Satya Nadella | BG2 w/ Bill Gurley & Brad Gerstner

30:51So it's fascinating. In fact, when was the last time any of us really went to a business application, right? We licensed all these SaaS applications. We hardly use them. And somebody in the org is sort of inputting data into it. But in the AI age, the intensity goes up because all that data now is easy, right? You're query away. I can literally say, hey, I'm meeting with Bill, tell me about all the companies that benchmarks invested in. It's both taking the web, anything that's in my CRM database, collating it together, giving me a note, what have you? So to some degree, all that I think can be monetized by us and by even these connectors.

31:33But more explicitly, like the thing that could happen really quickly, because there's been talk about it, like would you allow chat GBT on the Windows OS to just start opening random apps and take it? Now that's an interesting one, right? So that over the top computer use, who is going to permit that, right? So which is, is it the user or is it the operating system? Like on Windows, there is quite frankly not anything I can do to prevent that other than some security guardry else, right? So I could definitely, like because I think if they became a secure, like one of my big fears is the security risk, right?

32:13If the malware got downloaded and the malware started sort of actioning stuff, right? That's when it's really dangerous. So I think those are the ones that we will build into the OS itself, right? Which is some elevated access and privilege that this computer use stuff happens. But at the end of the day, the user will be in control on an open platform like Windows. And I'm sure Apple and you know, Google will have a lot more control, so they won't allow it. And so that's in some sense, you could say that's an advantage they have or depending on how AT rules on all of those, ultimately it will be an interesting thing to watch.

32:54We've flipped 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 email through a Microsoft client on that smartphone? Yeah, I mean, we kind of like, you know, for example, today, you know, one of the things I always think about is I don't know whether that was value leaking or did it actually help us, right? Which is we licensed the sync for outlook to Apple for Apple mail. It was kind of an interesting case. And I think that there was a lot of value leaked perhaps, but at the same time, I think that was one of the reasons why we were able to hold on to exchange, right?

33:43It would have been doubly problematic. Understood. If we had not done that. And so one of the things I think is going to your point, Bill, if we are building out, the reason we are going to do this is we have to have a trust system around Microsoft 365. We just cannot sort of say, hey, any agent comes in and does anything because after all, first, it's not our data. It's our customers data, right? It would be. And so, therefore, the customer will have to permit it. The IT folks in the customer will have to permit it. It's not like some blanket flag I can set. And then the second thing is it has to have a trust boundary.

34:22So I think what we will do is it's kind of, and it's an interesting way. It's kind of like what Apple intelligence is doing. Think of it as we will do that around M365. So if you go through a lot today, I'd highly recommend people download it. It's super interesting. Yeah. So such a, you know, clicking on this, you know, Mustafa has said that 2025 will be the year of infinite memory. And Bill and I have talked a lot, dating back to the start of this year, that we think the next 10x function, you know, it sounds like you agree on chat GBT is really, you know, this persistent memory, um, combined with being able to take some actions on our behalf.

35:05So we're already seeing the starts of memory. And I'm pretty convinced as well that 2025, where, you know, it seems like that one's pretty well solved. But this question of actions, when am I going to be able to say to chat GBT, book me the four seasons in Seattle next Tuesday at the lowest price, right? And Bill and I, you know, have gone back and forth on this one. And it would seem that computer use is the early, you know, test case for that. But do you have any sense is it, you know, does that seem like a hard one from here to you? Yeah, I mean, the most open ended, um, action space is still hard.

35:46But do you point? There are two things or maybe three things that are really exciting beyond, I'll just say, I'm sure we'll talk about it, the scaling laws itself and capabilities of the raw models. One is memory. The other is tools use or actions. And the other one I would say is even, um, entitlements, right, which is, you know, what can you like, you know, one of the most interesting products we have even is per view inside a Microsoft because increasingly, what do you have emissions to? What can you get? You know, you have to be able to access things in a safe way. Somebody needs to have governance on it and what have you?

36:25So if you put all those three things together and this agent is going to then be more governable. And when it comes to actions, it is verifiable. Uh, and then it has memory. Then I think you're off, uh, to a very different place where for doing more autonomous work, so to speak, I still think one of the things I always think is your bill, I like this co pilot as the UI for AI because even in a fully autonomous world, from time to time, you'll raise exceptions, you'll ask for permission, you'll ask for invocation, what have you? And so therefore this UI layer will be the organizing layer. In fact, that's kind of why we think of co pilot as the organizing layer for work, work artifacts and workflow.

