On February 19, renowned tech podcast host Dwarkesh Patel and Microsoft CEO Satya Nadella discussed Microsoft's AGI plans and quantum breakthroughs in a podcast episode. During the show, Nadella talked about the relationship between AI and economic growth, stating that the true benchmark for AGI is whether the global economy can grow at a rate of 10%. He emphasized that the value of AI lies not only in the technology itself but also in its business model and market positioning, noting that accurately assessing a business model is more challenging than catching technological trends. He also discussed the integration of AI and gaming, highlighting that gaming data will become an important resource for AI training, and shared Microsoft's breakthroughs in quantum computing. Here are the key points from the podcast: The milestone for AGI is not meaningless benchmark tests; the real benchmark is whether the world can grow at a rate of 10%. The development of AI should not merely be viewed as a technological advancement but as a key force driving economic and social transformation. The potential of AI lies in its demand for computational power and its push, which will create significant demand for massive-scale computing infrastructure. The value of AI lies in two aspects: super-scaled participants perform exceptionally well. If we look back at how people like Sam describe it, the intelligence of computation indicates that those who can perform massive calculations are the big winners. Another interesting phenomenon is that, despite the enormous potential of generative pre-trained models (like TikTokGPT), not everyone is excited about their applications in both training and inference stages. Not only the technology itself, but the value of AI also lies in its business model and market positioning. The AI market will not be a "winner-takes-all" landscape. The enterprise market requires diverse suppliers rather than a single monopolist. The development of AI will foster the coexistence of open-source and closed-source models, similar to the relationship between the Windows system and open-source software. To support the development of AI, Microsoft needs to establish a global network of massive-scale data centers to meet the demands of large-scale training, testing, and inference. This infrastructure needs to efficiently utilize computing and storage resources and support distributed computing and global deployment. The development of AI is not just a competition to build models but a race to create "commodities" that can drive global economic growth. Companies need to have a comprehensive perspective rather than focusing solely on a single technology or viewpoint. There may be instances of "overbuilding" in the construction of AI infrastructure, similar to the development trajectory during the internet era. Microsoft has made significant breakthroughs in quantum computing, particularly in the research of topological qubits. Advances in quantum computing will provide new possibilities for solving complex problems (such as materials science and chemistry) and will drive the development of AI. He anticipates that fault-tolerant quantum computers will be built in the coming years. Microsoft has launched the "Muse" project in the gaming sector, utilizing AI to generate game content and create a consistent user experience. Gaming data will become an important resource for AI training, and future games will achieve richer interactions and content generation through AI Microsoft's "three major bets" are AI, quantum computing, and mixed reality. Ultimately, how to combine these elements is not about technology for technology's sake, but about solving some fundamental problems we want as humans in life, as well as more issues we want in the economy to drive productivity. Nadella is cautious about AGI (Artificial General Intelligence), believing that human cognitive labor will not be completely replaced. The development of AI will create new cognitive tasks, and humans will work alongside AI rather than being replaced. If Nadella were to leave Microsoft, he would start a company focused on underserved areas, such as healthcare, education, or public services. These areas are where services are lacking in society, and if all this technology could translate into better healthcare, better education, and better public services, it would be a worthwhile field to invest in. Here is the full podcast transcript: Nadella 00:00 The U.S. claims to have reached a milestone in Artificial General Intelligence (AGI), but that is just a meaningless benchmark attack. The real measure is whether the world is growing at a rate of 10%, and how good we are at determining what is a winner-takes-all market and what is not. In a sense, this is the key to everything. If this technology is truly as powerful as people say, nations would not sit back and wait for private companies to develop it. We can liken it to the transistor matrix in quantum computing. Perhaps you would use quantum technology to generate synthetic data, which would then be used by AI to train Beta models. If intelligence is the logarithm of computation, then whoever can perform a large number of calculations correctly is the big winner, right? Dwarkesh Patel 00:44 Satya, thank you very much for joining the podcast. Next, we will introduce two breakthroughs that Microsoft has just achieved. Congratulations! On the same day, we made progress in the Maia 100 chip field in nature, as well as in the world human behavior model. But can we continue the conversation we just had? You were describing how what you saw in the 80s and 90s is happening again. Nadella 01:09 Dwarkesh, I'm glad to be here. It's great to be on your podcast; I've been a loyal listener. I love the way you conduct these interviews and the wide range of topics you explore. It reminds me of my early years in the tech industry in the early 90s, when there was a real debate about whether there would be risks or if we could really build servers using the x86 architecture. When I joined Microsoft, it was at the beginning of Windows NT. The entire tech stack was undergoing transformation, from the core silicon platform to the operating system and then to the application layer. Just like back then, the entire industry was embroiled in litigation. I think it could be said that cloud technology is like a desert, and distributed computing and cloud technology have indeed brought about significant changes. The client-server architecture and networks have also undergone tremendous changes. But this time it feels more like a complete stack, even more complete than what I was involved in the past Dwarkesh Patel 02:26 Think about it, in the 1980s and 1990s, which decisions ultimately became long-term winners and which did not. Especially when you mention your experience at Sun Microsystems, the internet bubble of the 1990s is an interesting case. People talk about the data center bill as a bubble, but at the same time, it was the construction during that time that gave us the internet we have today. What are the lessons that stand the test of time? What are the inherent long-term trends? What is transient? Nadella 02:54 I think if you look back at the four major technology transformations I was involved in, the first was the rise of client-server architecture, which includes the graphical user interface and the birth of the x86 architecture, which even allowed us to build servers. This is very clear to me. I remember the PDC (Professional Developers Conference) in 1991. At that time, Microsoft described the Win32 interface for the first time. I clearly realized what would happen if servers also adopted the x86 architecture. When you have scale advantages, that’s the long-term bet you have to make. What happens on the client side will also