Good morning,

This week’s Stratechery Interview is with Microsoft CEO Satya Nadella. I have previously interviewed Nadella in May 2024, October 2022, April 2020, and May 2019.

As I noted yesterday, I spoke to Nadella shortly after the conclusion of his keynote at Build, Microsoft’s annual developer conference. One notable thing about the keynote was the fact that Nadella was — outside of product demos — the sole presenter; one gets the sense he has shifted into a much more hands-on role at Microsoft over the last year.

The reasons why are clear: my first question to Nadella was if he was happy about where Microsoft was currently positioned as a company. We talk about the reasons for that question, the status of the company’s partnership with OpenAI, and whether Microsoft has invested sufficiently in AI infrastructure. Then we talk about the future of software, Microsoft’s business model in the age of AI, and if they can operate independently from the leading edge models. At the end we talk about Project Solara and whether Microsoft will ever pay residents to build data centers.

One note, with regards to a misunderstanding towards the end of the interview: there is no documentation I could find about being able to use Copilot Cowork with non-Anthropic models; Microsoft’s own documentation fits my understanding.

As a reminder, all Stratechery content, including interviews, is available as a podcast; click the link at the top of this email to add Stratechery to your podcast player.

On to the Interview:

An Interview with Microsoft CEO Satya Nadella About Finding Core Competencies

This interview is lightly edited for clarity.

Evaluating Microsoft’s Competitive Position

Satya Nadella, welcome back to Stratechery.

SN: It’s great to be with you, Ben.

So first off, I don’t know if you realize this, but at least according to my daughter, the defining word for the real grinders in Gen Z — first off, LinkedIn is like the social network.

SN: That’s great!

Number two, the word they all use is “build”, “I’m building, I’m building”, so who knew when I was at the first Build, I think, in 2010? Or was it 2011? Who knew you were such a trendsetter?

SN: (laughing) There you go, I’m thrilled that your daughter is building and is on LinkedIn.

Yeah, well, I’m not sure if she’s on there, she’s more making fun of people, so we’ll see how it works.

We last talked the summer of 2024 after Build, this was up in Seattle. To say a lot has changed since then is an understatement. I had a bunch of questions I wanted to ask you about the business as a whole, things going on, I’m going to start with those, then I have questions about the presentation at the end. But relative to that, I want to ask you one simple question: Are you happy with Microsoft’s current competitive position?

SN: You know, always this is the trickiest thing, you can sit here and say, “I’m happy” — that means you’re not ambitious enough and when you say, “If you’re not competitive, what the heck are you doing?”.

And plus you have like 57 different product lines.

SN: I’d say the thing in these platform shifts in particular is to, one, get the conceptual model of, “Where is the opportunity for us as a company?” — most people measure competitive position as if it’s a complete zero-sum game, and it’s never been the case. Which is, it is not the case with the cloud, it is not the case in client-server, and so to me, “What is Microsoft uniquely capable of doing in this new world” — that’s the key thing that we have to answer before we even get to the competitive position.

In that context, “What is it that we really have a shot at?”, which is we can be a trusted purveyor of a platform, which is what we’ve always done, that allows people to create more value on top of a platform, which is again the DNA we have. Even in a world where these frontier models seem to have no limit—

A very large appetite.

SN: They have large appetite. That is what I feel even this Build, this conference, we are at that state where we can now really turn this from any one frontier model to saying, “Hey, there is actually a way for a frontier ecosystem to emerge where there are many stakeholders who all actually are operating with their own frontier intelligence”, that is a place where I think we have a unique shot, a unique competitive angle, and most importantly, brand permission.

This is the other thing I’ve learned, Ben, which is every company thinks they can do everything, and then they realize that the world doesn’t need them to, the world wants them to do the one thing.

Is that a lesson that you had to learn?

SN: Yeah, absolutely. I’ve always said this, at Microsoft we are at our best when we do what the world expects us to do, we are at our worst when we do things out of envy, which is just because somebody else had some cool hit, somewhere, doesn’t mean we should go do that.

But enough about the Zune, right?

SN: (laughing) Yeah, Zune was a great device, but the world didn’t need Zune from us, and so that was the end of it.

This identification of your unique capabilities, is that one of the changes over the last two years where that has emerged?

SN: Yeah, in fact, it has emerged and also the world’s kind of gotten to it.

Has it been forced on you to an extent?

SN: Yeah, even my own conceptual understanding, I started by thinking of, “What are models?”, models are kind of like some stateless APIs, then I adjusted and said, “Oh, maybe there’ll be like databases” — they’re really more than that.

I don’t remember talking about this with you, but last time I talked to [Microsoft CTO] Kevin [Scott], we analogized it to processors at some point, and you actually did make a comparison in terms of the partnership to your partnership with Intel.

