Satya Nadella | BG2 w/ Bill Gurley & Brad Gerstner

I think the company of this generation has already been created, which is OpenAI in some sense. It's kind of like the Google or the Microsoft or the Meta of this era. Well, it's great to be with you. You know, when Bill and I were talking and looking back at your tenure as CEO, it was really quite astonishing. You know, you started at Microsoft in 1992.

For those who may not know, you took over online in 2007. You launched Bing Search in 2009. You took over servers and launched Azure in 2011. You became CEO in 2014. It was just before that when a now well-known essay entitled "The Irrelevance of Microsoft" was published.

Now, since then, you've taken Azure from $1 billion to a $66 billion run rate. The total revenues of the business are up 2.5 times. The total earnings are up over 3 times. And the share price is up almost 10 times. You've added almost $3 trillion of value to Microsoft shareholders.

And as you reflect back on that over the course of the last decade, what's the single greatest insight you thought you could pursue then to unlock value and change the course of Microsoft, which has been just an extraordinary success? Yeah, so the way I've always thought, Brad, about that entire period of time is that, in some sense, from 1992 to now, it's just one continuous period for me. Although, obviously, 2014 was a significant event with the accountability that came with it. What I felt was essential was to identify patterns: recognizing when we were successful and when we were not, and then doing more of the former and less of the latter. I mean, in some sense, it's as simple as that, because I lived through the time when I joined in 1992, just after Windows 3.1 was launched.

I think Windows 3.1 was released in May of '92, and I joined in November of '92. In fact, I was working at Sun, and I was considering going to business school. I received an offer from Microsoft, and I thought, "Maybe I'll go to business school." However, somehow, my future boss, who was hiring me, convinced me to come to Microsoft instead. It turned out to be the best decision because what really convinced me was the Professional Developers Conference (PDC) of '91 at the Moscone Center when I saw Windows NT for the first time. It wasn't called Windows NT at that time, and it was running on x86 architecture. I thought to myself, "God, what's happening on the client will happen on the server."

And this is a platform company and a partner company, and they're going to ride the wave. That was the calculus at the time. Then, of course, the web happened. We managed that transition, and we got a lot of things right.

Like, for example, I mean, we recognized the browser; we competed and eventually got that browser thing right. We missed search, right? We sort of felt like, wow, the big thing there was the browser because it felt more like an operating system, but we didn't understand the new category, which is that the organizing layer of the internet happened to be search. Then, we were kind of there, but we really didn't get it right. Obviously, the iPhone happened, and we got the cloud right.

So, if I look at it and then we are here, we are on the fourth one on AI. In all of those cases, I think doing things that do not come from envy or the desire to simply replicate what someone else has done is important. Sometimes it's okay to fast follow, and it may work out, but you shouldn't do things out of envy. That was one of the hardest lessons I think we've learned. You have permission to innovate, and you can often do it better. Both of those aspects matter to me: brand permission and the expectations of your customers. Jeffrey Moore once said to me, "Why don't you go do things that your customers expect you to do?" I love that idea!

Which is, cloud was one such thing, which is the customer. You know, in fact, when I first remember showing up in Azure, people would tell me, “Oh, it's a winner-take-all; it's all over, and Amazon's won it all.” I never believed that because, after all, I'd competed against Oracle and IBM in the servers. I always felt like, look, it's just never going to be winner-take-all when it comes to infrastructure. All you need to do is get into the game with a value proposition. So in some sense, a lot of these transitions for me have been about making sure you recognize your structural position.

You really get a good understanding of where you have permission from those partners and customers who want you to win, so go do those obvious things first. You could call it the basics of strategy, but that's how I feel it is—at least, I think it has been key. There are things that need to be cultivated, to your point, Brad, such as the sense of purpose and mission, and the culture that you need to have. I would say those are the necessary conditions to even have a real chance for shots on goal. However, I would emphasize that getting the strategy right by recognizing your structural position and permission is probably what I have done a reasonable job of.

Sachin, before we move on to AI, I have a couple of questions about the transition, echoing what Brad said. I mean, it’s definitive that you may be the best CEO hire of all time. I mean, $3 trillion is unmatched. So, I read an article that suggested, and maybe this isn't true, you can tell us, that you wrote a 10-page memo to the committee that was choosing the CEO. Is that true?

And what was in the memo? Yes, it is true. I think our CEO process was pretty open at that time. Quite frankly, it was definitely not obvious to me from the beginning. Remember, I never thought that the first Bill would leave and then Steve would leave, right? It's not like you join Microsoft and think, "Oh yeah, the founders are going to retire, and there's going to be a job opening that you can apply for."

I mean, that was not the mental model growing up at Microsoft. So when Steve decided to retire, I forget now, I think in August of 2013, it was a pretty big shock. At that time, I was running our Server and Tools business, as it was called, in which Azure was housed and so on, and I was having a lot of fun. I didn't even put up my hand first saying, "Oh, I want to be CEO," because it was not even something I was thinking would happen. Eventually, the board came around and asked, and there were a lot of other candidates at that time, even internally at Microsoft.

And so, at some point in that process, I just wanted to write. Quite frankly, it's fascinating. That memo—everything I said in it, right? You know, one of the terms I used in that memo, which I subsequently used even in the first email I sent out to the company, included "ambient intelligence" and "ubiquitous computing." I simplified it to "mobile first, cloud first later" because my PR folks came to me and asked, "What the heck is this?"

Nobody will understand what ambient computing is, or about ambient intelligence and ubiquitous computing. But that was the mobile-first, cloud-first approach. How do you really go where the secular shift is? Understanding our structural position involves considering Microsoft Cloud. What assets do we have? What is Microsoft 365?

In fact, one of the things I've always resisted is thinking of our cloud the way the market segments it, right? The market segments it—oh, here is IaaS. Even Brad, the way he described ads. To me, I don't allocate my capital thinking, "Here is the Azure capital. Here is the M365 capital."

Here is, you know, gaming. I kind of think of it as, hey, there's a cloud infrastructure. That's the core theory of the firm for me. On top of it, I have a set of workloads. One of those workloads happens to be Azure.

The other one is M365, Dynamics, gaming, what have you. In some sense, that was all in that memo and has pretty much played out. One of the assumptions at that time was that we had a 98% to 99% gross margin business in our servers and clients. People said, "Oh, good news! You can now move to the cloud, and maybe you'll have some margin." And so that was the transition.

My gut feeling was that it would result in less gross margin, but the total addressable market (TAM) is larger. We will sell more to small businesses and, in aggregate, we will see an increase in consumption, even in terms of upselling. We had previously sold a bit of Exchange, but if you think about it, Exchange, SharePoint, and Teams have all expanded. That was the basic premise in the memo I wrote.

Was there any element of cultural shift? I mean, there are CEO hires made all over the world all the time, and many of them fail. Intel is going through a second reorganization as we speak. As Brad pointed out, some people argue that Microsoft is the next IBM or DEC, suggesting that its better days are over. So, what did you do, and what would you advise the new CEOs coming in to reboot the culture and set it on a different path?

