Satya Nadella 谈 AI“末日论”放缓、Microsoft 的总体规划与谁将赢得 AI

Satya Nadella, chairman and CEO of Microsoft, has generated $250 billion—with a B—in market value for Microsoft. Since you've been the CEO for three and a half years, the stock is up about—I guess it's about—120%.

I'm good for my $80 billion. I am going to spend $80 billion building out Azure.

Maybe after the Industrial Revolution, this is the biggest thing.

That's our goal with our frontier model. Our model should be the best model that they can use as a base. We create technology so that others can create more technology. That's who we are. We're toolmakers.

All right. Hi, guys. Good to see you coming out.

Good to see you.

Good morning, guys. How are you?

Good.

1. Dario's blog, "pacing the frontier," common sense AI safety

Thanks for joining us. Crazy weekend, but here we are. Do we need to pace the frontier?

Let's start with the common-sense part first, which is that we should do what it takes to build stuff that serves humanity first and is in human control. It's kind of crazy that we have to start with that level of common sense, but I think it's a good place.

When I think about pacing, the first thing that I at least believe is that the broad diffusion of this technology is the most critical thing, because the benefits of this tech showing up everywhere are really what it's all about, right? At the end of the day, if you say, "Serving humanity," let it actually reach humanity in ways that serve humanity. That means you've got to have choice, you have to have competition, and you have to have all kinds of business models, whether they're open weights, closed weights, or what have you.

Then the other aspect that I think is not talked about when we talk about control is the control that, for example, customers—enterprises or businesses—have around this technology. Sometimes this is so opaque, right? I want my privacy. I want to be able to embed my knowledge in a set of weights that I control. I want to see all of the code that's being generated. I want to use it to do fine-tuning of my own models. My IP shouldn't leak.

There's an entire body of things that nobody's talking about as much, which is that I really want to make sure that this tech is in my control. Then we get to what I think is a real issue of safety, and we should take it seriously, which is that we should take all the time we want to test things.

In fact, I love this idea of having third-party testers.

Oh, wow.

I grew up in a company that's always done testing, so it's novel that we should say, "Wow, they're having embedded third-party testers." Why not? It's a great idea. In fact, the only thing I would say is that we should avoid these cozy arrangements of who's testing what and who has access to what. It should be broad.

Were you surprised, though, when both the essay landed and then it seemed like there was a circling of the wagons amongst the frontier companies?

My suspicion is that it comes genuinely from this place where, when you start seeing reward hacking and what's happening in these environments with these agent swarms, there is the mundane. There is some DevOps error where somebody misconfigured a container—

Right, right—or these API keys.

Or an API key. Yeah, exactly. There's no monitoring. There's internet access. There's classic, basic DevOps. Then there is real, novel new stuff, which is this reward hacking with these persistent agents and so on. That's a place where I'll admit that the science is not there.

I thought Jakob's post was a good one, where he said—we're growing intelligence, not building intelligence. It's an experimental science, and the more experimental the science, the more you really need to make sure you're doing those experiments in controlled environments.

If anything, the place where I would love to see more transparency is in incidents like the Hugging Face incident: what would it take? One of the fascinating things right now is insider risk. Think about it, right? This is all test-time compute, by the way. It's not like it's only going to happen during some training run. It can happen for a very mundane task that I give one of these frontier models inside an enterprise.

I was telling David this: suppose I say, "Hey, go optimize my working capital." It may fake my books, right? This is a new type of insider risk.

So what is the way to do that? I would say, "Go build maybe a causal model, like a semantic model, that actually checks and verifies." I think there's a lot of product-building—making things more robust, which is classic engineering—that we should be talking a lot more about transparently, versus saying, "Hey, this is so mystical that we can't figure this out."

Do you buy this argument that it's mystical?

I buy the argument that we do not understand the latent space. As you said, do we understand the brain? We don't. We do functional MRIs and neuroscience, and we're trying to figure this out continuously, getting a little better understanding.

I do think that, in that sense, we don't exactly have a complete understanding. That's why, by the way, I also don't believe in neuralese, right? That's why I think making sure that the chain of thought is in language that we can all understand—and, in fact, is transparent—is important, so that when I go back to an enterprise that's using all these models, if you have the full chain of thought—

Chain of thought.

