Microsoft CEO Nadella's latest 20,000-word insight: The C-end Agent business model still needs exploration, the advertising traffic model may face changes, and the key for the B-end lies in ecosystem integration

Microsoft CEO Satya Nadella recently discussed Microsoft's strategic transformation, investment in OpenAI, and the future of agents with well-known Silicon Valley investors Brad Gerstner and Bill Gurley, with the full text being 20,000 words.

Nadella believes that the current competition in the AI field will no longer be a winner-takes-all scenario, but rather a fierce competition among multiple companies across different levels of technology.

In particular, he thinks that Microsoft's Azure and other cloud services will play a significant role in the future AI competition, especially between infrastructure (such as cloud computing) and applications (such as AI models).

Agents are seen as more intelligent and personalized tools than traditional search engines, no longer just simple stateless query tools, but capable of maintaining state, remembering user history, and providing continuous interaction.

Satya Nadella mentioned that as consumer demand for AI increases, traditional search engines (such as Bing) face new challenges, and we will witness a shift from traditional search to AI-based question-and-answer systems in the future.

The emergence of agents may break the boundaries of traditional search engines, providing direct answers rather than just links, changing the way users interact. With tools like ChatGPT, users no longer need to make multiple queries to get answers but can receive more immediate and intelligent feedback.

For the consumer side, the business model for agents is still being explored, and traditional advertising and traffic-driven models may need to change, especially as agents can acquire and process data through simplified conversations.

Enterprise-level agent interfaces may become increasingly important, and Microsoft is already leveraging AI to connect with multiple systems, such as Adobe, SAP, and its own CRM (Dynamics). These interfaces can help AI acquire and integrate enterprise data, thereby providing more efficient services.

The following is the full text:

Brad Gerstner

It's great to be with you. When Bill and I were discussing, Satya, reflecting on your tenure as CEO, it has truly been a great learning experience. It has been quite challenging; you joined Microsoft in 1992. For those who may not know, you took over Microsoft's online business in 2007.

You launched Bing search in 2009, took over the server business and launched Azure in 2011, and became CEO in 2014. Just before that, a now-famous article titled "Microsoft's Irrelevance" had just been published.

Since then, you have increased Azure's annual revenue from $1 billion to $66 billion. The total revenue of the entire business has grown 2.5 times, total profits have increased by more than 3 times, and the stock price has nearly risen 10 times. You have added nearly $3 trillion in value for Microsoft's shareholders Looking back over the past decade, what do you think was the biggest change you could make at that time? How did you unlock value, change the direction of Microsoft, and achieve such extraordinary success?

Satya Nadella

Well, I've always thought about it this way, Brad, from 1992 to now, in a sense, it's been a continuous phase for me, although clearly, 2014 was a significant turning point, accompanied by corresponding responsibilities. But I feel that, ultimately, the patterns of success and failure can be matched, doing more of what leads to success and less of what leads to failure.

In a sense, it's that simple because I've experienced it. When I joined Microsoft in 1992, it had just released Windows 3.1; I remember Windows 3.1 was released in May 1992, and I joined in November 1992.

In fact, at that time, I was working at Sun Microsystems, considering going to business school, and then I received an invitation from Microsoft. I initially wanted to go to business school, but ultimately it was because my boss persuaded me to join Microsoft, which became one of the best decisions I've ever made.

The reason that made me decide to join was the PDC conference in 1991 at the Maskoni Center, where I saw Windows NT (it didn't have that name at the time) and x86, and I thought that what was happening on the client side would eventually happen on the server side. This is a platform company, a partner company, and they would ride this wave forward. So that was my consideration at the time.

Then, the emergence of the internet led us through a transformation, and we got a lot of things right. For example, we realized that browsers were our competitors, and ultimately we got the browser thing right.

But we made a mistake in the search area; we thought that browsers were the most important because they were more like an operating system, but we didn't understand the emergence of a new category, which is that the organizational layer of the internet is search.

Then we also got involved in mobile internet but didn't fully grasp it. Clearly, the emergence of the iPhone caused us to miss an opportunity. However, we got it right in the cloud computing space. So, if I look back at these things, we are now also experiencing the fourth transformation of AI.

In these processes, I think the most important thing is not to blindly imitate just because others are doing something. Sometimes, following trends can be okay, and the results can be good, but you shouldn't act out of jealousy. This is one of the hardest lessons we've learned. You do things because you have the authority and can do them better; both of these are very important to me.

Brand empowerment, for example, Jeffrey Morre once said to me, "Why don't you do what customers expect you to do?" I really like that phrase; cloud computing is a great example where customers were actually expecting us to do this.

In fact, when I first encountered Azure, many people told me that cloud computing was a winner-takes-all game, and Amazon had already won. I never believed that because, after all, I had competed in the server space with companies like Oracle and IBM, and I always felt that the infrastructure space could absolutely not be winner-takes-all You just need to enter this field and propose a valuable solution.

In a sense, many of these transformations for me are about ensuring that you can recognize your structural position in the market, truly understand the authority you have in the eyes of those partners and customers who want you to succeed, and first do the obvious things.

I think this could be called the foundation of strategy, but for me, this is key. The culture and sense of mission you mentioned are necessary conditions, even prerequisites for achieving goals. But I believe that it is precisely through identifying your structural position and gaining authorization that you can make the right strategic adjustments.

Bill Gurley

Wait a minute, Satya, before we talk about AI, I have a few questions about the transition. As Brad just mentioned, you might be one of the most successful CEO appointments in history. I mean, a $3 trillion market cap is unparalleled. First, I read an article that mentioned you wrote a 10-page memo for the CEO selection committee. Is that true? If so, what did that memo say?

Satya Nadella

Yes, that is indeed the case. Because I felt that our CEO selection process at the time was very open, and frankly, I wasn't sure I would become the CEO. Remember, at that time, I never thought Bill would leave, let alone Steve.

You can't join Microsoft thinking, "Oh, the founders will retire, there will be a vacancy, and I can apply." I didn't have that mindset at the time. So when Steve decided to retire, remember it was August 2013, that was a huge shock for me.

At that time, I was still in charge of our server and tools business (Azure was part of that), and I actually enjoyed that job, and I didn't proactively propose to become CEO because I didn't have that thought at the time. Then, the board began to consider this issue, and there were many other candidates, including senior executives from Microsoft.

