Satya Nadella — Microsoft’s AGI plan & quantum breakthrough

Satya, thank you so much for coming on the podcast. In a second, we're going to get to the 2 breakthroughs that Microsoft has just made—and congratulations on having them come out in Nature on the same day: the Majorana zero chip, which we have in front of us right here, and also the world human action models. But can we just continue the conversation we were having a second ago? You're describing the ways in which the things you were seeing in the '80s and '90s, you're seeing them happen again.

The thing that is exciting for me—it reminds me a little bit of my first few years in the tech industry, starting in the '90s, where there was real debate about whether it was going to be RISC or CISC, or, “Hey, are we really going to be able to build servers using x86?” When I joined Microsoft, that was the beginning of what was Windows NT. So everything from the core silicon platform to the operating system to the app tier—that full-stack approach—the entire thing is being litigated.

You could say cloud did a bunch of that, and obviously distributed computing and cloud did change client-server. The web changed massively. But this does feel a little more full-stack than even the past that at least I've been involved in.

When you think about which decisions ended up being the long-term winners in the '80s and '90s, and which ones didn't—and especially when you think about your time at Sun Microsystems—they had an interesting experience with the '90s dot-com bubble. People talk about this data center build-out as being a bubble, but at the same time, we have the Internet today as a result of what was built out then. What are the lessons about what will stand the test of time? What is an inherent secular trend? What is just ephemeral?

If I go back, the 4 big transformations that I've been part of are the client and client-server. That's the birth of the graphical user interface and the x86 architecture, basically allowing us to build servers. It was very clear to me. I remember going to what is now PDC in '91—in fact, I was at Sun at that time. In '91, I went to Moscone. That's when Microsoft first described the Win32 interface, and it was pretty clear to me what was going to happen: the server was also going to be an x86 thing.

When you have the scale advantages accruing to something, that's the secular bet you have to place. What happened in the client was going to happen on the server side, and then you were able to actually build client-server applications. So the app model became clear.

Then the web was the big thing for us, which we had to deal with. In fact, as soon as I joined Microsoft, the Netscape browser—or the Mosaic browser—came out, I think in December or November of '93. I think that's when Marc Andreessen and his crew had that. So that was a big game changer, in an interesting way, just as we were getting going on what was the client-server wave, and it was clear that we were going to win it as well.

We had the browser moment, and so we had to adjust. We did a pretty good job of adjusting to it because the browser was a new app model. We were able to embrace it with everything we did, whether it was HTML in Word, building a new thing called the browser ourselves and competing for it, or building a web server on our server stack and going after it.

Except, of course, we missed what turned out to be the biggest business model on the web, because we all assumed the web was all about being distributed. Who would have thought that search would be the biggest winner in organizing the web? That's where we obviously didn't see it, and Google saw it and executed super well.

So that's 1 lesson learned for me: you have to not only get the tech trend right, you also have to get where the value is going to be created with that trend. These business model shifts are probably tougher than even the tech trend changes.

Where is the value going to be created in AI?

1. AI won't be winner-take-all

That's a great one. I think there are 2 places where I can say with some confidence. One is the hyperscalers that do well, because the fundamental thing is, if you go back to even how Sam and others describe it, if intelligence is the log of compute, whoever can do lots of compute is a big winner.

The other interesting thing is, if you look underneath even any AI workload—take ChatGPT—it's not like everybody's excited about what's happening on the GPU side, which is great. In fact, I think of my fleet even as a ratio of the AI accelerator to storage to compute. At scale, you've got to grow it.

That infrastructure need for the world is just going to be growing exponentially.

Yeah.

So, in fact, it's manna from heaven to have these AI workloads because, guess what? They're more hungry for more compute—not just for training, but we now know for test-time compute.

When you think of an AI agent, it turns out the AI agent is going to exponentially increase compute usage because you're not even bound by just 1 human invoking a program. It's 1 human invoking programs that invoke lots more programs. That's going to create massive, massive demand and scale for compute infrastructure.

Our hyperscale business, our Azure business, and other hyperscalers—I think that's a big thing.

Then after that, it becomes a little fuzzy. You could say, “Hey, there is a winner-take-all model.” I just don't see it. This, by the way, is the other thing I've learned: being very good at understanding what are winner-take-all markets and what are not winner-take-all markets is, in some sense, everything.

I remember even in the early days, when I was getting into Azure, Amazon had a very significant lead. People would come to me, and investors would come to me, and say, “Oh, it's game over. You'll never make it. Amazon—it's winner-take-all.” Having competed against Oracle and IBM in client-server, I knew that the buyers would not tolerate winner-take-all.

Structurally, hyperscale will never be winner-take-all because buyers are smart. Consumer markets sometimes can be winner-take-all, but anything where the buyer is a corporation, an enterprise, or an IT department, they will want multiple suppliers. So you've got to be 1 of the multiple suppliers.

That, I think, is what will happen even on the model side. There will be open source. There will be a governor. Just like on Windows, 1 of the big lessons learned for me was, if you have a closed-source operating system, there will be a complement to it, which will be open source. To some degree, that's a real check on what happens.

I think in models there is 1 dimension of maybe there will be a few closed-source models, but there will definitely be an open-source alternative. The open-source alternative will actually make sure that closed-source winner-take-all is mitigated. That's my feeling on the model side.

And by the way, let's not discount that if this thing is really as powerful as people make it out to be, the state is not going to sit around and wait for private companies to go around the world. So I don't see it as winner-take-all.

Then above that, I think it's going to be the same old stuff: in consumer, in some categories, there may be some winner-take-all network effects.

After all, ChatGPT is a great example. It's an at-scale consumer property that has already got real escape velocity. I go to the App Store and see it's always there in the top 5, and I say, “Wow, that's pretty unbelievable.” So they were able to use that early advantage and parlay that into an app advantage. In consumer, that could happen.

In the enterprise, again, I think there will be different winners by category. That's at least how I analyze it.

