Microsoft Corporation (MSFT) Earnings Call Transcript & Summary

As the music dies down, the AI activity and IQ from our next guest is going to heat up. Did I set you up too much? I think I've heard good things about...

High expectations. High expectations. Yes.

Well, our next guest is Jared Spataro. I will have Jared walk us through his background. We would love to find out what you've been doing at Microsoft for the short span of 20 years. You look very young, and I find it very hard to believe that you've been at Microsoft 20 years. But tell us about your background and career.

Sure. So I was educated as a computer scientist and started there. It's always been my interest, and energy has been there. When I came to Microsoft, it wasn't at the beginning of my career, so I didn't start there. I worked at smaller firms prior to that. I was in product, essentially product management, and I went to Office. So I worked on the Office set of products. Over the course of the years, I started to migrate more towards the business side of what we do. So we put in product marketing as the business management side, and we typically will pair that up with an engineering lead. So now I'm responsible for a portfolio that we just call AI business solutions. But over the years, I ran everything from Office Commercial to Office Consumer to Windows to Office 365 to Microsoft 365. The work during the pandemic on Teams was all work that I did. And then just when I thought I was getting bored, this whole thing started to happen with AI, OpenAI and Copilot. So that's what I'm up to now.

Got it. So although you're a technical person by education, your title is Chief Marketing Officer. Is the bar so high at Microsoft to be in marketing?

Well, these days, I think it helps a lot to deeply understand the tech for sure. Like you're -- I don't know. I love Microsoft because it's a place where we expect our business people and the CMO, if you will, to not only understand the tech but to help shape it. And that's the part of my job I like the most, is trying to chart the future where things are going.

Great. On that topic, how do you envision AI evolving over time? So Microsoft has got a unique vantage point being super, super early. So explain to us how you envision the AI tech stack evolving over time, your vantage point and Microsoft's role in this evolution.

Yes. Let me talk tech stack, but I'll be brief. I don't want to go too deep, but you certainly can ask more questions. We start at the hardware level, obviously, all the way at the chip level, data center level. There's been a lot of innovation not just in chips but in data centers over the last couple of years. On top of that, we actually start next with the data layer. There's some important work that we've done, particularly with a set of investments that have gelled into what we call Fabric now. Then we have something called foundry, where we think about the model layer. This is where the LLM sits, and that's been important for us. Next up, we move to the dev layer. Here, we split things between what we think of as pro code and low code, so those things that are designed for makers versus professional developers. And then finally, we round things out a little bit differently than most because Copilot for us is kind of that front end, if you will. We think of it as a platform that we're really trying to draw users into not with a daily habit but in fact, with an hourly habit in the knowledge worker space. So that's how things compose. We spent a lot of work, if I were to take a step back though, trying to project out 3 to 5 years and say, well, what's the firm of the future look like. We've decided to call that idea of the frontier firm just using that frontier label from models. And we're pretty excited about what that looks like. We can describe it today as a firm that we would say is human led but agent operated. And we think that it behaves very differently. We think that humans play an important role, but the patterns of work between humans and agents are going to be very transformative. So that's kind of where we are. We see Microsoft's role in there is helping businesses, organizations get from today to that frontier firm. That's how we would frame up how I have conversations with CEOs and their teams, technical teams and business teams.

Got it. We all watched the launch of GPT-5. I'm sure you did. What is your reaction to GPT-5 launch? What do you make of the technological enhancements? And I have a follow-up question as it relates to Microsoft 365.

