was a good market. 

But the point is that even having a decent  share in what is a much more expansive market… 

You could say we had a high share  in client-server server computing. 

We have much lower share than that in hyperscale. But is it a much bigger business? By orders of  

magnitude. So at least it's existence proof that  Microsoft has been okay even if our share position  

has not been as strong as it was, as long as  the markets we are competing in are creating  

more value. And there are multiple winners. That's  the stuff. But I take your point that ultimately  

it all means you have to get competitive. I  watch that every quarter. That’s why I'm very  

optimistic about what we're going to do with  Agent HQ, turning GitHub into a place where  

all these agents come. As I said, we'll have  

multiple shots on goal on there. It need not be… Some of these guys  

can succeed along with us, so it doesn't need  to be just one winner and one subscription. 

I guess the reason to focus on this  question is that it's not just about GitHub,  

but fundamentally about Office and all  the other software that Microsoft offers. 

One vision you could have about how AI  proceeds is that the models are going  

to keep being hobbled and you'll need this  direct visible observability all the time. 

Another vision is that over time these models  which are now doing tasks that take two minutes,  

in the future, they'll be doing  tasks that take 10, 30 minutes. 

In the future, maybe they're doing  days worth of work autonomously. 

Then the model companies are charging thousands  of dollars maybe for access to, really,  

a coworker which could use any UI to communicate  with their human and migrate between platforms. 

If we’re getting closer to that, why  aren't the model companies that are  

just getting more and more profitable,  the ones that are taking all the margin? 

Why is the place where the scaffolding happens,  which becomes less and less relevant as the AI  

becomes more capable, going to be that important? That goes to Office as it exists now versus  

coworkers that are just doing knowledge work. That's a great point. Does all the value  

migrate just to the model? Or does it get split between  

the scaffolding and the model? I think that time will tell. 

But my fundamental point also is that  the incentive structure gets clear. 

Let’s take information work, or take even coding. Already in fact, one of my favorite settings  

in GitHub Copilot is called  auto, which will just optimize. 

In fact I buy a subscription and the auto  one will start picking and optimizing  

for what I am asking it to do. It could even be fully autonomous. 

It could arbitrage the tokens available  across multiple models to go get a task done. 

If you take that argument, the  commodity there will be models. 

Especially with open source models, you  can pick a checkpoint and you can take a  

bunch of your data and you're seeing it. I think all of us will start, whether  

it's from Cursor or from Microsoft,  seeing some in-house models even. 

And then you'll offload most of your tasks to it. So one argument is if you win the  

scaffolding—which today is dealing  with all the hobbling problems or the  

jaggedness of these intelligence problems,  which you kind of have to—if you win that,  

then you will vertically integrate yourself  into the model just because you will have  

the liquidity of the data and what have you. There are enough and more checkpoints that  

are going to be available. That's the other  thing. Structurally, I think there will always  

be an open source model that will be fairly  capable in the world that you could then use,  

as long as you have something that you can use  that with, which is data and a scaffolding. 

I can make the argument that if you're a  model company, you may have a winner's curse. 

You may have done all the hard work,  done unbelievable innovation, except  

it's one copy away from that being commoditized. Then the person who has the data for grounding and  

context engineering, and the liquidity of data  can then go take that checkpoint and train it. 

So I think the argument can be made both ways. Unpacking what you said,  

there's two views of the world. One is that there are so many different  

models out there. Open source exists. There  will be differences between the models that  

will drive some level of who wins and who doesn't. But the scaffolding is what enables you to win. 

The other view is that,  actually, models are the key IP. 

And everyone's in a tight race and there's  some, "Hey, I can use Anthropic or OpenAI." 

You can see this in the revenue charts. OpenAI's revenue started skyrocketing once they  

finally had a code model with similar capabilities  to Anthropic, although in different ways. 

There's the view that the model companies  are the ones that garner all the margin. 

Because if you look across this  year, at least at Anthropic,  

their gross margins on inference went from well  below 40% to north of 60% by the end of the year. 

The margins are expanding there despite  more Chinese open source models than ever. 

OpenAI is competitive, Google is  competitive, X/Grok is now competitive. 

All these companies are now competitive, and  yet despite this, the margins have expanded at  

the model layer significantly. How do you think about that? 

It's a great question. Perhaps a few years ago  people were saying, "Oh, I could just wrap a  

model and build a successful company." That has probably gotten debunked just  

because of the model capabilities,  and the tools used, in particular. 

But the interesting thing is, when I look at  Office 365, let's take even this little thing we  

built called Excel Agent. It's interesting.  Excel Agent is not a UI-level wrapper. 

It's actually a model that is in the middle tier. In this case, because we have all the IP from  

the GPT family, we are taking that and  putting it into the core middle tier of  

the Office system to teach it what it means to  natively understand Excel, everything in it. 

It's not just, "Hey, I just have  a pixel-level understanding." 

I have a full understanding of  all the native artifacts of Excel. 

Because if you think about it, if I'm  going to give it some reasoning task,  

I need to even fix the reasoning mistakes I make. That means I need to not just see the pixels,  

I need to be able to see, "Oh, I got that  formula wrong," and I need to understand that. 

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. 

I'm giving it essentially a markdown to  teach it the skills of what it means to  

be a sophisticated Excel user. It's a weird thing that it goes  

back a little bit to the AI brain. You're building not just Excel,  

business logic in its traditional sense. 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. In some sense, Excel will come with an analyst  

bundled in and with all the tools used. That's the type of stuff that will get  

built by everybody. So even for the  

model companies, they’ll have to compete. If they price stuff high, guess what, if I'm a  

builder of a tool like this, I'll substitute you. I may use you for a while. 