Yeah. One of the advantages I think I had was that I was a consummate insider, right? I mean, having spent pretty much all my professional career at Microsoft, in some sense, if I were to criticize our culture, it felt like I was criticizing myself. So, in a way, the break I got was that it never felt like someone from the outside was coming in to criticize the folks who were here; instead, it was mostly about pointing the finger right back at me because I was very much a part of the culture. I couldn't say anything that I was not part of.

And so I felt like, to your point, Bill, I distinctly remember, I think the first time Microsoft became the largest market cap company. I remember walking around the campus, all of us, including me, strutting around as if we were the best thing to humankind, right? It was as if all our brilliance was finally reflected in the market cap. Somehow, it stuck with me that, God, that is the culture we want to avoid. I always say, from ancient Greece to modern Silicon Valley, there's only one thing that brings civilizations, countries, and companies down, and that is hubris. One of the greatest breakthroughs for me was when my wife introduced me to a book by Carol Dweck a few years before I became CEO. I read it on growth mindset, primarily in the context of my children's education and parenting. I thought, God, this is the best! We are always talking about learning and fostering learning cultures and so on.

And this was the best cultural meme we could have ever picked. I attribute a lot of our cultural success to that meme because the other nice thing about it, Bill, is that it is not trademarked by Microsoft, nor is it some new dogma from a CEO. It's something that speaks to work and life. You can be a better parent, a better partner, a better friend, a better neighbor, and a better manager or leader. So we chose that, and the pithy way I’ve always characterized it is, "Hey, go from being the know-it-alls to the learn-it-alls."

And it's a destination you never reach because the day you say, "I have a growth mindset" means you don't have a growth mindset by definition. So, it has been very, very helpful for us. You know, it's like all cultural change; you have to give it time, oxygen, and breathing space. It's both top-down and bottom-up, and it balances out, right?

Which is, there's not a single meeting that I have with the company, or even my executive staff, where I don't start with mission and culture. Those are the two bookends. I've been very disciplined about my framework. To your point about that memo, for the last almost 11 years, the arc has been the same: mission, culture, the worldview—ambient intelligence and ubiquitous computing—and then the specific set of products and strategies.

That frame, I pick and choose every word. I'm very, very deliberate about it. I repeat it until I'm bored stiff, but I just stay on it. Raoul, speaking of that, you mentioned the phase shifts that we've been through. I've heard you say that as a large platform client company, most of the value capture is determined in those first three or four years of the phase shift when the market position is established, Sacha.

I've heard you say that Microsoft was essentially coming off having missed search and largely missed mobile. You also mentioned that Microsoft caught the last train out of town on cloud. As you started thinking about the next big phase shift, it seems that you and others on the team, including Kevin Scott, sniffed out early that Google was likely ahead in AI with DeepMind. You made the decision to invest in OpenAI. What convinced you to take this direction instead of continuing with the internal AI research efforts that you had underway? Yeah, that's a great point because there are a couple of things to consider here.

One is that we have been working on AI for a long time—quite a long time. Obviously, when Bill started Microsoft Research in 1995, I think, the first group he established was focused on natural user interfaces. The initial focus was on speech. Rick Rashid came on board, and in fact, Kai-Fu Lee worked here as well.

We had a lot of, I would say, focus on trying to crack natural user interfaces. Language was always something that was aired about. In fact, even Hinton worked on this. Some of the early work in deep neural networks (DNNs) happened when he was in residency at Microsoft Research, and then Google hired him. So we missed, I would say, even in the early 2010s, some of what could have been a doubling down around the same time that Google doubled down and acquired DeepMind.

And so that actually bothered me quite a bit. But I always wanted to focus, like, for example, Skype Translate was one of the first things I concentrated on because that was pretty cool. Like, that was the first time you could see transfer learning work, right? You could train it on one language pair, and it got better in another language, right? That was the first place where we could say, "Wow, machine translation is also a DNA; it's different." Ever since then, I've been obsessed with language. Along with Kevin, in fact, the first time Elon and Sam were looking for Azure credits and what have you.

We gave them some credits, and at that time, they were more into RL, Dota 2, and other games. It was interesting. Then we stopped—I'm not even sure exactly what happened next. I believe they went to GCP, and later, they came back to discuss what they wanted to do with language.

That was the moment, right, when they talked about transformers and natural language. And because I always felt like, look, if that’s our core business, it goes back a little bit to how I think, which is—what's our structural position? I always knew that if there were a way to have a nonlinear breakthrough in terms of some model architecture that exhibited, you know, like one of the things that Bill would always say throughout our career here was, there’s only one category in digital: it’s called information management. The way he thought about it was, you schematize the world, right?

Take people, places, things—just build a schema, right? We explored many avenues; there was this very infamous project called WinFS at Microsoft, which aimed to schematize everything. The idea was that by doing so, you could make sense of all information. However, it turned out to be impossible to achieve. Consequently, a breakthrough was needed.

And this was just impossible to do. Therefore, we needed some breakthrough. We thought maybe the way to achieve that is through our schema; after all, the human brain does it through language, monologue, and reasoning. Anyway, that's what led me to OpenAI, and quite frankly, the ambition that Sam, Greg, and the team had. That was the other thing, right?

Scaling loss. In fact, I think the first memo, weirdly enough, I read on scaling was written by Dario when he was at OpenAI and Ilya. And that's sort of what, like I said, led me to think, "Hey, wow, if this is going to have exponential performance, why not go all in and give it a real shot?" And then, of course, once we started seeing it work on GitHub Copilot and so on, it was pretty easy to double down.

But that was the intuition. One of the things that has happened, I think, in previous phase shifts is that some of the incumbents don't get on board fast enough. You even talked about Microsoft, perhaps missing mobile or search, or that kind of thing. I could argue, especially since I'm old and I've seen these shifts, that everyone's awake on this one. It's the most awake.

It's heavily choreographed. Everyone's maybe at the starting line at almost the same time. I'm curious if you agree with that and how you think about the key players in the race—Google, Amazon, Meta with LLaMA. Elon is introduced. Yeah, it's an interesting one. To your point about it, I always think about it, right?

If you take the late 90s, there was Microsoft, and there was daylight. And then there was the rest. Interestingly enough, now, people talk about the Magnificent Seven. There are probably even more than that, to your point about everyone being aware of it. They all have amazing balance sheets.

There are even, I think I’ll call it, if you think about OpenAI, in some sense, you could say it's MAG-8 because I believe the company of this generation has already been created, which is OpenAI. In some ways, it's kind of like the Google, Microsoft, or Meta of this era. Therefore, I think it's going to be very competitive. Additionally, I don’t think it's going to be a winner-takes-all situation, although there may be some categories that could be winner-takes-all.