SN: Exactly. So the question now is, it’s a better conceptual model to think of what we’re doing is you have to really build a learning machine, and any company has to build a learning machine, so what I want to build is essentially a multi-tenant learning system that allows everybody to have their own hill-climbing machine.

So that conceptual idea, now I’ve turned what is essentially frontier is not about any frontier model — I want to build whatever you did with M365 or with Azure into a platform which allows everybody to basically build their own hill-climbing machine right because the future of a firm at a foundational level they’ll have human capital they’ll have token capital and for the token capital they need their own hill-climbing machine.

MAI Models

All right, so I’ll jump to the end, you released seven new models, you emphasize the work you’ve done to build these models from scratch, not with distilling, not with using other models as teachers — so did you just articulate what the ambitions are with these models?

SN: Yeah, there are two sets of things. One is we wanted to build from ground up with clean lineage, the models that we will have that we can license and allow enterprises to continuously hill-climb, so that’s why we want that model. By the way you talked about distillation — the point is to not use distillation during any of our own hill-climbing but at the very end, in fact some of the things that we are doing is, after all, we have all the OpenAI IP, in fact some of the performance gains we get is by doing RKLD, which is reverse knowledge distillation, and RL on top of it. So we have effectively two frontiers, we have our own, we have the OpenAI, and we’re going to use these things to eval match.

And the clock is ticking to get to the right state you need to be while you still have that access.

SN: Yeah, and there’s five years of it. But the bottom line is at any given point in time, I want to make sure that I’m using the best, most efficient model for whether it’s in coding, whether it’s in security, making sure also in our case, we’ll have a harness that’s independent of these models, we have the GitHub Copilot harness that’s used everywhere across Microsoft. Our goal is to make sure we have a model lineage, which we control end-to-end, we then use OpenAI IP, even with all of the capability it has — ultimately, the tests are going to be the evals for us and our customers.

In the long run, the way it was framed today, and I thought it was very compelling, and it speaks to what you just said, was this idea of enterprises being able to take these models and in their own RL environments incorporate their data at a much deeper level than sort of a slap-on RAG implementation or basic post-training. Is that the end goal, though?

SN: Yeah, the end goal for me is the following, which is I go back and say, let’s say that they’re a generalist model — if you go back even, Windows could have a release, then another release, and Adobe and Autodesk could keep building and keep going up, what’s the moral equivalent of that? That is the thing. And then in the first time, we said fine-tuning, it kind of didn’t work because we didn’t have the tools, we didn’t have the data collection regime, none of that. But now we have it. So let’s say the generalist models keep getting better, MAI models, let’s say, or OpenAI models, then you have this RLE.

Right, but this deep customization of the models you’re talking about is only possible with MAI models.

SN: That’s correct, but the thing that we want to start getting everyone on is this multi-tenant hill-climbing system — so if you think about it, we literally turned your use of M365, which already is a multi-tenant system, into a hill-climbing system for you.

Okay, I’m gonna have to stop you, I’m going to give you an ELI5 opportunity, explain hill-climbing to the audience.

SN: Hill-climbing is basically when you think about, “What does AI do?” — AI is all about taking an objective and continuously learning how to go predict and create that output that is the representation of that objective, and do so continuously. So that’s why a metaphor of hill-climbing is the best way to describe learning.

And you want everybody to do this individually on their own hill.

SN: Individually on their own.

As opposed to like, hitching along.

SN: What is your moat as a company? Your moat as a company is your tacit knowledge. In a world where AI exists, and network effects of AI exist, you need your own hill-climbing machine in which the models are learning.

So the first thing we want you to do is, people don’t talk enough about this, but the private outputs, the evals, as I think about as, maybe the most important IP a firm creates are these private benchmarks and the private evals where you are tastefully recognizing what’s the output, the quality. And by the way, today’s failure cases are informing you to change the benchmark continuously, it’s not a static thing, that’s kind of how the evals work. And so if you have your private evals, then you have your own reinforcement learning environment that you’ve created, then you invite all the models to show up, and then you say, “Model A, generate the output that is maxing this eval using my environment and my trajectories and model B…”, and I can switch.

In that context, the MAI models is one more lineage that you can put into,c and what we proved today was even a very efficiently trained reasoning model or a coding model can hill-climb using your traces and that will be more token-efficient and it will be fundamentally a great advantage.

Exclusive to you the customer.

SN: Yeah, that’s right.

But is that just for now? If you fast-forward, is your vision that actually MAI models are fully competitive on the frontier with the other general models?

SN: They are. Even today, when you start saying that — the world will keep getting better in general.**

Well, I guess this goes back to, is this about how you need to do what you’re good at?