37:12Uh, but to your fundamental point, I don't think the models, I take even four or right, even not even going to all one. Four or is pretty good with function calling. So you can do in the enterprise setting, significant more, more so than consumer because consumer web function calling is just hard, uh, where at least in an open ended web, uh, you can do it for a couple of websites, but once you say, Hey, let's go do a book, me a ticket on anything and then just any of there's schema changes on the backend and so on the trip over. And you can teach you that, that's where I think all one can get better if it's a verifiable, autogradeable, uh, sort of process on rails.

37:54Uh, but I think we are maybe a year, year to two years away from doing more and more, but I wasn't, so, but I think at least from an enterprise perspective, going and doing, here's my sales agent, here's my marketing agent, here's my supply chain agent, which can do more of these autonomous tasks. Uh, we built 10 or 15 of them into dynamics, right? Even looking into sort of sub my supplier communications and automatically handling my supplier communications, updating my database, exchanging my inventories, my apply, those are the kinds of things that you can do today, I would say the Mustafa made this comment about near infinite memory in a, yeah, I'm sure you heard it here internally.

38:35Is there any clarification you can offer about that or is that more to come? I think that, I mean, at some level, the idea that you have essentially a type system, um, for your memory, right? That's the thing, right? Which is, it's not like every time I start, I get the idea, he made it sound like you guys had an internal technical breakthrough on this front. Yeah, I mean, we have, like, there is an open source project even, um, I think it's, I forget, like, it's, uh, it's the same set of folks who did all the type, uh, typescript stuff, uh, we're working on this. So what we're trying to do is, uh, essentially take memory and schematize it and sort of make it available such that you can go.

39:24Like each time I start, let's just imagine I'm prompting on some new prompt. I know how to cluster based on everything else I've done. And then that type matching and so on, I think is a good way for us to build up a memory system. So shifting maybe to enterprise AI such, uh, you know, the Microsoft AI business has already reported to be about $10 billion. You've said that it's all inference, um, and that you're not actually renting raw GPUs to others to train on because your inference demand is so high. So as we think about this, there's a lot of, I think, skepticism out there in the world as to whether or not major workloads are moving, you know, and, and, and so if you think about the key revenue products that are that are people are using today and how it's driving that inference revenue for you today and how that may be similar or different from Amazon or Google, I'd be interested in that.

40:21Yeah, I think that's a good. So the way for us, this thing has played out is you've got to remember, most of our training stuff with OpenAI is sort of more investment logic, right? So it's sort of not in our quarterly results. It's more in the other income, right, based on our investment. So, um, the, so that means the only little loss that's in maybe or loss other income or loss, right? That is right. That is right. Right now, that's how it shows up. And so the, um, so most of the revenue or all the revenue is pretty much our API business or, in fact, to your point, chatGPTs inference costs are there, right?

41:06So that's a different piece. So the fact is the big hit apps of this era are what? ChatGPT, Copilot, GitHub Copilot, and the APIs of OpenAI and Azure OpenAI, right? So in some sense, if you had to list out the 10 most sort of hit apps, you know, these would probably be in the four or five of them.

41:37And so OpenAI has had, which is we've had two years of runway, right? Pretty much uncontested to your point. Bill made the point about, hey, everybody's awake, but, and it might be, I don't think there will be ever again, maybe a two -year lead like this. Who knows? You know, it's all, you say that and somebody else, you know, drops some sample that's on me blows the world away. But that said, I think it's unlikely that that type of, you know, lead could be established with some foundation model. And, but we have that advantage. That was the great advantage we've had with OpenAI. OpenAI was able to really build out this escape velocity with ChatGPT.

42:18But on the API side, the biggest thing that we were able to gain was, you know, take, you know, Shopify or Stripe or Spotify. These are not customers of Azure. They were all customers of GCP or they were customers of AWS. So suddenly we got access to many, many more logos who are all, quote, unquote, digital natives who are using Azure in some shape of fashion and so on. So that's sort of one. And when it comes to the traditional enterprise, 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.

42:59But like, 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. And that's one of the reasons why the adverse selection problems here would have been lots of tech startups all looking for their H100 allocation in small batches, right? That, but having watched what happened to Sun Microsystems in the sort of dot com, I always worry about that, which is, whoa, if, you know, you just can't chase everybody building models. In fact, even in the, I think the investor side, I think the sentiment is changing, which is now people are wanting to be more capital light and build on top of other people's models and so on and so forth.