2. The failure of AI CEO messaging, monitoring agents, what will a slowdown mean for new AI products?

Chain of thought—and so then you can really go look at it deeply. In fact, you can have multiple models, and you can look at the chain of thought across those. I think these are all things that will become very important.

Satya, you've worked with technologists for decades. When you see, as a leader of one company, Microsoft, which has very crisp communications with the public, what's happening with Dario and his team—people coming out saying, "There's a 10% chance we all die"—what do you think is going through those technologists' minds?

Do you believe they actually believe that this is going to kill humanity, or are they going through some psychosis, or are they seeing something working on those frontier models that is terrorizing them? You're not a psychologist, but you have worked with technologists for a long time. Give us your read on what's going on in these organizations that's making people feel the need to resign and say we're all going to die.

It's hard for me to speak to what's happening in any of these places, but let's just say how I grew up. Even inside Microsoft, one of the biggest things you learn as an early engineering lead is how to deal with a showstopper bug.

Yeah.

Right. You're faced with a bug. What do you do? Do you stop and fix it, do you defer, or do you say, "Hey, this is such an edge case"? That's the judgment.

As the stakes go up, you want to take things seriously. Transaction processing—I remember working on databases—any bug where the transaction is going to get lost is something you take very seriously. Data loss is a reason to stop the thing.

I feel that the AI industry is culturally rediscovering this, maybe because it's possible that they see stuff that is a showstopper before the rest of us do. If you see a showstopper, stop the show to fix the bugs.

When you saw the Hugging Face run, and it was super performative, Dario did his whole post about civilization—what do you think? What's your take on the testing they ran? They could have run a test where they had 3,000 agents defend a bunch of websites. Instead, they instructed them to hack websites, and there was the hiding of information, all this anthropomorphizing of the agents.

The way I understand it, it was basically trying to do an eval for CyberGym. As I understand it, given that eval, it sort of figured out a way to, let's just say, reward-hack, and that's what led it to Hugging Face.

In fact, it speaks to what I think is the clear issue right now: you can have these things, if they're long-running, persistent agents, become essentially new insider risks. I would start from the very basics of saying, "Okay, what does containment look like?"

One of the things that I think is going to be a real issue, and a thing that needs great solutions, is truly aggressive monitoring of agent activity. That's behavioral—

Evidence.

Evidence. Everything has got to be auditable. Every object it accesses—if it goes and gets a secret, or it's going to chain a couple of things together—you should be able to see it when it's starting to chain a couple of vulnerabilities to go hack.

I think these are the ways that you really have to deal with these situations, versus saying, "This is so mystical that we can't figure this out." The core of my take is that we will have to get the engineering process around building out this experimental science to be more robust.

Yeah. Thanks. So I think that's a great point. I love how you differentiated in the Hugging Face episode between the mundane things they got wrong, like the misconfigured sandbox, and how Hugging Face had credentials just sitting in a public repository, and there was no monitoring. Then you have the genuinely novel behavior: the swarms of agents and the reward hacking.

That's the stuff that has everyone freaked out. I agree that we have to now figure out how to fix the bugs or fix the deeper problem that's coming from that reward hacking. To their credit, I think what the frontier labs are saying is, we are now going to slow down the pace of, let's say, raw power and shift toward reliability and predictability and what they call alignment, which I think is good business practice.

I guess, what do you think that means for what we see in terms of new products for the next year or two? Does it mean we just improve what we already have, or do we see new capabilities? What do you think this is going to mean?

A great question, David. I do think there's already a massive model overhang—capability overhang, in the sense that the models are very good, but broad diffusion requires a lot of things. It requires change management: if you're compressing workflows and changing workflows to happen differently, the amount of change management needed to incorporate these systems is what's taking time.

It also requires the ability to create these new form factors. If you think about coding agents, coding agents became really usable when you discovered that you could have an agent loop with a file system, and that was the breakthrough that just made coding agents work. Maybe now with CUA, with Astra and CUA, it could be a way for us to do computer use, or we could use long-trajectory tasks that can get completely automated.