I have so many follow-up questions. We have to get to quantum in just a second, but on the idea that maybe the models get commoditized: maybe somebody could have made a similar argument a couple of decades ago about the cloud—that fundamentally, it's just a chip and a box. But in the end, of course, you and many others figured out how to get amazing profit margins in the cloud.

You figured out ways to get economies of scale and add other value. Fundamentally, even forgetting the jargon, if you've got AGI and it's helping you make better AIs—right now, it's synthetic data and reinforcement learning; maybe in the future, it's an automated AI researcher—that seems like a good way to entrench your advantage there.

I'm curious what you make of that—the idea that it really matters to be ahead there.

At scale, nothing is a commodity. To your point about cloud, everybody would say, “Oh, cloud’s a commodity.” Except when you scale—that’s why the know-how of running a hyperscaler matters. You could say, “Oh, what the heck? I can just rack and stack servers.”

In fact, in the early days of hyperscale, most people thought, “There are all these hosters, and those are not great businesses. Will there be anything? Is there even a business in hyperscale?” It turns out there is a real business, just because of the know-how of running, in the case of Azure, the world’s computing across 60-plus regions with all the compute. It’s just a tough thing to duplicate.

So I was more making the point: is it one winner? Is it winner-take-all or not? That’s what you’ve got to get right. I like to enter categories that are big TAMs, where you don’t have to have the risk of it all being winner-take-all. The best market to be in is a big market that can accommodate a couple of winners, and you’re one of them. That’s what I meant by the hyperscale layer.

In the model layer, models ultimately need to run on some hyperscale compute. So that nexus, I feel, is going to be there forever. It’s not just the model; the model needs state. That means it needs storage, and it needs regular compute for running these agents and the agent environments.

That’s how I think about why the limit of one person running away with one model and building it all may not happen.

On the hyperscaler side, it’s also interesting to think about the advantage you as a hyperscaler would have, especially with inference-time scaling. If that’s involved in training future models, you can amortize your data centers and GPUs not only for the training, but then use them again for inference.

I’m curious what kind of hyperscaler you consider Microsoft Azure to be. Is it on the pre-training side? Is it on providing the o3-type inference? Or are you saying, “We’re going to host and deploy any single model that’s out there in the market,” and you’re sort of agnostic about that?

It’s a good point. The way we want to build out the fleet is, in some sense, to ride Moore’s law. I think this will be like what we’ve done with everything else in the past: every year, keep refreshing the fleet, depreciate it over whatever the lifetime of these things are, and get very, very good at placement of the fleet such that you can run different jobs at high utilization.

Sometimes there are very big training jobs that need to have highly concentrated peak FLOPs provisioned for them, and those also need to cohere. That’s great. We should have enough data center footprint to be able to provide that.

At the end of the day, these are all becoming so big. Even if you take pre-training scale, if it needs to keep going, at some point pre-training scale has to cross data center boundaries. It’s all more or less there. Once you start crossing pre-training data center boundaries, is it that different from anything else?

The way I think about it is: distributed computing will remain distributed. So go build out your fleet such that it’s ready for large training jobs, it’s ready for test-time compute, and it’s ready—in fact, if this RL thing happens, you build one large model, and then after that, there’s tons of RL going on.

To me, it’s more training FLOPs, because you want to create these highly specialized, distilled models for different tasks. So you want that fleet, and then the serving needs.

At the end of the day, the speed of light is the speed of light, so you can’t have one data center in Texas and say, “I’m going to serve the world from there.” You’ve got to serve the world by having an inference fleet everywhere in the world. That’s how I think of our build-out of a true hyperscale fleet.

By the way, I want my storage and compute close to all of these things, because it’s not just AI accelerators that are stateless. My training data itself needs storage, and then I want to be able to multiplex multiple training jobs. I want to be able to have memory, and I want to be able to have these environments in which these agents can execute programs. That’s how I think about it.

2. World economy growing by 10

You recently reported that your yearly revenue from AI is $13 billion. If you look at your year-on-year growth on that, in 4 years it’ll be 10x that. You’ll have $130 billion in revenue from AI if the trend continues.

If it does, what do you anticipate doing with all that intelligence, this industrial-scale use? Is it going to be through Office? Is it going to be you deploying it for others to host? You’ve got to have AGIs to have $130 billion in revenue. What does it look like?

The way I come at it, Dwarkesh, it’s a great question, because at some level, if you’re going to have this explosion, this abundance, this commodity of intelligence available, the first thing we have to observe is GDP growth.

Before I get to what Microsoft’s revenue will look like, there’s only one governor in all of this. This is where we get a little bit ahead of ourselves with all this AGI hype. Remember, the developed world is what—2% growth? And if you adjust for inflation, it’s 0%?

In 2025, as we sit here, I’m not an economist, but at least I look at it and say we have a real growth challenge. The first thing we all have to do is, when we say this is like the Industrial Revolution, let’s have that Industrial Revolution-type of growth.

That means, to me, 10%, 7%—developed-world, inflation-adjusted, growing at 5%. That’s the real marker. It can’t just be supply-side.

In fact, a lot of people are writing about this, and I’m glad they are: the big winners here are not going to be tech companies. The winners are going to be the broader industries that use this commodity that, by the way, is abundant. Suddenly, productivity goes up and the economy is growing at a faster rate. When that happens, we’ll be fine as an industry.

But that’s, to me, the moment. Our self-claiming of some AGI milestone—that’s just nonsensical benchmark hacking to me. The real benchmark is the world growing at 10%.

Okay, so if the world grew at 10%, the world economy is $100 trillion or something. If the world grew at 10%, that’s an extra $10 trillion in value produced every single year.

If that is the case, you as a hyperscaler—it seems like $80 billion is a lot of money. Shouldn’t you be doing $800 billion?

That is correct. But the classic supply side is, “Hey, let me build it and they’ll come.” That’s an argument, and after all, we’ve done that. We’ve taken enough risk to go do it.

But at some point, supply and demand have to map. That’s why I’m tracking both sides of it. You can go off the rails completely when you are hyping yourself on the supply side versus really understanding how to translate that into real value for customers.