43:47And if that's the case, you know, everybody who was looking for a H100 will not wonder, you know, want to look for it more. So that's kind of what we've been selective on. And your sense is that for the others that training of those models and those model clusters was a much bigger part of their AI revenue versus yours. I don't know. I mean, this is where I'm speaking for other people's thoughts. I don't, I mean, it's just I go back and say, what are the other big hit apps? Right. I don't know what they are. Like, I mean, where do they like, what models do they run? Where do they run them? And I would, like, that's kind of, I'm not, I mean, obviously Google's Gemini.

44:31I don't know that when I look at the down numbers of any of these AI products, there is chat GPT. Right. And then there is, you know, like, even Gemini, I am very surprised at the Gemini numbers. I mean, obviously, I think it will grow, you know, because of all the inherent distribution. But it's kind of interesting to say that they're not that many. In fact, we talk a lot more about AI scale, but there is not that many hit apps, right? There is chat GPT, get up co -pilot, there's co -pilot and there's Gemini. I think those are the four, I would say, in a Dow, like, is there anything else that comes to your mind?

45:10Well, I think there, you know, there are a lot of these startup use cases that I think are starting to get some traction, kind of bottoms up. A lot of them build on top of, of Lama. But, you know, but if you said, oh, and there's Meta, right? But if you said that, more, what are the apps that have more than five million Dow, right? I think, yeah, it's pretty interesting. I think Zuckerberg would argue Meta AI certainly, you know, has more, it's true. But I think you're right in terms of the non -affiliated apps you name them. And Zuck's stuff will run on his own cloud. I mean, he's not running on public cloud.

45:45So, on the enterprise side, obviously, the coding spaces, often to the races, and you guys are doing well, and there's a lot of venture back players there. On some of the productivity apps, I have a question about the co -pilot approach, and I guess Mark Binioste been kind of obnoxiously critical on this front and called it Clippy -2 or whatever. Do you worry that someone might think, kind of, first principles AI from ground up, and that some of the infrastructure, saying in Excel spreadsheet, isn't necessary to know if you did a AI first product? And the same thing, by the way, could be said about the CRM, right?

46:32There's a bunch of fields and tasks that may be able to be obfuscated for the user. Yeah, I mean, it's a very, very, very important question. The SaaS applications are biz apps, so let me just speak of our own dynamics thing. The approach at least we are taking is, I think the notion that business applications exist, that's probably where they'll all collapse, right, in the agent era, because if you think about it, right, they are essentially crud databases with a bunch of business logic. The business logic is all going to these agents, and these agents are going to be multi -repo crud, right?

47:21So they're not going to discriminate between what the backend is. They're going to update multiple databases, and all the logic will be in the AI tier, so to speak. And once the AI tier becomes the place where all the logic is, then people will start replacing the backends, right? We have people, you know, that's what, in fact, it's interesting. As we speak, I think we are seeing pretty high rates of wins on dynamics, backends, and the agent use. And we are going to go pretty aggressively and try and collapse at all, right, whether it's in customer service, whether it is in, you know, by the way, the other fascinating thing that's increasing is just not CRM, but even our what we call finance and operations, because people want more AI native biz apps, right?

48:14That means the biz app, the logic tier can be orchestrated by AI and AI agent. So in other words, co -pilot to agent to my business application should be very seamless. Now, in the same way, you could even say, hey, why do I need Excel? Like, interestingly enough, one of the most exciting things for me is Excel with Python is like GitHub with co -pilot, right? That's essentially, so what we have done is when you have Excel, like this, by the way, it would be fun for you guys, right? Which is you should just bring up Excel, bring up co -pilot, and start playing with it, because it's no longer like, oh, you know, it is like having a data analyst.

48:59And so it's no longer just making sense of the numbers that you have. It will do the plan for you, right? It will literally like how GitHub co -pilot workspace creates the plan, and then it executes the plan. This is like a data analyst who is using Excel as a sort of row column visualization to do analysis scratch -pad. So it's got a tool's use. So the co -pilot is using Excel as a tool with all of its action space, because it can generate, and it has Python interpreter. That is, in fact, a great way to reconceptualize Excel. And 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?

