Excellent. Thank you, all for joining us this morning. My name is Keith Weiss. I run the U.S. Software Equity Research franchise here at Morgan Stanley, and very pleased to have with us from Microsoft, CFO Amy Hood. Amy, thank you for joining us.
Thank you.
Excellent, so exciting times going on within the software landscape, and within Microsoft overall. I thought maybe to start off with, we could talk about the most recent quarter. Very strong Q2, fiscal Q2, in terms of bookings growth. You talked about 75% constant currency bookings growth. You talked about strong $100 million plus Azure AI contracts. Can you help us understand what that means about sort of the overall demand environment? What is Microsoft seeing out there when it comes to commercial demand?
Yeah. I would maybe take a step back and remind people that our bookings number is not just Azure. We'll spend a lot of time on that, I'm sure, in this moment. It's also long-term commitments for, I would say, what we would classify as like M365 or any of the per-user type logic. When you think about booking strength, what you wanna see overall, if you're in my seat, is you wanna see customer contracts being renewed. You wanna see products added. And on the per-user side, and if that happens, it's a good bookings quarter. So I would say it was a good bookings quarter on the per-user side of the house. Then you go to Azure, where we did talk about having certainly better than expected performance on the booking side.
And I think some people, there's a misconception a little bit that these longer-term contracts that we tend to call MACCs, but they're called different things by different people, is just about the larger companies that we sell to. It's a pretty broad concept. And those commitments were pretty consistent in terms of execution, from the smaller-sized companies doing smaller MACCs to large companies doing large MACCs. And then, of course, the relationship with OpenAI. So I would say, in general, that execution also felt very consistent. And I tend to think when it comes to the multi-year type of agreements that make up the majority of bookings, I tend to think about that as more long-term platform health type commitments as opposed to a temporal moment, right? I mean, it's a pretty consistent execution engine for us, and Q2 did feel good.
Got it. And in terms of duration, CRPO grew 21% on a constant currency basis. So it wasn't just duration, it was just kind of overall contract value improving.
Yeah. I think what's important, when you look at those disclosures on short-term and then longer than 12-month disclosure, is what you'll see: is, you know, obviously, per-user health tends to show itself on the shorter end of the spectrum, and MACCs tend to wait things a bit longer. But increasingly, we're also just seeing MACCs that aren't as long duration as they used to be. It's a good mix. It's, it's shorter-term things where customers are confident they're gonna use that amount in that time period, and then there's obviously longer ones. But, I, I don't know. I think this quarter felt pretty consistent in terms of the balance of, of duration.
Got it. Got it. So on the opposite side of the equation, Azure came in line with your guidance, but a little bit disappointing for investors. You're no longer talking to Q4, like you were or kind of earlier in the year. And you talked about a little bit of the go-to-market mix being off in terms of the core versus the Azure side of the equation. A couple things to unpack there. One, on the go-to-market side, can you dig in with us kind of what happened in sort of those incentive programs? What kind of pushed it to the right a little bit?
Yeah. I think we'll cover that in a bit of detail. You know, anytime you see platform shifts or technology shifts, and especially when you have a lot of product change, you know, a lot of excitement and a lot of products you want to make sure have healthy incentives on go-to-market and customer excitement, there's always a challenge on how you balance selling the new with selling the vast majority of your portfolio. Every time we go through one of these transitions, in the moment, it's always better to pivot toward the new. Always. You're trying to teach new sales motions. You're trying to educate your sellers. You're trying to educate customers. You're trying to educate partners. Pivoting incentives is the right thing to do. And the question is always the balance of that.
When you talk about selling through our scale motions, which is entirely through a vast, vast network of partners, you sometimes don't see the impact of that until a couple of quarters because we just, you know, we have this funny fiscal year end, in June. That means we started, those changes in July and August, and ultimately, you don't see impact on close rates and growth for months and months potentially. When you see the impact of that and you try to find the right balance, you just have to make sure you don't over-pivot. I mean, the right answer still is to pivot toward the new. If you're gonna make a mistake, it's to pivot toward the new. If you're gonna have enthusiasm, it should still be to pivot toward the new.