In the end, during the selection process, they asked us to write memos, and in fact, that memo was very interesting. Many of the things I mentioned in it, if I look back now, were very prescient. For example, I used terms like "environmental intelligence" and "ubiquitous computing" in that memo. I also used those words in my first email, although I later simplified it to "mobile-first, cloud-first," because my PR team came to me and said, "These terms are too difficult to understand; no one knows what environmental intelligence or ubiquitous computing is."

So, I used "mobile-first, cloud-first" to express how to seize the big trends, understand Microsoft's structural position, think about Microsoft's cloud computing business, what resources we have, and why M365 is so important.

In fact, I have always resisted the market's way of splitting the cloud computing business. I never allocate my capital as "this is Azure's capital, here is M365's capital, here is the gaming capital." I have always believed that the core of Microsoft is a set of infrastructure on which different workloads sit, one of which is Azure, another is M365, Dynamics, Games, etc.

Overall, a lot of the content is in that memo, and it has actually been realized. At that time, I also anticipated that although we had a gross margin of 98% to 99% in our server and client business, the gross margin might decline when transitioning to cloud computing, but the total market size would be larger.

We would sell more, especially to small and medium-sized enterprises, and even our sales in products like Exchange, SharePoint, and Teams would see growth, which have now been greatly expanded. This was the basic idea I mentioned in the memo at that time.

Bill Gurley

So, are there any elements of cultural change? I think there are many CEO appointments every year, but many of them fail. Just like Intel is currently undergoing its second reboot. And, as Brad pointed out, there was once a belief that Microsoft was like IBM or Digital Equipment Corporation (DEC), that its glory days were over. So, what did you do, and what advice would you give to the new CEO to help them reboot the culture and drive the company in a different direction?

Satya Nadella

I think one of my advantages is that I am a thorough "insider," right? I have spent almost my entire career at Microsoft. So, in a sense, if I criticize our culture, I am actually criticizing myself.

Therefore, to some extent, the breakthrough I achieved was that people never felt I was an outsider criticizing those present, but rather directed the blame more towards myself because I am almost a part of this culture, you see? I can't say anything that I am not involved in.

Remember, Bill, you mentioned this point, and I clearly remember the first time Microsoft became the most valuable company. I remember walking around the company campus, and all of us, including myself, seemed very proud, as if we were truly the greatest creation of humanity, and our intelligence was finally reflected in our market value.

I always felt that this culture was something we had to avoid because from ancient Greece to modern Silicon Valley, the only thing that can destroy civilizations, nations, and companies is "arrogance." So one of the most important turning points was when my wife recommended a book to me a few years before I became CEO—Carol Dweck's "Mindset." I initially read it for my children's education and development, but unexpectedly, this book gave me a lot of inspiration.

I think the concept of this book is fantastic. We have been discussing learning and a learning-oriented culture, and this is precisely the best cultural concept we can choose. Therefore, I attribute our cultural success to this concept because it is not only Microsoft's concept; it applies to all aspects of life.

You can use this mindset to become a better parent, a better partner, a better friend, a better neighbor, and a better manager and leader. So we adopted this concept, and one phrase I have always used is to transform "know-it-alls" into "learn-it-alls." This is a goal that can never truly be reached because once you say you have a "growth mindset," you can never truly possess it Therefore, this concept has been very helpful to us. Cultural change, like all cultural transformations, requires time and space to grow naturally. Moreover, this transformation is both top-down and bottom-up; they complement each other. In fact, every time I meet with the company, or even my executive team, I always start with "mission" and "culture," which are the two pillars of our discussions.

As for other aspects, I have also been very disciplined in adhering to my framework. As I wrote in my memo, for almost the past 11 years, the structure and philosophy I have adhered to are the same: mission and culture, which is the worldview.

For example, environmental intelligence, ubiquitous computing, followed by specific products and strategic frameworks. I choose each word very carefully and repeat it meticulously until I myself feel tired of it, but I still persist.

Brad Gerstner

Speaking of this, you mentioned the phase changes we have experienced. I have heard you say that as a large platform company, most of the value capture is actually determined in the first three to four years of phase changes, when the market position has already been established.

I have heard you say that Microsoft missed the search and also missed mobile, but you also mentioned that Microsoft caught the "last train" of cloud computing. So, when you start thinking about the next big phase change, it seems that you and others on the team, including Kevin Scott, realized early on that Google might be a step ahead in AI, after all, they have DeepMind.

You decided to invest in OpenAI. What convinced you of this direction rather than continuing to push for AI research within Microsoft?

Satya Nadella

That's a very good question. Because there are several aspects to this. First of all, we have been deeply exploring AI for a long time. Obviously, Bill founded Microsoft Research (MSR) in 1995, and I remember the first group was actually focused on natural user interfaces.

At that time, there were also many people at Microsoft Research, including Regret, Kaifu, etc., who were all trying to solve the problem of language understanding, including Hinton's early work, where he did some research on DNS while he was still at Microsoft Research before he went to Google. So, I think we missed the opportunity to synchronize with Google to increase our investment in AI early on, and we missed the chance when Google acquired DeepMind.

This makes me very regretful. However, as a leader, I have always focused on some other directions. For example, Skype Translator was the first project I focused on because it was very interesting—that was the first time we saw the effective application of transfer learning. That is to say, we could train on one pair of languages, and then it could perform better on another pair of languages.

This was the first time we could say, "Wow, machine translation can also be DNS," which was completely different from what we had done before. So, from that time on, I became fascinated with language, and Kevin did too. In fact, I remember the first time I met Elon and Sam, they mainly wanted some Azure credits, and at that time they were more focused on reinforcement learning (RL) And "Dota 2".

Then, after a while, we talked to them about natural language processing (NLP). At that time, they talked about transformers and natural language processing. I felt that this was core to our business and aligned with our long-standing structural positioning.

I have always thought that if there were some breakthrough model architecture that could produce nonlinear growth and demonstrate capabilities we had never seen before, it could be a tremendous opportunity for us.

Bill, you always say, "In the digital realm, there is only one category, and that is information management." You believe that information is organized in some way. We once did a very famous project called WinFS, which aimed to organize all information by creating schemas for everything.