49:42So therefore, you can generate anything. And so, yes, I think there will be disruption, but so the way we are approaching, at least our M365 stuff is one is build co -pilot as that organizing layer UI for AI. Get all agents, including our own agents. You can say the Excel as an agent through my co -pilot. Word is an agent. It's kind of specialized canvases, which is I'm doing a legal document. Let me take it into pages and then to Word, and then have the co -pilot go with it. Go into Excel and have the co -pilot go with it. And so that's sort of a new way to think about the work and workflow. You know, one of the questions I hear people ringing their hands about a lot today, such as the ROI, people are making on these investments.

50:31You know, you have over 225 ,000 employees. Are you leveraging AI to increase productivity, reduce costs, drive revenues in your own business? If so, kind of what are the biggest examples there? You know, and maybe to a finer point on that, you know, well, we had Jensen on. I asked him, you know, when he two or three X to his top line, what did he expect his head count to increase by? And he said 25%. And when asked why, he said, well, I have 100 ,000 agents helping us do the work. So when you two or three X, your revenue for Azure, you know, do you expect to see that similar type of leverage on on head count?

51:13Yeah, I mean, it's, it's top of mind and top of mind for both asset, Microsoft as well as customers. Here's the way I come at it. I love this thing of, I've been going to school on learning a lot about what happened in industrial companies with Lean. Yeah. Right. I mean, it's fascinating. Right. They're all GDP plus growers. It's unbelievable. Like, I mean, the discipline they have in how the good industrials can literally say, Hey, I'll add two to three, you know, 100 basis points of tailwind just by Lean, which is increased value, reduced waste, right? That's the practice. So I think of AI is the Lean for Knowledge work.

52:01You know, we are really going to school on it. Like, which is how do we really go look at that's why I think, you know, the good old, you know, we remember in the 90s, we had all this business process reengineering. I think it's back in a new way where people who can think into and process flows and say, Hey, what's the way to think about the process efficiency? What can be automated? What can be my, my, made more efficient? So that's a little bit of, I think, so customer service is the obvious one. Like, we are on course, we spend around $4 billion or so. This is everything from Xbox support to Azure support.

52:38This is really, I mean, this is serious one year because of the deflection rate on the front end, then the biggest benefit is the the agent efficiency, right, where the agent is happier, the customer is happier, and our costs are going down. And so that's, I think the most obvious place and that we have in our contact center application that's also doing super well. The other one is obviously get up co -pilot, that's the other, and with get up co -pilot workspace, right, that's the first place where even this, what is a genetic sort of side comes in, right? As you go from an issue to a plan to, or to a spec, to a plan, and then multi -file edit, right?

53:23So it's just completely changes the workflow for the engine team, as I said, and then the O365 is the, you know, the catch -all, right? So the M365 co -pilot is where, I mean, just to give you a feel, like even my own, right, every time I'm meeting a customer, I would say the workflow of the prep of the CEO office is not changed since 1990, right? Basically, I mean, in fact, one of the ways I look at it is, just imagine how did forecasting happen pre -PCs and post -PCs, right? There were faxes, then interoffers memos, and then PCs became a thing, and people said, hey, I'm just going to put an Excel spreadsheet and email and send it around, and people will enter numbers, and we will have a forecast.

54:11The same thing is happening in the AI era, right now, all over the place, right? I prep 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 tells me everything from my CRM, my emails, my teams meetings, and the web, right? It grounds it. I put it into pages, share it with my account team in real time. So just imagine the hierarchy, this entire thing of, oh, let me prepare a brief for the CEO goes away. It's just a query away. I generate a query, share a page if they want to annotate it, so I'm reasoning with AI, and collaborating with my colleagues, right?

54:50That's the new workflow, and that's happening all over the place. Somebody gave me this example from supply chain. Like somebody said, supply chain is like a trading desk, except it doesn't have real time information, right? That's kind of what it is. So it's like, you wait for the quarter to end, and then the CFO comes and bangs you on the head, as saying all the mistakes you made. What if that financial analyst essentially can be in real time, be available to you, and giving you like, oh, you're doing this contract for this data center, in this region, you should think about these terms. All that intelligence in real time is changing the workflow and work artifacts.

55:29So lots and lots of use cases all around. And I think you're to a fundamental point, our goal is to kind of create operating leverage through AI, right? So I think headcount will, in fact, one of the ways I look at it and say, is a total people costs will go down, our cost by head will go up, and my GPU per researcher will go up. That's kind of the way I look at it. That makes sense. Hey, let's shift ahead to something that you referenced earlier, just around what we're seeing out of model scaling and capex generally. I've heard you talk about Microsoft's capex. I imagine in 2014, when you took over, you had no idea that the capex would look like it does today.