SN: Correct. One, what we’re good at and also what’s the equilibrium of the world? Which is, if you believe there are only going to be two firms in the world, then of course, they only need two frontier models, but if you fundamentally believe that there are going to be as many firms as there are today and more, then what is the firm in the age of AI? It’s going to have human capital and token capital, how did that token capital get created? It’s not a bunch of API calls, it’s actually some set of weights even they have.

Right. And so do you want to accrue that advantage or do you want to give it to OpenAI and Anthropic?

OpenAI and Capex

Well, speaking of the OpenAI partnership, I mentioned you referred to it like the Microsoft-Intel partnership, and sometimes partnerships are the only way to get ahead. How do you think about that partnership now?

SN: I still think that it’s — I’m very proud of the fact that we came together, you remember the circumstances in which we came together were very different and the fact that there is a company now that may go public and be a trillion-dollar company—

This is my question — how long were the knockdown, drag out fights between in this corner, there’s Satya Nadella, the operator, and in this corner, there’s Satya Nadella, the investor, tussling over what to do?

SN: (laughing) At the end of the day, we are an operating company, investment is just more of an accident.

Yeah, but the shareholders are ultimately those investors!

SN: I’m glad and it’s a fantastic outcome for our shareholders too and what have you. But I think the way I came at this, Ben, is to say genuinely I’ve always approached it as, if there’s a partner that we can partner with and ourselves innovate, and they’re also successful, that’s fantastic.

I always go back to the story of having built SQL Server with SAP. SAP was successful, we were successful, we also then went on to do other things. And so therefore, I think OpenAI, I’m glad we worked with them, we’re working with them, they continue to be a premier partner. As I said, until 2032, we still have a lot as a customer of theirs, them as a customer of ours, as an IP partner. So every day OpenAI does well, Microsoft does well.

Is there a bit where everyone thought you were so far ahead because of your partnership with OpenAI, and now when we talk about things like your MAI models, it’s like actually “We got a little bit lulled to sleep because we offloaded too much to them, and now we’re having to recalibrate”?

SN: Lots of things, one is, like all things, there’s a lot more competition, there is OpenAI, there is Anthropic, there’s Google, there is tons of folks who are in there. And so I think for us, the beginning, it was great that we got started with OpenAI. Think about where we were in 2018 to where we are in 2026, here we are competing with Google and a bunch of people whose names I wouldn’t have known in 2018, and so that itself proves that to your very first question, “How competitive is Microsoft?” — I’m glad Microsoft took that shot. Here we are competing with a bunch of new people, a bunch of old people, and we have our own game.

So we already talked about Satya Nadella, the operator, and Satya Nadella, the investor. What about Satya Nadella, the capital allocator? There were a lot of reports in about early 2025 about Microsoft pausing and a reconsidering some data center investments, you guys have sort of spun that as, “Lots of speculative stuff”, “We’re streamlining”, etc. — but at the same time, your percentage of free cash flow committed to CapEx lags fairly significantly behind your peers. Four months ago, that was a compliment. Now, is it a diss? How are you feeling about that?

SN: The last time I checked, my free cash flow is getting allocated pretty well to capital return that makes sense.

Is there a case that you’ve underinvested?

SN: Not really. I think the key thing that at least we wanted to make sure is we were not upside down on building — we have a hyperscale busines, we have our own application business, and we have our own research compute to allocate, there are three buckets, we wanted to allocate with great discipline on all three.

So take the hyperscale business. Hyperscale businesses are about having a few big customers, but also having a massive long tail, so you can’t have a book of business that is just a few model companies — in fact, one model company — that was the fundamental decision.

And you wanted to get out of that business.

SN: Not just get out.

They’re still there, they’re a major tenant.

SN: They’re a major tenant. But, let’s face it, Anthropic over time or OpenAI over time will build their own, it makes sense. They would use — I’m not saying that they won’t use other cloud providers. So to me, it was clear as day that, what I wanted to do was not allocate all my compute only to one player and so that was the adjustment. And once you make that adjustment, you can’t build 10 gigawatts in Texas and say, “That’s it”, you’ve got to build a plant that is spread around the world, around the United States, and that adjustment is what we want to do on hyperscale.

The other thing that I have to do is make sure we’re doing also the long-term thing for our investors, which is, “Let’s invest in ourselves”, which is inference compute has exploded, whether it’s in GitHub or whether it’s in M365 and we needed to make sure we fund our own applications. And then our own research compute, these MAI models. So I just took the approach of putting these three, we will definitely want to allocate as we see progress on all this and we’ll see how it all shakes out. But to me, I’m not literally matching quarter-to-quarter.

By the way, the other interesting thing is the catch-up, we started early.

You were early, and you got a lot of the good spots, a lot of the good power generation.