I think these types of product innovations, where the model plus the harness allow us to do things that then lead to broad adoption, are important. I even go back to the ChatGPT moment for me. It was that RLHF at the very end that made a chat conversation possible.

Mhm.

And so I think that, yes, there's some science, and there is some form factor that then leads to broad diffusion. We now need to find the next level of these things that are doing real work in the real enterprise.

In that context, by the way, the other thing is it's going to be a multimodel world. At this point, just out of resilience, every enterprise now comes to me and says, "Hey, this model does refusals here; this model—I want weights here, I don't want them there." People are going to want multiple models, so one of the other things that we have to get right is some standards of interoperability.

Even KV cache—why the heck can't I use multiple model families and have KV cache reuse? We've had document standards; you and I lived through it. You kind of have things that are interoperable in the real world everywhere else. I think this industry also has to wake up and say, "Hey, in fact, if I were talking about the most important, pressing things, how do I have more standards on interoperability? How do I have a harness that is external to a model so that my memory is not tied to one model?"

This is the first time you're going to have a technology where your use of it and the data exhaust could not be yours. I mean, it's like if I sold you a database and said, "Hey, the data you put into your database is not yours; it's mine. It goes away if I take away the license." How would you feel about it? Therefore, I think we have some serious issues like that to deal with.

3. Economic incentives for frontier lab doomerism, where the AI profits are

I think that's a good segue. Sorry, let me just ask one question to connect the economic incentive argument to what's going on. The argument is the frontier labs are facing token price compression: $50 for OpenAI's million-token output versus, I think, someone estimated DeepSeek's new model can go as low as $0.15 for a million tokens of output. Let's call it $0.60—a 99% cost reduction.

4. Microsoft's master plan for AI, how they are allocating capital

If that's the big economic crux of what the frontier labs are facing, why would most enterprises be paying $50 when they could pay $0.60 for most of their tasks? Doesn't that also beg the question: are they in the wrong business model? I ask this for you as the CEO of Microsoft: what's the right business model? Do you want to be making the frontier model? Do you want to be running the compute and charging rent on your compute? Or do you want to be in the application layer? I know you talk about this a lot, but I would love your perspective from where we sit today.

Yeah, I think the fundamental thing that we're observing is good old-fashioned competition. For me, if I look back at it, we had some really great closed-source assets: Windows. What was the check against it? It was, of course, the Mac, but also Linux. We had a great closed-source product called SQL Server. What was the check against it? There was always a substitute called Postgres or MySQL.

I think that's what's happening. There's real competition between closed source and open source; the open-source check is real. That's good, quite frankly, because without it, I don't think we're going to have a broad frontier ecosystem or broad diffusion. Otherwise, we'll just be back to some mainframe lock-in that's just not a thing.

To your point, given that we will now hopefully continue to have a much richer choice in every layer, hopefully we can start building these AI products. Today, the royalty of an AI product all going to just the model layer doesn't make sense if you really want to build a product company. It just cannot be. In fact, it's the same thing: if you take the database, if there was no open-source check on closed source, the prices wouldn't have been at a place where people could have built the app tier successfully and with a margin.

I think the apps are going to become much more viable economically, which is great for the ecosystem. There are going to be all these other layers of middleware, call it: what's my memory system, and what's my harness and orchestration layer? There's going to be a very rich tools ecosystem there. The model companies will do fine. In fact, they can manage the token pricing based on their model family. If anything, I want them to work on even the KV cache standards such that we can use multiple model families.

In fact, it's better for them. I worked on Windows interoperability with Unix first.

In fact, it was counterintuitive. We used to think, "Oh my God, this interoperability means we'll be less used," except we were more used. We became, weirdly enough, more relevant because there were so many variants of Unix at that time that Windows interoperability made Unix better and Windows better. We were able to penetrate the enterprise primarily because we did that interoperability work. That's at least how I think about it.

Satya, we're in this interesting moment where, on the one hand, you have these experts asking for regulation, oversight, and governance. It typically leads to some restriction of freedom, and general society is put in a position where now we have to opine on whether this is right or wrong. But on the other side, most people's lived experience is not this magical productivity boost from AI.