56:18In fact, you've said it looks increasingly these companies look more like industrial company capex than traditional software companies. Your capex come from about 20 billion in 2020 to maybe as high as 70 billion in 2025. You've earned a pretty consistent return on that capex. Right? So there's actually a very high correlation when you look at your capex to revenue. Some people are worried that that correlation will break. And even you have said, maybe at some point there's going to be, capex is going to have to be spent ahead of the revenue. There may be an air pocket. We have to build for this resiliency.

57:02So how do you feel about the level of capex? 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, a couple of different things. One is this is where being a hyperscale up, I think structurally super helpful. Because in some sense, we've been practicing this for a long time, right? Which is, you know, hey, data centers have 20 -year life cycles. Power, you pay only when you use the kits are six years. You know how to sort of drive utilization up. These are, and the good news here is it's kind of like capital intensive, but it's also software intensive.

57:51And you use software to bring the ROIC of your capital higher. That's kind of like when people even in the early days said, hey, how can a hyperscaler ever make money? Because what's the difference between old holsters and the new hyperscalers? It is software, right? And that I think is what's going to apply even to this GPU physics even, right? Which is, hey, you buy, you build it out. In fact, one of the things that's happening right now is what I'll call catch up, right? Which is we built after all over the last 15 years, the cloud. Suddenly 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, right?

58:47So if you sort of say, oh, I need all three, you suddenly had to build up these AI accelerators in order to be able to provision for all of these applications. So that will normalize. So the first thing is the build out will happen. The workloads will normalize. And then it will be you will just keep growing like the cloud has grown. So that's sort of the one side of it. And that's where avoiding some of these adverse selection issues, making sure it's not just all supply side, you know, everybody's sort of building, only hoping demand will come, just making sure that there is real diverse demand all over the world, all over the segments I watch for all of that.

59:27So I think that that's I think the way to manage the ROIC. And by the way, the margins will be different, right? This goes back to the very early dialogue we had on when I think about the Microsoft cloud, the margin profile of a raw GPU versus the margin profile of fabric plus GPU or foundry plus GPU or or get up copilot add on to m365. So they're all going to be different. And so if you're having a portfolio matters here, right? Because if I look at even the Microsoft, why does Microsoft have a premium today in the cloud? We are bigger than Amazon growing faster than Amazon with better margins than Amazon because we have, you know, you know, all these layers.

1:00:13And that's kind of what we want to do even in the AI era. So actually there's been a lot of talk about model scaling. And obviously there was talk historically about kind of 10Xing the cluster size that you might do over and over again, not, you know, once and then twice. And next that AI is still making noise about going in that direction. There was a podcast recently where they kind of flipped everything on their head and they 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.

1:00:52I'm curious. Those are two kind of views of the same coin. But what's your view on on large LLM model scaling and training costs and where we're headed in the future? Yeah, I mean, you know, this, I mean, I'm a big believer in scaling laws. I'll sort of first say. And in fact, if anything, the bet we placed in 2019 was on scaling laws and I stay on, right, which is in other words, don't bet against scaling laws. But at the same time, let's also be grounded on a couple of different things. One is these exponentials on scaling laws will become harder just because as the clusters become harder, everything, I mean, the distributed computing problem of doing large scale training becomes harder.

1:01:41And 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, but they are, you know, continuing to, you know, pre -training, I think, is not over. It sort of continues. But the exciting thing, which again, OpenAI has talked about and Sam has talked about is what they've done with all one, right? So this chain of thought with auto grading and is just a fantastic. In fact, you know, basically, it is test time computer, inference time compute as another scaling law, right? So you have pre -training and then you have effectively this test time sampling that then creates the tokens that can go back into pre -training, creating even more powerful models that then are running on your inference, right?

1:02:31So therefore, that's, I think, a fantastic way to increase model capability. So the good news of test time or inference time compute is sometimes, you know, running of those all one models means the run, you know, there's two separate things. Sampling is kind of like training when you're using it to generate tokens for training for your pre -training, but also customers when they, you know, are using O1, they're using more of your meters. And so you are getting paid for it. And so therefore, there is more of an economic model, right? So therefore, I like it. In fact, that's where I said I have a good structural position with 60 plus data centers all over the world.