That’s why I look at my inference revenue. That’s one of the reasons why even the disclosure on the inference revenue—it’s interesting that not many people are talking about their real revenue. But to me, that is important as a governor for how you think about it.

You’re not going to say they have to symmetrically meet at any given point in time, but you need to have existence proof that you are able to parlay yesterday’s—let’s call it—capital into today’s demand, so that then you can invest again, maybe exponentially even, knowing that you’re not going to be completely rate-mismatched.

I wonder if there’s a contradiction in these 2 different viewpoints, because one of the things you’ve done wonderfully is make these early bets. You invested in OpenAI in 2019, even before there was Copilot and any applications.

If you look at the Industrial Revolution, these 6%–10% build-outs of railways and whatever things—many of those were not like, “We’ve got revenue from the tickets, and now we’re going to…” There was a lot of money lost.

So if you really think there’s some potential here to 10x or 5x the growth rate of the world, and then you’re like, “Well, what is the revenue from GPT-4?” If you really think that’s the possibility from the next level up, shouldn’t you just say, “Let’s go crazy, let’s do the hundreds of billions of dollars of compute”? I mean, there’s some chance, right?

That’s true. So if you really think there’s potential here to 10x or 5x the growth rate of the world, and then you’re asking, “What is the revenue from GPT-4?”—if you really think that’s the possibility from the next level up, shouldn’t you just say, “Let’s go crazy, let’s do the hundreds of billions of dollars of compute”? There’s some chance, right?

Here’s the interesting thing: that’s why even that balanced approach to the fleet, at least, is very important to me.

It's not about building compute. It's about building compute that can actually help me not only train the next big model but also serve the next big model. Until you do those 2 things, you're not going to be in a position to take advantage of even your investment.

That's kind of where it's not a race to just building a model; it's a race to creating a commodity that's getting used in the world. You have to have a complete thought, not just one thing that you're thinking about.

By the way, one of the things is that there will be overbuild. To your point about what happened in the dot-com era, the memo has gone out that, hey, you need more energy and you need more compute. Thank God for it. Everybody's going to race.

In fact, it's not just companies deploying; countries are going to deploy capital, and there will clearly be overbuild. I'm so excited to be a leaser because, by the way, I build a lot and I lease a lot. I am thrilled that I'm going to be leasing a lot of capacity in 2027 and 2028 because I look at the builds and I'm saying, "This is fantastic." The only thing that's going to happen with all the compute builds is the prices are going to come down.

3. Decreasing price of intelligence

Speaking of prices coming down, you recently tweeted after the DeepSeek model came out about Jevons' paradox. I'm curious if you can flesh that out. Jevons' paradox occurs when the demand for something is highly elastic. Is intelligence that bottlenecked on prices going down?

Because when I think about at least my use cases as a consumer, intelligence is already so cheap. It's like 2 cents per million tokens. Do I really need it to go down to 0.02 cents? I'm just really bottlenecked on it becoming smarter. If you need to charge me 100 times more, do a 100 times bigger training run. I'm happy for companies to take that. But maybe you're seeing something different on the enterprise side or something.

What is the key use case of intelligence that really requires it to get to 0.002 cents per million tokens?

I think the real thing is the utility of the tokens. Both need to happen: Intelligence needs to get better and cheaper. Anytime there's a breakthrough, like even what DeepSeek did, the efficient frontier of performance per token changes, the curve gets bent, and the frontier moves. That just brings more demand. That's what happened with cloud.

Here's an interesting thing: We used to think, "Oh my God, we've sold all the servers in the client-server era." Except once we started putting servers in the cloud, suddenly people started consuming more because they could buy it cheaper, it was elastic, and they could buy it as a meter versus a license. It completely expanded.

I remember going, let's say, to a country like India and talking about, "Here is SQL Server." We sold a little, but the cloud in India is so much bigger than anything that we were able to do in the server era. I think that's going to be true.

If you think about it, if you want to really have, in the Global South, in a developing country, these tokens available for healthcare that were really cheap, that would be the biggest change ever.

I think it's quite reasonable for somebody to hear people like me in San Francisco and think, "They're kind of silly. They don't know what it's actually like to deploy things in the real world." As somebody who works with these Fortune 500 companies and is working with them to deploy things for hundreds of millions or billions of people, what's your sense of how fast deployment of these capabilities will be?

Even when you have working agents, even when you have things that can do remote work for you, with all the compliance and all the inherent bottlenecks, is that going to be a big bottleneck, or is that going to move past pretty fast?

It is going to be a real challenge because the real issue is change management, or process change. Here's an interesting thing: One of the analogies I use is, just imagine how a multinational corporation like us did forecasts before PCs, email, and spreadsheets.

Faxes went around. Somebody then got those faxes and did an interoffice memo that went around, and people entered numbers. Ultimately, a forecast came, maybe just in time for the next quarter.

Then somebody said, "Hey, I'm just going to take an Excel spreadsheet, put it in email, send it around. People will go edit it, and I'll have a forecast." So the entire forecasting business process changed because the work artifact and the workflow changed.

That is what needs to happen with AI being introduced into knowledge work. In fact, when we think about all these agents, the fundamental thing is that there's new work and a new workflow.

For example, even prepping for our podcast, I go to my Copilot and I say, "Hey, I'm going to talk to Dwarkesh about our quantum announcement and this new model that we built for game generation. Give me a summary of all the stuff that I should read up on before going."

It knew the 2 Nature papers and took that. I even said, "Hey, go give it to me in a podcast format." It did a nice job of 2 of us chatting about it. That became the new workflow, and then I shared it with my team. I took it and put it into Pages, which is our artifact, and then shared it.

So the new workflow for me is, I think, working with AI and working with my colleagues. That's a fundamental change-management process for everyone who's doing knowledge work: suddenly figuring out these new patterns of, "How am I going to get my knowledge work done in new ways?"

That is going to take time. It's going to be something like this in sales, finance, and supply chain.

For an incumbent, I think this is going to be one of those things where—let's take one of the analogies I like to use—manufacturers did this with Lean. I love that because, in some sense, if you look at it, Lean became a methodology for how one could take an end-to-end process in manufacturing and become more efficient.