At best, it's integrating our Apple Watch data to tell us why we're sleeping less. That's functionally the bar for most people. Or, "Why is my kid not into ChatGPT?" Can you just help us bridge this? You see so many enterprise applications. Where's the magic? Where are the gains in profits? Where are the huge upside breakthroughs that AI is creating that will somehow make all of this tension understandable for everybody?

Yeah, it's a great point. I think this is the real question: how do we truly see this in the productivity stats? How do we really see it in GDP growth that's broad-based, not just supply-side?

The one example that I love and get back to—in fact, healthcare is a good one—is healthcare and even the simple doctor-patient interaction. In our case, we have this thing called DAX Copilot. That's the most tangible example I can always point to: when a doctor can spend more time with the patient, caring for them, versus just entering information into an EMR system, that's a good productivity gain.

If it can triage the inbox for the doctor so that they can be more responsive, that's helpful for the patient and the care system. Even the administrator, in fact, the insurer, because it's the triangulation of the payer, the patient, and the health system. Most of healthcare is sort of all workflow costs, so taming that workflow complexity is a helpful thing.

But do you see that in Microsoft with the people that you're helping?

Yeah, absolutely. We see that. And, by the way, even in simple Copilot cases, most people think about jobs, which I think there is going to be displacement—there is—but the bottom line is, what are the new jobs that get created? That's going to be one of the key aspects of it.

But also, a lot of knowledge work, unfortunately, is drudgery. I get up in the morning and think, "Man, all I do is email triage." What if even just these workflows that are taking away time from things that you could be spending time on were eliminated?

Okay, well, you're bringing up this great point. If you go all the way back to the turn of the century, the Industrial Revolution, when we had a 7-day workweek, a lot of people forget: why did we introduce the weekend? It was to manage the tension between different religious groups that had to work in the same factory.

And when you look at long-run GDP, outside of some exogenous events, it sort of is between 200 and 400 basis points. So what happens is, as productivity boosts come in, human work steps back, and you kind of accomplish the same amount of work. Do you think that happens here? Is there a risk that we have a 3-day workweek and we're still growing at 2.5%?

Yeah, that's a great question. Or will we find new things? This is where the excitement I have for what the real impact of AI would be comes in. Instead of just thinking, hey, it has helped me augment some workflow or simplify something that's happening today, is it inventing new things? Is it speeding up drug discovery?

Let's again go back to my example of how the working-capital management of a small business has become so much more efficient. Suddenly, it's no longer just, oh, I have an ERP or a QuickBooks-like thing, but I'm truly making decisions based on the ability to introspect my invoices, my emails, and what have you, and somehow optimize my working capital. That's productivity that didn't exist. I do hope that we will start seeing GDP growth, which we did see in the industrial era during the first phase of it.

Yeah.

Right. So that, I think, is what is needed. In order for all of this to play out, quite frankly, we do need to see at least 7% or 8% GDP growth that is real and broad-based.

What's the business Microsoft is in in relation to AI? Obviously, Azure has been crushing it. You're turning away customers, and you're doing $175 billion in capex buildout, but your capex is far below what Meta is doing, far below what Google is doing. They're doing secondary raises and raising debt—$350 billion. The frontier labs are spending $500 billion.

You were so early to the party with the prescient OpenAI investment, but then Copilot didn't exactly land. I don't think it got great reviews. You don't have a frontier model. What's the business? Please come back.

No, but what's the business here? What's the—

Do you need to have a frontier model?

Did we tell you there was one journalist on the panel?

No, no, no. I mean it sincerely because I'm just curious. You're a great strategist—we know that about you. Microsoft missed the mobile revolution. Is Microsoft going to miss the AI revolution? You don't have a frontier model. I always found it perplexing that you didn't. What's the strategy there, in all seriousness? Do you think open source is going to win? You should have that play.

Yeah. So, let me walk you through where we are and what we're up to on each of these.

By the way, on the capex side and the buildout side, we started early. If you cumulatively look, it's a good—I'm not saying that speaking about a lot of capex right now is not a feature; it's a bug—but that said, if you actually add up the math, given when we started, we started multiple years before people woke up to even needing to build. That's one aspect of it.

The other aspect of it is that we are calibrating our capex in such a way that we don't want to build for 1 or 2 customers. We want to build for the long tail. That's, I think, most important. If you're a hyperscaler, you're not a supplier to 2 model companies. That's not a business. You have to build a system that is great for lots of third parties and our own in that context.