happen on the server side, and then you can actually build client-server applications. This is a shift in the application model. The web was also a big deal for us. In fact, when I joined Microsoft, browsers (like Netscape or Mosaic) had just emerged, probably around November or December of 1993. That was a significant turning point. Nadella 04:24 Just as we were starting the client-server wave, it was clear that we would also win this transformation. We had our moment with browsers, so we had to adapt, and we did well. Because browsers are a new application model, we were able to embrace it and adapt everything we did to it. Whether it was hypertext markup language or developing a new product called a browser and competing in that space, we invested a lot of energy in the web server stack. Of course, we missed the biggest business model on the web because we all thought the web was about distribution. Who would have thought that search would become the biggest winner in organizing the web? That’s where we clearly didn’t see it, while Google saw it and executed very well. So, the lesson I learned is that you not only need to correctly grasp technological trends but also understand where that trend will create value. The shift in business models can be more challenging than the technology trends themselves. AI Will Not Be a Winner-Takes-All Dwarkesh Patel 05:48 Where is the value of AI created? Nadella 05:51 That’s a great question. I think, at least in my work, there are two aspects I can confidently talk about. First, super-scaled participants are performing very well. Fundamentally, if you look back at the way Sam and others described it, the intelligence of computing logarithmically indicates who can do a lot of computation and who the big winners are. Another interesting phenomenon is that even in any AI workload, such as TikTokGPT, not everyone is excited about both sides of the GP (Generative Pre-trained Model) Nevertheless, I believe that the ratio of AI accelerating storage and computing is very important. At scale, you must expand it. Therefore, the world's demand for infrastructure will grow exponentially. Having these AI workloads is like a boost from heaven, as their demand for computing is extremely strong, not only for training but now also for testing time. AI agents will exponentially increase computing usage because they are no longer limited by a person calling the program. This will create huge demand and scale for computing infrastructure. Therefore, I believe our hyperscale business, especially in Asia, will become another hyperscale enterprise. This is an important development direction. However, the situation afterward becomes somewhat unclear. You might say there is a winner-takes-all model, but I don't think so. By the way, this is another thing I've learned: being very good at judging what is a winner-takes-all market and what is not. For example, in the early days of my entry into Azure, Amazon had a very significant lead in the cloud services space. Many people and investors came to me, thinking the game was over and I would never succeed. Amazon was the winner-takes-all. But in competing with Oracle, IBM, and client-server, I realized that buyers would not tolerate a winner-takes-all. Structural hyperscale will never become winner-takes-all because buyers are smart. The consumer market may sometimes have winner-takes-all, but in the enterprise market, the buyers are companies, enterprises, and IT departments that want multiple vendors. Therefore, you now have to be one of the multiple vendors. I believe this will also happen even in terms of models. Therefore, there will be open-source models, and there will be regulators similar to those on Windows. Nadella 08:39 An important lesson I've learned is that if you have a closed-source operating system, it will have a complement, which is open-source. To some extent, this is a real test of what is happening. Therefore, I believe there will be a dimension in models that may be closed-source, but when open-source models actually ensure that closed-source winners are mitigated, there will definitely be an open-source alternative. Nadella 09:09 At least this is my view on models. By the way, if this technology is really as powerful as people imagine, let's not underestimate it. Countries will not sit back and wait for private companies to develop it. Therefore, I don't think this is a winner-takes-all issue. I think it will be the old-fashioned way. In certain categories, consumers may experience winner-takes-all network effects. For example, ChatGPT is a great example. It is a scaled consumer product that has achieved real escape velocity. I go to the App Store and see it always ranking in the top five, which is incredible. Therefore, they are able to leverage early advantages and turn them into application advantages. This phenomenon may occur in both consumer and enterprise markets, but I would think they will be different categories of winners. At least this is how I analyze it. Dwarkesh Patel 10:11 I have many follow-up questions. We must reach quantum in a second. But regarding the idea that models might be commercialized, similar arguments were made decades ago about cloud services, which are essentially like a chip and a box. But in the end, you and many others found that you had amazing profit margins in the cloud and discovered ways to achieve economies of scale and add other value. Fundamentally, even forgetting the jargon, if you have AGI (Artificial General Intelligence), it's like helping you create better AI, now with synthetic data and reinforcement learning. Perhaps in the future, this will be an automated AI research. Of course, this seems like a good way to consolidate your advantage there. I would love to know your thoughts on this issue. Nadella 10:57 Scale. Nothing is commoditized, right? Regarding cloud services, I mean, everyone says the cloud is a commodity, but when you scale, that's why you need to know how to run a super scaler. You can simply say, I'm just racks and stacked servers, but it turns out that even super-scaled businesses have real business value. Just because of the knowledge of running it, in the case of Azure, the computing across more than 60 regions globally and all that computing is something hard to replicate. Nadella 11:25 So what I emphasize more is, is it a winner, right? Or is it winner-takes-all or not? Because you want to do it right, because you want to win in your category. I like to get into those large camp categories where you don't even have to risk being the winner. If you think the best news is in a large market that can accommodate several winners, and you are one of them, that's the real victory. Nadella 12:06 So that's what I mean, I mean super scale, at model layer 1, the model ultimately needs to run on some super-scaled computing. This will become a long-term demand because the model needs state. That means it needs storage and requires periodic computation to run these agents in an agent environment. Therefore, I think it's unlikely that one person runs away with a model and builds all the models. Dwarkesh Patel 12:43 In terms of super scalar, by the way, your advantage as a super scaler is, especially in reasoning time scaling, if this involves training future models, you can amortize your data center and GPUs not only for training but also reuse them for reasoning. Dwarkesh Patel 13:02 I would love to know what you think Microsoft's and Azure's role in super scalar is. Is it pre-transaction? Does it provide O3 type reasoning? Or are you just hosting and deploying any single model on the market, and you take an agnostic stance on it? Nadella 13:16 Next, I mean, have we built at least the way we want to build the fleet, just like we did in the past? This is not something you