SN: Yeah, and also two years of cash flow.

Yeah, for sure. Well, speaking of the balance between the three, in January 2026, you missed Azure earnings by like 0.1%, so it was very small, and you said on the call, you allocated more compute to internal R&D and applications.

Setting aside the earlier question about whether or not you erred by the total amount of capacity, you talked in that call about having a portfolio approach in terms of investment, balancing Azure, and those two other businesses. That’s all well and good, but if there is a constraint, you do have to choose, do you think you made the right choice then? And is that the choice you’ll make going forward? Where you are at the end of the day, you have a higher lifetime value, higher margin on your own businesses, and that’s going to be number one.

SN: Yeah, and also research compute. Ben, I think that for all of us, quite frankly, we have to really, at the end of the day, that’s why I think quarterly earnings are interesting, which is, of course, The Street should hold every one of us very accountable for “What did you do for me lately?”.

But was that a very particular, annoying, being held accountable for the wrong thing?

SN: It’s their job, everyone’s got to do their job, and so I can’t accuse them of them asking, “Hey, what did you do for me this quarter?”, that’s the question they rightfully should ask. And the right answer for me is, “I’ve done enough for you this quarter, and we’re also making sure that 10 quarters from now, Microsoft’s continuing to thrive”, and that’s the job, and sometimes there’s a little bit of disconnect on it.

But when I look at the three things, you just have to be disciplined that you’re doing what you can add value, it can’t be, “Oh, I’m misallocated”. To your point, you get punished if you do things where you’re not producing. So that’s why research compute, here is now an MAI model output. Today, it’s just not a model output as an academic thing, that’s now in differentiating our Foundry where we now are able to license it, it’s going to grow Foundry revenue. And so as long as I’ve felt that as long as Microsoft can continue to invest in ways that show results, then we will have the ability to do the right thing in the long run and in the short run deliver results.

For the last quarter, was there a bit of, “Let’s give a little bit more compute to Azure?”

SN: Last quarter, no. In fact, that one was just a little more of the compute — we are supply-constrained.

I know, but that’s what makes it so interesting.

SN: We are not at all, like at this point, if anything, the thing that we do not want to do is to disappoint especially our enterprise customers on Azure.

That was the question, right? Because if they look at that quarter and they’re like, “Hmm, Microsoft’s saying we’re supply-constrained and also we’re prioritizing our higher margin, higher lifetime value businesses, where does that leave me? I’m competing against my supplier”.

SN: That’s one of the reasons why we had to make some very hard choices around, for example, raw GPUs. We’re not selling raw GPUs to a bunch of Neolabs, for example. I wish I could add more Neolabs on Azure, we just cannot. And so therefore, we are being very disciplined on some business that we turn away.

Were those some of the conversations you had to have?

SN: Yeah, and so to me, in a world where you have constraints, you want to basically make sure you’re building for both what the world expects and the customers who have trusted you in the longest and so we will definitely make sure that Azure has capacity, it’s just that we are not going to go for what I’ll call in this context, “easy money”. Which is, you can always, in today’s day and age, if you want to have short term Azure revenue, it’s pretty easy.

Oh yeah, we’ve seen that, to say the least.

SN: Yeah, all you gotta do is turn up, you know, and go sell it to a Neolab.

So when it comes to AI infrastructure specifically, as you look out in the long run, you mentioned it may very well be rational for the frontier labs to build their own hardware, for example. You have all these Neolabs, you have whatever controls [Nvidia CEO] Jensen [Haung]’s allocation of GPUs, you have different ASICs, what is your true differentiation as a hyperscaler? Is it just lower cost of capital?

SN: First of all, think of our hyperscale business as this portfolio, everything from what we are trying to get done is build a system which we have to be competitive in when it comes to tokens-per-dollar-per-watt, that’s one side of it. We can unpack that and what our thesis is there.

Well, I just noticed when you were talking about some of your chips, sometimes it was tokens-per-watt, sometimes it was tokens-per-dollar.

SN: Yeah, I think of all three, right? It’s like tokens as a function of both power and dollars and so that’s a systems thing that we have to be world class at and be competitive at. And I would be able to claim, and that’s where I think [Microsoft AI CEO] Mustafa [Suleyman] talked about it, like unless and until you build your own model, you can’t, there’s no point. I believe that you don’t want to build accelerators without building a model, you kind of have to co-design. In the long run, the only way to be super efficient on that is to think about, the network is a great example, which is you want the network, the model, all to come together in ways that make sense, so therefore that’s one side.