But in fact, that is impossible to achieve. Therefore, we need some breakthroughs. I thought at the time that perhaps organizing information into language in some way, similar to how the human brain organizes information through language and reasoning, could be the answer.

That is why we decided to invest in OpenAI; in fact, Sam, Greg, and the team's ambitions were another reason that prompted my decision.

As for the "scale law," I remember the first time I saw the memo about the "scale law" was when Dario and Ilia wrote it at OpenAI.

At that time, I thought, "If this field can really bring exponential performance improvements, why not go all out and give it a real chance?" Then, when we saw its effectiveness in GitHub Copilot, we felt it was truly feasible, and we decided to increase our investment. So that was the initial motivation.

Bill Gurley

I think that in the past phase changes, some incumbents did not keep up with the pace quickly. You even mentioned that Microsoft may have missed opportunities in mobile or search. It can be said that, especially at my age, I have witnessed these changes firsthand, and now everyone seems to have woken up, or this round of changes is like a carefully choreographed performance, with everyone almost at the same starting line.

I am curious to know if you agree with this or how you view the key players in the competition, such as Google, Amazon, Meta, Llama, and Elon entering the game.

Satya Nadella

That's an interesting perspective. As you said, I have also been thinking about this issue. If you look back to the late 1990s, Microsoft was at the top with almost no competitors. But now, everyone talks about the so-called "MAG 7," and perhaps even more than that. As you said, everyone has realized this, and everyone has strong balance sheets.

It can even be said that OpenAI can be seen as the 8th player in a sense. Because this generation of companies has been established in some way—OpenAI is like the Google, Microsoft, or Meta of this era. So, I believe the competition ahead will be very intense I also believe that this will not be a "winner-takes-all" situation, although there may be instances where this occurs in certain areas. For example, in the hyper-scale domain, it will absolutely not be a winner-takes-all. The global market, including those outside of China, will require multiple suppliers providing cutting-edge models, distributed across the globe.

In fact, I think Microsoft has a very good structural advantage in this regard—you remember Azure, right? Its structure is somewhat different. We built Azure for enterprise workloads, focusing on data residency issues, supporting over 60 regions, even more than other cloud service providers.

So, we are not building the cloud for a single large application, but for a variety of heterogeneous enterprise workloads. I believe this will be the main battleground for future inference demands, combined with data centers and application servers. Therefore, I think there will be multiple winners in infrastructure, and the same goes for models; each hyper-scale cloud provider will have a bunch of models, around which there will be an application server.

Every application today, including Copilot, is a multi-model application. In fact, a brand new application server will emerge, just like there used to be mobile application servers and web application servers, now we have AI application servers.

For us, this is Foundry; we are building it, and other companies will build similar things. In the future, there will be multiple such servers.

At the application level, I believe network effects will always exist at the software layer. So at the application layer, there will be different network effects, including both consumer and enterprise sides.

Therefore, fundamentally, I think you must analyze from a structural perspective; there will be very fierce competition between different levels, with the top 7, 8, 9, and 10 companies competing intensely at different technological levels.

As I always tell our team, pay attention to those latecomers, the suddenly emerging entrepreneurs. You need to closely watch which emerging companies will bring you change; at least OpenAI is one of them. So far, it has achieved scale and speed.

Brad Gerstner

Speaking of this, if we focus on the application layer, let's first talk about consumer AI. Bing is a very large business; you and I have discussed that "10 blue links" may be the most successful business model in the history of capitalism, but it faces a huge threat from a new model, which is that consumers now just want answers.

For example, my kids say, why should I go to a search engine when I can get the answer directly? So, do you think Google and Bing can continue to grow in the age of answers?

What does Bing or your consumer business under Mustafa's leadership need to do to compete with ChatGPT? After all, from a consumer perspective, ChatGPT is already a very prominent presence Satya Nadella

Yes, I think your first point is very correct, which is the combination of chat and answers. This is exactly what the ChatGPT product is. As you said, it is not just a search engine, but a stateful intelligent agent that truly breaks the limitations of traditional search.

Traditional search engines are stateless. When you search, although there is a history, each search is a new query. Now, these intelligent agents will become more intuitive, persistent, and "memorable."

Therefore, this is also why I am so happy—I've been trying to reach a search agreement with Apple for 10 years. So, when Tim finally reached a collaboration with Sam, I was really excited. For us, ChatGPT securing this agreement is more meaningful than for anyone else because we have established a business and investment relationship with OpenAI.

In this regard, I think distribution channels are still very important. Google has a huge advantage in this area, after all, they are the default search engine on Apple and also on Android. They reach a massive user base.

However, once habits are formed, they are hard to change. As you mentioned, even though I now prefer using Copilot, my usage habit is still to directly input queries in the browser. Sometimes, even when I use Copilot, the functionality of search engines still has its unique value. For example, when I am dealing with navigation information, I will search on Bing, while for other questions, I tend to use Copilot.

I believe this change is happening universally. We are just one or two steps away from completely migrating certain business queries to chat formats. When business intent also migrates to chat platforms, traditional search engines may face real challenges.

Currently, business intent has not fully shifted, so traditional search engine operations are still running. But once this business intent shifts, traditional search will face significant challenges. Therefore, I think this is a long-term structural change.

Under Mustafa's team management, we have three core products: Bing, MSN, and Copilot. So we believe he has clearly defined the roles of these three, which together form an ecosystem.

One is the traditional search engine, one is news and information streams, and one is the new intelligent agent interface. They have a social contract with content providers; we need to bring them traffic, while possibly needing paywalls, ad support, and other models. This is how we are managing, and we already have our own distribution channels.

One of the few advantages we still have is Windows. Although we missed the browser market, with Chrome becoming the dominant browser, this was a failure for us, but we are regaining the market through Edge and Copilot. Windows, at least in some aspects, is still an open system for us, which means that both ChatGPT and Gemini have the opportunity to leverage their strengths on it. Microsoft does not restrict their performance; rather, it can bring more competition and innovation Bill Gurley

Satya, everyone is talking about these intelligent agents. If you look to the future, you can imagine that many players will want to take action on data from other applications and systems.

Microsoft's position in this regard is quite interesting because you control the Windows ecosystem, but your applications also appear in the iPhone and Android ecosystems. What do you think about this situation?