have to keep updating every year. Regardless of the lifecycle value of these things, you will depreciate it and then do very well in the fleet position so that you can run different workloads at high utilization Sometimes they are very large training jobs that require a highly concentrated peak failure, while also needing cohesion. That's good. So we should have enough data center footprint to achieve this goal. Nadella 14:14 Ultimately, these become so large that even if you say to maintain the peak, for example, taking the pre-training scale. If it needs to continue, even the pre-training scale must at some point cross the boundaries of data centers. Once you start crossing the boundaries of pre-training data centers, is it different from anything else? My thought is that distributed computing will remain distributed. So, to scale up. Build your fleet and make it ready for large training jobs, ready for test time computation, ready to go. In fact, if this reinforcement learning thing happens, you build a large model, and then after that, reinforcement learning continues and tests me. It's a bit like, again, more training failures because you want to create these highly specialized distilled models for different tasks. So you need that fleet, and then there's the service demand. At the end of the day, the speed of light is the speed of light. You can't have a data center in Texas and then say, I'm going to serve the world from there. You have to base it on having a reasoning fleet around the world to serve the world. So that's my thought on building a truly hyperscale fleet. Nadella 15:29 Oh, by the way, I want my storage and compute to be close to all of this because it's not just that AI accelerators are stateless. I need to be able to store not just my training data itself. Then I want to be able to reuse multiple training jobs. I want to have memory. I want to have environments where these agents can execute programs. That's my thought. World Economic Growth 10% Dwarkesh Patel 16:02 You recently reported that Microsoft earns $13 billion annually from its AI business. But if you look at the growth trend over the past four years, if this trend continues, you will earn $130 billion from the AI business. If so, how do you expect us to leverage these intelligent technologies for industrial-scale deployment? Will it be deployed by Microsoft like traditional office software and hosted by others? Do you think AGI (Artificial General Intelligence) will be the key driver of revenue growth? What would that scenario look like? Nadella 16:30 Yes, that's a great question. If we are to achieve this explosive growth and make intelligent technologies abundant and accessible, the first thing we need to observe is GDP growth. Before I discuss Microsoft's revenue, I think this is the starting point for everything. We cannot be misled by the hype of AGI while ignoring the actual economic impact. For example, the economic growth rate in developed countries is 2%, and if you consider inflation, the real growth rate is close to zero. In 2025, when we sit here discussing, I'm not an economist, but I think we are facing real growth challenges Therefore, the first thing we all need to do is to drive economic growth, just like the Industrial Revolution. For me, this means that the inflation-adjusted growth rate of developed countries should reach 5% or higher. This is the real goal. We cannot just stay on the supply side; we must truly understand how to translate these technologies into actual value for customers. I believe that the ultimate winners will not be the tech companies, but the industries that can widely leverage these technologies. When productivity increases and the economy grows at a faster pace, we in the tech industry will also benefit. But that is where our responsibility lies. This requires some milestones in AGI, rather than meaningless benchmarks. For me, the real benchmark is whether the world can grow at a rate of 10%. Dwarkesh Patel 18:24 If the total world economy is $100 trillion, a 10% growth means creating an additional $10 trillion in value each year. If that’s the case, for a mega company like Microsoft, $80 billion doesn’t seem like a large number. Shouldn’t it reach $800 billion? If in a few years we can really drive world economic growth at such a pace, then is the key bottleneck whether you have enough computing power to deploy these AIs to accomplish all this work? Nadella 18:53 That’s correct. But at the same time, I think there’s a balance that needs to be struck. Like classic supply-side thinking, we can build the infrastructure first and then wait for demand to come. We have already taken enough risks to do this. But at some point, supply and demand must match. That’s why I focus on both aspects. If you only focus on the hype on the supply side and ignore how to translate it into actual value for customers, you may completely go off track. That’s also why I pay attention to Microsoft’s AI business revenue and even disclose these revenue figures. Interestingly, few people talk about their AI business revenue, but for me, it’s important. It helps me think about how to convert yesterday’s capital into today’s demand, knowing that we won’t fully match supply and demand, we can still proceed with exponential investments. Dwarkesh Patel 20:07 I wonder if there’s a contradiction between these two viewpoints. Because your investment in OpenAI in 2019, even before any practical applications emerged, was an early bet. If you look back at the Industrial Revolution, the construction rate of railroads and other infrastructures was about 6.10%, many of which did not immediately generate revenue and might even have incurred losses. But if you really believe these technologies have the potential to increase the world growth rate by 10 times or 5 times, then you would think about where the GDP revenue would go. If you truly believe this is the next level of possibility, shouldn’t you be bolder and make hundreds of billions of dollars in computing investments? After all, there are indeed opportunities here. Nadella 20:52 You’re right. That’s the interesting thing. Frankly, the real question is, at least for us, why it’s very important to take a balanced approach to infrastructure. It’s not just about building computing power, but about building the capability that can help us train the next big model and serve the next reasoning model Before you accomplish these two things, you can't really leverage your investments. Therefore, this is not just a competition to build models, but a competition to create a commodity that can be used by the world to drive economic growth. You must have a complete mindset, not just a single perspective. By the way, I think one of the issues is that it may be overbuilt. Your point also illustrates what has happened in the internet age. What I see is that now that the memorandum has been released, everyone needs more energy and computing power. Thank God we are prepared. In fact, I am not only concerned about the number of corporate deployments; countries will also invest capital. I am glad that we will lease a large capacity in 2027 and 2028 because when I look at the bills, I realize that the only possible change in all computing bills is the price increase. Declining AI Costs Dwarkesh Patel 22:23 Regarding the issue of price decline, you mentioned the Jevons Paradox after the release of the Deep Seek model, and I would love to hear your further explanation. Gen's products appear at a time of highly elastic demand; does intelligence become a bottleneck for price decline? At least from my perspective as a consumer, intelligence seems quite cheap—only 2 cents per million tokens. Do I really need to bring it down to zero? If I need to train 100 times larger and pay 100 times the cost for