For example, on the hyperscale side, absolutely not. I mean, the world will demand, even outside of China, multiple providers of frontier models distributed all over the world. In fact, one of the best structural positions that I think Microsoft has is due to the fact that, if you remember, the Azure structure is slightly different. We built out Azure for enterprise workloads with a lot of data residency; we have over 60 regions, more regions than others. So it was not like we built our cloud for one big app; we built the cloud for a lot of heterogeneous enterprise workloads, which I believe, in the long run, is where all the inference demand will be, with nexus to data and the app server, and so on.

So, I think there are going to be multiple winners at the infrastructure layer. There are going to be various models as well; each hyperscaler will have a range of models along with an app server. Every application today, including Copilot, is essentially a multi-model app. In fact, there is now a completely new app server. Just as there were mobile app servers and web app servers, there is now an AI app server.

And for us, that's Foundry, and we're building one, and others will build as well. There will be multiple of those. Then, in apps, I think there will be more. I would say network effects are always going to be at the software layer, right? So at the app layer, there will be different network effects in consumer, in the enterprise, and so on. To your fundamental point, I think you have to analyze it structurally by layer.

There is going to be fierce competition among the seven, eight, or ten of us at different layers of the stack. As I always tell our team, we need to watch for those who come in and add to it, right? That’s the game you’re all in—always looking for the new entrepreneur who will emerge out of the blue. At this point, I would say OpenAI is one such company that has achieved escape velocity. Yeah.

If we think about the app layer for a second, let's start with consumer AI a little bit here. Satya, Bing is a very large business. You and I have discussed that the ten blue links were maybe the best business model in the history of capitalism, but it's massively threatened by a new modality where consumers just want answers. For example, my kids say, "Why would I go to a search engine when I can just get answers?" So, do you think Google and Bing can continue to grow their legacy search businesses in the age of answers? Additionally, what does Bing need to do, or what do your consumer efforts under Mustafa need to focus on in order to compete with ChatGPT, which really seems to have broken out from a consumer perspective?

Yeah, I mean, I think the first thing is what you said last, which is chat meets answers. And that's ChatGPT; both the brand and the product are becoming stateful, right? I mean, ChatGPT now is not... you know, in fact, search was a stateless model. There was search history, but I think more so, these agents will be a lot more stateful. So, in fact, that's why I was so thrilled.

I've been trying to secure an Apple search deal for about ten years. So when Tim finally made a deal with Sam, I was the most thrilled person! It's better for ChatGPT to get that deal than anyone else because we have that commercial and investor relationship with OpenAI. To that point, the way I look at it is that distribution matters. This is where Google has an enormous advantage; they have the distribution on Apple.

They're the default. They are obviously the default on Android. They touch, so therefore I think, and the habits don't go away, right? But at times, you just go to the browser URL and type in your query, right? I mean, even now, even though I want to go to Copilot, my usage is mostly Copilot. And if I have to think about Bing versus Copilot, it's kind of interesting, right?

Some of the navigational tasks I perform using Bing; for pretty much everything else, I rely on Copilot. I believe that shift is happening universally. We are perhaps just one or two of these agents away for shopping or travel from significantly impacting some of the commercial queries. That will be the moment when traditional search experiences a major change, as I think some of the commercial intent will also migrate into chat. Currently, businesses have largely withstood this shift because the commercial intent has not yet migrated.

But once commercial intent migrates, that's when it suddenly moves. So I think, yes, this is a secular shift. The way we are managing it is that we have three properties as well, right? There is Bing, MSN, and Copilot. We believe he has a clear vision of what these three things are.

They're all part of one ecosystem. One is a feed, one is search in the traditional way, and the other is this new agent interface. They all have a social contract with content providers.

We need to drive traffic. We may need to implement paywalls and ad-supported models. All of these factors are things we are trying to manage. We also have our own distribution.

The one advantage we still have is Windows. We get to relitigate. We lost the browser, right? Even Chrome became the dominant browser, which is a real travesty because we had won against Netscape only to lose to Google. However, we are getting it back now in an interesting way, both with Edge and with Copilot.

Guess what? Now even Gemini has to earn. The good news about Windows, at least, is that it's an open system, right? He has a shot. Gemini has a shot. You don't have to call Microsoft.

You can do your best work and go over the top. But that also means we get to as well. Losing can be great sometimes because you can win it all back. To me, even with Windows distribution, I always say Google makes more money on Windows than all of Microsoft. I mean, literally.

I mean, wow! This is the best news for Microsoft shareholders that we lost so badly. Now we can contest it and win back some ground. Hey, Satya, one thing—everybody's talking about these agents, and if you think ahead a bit, you can imagine all kinds of players wanting to take action on other apps and data that may be on a system. Microsoft is in an interesting position because you control the Windows ecosystem, but you also have apps on the iPhone ecosystem and the Android ecosystem. How do you think about this? This is partially a terms of service question and partially a partnership question. Will Apple allow Microsoft to control other apps on iOS? Will Microsoft let ChatGPT instantiate apps and access data from apps on Windows OS?

I mean, you get the question. It goes all the way to when you start thinking about search and commerce. For example, will Booking.com let Jim and me run transactions on it without their permission or knowledge? Yeah, I think that this is the most interesting question, right? To some degree, it's unclear exactly how this will happen. There is a slight, very old-school way of thinking about some of this. If you remember, how did business applications or work kinds manage to achieve interoperability, right?

They did manage to achieve that interoperability using connectors, and people had connector licenses. So, a business model emerged, right? I mean, SAP was one of the most classic examples where you could say, “Hey, you can access SAP data as long as you have connectors.” There's a part of me that thinks something similar will emerge when agent-to-agent interfaces occur. It's unclear exactly what will happen with consumers because the value exchange there was largely based on advertising, traffic, and what have you.

Some of those things go away in an agentic world. So, I think the business model is slightly unclear to me on the consumer side. But on the enterprise side, I believe what will happen is that everyone will say, "Hey, in order for you to either act within my action space or to extract data from my sort of schema, so to speak, there needs to be some kind of an interface to my agent that is licensed." I think that's a reason why, today, for example, when I use Co-Pilot at Microsoft, I have connectors to Adobe, to my instance of SAP, and obviously to our instance of CRM, which is Dynamics. It's fascinating.

In fact, when was the last time any of us really used a business application? We license all these SaaS applications, but we hardly use them, and someone in the organization is just inputting data into them. However, in the age of AI, the intensity increases because all that data is now easily accessible. You're just a query away. I can literally say, "Hey, I'm meeting with Bill. Tell me about all the companies that Benchmark has invested in."

It's about taking the web and everything that's in my CRM database, collating it all together, and providing me with notes or insights. To some degree, I think all of this can be monetized by us and by these connectors. More explicitly, one thing that could happen quickly—since there has been talk about it—is allowing ChatGPT on the Windows operating system to start opening random apps and taking actions. Now, that's an interesting idea, right? So when it comes to that type of computer use, who is going to permit it?

So, which is it: the user or the operating system? On Windows, quite frankly, there isn't much I can do to prevent that, other than implementing some security guardrails. Right? I could definitely see that, because one of my biggest fears is the security risk. Right?