Yes, this one has been -- this has been interesting for me because for me, it separated out a little bit people who understand what's happening versus those who are kind of superficially just tracking the technology. Let me explain what I mean. GPT-5 was a systems launch. A lot of people focused on, hey, wait a second, we don't see -- we were looking for nonincremental model differences. And I understand that. I understand the expectation. But if you get to the heart of what was going on, the biggest thing that was introduced with GPT-5 was at the orchestration layer, what they call the router, move from being a deterministic hard-coded thing to being an LLM model that had been posttrained to orchestrate other models and tool usage. And so today, with GPT-5, when you send in a prompt, it's deciding do I route this to a reasoning model. Do I route this to a cheaper model? Going forward, that system, I can't explain how important that system is going to be for actually realizing the value of agents and the value of what's happening. So I understand the chatter about I was looking for more in terms of the model. This was a really important systems launch, and it exposes, in their case, for that ChatGPT, it exposes the broad swath of free users for the very first time to reasoning models at the right time in an economical type of way. So I think we may talk a little bit more about that, but that systems-based approach, I think, is really, really important because it's the beginning of a system that is going to prove to be very influential.

So you just got me thinking, so the beginning of a system that's going to prove to be very influential. Where do you go forward with this? How do you make this more useful in terms of Microsoft approach to this technology?

Man, it's a great question. We think of -- here's what the state of the art would tell us right now. There is a question about can you create essentially a super app that is so smart model-wise that it can do everything you could ever imagine doing with superhuman capability. And it turns out that the answer, we would think at Microsoft and I think many of the model producers feel, is no. In fact, it's better to go and train agents that are domain- or area-specific that prove to be superhuman and then essentially to have a system that pulls on the right agents at the right time to answer the problem. So we look at this in commercial terms and say, man, we were just given a gift as an industry with essentially an orchestrator, conductor of the symphony that with time is going to be able to register up the right agents and those agents will have real domain expertise and then call on those agents in an orchestrated way to get incredibly sophisticated jobs, not just tasks, but jobs done. That for us is the future of this frontier firm. We expect that companies will produce financial analyst agents. They're going to produce expert scientist agents. They're going to produce actuarial agents. And this router is going to call on them to do the right things in business process. So we look at this moment. That's one of the reasons we sim-shipped GPT-5. We worked so hard to sim-ship that in Copilot, is it's a real watershed moment for us.

Yes. So the bleed over to M365 Copilot, how does this manifest going forward, the GPT-5 improvements in M365 Copilot?

Well, same day, we replaced the back end of Copilot with GPT-5. You -- currently, if you use Copilot, you have to select a button. That was just us making sure putting running water through the pipes, but in the next couple of weeks, you'll see that turn to the default. And what it means for us is it's the realization of a strategy that we simply would call Copilot plus agents. We think of it this way. Copilot is to agents like the iPhone is to apps or in other words, just like the iPhone has become a platform or a window into the world of apps over the last decade, we believe that, in a commercial sense, Copilot will become a window or a conduit into the world of agents. And what we envision happening is Copilot will help you at the right time select the right agents to solve the business problem that's in front of you. That's what our customers are asking for. They're saying, "Hey, I love the power I'm seeing, but I need it to be specific to my business. I need it to be specific to my industry." and Copilot is architected to do exactly that.

Got it. Any examples of that? So you talked about the financial analyst. Any other examples besides putting me out of a job?

Sure. I can give you a bunch of different things. Across the industries right now, customers are working with us, either pro code or with something we call Copilot Studio to create purpose-built agents. These agents typically are being trained on like a reasoning model, and then you can take in their own content and train it. So I'll take pharmaceuticals as an example, and you can take something like the clinical trials or you could take a part of the scientific process. There are a bunch of different jobs that happen along the value chain of clinical trials, and what our customers are starting to do is to build agents to do those jobs. Someone, for instance, is as simple as a technical writer. Another person is -- prepares the submission to the FDA. Other people summarize, for instance, the actual trial data. I mean there's a bunch of things that happen. You can stack those agents up, and you can essentially design an end-to-end process where Copilot can do the orchestration for you of, hey, I want to run this and pick up where the data is today, and I want to start to run it down to the process of getting to an FDA submission. It's not done by Copilot itself. It's done by Copilot orchestrating these specialized agents to do their jobs. And that really moves us into this realm of not just agents but digital workers, digital employees.