We're pretty thrilled with the progress we're making, even with Copilot. If you look at the subscriber numbers we gave, this goes back, in fact, to Chamath's fundamental point, which is these are real enterprises using it for real workflows. We now have 30-plus million—not over 4 billion, remember the total knowledge-worker base. Most people talk about 3 billion or 4 billion people on the internet. The entire Office 365, or Microsoft 365, is the standard when it comes to knowledge work. There are 450 million, including all students in the world.

When we talk about the market, quote-unquote, as defined, it's maybe 300 million, or even 250 million, real enterprise users. Of that, we've got penetration of close to 30 million, and it's growing.

On the model side, we're thrilled about our investment in OpenAI and the access we have to their IP, which we've had for a long time. We're going to use that, but we are well on our way to building our MAI models. We have a fast cyber model that, with our harness orchestrating other models, outperforms on CyberGym even a Mistral. The same thing we're seeing in coding, and the same thing we're seeing in knowledge work.

Our goal is to hill-climb from the bottom, by the way, not distilling anything—from the very bottom, using our RLs and our data—and then also have a differentiated position with enterprises. Going back to addressing some of the things that they want: Hey, can I have the weights? Can I have the weights that I can then add to my knowledge? These are the things that we will do with our foundation.

Your best advice, I think, to enterprises is that AI sovereignty is important. Putting your data into a frontier model is probably not a good idea, and then you're going to be that harness for them to help them. My advice is more like: use all, but be independent of all.

For example, my acid test is that you should always evaluate the things that matter to you. What's the outcome you want? You should run that outcome through all the models. Then here's the test I would do: I would pull out a model and see whether I can retain the eval. If I can't, that means you really are dependent on something that may or may not be yours.

Right.

Right. So my fundamental enterprise architecture would say you should have a model system that fundamentally allows you to continuously hill-climb on your own on evals while using all models—closed or open. If you want, you can even fine-tune any of these models, and you can even substitute models.

Just to build on Jason's question, you had this incredible moment, I think we put it here, where you said, you know, we're good for our $80 billion. But just to expand the question, there's effectively this sort of bank of AI that has emerged, and there's this financing mechanism that is so important to the entire ecosystem and now, broadly, to the entire economy.

You've been very disciplined. You have an enormous balance sheet. You're also an investment-grade issuer, so you could do what Jensen did, but you've taken a very different capital-allocation approach: much larger bets, very concentrated, and you've kind of stayed within your own ecosystem. Just talk us through your mindset as a capital allocator at Microsoft and with that balance sheet.

Yeah. So, the way I'm looking at our book of business—whether it's the hyperscaler, our model, or our app tier—and the shape of the demand, what's the way to build out for it?

If you think about these assets, there are 2 classes of them. There are the long-lead, long-duration assets, like the land, power, and cold shell. Then there is the kit. The kit is the short-term asset that you can much more easily be demand-driven on. In other words, I have to forecast 2 or 3 years out on demand, and then—

The kit means the racks, the chips?

The racks, the chips, and what have you. That's 60% of the cost, or whatever. Therefore, what we do is build as much as we can, lease, and even rent. Right now, we're even renting quite a bit because we were short on supply. The overall goal is to build more, lease some, and, if we really need to surge, even rent.

That's the approach on the assets. With the chips themselves, we will first of all try to make sure that we're matching demand. As I said, my goal is not to have just 2 or 3 customers. It's great to have OpenAI as one of our largest customers, and it's great that they're growing, but we need more.

Is the kit over-earning right now? Do we need the industry to push for more diversification—more silicon, more memory, more vendors?

Yeah. What's happening is that the workloads that are now at scale obviously grew up from what GPUs were, but now the shape is so well understood that you're able to optimize for a very different world. You can start building and saying, well, there are these multiple phases in an inference or a training phase, so why not build silicon that's optimized for these? That's just going to lead to a systems architecture that, by definition, I think is going to have a lot more diversity.

I know you have Jensen coming. If you look at his own architecture, it's changing quite drastically.

Quite drastically.