that, I am willing to let the company bear it. But perhaps you see a different situation in the enterprise sector. So, what are the key use cases for intelligence that truly need to bring the cost per million tokens down to 0.02 cents? Nadella 23:08 I think the real issue lies in the utility of tokens. In a sense, both need to happen: intelligence needs to become better and cheaper. Whenever there is a technological breakthrough, such as in deep learning or other fields, the effective frontier changes, the performance curve bends, and the frontier moves. This will bring more demand. Nadella 23:35 This is the evolution of cloud computing. We once thought that all servers in the client-server era were sufficient, but once we started putting servers in the cloud, people began to consume more because cloud services are cheaper and more elastic, allowing users to purchase on demand rather than buying licenses. This completely changed the market. I remember in India, the sales of SQL servers were limited, but the cloud market in India is much larger than what we could achieve during the server era. I believe this situation will continue to occur. Nadella 24:16 If these tokens can be used for healthcare in the Global South or developing countries at a very low price, it would be a tremendous transformation. Dwarkesh Patel 24:32 I think it's reasonable to hear people like me in San Francisco and think they are a bit unrealistic. They don't know how to deploy things in the real world. As someone who works with Fortune 5 companies and deploys projects for hundreds of millions of people, how fast do you think the deployment of these functions will be, even with work agents, even with capabilities that can work remotely and solve all compliance and inherent bottlenecks? Will this be a major bottleneck or will it pass quickly? Nadella 26:08 This is a real challenge because the core issue is change management or process transformation. How did multinational companies predict the emergence of pre-PC, email, and spreadsheets? Fax used to be the primary means of communication, and then people started using email and spreadsheets, which changed the entire forecasting business process as work artifacts and workflows changed. This is what needs to happen to bring AI into knowledge work. Nadella 27:14 When we consider all these agents, the most fundamental thing is to have a new job and workflow. For example, when I was preparing for this podcast, I found my Copilot and said, "I want to talk to Rakesh about our quantum announcement and the new model for game generation. Can you give me a summary of everything I need to read before I go?" It can do that and even present it in podcast form. Then I shared this with my team. For me, the new workflow is thinking with AI and working with my colleagues. This is a fundamental transformation that requires everyone engaged in knowledge work to rethink how they do their jobs in new ways. This will be similar to the transformations in sales, finance, and supply chain. For existing businesses, this will be a challenge, just like the transformation in manufacturing regarding lean production. Nadella 28:36 I love this analogy because lean production became a methodology for how to take an end-to-end process in manufacturing and become more efficient. This is continuous improvement, reducing waste to increase value. This is the source of knowledge work. It's like lean knowledge work, which requires hard work from management teams and individuals, and it takes time. Dwarkesh Patel 29:00 Regarding this analogy, I want to ask: The physical transformation, like Industry 4.0, reveals bottlenecks that people were not aware of before truly focusing on processes and workflows. You mentioned your own workflows and how they have changed because of AI. I'm curious, as these AI agents become increasingly intelligent, how do we add more color to running a large company? Nadella 29:29 That's a very interesting question. I've been thinking, for example, our emails today are very burdensome. I come in the morning, and my inbox is already full, and I need to reply. I can't wait for these Copilot agents to auto-fill my drafts so I can start reviewing and sending. Nadella 29:51 But in reality, I feel like I already have at least 10 agents in Copilot that I use for queries on different tasks. I think there will be a new inbox created where these agents will have to report some exceptions to me, notify me, or request instructions. So I'm thinking that in the future, there will be a new framework where agent managers will be at the core of this framework. Nadella 30:24 This is not just a chat interface. I need something smarter than a chat interface to manage all the agents and their conversations. That's why I think the co-pilot is the user interface (UI) of AI, and it's a big deal. Each of us will have it. You can think of it as a combination of knowledge work and knowledge workers: knowledge work may be done by many agents, but you still need knowledge workers to deal with all these agents. I think this is a key interface for the future. Microsoft's Quantum Breakthrough Dwarkesh Patel 31:01 Yes, I'm a bit curious, as one of the few leaders in the world who can reach 200,000 people, you have the wisdom of companies like Microsoft and all its employees around you. You have to manage it, understand how to interface with it, and how to make the most of it. Hopefully, more people will have such experiences in the future. Dwarkesh Patel 31:24 I'm curious what your inbox would look like if that means everyone's morning sales inbox would be like yours. But before we start, I want to continue asking you more about AI. However, I really want to ask you about the significant breakthrough announced by Microsoft Research in the field of quantum. Can you explain what happened here? Nadella 31:39 This is another 30-year journey for us, it's incredible. I am the third CEO of Microsoft to be excited about quantum. I think the fundamental breakthrough here, or the vision we've always had, is that you need a physical breakthrough to build a practical-scale quantum computer. So we chose this path, which is a bit like saying that the only way to have less noise or more reliable qubits is to bet on a physical property that is, by definition, more reliable. That's why we chose to target topological qubits, which were theorized in the 1930s. The question is, can we really manufacture these things physically? The biggest breakthrough is that we now finally have proof of existence—the Majorana zero mode in a new phase of matter's physical breakthrough. So we liken it to the transistor matrix of quantum computing, and we can effectively have a new phase, the topological phase, which is more dependent on direction. This means we can now even reliably hide quantum information and measure it, and then we can manufacture it. Now we have the core foundational manufacturing technology to start building a mile on a chip. That Myrana 1 will basically be the first chip capable of physically reaching one million qubits and then error-correcting on thousands of logical qubits. That's emerging. So, you can't suddenly have the capability to build a truly practical-scale quantum computer. It feels more feasible to me now. We've been working on this because without such a thing, you can still reach milestones, but you can never build a practical-scale computer. That's why we're excited about it. Dwarkesh Patel 34:23 Amazing. By the way, I believe this is it. Nadella 34:25 Yes, yes. I forgot it now. Did we call it Myorana? Yes, that's right. Myrana 1. I'm glad we named it that. It's incredible to think that we can build something like a million-qubit computer on such a large scale. That's the key; unless we can do that, you can't dream of building a practical-scale