1:03:10It's a different hardware architecture for one of those scaling versus the other for the pre -training versus the other. Exactly. And I think the best way to think about it is it's a ratio, right? So going back to sort of Brad's thing about ROIC, this is where I think you have to sort of really establish a stable state. In fact, you know, whenever I've talked to Jensen, I think he's got it right, which is, look, you kind of want to buy some every year, not buy, like think about it, when you depreciate something over six years, the best way is what we have always done, which is you buy a little every year and you age it, you age it, you age it, right?

1:03:46You use the leading node for training and then the next year it makes it goes into inference. And that's sort of the stable state. I think we will get into across the fleet for both utilization and the ROIC and then the demands meet supply. And like basically, to your point about everybody saying, oh, wow, have the exponential stopped. One of the other things is the economic realities will also sort of stop. 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, but I'm not able to sell that inventory and the other problem is the winner's curse, right?

1:04:26Which is, if you don't even have to publish a paper, the other folks have to just look at your capability and do up either a distillation, it's just impossible. It's kind of like piracy, right? I mean, you can sort of all kinds of terms of views, but it's impossible to control distillation. That's one. Second thing is, you know, you don't even have to do anything. You just have you reverse engineer that capability and you do it in a more compute efficient way. And so given all this, I think there will be a governor on how much people will kind of chase right now a little bit of it. Everybody wants to be first.

1:05:03It's great, but at some point, all the economic reality will set in on everyone. And the network effects are in the app layer. So why would I want to spend a lot on some model capability with the network effects are all on the app layer? What I heard you say, I believe, you know, so Elon has said that he's going to build a GPU cluster. I think meta has said the same thing. I think the pre -training 200 and any kind of joke to about a million, but I think he joked about a billion. But you know, the fact of the matter is have your versus the start of the year, Sasha, based on what you've seen around pre -training and scaling, have you changed your infrastructure plans around that?

1:05:50And then I have a separate question with regard to 01. I am building to what I would say is a way like it's a little bit of the 10x point, right, which is, hey, how do you, we can argue the duration? Like is it every two years? Is it every three years? Every four years? There is an economic model. And this is where I think a little bit of disciplined way of thinking about how do you clear your inventory? Such that it makes sense, right, which is, or the other way is the depreciation cycle of your kit, right? There is no way you can sort of buy, you can pre - unless you find the physics of the GPU works out where suddenly it flows through my PNL and it's actually, you know, it's in the same or better margin than a high -press scale.

1:06:42That's simple. Like, so that's kind of what I'm going to do. I want to keep going and building basically to, hey, how do I drive inference demand and then keep increasing my capability and be efficient at it? I absolutely, and Sam may have a different objective and he's been open to it, right? He's sort of like he may say, hey, I want to build because I know I'm deeply, have deep conviction on what AGI looks like or what have you. And so be it. So therefore, that's where I think a little bit of our tension is even. And 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 going on.

1:07:27Is that fair? Well, what we won't do is do it twice, right? Because after all, we have the IP from, right? It'd be silly for Microsoft today, given the partnership with OpenAI to do two unnecessary, I mean, I'm just doing a second training, sir. Yes. Correct. So we are ready. And that's why we have concerned. And by the way, that's the strategic discipline we have had, right? Which is, you know, that's why, you know, I always stress to Sam, like we bet the farm on OpenAI and said, hey, we will concentrate our compute. And we did it because we had all the rights to the IP. And so that's sort of the give -gets on it.

1:08:08And we feel fantastic about it. And so then what Mustafa is basically saying is, hey, we will also do, in fact, a lot of focus on our end is post -training and even on the the verification of what have you. So that's a big thing. So we'll focus a lot of our compute resources on adding more model adaptations and capabilities that make sense. While also having a principled pre -training stuff that sort of gives us capability internally to do things, we anyway have different model bates and model classes for different use cases that we will continue go ahead and develop as well. Is your question to Brad's question about your answer to Brad's question about the balancing of GPO, GPURY?

1:08:54Does that answer the question as to why you've outsourced some of the infrastructure to CoreWeave in that partnership that you have? That we did because we all got caught with the hit called ChatGPD and OpenAI APIs. Yeah, we were completely. I mean, it was impossible. There's no supply chain planning I could have done in what is it? Like none of us knew what was going to happen. What happened in November of 22? That was just a bolt from the blue. So therefore, we had to catch up. So we said, hey, we're not going to, in fact, worry about too much inefficiency. So that's why whether it's CoreWeave or many others, we bought all over the place.