So as long as there's competition…  There's always a winner-take-all thing. 

If there's going to be one model that  is better than everybody else with  

massive distance, yes, that's a winner-take-all. But as long as there's competition where there are  

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. At Microsoft, the way I look at it is  

that we are going to be in the hyperscale  business, which will support multiple models. 

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

Essentially, I think of ourselves as  having a frontier-class model that we can  

use and innovate on with full flexibility. And we'll build our own models with MAI. 

So we will always have a model level. And then we'll build—whether it's in security,  

whether it's in knowledge work, whether it's  in coding, or in science—our own application  

scaffolding, which will be model-forward. 

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

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

But before we move onto those topics,  I still wonder whether this is not  

forward-looking on AI capabilities, where  you're imagining models like they exist today. 

It takes a screenshot of your screen, but it can't  look inside each cell and what the formula is. 

I think the better mental model here is  just imagining that these models will be  

able to use a computer as well as a human. A human knowledge worker who is using Excel  

can look into the formulas, can use alternative  software, can migrate data between Office 365  

and another piece of software if the  migration is necessary, et cetera. 

That's kind of what I'm saying. But if that's the case, then the  

integration with Excel doesn't matter that much. No, no, don't worry about the Excel integration. 

After all, Excel was built as a tool for  analysts. Great. So whoever is this AI that  

is an analyst should have tools that they can use. They have the computer. Just the way a human can  

use a computer. That's their tool. The tool is the computer. So all  

I’m saying is that 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. 

Just to make sure we're talking  about the same thing, is it a thing  

that a human like me using Excel… No, it's completely autonomous. So  

we should now maybe lay out what I  think the future of the company is. 

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

In fact, in the future I'll even have a  Copilot, and that Copilot will also have agents. 

But it's still me steering everything, and  everything is coming back. That's one world.  

The second world is the company just literally  provisions a computing resource for an AI agent,  

and that is working fully autonomously. That fully autonomous agent will have  

essentially an embodied set of those  same tools that are available to it. 

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. 

In fact, I kind of look at  it and say that our business,  

which today is an end-user tools business,  will become essentially an infrastructure  

business in support of agents doing work. It's another way to think about it. 

In fact, all the stuff we built underneath  M365 still is going to be very relevant. 

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. It's a new infrastructure. 

To make sure I understand, you're  saying theoretically a future AI that  

has actual computer use—which all these model  companies are working on right now—could use,  

even if it's not partnered with Microsoft  or under our umbrella, Microsoft software. 

But you're saying, if you're working  with our infrastructure, we're going  

to give them lower-level access that makes it  more efficient for you to do the same things  

you could have otherwise done anyways? 100%. What happened is we had servers,  

then there was virtualization,  and then we had many more servers. 

That's another way to think about this. Don't think of the tool as the end thing. 

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

That entire substrate is the  bootstrap for the AI agent as well,  

because the AI agent needs a computer. In fact, one of the fascinating  

things where we're 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 365. 

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

That's why we're going to have essentially an  end-user computing infrastructure business,  

which is going to just keep growing because it's  going to grow faster than the number of users. 

That's one of the other questions people ask me,  "Hey, what happens to the per-user business?" 

At least the early signs maybe, the way  to think about the per-user business  

is not just per user, it's per agent. And if you say it's per user and per agent,  

the key is what's the stuff  to provision for every agent? 

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

All those things, observability and  so on, are the management layers. 

That's all going to get baked into that. The way to frame it—at least the way I currently  

think about it and 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. 

But at the same time, they're also  training these models to do migration. 

Because that is probably the most immediately  valuable thing: converting mainframe-based  

systems to standard cloud systems, converting  Excel databases into real databases with SQL,  

or converting what is done in Word and Excel  to something that is more programmatic and  

more efficient in a classical sense  that can be done by humans as well. 

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

That seems to be what everyone is going  to do with AI, for the next few years at  

least, to massively drive value. How does Microsoft fit into that  

if the models can utilize the tools  themselves to migrate to something? 

Yes, Microsoft has a leadership position  in databases and in storage and in all  

these other categories, but the use  of an Office ecosystem is going to  

be significantly less just like the use of a  mainframe ecosystem could be potentially less. 

Now mainframes have grown for the last two decades  actually, even though no one talks about them  

anymore. They've still grown. 100%, I agree with that. 

How does that flow? At the end of the day, there  

is going to be a significant amount of time where  there's going to be a hybrid world, because people  

are going to be using the tools that are going  to be working with agents that have to use tools,  

and they have to communicate with each other. What's the artifact I generate that  

then a human needs to see? All of these things will be real  

considerations in any place, the outputs, inputs. I don't think it'll just be about,  

"Oh, I migrated off." The bottom line is that I  

have to live in this hybrid world. But that doesn't fully answer your  

question because there can be a real new  efficient frontier where it's just agents  

working with agents and completely optimized. Even when agents are working with agents,  

what are the primitives that are needed? Do you need a storage system? 

Does that storage system need to have e-discovery? Do you need to have observability? 

Do you need to have an identity system  that is going to use multiple models  

with all having one identity system? These are all the core underlying rails  

we have today for what are the  Office systems or what have you. 

And that's what we will have in the future as  well. You've talked about databases. I mean man,  

I would love all of Excel  to have a database backend. 

I would love for all that to happen immediately. And that database is a good database. 