Then the other side for us is the differentiation has to come from, “If I’m building agents on top of this infrastructure, what agents does Microsoft produce?”. I have three domains in which we are going to try and major on: coding, security, and knowledge work. Luckily these are three massive domains where tokens make sense — I’m not saying there won’t be others, science is another one we will enable but I think there will be others who will do great work in there. But to me the three primary domains in which all this is going to be exercised use. So when I think about the portfolio of building a system plus model plus these three domains, then I feel like that’s where our differentiation will come from.

But is that just a re-articulation of circling back to, in the long run, our true differentiation is from our higher margin, our own businesses, higher LTV? Where does that leave just customers who—

SN: I think it’s not higher margin. The overall margin dollars from our infrastructure business may be higher. In fact, they already are getting close to being higher than our total margin dollars from our high margin businesses.

So I think that Microsoft has always benefited from having a portfolio of businesses, and we’ve been comfortable managing through it, where it’s not one margin profile. But in aggregate, we will have high ROIC, and we will make sure that we have an infrastructure business that’s got ROIC that’s commensurate with an infrastructure business, and we have a business that builds on top of it, which I’d like call it like the new apps are agents. So we’ll have agent businesses in security, in coding, in knowledge work, as the three big domains.

We’ll get to agents in a little bit, but I didn’t expect to ask this question, big news this week, will you ever issue equity to fund this build out?

SN: Yeah, I just saw the news, I think Google just did it.

Were you as surprised as everyone else?

SN: I’m not sure, exactly, I’ve not studied it, it came last night, I think, so I’ve got to go understand what’s happening. But, it’s like maybe it’s the thing to do is everybody is going public or reissuing equity, maybe that’s the season.

Gobble up some of the money.

The Software Business

Is software dead?

SN: I think software is alive, but the way I think this entire meme has come about is, like, if you take the SaaS question in particular, right? We built in a particular way where I had a data model, and then I had a business logic tier, and then I had a UI tier, I coupled the three, then had a business model.

Integration is a beautiful thing.

SN: Look at this, Ben, right now, we took what is the database that no one knows about underneath Microsoft 365 and said, “Oh, WorkIQ is available, it’s just a skill/MCP, and it’s out there”, and suddenly people are falling in love with, “I can now interrogate and have an agent continuously hit this database to reason over and plan over, act over from any place”.

By the way, it requires a new business model. So, for example, when Cowork is using WorkIQ, that’s going to be a usage-based business model, so I think what needs to happen is we now need to take what we built, rebuild it for the agent era and change the levers of the business model such that you have a per-user business model and you have a consumption business model.

So the hybrid business model, you do think that is going to be the future?

SN: 100%. And once you have that then I think what happened between servers — even I had not understood it when we moved to the cloud, even I was a little worried about, “Oh man, we move to the cloud, we’ll sell the same servers”, and it turned out we sold a lot more subscriptions because people who never bought servers from us were buying subscriptions.

I think that’s what’s happening already with agents, I see that on GitHub, I see that on M365, I see that on security, because everyone is building these agent systems that are continuously “working” and so what we built and thought of as the end-user compute is completely getting rebuilt.

Is there a bit where, if you have to zoom out a hybrid system where a combination of per-seat but also usage, where does E7 fit in this idea, it’s like double the price, it seems it’s an attempt to respond to maybe a secular decrease in seats by increasing ARPU? Is that the right way to think about it?

SN: The way you think about this is, see per-seat is a very important element still because what is per-seat? Per-seat is basically a set of usage entitlements, so anyone who is budgeting really will push you.

That’s right, people don’t like usage, we’re seeing that right now, it could explode.

SN: Exactly, so therefore you just want to take packaging or bundling of usage into proceeds so that there’s some way for people to budget. So I kind of think about the E7, E5, these things will continue and then you’ll always have the outcall consumption. People also talk about, “Hey, maybe people want outcome-based pricing”. Outcome-based pricing, we’ll be thrilled about some of that, but remember, outcome-based pricing is also called royalty. When a customer has a great outcome, they necessarily don’t want to share their outcome so I think what is really being thought about is, ultimately, there is real marginal cost to software, that’s kind of what it is, and that’s going to be priced through.

When did that really click for you, the implications of that?

SN: I think that I would say agents. Before agents, if it is still human interaction—

Right, you can imagine a world where just like basic inference got super cheap and easy.

SN: Exactly, the Moore’s Law itself. Like, if you think about it, if I just used Moore’s Law, get software efficiency, I used software for efficiency and drive that home for customers to have more functionality. In fact, I used to always think about, “Hey, how much more value did we add in M365 and not raise price?” — we didn’t raise prices for a decade plus. That’s all thanks to the software efficiencies on top of hardware.