And so we're gonna continue to do that, but also make sure that our partners and even our sellers continue to understand there's two motions that have to land. And they have to land no matter whether you're selling an AI workload or a non-AI workload. They ultimately come together. You're gonna have a workload that has an AI layer and CPUs and storage and compute. It's gonna be one motion, and we have to remember and help make sure partners continue to move customers forward so they're ready. And so, you know, we've made some changes. We've changed some incentives. We did that relatively quickly, and then we'll continue to monitor and tweak as we see how things go.
Got it. Hey, I wanna dig into that last point that you made about when you sell the Gen AI solution, it pulls through additional 'cause I think that's a concern that's arisen in a lot of my investor conversations is this Gen AI instead of selling the core? But it sounds like you guys are seeing both, that both get pulled through when people are building out these workloads.
Yes. I think what we're seeing in terms of the app patterns, when developers, partners, ISVs build these types of solutions, is that it uses the entire stack. I mean, I think Satya's tried to mention that numerous times in his earnings comments and I think in lots of his interviews. But we see it. We see it in terms of how apps are being developed, in terms of how the apps we're developing are developed. They tend to use every layer of the stack, and it's part of the benefit of having a complete stack is that when we sell, I guess the AI layer, it does pull through.
Got it.
We're seeing it.
I think that, like, the question that I get most often from investors after, or have gotten most after, over Q2 is, after your June quarter, you talked to us about increasing supply in the back half of the year, giving you confidence.
Yeah.
An acceleration in Azure. After this quarter, you pulled away from guiding to Q4, so not saying anything about Q4.
Yeah.
What's changed? What, why do you have more confidence in June than you do now in terms of what's gonna happen in Q4?
Yeah. It's a great question. I think it gives me a chance to talk a little bit about connecting the supply with where we saw weakness. AI results were better than we thought in Q2. Every bit of the H2 confidence we had on selling all the incremental capacity is still there. We've said we're still impacted by being short supply. We hope to have that in balance by the end of the fiscal year.
Mm-hmm.
If you take those comments, really nothing changed about AI supply, AI sales, AI revenue growth, AI momentum, and AI revenue expectations through those, through the fiscal year. We feel really good about that.
Okay.
And so when you say what's different, which is the right question, it goes back to the non-AI workload execution. And so, in general, when I look at that and, you know, Q3, we've guided to 31-32, which is an improvement over Q2. And I look and say it's really a focus on making sure that execution on the non-AI ACR improves from what we saw. We clearly didn't expect to see that weakness in Q2. We need to fix it as we just talked about. We've made some good changes. We'll watch, and then we'll talk a little bit more about it, I'm sure, in April.
Got it. I mean, it would seem like there's a possibility for, like, part one to help fix part two in that if you're seeing good attach of non-AI onto sort of AI workloads as the capacity opens up, as you're able to provision more of those AI workloads, it should pull through more of the core?
You have to remember. I mean, it's a big business, Azure. And a lot of it is not AI related. It's related to the constant migrations of existing workloads and the opportunity that still exists to get that done.
Right.
And so when we talk about the Q2 execution challenges, it's really about those motions, the motions that drove the past 15 years of Azure growth. It's the fundamental shift from sort of on-prem to cloud. And I know that's maybe not as exciting to talk about, but it is the fundamental pattern that we continue to see and there's tons of opportunity in. And so continuing to execute on that is super important, super important to customers to make sure they continue to get value, get resiliency, get security, can run better app loads. I mean, it's just, it has to get done. And so being able to do that and sell the AI workloads is certainly the opportunity ahead. It's not an or. It's not a choice between budgets. It's that both have to happen and customers need both to happen.
Got it. I wanna dig into the capacity constraints and sort of the buildout and capacity. You guys have been ramping up CapEx for a couple of years now and spending really big dollar amounts and big growth in those dollar amounts. Can you help us understand sort of the process of solving a capacity constraint like that? Like, it's not just buying a bunch of GPUs. I mean, you need the data centers. And, what's the process in getting that fixed?