quantum computer. Dwarkesh Patel 34:55 You mean the eventual million qubits will fit on a chip size. That's fine. Amazing. So, other companies have announced a hundred physical qubits, like Google, IBM, etc. When you say you've announced one, but you say your limitations are more scalable. Nadella 35:13 Yes, by the way, we also did one thing, which is that we separated software and hardware. So, we are building our software stack. In fact, we now have some different approaches, including neutral atoms, ion traps, and we are also collaborating with others who have good methods in photonics and so on. This means there will be different types of quantum computers. In fact, we have 20, and I think we finally announced 24 logical qubits. So, we have also made some amazing breakthroughs in error correction, which allows us to build these 20-plus qubits even on neutral atoms and ion traps. I think this will continue throughout the year. You will see us making progress, proving that standard. But we also said, let’s follow the first principles and build our own super quantum computer, betting on topological qubits. That’s what this breakthrough is about. Dwarkesh Patel 36:21 Amazing. A million topological qubits, tens of thousands of logical qubits. What’s the timeline expected to reach that level? What’s the “quantum Moore's Law” here? If you have the first transistor, it looks like…? Nadella 36:34 Obviously, we have been working on this for 30 years. I’m glad we now have manufacturing, physical breakthroughs, and manufacturing decomposition. I hope we have a quantum computer because the first thing a quantum computer allows us to do is to build a quantum computer, as simulating qubits one by one to build these new quantum gates will be much easier. But anyway, for me, the next real thing is now that we have the manufacturing technology, let’s go build the first fault-tolerant quantum computer. That would be the logical thing to do. So, I would say maybe in 2027, 2028, or 2029, we will be able to really build this. So now we have this gate, I can put this thing into an integrated circuit, and then put these integrated circuits into a real computer. I think that’s the next logical step Dwarkesh Patel 37:31 What do you see in 2027 or 2028? You made it work. Is it something you got through an API? Is it something you use internally for research, materials, and chemistry? Nadella 37:43 One thing I've always been excited about is that even in today's world, because we have this quantum program, we have HR, we can say, hey, this is a... you know, some API. Maybe the breakthrough we made two years ago was combining the HPC (High-Performance Computing) stack and the AI stack with quantum. If you think about it, AI is like a simulator's simulator, and quantum is like nature's simulator. What does quantum want to do? By the way, quantum will not replace classical, right? Quantum is good at doing things that quantum can do, and classical computing will continue to exist because you can't... quantum will be great for anything that is not data-intensive, but it has more exploration in the state space, right? So it should be data-light, but you want to explore the exponential state space. Simulation is a great field, and so is chemistry, physics, and biology. So one thing we have started doing is really using AI as a simulation engine, but you can train it. So my thought is, if you have AI plus quantum, maybe you would use quantum to generate synthetic data, which then gets used by AI to train better models that know how to simulate things like chemistry or physics. These two things will work together. So even today, this is our combination of HPC and AI. And I hope to replace some fragments of HPC with quantum. Dwarkesh Patel 40:26 Okay, back to Satya. Can you tell me a bit about how you make these research decisions? These decisions will really pay off in 20 or 30 years, especially in a company the size of Microsoft. Clearly, you are very familiar with the technical details of this project. Can you do this with everything that Microsoft Research does? What do you think the current metrics will yield in 20 years? How do you decide to emerge organically through the organization? Or how do you keep track of all this? Nadella 40:56 Yes, I think what's great is the vision Bill had when he founded Microsoft Research (MSR) in 1995. Yes, that's right. Nadella 41:09 I think in the long history of these curiosity-driven research organizations, MSR has built this institutional strength over the years. So when I think about capital allocation, budgets, or anything else, you first put the chips in, and then say, hey, look, this is MSR's budget, we have to do this every year, knowing that some... you know, most of these bets will not pay off in any limited timeframe. This may be something that benefits Microsoft's sixth CEO. I think, you know, this is a technology, I think, is taken for granted. What I really want to know is, when something like quantum or new models comes along, what do you have, what can you leverage? So as a current holder, if you look at the history of technology, no one would not invest Nadella 42:13 It's like you need to have a culture that knows how to embrace innovation and scale it. That's the hard part, and frankly, it's a bit fascinating for CEOs and management teams, right? I mean, it's equally about good judgment and good culture. You know, sometimes you get it right, and sometimes you get it wrong, right? I mean, I can tell you a thousand projects from MSR that, you know, we probably should have led, but we didn't. I always ask myself why, because we couldn't get enough conviction or a complete idea of how to not only turn innovation into useful products but also have a business model. Microsoft's Game World Model Dwarkesh Patel 43:35 Let's dive into another major breakthrough you've just achieved. It's surprising that these breakthroughs were announced on the same day, including your game world model. I hope you can tell me more about this. Nadella 43:43 We call it "Muse" because they will become models of human behavior or world actions. It's very cool. You see, obviously Dolly and Sora are incredible in generative models. One thing we wanted to do is leverage game data to see if we can generate content that is both consistent and reflects the diversity of the game, and can sustainably generate user mods. So, we partnered with one of the game studios, which was a natural collaboration. The exciting thing is that we will soon have a game library that starts using these models, or really trains these models to generate content and then starts playing the games. In fact, when Phil Spencer first showed it to me, he had an Xbox controller, and this model basically takes input and generates output based on that input, keeping in line with the game. For me, that was a huge moment, just like when we first saw ChatGPT generate complete sentences, or Dolly generate paintings, or Sora. This is a very significant moment. Dwarkesh Patel 45:15 I had the opportunity this morning to watch videos of these models in a live demonstration with your chief researcher, Katie. It wasn't until I spoke with her that I truly realized how incredible this is. In the past, we used AI to shape agents, but now we are using the same technology to simulate the world around the agents and achieve a consistent real-time experience. We will overlay the video in the podcast, giving people a chance to see it for themselves. I guess it will go live at the time of release, so the audience can watch as well. Dwarkesh Patel 45:43 That's incredible in itself. Through your strategic investments, Microsoft has poured billions of dollars into the gaming sector and acquired a significant amount of IP. Looking back, if you could integrate all this data