There are both service-level issues and partnership issues here. Will Apple allow Microsoft to control other applications on iOS? Will Microsoft allow ChatGPT to launch applications on Windows and access application data? This question extends to areas like search and e-commerce— for example, would Booking.com allow Gemini to transact without their permission or knowledge?

Satya Nadella

Yes, I think this question is very interesting. To some extent, it is still unclear how this will be realized. Indeed, there is a very traditional way of thinking; think back to how enterprise applications achieve interrupt operations. They typically do this through connectors, and users need to purchase connector licenses.

Thus, a certain business model emerged. SAP is a classic example; you can access SAP data by owning a connector. I think when interfaces between intelligent agents emerge, similar patterns may reappear. But for consumers, this model is still not very clear, because on the consumer side, value exchange is often achieved through advertising and traffic, and these methods may change in the world of intelligent agents.

So, the business model on the consumer side still seems somewhat unclear to me. But on the enterprise side, I believe there will ultimately be a situation where everyone will say that in order to enter my operational space or extract data from my architecture, it must go through some form of intelligent agent interface, and this interface is licensed.

For example, today when I use Copilot at Microsoft, I have connectors to Adobe, SAP instances, and our CRM (Dynamics). This model is very interesting. In fact, think about it, we haven't really used those enterprise applications for a long time.

We license many SaaS applications, but in reality, very few people use them personally; more often, certain people within the company are inputting data. But in the AI era, this situation has changed because all data has become more accessible.

You can easily query, for example, "I want to meet with Bill, tell me all the companies Benchmark has invested in." At this point, AI will extract relevant information from the web and the CRM database, integrate it, and provide a summary or notes.

Bill Gurley

To some extent, all this content can be monetized through us or these connectors. But a clearer point is whether ChatGPT can directly open random applications on the Windows operating system and access data; this question has already been widely discussed What do you think about this issue?

Satya Nadella

That's an interesting question. Who can allow such behavior? Is it the user or the operating system? Frankly, on Windows, I have no way to prevent such behavior except through some security measures.

So, theoretically, I can ensure that such behavior is safe through certain means. My biggest concern is security risks. If malware is downloaded and starts executing operations in the system, that poses a huge risk. Therefore, I think we will integrate this permission control into the operating system, setting some higher access permissions and permission management.

However, ultimately, users will be able to control these behaviors on an open platform like Windows. I believe Apple and Google will have more control permissions, so they won't allow such behavior to occur.

From this perspective, you could say that open platforms like Windows have this advantage, while the closed systems of Apple and Google have their own advantages, and in the end, we need to see how each party defines these rules.

Bill Gurley

We can look at this issue from another angle and then continue the discussion. If it were the Android operating system, or what we call Android AI, or iOS AI, could it read emails on the phone through the Microsoft client?

Satya Nadella

Yes, I've been thinking about this issue. For example, today we have licensed Outlook synchronization for Apple Mail. This case is interesting; I think there might be some value leakage, but at the same time, it's also one of the reasons we can keep Exchange.

If we hadn't done this licensing at that time, it might have been more troublesome. Therefore, I think, going back to your question, Bill, when we build Microsoft 365, we must design it around a trust system. We can't let any agent come in and do anything because, first of all, it's not our data; it's the customer's data. So, the customer must agree, and the customer's IT department must also allow it. This is not a switch I can set arbitrarily.

The second point is that it must have a trust boundary. So, I think we will implement such functionality on M365, and such operations will be similar to Apple's smart management. Imagine that we will establish a similar trust and governance structure for M365.

Bill Gurley

You've talked a lot today, and I strongly recommend everyone download and delve into it because it's really fascinating.

Brad Gerstner

So, Satya, continuing to delve into this topic. Mustafa mentioned that 2025 will be the era of "infinite memory." Bill and I have been discussing since the beginning of this year that the next tenfold leap is likely the persistent memory brought by ChatGPT, which can perform operations with our authorization We have already seen the preliminary implementation of memory, and I am very confident that this issue will be basically resolved by 2025. However, regarding the issue of executing operations, when will we be able to tell ChatGPT, "Please help me book the lowest price room at the Four Seasons Hotel in Seattle for next Tuesday"? Bill and I have had many discussions about this, and it seems that the computer usage scenario is an early test case for this issue. So what are your thoughts? Do you think this is a difficult problem?

Satya Nadella

Yes, I agree with your view that the most open and limitless operational space is still very challenging. But as you mentioned, there are indeed two or three very exciting points that go beyond the scalability and raw capabilities of the model itself. One of them is memory, another is tool usage or executing operations, and one more I want to mention is permission management.

That is to say, what can you do? For example, with Microsoft's Purview product, it increasingly focuses on what permissions you have, what data you can safely access, and who manages and governs it.

So when you put all of these together, the behavior of the agent becomes more manageable. When executing operations, it is verifiable and has memory functionality, then you enter a completely different stage where it can handle more autonomous tasks.

However, I have always believed that even in a fully autonomous world, we will still face exceptions where you may need to request permission or call other operations. Therefore, we still need a UI layer to organize this work. For this reason, we view Copilot as the organizational layer for work documents and workflows.

But back to your core question, I believe that even if the model reaches 4.0 (not even 0.1, 4.0 is already very good), the functionality calls will still be limited. Especially on the consumer side, web functionality calls remain very challenging.

At least on the open web, it can perform operations on several websites, but once it involves tasks like booking flights or hotels, it can run into problems if the backend architecture changes, although it can improve through learning. But I think it will still take one to two years to complete more functionalities.

From a business perspective, there has already been some progress in creating sales agents, marketing agents, supply chain agents, etc. For example, in Dynamics, we have already implemented 10 to 15 agents that can automatically handle tasks like vendor communications, updating databases, and adjusting inventory. These can all be accomplished today.

Bill Gurley

Mustafa mentioned comments about near-infinite memory, well, I believe you should have heard or discussed this internally. Can you provide some clarification on this? Or is this part not yet public?

Satya Nadella

I think, to some extent, the memory system is like having a type system, right? That is the key. It doesn't start over every time. You have to organize it Bill Gurley

I understand. He means you have a technological breakthrough in this area?

Satya Nadella

Yes, actually we did an open-source project, I remember it was done by the TypeScript team. What we are trying to do is to structure memory processing so that every time I start up, I can cluster based on previous operations and then perform type matching, allowing us to gradually build a memory system.