If the malware gets downloaded and starts executing actions, that's when it becomes really dangerous. I think those are the types of threats that we will build safeguards against within the operating system itself, which requires elevated access and privilege that this computer has. However, at the end of the day, the user will be in control on an open platform like Windows. I'm sure Apple and Google will have much more control and won't allow such issues to arise. In some sense, you could say that's an advantage they have. It will ultimately be interesting to see how the rules develop around all of this.

We can flip that around, and then we can move on. But would you allow the Android OS, or let's just call it the Android AI, or the iOS AI, to read emails through a Microsoft client on that smartphone? Yeah, I mean, we kind of like— for example, today, one of the things I always think about is I don't know whether that was value leaking or if it actually helped us, right? We licensed the sync for Outlook to Apple for Apple Mail.

It was an interesting case, and I think that there was a lot of value leaked, perhaps. At the same time, I believe that was one of the reasons why we were able to hold on to the exchange. It would have been doubly problematic if we had not done that. So, one of the things I think, going to your point, Bill, is that if we are building this out, we need to create a trust system around owned Microsoft 365. We simply cannot say, "Hey, any agent can come in and do anything," because, after all, it's not our data; it's our customers' data, right?

It would be, and therefore, the customer will have to permit it. The IT team at the customer organization will have to grant permission. It's not like there is a blanket flag I can set. Secondly, it has to have a trust boundary. So, I think what we will do is that it’s an interesting approach, somewhat similar to what Apple is doing with artificial intelligence.

Think of it as something we will do around M365. I played with it a lot today. I'd highly recommend that people download it. It's super interesting.

So, Sachin, you know, clicking on this, Mustafa has said that 2025 will be the year of infinite memory. Bill and I have talked a lot, dating back to the start of this year, about how we think the next 10x function, and it sounds like you agree, is really this persistent memory combined with the capability of taking some actions on our behalf. We're already seeing the beginnings of memory, and I'm pretty convinced as well that 2025 seems like a problem that is pretty well solved. But this question of actions remains: when am I going to be able to say to ChatGPT, "Book me the Four Seasons in Seattle next Tuesday at the lowest price," right?

Bill and I have gone back and forth on this one. It would seem that computer use is the early test case for that. But do you have any sense? Does that seem like a hard one from here to you? Yeah.

The most open-ended action space is still challenging. However, regarding your point, there are two or perhaps three things that are really exciting. Beyond what I'm sure we'll discuss—namely, the scaling laws and the capabilities of the raw models—there is also the use of tools or actions. Additionally, I would say the concept of entitlements is noteworthy. For instance, one of the most interesting products we have within Microsoft is "per view," because increasingly, it raises the question: what permissions do you have?

What can you get? You know, you have to be able to access things in a safe way. Somebody needs to have governance over it and so on. So if you put all those three things together, this agent is going to be more governable. When it comes to actions, they should also be verifiable.

And then it has memory. I think you're off to a very different place, where you can do more autonomous work, so to speak. One of the things I always consider is your build. I like this co-pilot as the user interface for AI because, even in a fully autonomous world, there will be times when you need to raise exceptions or ask for permission for certain actions. Therefore, this UI layer will serve as the organizing layer. In fact, that’s why we think of co-pilot as the organizing layer for work, work artifacts, and workflow.

But to your fundamental point, I don't think the models—take, for example, version 4.0—are quite ready for certain tasks. Not even going to 0.1, but 4.0 is pretty good with function calling. In an enterprise setting, it can perform significantly better than in a consumer context. Function calling on consumer web applications is just challenging; at least in an open-ended web environment, you can work with a couple of websites. However, once you say, "Hey, let's book a ticket on anything," if there are schema changes on the back end, it will trip over.

You can teach it that. That's where I think AI can get better if it's a verifiable, auto-gradable sort of process on rails. However, I believe we are maybe a year or two away from achieving more and more advancements in this area. At least from an enterprise perspective, we're looking at having a sales agent, a marketing agent, and a supply chain agent that can perform more autonomous tasks. We built 10 or 15 of them into Dynamics, right? This includes managing my supplier communications, automatically updating my databases, and changing my inventories as needed.

Those are the kinds of things that you can do today, I would say. I made this comment about near-infinite memory, and I'm sure you’ve heard it or hear it internally. Is there any clarification you can offer about that, or is that more to come? I think that, at some level, the idea is that you have essentially a type system for your memory, right? That's the key point, right? It's not like every time I start, you have to organize it.

I get the idea. He made it sound like you guys had an internal technical breakthrough on this front. Yeah, we have done something; there's even an open-source project. I think it's the same set of folks who worked on all the TypeScript stuff who are working on this. What we're trying to do is essentially take memory, schematize it, and make it available in such a way that you can access it.

Like each time I start, let's just imagine I'm working on a new prompt. Yeah, I know how to cluster based on everything else I've done. I think that type matching and similar techniques are good ways for us to build a memory system. Shifting to enterprise AI, Sacha, you know, the Microsoft AI business has already been reported to be worth about $10 billion.

You've mentioned that it's all inference and that you're not actually renting raw GPUs to others for training because your inference demand is so high. As we consider this, there's a lot of skepticism in the world regarding whether or not major workloads are shifting. If you think about the key revenue products that people are using today and how they are driving your inference revenue, I'd be interested in how that may be similar or different from Amazon or Google. Yeah, I think that's a good question. The way this has played out for us, you have to remember that most of our training efforts at OpenAI are more about investment logic, right? So, it's not really reflected in our quarterly results.

It's more in the other income based on our investment. So, that means the only thing that shows up in revenue — or loss, other income or loss, right? That is right. That is right. Right now, that's how it shows up.

And so, most of the revenue, or all the revenue, is primarily from our API business. In fact, to your point, ChatGPT's inference costs are a significant factor. So, the reality is that the major apps of this era are ChatGPT, Copilot, GitHub Copilot, and the APIs of OpenAI and Azure OpenAI. In some sense, if you had to list the 10 most successful apps, these would likely be among the top four or five.

And so, that’s the biggest driver. The advantage we have had, and OpenAI has had, is that we’ve had two years of runway, pretty much uncontested. To your point, Bill made the observation that, hey, everybody's awake now, and it might be true. I don’t think there will ever be another two-year lead like this again. Who knows?

You know, you say that, and then somebody else drops a sample and suddenly blows the world away. That said, I think it's unlikely that such a lead could be established with some foundation model. However, we have that advantage. That was the great advantage we had with OpenAI. OpenAI was able to really build out this escape velocity with ChatGPT.

But on the API side, the biggest thing we were able to gain was, you know, taking Shopify, Stripe, or Spotify. These are not customers of Azure; they were all customers of GCP or AWS. So suddenly, we gained access to many more logos, who are all, quote unquote, digital natives using Azure in some shape or form. That's one aspect. When it comes to traditional enterprises, I think it's scaling.