into a large model to provide users with the experience of accessing and browsing multiple game worlds simultaneously, it seems like a significant direction for the future of gaming. If this is the future of gaming, then our past investments seem very wise. Do you have any premonitions or coincidences about this? Nadella 46:16 No, I wouldn't say we invest in games to build the models we invest in. Frankly, we have an interesting history: we developed the first game before we developed Windows. "Flight Simulator" was already a Microsoft product long before we built Windows. So, games have a long history in the company, and we engage in gaming for the sake of gaming itself. I have always opposed using some means as a tool for another purpose in business; we must be our own purpose. We are not a corporate group; we are a company that must integrate all these assets together to become better owners by adding value. For example, cloud gaming is a natural choice for our investment because it will expand the accessibility of games, allowing people to play anywhere. The same goes for AI and gaming. We certainly believe this could help change the gaming industry; it's like the CGI moment for long-term gaming. As the world's largest game publisher, this will be very helpful. But at the same time, we must produce high-quality games. Without first focusing on that, you cannot become a game publisher. However, this data asset will be interesting, not only in the gaming environment but it will also become a universal action model and world model. That's fantastic. Nadella 47:40 I think game data could be Microsoft's "oil." So, I'm excited about that. Dwarkesh Patel 47:52 Yes, I mean, you can have a unified experience across many different types of games. How is this separate from what Microsoft has done in the past, like mixed reality? Maybe this gives smaller game studios a chance to build these AAA action games in the next 5 to 10 years. Nadella 48:13 I see these three things as cornerstones in an interesting way. Even 5, 6, 7 years ago, I said that the three big bets we wanted to make were AI, quantum, and mixed reality. I still believe in them because they all address some significant issues. The dream of mixed reality is, can you create a real sense of presence like we do with podcasts? Nadella 48:45 I think we still believe, honestly, that this is one of the harder challenges. I thought it would be easier to solve, but it might be harder because of its social aspects, like wearables, etc. We are excited about how Adriel and Palmer will now advance the IVAS program because it's a great use case. So we will continue to work on that. And there’s the 2D surface, like the team, right? During the pandemic, we really developed the ability to create a basic sense of presence through 2D, and I think that capability will continue to evolve. This is a secular part. The quantum and AI we talk about are another. Nadella 49:32 So these are the three things I observe and think about: how do you ultimately bring these things together, not for technology's sake but to solve some fundamental issues we want in life as humans, as well as more issues we want in the economy, driving our productivity If we can do this in some way, then I think we will really make progress. Dwarkesh Patel 49:54 If you write a book, you have to have some explanation for why these three parts are happening at the same time, right? It's like there's no entry point. You would think that quantum and AI should happen in 2028 and 2025, etc. Nadella 50:04 That's right. But to some extent, I think the simple model I have is: Is there a systemic breakthrough? For me, that systemic breakthrough is the quantum stuff. Is there a business logic breakthrough? That feels a bit like AI to me, like can I fundamentally change the logic instead of forcing code to be written, can you have a learning system? That's one aspect of AI. Then there's the UI aspect, which is presence. Dwarkesh Patel 50:36 Back to AI. In your 2017 book "2019," you invested in OpenAI quite early, even before 2017. You mentioned in the book that some might say we are nurturing a new species, a species whose intelligence may have no limits. Now, of course, it was too early to talk about this in 2017. So far, we've been discussing agents, office co-pilots, capital expenditures, and so on in a granular way. But if you just amplify and consider that statement you made, imagining yourself as a super scalar person, as someone researching in these models, training and researching to establish a new species. From a big picture perspective, what are your thoughts on this? Do you think we are moving towards superhuman intelligence? Nadella 51:22 Safa recently used the term "new species" to describe it. My view is that you absolutely need trust. Before we claim it is a species, I think the fundamental issue we must get right is establishing real trust, whether at the individual level or the societal level. This is a conundrum because I think the biggest limiting factor here will be the infrastructure, how our laws refer to it. We are talking about all the computational infrastructure, how the legal infrastructure evolves to address this issue? The whole world is built on humans owning property, having rights, and bearing responsibilities, etc. This is a fundamental thing to first say, what does this mean for anything that humans currently use as a tool? If humans are going to delegate more power to these things, how will that structure evolve? Until that question is truly resolved, I think merely talking about technological capability is not enough. Dwarkesh Patel 52:51 It's like we won't be able to deploy this intelligence until we ultimately figure out how to deploy it. Nadella 52:56 Yes, there is no way to deploy this intelligence as it is today unless someone compensates for it as a human. I think that's also one of the reasons why I believe that even the most powerful AI is essentially using some form of authorization, coming from some human. Yes, you can talk about alignment issues, but that's exactly why I think you have to make that alignment really work and be verifiable in some way But I just think you can't deploy outdated intelligence. For example, the AI takeoff issue might be a real problem, but before it becomes a real problem, the real issue will be in court. Because the court won't allow certain people to say, "Hey, I did it." Dwarkesh Patel 53:44 Yes, there are many societies in the world, and I wonder if there is any society that might not have a more acceptable legal system. If you can't take off there, then you might worry that it doesn't have to happen in the U.S., right? Even if you are legal. Nadella 53:57 Yes, but even in anything, I think we can't assume that there is no society that would care about it. There might be rogue actors. I'm not saying there won't be rogue actors. What I mean is, there are cybercriminals and rogue nations that will be there. But to think that the entire human society doesn't care about it is also not true. So I think we all will care. We know how to deal with rogue actors and rogue nations today. The world won't sit idly by. That's why I'm glad we have a world order in which, even so, any rogue actor from a rogue nation has consequences. Dwarkesh Patel 54:42 If you have this picture, you can have 10% economic growth, and I think it depends on whether something like AGI is effective, right? Because trillions of dollars in value sound more like human wages are $60 trillion. The economy will get this. Scale is like you have to automate labor or supplement labor in very significant ways, if possible. Once you figure out its legal consequences, even within your term, we are very likely to figure it out. You're thinking about the deployment of superintelligence. Ensuring