Brad Gerstner

Let's change the topic and talk about enterprise AI. You mentioned that Microsoft's AI business has about $10 billion in revenue, and this part is all inference tasks, not renting raw GPUs for training. How do you see the current market regarding significant workload shifts? What are your current revenue products?

Satya Nadella

Yes, actually most of the training collaboration with OpenAI is more about investment aspects, which won't directly appear in our quarterly financial reports, but rather in other revenue items based on our investments.

Brad Gerstner

So this means your revenue or losses are mainly reflected in other income or loss sections, right?

Satya Nadella

That's right. So, most of the revenue or almost all of it comes from our API business, or rather, the inference costs of something like ChatGPT are also part of it.

This is a different part. Now, what are the "hit" applications of the era? ChatGPT, Copilot, GitHub Copilot, and the APIs of OpenAI and Azure OpenAI. You could say that if you list these most popular applications, it should be about these few. So that's the biggest driving force.

Our advantage with OpenAI is that we have a two-year first-mover advantage, with almost no competitors. As Bill mentioned, everyone has woken up, but I think there may never be another opportunity to have such a two-year lead again. Who knows?

You're right, there will always be other teams releasing some samples that suddenly break the world. However, I think it's very difficult to establish such a leading position through a foundational model. But we have this advantage, especially OpenAI, which can leverage this to establish the acceleration track for ChatGPT.

Brad Gerstner

Do you think other companies training these models and model clusters account for a larger proportion of their AI revenue, unlike you?

Satya Nadella

I'm not sure. I can only say, look at what "hit" applications other companies have; I'm not clear on which models they specifically run and where they run them. I guess Google's Gemini is one of them. Based on the DAU numbers of any AI product, ChatGPT is one of them, and even Gemini surprised me Although I believe it will grow due to its own distribution capabilities, an interesting point is that despite everyone talking about the scale of AI, there aren't many truly "explosive" applications. For example, ChatGPT, GitHub Copilot, Copilot, and Gemini are probably the most well-known ones.

Brad Gerstner

Well, there are indeed many startups that are moving down this path and gradually gaining some attention, many of which are built on Llama.

Satya Nadella

However, if you say, oh, what about Meta? But if you ask how many of the 10 most influential applications have over 5 million DAU, how many can you list?

Brad Gerstner

I think Zuckerberg might argue that Meta's AI definitely has over 5 million DAU, but as for the independent applications you mentioned, it is indeed as you said, Zach's technology runs entirely on its own platform.

Satya Nadella

He does not rely on public cloud.

Bill Gurley

Satya, speaking of the enterprise side, the programming space has also started to accelerate, and you all are performing well in this area, the market is very interested. I have a question about Copilot's approach. I know Mark Benioff has some criticisms about it, calling it "Clippy version," or something else.

Are you concerned that some people think AI should start from scratch and rebuild the entire infrastructure, for example, whether tools like Excel are still necessary, or can you eliminate these unnecessary elements through AI-first products? The same situation applies to CRM. In fact, many fields and tasks could potentially be simplified or hidden through AI.

Satya Nadella

This question is very important. For SaaS applications or business applications, I can talk about our own approach on Dynamics. Our thinking is that with the arrival of the agent era, business applications may undergo an integrated change.

Because if you think about it, they are essentially composed of a shared database of business logic, and this logic will shift to these agents, which will perform multi-database updates. All business logic will be centralized in the AI layer; in other words, AI will become the core of all business logic. Once the AI layer becomes the core of business logic, all backends will be replaced.

Our current win rate on Dynamics is very high, especially in terms of agent usage. We are actively pushing this work forward, aiming to integrate it into the entire process. Not just CRM, in fact, our finance and operations parts are also undergoing similar changes, as everyone wants to see more AI-native business applications. That is to say, the logic layer of business applications can be orchestrated by AI and agents, making enterprise applications more seamless Additionally, you might ask, why do we still need Excel? In fact, one of the things I’m most excited about is the combination of Excel and Python, which is just like the combination of GitHub and Copilot. What we are doing is using Excel in conjunction with Copilot, not just simply processing data, but allowing it to plan the entire process for you like a data analyst.

It can automatically generate plans and then execute those plans. It’s like a data analyst using Excel for data analysis; it’s not just a “row and column” view, but also a “tool” that can perform actual operations.

Brad Gerstner

One of the most discussed questions I heard today is about the ROI of these investments. You have over 225,000 employees. Are you leveraging AI internally to improve productivity, reduce costs, or drive revenue growth? If so, can you provide some specific examples?

Additionally, regarding what Jensen said earlier, he mentioned that when revenue grows two to three times, employee numbers are expected to grow by 25%. If Azure's revenue grows 2 to 3 times, do you also expect employee numbers to show similar growth?

Satya Nadella

Yes, in fact, this is something that Microsoft is very focused on right now, and it’s a topic that customers are very concerned about. Here’s how I see it: I really enjoy learning from the experiences of industrial companies in lean management. Right? This is really interesting. For example, the growth of these companies often exceeds GDP growth, which is remarkable.

These good industrial companies can improve by 200 to 300 basis points through lean management, adding value and reducing waste. That’s the practice of lean management. So I think AI is like lean management in knowledge work.

We are really learning from these industrial companies, such as how to observe process efficiency, how to find parts that can be automated, and how to make processes more efficient. So customer service is one of the most obvious examples.

We have invested about $4 billion in this area, covering everything from Xbox support to Azure support. In fact, this is a very serious investment. Through the front-end guidance rate, we are able to improve the efficiency of agents, and most importantly, agents are happier, customers are more satisfied, and our costs are decreasing.

This is one of the most obvious examples, and another is GitHub Copilot. It’s also a very typical example. In the GitHub Copilot workspace, you start with a question, then move to a plan, execute or specify a plan, and then do multi-file editing. It completely changes the workflow of teams.

Then there’s 365, M365’s Copilot, which can also serve as an example. For instance, from my personal experience, the preparation work in the CEO's office for meetings with clients has hardly changed since 1990 In fact, I see it this way—imagine how financial forecasting was done before computers existed? We used fax machines, internal memos, until personal computers became popular, and people started using Excel spreadsheets to send emails, exchanging numbers and forming forecasts.