It's that continuous improvement, which is reducing waste and increasing value. That's what's going to come to knowledge work. This is like Lean for knowledge work, in particular.

That's going to be the hard work of management teams and individuals who are doing knowledge work, and that's going to take time.

Can I ask you just briefly about that analogy? One of the things Lean did is physically transform what a factory floor looks like. It revealed bottlenecks that people didn't realize until they were really paying attention to the processes and workflows.

You mentioned briefly how your own workflow has changed as a result of AI. I'm curious if we can add more color to what it will be like to run a big company when you have these AI agents that are getting smarter and smarter over time.

It's interesting you ask that. I was thinking, for example, today, if I look at it, we are very email-heavy. I get in in the morning, my inbox is full, and I'm responding, so I can't wait for some of these Copilot agents to automatically populate my drafts so that I can start reviewing and sending.

But I already have at least 10 agents in Copilot, which I query for different tasks. I feel like there's a new inbox that's going to get created: my millions of agents that I'm working with will have to raise some exceptions to me, send notifications to me, and ask for instructions.

So at least what I'm thinking is that there's a new scaffolding, which is the agent manager. It's not just a chat interface. I need something smarter than a chat interface to manage all the agents and their dialogue.

That's why I think of Copilot, as the UI for AI, as a big, big deal. Each of us is going to have it. Basically, think of it as: There is knowledge work, and there's a knowledge worker. The knowledge work may be done by many, many agents, but you still have a knowledge worker who is dealing with all the knowledge workers.

That, I think, is the interface that one has to build.

4. Microsoft's Quantum breakthrough

You're one of the few people in the world who can say that you have access to 200,000—you have this swarm of intelligence around you in the form of Microsoft, the company, and all its employees. You have to manage that, and you have to interface with that, figuring out how to make the best use of it. Hopefully, more of the world will get to have that experience in the future.

I'd be curious about how your inbox—if that means everybody's inbox—will look like yours in the morning.

Okay, before we get to that, I want to keep asking you more about AI, but I really want to ask you about the big breakthrough in quantum that Microsoft Research has announced. Can you explain what's going on?

This has been another 30-year journey for us. It's unbelievable. I'm the third CEO of Microsoft who's been excited about quantum. The fundamental breakthrough here, or the vision that we've always had, is that you need a physics breakthrough in order to build a utility-scale quantum computer that works.

We took the path of saying the one way to have a less noisy or more reliable qubit is to bet on a physical property that, by definition, is more reliable. That's what led us to the Majorana zero modes, which were theorized in the 1930s. The question was, can we actually physically fabricate these things? Can we actually build them?

The big breakthrough, effectively—and I know you talked to Chetan—is that we now finally have an existence proof and a physics breakthrough of Majorana zero modes in a new phase of matter, effectively. This is why we like the analogy of thinking of this as the transistor moment of quantum computing, where we effectively have a new phase, which is the topological phase. That means we can now reliably hide the quantum information, measure it, and fabricate it.

Now that we have it, we feel like, with that core foundational fabrication technique out of the way, we can start building a Majorana chip. Majorana One, I think, is going to basically be the first chip that will be capable of 1 million physical qubits. On that, there will be thousands of logical, error-corrected qubits. Then it's game on.

You suddenly have the ability to build a real utility-scale quantum computer, and that, to me, is now so much more feasible. Without something like this, you will still be able to achieve milestones, but you'll never be able to build a utility-scale computer. That's why we're excited about it.

Amazing. And by the way, I believe this is it right here.

That is it. Yes.

I forget now, are we calling it Majorana?

Yes, that's right. Majorana One. I'm glad we named it after that.

To think that we are able to build something like a 1-million-qubit quantum computer in a thing of this size is just unbelievable. That's the crux of it: unless and until we could do that, you can't dream of building a utility-scale quantum computer.

And you're saying the eventual 1 million qubits will go on a chip this size? Okay, amazing. Other companies have announced 100 physical qubits—Google's, IBM's, and others. When you say you've announced one, are you saying that yours is way more scalable in the limit?

Yeah. The one thing we've also done is taken an approach where we've separated our software and our hardware. We're building out our software stack, and we now have, with the neutral-atom folks and the ion-trap folks—we're also working with others who have pretty good approaches with photonics and what have you. That means there will be different types of quantum computers.

In fact, I think the last thing that we announced was 24 logical qubits. We've also had some fantastic breakthroughs on error correction, and that's what is allowing us, even on neutral-atom and ion-trap quantum computers, to build these 20-plus-qubit systems. I think that will keep going even throughout the year; you'll see us improve that yardstick.

But we also said, “Let's go to first principles and build our own quantum computer that's betting on the topological qubit.” That's what this breakthrough is about.

Amazing. The 1 million topological qubits, thousands of logical qubits—what is the estimated timeline to scale up to that level? What does Moore's law here, if you've got the first transistor, look like?

We've obviously been working on this for 30 years. I'm glad we now have the physics breakthrough and the fabrication breakthrough. I wish we had a quantum computer because, by the way, the first thing the quantum computer will allow us to do is build quantum computers. It’s going to be so much easier to simulate, atom by atom, the construction of these new quantum gates.

But in any case, the next real thing is, now that we have the fabrication technique, to go build that first fault-tolerant quantum computer. That will be the logical thing. I can say now, “Oh, maybe ’27, ’28, ’29, we will be able to actually build this.”

Now that we have this one gate, can I put the thing into an integrated circuit and then actually put these integrated circuits into a real computer? That is where the next logical step is.

And what do you see as happening in ’27, ’28? You've got it working—is it a thing you access through an API? Is it something you're using internally for your own research in materials and chemistry?

It’s a great question. One thing that I've been excited about is that, even in today's world, we had this quantum program, and we added some APIs to it. The breakthrough we had maybe 2 years ago was to think of this HPC stack, AI stack, and quantum together.