It's definitely not just buying GPUs for sure, 'cause that would be a problem that you could solve pretty quickly. Capacity constraints fundamentally start with an inflection point in demand. So the first question isn't about capacity. It's about demand planning. So two years ago-ish, I mean, maybe it's been longer than that. Time sort of flies, in this era. We saw the work, that was coming out of our partnership with OpenAI and said this is gonna be an inflection point way before I think the market was aware and could see the applications and could see what we thought was so exciting about the work. When that happens, you say, "Well, we're gonna see a data inflection point. We're gonna see demand change.
And how quickly can we get capacity online?" The answer is short-term, you can get a decent amount of capacity online 'cause you take every bit of capacity that you'd used for Commercial Cloud and you start jamming GPUs, CPUs, storage into every corner of every data center we had built and powered up. It does two things when you do that. We're really excited 'cause we have people be able to use the technology. The challenge in that is that every bit of room you had built on the Commercial Cloud side and the 'cause right, you build a demand curve, you have standard deviations of outcomes, you make sure you have capacity to be able to grow within normal surge deviations. If you fill that with a new, entirely new workload, you suddenly have a challenge that you're running incredibly tight.
That happened really very, very quickly. So even if you start the concept of building data centers from that point, it's about land. It's about construction. It's been about power. That process isn't fast. And so as I think people know, listen, we did what we could in terms of leases to be able to deliver revenue growth. Even leases take a long time to come online. They're going through the same process as we are. What you've seen in that spend, over the past two years, and we've talked about this, it is more long-term assets. So think about that as everything I just talked about, right? Land, construction, buildings, 15-year plus type assets. And when you start about two-ish plus years ago, things start coming online in volume going forward, just timeline-wise.
And so what we're really doing is replacing not just what I think people consider is like, "Oh, it's all AI basis." It's not. It's Commercial Cloud. It's a global footprint. It's the AI footprint. And it's building up the room we need, and should have always had to make sure that we, you know, can adjust to much smaller changes in demand if we need to. And I think what we'll see is as we get to the end of this fiscal year, which I've talked about, I'll feel good that we've got enough of the long assets coming online to be able to better match. And I feel like we'll be in a good balance place. Then going forward, when that happens, you shift from these longer-term assets, which we'll still need because we've got $300 billion of RPO to deliver to customers.
So we have to keep building and have to keep adding and need capacity and it needs to grow to deliver the revenue we've already sold, much less the revenue we'll continue to sell. And so you'll see it pivot, right, to be a little bit more weighted towards servers. I say servers broadly, CPUs, GPUs, other short-lived assets. And that's more correlated to revenue growth because we're no longer having to build a global footprint. The other thing I would say that was somewhat unique and quite different from the Commercial Cloud transition that I don't know that everybody was in this room when we started that one. I guess I certainly am the one still here is, that transition rolled out almost geo- by- geo. And you heard us, like, announce Azure regions, like, "Hey, we have a new region." And we still do that.
I'm still excited when we add new regions on a global basis. When we add the AI workload transition, it will land, you know, globally, not in a all at once. That's super important if you're gonna build workloads for our customers around the world. I think, you know, it's a long answer, but I think people need to understand the context a little better to why things end up looking the way they look, when you need to catch up. I'm still really glad that we used every bit of space we had. Just to be clear, like, I'm super glad we used every bit of space we had around the world to deliver, and to be a leader. That's still the absolute right choice. And now we need to make sure we've got the room to grow to stay that way.
Got it. So if I take that answer and think about it in context of the guidance that you gave into the back half of the year of relatively flat CapEx spending in line with what we saw in Q2, there's it sounds like there's a level deeper within that, that because the 15-year depreciation asset has been built out, we're gonna see a shift in that 50/50 away from the long-term depreciation asset. And you're gonna start filling up those data centers with the server kits, with the six-year depreciation asset.
Yeah. That's a good way. Yes. And I think you have to remember it's gonna be, it was even more than 50%, just so people are clearer, when we talked about Q2. So what you'll see is over time that'll start to shift. And, and the shift will be dependent. We, you know, it's gonna be bumpier in some ways in terms of percentages. Leases come online. It shifts it up in terms of long life. But in, in general, over a multi-period time period, you will see it shift toward kits, just because, once you get more in balance, that's logically how CapEx would land.