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

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 agentic world, that will grow the  

underlying infrastructure business. It happens that the consumption  

of that is all being driven by agents. You could say all that is just-in-time generated  

software by a model company. That could also be  true. We will be one such model company too. We  

will build in... The competition could be that we  will build a model plus all the infrastructure and  

provision it, and then there will be competition  between a bunch of those folks who can do that. 

Speaking of model companies, you say not  only will you have the infrastructure,  

you'll have the model itself. Right now, Microsoft AI's most  

recent model that was released two  months ago is 36 in Chatbot Arena. 

You obviously have the IP rights to OpenAI. To the extent you agree with that,  

it seems to be behind. Why is that the case, especially given  

the fact that you theoretically have the right to  fork OpenAI's monorepo or distill their models,  

especially if it's a big part of your strategy  that you need to have a leading model company? 

First of all, we are absolutely  going to use the OpenAI models  

to the maximum across all of our products. That's the core thing that we're going to continue  

to do all the way for the next seven years,  and not just use it but then add value to it. 

That's where the analyst and this Excel  agent, these are all things that we will  

do where we'll do RL fine-tuning. We'll do some mid-training runs  

on top of a GPT family where we have  unique data assets and build capability. 

With the MAI model, the way that I think  we’re going to think about it is that the  

good news here with the new agreement is  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. 

But at the same time, we're also  going to use this time to be smart  

about how to use both these things. That means we will, on one end,  

be very product-focused, and on the  other end, be very research-focused. 

Because we have access to the GPT family, the last  thing I want to do is use my flops in a way that  

is just duplicative and doesn't add much value. 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 is at number nine in the image arena. 

We're using it both for cost  optimization, it's on Copilot,  

it's in Bing, and we're going to use that. We have an audio model in Copilot. 

It's got personality and what have you. We optimized it for our product. So we will  

do those. Even on the LMArena, we started  on the text one and it debuted at like 13. 

By the way, it was done  only on around 15,000 H100s. 

It was a very small model. So it was, again, to prove out  

the core capability, the instruction  following, and everything else. 

We wanted to make sure we could  match what was state of the art. 

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

The next thing we will do is an omni-model where  we will take the work we have done in audio,  

what we have done in image,  and what we have done in text. 

That will be the next pit stop on the MAI side. So when I think about the MAI roadmap, we are  

going to build a first-class  superintelligence team. 

We are going to continue to drop, and  do it in the open, some of these models. 

They will either be used in our products,  because they're going to be latency-friendly,  

cost-friendly, or what have you, or  they'll have some special capability. 

And we will do real research in order to  be ready for the next five, six, seven,  

eight breakthroughs that are all needed on  this march towards superintelligence—while  

exploiting the advantage we have of having the  GPT family that we can work on top of as well. 

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

What does Microsoft do to make sure they  are leading, or have a leading AI lab? 

Today, 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. You've seen Meta spend north  

of $20 billion on talent. You've seen Anthropic poach the entire  

Blueshift reasoning team from Google last year. You've seen Meta poach a large reasoning and  

post-training team from Google more recently. These sorts of talent wars are  

very capital intensive. Arguably, if you're spending  

$100 billion on infrastructure, you should  also spend X amount of money on the people  

using the infrastructure so that they're more  efficiently making these new breakthroughs. 

What confidence can one get that Microsoft  will have a team that's world-class that  

can make these breakthroughs? Once you decide to turn on  

the money faucet—you're being a bit capital  efficient right now, which is smart it seems,  

to not waste money doing duplicative work—but once  you decide you need to, how can one say, "Oh yeah,  

now you can shoot up to the top five model?" 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 assembled. 

We have Mustafa coming in, we have Karen. We have Amar Subramanya who did a lot of the  

post-training at Gemini 2.5 who's at Microsoft. Nando, who did a lot of the multimedia  

work at DeepMind, is there. We're going to build a world-class team. 

In fact, later this week even, Mustafa  will publish something with a little  

more clarity on what our lab is going to go do. The thing that I want the world to know, perhaps,  

is that we are going to build the infrastructure  that will support multiple models. 

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. That's one  job. Secondly, in our own model capability,  

we will absolutely use the OpenAI model in our  products and we'll start building our own model. 

And we may—like in GitHub Copilot where Anthropic  is used—even have other frontier models that are  

going to be wrapped into our products, as well. I think that’s how 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. 

We'll start 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. 

There's a question going forward.  Right now, we have models that have  

this distinction between training and inference. One could argue that there's a smaller and smaller  

difference between the different models. Going forward, if you're really  

expecting something like human-level  intelligence, humans learn on the job. 

If you think about your last 30 years,  what makes Satya tokens so valuable? 

It's the last 30 years of wisdom and  experience you've gained in Microsoft. 

We will eventually have models, if they  get to human level, which will have  

this ability to continuously learn on the job. 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. And unlike humans, they can amalgamate  

their learnings to that model. So there's this sort of continuous  

learning exponential feedback loop, which almost  looks like a sort of intelligence explosion. 

If that happens and Microsoft isn't  the leading model company by that time…  

You're saying that well, we substitute one model  for another, et cetera. Doesn’t that then matter  

less? Because it's like this one model  knows how to do every single job in  

the economy, the others in the long tail don't. Your point, 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 does continuous  

learning, that's game set match and you stop shop. The reality that at least I see is that in the  

world today, for all the dominance of  any one model, that is not the case. 

Take coding, there are multiple models. In fact, everyday it's less the case. 

There is not one model that  is getting deployed broadly. 

There are multiple models that are getting  deployed. It's like databases. It's always  

the thing, "Can one database be the one that  is just used everywhere?" Except it's not.  