In fact, if you think about it, AI is like an emulator of the simulator. Quantum is like a simulator of nature. What is quantum going to do? By the way, quantum is not going to replace classical. Quantum is great at what quantum can do, and classical will also…

Quantum is going to be fantastic for anything that is not data-heavy but is exploration-heavy in terms of the state space. It should be data-light but involve exponential states that you want to explore. Simulation is a great one: chemical physics, what have you, biology.

One of the things that we've started doing is really using AI as the emulation engine, but you can then train it. The way I think of it is, if you have AI plus quantum, maybe you'll use quantum to generate synthetic data that then gets used by AI to train better models that know how to model something like chemistry or physics or what have you. These 2 things will get used together.

Even today, that's effectively what we're doing with the combination of HPC and AI. I hope to replace some of the HPC pieces with quantum computers.

Can you tell me a little bit about how you make these research decisions which, in 20 years' time or 30 years' time, will actually pay dividends, especially at a company of Microsoft's scale? Obviously, you're in great touch with the technical details in this project. Is it feasible for you to do that with all the things Microsoft Research does?

How do you know the current bet you're making will pay out in 20 years? Does it just have to emerge organically through the organization, or how are you keeping track of all this?

The thing that I feel was fantastic is that when Bill started MSR back in 1995, I think, in the long history of these curiosity-driven research organizations, to just have a research organization that is about fundamental research—MSR, over the years, has built up that institutional strength.

When I think about capital allocation or budgets, we first put the chips in and say, “Here is MSR's budget.” We've got to go at it each year knowing that most of these bets are not going to pay off in any finite time frame. Maybe the 6th CEO of Microsoft will benefit from it. In tech, that is, I think, a given.

The real thing that I think about is, when the time has come for something like quantum or a new model or what have you, can you capitalize? As an incumbent, if you look at the history of tech, it's not that people didn't invest. It's that you need to have a culture that knows how to take an innovation and scale it.

That's the hard part, quite frankly, for CEOs and management teams. It's as much about good judgment as it is about good culture. Sometimes we've gotten it right; sometimes we've gotten it wrong. I can tell you the thousand projects from MSR that we should have probably led with, but we didn't.

I always ask myself why. It's because we were not able to get enough conviction and that complete thought of how to not only take the innovation but make it into a useful product with a business model that we can then go to market with.

That's the job of CEOs and management teams: not to just be excited about any one thing, but to be able to actually execute on a complete thing. And that's easier said than done.

When you mentioned the possibility of 3 subsequent CEOs of Microsoft, if each of them increases the market cap by an order of magnitude, by the time you've got the next breakthrough, you'll be like the world economy or something.

Or remember, the world is going to be growing at 10%, so we'll be fine.

5. Microsoft's gaming world model

Let's dig into the other big breakthrough you've just made. It's amazing that you have both of them coming out the same day in your gaming world models. I'd love it if you could tell me a little bit about that.

We're going to call it Muse. It's going to be the model of this world action, or human action model. This is very cool. One of the things is that, obviously, DALL·E and Sora have been unbelievable in what they've been able to do in terms of generative models. One thing that we wanted to go after was using gameplay data. Can you actually generate games that are both consistent and have the ability to generate the diversity of what that game represents, and then are persistent to user mods? That's what this is.

They were able to work with one of our game studios, and this is the other publication in Nature. The cool thing I'm excited about is that we're going to have a catalog of games soon that we will start using these models, or we're going to train these models to generate, and then start playing them.

In fact, when Phil Spencer first showed it to me, he had an Xbox controller, and this model basically took the input and generated the output based on the input. It was consistent with the game. That, to me, is a massive moment of “wow.” It's kind of like the first time we saw ChatGPT complete sentences, or DALL·E draw, or Sora. This is one such moment.

I got a chance to see some of the videos in the real-time demo this morning with your lead researcher, Katja, on this. It was only once I talked to her that it really hit me how incredible this is, in the sense that we've used AI in the past to model agents, and just using that same technique to model the world around the agent gives consistent, real-time worlds. We'll superimpose videos of what this looks like atop this podcast so people can get a chance to see it for themselves. I guess it'll be out by then, so they can also watch it there.

This in itself is incredible. Through your span as CEO, you've invested tens of hundreds of billions of dollars in building up Microsoft Gaming and acquiring IP. In retrospect, if you can just merge all of this data into one big model that can give you this experience of visiting and going through multiple worlds at the same time, and if this is the direction gaming is headed, it seems like a pretty good investment to have made. Did you have any premonition about this?

I wouldn't say that we invested in gaming to build models. We invested, quite frankly, because—here's an interesting thing about our history: we built our first game before we built Windows. Flight Simulator was a Microsoft product long before we even built Windows. Gaming has a long history at the company, and we want to be in gaming for gaming's sake.

I always start by saying I hate to be in businesses where they're a means to some other end. They have to be ends unto themselves. And then, yes, we're not a conglomerate. We are a company where we have to bring all these assets together and be better owners by adding value. For example, cloud gaming is a natural thing for us to invest in because that will just expand the TAM and expand the ability for people to play games everywhere.

The same thing with AI and gaming: we definitely think that it can be helpful in maybe changing—it's kind of like the CGI moment, even for gaming long-term. And it's great. As the world's largest publisher, this will be helpful. But at the same time, we've got to produce great-quality games. You can't be a gaming publisher without, first and foremost, being focused on that.

But the fact that this data asset is going to be interesting, not just in a gaming context, but that it's going to be a general action model and a world model, is fantastic. I think about gaming data as perhaps what YouTube is to Google: gaming data is to Microsoft. Therefore, I'm excited about that.

Yeah, and that's what I meant, just in the sense that you can have one unified experience across many different kinds of games. How does this fit into the other things, separate from AI, that Microsoft has worked on in the past, like mixed reality? Maybe giving smaller game studios a chance to build these AAA action games? Just 5 or 10 years from now, what kinds of ways could you imagine?

In an interesting way, even 5, 6, 7 years ago, I said the 3 big bets that we want to place are AI, quantum, and mixed reality. And I still believe in them, because in some sense, what are the big problems to be solved? Presence. That's the dream of mixed reality. Can you create real presence? Like you and I doing a podcast like this.