I'm going to throw out a bunch of seemingly unrelated terms and ask you to see if you can explain where this concept kind of fits in this, RAG, context -- broader context windows, reinforcement learning, test time compute. In that realm of seemingly unconnected things that I just blabbered out, where would this enhancement to GPT-5 fit in? Is it a broader context window or none of the above, and it's like that's mad rambling, Kash, and it does not make any sense?

No, it's good rambling. I have to put them together though for a second. It was RAG and...

RAG, broader context window, reinforcement learning, so those are all things that are supposed to enhance the core model in some way, add context to it, provide information to it, hold it more in memory. Whereas what you're talking about is getting deeper into domains. Does that happen because of any of these things? Or does the model itself come with the deeper domain and you just have to call upon that, okay, I want to pull upon the financial analyst domain, boom, it goes and pulls [ that ]?

Yes. Okay. I think I can construct in a way, if you're following the conversation out there. Think of what we're describing as a way to create a system of agents that has an orchestrating agent in Copilot, essentially directing traffic or orchestrating a team of specialized agents who do work. Now take, for instance, the things that we just threw out there. All those things that Kash has talked about are ingredients to make typically those specialized agents more effective. So RAG is a way to, for instance, add very specialized context to an LLM so that it can reason over data. The context window creates a bigger kind of space of context for people to be able to do that. So the thing we're talking about here is a system of orchestration. A lot of what we've heard from the industry so far has been almost like incremental improvements on an agent per agent basis. And this is about how do you get nonincremental improvements because you have a set of agents that are working together in a structured way.

Got it. Not step function, but orders of magnitude better.

Yes, that's exactly right. Yes.

That was what was echoed by our previous panelists from Sequoia Capital that he said that people misinterpreted the GPT-5 as an incremental advance over GPT-4.5, and he made it look like it's a lot more than that, which seems to be very consistent with how you characterized it. I've watched Build, and I was fascinated. I was not prepared to be riveted to every second of what happened at Build, and I was riveted. In particular, the MCP, support for MCP protocol, the agent-to-agent protocol and what seemed like you were building a middleware layer, just my mind flashed back to BEA Systems. That may not mean anything to you, but maybe some of us here in the audience might actually recollect BEA Systems. They produced middleware for SaaS web-based applications. Of course, the company went nowhere, but they were -- at the beginning, you could see the future if you envision, right? It felt like you're building middleware for agentic applications. So talk to us, if you don't mind, the implications of these frameworks A2A, MCP as it pertains to AI apps ecosystem development? How important is this?

Yes, it's a great question. To answer the question, I'll back up a little bit and just say there is a huge delta, a gap between the capabilities, the raw capabilities we see of the technology, particularly the models right now and the value that commercial customers are realizing. I think all of us understand that. Otherwise, there would be different things happening in the market. The thing that we see that is the gap to be filled essentially is how do you practically take these very powerful workhorses, these engines and put them to work in the very complex plumbing that is today's enterprise. And as you're talking about MCP and A2A and these protocols that allow agents to talk to one another, that's exactly what we're doing. We see our role at Microsoft as taking the best innovations from all over the industry from all sorts of different providers of those innovations and providing a platform for our customers to build solutions that actually just kick out business results. And it could be whatever the industry is, if it's consulting or if it's pharmaceuticals or whatever happens to be, really produces results. But to do that, you actually have to start to fill the gaps in there. And so for instance, having agent-to-agent communication even for what we were just talking about with the orchestration, it's key. It's foundational. It's really important. Might be -- might not be the most exciting thing, but we're trying to really build that up, build that stack at every layer so that there's connections there and then it's easier and easier to program.

That -- when Satya Nadella talks about foundry, the foundry layer has all these protocols being supported. And then if these protocols are supported, then you would need to pull into a data layer that would have all the semantics, the context, et cetera, to make the agents actually do the work, right?

That's right. And that's -- for us, that's where, as I rattled off, Fabric is that data layer. You don't have to move all your data. Fabric can both do in place as well as data that you move over. Foundry, whatever models you want to with, for instance, communication between them, you've got exactly right. Yes.