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

I think that there are going to be some  network effects of continual learning—I  

call it data liquidity—that any one model has. Is it going to happen in all domains? I don't  

think so. Is it going to happen in all geos?  I don't think so. Is it going to happen in all  

segments? I don't think so. It'll happen in all  categories at the same time? I don't think so.  

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

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

and at the scaffolding layer, and then  being able to compose these things not  

just as a vertical stack, but to be able to  compose each thing for what its purpose is. 

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

If you do that, what if you fall behind? In fact, all the infrastructure  

you built will be a waste. You kind of need to build an  

infrastructure that's capable of supporting  multiple families and lineages of models. 

Otherwise the capital you put in, which is  optimized for one model architecture, means  

you're one tweak away, some MoE-like breakthrough  that happens, and your entire network topology  

goes out of the window. That's a scary thing.  Therefore you kind of want the infrastructure  

to support whatever may come in your own model  family and other model families. You've got to  

be open. If you're serious about the hyperscale  business, you've got to be serious about that. 

If you're serious about being a model  company, you have to basically say,  

"What are the ways people can do things on top of  the model so that I can have an ISV ecosystem?" 

Unless I'm thinking I'll own every  category, that just can't be that. 

Then you won't have an API business and that, by  definition, will mean you'll never be a platform  

company that's successfully deployed everywhere. Therefore the industry structure is such that  

it will really force people to specialize. In that specialization, a company like Microsoft  

should compete in each layer by its merits, but  not think that this is all about the road to game  

set match, where I just compose vertically  all these layers. That just doesn't happen. 

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

You were the earliest in 2023, so you went  out there, you acquired all the resources  

in terms of leasing data centers, starting  construction, securing power, everything. 

You guys were on pace to  beat Amazon in 2026 or 2027. 

Certainly by 2028 you were going to beat them. Since then, let’s call it, in the second half  

of last year, Microsoft did this big pause, 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. 

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

you guys are expanding like crazy. But there are sites that you just stopped  

working on. Why did you do this? This goes back a little bit to,  

what is the hyperscale business all about? One of the key decisions we made was that  

if we're going to build out Azure to  be fantastic for all stages of AI—from  

training to mid-training to data gen to  inference—we just need fungibility of the fleet. 

So that entire thing caused us basically  not to go build a whole lot of capacity  

with a particular set of generations. Because the other thing you have to  

realize is that having up to now 10x'ed  every 18 months enough training capacity  

for the various OpenAI models, we realized  that the key is to stay on that path. 

But the more important thing is to have a balance,  to not just train, but to be able to serve these  

models all around the world. Because at the end of the day,  

the rate of monetization is what  will then allow us to keep funding. 

And then the infrastructure was going  to need us to support multiple models. 

So once we said that that's the case, we  just course-corrected to the path we're on. 

If I look at the path we're on, we  are doing a lot more starts now. 

We are also buying up as much managed  capacity as we can, whether it's to build,  

whether it's to lease, or even GPUs as a service. But we're building it for where we see the demand  

and the serving needs and our training needs. We didn't want to just be a hoster for one company  

and have just a massive book  of business with one customer. 

That's not a business, you should be  vertically integrated with that company. 

Given that OpenAI was going to be  a successful independent company,  

which is fantastic. It makes sense. And even  Meta may use third-party capacity, but ultimately  

they're all going to be first-party. For anyone who has large scale,  

they'll be a hyperscaler on their own. To me, it was to build out a hyperscale  

fleet and our own research compute. That's what  the adjustment was. So I feel very, very good. 

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

We just saw the GB200s, the GB300s are coming. By the time I get to Vera Rubin, Vera Rubin Ultra,  

the data center is going to look very different  because the power per rack, power per row,  

is going to be so different. The cooling requirements are  

going to be so different. That means I don't want to  

just go build out a whole number of gigawatts  that are only for a one-generation, one family. 

So I think the pacing matters, the  fungibility and the location matters,  

the workload diversity matters, customer diversity  matters and that's what we’re building towards. 

The other thing that we've learned  a lot is that every AI workload does  

require not only the AI accelerator, but  it requires a whole lot of other things. 

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

Therefore, we want to build out Azure as being  fantastic for the long tail of the 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. 

But that can't crowd out the rest of the  business, because we're not in the business  

of just doing five contracts with five customers  being their bare-metal service. That's not a  

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

What we have said is that we're in the  hyperscale business, which is at the end  

of the day a long tail business for AI workloads. 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. 

And that, I think, is the balance you see. Another question that comes around  

this whole fungibility topic. Okay, it's not where you want it, you would  

rather have it in a good population center, like  Atlanta. We're here. There's also the question of,  

how much does that matter as the horizon of AI  tasks grows? 30 seconds for a reasoning prompt,  

or 30 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. 

Why does it matter if it's location A, B, or C? It’s a great question. That's exactly it. In fact,  

that's one of the other reasons why we want to  think about what an Azure region looks like and  

what is the networking between Azure regions. This is where I think as the model capabilities  

evolve and the usage of these tokens evolves,  whether it's synchronously or asynchronously,  

you don't want to be out of position. Then on top of that, by the way,  

what are the data residency laws? There’s the entire EU thing, where  

we literally had to create an EU Data Boundary. That basically meant that you can't just roundtrip  

a call to wherever, even if it's asynchronous. Therefore you need to have maybe regional  

things that are high density, and  then the power costs and so on. 

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

One, for tokens per dollar per watt. What  are the economics? Overlay that with,  

what is the usage pattern? Usage pattern in terms  

of synchronous, asynchronous. But also what is the compute storage? 