I think it's still proving to be the harder one of those challenges, quite honestly. I thought it was going to be more solvable. It's tougher, perhaps, just because of the social side of it: wearing things and so on.

We're excited about what we're going to do with Anduril and Palmer now, including how they'll take forward the IVAS program, because that's a fantastic use case. And so we'll continue on that front. But also, the 2D surfaces. It turns out things like Teams, thanks to the pandemic, have really given us the ability to create essentially presence through even 2D. And that, I think, will continue. That's one secular piece.

Quantum we talked about, and AI is the other one. So these are the 3 things that I look at and say: how do you bring these things together? Ultimately, not as tech for tech's sake, but solving some of the fundamental things that we, as humans, want in our life. More than that, we want them in our economy, driving our productivity. If we can somehow get that right, then I think we will have really made progress.

When you write your next book, you've got to have some explanation of why those 3 pieces all came together around the same time, right? There's no intrinsic reason you would think quantum and AI should happen in 2028 and 2025 and so forth.

That's right. At some level, I look at it and say: the simple model I have is, is there a systems breakthrough? To me, the systems breakthrough is the quantum thing. Is there a business-logic breakthrough? That's AI to me: can the logic tier be fundamentally reasoned differently? Instead of imperatively writing code, can you have a learning system? That's the AI one. And then the UI side of it is presence.

6. Legal barriers to AI

Going back to AI for a second: in your 2017 book, you invested in OpenAI in 2019, very early—2017 is even earlier—you say in your book, “One might also say that we're birthing a new species, one whose intelligence may have no upper limits.”

Now, super-early, of course, to be talking about this in 2017. We've been talking in a granular fashion about agents, Office Copilot, capex, and so forth. But if you zoom out and consider this statement you've made, and you think about you as a hyperscaler, as the person doing research in these models as well, providing training, inference, and research for building a new species, how do you think about this in the grand scheme of things?

Do you think we're headed towards superhuman intelligence in your time as CEO? I think even Mustafa uses that term. In fact, he's used that term more recently: this “new species.”

The way I come at it is, you definitely need trust. Before we claim it is something as big as a species, the fundamental thing that we've got to get right is that there is real trust, whether it's personal or societal-level trust, that's baked in. That's the hard problem.

I think the biggest rate limiter to the power here will be our legal infrastructure. We're talking about all the compute infrastructure, but how does the legal infrastructure evolve to deal with this? This entire world is constructed with things like humans owning property, having rights, and being liable. That's the fundamental thing that one has to first say: okay, what does that mean for anything that humans are now using as tools?

And if humans are going to delegate more authority to these things, then how does that structure evolve? Until that really gets resolved, I don't think just talking about the tech capability is going to be enough.

As in, we won't be able to deploy these kinds of intelligences until we figure out how to…?

Absolutely.

Because at the end of the day, there is no way. Today, you cannot deploy these intelligences unless and until there's someone indemnifying it as a human. To your point, I think that's one of the reasons why I think about even the most powerful AI as essentially working with some delegated authority from some human. You can say, “Oh, that's all alignment and this, that, and the other.” That's why I think you have to really get these alignments to work and be verifiable in some way, but I just don't think that you can deploy intelligences that are out of control.

For example, this AI takeoff problem may be a real problem, but before it is a real problem, the real problem will be in the courts. No society is going to allow for some human to say, “AI did that.”

Yes. Well, there's a lot of societies in the world, and I wonder if any one of them might not have a legal system that might be more amenable. And if you can't have a takeoff, then you might worry. It doesn't have to happen in America, right?

We think that no society cares about it, right? There can be rogue actors; I'm not saying there won't be rogue actors. There are cybercriminals and rogue states; they're going to be there. But to think that human society at large doesn't care about it is also not going to be true. I think we all will care.

We know how to deal with rogue states and rogue actors today. The world doesn't sit around and say, “We'll tolerate that.” That's why I'm glad that we have a world order in which anyone who is a rogue actor in a rogue state has consequences.

Right. But if you have this picture where you can have 10% economic growth, I think it really depends on getting something like AGI working, because tens of trillions of dollars of value—that sounds closer to the total of human wages, around $60 trillion of the economy. Getting that magnitude, you kind of have to automate labor or supplement labor in a very significant way.

If that is possible, and once we figure out the legal ramifications for it, it seems quite plausible, even within your tenure, that we figure that out. Are you thinking about superhuman intelligence? Is the biggest thing you do in your career this?

You bring up another point. I know David Autor and others have talked a lot about this. I think the other question that needs to happen is, let's at least talk about our democratic societies. I think that in order to have a stable social structure and democracies function, you can't just have a return on capital and no return on labor. We can talk about it, but that 60% has to be revalued.

In my own simple way, maybe you can call it naive, we'll start valuing different types of human labor. What is today considered high-value human labor may be a commodity. There may be new things that we will value, including that person who comes to me and helps me with my physical therapy or whatever is going to be the case that we value. But ultimately, if we don't have return on labor, and there's meaning in work and dignity in work and all of that, that's another rate limiter to any of these things being deployed.

7. Getting AGI safety right

On the alignment side, 2 years ago, you guys released Sydney Bing. Just to be clear, I think given the level of capabilities at the time, it was a charming, endearing, kind of funny example of misalignment. But that was because, at the time, it was like chatbots: they could go think for 30 seconds and give you some funny or inappropriate response.

But if you think about that kind of system—that, I think, to a New York Times reporter, tried to get him to leave his wife or something—if you think about that going forward, and you have these agents that are, for hours, weeks, months going forward, just like autonomous swarms of AGIs, who could be in similar ways misaligned and screwing stuff up, maybe coordinating with each other, what's your plan going forward so that when you get the big one, you get it right?

That is correct. That's one of the reasons why, when we usually allocate compute, let's allocate compute for the alignment challenge. More importantly, what is the runtime environment in which you are really going to be able to monitor these things? The observability around it?