Got it. How dependent is the Microsoft Copilot stack on OpenAI's IP? And are you diversifying your tech stack beyond OpenAI?

This must be the question of the hour, I would imagine, right now. OpenAI is our first...

There's something about this MAI model that Microsoft is doing.

Yes. OpenAI is our first and best partner. No doubt about it. We have a very unique relationship, but it's a very important relationship for both companies. And the Microsoft 365 Copilot runs on the OpenAI models. GPT-5 sim-ship was another vote of confidence. We think that they've got the best technology out there. So it's really important to understand that we feel like it's a great partnership. That said, you just talked about the tech stack, and you did a great job of hitting at the layer of foundry. We have been now for some time, for some number of quarters and particularly earlier this year, very open about the fact that we want to give enterprises access to just the best technology out there. So that set of announcements at Build, what most people didn't understand is using foundry, you could choose any model, make an agent and essentially mount it in M365 Copilot already. That means you could choose a Llama as your model. You could choose whatever you wanted to out there from any other provider as a model and put it in there today. So that's been an important part of the way that we are pursuing the business. But OpenAI, first and best partner, we have a very unique and very important relationship.

So the role of MAI is just one other model. It's a Microsoft first-party model. Yes. Where is that, I mean, going? What are the plans for MAI in the context of the foundry layer?

Yes, we would say it this way. We've always said for a long time, Satya was one of the first people to say this, that we believe that models will become commoditized over time. There are many ways to hear that statement, but one way to understand the statement is there is some threshold that when you pass, you're already so valuable that what you're putting on top of that becomes essentially diminishing returns for many tasks that you need to do. So us developing our own models is -- it's very consistent with that idea that we believe that these, at some point, are going to become commoditized and that you're going to say, hey, there's diminishing returns. We don't need something more powerful for the general purpose. So frontier models will always have their place, pushing science and pushing the boundaries. And yes, we've said that we're going to develop our own models. We've said that we are not trying to develop frontier models, which is interesting. And so that -- we may change that perspective. You shouldn't take that as anything more than kind of how we've talked about what the MAI models are right now. We've said that it's not meant to compete, for instance, against GPT-5. But the model space continues to change, and we hold that the -- it hasn't reached a place where we feel like it's just slowed down in its overall progress.

We've heard so much about the AI, the threat, potential threat of AI disrupting seat count or put pressure on seats due to layoffs or maybe intentions to hire at a slower rate because of AI's ability to write code, do design and who knows what else it's going to do. Where do you stand on this debate?

We're trying to take both positions, I'll say. So let me talk about both positions. Here's what we'd say. Oftentimes, when people come in and say, "Oh, my gosh, the world is going to change," they think it's going to change overnight or at least that's kind of where the mind jumps to, some amazing technology that's going to change overnight. There's no doubt that if we -- if our hypothesis is true that there will be human-led, agent-operated enterprises, that agents will become more important -- more and more important in doing work and that, that will have an impact on the work that humans do, for sure. But how fast that will happen, what the humans will do, et cetera, we can't predict the future with perfect fidelity. And so we would say this. We have a business model today that has a per-user licensing basis to it with M365. We like that. We continue to grow it. We grow seats. We grow ARPU. At the same time, we increasingly have an agent-based business model where we are charging in 2 different ways per agent, and we charge per consumption. Our customers can choose. So we kind of have our foot in the 2 camps right now, and we think that's the right place for us to be as a business because we're just going to watch how the dynamic plays out over time.

What are the productivity benefits that you're seeing with your customers on account of adopting Copilot at this stage? It's been out for a couple of years now.