Got it. I wanna talk to some of the tea leaves, excuse me, that investors look at to try to kinda understand what's going on within Microsoft. I think one of the big ones was the announcement of Stargate and the changing in the nature of the relationship between OpenAI and Microsoft, where you guys moved from an exclusive relationship to one of where you have, like, right of first refusal. Why did that make sense for Microsoft? Like, why does Microsoft wanna be in maybe not front and center when it comes to Stargate, but sort of a little bit more of a backseat participant, if you will?
Yeah. I think it's important to understand, you know, listen, the partnership with OpenAI is important. It has been incredibly beneficial to both of us, and it remains so. And I think a lot of ways we hadn't really disclosed the nature of that partnership. And I think we put out a statement sort of sharing more details so that when Stargate came out, so that people could understand what wasn't changing about the relationship, which is probably more important than what did.
Mm-hmm.
The nature of our IP relationship, the nature of our go-to-market relationship, the nature of, we're both successful when each, each of us are successful, the nature of our supplying them. And if you think about a right of first refusal, you know, we've already had them use other vendors when we couldn't supply all the demand needed based on our conversation we just had because the goal is to make sure they grow. Their success is paramount. And so if that requires them going and buying additional compute that we can't supply, that's good for both of us. And I, I just think it's important to realize it's not like a either/or. It's, it's an and. I would also say that when you think about, and we talked about the relationship being through 2030 when we started the relationship, I think in 2018.
And so as you go through that process, I do think everybody's planning for what happens for a decade or two decades. And that's important for both of us to do. And what's great is that we're building a really flexible fleet, that can be used for any type of workload, on a global basis. And we look forward to continuing to be their primary partner, and be able to supply them, through that agreement and structure through 2030. It's a good thing for both of us.
Got it. One of the things that both you and Satya have talked about a lot on the conference calls is the nature of the AI workloads that we're seeing on Azure tend to veer towards inference. The vast majority are inference and training workloads, and I think Satya's even said he's turned away some training workloads. Can we look at that Stargate announcement through that type of lens, that you guys are more interested in the inference side of the equation versus the pre-training side?
I think it's maybe to take a different approach. What we're seeing in terms of, I think in the question, it conflates two very different concepts, which is today when we talk about our $13 billion AI revenue number, it's primarily inference and post-training workloads done in the fleet, right? Plus our Copilot revenue. And that's because of the nature of the relationship and how we sell to OpenAI plus every other customer we have. And I think sometimes people take that to mean, wait, where's training in that? And the answer is training revenue that we work with OpenAI does not and is not in that number, right? So that's. We've tried to be very clear on that.
Now let's talk about a separate thing, which is what are we trying to build for the next two decades and the opportunity we see, which is to build the world's leading AI platform. We happen to call that thing Azure. It has not just the A, the AI platform, but the stack underneath it. It is global. It is distributed. It will be able to serve every type of workload because as we're seeing, especially from some of the post-training work, that being able to have that flexible workload is super important. It'll make sure utility, utilization remains high. It'll make sure it's sellable. It'll make sure it has a long life. And those are really important attributes of a global fleet.
And so, our focus on that, and by the way, it's primarily what is powering our workloads and everybody else's OpenAI's, et cetera, et cetera, is that that is a durable asset addressing a giant TAM. And our focus on building that is because the return on invested capital on that investment is well understood by us. Looks much like what we've seen in the Commercial Cloud. It's run as a single fleet. We know how to optimize it. You'll see software improvement. You'll see hardware improvement. You'll see model improvement, and you'll see efficiency. And building a fleet to do that improves the returns. And so I think that's where our focus is, because you don't wanna solve for one or two years. You wanna solve for something, I think, that's durable in terms of execution.
Got it. Got it. Makes a ton of sense. So if you think about the comment that you made that exiting FY 25, you expect to see supply and demand relatively in balance. It speaks to a comfort that you have with kind of your capacity planning on a go-forward basis. I think investors worry a lot about availability of power, right? Ability to sort of build out data center capacity. Do you see constraints in that way, or do you guys have, like, a good roadmap in terms of what gets you comfortable with having the necessary capacity on a go-forward basis?