Because the latencies may matter for certain  things. The storage better be there. 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. All of those considerations are  

what will shape the hyperscale business. Prior to the pause, what we had forecasted for  

you, by 2028 you were going to be 12–13 gigawatts. Now we're at nine and a half or so. 

But something that's even more relevant—and  I just want you to more concretely state that  

this is the business you don't want to be  in—is that Oracle's going from 1/5th your  

size to bigger than you by the end of 2027. While it's not a Microsoft-level quality of  

return on invested capital, they're  still making 35% gross margins. 

So the question is, maybe it's not Microsoft's  business to do this, but you've created a  

hyperscaler now by refusing this business, by  giving away the right of first refusal, et cetera. 

First of all, I don't want to take away  anything from the success Oracle has had in  

building their business and I wish them well. The thing that I think I've answered for you  

is that it didn't make sense for us to  go be a hoster for one model company  

with limited time horizon RPO. Let's just put it that way. 

The thing that you have to think through  is not what you do in the next five years,  

but what you do for the next 50. We made our set of decisions. 

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

We have a decent book of business. We wish them a lot of success. 

In fact, we are buyers of Oracle capacity.  We wish them success. But at this point,  

I think the industrial logic for what we are  trying to do is pretty clear, which is that it's  

not about chasing… First of all, I track, by  the way, your things whether it's AWS or Google  

and ours, which I think is super useful. But it doesn't mean I have to chase those. 

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

What is this book of business that  Microsoft uniquely can go clear, which  

makes sense for us to clear? That's what we'll do. I have a question even stepping back from this,  

I take your point that it's a better business  to be in, all else equal, to have a long tail  

of customers you can have higher margin from  rather than serving bare metal to a few labs. 

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

If we believe we're on the path to smarter  and smarter AIs, then why isn't the shape of  

the industry that the OpenAIs and Anthropics and  DeepMinds are the platform on which the long tail  

of enterprises are actually doing business? They need bare metal, but they are the platform. 

What is the long tail that  is directly using Azure? 

Because you want to use  the general cognitive core. 

But those models are all going to be available  on Azure, so any workload that says, "Hey,  

I want to use some open source model and an  OpenAI model," if you go to Azure Foundry today,  

you have all these models that you can  provision, buy PTUs, get a Cosmos DB,  

get a SQL DB, get some storage, get some compute. That's what a real workload looks like. 

A real workload is not just  an API call to a model. 

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

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

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

the hyperscale business. And it's not on  any one model, but all of these models. 

So if you want Grok plus, say, OpenAI plus  an open source model, come to Azure Foundry,  

provision them, build your application. Here is  a database. That's kind of what the business is. 

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

And that's the argument about  how much of that business you  

want to be in and not be in and what that is. It's a very different segment of the business,  

which we are in, and we also have limits to how  much of it is going to crowd out the rest of it. 

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

There are sort of two questions here. One is, why couldn't you just do both? 

The other one is, given our estimates  on what your capacity is in 2028,  

it's three and a half gigawatts lower. Sure, you could have dedicated that to  

OpenAI training and inference capacity,  but you could have also dedicated that  

to actually just running Azure, running  Microsoft 365, running GitHub Copilot. 

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

Or I may want to build it in a different location. I may want to build it in the UAE,  

I may want to build it in India,  I may want to build it in Europe. 

One of the things is, as I said, where we have  real capacity constraints right now, given  

the regulatory needs and the data sovereignty  needs, we've got to build all over the world. 

First of all, stateside capacity is super  important, and we want to build everything. 

But when I look out to 2030, I have a  global view of what is Microsoft's shape  

of business by first-party and third-party. 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. That's all going into my calculus. 

You're rightfully pointing out the pause,  but the pause was not done because we said,  

"Oh my God, we don't want to build that." We realized that we want to build what we want  

to build slightly differently by both workload  type as well as geo-type and timing as well. 

We'll keep ramping up our gigawatts, and the  question is at what pace and in what location. 

And how do I ride Moore's law on it, which is,  do I really want to overbuild three and a half  

in 2027 or do I want to spread that in 2027-28  knowing even… One of the biggest learnings we had  

even with Nvidia is that their pace increased  in terms of their migrations. That was a big  

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

In fact, Jensen's advice to me was two things. One is, get on the speed-of-light execution. 

That's why the execution in  this Atlanta data center.... 

I mean, it's like 90 days between when we  get it and to hand off to a real workload. 

That's real speed-of-light  execution on that front. 

I wanted to get good on that. And then that way I'm building  

each generation in scaling. And then every five years,  

you have something much more balanced. So it becomes literally like a flow  

for a large-scale industrial operation like  this where you're suddenly not lopsided,  

where you've 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, or it may not  

be great for inference because I can't serve,  even if it's all asynchronous, because Europe  

won't let me round-trip to Texas. So that's all of the things. 

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

You've announced deals with Iris  Energy, with Nebius, and Lambda Labs,  

and there's a few more coming as well. You're going out there and securing capacity  

that you're renting from the neoclouds  rather than having built it yourself. 

It's fine for us because now when you have line  of sight to demand, which can be served where  

people are building, it's great. In fact we will take leases,  

we will take build-to-suit, we'll even take  GPUs-as-a-service where we don't have capacity  

but we need capacity and someone else has that. And by the way, I would even sort of welcome every  

neocloud to just be part of our marketplace.  Because guess what? If they go bring their  

capacity into our marketplace, that customer who  comes through Azure will use the neocloud, which  

is a great win for them, and will use compute,  storage, databases, all the rest from Azure. 