Like, I mean, literally it is, you know, people are playing with Copilot on one end and then are building agents on the other end using Foundry. These things are design wins and project wins, and they're slow, but they're starting to scale. And again, the fact that we've had two years of runway on it, I think I like that business a lot more. That's one of the reasons why the adverse selection problems here would have led to lots of tech startups all looking for their H100 allocation in small batches, right? Having watched what happened to Sun Microsystems during the dot-com era, I always worry about that, which is, whoa, you just can't chase everybody building models. In fact, even on the investor side, I think the sentiment is changing, as now people want to be more capital-light and build on top of other people's models and so on and so forth.

If that's the case, everybody who was looking for H100 will not want to look for it anymore. So that's kind of what we've been selective about. And your sense is that for the others, the training of those models and those model clusters was a much bigger part of their AI revenue compared to yours? I don't know. I mean, this is where I'm speaking for other people's results.

I don't know. I mean, it's just that I go back and think about what the other big hit apps are, right? I don't know what they are. Like, I mean, where do they come from? What models do they run, and where do they run them?

And I would imagine that, like, that's kind of... I'm not— I mean, obviously, Google Gemini. I don't know. When I look at the download numbers of any of these AI products, there is ChatGPT, and then there is even Gemini. I'm very surprised at the Gemini numbers. I mean, obviously, I think it will grow because of all the inherent distribution.

But it's interesting to say that there aren't that many. In fact, we talk a lot more about AI scale, but there are not that many hit apps, right? There’s ChatGPT, GitHub Copilot, and Gemini. I think those are the four I would mention in a DAO. Is there anything else that comes to your mind? Well, I think there are a lot of these startup use cases that are starting to gain some traction, kind of from the bottom up.

A lot of them build on top of Llama. But if you said, "Oh, and there's Meta," right? But if you asked, "What are the apps that have more than 5 million DAU?"

Right, I think it's going to be interesting. Zuckerberg would argue that Meta certainly has more, et cetera. But I think you're right in terms of the non-affiliated apps you mentioned. And Zuck's services all run on his own cloud; I mean, he's not using a public cloud. Satya, on the enterprise side, obviously, the coding space is often highly competitive, and you guys are doing well with a lot of venture-backed players in that area.

On some of the productivity apps, I have a question about the co-pilot approach. I guess Marc Benioff has been somewhat obnoxiously critical on this front and referred to it as "Clippy 2" or whatever. Do you worry that someone might consider a first-principles AI built from the ground up, and that some of the infrastructure, say, in an Excel spreadsheet, isn't necessary to know if you created an AI-first product? The same thing, by the way, could be said about CRM systems, right?

There are a number of fields and tasks that may be able to be obfuscated for the user. Yes, I mean, it's a very, very important question. Regarding SaaS applications or business applications, let me just speak about our own Dynamics product. The approach we are taking, I think, revolves around the notion that business applications may all collapse in the agent era. Because if you think about it, they are essentially CRUD (Create, Read, Update, Delete) databases with a lot of business logic.

The business logic will all be handled by these agents, which will operate as multi-repo CRUD systems. They will not discriminate between different backends; instead, they will update multiple databases. All the logic will reside in the AI tier, so to speak.

And once the AI tier becomes the place where all the logic resides, then people will start replacing the backends, right? That's what's actually interesting. As we speak, I think we are seeing pretty high rates of wins on dynamic backends and the use of agents. We are going to go pretty aggressively and try to collapse it all, right? Whether it's in customer service, or, by the way, the other fascinating area that's increasing is not just CRM, but also what we call finance and operations, because people want more AI-native business applications, right?

That means the business application, the logic tier, can be orchestrated by AI and AI agents. In other words, the interaction from the co-pilot to the agent to my business application should be very seamless. Similarly, you could even ask, "Why do I need Excel?" Interestingly enough, one of the most exciting things for me is that Excel with Python is like GitHub with Co-Pilot, right? That's essentially what we have done. When you have Excel, by the way, this would be fun for you guys, right?

You should just open Excel, bring up Copilot, and start experimenting with it. It's no longer just about making sense of the numbers you have; it's like having a data analyst at your disposal. It will actually create the plan for you, right? It will essentially function like GitHub Copilot, which generates a plan and then executes it. This is akin to a data analyst using Excel as a tool for visualizing rows and columns to conduct analysis. So, it’s a different kind of tool for you.

So the co-pilot is using Excel as a tool with all of its action space because it can generate and has a Python interpreter. That is, in fact, a great way to reconceptualize Excel. At some point, you could say, "Hey, I'll generate all of Excel." And that is also true. After all, there's a code interpreter, right?

So, therefore, you can generate anything. And so, yes, I think there will be disruption. The way we are approaching our M365 products is by building Copilot as that organizing layer and user interface for AI. We aim to integrate all agents, including our own. You could say that Excel is an agent for my Copilot, and Word is also an agent.

It's kind of specialized canvases, which is what I'm doing for a legal document. Let me take it into Pages and then into Word, and then have the co-pilot assist with it. I'll go into Excel and have the co-pilot help with that as well. So, that's a new way to think about the work and workflow. You know, one of the questions I hear people wringing their hands about a lot today, Satya, is the ROI people are getting from their investments.

You have over 225,000 employees. Are you leveraging AI to increase productivity, reduce costs, and drive revenues in your own business? If so, what are the biggest examples? To be more specific, when we had Jensen on, I asked him that when he 2x or 3x'd his top line, what did he expect his headcount to increase by? He said 25%.

And when asked why, he said, "Well, we have 100,000 agents helping us do the work." So when you 2 or 3 times your revenue for Azure, do you expect to see a similar type of leverage on headcount? Yeah. I mean, it's top of mind for both Microsoft and its customers. Here's the way I come at it.

I love this idea of going to school and learning a lot about what happened in industrial companies with lean practices. Yeah, right? I mean, it's fascinating, right? They're all GDP-plus growers.

It's unbelievable. I mean, the discipline they have in how the good industrials can literally say, "Hey, I'll add 200 to 300 basis points of tailwind just by lean," which means increasing value and reducing waste. That's the practice. So I think of AI as the lean for knowledge work. We are really diving deep into it, looking at how we can really enhance our processes. That's why I think of the good old days. We remember in the '90s, we had all this business process re-engineering.

I think it's back in a new way, where people who can think about end-to-end process flows can ask, "Hey, what's the best way to approach process efficiency? What can be automated? What can be made more efficient?" That's part of my perspective. Customer service is the most obvious example. We are on course to spend around $4 billion or so.

This includes everything from Xbox support to Azure support. This is really, I mean, this is a serious one-year consideration because of the deflection rate on the front end. Then, the biggest benefit is agent efficiency, right? When the agent is happier, the customer is happier, and our costs are going down. So, I think that's the most obvious place we have in our contact center application that's also performing exceptionally well.

The other one is obviously GitHub Copilot. That's the other. With GitHub Copilot Workspace, that's the first place where even this agentic side comes in. You go from an issue to a plan, or to a specification, to a plan, and then to multi-file editing. It completely changes the workflow for the engineering team.