1:09:43Fair enough. And that is a one -time thing. And then now it's all catching up. So that was just more of trying to get caught up with the man. Are you still supply constraints, Sasha? I am power. Yes, I am not chip supply constrained. We were definitely constraining 24. What we have told the street is that's why we are optimistic about sort of the first half of 25, which is the rest of our fiscal year. And then after that, I think we'll be in better shape going into 26 and so on. We have good line of sight. So I'm hearing with respect to this level two thinking, the O1 test -time compute, post -training work that's being done on that is leading to really positive outcomes.

1:10:38And when you think about that, that's also pretty compute intensive. Because you're generating a lot of tokens, you're recycling those tokens back into the context window, and you're doing that time and time again. And so that compounds very quickly. Jensen said he thought looking at O1, the inference was going to a million or a billion X, just that it was the demand for inference is going to go up 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, which is in some sense, it will be, it's very helpful to think about the full workload there.

1:11:21The full workload, like in the Agentec world, you have to have the AI accelerator. One of the fastest growing things of, in fact, OpenAI itself is the container service. Because after all these agents need a scratch pad for doing some of those auto grading even, to generate the samples. That is where they run a code interpreter. That, by the way, is a regular Azure Kubernetes cluster. Interesting with there's a ratio of even what is regular Azure compute and its nexus to the GPU, and then some data service. 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.

1:12:06AI is now 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 -agent performs actions, then classic app server plus the AI app server, plus the database are all required. And so I go back to my fundamental thing, which is, hey, we built this 60 plus AI regions, I mean, Azure regions, they all will be ready for full -on AI applications. And that's, I think, what will be needed. That makes a lot of sense. So let's talk a little bit. We've talked around to Open AI a lot during this conversation, but you're managing this balance between a huge investment there in your own efforts.

1:12:59At Ignition, you showed a slide highlighting the differences between Azure Open AI and Open AI Enterprise. And a lot of those were about the Enterprise grade things that you bring to the table. So when you look at that tension, the competition that you have with Open AI, do you think about them as ChachiBT is likely to be that winner on the consumer side? You'll have your own consumer apps as well. And then you'll divide and conquer when it comes to Enterprise. How do you think about competing with them? The way I think about at this point, given Open AI is a very at -scale company, right? So it's no longer, it's a really very successful company with even multiple lines, if you will, of business and segments and what have you.

1:13:57And so I come at it very principally, like I would with any other big partner, right? Because I don't think of them. So 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. And because we give them systems IP, they give us model IP, right? So that's another side of it where we are very deeply interested in each other's success. The third is I think of them as a big customer. And so therefore, I want to serve them, like I would serve any other big customer. And then the last one is the Co -op Petition, right?

1:14:43Which is whether it's Co -Pilot in the consumer space, whether it's Co -Pilot with M365 or whatever else, we sort of say, hey, where is the Co -op Petition? Where is, and that's where I kind of look at it and say, you know, ultimately these things will have some overlap. But I also in .context, the fact that they have the Apple deal is in some sense for the MSFT shareholder, you know, a creative, right? Even in like the fact that they're their APIs, like to your point about the API differences, hey, you choose, right? The customers can choose which API front or like some of the, you know, there's differences, right?

1:15:22Azure 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 an Azure and Azure Mac. But if you're on AWS and you want to just use just the API in a stateless way, great, just to use even open AI. So I think in an interesting way, there's sometimes having these two types of distributions is also helpful to the MSFT cost. Such as the, I would say the kind of curious part of the Silicon Valley community and even writ larger, I would say the entire business community is, I think, infatuated with the relationship between Microsoft and APNAA.

1:16:02I 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 a supposedly a restructuring, conversion to profit, I guess, Elon's launched a mis -sub in there as well. What can you tell us? Yeah, I mean, I think those build obviously all for the OpenAI board and Sam and Sarah and Brad and that team to decide what they want to do and we want to be supportive. I mean, so this is where we're an investor. Let me, I would say the one thing that we care deeply about is OpenAI continues to succeed. I mean, it's in our interest and I also think it's a company that is an iconic company of this platform shift and the world is better with OpenAI doing well.