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

You mentioned how this depreciating asset, in five  or six years, is 75% of the TCO of a data center. 

And Jensen is taking a 75% margin on that. 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. And when you look at where they are,  

Google's way ahead of everyone else. They've been doing it for the longest. 

They're going to make something like five  to seven million chips of their own TPUs. 

You look at Amazon and they're trying to make  three to five million [Lifetime shipment units]. 

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

You've had a program for just as long. What's going on with your internal chips? 

It’s a good question. A couple of things.  One is that the thing that is the biggest  

competitor for any new accelerator is kind  of even the previous generation of Nvidia. 

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

The bar I have, even for our own… By the  way, I was just looking at the data for  

Maia 200 which looks great, except that one  of the things that we learned even on the  

compute side… We had a lot of Intel, then  we introduced AMD, and then we introduced  

Cobalt. That's how we scaled it. We have good  existence proof of, at least in core compute,  

how to build your own silicon and then manage a  fleet where all three are at play in some balance. 

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

It makes sense because Nvidia is innovating  and it's the general-purpose thing. 

All models run on it and customer demand is there. Because if you build your own vertical thing,  

you better have your own model, which is either  going to use it for training or inference,  

and you have to generate your own demand  for it or subsidize the demand for it. 

So therefore you want to make  sure you scale it appropriately. 

The way we are going to do it is to have a close  loop between our own MAI models and our silicon,  

because I feel like that's what gives you  the birthright to do your own silicon,  

where you literally have designed the  microarchitecture with what you're doing,  

and then you keep pace with your own models. In our case, the good news here is that OpenAI  

has a program which we have access to. So therefore to think that Microsoft  

is not going to have something that's— What level of access do you have to that? 

All of it. You just get the IP for all of that? 

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

That's it. Oh, okay. Interesting. 

By the way, we gave them a bunch  of IP as well to bootstrap them. 

This is one of the reasons why they… Because  we built all these supercomputers together. 

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

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

And we first want to instantiate what they  build, for them, but then we'll extend it. 

So 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. Obviously Jensen's doing super well with his  

margins, but the TCO has many dimensions  to it and I want to be great at that TCO. 

On top of that, I want to be able to  really work with the OpenAI lineage  

and the MAI lineage and the system design,  knowing that we have the IP rights on both ends. 

Speaking of rights, you had an interview a  couple days ago where you said that in the  

new agreement you made with OpenAI you have  rights, the exclusivity, to the stateless  

API calls that OpenAI makes. We were sort of confused about  

if there's any state whatsoever. 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. Is that now not stateless if  

ChatGPT is storing stuff on sessions? That's the reason why. The strategic decision we  

made, and also accommodating for the flexibility  OpenAI needed in order to be able to procure  

compute for… Essentially think of OpenAI having  a PaaS business and a SaaS business. The SaaS  

business is ChatGPT. Their PaaS business is  their API. That API is Azure-exclusive. The  

SaaS business, they can run it anywhere. And they can partner with anyone they  

want to to build SaaS products? If they want a partner and that  

partner wants to use a stateless API, then Azure  is the place where they can get the stateless API. 

It seems like there's a way for them to build  the product together and it's a stateful thing… 

No, for even that they'll have to come to Azure. Again, this is done in the spirit of "what is it  

that we value as part of our partnership." And we made sure that, at the same time,  

we were good partners to OpenAI given  all the flexibility they needed. 

So for example, Salesforce wants to integrate  OpenAI. It's not through an API. They actually  

work together, train a model together  and deploy it on, let's say, Amazon now. 

Is that allowed or do they have to use your… For any custom agreement like that, they will have  

to come run it… There are some few exceptions,  the US government and so on, that we made,  

but other than that, they'd have to come to Azure. Stepping back, when we were walking back and forth  

through the factory, one of the things  you were talking about is that Microsoft,  

you can think of it as a software business, but  now it's really becoming an industrial business. 

There's all this capex,  there's all this construction. 

If you just look over the last two  years, your capex has sort of tripled. 

Maybe you extrapolate that forward, it actually  just becomes this huge industrial explosion. 

Other hyperscalers are taking loans. Meta  has done a $20 billion loan at Louisiana.  

They've done a corporate loan. It seems clear  everyone's free cash flow is going to zero,  

which I'm sure Amy is going to beat you up if  you even try to do that, but what's happening? 

I think the structural change is what  you're referencing, which is massive. 

I describe it as we are now a capital-intensive  business and a knowledge-intensive business. 

In fact, we have to use our knowledge to  increase the ROIC on the capital spend. 

The hardware guys have done a  great job of marketing Moore's Law,  

which I think is unbelievable and it's great. 

But if you even look at some of the stats I even  did in my earnings call, for a given GPT family,  

the software improvements of really throughput  in terms of tokens-per-dollar-per-watt that we're  

able to get quarter-over-quarter,  year-over-year, it’s massive. 

It's 5x, 10x, maybe 40x in some of these  cases, just because of how you can optimize. 

That's knowledge intensity coming  to bring out capital efficiency. 

That, at some level, is what we have to master. Some people ask me, what is the difference between  

a classic old-time hoster and a hyperscaler?  Software. Yes, it is capital intensive,  

but as long as you have systems know-how, software  capability to optimize by workload, by fleet... 

That's why when we say fungibility,  there's so much software in it. 

It's not just about the fleet. It's the ability to evict a  

workload and then schedule another workload. Can I manage that algorithm of scheduling around? 

That is the type of stuff that  we have to be world-class at. 