Correct, yes. Correct. They fall largely into 3 categories. Category 1, we would call personal productivity and that's about making someone who already does the job mostly do that job more efficiently. And here, we've -- for the most part, we feel great about what we can do. We think, in many tasks, we can improve efficiency by 20% to 30%. And if you do kind of like control and treatment type group experiments, you can show that. The most difficult thing about that is it's tough to drive ROI on saying Kash is 30% more productive like unless he's a salesperson and carries a quota, quite frankly, because a lot of knowledge work doesn't translate directly into top line, bottom line. It's a team that has to work. So we continue to do that work. We feel good about it, but it is hard to make the ROI argument for it. Most of what's happening in terms of actual OpEx savings is happening in process-based applications. So by that, I simply mean people are picking a process out there that runs, that they measure, that they have KPIs with. And they're inserting AI in 2 ways, either incrementally improving that, which is important to some sort of continuous improvement type of way or, in some cases, they're entirely redesigning the process and saying, oh, we think we can do this differently. So to make that very real for you, a lot of people have started, and we've heard that a lot but in customer support and customer service. So incremental improvements, for instance, are throughput by customer human customer service agent. We've done some work at Microsoft, where we've improved throughput of our agents by about 12%. That's real money saved for sure. But there have been other places in customer support, where we just entirely redesigned the system and we tried to do what we call deflection upfront. Never have it get to a human agent. Can we answer your question essentially with AI upfront? That's a different process, and we've been able to say about half of the savings we've realized have come from that deflection. So just real nuts and bolts. It may not be the most exciting thing to talk about, but at Microsoft, we're saving a lot of money, and we see our customers doing it as well.

Well, I use Copilot to draft e-mails, and I make very little corrections to it. And I also ask it to summarize my calendar for the day and what are the priorities. I even ask it to prioritize the e-mails that I need to respond to. And sometimes it'd look at an e-mail and say, oh, #1 important e-mail would be like some client asking me to comment on a wildly speculative thing that's happening that it thinks it's very important because it's worded in such -- you have to please get back to me on this, right? But on that, I think it's been just generally a tool that has been growing. And I start to use Word, not that I use PowerPoint, but I get into it. Copilot starts to show up, and, "Can I summarize this for you?" It's turned -- it started to show proactivity, just like ChatGPT, say, "Would you like me to do more research on this?" It's starting to show up a little bit more and more and more.

Yes. And we're excited about what you'll see from us this fall because we finally feel like with GPT-5, we're starting to get away to be pretty sophisticated about some of the things you're talking about. I'll give another couple of examples quickly or maybe one. A lot of people don't think about it, but we often work outside of our native tongue. You may have to do something in Italian. People have to do things in English that isn't native. We see lots of people using Copilot actually to help them communicate in something that is not their native tongue incredibly eloquently as an example. It's a small thing, but it really matters. It's a great example of something that's not currently being measured but makes a big difference at the personal level.

I think some people would say that Wall Street uses too much jargon in cRPO, RPO, KPI. I mean, translate all that stuff. Lay it for me in English, right? We actually have a jargon dictionary in the Goldman Sachs research. You had to [ head to LinkedIn ]. It's like a massive one, a pretty long one because every industry analyst has their own like jargon list. 1 day, that goes away. As of the Q4 earnings, which were phenomenal, Copilots are now more than 100 million MAUs across commercial and consumer. Where are we in this adoption cycle, the first inning, second inning, third?

We're very early. Yes. We would say we're very early. Our -- in the commercial side, which I'm responsible for, the motion is land and expand. The last data point that we released was a couple of quarters ago, and we talked about the fact that 70% of the Fortune 500 are using Copilot in a pretty extensive way. And those are lands, and we're expanding from there. This past quarter, when we reported, it was our best quarter ever, both in terms of seat adds for Copilot and the number of customers coming back to add seats in a big, significant way. So we feel like we're early. One of the things I said in an earlier meeting today that I'll just repeat is the biggest difference that I would flag for the group, Kash, is that, over the last 6 months, I've seen a marked change in most enterprises, whereas 6 months ago, they were like, well, maybe we'll do something where we provide kind of some sort of assistant at work for everyone. Most companies at this point have decided they're going to do that and that they're just trying to decide who will I go with or what will my estate look like. And that's a big change because once you start to lay down that baseline layer, I think it improves the literacy, helps people to find other ways to use the technology.