1:16:56Therefore, that's sort of the fundamental position. Then after that, the pace with which the tension to your point comes from, like in all of these partnerships, some of it is that co -op petition tension, some of it is, you know, Sam's somebody who is an unbelievable entrepreneur with great amount of sort of vision and ambition and the pace with which he wants to move. And so therefore, we have to balance that all out, which is what he wants to do. I have to accommodate for so that he can do what he does to do and he needs to accommodate for the discipline that we need on our end. Given the overall constraints we may have.

1:17:39And so I think we'll work it out. But I mean, the good news here, I think, is in this construct. We have come a long way. I mean, this five years has been great for them. It's been great for us. And at least from my part, I'm going to keep going back to that. And I want to prolong it as long as I can. It will only be who wants to have a long -term stable partnership. When you think about the separate funding, and untangling the two businesses, Sasha, are you guys motivated to do that relatively quickly? I've talked about thinking that the next step for them would be great to have them as a public company.

1:18:20It's such an iconic business, an early leader in AI. Is that the path that you see for these guys on the way forward? Or do you think that it stays kind of in the relationship that we are today? And that's the place for Brad. I want to be careful not to overstep, right? Because in some sense, I'm neither in the board, we're investors like you. At the end of the day, it's their board and their management decision. And 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.

1:19:13We want to make sure we protect our interests in all of this. And if anything boils to them going forward. But I think, at this point, people like Sarah and Brad and Sam are very smart folks on this and 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. And thank you for so much time today. But I want to wrap on this topic of open versus closed, you know, and how we should cooperate to usher in safe AI. And so maybe I'll just leave it open, ended to you. You know, talk to us a little bit about how you think about some of these differences and debates and the importance of doing this.

1:19:59One anecdote I would just throw out there is Reuters recently reported that Chinese researchers developed an AI model for potential military use on the back of Meta Lama. And there are a lot of supporters like Bill and I of open source. But we've also heard critics. And you said everybody can distill a model out there. So 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 about, you know, us coming together really as a nation and as a collection of companies to usher in safe AI? Yeah, I think two things. I think that I always have thought of open source versus closed source as two different tactics in order to create network effects.

1:20:49Right? I've never thought of them as just religious battles. I've thought of them as more like, hey, two different, I mean, that's why I think what Meta and Mark are doing is very smart, right? Which is in some sense, he's trying to commoditize even his complement, right? It makes a ton of sense to me. If I were in issues, I would do that, right? Which is get the entire world converged. I mean, I think he talks openly and very eloquently about how he wants to be the Linux of LLAMs. And I think it's a beautiful model. In fact, there is even a model there. I think there, you know, sometimes to you going back to some of your economics question, I think there is like the game theoretically a consortium could be a superior model quite frankly.

1:21:37There's any one player trying to do it. Like this has unlike the Linux foundation where the contributions were mostly apex contributions, right? Which is if I always say Linux wouldn't have happened, but for I guess, you know, or in fact, the Microsoft's one of the largest committers to Linux. And so was IBM, so was Oracle and what have you? And I think that they may be a real place for and open sources of 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 closed source.

1:22:15After all, we have had lots of closed source products. Then Syfty is an important but orthogonal issue because after all, regulations will apply and safety will apply to both sides. And you know, one could make arguments and hey, everybody's inspecting it and, you know, there will be more safety on one side or the other. So I think of these as perhaps best dealt with in capitalism at least, it's better to have multiple models and let there be competition and different companies will choose different paths. And then we should be pretty hard core and the governments will demand that. I think in attack, you know, now there's no chance of saying, hey, we'll see what happens to the unintended consequences later.

1:23:00I mean, no government, no community, no society is going to tolerate that. So therefore, these AI safety, you know, institutions all over will hold a same bar and also national security to your point. If there is sort of national security leakage challenges, the people will worry about that too. 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 well, it's hard to believe that we're only 22 months into the post chat GPT era. You know, but you know, it's interesting when I reflect back on your, you know, framework around phase shifts, you have to put Microsoft in a really good position as we emerge into the age of AI.

1:23:51And so congrats on the run over the last 10 years. It's been really, you know, a sight to be hold, 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, you know, really forging ahead for team America around AI. You know, I feel, I think we both feel pretty, 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. You can't thank you enough for the time, Sacha. Really appreciate it. Thank you so much. Thank you, Brad. Bill, thank you. Take care, Sacha.

1:24:36As a reminder to everybody, just our opinions, not investment advice.