I laugh. We started by you asking, am I spending too much? And I think maybe you're asking if I'm spending too little. But let me try to take a slightly different tactic to the answer. You know, over the short term, you have to remember these, we talked about lead times being quite long, right? And so we have demand plans that cover anywhere from zero to 10 years given lead times. And really what we're focused on is making sure we have the right capacity across geos built to that demand plan. And, you know, constraints move in those, in those periods. Short term, we feel really good about our availability to get full data centers with power, with chips ready and functioning on timelines that make a ton of sense to us.
Over the long term, you may say, well, do we have enough labor, not for Microsoft, but as an industry to be able to build the capacity that could be needed in a decade? And that's something we've talked about quite publicly is making sure we as a, as an industry and as a country have enough ability to do that and make sure that over the long term we have the skills necessary and the power necessary, to make sure we meet the opportunity.
Got it. I wanna ask about DeepSeek. And it is definitely a it was an announcement and a sort of a level of innovation that there was a wave that went through the investor community when when thinking about what that implies for generative AI on a go-forward basis. Was it as surprising like within Microsoft or was this more aligned to kind of the cost curves that you guys were already thinking about when it comes to these large language models the level of performance improvements? Or said another way did it change your either your capacity planning or how you're thinking about the monetization of these models on a go-forward basis?
No, I think, you know, if I go back, I think Satya had actually done a podcast before that talking about the role distillation would play in model costing. So I don't know that I would say per se, let me separate the concept of DeepSeek from a broader concept that we believe that models would get more efficient, costs would come down, software improvements would change. Like, I think we understood that. Now what's important to realize, I think, is that as a platform company, what matters the most is making sure you have options for developers to be able to use the model they want to build the thing they want. I think we have, I don't know, over 1,800 models today that run in and on and through our Azure marketplaces.
If you think about that, I don't know that anybody could name 1,800 models they think exist because people kinda remember a few. But the important part is that each of those, including open source, may be the best model for a use case that you have. And so what's important in that is that you also wanna make sure that people can get an improving return on their investments in those, applications they're building. Lower cost and higher output is a good thing for demand. So if you separate it from DeepSeek and say, in general, is having a proliferation of models built for single purpose that bring down costs and have high impact a good thing for demand, the answer is yes. And it's especially good if you think about saying, well, we have and we feel great about having the leading models from OpenAI.
We're still incredibly proud of that, and it's important. But we also have other models, including ones we build, to make sure that there's choice. So if cost comes down, value goes up, demand improves. And so for us, I think we feel good about that.
Got it. So I think Satya talked about that in terms of Jevons Paradox, my favorite paradox.
Of many paradoxes.
Of all the paradoxes, I think.
I'd like you to name your second favorite paradox.
We'll do that at a different session.
That's what I suspected.
just,
Come on, Keith. Sometimes it's too easy.
Yeah. So in that dynamic is what gives Satya that confidence is that we've seen this dynamic roll through before with sort of the commoditization of underlying compute, underlying storage. Like when you're bringing down those input costs, you've seen the result of that. It drives further workload growth.
Yeah. I think, one of the things that I think builds confidence is watching this model be the case through the Commercial Cloud transition too. I mean, costs came down to build new workloads. And so I think if you try to get away from some of the technical arguments on why a model works differently or if the output is you get more for less, it usually is good for demand, especially consumption demand.
Mm-hmm.
Because once you deploy workloads, they generally keep running, keep consuming. You add more workloads. You add more workloads. You add more workloads. That's the transition that we saw before, and what we saw was that, and with confidence and with new, new capabilities, people found new workloads that we didn't know would even exist in some ways, not in the AI wave, in the, I don't know, the old cloud wave. I'm not sure how we're talking about the thing that still is a massive opportunity. This pattern tends to repeat itself, and especially if you're a hyperscaler, making sure you can be the most efficient deliverer of that is important to success.
Got it. I wanna switch gears a little bit and talk about Microsoft 365 Copilot, the other kind of big AI initiative from Microsoft that investors focus a ton on. I would say from an investor perspective, we've kind of come into the trough of a disillusionment, right? In terms of we were expecting a lot out of a new product really quickly. And as often seen in investors, we proved to be impatient, kind of in seeing that come through. From the Microsoft perspective, how is the Microsoft 365 Copilot rollout going versus your expectations?