And so I think there's going to be a lot more choice even in that layer. We have Jensen's stuff, which is, I think, our primary thing. We have our own chips, and OpenAI is building its chip, so that's also going to be there. AMD is in there. My thing is to run the OpenAI models, the Anthropic models, or our own models on a heterogeneous kit.

Sacks, I want to let you get in here before we run out of time.

5. China's slow down, changing AI perception, data center benefits

Yeah. So, you know, we've heard now from the various frontier-lab leaders—Sam, Dario, Elon, and Demis—that we need to prioritize alignment. We're talking predictability, reliability, and robustness, as opposed to maybe just raw, raw power.

Do you think the Chinese labs will follow suit?

I think that's the dialogue that should be prioritized, because at some level, my own premise would be that China should also deeply care about the same safety concerns. If the United States cares about them, why should it be different for them? It's not like they won't have the same hacking problem.

It's not as if they don't want to make sure that their citizens are benefiting from AI, just like we want our citizens to benefit from AI. So I think there's a possibility of international norms around it. If we really are concrete about what's the risk, why is this risk so idiosyncratic that the only people who are worried about it are Americans? It doesn't make sense, right? It's not like it's something that says, "Oh, I'm only going to show up in the United States." If it's going to go wrong, it's going to go wrong everywhere at the same time. So I think the Chinese should care. I mean, they're a superpower.

Well, you use the word idiosyncratic, and I think that is the right word. I don't think we know yet: Is this conversation we're having in the U.S. over the past week idiosyncratic to us because we have the strong, I guess you could say, doomer-type school of thought, or is it something that the rest of the world will basically feel as well?

It's a great question.

And if they do, then presumably they'd want to act on it as well.

Yeah, I just feel my take there is that we are ahead.

And we are who we are, which is: We argue, we compete, and we are more transparent, which are all virtues, as far as I'm concerned. Therefore, the fact that this debate is happening here means the world will be better off for it, right? To some degree, us setting—if anything, I would love the U.S. to lead in the norms that allow us to diffuse this technology broadly and create safety standards that work for the world, including China.

What do you think we should be doing that we're not doing, and what are you doing at Microsoft to change the narrative—the populist sentiment that we have to shut down superintelligence, stop building data centers, et cetera?

So, to me, I think this is squarely focused on answering Chamath's question from earlier: Whom is it benefiting? Give me concrete stories, right? We talked about the productivity benefits a bit, whether it's in health care or general knowledge work or coding. But I'll give you another example.

I was looking at data centers because, after all, we didn't talk much today about that. But there's a challenge: How does one earn permission to open a data center in a region? In fact, we now have some of the best longitudinal data for a data center we built out in Quincy, Washington, for 20 years. 2008 is when we started it.

When I look at that data and what it has meant for that community, the tax revenues have gone up 12 times, the tax rates have gone down by a third. The growth is higher than Seattle in Quincy. This is a rural town. They have a new school, a new hospital, a new town center, and a new aquatic center. Wow.

We have 2, and most people say, "Oh, there aren't that many jobs."

In fact, there have been 1,200 construction jobs in that region all through that 20-year period, right? It's not like you just build it and leave. You continuously refurbish, build, and expand.

And how big is that data center?

I think it's now going to be at least 400 or 500 megawatts.

And it will keep expanding.

And so these are—so that's real for that community. Earning permission isn't just saying, "Hey, these are all the benefits," but seeing it.

But how do you get people to tell that story? Because that's what's missing today. Those stories aren't being organically told, and if a Microsoft executive gets onstage and says, "Don't worry, it's good for the community."

Yeah. No, I don't think so. I think storytelling is one thing. The other is that we need more people outside of the tech industry to say, "Yeah," because if you go to Quincy, Washington, they will tell you, "Thank God for this data center."

To me, that's when it's tangible, because that's the only way to earn permission. At some level, the skepticism of any of us in the tech industry just saying things is so high that I think we have to do the hard yards of actually doing things in the world, which allow people to say, "Okay, I now believe you."

It's a new muscle.

It's a new muscle. It's a new muscle.

So I think you're a good spokesperson to flex that muscle. I hope you do it more. Thank you for being with us.

Thank you so much.

We appreciate you.

Thank you, sir. Appreciate your time.