So yes, I think we'll still remain a software  company, but yes, this is a different business  

and we're going to manage. At the end of the day,  

the cash flow that Microsoft has allows  us to have both these arms firing well. 

It seems like in the short  term you have more credence  

on things taking a while, being more jagged. But maybe in the long term you think the people  

who talk about AGI and ASI are correct. Sam  will be right, eventually. I have a broader  

question about what makes sense for a hyperscaler  to do, given that you have to invest massively in  

this thing which depreciates over five years. So if you have 2040 timelines to the kind of  

thing that somebody like Sam anticipates  in three years, what is a reasonable  

thing for you to do in that world? There needs to be an allocation to,  

I'll call it, research compute. That needs to be done like you did R&D. 

That's the best way to even  account for it, quite frankly. 

We should think of it as just R&D expense and  you should say, "What's the research compute  

and how do you want to scale it?" Let's even say it's an order of  

magnitude scale in some period. Pick your thing, is it two years?  

Is it 16 months? What have you. That's sort of one  piece, which is table stakes, that's R&D expenses. 

The rest is all demand driven. Ultimately, you're allowed to build  

ahead of demand, but you better have a demand  plan that doesn't go completely off kilter. 

Do you buy… These labs are now projecting  revenues of $100 billion in 2027–28  

and they're projecting revenue to keep  growing at this rate of 3x, 2x a year… 

In the marketplace there's all kinds of  incentives right now, and rightfully so. 

What do you expect an independent lab that  is sort of trying to raise money to do? 

They have to put some numbers out there  such that they can actually go raise  

money so that they can pay their bills  for compute and what have you. And it's  

a good thing. Someone's going to take some risk  and put it in there, and they've shown traction. 

It's not like it's all risk without seeing  the fact that they've been performing,  

whether it's OpenAI, or whether it's Anthropic. So I feel great about what they've done,  

and we have a massive book of business with these  chaps. So therefore that's all good. But overall  

ultimately, there's two simple things. One is you have to allocate for R&D.  

You brought up talent. The talent for  AI is at a premium. You have to spend  

there. You've got to spend on compute. So in some sense researcher-to-GPU  

ratios have to be high. That is sort of what it takes  

to be a leading R&D company in this world. And that's something that needs to scale,  

and you have to have a balance sheet that allows  you to scale that long before it's conventional  

wisdom and so on. That's kind of one thing. But  the other is all about knowing how to forecast. 

As we look across the world, America  has dominated many tech stacks. 

The US owns Windows through Microsoft,  which is deployed even in China,  

that's the main operating system. Of course, there's Linux,  

which is open source, but Windows is deployed  everywhere in China on personal computers. 

You look at Word, it's deployed everywhere. You look at all these various technologies,  

it's deployed everywhere. And Microsoft and other  

companies have grown elsewhere. They're building data centers  

in Europe and in India and in all these other  places, in Southeast Asia and LatAm and Africa. 

In all of these different places, you're building  capacity. But this seems quite different. Today,  

the political aspect of technology,  of compute… The US administration  

didn't care about the dot-com bubble. It seems like the US administration,  

as well as every other administration  around the world, cares a lot about AI. 

The question is, we're sort of in a bipolar world,  at least with the US and China, but Europe and  

India and all these other countries are saying,  "No, we're going to have sovereign AI as well." 

How does Microsoft navigate the difference to the  90s—where there's one country in the world that  

matters, it's America, and our companies  sell everywhere and therefore Microsoft  

benefits massively—to a world where it is bipolar? Where Microsoft can't just necessarily have the  

right to win all of Europe or India or Singapore. There are actually sovereign AI efforts. 

What is your thought process here  and how do you think about this? 

It's a super critical piece. I think that  the key, key priority for the US tech sector  

and the US government is to ensure that  we not only do leading innovative work,  

but that we also collectively build  trust around the world on our tech stack. 

Because I always say the United States is just  an unbelievable place. It's just unique in  

history. It's 4% of the world's population,  25% of the GDP, and 50% of the market cap. 

I think you should think about  those ratios and reflect on it. 

That 50% happens because quite frankly the  trust the world has in the United States,  

whether it's its capital markets  or whether it's its technology and  

its stewardship of what matters at any  given time in terms of leading sector. 

If that is broken, then that's not  a good day for the United States. 

We start with that, which I think President  Trump gets, the White House, David Sacks,  

everyone really, I think, gets it. So therefore I applaud anything  

that the United States government and the  tech sector jointly does to, for example,  

put our own capital at risk, collectively  as an industry, in every part of the world. 

I would like the USG to take credit  for foreign direct investment by  

American companies all over the world. It's the least talked about, but the best  

marketing that the United States should be doing  is that it's not just about all the foreign direct  

investment coming into the United States, but the  most leading sector, which is these AI factories,  

are all being created all over the world. By  whom? By America and American companies. And  

so you start there, and then you even build  other agreements around it, which are around  

their continuity, their legitimate sovereignty  concerns, around whether it's data residency,  

for them to have real agency and  guarantees on privacy, and so on. 

In fact, our European  commitments are worth reading. 

We made a series of commitments to Europe on how  we will govern our hyperscale investment there  

such that the European Union and the  European countries have sovereignty. 

We're also building sovereign  clouds in France and in Germany. 

We have something called Sovereign Services on  Azure, which literally gives people key management  

services along with confidential computing,  including confidential computing in GPUs, which  

we've done great innovative work with Nvidia. So I feel very, very good about being able to  

build, both technically and through policy,  this trust in the American tech stack. 