We do deal with a lot of these things today on the classical side of things as well, like cyber. We don't just write software and then just let it go. You have software and then you monitor it. You monitor it for cyberattacks, you monitor it for fault injections, and what have you. Therefore, I think we will have to build enough software engineering around the deployment side of these, and then inside the model itself, what's the alignment?

These are all—some of them are real science problems. Some of them are real engineering problems, and then we will have to tackle it.

That also means taking on our own liability in all of this. So that's why I'm more interested in deploying these things where you can actually govern what the scope of these things is, and the scale of these things is. You just can't unleash something out there in the world that creates harm, because the social permission for that is not going to be there.

When you get the agents that can really just do weeks' worth of tasks for you, what is the minimum assurance you want before you can let it run a random Fortune 500?

I think when I use something like Deep Research, even, the minimum assurance I think we want is before we especially have physical embodiment of anything. I think that's kind of one of those thresholds when you cross. That might be one place.

Then the other one is, for example, the permissions of the runtime environment in which this is operating. You may want guarantees that it's sandboxed, it's not going out of that sandbox. I mean, we already have web search and we already have it out of the sandbox. But even what it does with web search and what it writes—for example, to your point, if it's just going to write a bunch of code in order to do some computation, where is that code deployed? And is that code ephemeral for just creating that output, versus just going and springing that code out into the world? Those are things that you could, in the action space, actually go control.

And separate from the safety issues, as you think about your own product suite, and you think about, if you do have AIs this powerful, at some point, it's not just like Copilot—an example you mentioned about how you were prepping for this podcast—it's more similar to how you actually delegate work to your colleagues. What does it look like, given your current suite, to add that in?

I mean, there's one question about whether LLMs get commodified by other things. I wonder if these databases or canvases or Excel sheets or whatever—if the LLM is your main gateway into accessing all these things—is it possible that the LLMs commodify Office?

It's an interesting one. The way I think about the first phase, at least, would be: Can the LLM help me do my knowledge work using all of these tools or canvases more effectively? One of the best demos that I've seen is a doctor getting ready for a tumor board workflow. She's going into a tumor board meeting, and the first thing she uses Copilot for is to create an agenda for the meeting because the LLM helps reason about all the cases, which are in some SharePoint site.

It says, “Hey, these cases—obviously, a tumor board meeting is a high-stakes meeting where you want to be mindful of the differences in cases so that you can then allocate the right time.” Even that reasoning task of creating an agenda that knows how to split time—super. So, I use the LLM to do that.

Then I go into the meeting. I'm in a Teams call with all my colleagues. I'm focused on the actual case versus taking notes, because you now have this AI Copilot doing a full transcription of all of this. It's not just a transcript, but a database entry of what is in the meeting that is recallable for all time.

Then she comes out of the meeting, having discussed the case and not been distracted by note-taking. She's a teaching doctor; she wants to go and prep for her class. And so she goes into Copilot and says, “Take my tumor board meeting and create a PowerPoint slide deck out of it so that I can talk to my students about it.”

So that's the type. The UI and the scaffolding that I have are canvases that are now getting populated using LLMs. And the workflow itself is being reshaped; knowledge work is getting done.

Here's an interesting thing: If someone came to me in the late '80s and said, “You're going to have a million documents on your desk,” I would say, “What the heck is that?” I would have literally thought there was going to be a million physical copies of things on my desk.

Except, we do have a million spreadsheets and a million documents. I don’t; you do. They’re all there. That’s what’s going to happen with agents, too.

There will be a UI layer. To me, Office is not just about the office of today; it’s the UI layer for knowledge work. It’ll evolve as the workflows evolve. That’s what we want to build.

I do think the SaaS applications that exist today—these CRUD applications—are going to be fundamentally changed because the business logic will go more into this agentic tier. In fact, one of the other cool things today in my Copilot experience is when I say, “Hey, I’m getting ready for a meeting with a customer,” I just go and say, “Give me all the notes for it that I should know.”

It pulls from my CRM database, it pulls from my Microsoft Graph, creates a composite artifact, essentially, and then it applies logic to it. That, to me, is going to transform the SaaS applications as we know them today.

SaaS as an industry might be worth hundreds of billions to trillions of dollars a year, depending on how you count. If that can really just get collapsed by AI, is the next step in your next decade 10X-ing the market cap of Microsoft again? Because you’re talking about trillions of dollars…

It would also create a lot of value in SaaS. One thing we don’t pay as much attention to, perhaps, is the amount of IT backlog there is in the world.

These code-gen things, plus the fact that I can interrogate all of your SaaS applications using agents and get more utility, will create the greatest explosion of apps. They’ll be called agents, so that for every vertical, in every industry, in every category, we’re suddenly going to have the ability to be serviced.

There’s going to be a lot of value. You can’t stay still. You can’t just say the old thing of, “Oh, I schematized some narrow business process, and I have a UI in the browser, and that’s my thing.” That ain’t going to be the case.

You have to go up-stack and say, “What’s the task that I have to participate in?” You’ll want to be able to take your SaaS application and make it a fantastic agent that participates in a multi-agent world. As long as you can do that, then I think you can even increase the value.

8. 34 years at Microsoft

Can I ask you some questions about your time at Microsoft?

Yeah.

Is being a company man underrated? You’ve spent most of your career at Microsoft, and you could say that one of the reasons you’ve been able to add so much value is that you’ve seen the culture, the history, and the technology. You have all this context from rising up through the ranks. Should more companies be run by people who have this level of context?

That’s a great question. I’ve not thought about it that way. Through my 34 years now at Microsoft, each year I’ve felt more excited about being at Microsoft, rather than thinking, “Oh, I’m a company person,” or what have you.

I take that seriously, even for anybody joining Microsoft. It’s not like they’re joining Microsoft as long as they feel that they can use this as a platform for both economic return, but also a sense of purpose and a sense of mission that they can accomplish by using us as a platform. That’s the contract.

So I think, yes, companies have to create a culture that allows people to come in and become company people like me. Microsoft got it more right than wrong, at least in my case, and I hope that remains the case.