Got it. How is Copilot usage tracked relative to your expectations? How are those usage patterns changing?

Well, let me just say my expectation -- we'll talk about this group's expectations for a moment. My expectations would be, when you look at this, it still is our fastest growing M365 portfolio product that we've ever had. And that's saying something because M3 and ME5, like those are big important products that continue to really grow. So I really need to stress if you want to get a sense for it, like it's growing faster than either 1 of those 2. In terms of usage intensity, we are really encouraged. When people start to use it, they do what you just described. They start to use it, and then we see the intensity increase in fact, user by user as a general trend. So we like what we're seeing there. The only workload that we feel like is more intense like during a ramp-up period ended up being Teams, and that's because we have the pandemic, and it was -- there is a single world event. So it's got a lot of intensity. That said, if I were to sit down with each one of you, I think mostly what I hear from financial analysts is what's taken you guys so long. You have 430 million seats out there. Like what's the seat count? What's the penetration? What does this look like? And it's just early innings. It takes a while. This is a business where we've been trying to say we feel like we've got a great position, but we've got to land. We've got to expand. We've got to help people change the way they work. There's some transformation required.

I have to say another Copilot usage, which really was absolutely productive for me, was -- there was a point in time, March or April, I forget. There was this thing, oh, Microsoft is going to cut CapEx, their whole hyperscalers or blah, blah, blah. I had the benefit of your most amazing IR team on the road, and I had conviction. And I -- we wrote up a report, and I wrote up in my e-mails my thoughts on CapEx. And one client -- and there are multiple e-mails sent to multiple clients on this topic, covering multiple aspects. And 1 client called me like 2 minutes before the company was reporting -- is going to be reporting earnings. And he said I want your thoughts on Microsoft CapEx. I said I'll call you right back. I just went to Copilot and said summarize everything that I've said about Microsoft CapEx in 5 bullet points. Boom, I sent it to him, and he was ecstatic. He was like I really appreciate you getting back to me so quick. Thank you, Copilot.

I love it. Copilot win.

I'll not mention the client's name. I don't think he's here. He may be here. I'm not...

I will say that's one of the big differentiators for us, is access to your work data cannot be -- I don't think it can be overplayed. And it's really important to understand...

I checked the e-mail, though, before I sent it.

Good. That's the right thing to do. AI can make mistakes. That work data is very valuable.

Yes, yes. What is -- who's Copilot's competition? When you try to sell this, who do you run into? Is it like ChatGPT modified by the customer to pose like Copilot or...

Yes, there are 2 classes of competitors that we see. We see other chatbots, typically by the model makers out there. And then interestingly enough, we see competition from roll your own. IT is really excited about creating the ACME Corporation chatbot based on GPT-5, like so excited about doing this. So that's an interesting dynamic that I'm not sure I would have predicted. Now I feel like I understand what's going on there. ITs, they're builders. They like to solve problems and build, so we see both those things. For the most part, we really believe that those folks who are rolling their own are not going to be able to keep up. They won't keep up with the iterations of the model. They won't keep up with the innovations and user experience, et cetera. But right now, it is a portion of what we see as also challenging those of us who are making our own kind of assistants at work.

I mean there's this tendency, the beginning point of any new tech cycle that when you have the building blocks, you're excited to build your own thing. And you feel like you're a genius developer, and everybody is a genius developer, creative, et cetera. When things kind of settle down, you don't -- that's not the core competence, and you want to rely on a provider like Microsoft that has watched hundreds of customers and has assimilated the problems and makes the product better. And what if the person that developed this just leaves? You're stuck. Who do you call, right? Not so simple.

Yes. That's right.

The decision on part of Microsoft to price M365 Copilot at $30 a month, does this -- how do you expect this to change over time?