As we talked about, even in Q2, it was better than we thought it would be. And so I really feel like this is one where, you know, deployment and adoption does take some time. But what we're seeing is, which I think is the most interesting, is in customers who initially purchased, they're buying more and using more. That pattern is a good pat. I mean, if you just simply say for a per user business, expanding seats and having them used more is an incredibly good thing for value. And in many ways, it's running it, it's faster than the other products we've released into the enterprise before in our per user businesses. I feel really good about that execution. I'm excited by Copilot Chat that we released in January. I think days sort of rolled together here a little bit, but I think that was January.
It is incredibly important to have 400 million plus commercial M365 users be able to learn the productivity that's possible just through Copilot Chat and watching that habit form, and usage increase. It'll be a really healthy funnel for us to continue to sell the full value SKU. I do think that the early reception to that has been really exciting for us, and so, I'm excited to watch, and continue to monitor usage growth there in particular.
Got it. Got it. And can you help us think about the product evolution? There's been a debate in the marketplace about Copilot versus agents. And when I hear Microsoft talk about it, to me, it feels more like a continuum that the Copilots are going to gain more agency and act more like agents on a go-forward basis. How are we gonna see that within the product portfolio?
Yeah. I think, maybe two things I would give to frame 'cause I think you're asking a question that will have some duration to it, right? Which is important. Copilot Studio, which you hear us talk a lot about, and I'm not sure people engage in it as much as I would think is warranted. It's the way customers will build agentic AI for themselves and using their data and maybe having very specific use cases and our ability to have Copilot Studio be the interface which they can do that in a low code environment is super important.
and having the UI be the Copilot that they're used to, whether that's from Copilot Chat or the full value Copilot that we sell, having that continuity is really important because then if you have multiple agents, right, the interface is familiar, the understanding as a worker and as a productivity tool is, is clear. And I feel like that strategy continuum where you have the Copilot is almost the UI, Copilot Studios, how developers can build and companies can build out their own agents, having that be familiar, having it run, on Azure as the backend is a really thoughtful, way for us to make sure that we can cover the continuum. and I, and I think that's worth, you know, investors focusing on that logic, if I was to say how to think about where we view agents and, and where they're going.
Got it.
The capabilities that already exist today.
Got it. I'm gonna just sneak in just one last question on, on margins. What we've seen throughout FY25 thus far is, as more Gen AI comes online, we're seeing pressures on, on gross margins and I guess more cloud come online. But you've been able to offset a lot of that with OpEx efficiencies. Is that a paradigm we could see on a go-forward basis? And, and to what extent is Microsoft utilizing the technologies themselves, utilizing Gen AI, enabling you to drive more of that OpEx efficiency?
Yep. You get a lot of questions in one question, Keith.
Sure.
I wanna compliment you on that. That, that final question has about six parts. So, I'm gonna start by talking about margins. You're right that this AI wave has put pressure on Azure gross margins as we've talked about. What I would say is incredible work by the engineering and platform teams and the architecture we committed to is that our margins very early in the AI process are monumentally better than they were the first time we went through this transition on Azure. And I'm really, it's thoughtful, it's a fungible architecture, it, it can be used by any workload, utilization's gonna be high. It, it starts from a much better place. And so even through this surge, we'll start to see and continue to make sure that we get efficiencies on the AI platform part of Azure as we, as we go forward, which is good.
Mm-hmm.
Then you're right. We have made a concerted effort to also make sure we look at our operating expenses and ask ourselves a couple things. Are we putting them toward the highest growth, highest propensity growth areas for us? And the answer is you can always do better. And the environment changes every little bit. So we're continuing to move. Secondly, tools like GitHub Copilot and tools like Microsoft 365 Copilot and tools that we're deploying across customer service, customer support, our sales teams in finance all have provided incremental opportunities for savings. And so as we continue to deploy our own AI workloads, to our departments as well as see and focus on moving and more moving our resources, I feel good as I've said. I feel better about margins in FY25 than I felt when we started in June.
And we've committed to that and team's done really good work. And I continue to believe there's room for us to keep pushing on that front, whether it's AI productivity or good old-fashioned portfolio work. So, you know, we'll stay on it.
Outstanding. Amy, thank you so much for joining us. As always, a fascinating conversation.
Thank you, Keith.