But now where you are, and if you have a thousand autonomous agents that are all working continuously 24/7 hitting Work IQ, then that is a lot and so that is where I think, and so the real test for me Ben is, that’s why evals, outcomes — no customer will use consumption or their seats if it’s not creating value for them. Therefore, they now are going to be a lot more disciplined on, “What exactly did this stuff do for me?”, “How do I measure it?”, “How do I get into the efficient?”.

And if you think back to going back to the 80s or 90s, where back then it’s like, “Don’t waste time on optimization, the next processor will come out and solve all your problems”, is that now totally the wrong paradigm?

SN: In some sense, you want that to happen, but you can’t just count on that.

It will happen, but your prices will explode.

SN: Exactly, and more importantly, you will be found out if you don’t optimize. Take that example we showed with Land O’Lakes today, which is, here’s an agent, and there is an outcome you care about, I was able to use a model that is using 500B, I was able to use a 5B, and have it really deliver the same outcome, why would I not use that?

That does seem to be a very different thing about this period. It seems clear that’s going to be a huge thing in enterprise going forward, using the right model, optimizing, it’s like we didn’t get to the optimization stage of the PC era.

SN: That’s right.

I don’t think we ever did get there.

SN: We never got there.

Stuff’s still bloated as ever, because everyone just assumes it’s going to get faster, it’s going to be fine.

SN: Exactly, because things were not priced for it. Once you have consumption, everyone will optimize.

For E7, it does seem like the real lure there is Cowork. It’s like this new capability, it’s super powerful, it’s taking Anthropic’s Cowork, which is on your PC, now it’s in the cloud, has all the niceties around that, permissions, controls, all those sorts of things. Is that why it’s there? Is that the hook?

SN: Yeah, there’s also the Agent 365, so there’s a whole lot. Like always, these things, we’re going to take everything from what I’ll talk about as what is an end-user thing and an IT thing, bring it all together.

You guys know bundling.

SN: And security. Yeah, definitely, and they’re all about, ultimately, how do we get the value equation right such that the customer can cover, because right now, it’s kind of fascinating. You have an agent, you immediately say, “Oh, I’ve got to secure it, I’ve got to have observability on it, I need a sandbox for it”. So it’s just that if you don’t bundle, you kind of are sending the customer down the chase of five different things.

With that, though, the reason I find that striking is you’ve talked a lot about — to what extent do you think the point of integration that really matters is it does seem to be increasingly between the models and the harness themselves? You’ve talked about things like your CoreAI initiative and GitHub Copilot, a lot of which is, “We’re going to build the harness and you can slip the models in and out”, and that works right now for Copilot and you can choose your model and even then, from what I’ve heard, not quite as easy as you might think it might be, but it’s still there, the selector’s there. Cowork seems like, “Yeah, that’s right, it has to be the whole package and it’s important for us to have a selling point on E7” — that this feels like maybe it’s not easily substitutable.

SN: No, it is. The same thing on Cowork. In fact, right now, the Cowork that I’m using is already mostly defaulted GPT.

Okay, so it is going to be fully interchangeable?

SN: We’re using the same harness that we use in GitHub and the same thing in security, too. So we have the same harness that’s a multi-model harness in which we will rotate through — obviously MAI by default gets trained in our harness, but we will have GPT, we will have Anthropic in there and any open weight model. We will allow anyone to take any of the models they fine-tune or build. In fact, they can take an open weight model from Fireworks, tune it, put it into Copilot, no problem.

All right, so I am misinformed, so I will take the L on that. Explain what is Cowork then and what is the connection with Anthropic as far as that product goes?

SN: Cowork, to me, it’s kind of like Copilot. I took the term Cowork, it’s part of there and it’s definitely got the Anthropic models in there. Cowork is — think of it as a form factor, the best way to describe it is we built a chat interface first for Copilot, then we now have built Cowork for Copilot, and now we’re building autopilots, as I described it there, think of it as the enterprise-grade OpenClaws. So basically, I think of these as different form factors of agents — chat was the first thing, Cowork is the next thing and in fact, you can even go back to the developer thing. Developers, how did we start? We started with code completions first, then we went to—

I get all this, but I’m genuinely confused here, because I go back to the blog post. It says, “Working closely with Anthropic, we took what they’ve done with Cowork…”.

SN: Yeah, that’s what we launched first. All I’m saying is it’s evolved. It’s kind of like, Copilot today.

Got it, which started out with ChatGPT.

SN: ChatGPT, now it has both Opus and GPT models.

Got it, okay.

SN: So, they’re going to be all over.

All right. So, I wasn’t completely off the reservation.

SN:That’s right.

I failed to catch up, I will accept that.