As I said, the O365 is the catch-all, right? So the M365 Copilot is where, I mean, just to give you a feel, even my own experience, right? Every time I'm meeting a customer, I would say the workflow of the preparation in the CEO office has not changed since 1990, right? Basically, I mean, in fact, one of the ways I look at it is to imagine how forecasting happened pre-PCs and post-PCs, right? There were faxes.

Right. Then inter-office memos became common, and then PCs became a thing. People said, "Hey, I'm just going to put an Excel spreadsheet in an email and send it around. People will enter numbers, and we will have a forecast." The same thing is happening in the AI era right now, all over the place, right? I prepare for a customer meeting where I literally go into Co-Pilot, and I say, "Tell me everything I need to know about the customer." It provides me with information from my CRM, my emails, my team's meetings, and the web, right? It grounds it.

I put it into Pages and share it with my account team in real time. So just imagine the hierarchy: this entire concept of, "Oh, let me prepare a brief for the CEO," goes away. It's just a query away. I generate a query and share a page if they want to annotate it. So I'm reasoning with AI and collaborating with my colleagues.

That's the new workflow, and it's happening all over the place. Someone gave me this example from the supply chain. They said supply chain is like a trading desk, except it doesn't have real-time information. That's kind of what it is.

It's like you wait for the quarter to end, and then the CFO comes and hits you over the head, listing all the mistakes you've made. What if the financial analyst could essentially be available to you in real time, providing insights like, "Oh, you're working on this contract for a data center in this region. You should consider these terms." All that intelligence in real time is changing the workflow and work artifacts. There are numerous use cases all around. I think your fundamental point is that our goal is to create operating leverage through AI, right?

So, I think headcount will, in fact, be one of the ways I look at it. I would say our total people costs will go down, our cost per head will go up, and my GPU per researcher will increase. That's kind of the way I see it. That makes sense. Hey, let's shift ahead to something that you referenced earlier regarding what we're seeing from model scaling and capital expenditures in general. You know, I've heard you talk about Microsoft's capital expenditures.

I imagine in 2014, when you took over, you had no idea that the capital expenditures would look like they do today. In fact, you've said that these companies increasingly look more like industrial companies in terms of capital expenditures than traditional software companies. Your capital expenditures grew from about $20 billion in 2020 to maybe as high as $70 billion in 2025. You've consistently earned a pretty strong return on that capital expenditure, and there's actually a very high correlation when you look at your capital expenditures relative to revenue. Some people are worried that this correlation will break, and even you have acknowledged that, at some point, capital expenditures may need to be spent ahead of the revenue. There could be an air pocket we need to build for this resiliency. So, how do you feel about the level of capital expenditures?

Does it cause you any sleepless nights? And when does it begin to taper off in terms of this rate of growth? Yeah. I mean, there are a couple of different things, right? One is, this is where being a hyperscaler, I think, is structurally super helpful because, in some sense, we've been practicing this for a long time. You know, data centers have 20-year life cycles, you pay for power only when you use it, the equipment has a lifespan of six years—you know how to drive utilization up.

And the good news here is that it's kind of capital-intensive, but it's also software-intensive. You use software to bring the return on invested capital (ROIC) of your capital higher, right? That's similar to when people, even in the early days, asked, "Can a hyperscaler ever make money?" The difference between old hosting companies and the new hyperscalers is indeed software, right?

And that, I think, is what's going to apply even to this GPU physics event, right? Which is, yeah, you buy it, you build it out. In fact, one of the things that's happening right now is what I'll call catching up, right? We built, after all, over the last 15 years, the cloud. Only a new meter showed up in the cloud. It's called the AI accelerator, because every app now needs a database, a Kubernetes cluster, and a model that runs on an AI accelerator.

So if you say, "Oh, I need all three," you suddenly have to build up these AI accelerators in order to provision for all of these applications. That will normalize. The first thing is that the build-out will happen, the workloads will normalize, and then it will continue to grow like the cloud has grown. That’s one side of it. Additionally, it's crucial to avoid some of these adverse selection issues and ensure that it is not just a supply-side phenomenon, where everybody is building, only hoping demand will come. We need to make sure that there is real, diverse demand all over the world and across various segments.

I watch for all of that. So, I think that the way to manage the ROIC is important. By the way, the margins will be different. This goes back to the very early dialogue we had. When I think about the Microsoft Cloud, the margin profile of a raw GPU differs from the margin profile of Fabric plus GPU, Foundry plus GPU, or the GitHub Copilot add-on to M365. So, they are all going to be different.

And so, if you’re having a portfolio, that matters here, right? Because if I look at Microsoft, why does Microsoft have a premium today in the cloud? We are bigger than Amazon, growing faster than Amazon, and have better margins than Amazon, because we have all these layers. That’s kind of what we want to do even in the AI era. Satya, there’s been a lot of talk about model scaling. Obviously, there has been discussion historically about 10x-ing the cluster size that you might do over and over again, not just once and then twice.

And NX.AI is still making noise about going in that direction. There was a podcast recently where they kind of flipped everything on its head and said, "Well, if we're not doing that anymore, it's way better because we can just move on to inference, which is getting cheaper, and you won't have to spend all this CAPEX." I'm curious—those are two different views of the same coin. What's your perspective on large LLM model scaling, training costs, and where we're headed in the future?

I mean, I'm a big believer in scaling laws; I'll sort of first say that. In fact, if anything, the bet we placed in 2019 was on scaling laws, and I stand by that. In other words, don't bet against scaling laws. But at the same time, let's also be grounded in a couple of different things. One is that these exponentials in scaling laws will become harder, simply because as the clusters become more complex, everything—the distributed computing problem of large-scale training—becomes more challenging. And so that's kind of one side of it.

So there is, but I would just still say, and I’ll let the OpenAI folks speak for what they’re doing, that they are continuing to enhance their pre-training. I think it is not over; it sort of continues. But the exciting thing, again, OpenAI has talked about, and Sam has mentioned, is what they’ve done with O1, right? This chain of thought with auto grading is just fantastic. In fact, basically, it is test-time compute or inference-time compute as another scaling law, right? So you have pre-training, and then you have effectively this test-time sampling that creates the tokens that can go back into pre-training. This process creates even more powerful models that then are running during your inference, right?

So, therefore, I think that's a fantastic way to increase model capability. The good news about test time or inference time compute is that sometimes running those O1 models involves two separate things. Sampling is somewhat like training when you're using it to generate tokens for your pre-training. Additionally, when customers are using O1, they are using it to generate tokens for their pre-training, and they are also getting paid for it.

And so, therefore, there is more of an economic model, right? So, I like it. In fact, that's where I said I have a good structural position with over 60 data centers all over the world. It's a different hardware architecture for one of those scaling versus the other, for the pre-training versus. Exactly.