AGI Safety Dwarkesh Patel 56:29 Regarding the deployment and alignment issues, two years ago you released Sydney Bing. Frankly, given the technological capabilities at the time, I thought it was a bit of a charming, lovable but amusing misalignment example. At that time, some chatbots might give some interesting but inappropriate responses in 30 seconds. But if you think about the future, these agents might run for hours, weeks, or even months, like a group of autonomous AGI (Artificial General Intelligence), which might misalign in similar ways, mess things up, and even coordinate with each other. So what are your next plans? For example, how will you handle it when you get a powerful model? Nadella 57:25 Yes, you're right. That's also why we believe that even when allocating computing resources, we need to allocate computing resources for alignment challenges. More importantly, we need a mechanism that can monitor the runtime environment of these systems to ensure their observability. By the way, just like we deal with many issues in the classic computing field, such as cybersecurity. We don't just write software and then let it go. We continuously monitor the software, guard against cyber attacks, conduct fault injections, etc. Therefore, I think we must build sufficient software engineering capabilities around these deployments What is alignment within the model itself? These are both real scientific questions and real engineering questions. We must address these issues. By the way, this also means we need to take responsibility for ourselves. Nadella 58:27 That's also why I'm more interested in deploying these things, because you can actually manage the scope and scale of these things. You can't release something into the world that could cause harm because there is no societal permission for that. Dwarkesh Patel 58:45 Yes, what do you do when you actually have agents that can complete tasks for you over weeks? What kind of minimum guarantees do you want to ensure that comes from my random wealth? Nadella 58:58 When I use something like deep research, I think the minimum guarantee we want is before we have a physical embodiment of AI that we specifically own. I think that's one of the thresholds, and when you cross it, there might be another, for example, the permissions of the runtime environment. You might need to guarantee that it is sandboxed and that it doesn't go in and out of the sandbox. Dwarkesh Patel 59:32 I mean, we already have web search, you know, we have sandboxes now, right? Nadella 59:37 But even with the web, it has implications for web search and what it writes. So, for example, like your point, if it's just going to write code to kick off some computation, then where is that code deployed, and that code is ephemeral, only used to create that output, rather than just popping that code out into the world. These are all things you can actually do in the action space. Dwarkesh Patel 1:00:05 Aside from security issues, when you think about your own product suite, you might wonder, if you really had such powerful AI at some point, it wouldn't just be like the co-pilot example you mentioned about how you prepare for this podcast. Rather, it would be more like how you actually delegate work or tasks to colleagues. Considering your current suite, what would it look like to add it? I mean, you know, there's a question of whether LLMs (large language models) will become a commonality for other things. I wonder if they are like databases or canvases or Excel spreadsheets or something else, and if LLMs are the main entry point to access all these things, can LLMs modify Office? Nadella 1:00:46 That's an interesting question. I think my view on the first stage is whether LLMs can help me use all these tools or canvases more effectively to accomplish my knowledge work. One of the best demonstrations I've seen is a doctor preparing for a tumor board workflow. One of the first things she used the co-pilot for was to create an agenda for the meeting because the LLM helped reason through all the cases on certain SharePoint sites and said, "Hey, these cases are obviously... you know, tumor board meetings are high-stakes meetings, and you want to focus on the differences in cases so you can assign the right types." So even the reasoning task of creating an agenda, even knowing how to break down super types, I used the LLM to do that Nadella 1:01:47 Then I entered the meeting. I was on a team call with all my colleagues. Guess what? I focused on the actual cases instead of taking notes. Because now you have this AI co-pilot doing the full transcription of all this, basically an intelligence that is not just a transcription, but a database entry of all types that I can recall from the meeting. Nadella 1:02:11 Then she came out of the meeting, discussed the case without being distracted by notes. She is a teaching physician, and she wanted to prepare for her course. So she started the co-pilot and said, “Hey, attend my tumor board meeting and then create a PowerPoint slide from it so I can talk about it with my students.” So my UI and scaffolding are now filled with LLMs. The workflow itself is being reshaped, and knowledge work is being accomplished. Nadella 1:02:50 This is an interesting thing. If someone had come to me in the late '80s and said, “You will have a million documents on your desk.” We would have said, “What is that about?” I mean, I really would have thought, oh, I would have a million physical copies on my desk. Except we do have a million spreadsheets and a million documents, I know you have them, they are all there. So I think even the agents will have this happen. So there will be a UI layer. Nadella 1:03:21 For me, Office is not just today’s Office. It is the UI layer of knowledge work. It will evolve as workflows evolve. That’s what we want to build. I do believe that these SaaS applications that exist today—these rough applications will fundamentally change because business logic will move more into this agent layer. In fact, in my co-pilot experience, another cool thing today is when I say, “Hey, I’m preparing for a meeting with a client, give me all the notes I should know.” It pulls from my CRM database, it pulls from my Microsoft Graph, creating a composite, essentially an artifact. This means that it even applies logic. For me, this will largely change what we know today as SaaS applications. Dwarkesh Patel 1:04:10 So SaaS as an industry could be worth hundreds of billions to trillions of dollars each year, depending on how you calculate it. If this is really going to be disrupted by AI, then this will be a significant step for Microsoft in the next decade, and Microsoft’s market value could rise again. Because, you know, if you are really talking about trillions of dollars... Nadella 1:04:30 I think this will also create a lot of value. Certainly in the SaaS space. Remember one of the big issues is how much backlog there is in the world, right? So one way to address this is through these code-generating things, plus I can use agents to query all my SaaS applications and gain more utility, which will be the biggest explosion of applications. They will be called agents. This way, you can, in every industry or every category of every vertical, suddenly have the capability to get services So there will be a lot of value. I think you can't just stand still, like, you can't say the old saying: "Oh, I've outlined some narrow businesses from the process, I have a UI in the browser, that's my thing." It won't be like that. Nadella 1:05:25 You have to stack up and say, "What are the tasks I need to be involved in?" Therefore, you will be able to use your SaaS applications and make them excellent agents for participating in a multi-agent world. As long as you can do that, I think you can even