[Editor’s Note: the FAQ for Cowork still says it uses Anthropic models, just like the original blog post]

SN: Every product of ours, you’ll have both Anthropic and OpenAI models, and MAI models, and your ability to put your own models, and that, I think, is the fundamental promise. Oh, by the way, I should mention this. The amount of auto — I don’t know how much you’re doing selection, I’m mostly auto — and so then one of the biggest pieces of work at Microsoft is all the training models to do auto-routing. That, by the way, is perhaps one of the biggest continuous learning things.**

It’s interesting because I probably approach it more from a consumer perspective, so I just literally choose the app that I want to do something in or call from the CLI.

GitHub Copilot

What happened to Github Copilot? You’re talking about it very positively, but I think a negative spin would be two or three years ago, you were first to market with autocomplete, everyone assumed you got there, you won, and now it’s like, “We’re going to catch up with GitHub Copilot”.

SN: I think what happened is this is one of those classic cases — remember, it was a tools business before, and now it is the business, who would have thought that coding is everything?

Right, it should have been everything, but it seems like for some period of time, it wasn’t?

SN: For us, I think what has happened is we have continued — there are two things that are happening in GitHub, before I even talk about Copilot, I should talk about GitHub. All these coding agents have shown up to work, and where have they shown up? In GitHub. And so the first thing that, quite frankly, I wish we had anticipated better, was the amount of agenting.

The whole GitHub reliability thing is like one thing, but for Copilot specifically.

SN: I’ll say the first thing, that’s kind of, at some level I take that job seriously, because job number one before you want to get to Copilot is go make sure that we are scaling, so let’s leave that alone.

There’s a lot of people very unhappy about that.

SN: Yeah, and we’re going to work it and they should have higher expectations of us and we need to deliver for them.

Then the next thing is on the Copilot side, you’re absolutely right, we started by saying, “This must be just a code completions thing in the IDE”, we added chat, we added tasks, and guess what? Let’s give credit where it needs to be given. Anthropic showed up with a model.

Well, this is like Cursor’s story, they ate your lunch even before Anthropic did. Or you’re saying that that was also an Anthropic story?

SN: Not really, I mean it’s kind of like Cursor/Microsoft, it’s like Borland v us, it’s not like that was not the end all be all.

It was really the Anthropic coming in with a completely different approach, a more agentic approach.

SN: That’s right, with a different approach. With a model and what they’ve done there, and essentially the agent loop is what the change was. In fact, if you look at it, Cursor never, total volume-wise—

They got eaten by the same thing, they’re facing the same challenges.

SN: Also even the market share and so on — Cursor did fantastic, they forked VS Code, did a good job, lots of credit to them. But the real thing was agentic coding became real and now the good news is the agentic coding really drives — people want choice, we will be there, we will have our own models. GitHub itself and Copilot itself will have both the Anthropic and Claude. In fact, the rubber duck feature is my most favorite feature, which is I can use it to check the others.

Windows vs. Project Solara

The headline announcement from this week, I guess is these new Nvidia-based PCs running Windows. However, the announcement I found much more interesting — or not an announcement, preview — Project Solara, viewing these devices as ways to access agents in the cloud, totally different center of gravity. I don’t know if it was you that said it or the presenter, something which I thought was really compelling, which is a limitation of wearables is if you have to interact with them continuously, they get very tiring, so their utility is fundamentally limited. But if you can ask an agent to do something, then you can go do something else and meanwhile, it’s running in the background. Super compelling. I guess the question is, this feels totally different than Windows — it was weird to start this keynote talking about Windows and the AI PC, and that’s nice, and local inference, but this is like, “Actually, what if everything was in the cloud?”.

SN: Yeah, I always find this frame back from 2014 of ubiquitous computing and ambient intelligence and it’s becoming more and more real each day.

First of all, the first part of it was, “I’m so thrilled to have these Windows machines”, and the fact that Jensen had that beautiful slide, the picture of him with all the desktops, I was like “God, yes, I’ve been waiting for it”, which is it’s great, so I think because it makes sense, it makes logical sense to have powerful silicon systems with power that really have it with unmetered intelligence.

When I worked at Windows, I had to like furtively hide my iPhone and then it was okay to show up on campus with an iPhone, now I’m here with a MacBook Air — next time I interview you do I have to feel bad that I don’t have an Nvidia AI PC?

SN: You will always have choice, Ben, and I hope you choose the right thing. I’m excited about that stuff because I think there’s unmetered intelligence, even there was one little feature that we showed, which is that ability to have eight agents running continuously, analyzing logs and so on, but all of them were unmetered.

Right, but that feels like it’s a side project, side quest.

SN: It’s kind of like a billion users all having that, that’s not a side quest. To me, it’s as fundamental as like I think the people are going to want for their knowledge work, for their security work, for their coding work, machines—

They’ll want for themselves. Is this actually the new consumer/enterprise separation?