And I think the best way to think about it is that it's a ratio, right? So going back to Brad's point about ROIC, this is where I believe you really need to establish a stable state. In fact, whenever I hear Dr. Jensen, I think he's got it right, which is, look, you kind of want to buy some every year, not just one. Think about it, right?

When you depreciate something over six years, the best way is what we have always done, which is to buy a little every year and age it, age it, age it, right? You use the leading node for training, and then the next year it goes into inference. That's sort of the stable state I think we will reach across the fleet for both utilization and the Return on Investment (ROI), and then the demands will meet supply. And basically, to your point about everybody saying, "Oh, wow, have the exponentials stopped?" one of the other factors is that economic realities will also sort of come to a halt.

And at some point, everybody will look and say, "What's the economically rational thing to do?" Which is, hey, even if I double every year's capability, if I'm not able to sell that inventory, that presents a problem. The other issue is the winner's curse. Even if I double every year's capability, if I'm not able to sell that inventory, it still becomes a concern. Additionally, you don't even have to publish a paper; other folks can just look at your capability and either distill it or assess its value.

It's just impossible. It's kind of like piracy, right? I mean, you can sort of enforce all kinds of terms of use, but it's impossible to control distillation. That's one. The second thing is, you know, you don't even have to do anything.

You just have to reverse engineer that capability, and you can do it in a more computationally efficient way. Given all this, I think there will be a limit on how much people can do right now. A little bit of it is that everybody wants to be first, which is great. However, at some point, the economic reality will set in for everyone. The network effects are in the app layer.

So why would I want to spend a lot on some model capability when the network effects are all on the app layer? What I heard you say, I believe, is that Elon has mentioned he’s going to build a million GPU cluster. I think Meta has said the same thing. I think the pre-training—and then he kind of joked about a million.

But I think he joked about a billion. But, you know, the fact of the matter is, have you, versus the start of the year, Satya, based on what you've seen around pre-training and scaling, changed your infrastructure plans around that? And then I have a separate question with regard to O1. I am building what I would call a way of thinking a little bit about the 10x point, right, which is, hey, how do you... we can argue the duration, like is it every two years, every three years, every four years? There is an economic model. This is where I think we need a more disciplined way of thinking about how to clear your inventory so that it makes sense, right? Or the other way is looking at the depreciation cycle; it makes sense, right?

Or the other way is the depreciation cycle of your kit. There is no way you can sort of buy, unless you find that the physics of the GPU works out, where suddenly it flows through my P&L and it's actually, you know, it's in the same or better margin than a hyperscaler. That's simple. So that's kind of what I'm going to do. I'm going to keep going and building basically to figure out how to drive inference demand and then keep increasing my capability and be efficient at it. I absolutely understand that Sam may have a different objective, and he's been open about it.

He may say, "Hey, I want to build because I have a deep conviction about what AGI looks like," or something to that effect. So, that's where I think a bit of our tension lies. That said, to clarify something, I heard Mustafa say on a podcast that Microsoft is not going to engage in the biggest model training competition that’s currently happening. Is that correct? Well, what we won’t do is it twice, right?

Because after all, we have the IP from, right? Right. Fair. Fair. For Microsoft today, given the partnership with OpenAI, to do two unnecessary—you're going to create a redundant training set. Yes.

Correct. So we are very focused, and that's why we have control. By the way, that's the strategic discipline we have maintained, right? That's why I always stress to Sam that we bet the farm on OpenAI and said, "Hey, we'll concentrate our compute." We did this because we had all the rights to the IP. So that's sort of the give and take on it.

And we feel fantastic about it. What Mustafa is basically saying is that we will also focus a lot of our efforts on post-training and verification, among other things. That's a significant aspect. We will dedicate many of our computing resources to adding more model adaptations and capabilities that make sense while maintaining a principled pre-training process. This will internally give us the capability to handle various tasks. We already have different model weights and model classes for different use cases, which we will continue to develop. Regarding your question about Brad's inquiry on balancing GPU ROI, does that address why you've outsourced some of the emphasis?

Does that answer the question as to why you've outsourced some of the infrastructure to CoreWeave, that partnership that you have? We did this because we all got caught off guard by the impact of ChatGPT and OpenAI APIs. It was just too much for us to handle. Fair enough. Yeah, we were completely overwhelmed; I mean, it was impossible.

There's no supply chain planning I could have done. You know, what is it? None of us knew what was going to happen, right? What happened in November of 2022 was just completely out of the blue. Therefore, we had to catch up. We said, "Hey, we're not going to worry too much about inefficiency."

So that's why, whether it's cold weave or many others, we bought from various places. Fair enough. It makes sense. That was a one-time thing, and now everything is catching up.

And yeah, that was just about trying to get caught up with demand. Are you still supply constrained, Satya? I am Power, yes. I am not chip supply constrained. We were definitely constrained in 2024. What we have told the street is that’s why we are optimistic about the first half of 2025, which is the rest of our fiscal year.

Then after that, I think we'll be in better shape. Going into 2026 and so on, we have a good line of sight. I'm hearing that with respect to this level two thinking, the zero-one test time compute and post-training work being done is leading to really positive outcomes. When you think about that, it’s also quite compute intensive because you're generating a lot of tokens, recycling those tokens back into the context window, and doing that repeatedly. So, that compounds very quickly.

Jensen said he thought that looking at 01, the inference was going to scale to a million or a billion times, meaning that the demand for inference is expected to increase dramatically. In that regard, do you feel like you have the right long-term plan to scale inference to keep up with these new models? Yeah, I mean, I think there are two things there, Brad. In some sense, it's very helpful to think about the full workload. In the agentic world, you must have the AI accelerator. One of the fastest-growing aspects of, in fact, OpenAI itself is the container service because, after all, these agents need a scratch pad for doing some of that auto-grading and even to generate the samples, right?

And so, that is where they run a code interpreter. By the way, that is a regular Azure Kubernetes cluster. In an interesting way, there's a ratio of what constitutes regular Azure compute and its connection to the GPU, along with some data services. To your point, when we say "inference," that's why I look at it and say people think about AI as separate from the cloud. AI is now a core part of the cloud.

And I think in a world where every AI application is a stateful application, it's an agentic application. That agent performs actions, and thus, a classic app server plus the AI app server plus the database are all required. So I go back to my fundamental point, which is, hey, we built this 60 plus AI regions — I mean, the Azure regions — and they will all be ready for full-on AI applications. That's, I think, what will be needed. That makes a lot of sense. So let's talk a little bit. We've discussed OpenAI a lot during this conversation, but you're managing a balance between a huge investment there and your own efforts. At Ignition, you showed a slide highlighting the differences between Azure OpenAI and OpenAI Enterprise.

A lot of those discussions were about the enterprise grade, you know, things that you bring to the table. So when you look at that tension and the competition you have with OpenAI, do you think of ChatGPT as likely being the winner on the consumer side? You’ll have your own consumer apps as well. Then you’ll divide and conquer when it comes to enterprise. How do you think about competing with them?

