Satya, thank you so much for taking the time and I'm delighted to have you as my first guest on DB Live.
First of all, congratulations, Deedra, on your new show and it's great that you're hosting it here at our Copilot event today.
So, I got to experiment a little bit with the new Copilot and it really feels like this new interface. You could just open that up and do everything on Copilot on your computer. How long have you been working on this? Why release it now?
Yeah. So this is a big milestone for us because as you sort of described it, the idea for us always about how do we really have AI really intersect with what is the new way we all will work and that changing nature of work and workflow, right? And so it started obviously with, you know, essentially using chat as a new way to even search for information quickly, things like new form factors like Copilot started doing things like task delegation. And today marks, I would say, the next big step up, right? Where you now have not only chat and Copilot, but you now have code. Because one of the other artifacts, right? I mean, one of the dreams in fact Microsoft was founded on was what's the real difference between creating a website, an app, or a document? Guess what, there is none now. In fact, I can build a complete application and have it hosted like I would on OneDrive, a document. And then the other thing that we're also really excited about is this autopilot, right? We knew this day would come as the models continue to scale and new capabilities. What's a way to have a long-running business process agent, right? Where it completes the job, but you're still the human in the loop both at the input and at the output. And so bringing all these things together into one powerful, you know, application suite is what, in fact, you could look back and say, you know, that's how Office came about in the PC. Copilot in the AI era is the coming together, in fact, of all of these functions.
So I feel like the real star of the show is autopilot. When I was looking through it, I said just give me autopilot because I as a consumer have already been using sort of personal agent. Talking about Muse from Meta, Instinct, Grokbot. Feels like your timing is so interesting here as consumers are experimenting with these tools. To me, that's what autopilot felt like. How did you get there? And I know that you're rolling it out sort of over time, but because you're seeing the adoption in the consumer space happen so quickly, do you think you need to roll out autopilot faster?
Yeah. Well, to me, first of all, it's a little more a directed graph, if you will, because if you think about it, right, we launched, in fact, Copilot task just around the same time as Open Claw came out and we saw a ton of activity in the consumer space. But, you know, autopilots are fantastic until you have to do something with the output of an autopilot. So this is where I don't think of any one form factor as the be-all end-all, right? This is where, like, I'll give you another example, right? Even with Copilot and Excel, right? Copilot can create fantastic spreadsheets, but at some point I want to reason and that's the... so there's an outer loop that's kind of Copilot, the inner loop is the agent loop inside of Excel. Same thing with autopilot. I think of autopilot as something that has an identity, it does complete jobs, but then when it hands off or when I am trying to instruct it, I need my Copilot and my Copilot. And so this is where I don't think of this as any one thing, but it's the composition. But to your point, having autopilot is about that next generation of, I would say, automation in the enterprise that I think is going to have very high ROI for businesses adopting it.
And as you said, we sort of saw the explosion and real adoption of agents earlier this year with Open Claw, but you don't think that that's going to happen sooner when you see so many, you know, even consumers outsource those tasks to Muse, for example? You don't think that autopilot could just take the job of chat and Copilot?
I think that trust is going to be the maximum biggest issue for us. In fact, it's... you're going to hit that on the consumer side. In fact, if I look back at even our experience in the consumer space, that is it, right? Okay, can I really trust it with all of my credentials when it does autonomous activity? How do I make sure that this is something that I can feel that I'm in control, I can audit? And in the enterprise, this is everything. That's why when you even think about Copilot, we don't think about it in an isolation like any form factor like autopilot. Agent 365 is probably the most important. You know, it's the product that is getting the fastest adoption in the enterprise. That ability to observe everything, govern everything, set policy, have security, have the fin... By the way, the other thing that you'll notice is that this thing, once the subsidies stop for everybody, this is going to... there's marginal cost to tokens. And so one of the core business model realizations is that this is... there's a certain amount of subsidy anyone can do. But at the end of the day, this all has to be about the return that the user gets whether it's in the consumer space or in the enterprise space.
When you talk about subsidies, are you saying that an always-on agent is going to be too expensive to sustain?
It is going to consume tokens and so there needs to be ultimately a business model. And the business models, as far as I can tell, will always be the same, which is somebody pays subscriptions, somebody pays pay-as-you-go, or somebody pays, you know, through some sort of two-sided thing like an ad or what have you. But ultimately, or transactions or what have you, but there are a finite number of business models that have to ultimately account for the fact that software for the first time has marginal cost.
Right. So you have sort of two sides the pricing to this for Copilot and for autopilot. You have usage-based pricing and for chat you have more seat-based pricing. Tell me how you landed on that and how you think that's going to help companies.
Our goal was to make sure that we can have broad penetration so that, you know, users of Office 365 or Microsoft 365 today can benefit from having Copilot broadly deployed. And that's where the seats come in and seats are nothing other than entitlements to usage, right? I mean, seats are a more convenient way to budget and buy versus just having usage-based billing. And so therefore, we want to always say, let's really make sure we're taking advantage of the advancements in AI and pricing as well coming down and adding more and more value. So the way I think about our subscriptions are it's ultimately going to deliver more value every day of the week as models improve. And then on top of that we have usage-based functionality that any user at any time can use with no limits. And so that combination gives us, I think, the right business model to be able to both deliver broad access, deliver the ROI that's expected, and most importantly continuously be aligned with where customers will see value. Right? So this is not about us trying to say, well, it's all about usage or it's all about seats, but being able to be aligned with the ROI the customers see.