Now, the era of AI has arrived, and things have changed. When I prepare for client meetings, I enter Copilot and ask, “Please tell me everything I need to know about this client.” It collects information from my CRM, emails, Teams meeting notes, and the web and provides it to me. I can create pages based on this information and share them with my team in real-time.

Imagine that the reporting methods used in the CEO's office are no longer necessary; this content can be generated through queries and even shared in real-time, allowing team members to annotate directly on it. Therefore, I work collaboratively with AI and in real-time with my colleagues. This is a new workflow that is becoming widespread across various fields.

For example, there’s a case in the supply chain field: someone said the supply chain is like a trading desk, just lacking real-time information. You have to wait until the financial quarter ends, and then the CFO will come to blame you for previous mistakes.

So what if financial analysts could provide you with feedback in real-time? For instance, if you are drafting a contract for a data center, AI might remind you of which terms to consider. All this real-time intelligent feedback is changing workflows and work products. So, we see many similar cases.

I think your core question is how to achieve operational leverage through AI. That’s the goal we want to reach. We expect to reduce labor costs through AI, but everyone’s output will be higher. My researchers, perhaps each of their GPU usage efficiency will be higher. That’s my perspective on this matter.

Brad Gerstner

That makes sense. Okay, let’s change the topic and talk about the model expansion and capital expenditure issues you mentioned earlier. I’ve heard you talk about Microsoft’s capital expenditure. Imagine when you took over in 2014; you probably didn’t expect capital expenditure to turn out like it is today.

In fact, you mentioned that today these companies look increasingly like industrial companies in terms of capital expenditure, rather than traditional software companies. Your capital expenditure grew from $20 billion in 2020 to potentially $70 billion by 2025. The correlation between your capital expenditure and revenue is very high, which is interesting.

Some people are concerned that this correlation might break down, and even you have mentioned that in the future there might be a situation where “capital expenditure needs to lead,” and we may need to prepare for this elasticity. So how do you view this level of capital expenditure? Does it keep you up at night? When do you think this growth rate will start to slow down?

Satya Nadella

Well, there are several aspects to consider. First, as a hyperscale company, we actually have structural advantages in this regard because we have been practicing all of this for a long time. Data centers have a 20-year lifecycle, and you only need to pay for power when you are using the equipment, while the hardware usage cycle is about 6 years, and you know how to improve equipment utilization These are all things we know. And the good news is that this is not only capital-intensive, but it is also software-intensive. You can improve the return on invested capital (ROIC) through software.

In fact, many people initially wondered how ultra-large companies like Microsoft could make money. What is the difference between the new ultra-large companies and the old-style hosting companies? The answer is: software. This also applies to the construction of AI accelerators—through leading technology construction, we can better utilize capital.

In fact, a current trend is the so-called "catch-up." Over the past 15 years, we have continuously built and expanded infrastructure, but suddenly, a new demand has emerged in cloud computing, which is the demand for AI accelerators. Because now every application needs a database, a Kubernetes cluster, and an AI accelerator. If you want to provide these three services simultaneously, you must build AI accelerators on a large scale. This demand will gradually become standardized. First is construction, followed by the standardization of workloads. Ultimately, everything will continue to grow like cloud computing.

So, we will continue to grow, ensuring that the demand for these workloads is diversified, avoiding adverse choices based solely on supply-side construction, and ensuring that real demand can be generated globally. I will focus on these factors. This is how to manage return on capital.

Additionally, regarding profit margins, there will definitely be differences. We discussed earlier that the profit margin of Microsoft Cloud is different from the profit margin of GPU original hardware. These will have different profit margins, such as the layers combining cloud architecture, GPU, and applications like GitHub Copilot or M365. These all have different profit margins. Therefore, in the AI era, our strategy is also to continue to maintain a diversified product portfolio to ensure profit maximization.

In fact, one of Microsoft's advantages in cloud computing is that not only are we larger than Amazon, but our growth rate also exceeds Amazon's, and our profit margins are higher than Amazon's. This is precisely because we have made deep layouts at multiple levels. This is the strategy we hope to continue in the AI era.

Bill Gurley

Because there has been a lot of recent discussion about model scaling, and historically, there has also been discussion about how to scale clusters multiple times rather than scaling to a certain size all at once. Recently, in a podcast, they completely changed their thinking, saying that if we no longer do it this way, it might actually be better because we can go directly into the inference stage, which has become cheaper and does not require a large capital expenditure. I am curious about how you view the scaling and training costs of large-scale LLM models, and how you think they will develop in the future, even though these two viewpoints are two sides of the same coin.

Satya Nadella

Well, you know, I firmly believe in the law of scale. First, I have to say that, in fact, the bet we made in 2019 was based on the law of scale, and I still believe in it In other words, do not oppose the law of scale, but at the same time, we must remain realistic based on several different factors. One is that as the scale of the cluster increases, the exponential growth of the law of scale will become more difficult, as distributed computing problems become more complex during large-scale training. So, that is one aspect of it.

However, I still believe that, despite this, OpenAI's friends can illustrate their approach, but they are still continuing with pre-training, and I think that will not stop; it is still ongoing.

What is exciting is what OpenAI has publicly discussed, and what Sam has also mentioned, which is their work on 0 and 1. This chain of thought through automatic grading and testing reasoning is actually a huge advancement. Essentially, the reasoning computation time itself is also a kind of law of scale.

So you have pre-training, and then you effectively generate tokens through this testing time sampling, and send them back to pre-training, creating more powerful models that can run during the reasoning phase. Therefore, I think this is an excellent method to enhance the model's capabilities.

The computational benefits of testing time or reasoning time lie in the fact that when running these O1 models, there may be two independent things involved: sampling is similar to training, using it to generate tokens for training; and customers using O1 are actually using more resources. So you get rewarded from that. Therefore, this economic model is feasible. So I think this is a good way.

In fact, this is also what I have always said; I have more than 60 data centers globally, which is a good structural advantage.

Bill Gurley

The hardware architectures for these two expansion directions are different, right? One is for pre-training, and the other is for reasoning.

Satya Nadella

Yes, I think the best way to understand it is that there is a proportional relationship between them. So going back to what Brad said about ROIC, this is actually where you have to establish a steady state.