SN: The enterprise — the business model, we had this long conversation about enterprises continuously optimizing — in fact, I think the biggest value prop of a Windows machine in the enterprise will be unmetered intelligence. So people are going to say, “Oh wow, instead of having my cloud bill keep going up, I’m going to have Windows machine and amortize it that way”, so I think that there is going to be a real value to — because in a world where you have infinite amount of tokens you want to consume, you want to optimize, and why would I not optimize using everything?

I don’t know, I just feel like — as you know, I’ve been very impressed with the job you’ve done with Microsoft, ending the stranglehold Windows had on the company, I still remember I was actually in the Bay Area, I was sitting at the bar at The Westin by the airport typing The End of Windows, recounting all these things you did to not kill Windows, but not make it the center of gravity for the company.

SN: And that I think is what goes to Solara. I don’t think Windows, we are trying to make Windows—

For sure.

SN: Solara, to your point, I thought it was a great question, because the thing that I want us to take a shot at is the following which is, “Can you think of a platform and platform rules, by the way, which are built for the agent era?” — because right now, what is everyone else who are “platform owners” who will try to move from the phone to this wearables will try to bring their apps to the same game, right? I want to open that up, so I would like, for example, like what we were able to do with Teams devices, and that’s where we built some of this sort of distribution capability, so I want to use that connected to this agent world so I’m excited I’m in MediaTek, Qualcomm.

Well I have a great analogy for you, I think. So there’s a bit where I think you just circle back to the great job you’ve done as CEO — this is the butter-up portion of the interview — there is a bit where I think you benefited from following the follower as it were. Steve Ballmer’s one that had to go after Bill Gates and he for better or worse created the conditions for you to succeed, I think is one way to put it, is it possible that for this, your opportunity device space — like can Apple ever really make an agent that works everywhere as long as they’re stuck on the phone?

SN: That’s a great question. That is the question for all of us which is you know the reality is it’s easy to say for someone who’s been so successful with something that in face continues to have a lot of success and say, “I’m going to burn it all down and build something else”.

But to the point, the way they’re architectured, everyone’s vertical.

SN: Exactly, it’s not natural. Like you think about it, we’re saying, “Building agents is easy”, the SOCs are jumping out everywhere, they’re there, the silicon is easy, the system is easy, the operating system is built, and now you’re telling me that I have only one choice for an ambient thing in a hotel, in a restaurant, in a healthcare setting? It makes no sense. So therefore, I imagine that building these ambient devices using Project Solara will be as easy — if you’re successful a year from now, everybody, even in the enterprise, is going to say, “Oh, I’m just going to order a bunch of these things from a no-name ODM who just built it for me”.

I think it’s super smart to start at the enterprise only. Do you have dreams that maybe this will eventually spill over?

SN: Right now, I want us to again do what I think is natural, like where am I seeing people—

Well, that’s where you have the Microsoft 365 environment, you have all the context there.

SN: And also the agents, where would people build agents? The thing is, the consumer one will be like, “I need the one agent I want”, so it’s not like I’m not building a Copilot device, I’m building an agentic platform where the healthcare provider can have their own agent, so that’s the right place for Microsoft to start, let’s see how it goes.

Datacenters

One last question. You had a data center segment appropriately focused on communities, you talked about things like paying your way for electricity, not using water, building up the tax base, education, etc. Why not just pay the residents? Just pay them a dividend?

SN: I’m open to all ideas here, I’m not close-minded at all because at the end of the day, I think the fundamental thing you’re asking about is, “How does this industry, including Microsoft, have permission to do what we’re doing in terms of infrastructure build out?”.

My theory is we get to everything backwards in the US, this is how we back into UBI [Universal Basic Income], is we’re just paying people to build data centers.

SN: Yeah. And I mean, one thing that I have an issue with things like UBI and so on are the—

I’m anti-UBI. That’s how you get there while being anti-UBI.

SN: I want people and communities to have control, have agency, humans to have real dignity in their work and you’re 100% right in saying, “Look, we have to do what it takes to get that permission”. And so right now, there’s so much about our industry that’s so glorious, so good, so great.

What about the you’re going to lose your job part?

SN: Yeah, that’s the problem. Self-obsession about our own glory and our own — if you’re not creating opportunity, why would anybody want you to succeed? That’s the fundamental memo that needs to be re-sent to everyone across our industry, and then we have to live up to it.

Satya Nadella, great to talk to you again.

SN: Thank you so much, Ben, as always.

---

This Daily Update Interview is also available as a podcast. To receive it in your podcast player, visit Stratechery.

The Daily Update is intended for a single recipient, but occasional forwarding is totally fine! If you would like to order multiple subscriptions for your team with a group discount (minimum 5), please contact me directly.

Thanks for being a supporter, and have a great day!