How do you see this shaking out as you have  this network effect with continual learning  

and things on the model level? Maybe you have equivalent  

things at the hyperscaler level as well. Do you expect that the countries will say,  

"Look, it's clear one model or a couple models  are the best, and so we're going to use them,  

but we're going to have some laws around the  weights having to be hosted in our country"? 

Or do you expect that there will be this push so  that it has to be a model trained in our country? 

Maybe an analogy here is that semiconductors  are very important to the economy,  

and people would like to have their sovereign  semiconductors, but TSMC is just better. 

And semiconductors are so important to  the economy that you will just go to  

Taiwan and buy the semiconductors. You  have to. Will it be like that with AI? 

Ultimately, what matters is the use of AI  in their economy to create economic value. 

That's the diffusion theory, which  ultimately, it's not the leading sector,  

but it's the ability to use the leading technology  to create your own comparative advantage. 

So I think that will  fundamentally be the core driver. 

But that said, they will want continuity of that. So in some sense, that's one of the reasons why,  

I believe, there's always going to be a  check to "Hey, can this one model have all  

the runaway deployment?" That's why open source  

is always going to be there. There will be, by definition, multiple  

models. That'll be one way. That's one way for  people to sort of demand continuity and not have  

concentration risk, that’s another way to say it. And so you say, "Hey, I want multiple models,  

and then I want an open source." I feel that as long as that's there,  

every country will feel like, "Okay, I don't  have to worry about deploying the best model  

and broadly diffusing because I can always take  what is my data and my liquidity and move it  

to another model, whether it's open source  or from another country or what have you." 

Concentration risk and sovereignty, which is  really agency, those are the two things that  

will drive the market structure. The thing about this is that  

this doesn't exist for semiconductors. All refrigerators, cars have chips made in Taiwan. 

It didn't exist until now. Even then, if Taiwan is cut off,  

there are no more cars or no more refrigerators. TSMC Arizona is not replacing any  

real fraction of the production. The sovereignty is a bit of a scam, if you will. 

It's worthwhile having it, it's  important to have it, but it's not  

real sovereignty. We're a global economy. I think it’s kind of like saying, "Hey,  

at this point, we've not learned anything about  what resilience means and what one needs to do." 

Any nation state, including the United  States, at this point will do what it  

takes to be more self-sufficient on  some of these critical supply chains. 

So I, as a multinational company, have to  think about that as a first-class requirement. 

If I don't, then I'm not respecting what is in  the policy interests of that country long-term. 

I'm not saying they won't make  practical decisions in the short term. 

Absolutely, globalization can't just be rewound. All these capital investments cannot be made  

in a way, at the pace at which… But at the same  time, think about it, if somebody showed up in  

Washington and said, "Hey, we're not going to  build any semiconductor plants," they're going  

to be kicked out of the United States. The same thing is going to be  

true in every other country, too. So therefore we have to, as companies,  

respect what the lessons learned are, whether  it's that the pandemic woke us up or whatever. 

But nevertheless people are saying,  "Look, globalization was fantastic. 

It helped supply chains be  globalized and be super efficient. 

But there's such a thing called  resilience, and we want resilience." 

So therefore that feature will get built. At what pace, I think,  

is the point you are making. You can't snap your fingers  

and say all the TSMC plants now are all in  Arizona with all their capability. They're  

not going to be. But is there a plan? There  will be a plan. And should we respect that?  

Absolutely. So I feel that that’s the world. I want to meet the world where it is and on what  

it wants to do going forward, as opposed  to saying, "Hey, we have a point of view  

that doesn't respect your view." Just to make sure I understand,  

the idea here is that each country will want  some kind of data residency, privacy, et cetera. 

And Microsoft is especially privileged here  because you have relationships with these  

countries, you have expertise in setting  up these kinds of sovereign data centers. 

Therefore Microsoft is uniquely fit for a  world with more sovereignty requirements. 

I don't want to sort of describe it  as somehow we're uniquely privileged. 

I would just say I think of that as a business  requirement that we have been doing all the hard  

work all these decades, and we plan to. So my answer to Dylan's previous question  

was that I take—whether it's in the United  States, or when the White House and the USG says,  

"We want you to allocate more of your wafer  starts to fabs in the US"—we take that seriously. 

Or whether it is data centers and the  EU boundary, we take that seriously. 

So to me, respecting what are legitimate  reasons why countries care about sovereignty,  

building for it as a software and  a physical plant, is what we'll do. 

As we go to the bipolar world—US,  China—it's not just you versus Amazon,  

or you versus Anthropic, or you versus Google. There is a whole host of competition. 

How does America rebuild the trust? What do you do to rebuild the trust? 

To say, "Actually, no, American companies  will be the main provider for you." 

And how do you think about competition with  up and coming Chinese companies, whether it  

be ByteDance and Alibaba or Deepseek and Moonshot? To add to that question, one concern is how we're  

talking about how AI is becoming this industrial  capex race where you're rapidly having to build  

quickly across all loads of supply chain. When you hear that, at least up until now,  

you just think about China. This is their  comparative advantage. And especially if we're not  

going to moonshot to ASI next year, but it's going  to be decades of buildouts and infrastructure,  

how do you deal with Chinese competition? Are they privileged in that world? 

It’s a great question. In fact, you just  made the point of why trust in American  

tech is probably the most important feature. It's not even the model capability, maybe. 

It is, "can I trust you, the company,  can I trust you, your country, and its  

institutions to be a long-term supplier?" That may be the thing that wins the world. 

That's a good note to end on. Satya, thank you for doing this. 

Thank you so much. Thank you. Thank you. 

It's awesome. You two guys are quite the team.