We feel good about the price point. I'll say without going into great detail, we've been able to hold that price point quite nicely. We have lots of, you can imagine, discounting pressures. We do big deals and things like that, but we've held the price point. So that's good. The product is growing into the price point, frankly. That's what happens. You start with technology. It gets better over time. We really feel like, for instance, our researcher and analyst agents that are a part of the price point are examples of us adding additional value and capabilities into it. Our initial pricing of that was based on some research that we did really early out of the gates, looking at what the value would be and trying to not just have it all get squeezed out by customer expectations that would be included into the suite. It was our feel that with 430-plus million users as a base, we were looking for a way to monetize the great gains that we give them. So that's kind of how we started.

Got it. Okay. E5 upgrades have been a big driver of ARPU increases. Could you quantify for us or maybe give us a range what percentage of the base is on E5 today? And how large is the opportunity that's left ahead of you?

Yes. We occasionally publish numbers here. Our last published number actually is not very new, so it's about 3 years old. And 3 years ago, we talked about E5 being 12% of the base. So that's the official number that we gave out. We don't have an updated number that we've shared. But suffice it to say, there's still a lot of upside with E5. So again, you can compare the 12% and think of our latest number being about -- a little over 430 million, the base.

The denominator. Yes. That's right. And Copilot has to be a catalyst for the...

Absolutely. Well, I should also say what we're really excited is there's real synergy. First, E5 and Copilot are the 2 ARPU drivers for the business, but there's great synergy. As you start to think seriously about moving to Copilot, one of the big questions you have is how you're managing your data estate. And that plugs into Purview, which is a component of E5. And so security has always been the driver of E5, but we're increasingly seeing the data labeling, sensitivity, those types of things, compliance really driving with Purview. So we've got a nice connection there between those 2 drivers.

Got it. We've got about 2 minutes left. Are we okay to get a question from the client? Or should I just go on? Anybody. We generally don't get any questions because it's such a large room and by the time the mics get to the client -- is that the case? Or who's going to be the bold one? Yes. All right.

Okay. So I'm going to make the question that people would like to hear. Your partner, OpenAI, just signed a gigantic multiyear contract with another cloud provider. Why not Azure?

I'm really not in the best position to answer that one, and I'm not joking. I don't run Azure, and I don't have much to say to be honest with you.

Yes. That's...

It's good question.

You're not the right person for that. It's a good question. Yes. You should contact Microsoft IR for that. Anybody else? Yes. This is a question that has to be addressed to Jared given the scope of his work. So yes.

Talk about Microsoft [indiscernible] enterprise [indiscernible] software application [indiscernible] in the broader ecosystem of the enterprise.

Yes, beautiful. Well, we wouldn't say that we can perfectly predict the future, but what we see happening right now is that the patterns associated with knowledge work are all being redone. Like they're being rewired at this moment. We definitely have an ambition just like we did with Office of being the front end of those patterns. We think Office was very unique in that it brought together some things that people didn't think went together. At the time when we brought together the Office Suite, people didn't think everybody needed Excel or everybody needed PowerPoint or everybody needed Word. Those things were kind of apportioned out depending on your role. So as we look at building that front end, we're trying to find the same analogous set of patterns that we can bring to the knowledge worker of the AI era. I'll just start there. That's the ambition. Now what happens to this question of SaaS providers? Are they all dead? Where did they go? We would simply say, gosh, those businesses are valuable. They have valuable data in them. If I choose any one of the SaaS providers, they typically have really valuable logic, business logic inside them as well. That said, people are going to increasingly, we believe, not want to go to those systems directly to do work, but they're going to want agents to be able to do that work for them. So I don't know exactly how it plays out. We don't think that those companies go away automatically, but we think the workflows, the data flows, even the architecture is going to change in very significant ways. And I would say it this way. We're trying to really be in position to ride that wave, and I think every company needs to think about what position they should be in.

With that, let's give a round of applause for Jared. Thank you so much. Great discussion.

Great to be here.