The way I think about it at this point, given OpenAI, is as a very large-scale company. It's a truly successful company with multiple lines of business and segments. I approach it very Principally, like I would with any other major partner, right? Because I don't think of them differently. I think of them as, hey, as an investor, what are their interests and our interests, and how do we align them?

I think of them as an IP partner. Because we provide them with systems IP, they provide us with model IP, right? So that's another aspect where we are very deeply interested in each other's success. I view them as a significant customer, and therefore, I want to serve them as I would serve any other major customer. Lastly, there's the concept of co-opetition, whether it’s Co-Pilot in the consumer space, Co-Pilot with M365, or something else. We often discuss, "Where is the co-opetition?"

Where is that? And that's where I kind of look at it and say ultimately, these things will have some overlap. However, in that context, the fact that they have the Apple deal is, in some sense, accretive for the Microsoft shareholder. Even considering the differences in their APIs, customers can choose which API front to use. There are differences; Azure has a particular style. If you're an Azure customer and you want to use some of the features, there are indeed differences. Azure has a particular style.

And if you're an Azure customer and you want to use other services of Azure, then it's easiest to have Azure and Azure Mac. But if you're on AWS and you want to use just the API in a stateless way, great, just use OpenAI. In an interesting way, having these two types of distributions is also beneficial for Microsoft's costs. I would say that the curious part of the Silicon Valley community, and even more broadly, the entire business community, is infatuated with the relationship between Microsoft and OpenAI. I was at DealBook last week, and Andrew Sorkin pushed Sam really hard on this.

I imagine there's a lot you can't say, but is there anything you can say? There's supposedly a restructuring and a conversion to profit. I guess Elon has launched a missive in there as well. What can you tell us? Yeah, I mean, I think those, Bill, arrive in there as well.

What can you tell us? Yeah, I mean, I think those, Bill, are obviously all for the Opening Eye Board and Sam, Sarah, and Brad, and that team to decide what they want to do. We want to be supportive. I mean, this is where we're an investor. I would say the one thing that we care deeply about is that OpenAI continues to succeed, right?

I mean, it's in our interest. I also think it's a company that is iconic in this platform shift. The world is better with OpenAI doing well, and that's sort of the fundamental position. After that, the pace of the tension— to your point— comes from, like in all of these partnerships. Some of it is that cooperative tension, and some of it is that Sam is an unbelievable entrepreneur with a great amount of vision and ambition at the pace he wants to move. Therefore, we have to balance that all out.

And so, therefore, we have to balance that all out, right? What he wants to do, I have to accommodate for so that he can do what he needs to do. He also needs to accommodate for the discipline that we require on our end, given the overall constraints we may have. I believe we'll work it out. The good news here, I think, is that in this construct, we have come a long way.

I mean, these five years have been great for them. It's been great for us. And at least for my part, I'm going to keep going back to that. I want to prolong it as long as I can. It will only benefit us to have a long-term, stable partnership.

When you think about the separate funding and untangling the two businesses, are you guys motivated to do that relatively quickly? I've mentioned that the next step for them would be to have them as a public company. It's such an iconic business and an early leader in AI. Is that the path you see for them moving forward, or do you think it will stay in the relationship we have today?

And that's the place where, Brad, I want to be careful not to overstep, right? Because in some sense, we're not on the board; we're investors like you. At the end of the day, it's their board and their management decision. So, at some level, I'm going to take whatever their cues are. In other words, I'm very clear that I want to support them with whatever decision they make.

And to me, perhaps even as an investor, it's that commercial and IP partnership that matters the most. We want to make sure we protect our interests in all of this, and if anything, bolster them going forward. But I think at this point, people like Sarah, Brad, and Sam are very, very smart folks on this. What makes the most sense for them to achieve their objectives on the mission is what we would be supportive of. Well, maybe we should wrap up and thank you for your time today. I want to conclude on the topic of open versus closed systems and how we should cooperate to usher in safe AI. So, I'll leave it open-ended for you to share your thoughts on some of these differences and debates, as well as the importance of addressing them. One anecdote I would like to mention is that Reuters recently reported that Chinese researchers developed an AI model for potential military use based on Meta's LLaMA.

Right. You know, there are a lot of supporters, like Bill and me, of open source. But we've also heard critics. You said that everybody can distill a model, and we are going to see some of these put to uses that we're not going to be happy about. So how do you think we can come together as a nation and as a collective of companies to usher in safe AI?

Yeah, I think two things. I have always viewed open source versus closed source as two different tactics for creating network effects, rather than just as religious battles. I've thought of them more like two different approaches. This is why I think what Metamark is doing is very smart; in some sense, he’s trying to commoditize even his complement.

It makes a ton of sense to me. If I were in his shoes, I would do that, right? Which is to get the entire world converged. I mean, I think he talks openly and very eloquently about how he wants to be the Linux of LLMs. And I think it's a beautiful model.

In fact, there is even a model there, I think, that sometimes, going back to some of your economics questions, a consortium could be a superior model, quite frankly, compared to any one player trying to do it. Unlike the Linux Foundation, where the contributions were mostly operational expenditures, I always say that Linux wouldn't have happened without, I guess, Microsoft being one of the largest committers to Linux. So were IBM, Oracle, and others. I think there may be a real place for open source, and open source is a beautiful mechanism for that, right?

Which is when you have multiple entities coming together, and so on, and it's a smart business strategy. Then closed source may make sense in a closed source context. After all, we have had many closed source products. Safety is an important but orthogonal issue because, after all, regulations will apply, and safety will be relevant to both sides. One could make arguments that, hey, if everybody's inspecting it, there will be more safety on one side or the other.

So I think of these issues as perhaps best dealt with in a capitalist framework. At least, it's better to have multiple models and foster competition, with different companies choosing different paths. We should be quite strict, and the governments will demand that. In the tech sector today, there's no chance of saying, "Hey, we'll see what happens with the unintended consequences later." No government, community, or society is going to tolerate that. Therefore, these AI safety institutions worldwide will uphold the same standards. Additionally, regarding national security, if there are challenges related to potential national security leaks, people will be concerned about that as well.

So, therefore, I think states and state policy will have a lot to say about which of these models and what the regulatory regime will look like. Well, it's hard to believe that we're only 22 months into the post-ChatGPT era. But it's interesting. When I reflect back on your framework around phase shifts, you have to put Microsoft in a really good position as we emerge into the age of AI. And so, congrats on the run over the last 10 years.

It's been really, you know, a sight to behold. But, you know, it's great. I think both Bill and I get excited when we see the leadership— you, Elon, Mark, Sundar, et cetera— really forging ahead for Team America around AI. You know, I feel, and I think we both feel, pretty incredibly optimistic about how we're going to be positioned vis-à-vis the rest of the world. So, thanks for spending some time with us.

I can't thank you enough for the time, Satya. I really appreciate it. Thank you so much. Bye-bye. Thank you, Brad and Bill.

Take care, Satya. Cheers! As a reminder to everybody, these are just our opinions, not investment advice.