add value. Dwarkesh Patel 01:05:43 Can I ask you some questions about your time at Microsoft? Is being a company person underrated? Most of your career has been spent at Microsoft. It could be said that this is one of the reasons you are able to add so much value, because you have seen the company's culture, history, and technology, and gained all this background through promotions. Should more companies be run by people with this background? Nadella 01:06:04 Good question. I haven't thought about it from that angle, but indeed, I have worked at Microsoft for 34 years, and basically every year I am more excited about my work at Microsoft, rather than thinking, "I am a company person." It's not that I joined Microsoft to use it as a platform for economic returns, but there is a sense of mission and purpose to achieve through the platform of Microsoft. It's a kind of covenant. I think companies must create a culture that allows people like me, "company people," to integrate. Microsoft does this well, at least in my case, and I hope it continues. Dwarkesh Patel 01:07:06 How do you think the sixth CEO you are talking about will leverage the research you are starting now? What are you doing to reserve space for future successors so they can be part of the future? Nadella 01:07:16 Yes, this year marks the 50th anniversary of Microsoft, which has made me think a lot. I believe longevity itself is not the goal; the key is relevance. So, what I and all our 200,000 employees do every day is to ask whether we are doing things that are useful and relevant to the developments we see in the world, not just today, but tomorrow. We must realize that we live in an industry without franchise value. So, we must approach our work with this attitude: we are doing what we believe is relevant. This is also why we must focus on the future while having a high tolerance for failure. Dwarkesh Patel 01:08:36 You just mentioned that there are two months left until Microsoft's 50th anniversary. If you look at the top ten companies by market capitalization or the top five, basically all the other companies besides Microsoft are younger than Microsoft. This is a very interesting observation about why the most successful companies are often very young. You know, the average lifespan of a Fortune 500 company is 10 to 15 years. What has Microsoft done to maintain relevance for so many years? How do you continue to "give back"? Nadella 01:09:09 I like the word "refund." I think that's the mindset. People talk about the founder model, but for me, it's more like a "refund model." The ability to look at things in a new way is key for me. So, to your question, can we culturally create an environment where "refund" becomes a habit? Every day we come in and say, yes, we feel we have a stake in this place, to change the core assumptions of what we do, how we connect with the world around us, and what value we provide. Many times, companies feel overly constrained by business models or other factors, and you just need to loosen up. Dwarkesh Patel 01:10:08 If you left Microsoft, what company would you start? Nadella 01:10:12 If I left Microsoft, I would start a company focused on underserved areas. I never leave Microsoft, but if you made me think about it, I would choose a field like healthcare, education, or public services. These are areas where services are lacking in society, and if all this technology could translate into better healthcare, better education, and better public services, I would be better off. That would be a worthwhile area to invest in. Does Nadella believe in AGI? Dwarkesh Patel 01:11:31 I'm not sure after hearing your answers to different questions whether you think AGI (Artificial General Intelligence) is something meaningful, whether there will be something that can automate all cognitive labor, at least starting from anything humans can do on a computer? Nadella 01:11:46 I think there's a problem with how people talk about the definition of cognitive labor because cognitive labor is not static. Today's cognitive labor may be automatable, but what is the new cognitive labor? Both of these things need to be considered. That's why I think we shouldn't conflate knowledge workers with knowledge work. Today's knowledge work may be automatable, but what will the new cognitive labor be? Dwarkesh Patel 01:12:50 But what will AI get? Nadella 01:12:52 Once you get to the second thing, there will be a third thing. That's why I think we have tools that have historically changed cognitive labor. Why are we worried that all cognitive labor will disappear? I mean, humans have only had a 200-year history, and we value something narrow called cognitive labor. But if things like quantum and AI really help us do a lot of new materials science, will that really undermine other things humans can do? Why can't we live in a world with powerful cognitive machines while knowing our cognitive abilities are not being stripped away? Dwarkesh Patel 01:13:13 I believe you've heard these examples before, but just like with horses, there's still the idea that they are beneficial for certain things. There are terrains you can't drive on, but like horses, you see them on the streets, and they employ millions of horses, as if nothing has happened So, will similar things happen to humans? Satya Nadella 01:13:28 In a very narrow dimension, humans have only a 200-year history, and we value some narrow things called cognitive labor. But if things like quantum and AI really help us create many new materials science, will that really undermine other things humans can do? Why can't we live in a world with powerful cognitive machines while knowing our cognitive abilities are not being taken away? Dwarkesh Patel 01:14:14 I won't ask you this question, but in a different context, so maybe you can answer this question without any awkwardness. Suppose at the Microsoft board meeting, could you see an AI being added to the board? Does it have similar judgment, background, and overall understanding to be a useful advisor? Satya Nadella 01:14:32 That's actually a great example. One of the things we are adding is a facilitator agent in the team. The goal there, still in the early stages, is for it to be a facilitator with long-term memory, not just in the context of meetings, but in the context of the projects and teams I am working on, to become a great facilitator. I even hope that in board meetings, it can easily keep us on track. After all, board members come once a quarter, and they try to digest what is happening in a complex company like Microsoft. I think a facilitator agent that truly helps humans stay on topic and focus on important issues would be great. It's a bit like literally having something with infinite memory that can even help us cope with human bounded rationality. Dwarkesh Patel 01:15:46 Speaking of materials and chemistry, I remember you recently mentioned that you hope advancements in these fields happen in the next 250 years. Now when I think about what might happen in the next 250 years, I think of space travel, space elevators, immortality, and curing all diseases. What do you think? Satya Nadella 01:16:08 I hope one of the reasons I raise this question is that I love the Industrial Revolution. If you change the entire system from a carbon-based system to a different system, it means you have to fundamentally reinvent chemistry or everything that has happened over the last 50 years. This is where I hope we have quantum computers. These quantum computers help us obtain new materials, and then we can manufacture these new materials to help us tackle all the challenges on Earth. Then I am fully supportive of interstellar travel. Dwarkesh Patel 01:16:44 That's amazing. Satya, thank you very much for your time. Satya Nadella 01:16:50 Thank you, it was great