In fact, every time I talk to Jensen, his view is very correct: you want to buy some every year, rather than making a one-time purchase. Think about it, when you set the depreciation cycle of the equipment to 6 years, the best way is to buy a little each year, gradually accumulating, right?

You use leading nodes for training, and the second year it enters the reasoning phase. This is the steady state. I believe we will ultimately achieve this steady state across the entire equipment population, not just in terms of utilization, but also ROIC. Ultimately, demand will match supply.

As you mentioned, everyone is asking whether exponential growth has stopped; economic realities will also come into play. At some point, everyone will examine and make economically rational decisions. Even if I double my capabilities every year, if I cannot sell these products, it is meaningless.

Another issue is the winner's curse. You do not necessarily need to publish papers; others just need to look at your capabilities, and then they can distill or replicate them in other ways. It is like piracy. You can set various usage terms, but in reality, these things are hard to control. Distillation is like that; moreover, you do not have to do anything, just reverse engineer your capabilities and achieve it in a more efficient computational manner Therefore, considering all of this, I believe there will be a limit. Everyone is currently chasing a bit of an edge, but eventually all economic realities will emerge, and the network effects are at the application layer. So if the network effects are all at the application layer, why should I invest a large amount of money in some model capabilities?

Brad Gerstner

What I hear you saying is that Elon once mentioned he wanted to build a million GPU cluster, and I think Meta has said something similar.

Bill Gurley

I remember he mentioned needing 200 for pre-training, and then joked about a million.

Brad Gerstner

But I remember he jokingly mentioned a billion cluster. But in fact, Satya, based on your views on pre-training and scaling, have you changed your infrastructure planning?

Satya Nadella

My current approach to building is to take a relatively close to 10x mindset. That is, we can discuss cycles, like every two years, every three years, or every four years, all having an economic model. I think there needs to be a relatively disciplined way to think about how to clear inventory to make it meaningful, or you can also consider it from the depreciation cycle of the equipment.

You can't just buy a large amount of equipment at once unless you can find the physical characteristics of the GPU that can match my financial situation, and its profit margins are as good as those of the hyperscalers. Simply put, my approach is to continue building how to drive inference demand, then enhance my capabilities, and maintain efficiency.

I certainly know Sam may have different goals, he has a deep belief in AGI, or profound views in other areas, so go for it. So, I think this is also part of our focus.

Bill Gurley

But I heard Mustafa mention in a podcast that Microsoft will not participate in the current large model training competition, is that accurate?

Satya Nadella

Well, we won't do redundant work. After all, given our partnership with OpenAI, it doesn't make sense for Microsoft to conduct a second training.

Bill Gurley

Right, exactly.

Satya Nadella

So we are very cautious. This is also a kind of discipline in our strategy. In fact, this is what I have been emphasizing to Sam: we are betting everything on OpenAI, saying we want to concentrate computing resources, and because we own all the IP rights, we feel very good about making such a choice.

So, what Mustafa means is that we will concentrate more resources after training, even in validation and other aspects. Therefore, we will focus on adding more model adaptations and capabilities while ensuring we also have principled pre-training work, so that we can make corresponding adjustments internally. We will continue to develop model weights and categories that adapt to different use cases Bill Gurley

Regarding the question raised by Brad about balancing GPO and Gpuroi, does your answer also explain why you outsourced part of the infrastructure to Core Weave and established a partnership?

Satya Nadella

We did this because everyone was caught off guard by the impact of ChatGPT and OpenAI. Yes, absolutely. I mean, there was no way to plan for the supply chain, nothing could be anticipated. Who could have predicted what happened in November 2022 over twenty years ago? It was a bolt from the blue. So we had to catch up. We said at the time that we wouldn't overly worry about efficiency issues. So whether it's Core Weave or other companies, we are purchasing from all over. Makes sense, right? This is a one-time thing, and now we are catching up. So it's more like a process of catching up.

Brad Gerstner

So do you still have supply constraints now, Satya?

Satya Nadella

There are no longer any constraints on chip supply. We did experience some supply constraints in 2024. We announced this publicly, so we are optimistic about the first half of 2025, which will be the remainder of our fiscal year. After that, I think by 2026, our situation will be better. So we have a good supply chain.

Brad Gerstner

I've heard that you've achieved very positive results regarding your investments in O1, test time calculations, and post-training work. The things you mentioned are also very compute-intensive because you need to generate a large number of tokens and then backfill those tokens into the context window repeatedly. This computational demand can accumulate rapidly.

Jensen has stated that he believes the demand for O1 inference computing will grow by a million times or even a billion times. Do you feel you have enough long-term plans to scale inference computing to keep up with the demands of these new models?

Satya Nadella

Yes, I think there are two points to focus on here, Brad. In a sense, understanding the entire workload is very helpful. Within the overall workload, in the application of agent models, AI accelerators must be present. In fact, OpenAI's own container service is the fastest-growing part.

After all, these agents need a "temporary workspace" to perform some automated grading or even generate samples. This is where they run the code interpreter. By the way, this is a standard Azure Kubernetes cluster. So from a certain perspective, the ratio of regular Azure compute to GPU and some data services can also be seen as part of the overall computing model.

Therefore, back to your question, when we talk about inference, we are actually talking about a combination of these things. That’s why I believe AI is no longer a separate domain from cloud computing; it has now become a core part of cloud computing In an era where every AI application is stateful and agent-based, the classic application server combined with AI application servers and databases is all that is needed when these agents perform specific operations.

So, I return to my fundamental point, which is that we have built over 60 AI regions, and all Azure regions are ready to support comprehensive AI applications. This is the infrastructure needed for the future.

Brad Gerstner

That makes a lot of sense. We have mentioned a lot about OpenAI in this conversation, but you are managing the balance between your significant investment in OpenAI and your own "ignite" plan. You showed a slide highlighting the differences between Azure, OpenAI, and OpenAI Enterprise, much of which involves enterprise-level capabilities, which are your unique advantages.

So, how do you think about this competitive relationship when you see it? Do you believe ChatGPT might be the ultimate winner on the consumer side? You will also have your own consumer applications, which may later collaborate in the enterprise market. How do you view the competitive relationship with OpenAI?

Satya Nadella

So far, my view is that OpenAI, as a large-scale company, is no longer just a startup. It is now a very successful company with multiple business lines and market areas.