The sixth CEO that you’re talking about, who’ll get to use the research you’re starting now, what are you doing to retain the future Satya Nadellas so that they’re in a position to become future leaders?

It’s fascinating. This is our 50th year, and I think a lot about it. The way to think about it is: longevity is not a goal; relevance is.

The thing that I have to do, and all 200,000 of us have to do every day, is ask: Are we doing things that are useful and relevant for the world as we see it evolving, not just today, but tomorrow?

We live in an industry where there’s no franchise value, so that’s the other hard part. If you take the R&D budget that we will spend this year, it’s all speculation on what’s going to happen 5 years from now.

You have to basically go in with that attitude, saying, “We are doing things that we think are going to be relevant.” That’s what you have to focus on.

Then know that there’s a batting average. You have to have a high tolerance for failure. You have to take enough shots on goal to be able to say, “Okay, we will make it to the other side as a company.” That’s what makes it tricky in this industry.

Speaking of that, you just mentioned that you’re 2 months away from the 50th anniversary of Microsoft’s founding. If you look at the top 10 companies by market cap, or the top 5, basically everybody else but Microsoft is younger than Microsoft.

It’s an interesting observation about why the most successful companies often are quite young. The average Fortune 500 company will last 10 to 15 years. What has Microsoft done to remain relevant for this many years? How do you keep refounding?

I love that Reed Hoffman uses that term, “refounding.” That’s the mindset. People talk about founder mode, but for us mere-mortal CEOs, it’s more like refounder mode.

To be able to see things again in a fresh way is the key. To your question, can we culturally create an environment where refounding becomes a habit? Every day we come in and say, “We feel we have a stake in this place to be able to change the core assumptions of what we do and how we relate to the world around us. Do we give ourselves permission?”

I think many times companies feel over-constrained by either the business model or whatever. You just have to unconstrain yourself.

If you did leave Microsoft, what company would you start?

The company I would start? Man. That’s where the company man in me sort of says, “I’ll never leave Microsoft.”

If I were thinking of doing something, I think picking a domain that has… When I look at the dream of tech, we’ve always said technology is about the biggest, greatest democratizing force. I feel like finally, we have that ability.

If you say those tokens per dollar per watt are what we can generate, I would love to find some domain in which that can be applied, where it is so underserved. That’s where healthcare and education would come in.

Public sector would be another place. If you take those domains, which are the underserved places, where my life as a citizen of this country or a member of this society, or anywhere, would I be better off if somehow all this abundance translated into better healthcare, better education, and better public-sector institutions serving me as a citizen? That would be a place.

9. Does Satya Nadella believe in AGI?

One thing I’m not sure about, hearing your answers on different questions, is whether you think AGI is a thing. Will there be a thing that automates all cognitive labor, like anything anybody can do on a computer?

This is where I have a problem with the definitions of how people talk about it. Cognitive labor is not a static thing. There is cognitive labor today. If I have an inbox that is managing all my agents, is that new cognitive labor?

Today’s cognitive labor may be automated. What about the new cognitive labor that gets created? Both of those things have to be thought of, which is the shifting…

That’s why I make this distinction, at least in my head: don’t conflate knowledge worker with knowledge work. The knowledge work of today could probably be automated.

Who said my life’s goal is to triage my email? Let an AI agent triage my email. But after having triaged my email, give me a higher-level cognitive labor task of, “Hey, these are the 3 drafts I really want you to review.” That’s a different abstraction.

But will AI ever get to the second thing?

It may, but as soon as it gets to that second thing, there will be a third thing. Why are we thinking that somehow, when we have dealt with tools that have changed what cognitive labor is in history, all cognitive labor will go away?

I’m sure you’ve heard these examples before, but the idea that horses can still be good for certain things—there are certain terrains you can’t take a car on. But the idea that you’re going to see horses on the street, that they’re going to employ millions of horses, it’s just not happening.

And then the idea is, could a similar thing happen with humans? But in one very narrow dimension?

It’s only 200 years of human history where we have valued some narrow sort of things called “cognitive labor,” as we understand it.

Let's take something like chemistry. If quantum plus AI really helped us do a lot of novel materials science and so on, that's fantastic—to have novel materials science being done by it. Does that take away from all the other things that humans can do? Why can't we exist in a world where there are powerful cognitive machines, knowing that our cognitive agency has not been taken away?

I'll ask this question not about you, but in a different scenario, so maybe you can answer it without embarrassment. Suppose on the Microsoft board, could you ever see adding an AI to the board? Could it ever have the judgment, context, and holistic understanding to be a useful advisor?

It's a great example. One of the things we added was a facilitator agent in Teams. The goal there—it’s in the early stages—is: can that facilitator agent use long-term memory, not just in the context of the meeting, but with the context of projects I'm working on, the team, and what have you, to be a great facilitator?

I would love it even in a board meeting, where it's easy to get distracted. After all, board members come once a quarter, and they're trying to digest what is happening with a complex company like Microsoft. A facilitator agent that actually helped human beings all stay on topic and focus on the issues that matter—that's fantastic.

That's kind of literally having, to your point about even going back to your previous question, something that has infinite memory that can even help us. After all, what is that Herbert Simon thing? We are all bounded rationality. So if the bounded rationality of humans can actually be dealt with because there is a cognitive amplifier outside, that's great.

Speaking of materials and chemistry, I think you said recently that you want the next 250 years of progress in those fields to happen in the next 25 years. Now, when I think about what's going to be possible in the next 250 years, I'm thinking about space travel, space elevators, immortality, and curing all diseases. Next 25 years, you think?

One of the reasons why I brought that up was, I love that thing of the Industrial Revolution—it was 250 years. We have to take this entire change from a carbon-based system to something different. That means you have to fundamentally reinvent all of what has happened with chemistry over the last 250 years.

That's where I hope we have this quantum computer. This quantum computer helps us get to new materials, and then we can fabricate those new materials that help us with all of the challenges we have on this planet. And then I'm all for interplanetary travel.

Amazing. Satya, thank you so much for your time.

Thank you so much. It's wonderful. Thanks.

Great, thank you.