Are you preparing for a world in which a lot of your users... I know you said there's different use cases that you want chat, Copilot, autopilot. Are you prepared for a world in which everyone wants to use autopilot? Maybe it's so successful. What does that do to cost?
Yeah, I mean, it's definitely going to have... but again, the same thing, right? Which is there will be... autopilot will start with a certain cost footprint and it'll only reduce after that. And so therefore, when I think about anything new from Microsoft, you will always... my expectation is that you'll have some set of usage rights that come right at the seat and then some that are going to be usage-based, if you will. And so that combination is what gives us confidence that we can keep innovating at the speed at which, you know, these new model capabilities enable new form factors, but at the same time have the promise to customers that, you know, this is coming everywhere.
So Microsoft is in a place where you're able to use lots of different models. Which ones are you using? Where are you seeing higher adoption in chat, for example, versus Copilot versus code?
You know, the last time I picked a model, it's been a long time because in an interesting way, auto has become the product essentially, right? Because the idea that you go in and pick models is becoming pretty... both, by the way, in GitHub Copilot as well as in Copilot. In fact, we are even not just doing model routing, but we are kind of getting to a state where we have learned routers, right? So the real model is my model that routes hard model for hard tasks. Correct. And it's much more dynamic when you have even swarms of agents, even, right? When or you have sub-agents and so on. So in some sense you want to be able to cross over multiple model families. You want to be able to minimize and maximize KV cache hit ratios, which is much easier in coding. We want to crack that even for knowledge work and cyber and what have you. But overall, I feel from a customer adoption perspective, where the behavior is going is not about any given model, but it's about sort of the intensity of the task that then really picks the model. So it's not the user, the model... your agent picks the model.
Is smart enough to know which model is needed. So if I was looking at a dashboard of model usage across Microsoft customers, which are the models being used the most? Are you seeing open-weight take more of that share?
I think the open-weight for us is still, I would say, the MAI models. In fact, interestingly enough, when I look at some of our own models that are really, you know, becoming the ones that hand off even to the more powerful models because we're building this model system where there is the model router plus a model that we have trained together in our harness and our harness is capable of calling all the other models. So that's sort of by inherently getting the most usage. But OpenAI for us, you know, especially with Astra now, is again really seeing fantastic traction. Obviously we have the models from Anthropic. We even have Grok in there in Excel. In fact, the other day I was looking at how that's performing. So it's great, in fact, to have all these models be available and quite frankly competing. So I like the fact that now every model needs to compete for handling tasks inside of the Copilot system. That's a great benefit to customers.
You mentioned the American models. What about the Chinese models hosted here that have become very efficient and very capable?
Yeah, we do have in our Foundry product that is a lot of the open-weight models, but when it comes to Copilot and the pro, most of the are all American closed-source models right now. But we... in fact, the thing that I want us to be able to, which we support, is anybody who has an open-weight model, whether it's Western or others, can put that into circulation. So it's up to customers. So the customers, in fact, that's one of the other great benefits is given the open harness we have, we're not locked into any one model. So literally a customer can deploy Copilot and bring their own model to Copilot and then the harness or the router will decide which model to use and that includes the open-source ones. Are they taking up a bigger share? On the API side, in our case, they are slow on a growth basis, yes, but in absolute terms, we are dominated still by what is happening with OpenAI.
What do you think that looks like a year from now?
I think that there will still be... like, it'll all depend on how rapidly do these things converge in capability. And so I do think that, like all things, right, there's a place for open models, there's a place for closed models, and the Pareto, it'll be on pricing. And so therefore, it'll all depend on how much control do customers want, whether they fine-tune something. So I think of, at least in the enterprise, open-weights are going to be more useful when a customer takes it and does something very specific for a business process.
Now, I was comparing autopilot, the autopilot part of the new Copilot, to what I've been using myself on the consumer side with Muse and Instinct. And I was remarking that your timing is so good because people, those habits are starting to develop. OpenAI and Anthropic have been increasingly focused on the enterprise. Do you worry about them creating something and enterprise wanting to use a product like this from them?
You know, first of all, I think again, let's sort of call it as what it is, right? Open Claw came out and I think inspired all of us to say, wow, there is a way to think about these autopilots. But what we needed to do, for example, in our case, was really make sure that there's trust in their runtime environments, right? So all of the work that's gone in to create this hosting environment with Agent 365 that really has the auditability of the activity of an autopilot. That's what is going to be the key in the enterprise. Of course, there's going to be fierce competition and there's going to be innovation. That's fantastic to see. But, you know, the point you're also making is that this is again maybe a new emergence of a golden era of product making, which is fantastic to see.
Let's talk a little bit about a big meeting this week between President Trump and President Xi Jinping. You're going to be at the dinner on Thursday evening. Is that right?
What do you think the two presidents should or need to agree on in AI?
I think what Secretary Bessent said recently, I think, is a good start, right? Which is in some sense the two superpowers, so to speak, having some set of norms. I think the one that Secretary Bessent talked about is some notifications.