So I think about this issue from a principled standpoint, just like I would with any other large partner, because I do not see them as competitors; I see them as an investment partner, looking at how our interests align. I consider them as IP partners because we provide system IP, and they provide model IP. So this is also an aspect of our mutual deep concern for each other's success.

Third, I view them as a major customer, and therefore, I want to serve them as I would any other major customer.

Finally, there is collaboration. Whether it's the consumer-side Copilot or collaboration with products like M365, we will engage in deep cooperation here. So, when I look at this competition, ultimately there will be some overlap in these areas. But in this context, OpenAI having a partnership agreement with Apple, from a certain perspective, actually creates value for Microsoft shareholders.

As you mentioned the API differences, customers can choose to use them based on their needs. For example, if you are an Azure customer and want to use other Azure services, using Azure's API and related services will be more convenient. But if you are on AWS, simply using the API and statelessly using OpenAI is also quite good. So, in a sense, having these two distribution methods is also beneficial for Microsoft Bill Gurley

It can be said that this is indeed a very attractive topic in the Silicon Valley community and even in the broader business community. I think the relationship between Microsoft and OpenAI has been a focal point of attention. Last weekend, I heard Andrew Sorkin pressing Sam hard on this issue during Dealbook.

While there may be many things you cannot disclose, can you share something? Is OpenAI undergoing a restructuring, and are there plans to transition to a profit model? I guess Elon has also expressed some opinions on this. Can you share some insights?

Satya Nadella

Well, I think these questions should certainly be decided by OpenAI's board, Sam, Sarah, and Brad and their team, who will make choices based on their judgment, and we will provide support. From our perspective, what we deeply care about is that OpenAI continues to succeed because it benefits us. I also believe that OpenAI is an iconic company in this platform transformation, and the world is better because of OpenAI's success. So that is our basic position.

The next question is the kind of tension you mentioned. Like in all these partnerships, part of it is the tension of collaboration, and part of it is Sam as a visionary and ambitious entrepreneur, who has very clear goals and a fast pace of action. His pace is quick, so we need to balance that.

This also means we need to understand and support what he wants to do within our own constraints, and he also needs to understand and adapt to the discipline we need in some areas. So I believe we will find a balance.

But I think the good news is that we have come a long way within this collaborative framework. The past five years have been very good for both them and us. And from my perspective, I will continue to hold this position and hope to extend this partnership as much as possible. We can only benefit both sides with a long-term stable partnership.

Brad Gerstner

When you consider independent financing and untangling the connections between the two companies, are you planning to move forward quickly? I once mentioned that perhaps the best development path for OpenAI is to become a publicly traded company. After all, being a leader in the AI field with such an iconic business has a positive impact on their future development. What do you think about this possibility? Do you believe OpenAI's future development will continue to maintain the current partnership, or will there be greater changes?

Satya Nadella

I think I need to be careful not to overstep. Because in a sense, we are not members of OpenAI's board; we are just investors, like you. Ultimately, these decisions are made by their board and management. So to some extent, I will respond based on their judgment In other words, I am very clear that we want to support any decisions they make. For me, as an investor, the most important thing is our business cooperation and intellectual property partnership. We need to ensure that we protect our interests throughout this process and continuously strengthen these collaborations in the future.

But I believe that smart people like Sarah, Brad, and Sam will make decisions that best align with their mission objectives. We will also support the decisions they make in this process.

Brad Gerstner

So, perhaps we should wrap up. Thank you very much for your time today. I would like to summarize with the topics of "open" and "closed," discussing how we can collaborate to ensure the safety of AI. Maybe I can pose an open-ended question to you about how you view the differences between open source and closed source, and the collaboration in promoting safe AI.

You also mentioned that people can extract models for training, and ultimately some applications of these models may be ones we do not want to see. How do you view the role of a country and a group of companies in jointly promoting the development of safe AI?

Satya Nadella

I think this question has two aspects. First, I have always believed that open source and closed source are two different paths to creating network effects. I have never viewed them as purely a "religious battle," but rather from a business strategy perspective, seeing them as two different choices.

That is also why I think what Meta and Mark are doing is very smart. In a sense, he is trying to commoditize his advantages. It makes a lot of sense to me, and if I were at Meta, I would do the same. He has openly and persuasively talked about wanting Meta to be the "Linux" of LLMs (large language models).

I think this is a wonderful model, and in fact, there is a potential business model here. From an economic perspective, I believe that theoretically, an alliance driven by multiple participants may be better at doing this than any single company. For example, under the Linux Foundation, the main funding source for contributors is operational expenses.

I have always said that the success of Linux is not only due to its open-source spirit but also relies on the support of companies like Microsoft, IBM, and Oracle. Open source provides a great mechanism for this collaboration, while closed source may have advantages in other areas.

As for closed source, we have also experienced many successful closed-source products. Regarding security, it is a very important issue, but it is a separate issue. After all, legal and security standards apply to all products, whether open source or closed source.

So, I believe that under a capitalist system, it is best to maintain multiple business models and allow competition, letting different companies choose the paths that suit them. The government should also impose strict regulations on this As for the safety of AI, there is no doubt that we can no longer wait for the so-called "wait and see" approach. No government, community, or society can tolerate this attitude. Therefore, these AI safety organizations will set unified standards for all models. If there are issues related to national security leaks, everyone will be very concerned about it. Thus, I believe that the governments and national policies of various countries will have a significant impact on the development of these models, and regulatory systems will be established accordingly.

Brad Gerstner

It's hard to believe that we have only entered the ChatGPT era for 22 months. Looking back, your framework for phased transformation places Microsoft in a very advantageous position as we enter the AI era. So, the performance over the past decade is truly commendable and remarkable.

At the same time, I feel that when Bill and I see your leadership along with Elon, Mark, Sundar, and others, we are very excited because you are indeed driving the "American team" forward in the field of AI. Both of us are confident about how to position ourselves globally in the future. So, thank you very much for taking the time to communicate with us.

There is a new Newin, original title: "In-depth | Microsoft CEO Nadella's latest 20,000-word insight: The C-end agent business model still needs exploration, the advertising traffic model may face changes, and the B-end key lies in ecological integration."

Risk warning and disclaimer

The market has risks, and investment should be cautious. This article does not constitute personal investment advice and does not take into account the specific investment goals, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Investment based on this is at one's own risk