21 Jan 2026 · 31m · Satya Nadella, Jason Calacanis, David Sacks · 20 quoted facts
Satya Nadella argues that AI is the biggest change in knowledge work since the PC. People will act as managers of many AI agents, which they direct broadly and correct in detail, and companies will use many models rather than one. He says the real payoff comes from wide, intense use of the technology, not from owning it. He also says the American tech stack wins when others build businesses on top of it.
The central tension is between structural change and keeping the business running. Microsoft has kept headcount flat while revenue grew, merging roles into full-stack builders. Nadella must also defend Windows and Copilot quality, face fiercer rivals, and still hire young people. He treats the competition as non-zero-sum, expects open models to rival closed ones, and sees AI adoption spreading both top down and bottom up.
Nadella says coding moved from next-edit suggestions to chat, then actions, then autonomous agents running in the foreground, background, cloud or locally. Developers use all of these at once, so he expects other knowledge work to mix several forms too.
Nadella says the key organizational question is who did what to whom. Microsoft's Agent 365 extends human identity and endpoint protection to agents, so each action can be traced to either a delegating human or a separate agent identity.
At LinkedIn, product managers, designers and front-end engineers were merged into full-stack builders with broader scope. Nadella says this changes the work and the workflow, which now starts with evaluations, then science, then infrastructure.
Nadella cites economist Diego Comin's work that countries got ahead by adopting the latest technology and building on it. He measures a platform by the employment and partner revenue around it. Sacks adds that the SharePoint ecosystem earned about seven times Microsoft's own revenue.
He says Microsoft owns OpenAI's IP, yet builders will use many models and orchestrate them. A healthcare orchestrator that assigned roles to models beat any single frontier model. He compares models to databases, which diversified, and thinks each firm may eventually embed its knowledge in its own model.
Nadella says Microsoft already has the IP to OpenAI's models. He adds that its strategy rests on running huge cloud infrastructure and offering a platform where customers use many models.
It means giving an AI agent a large task, then giving it extra instructions while it works. He says coders already do this, and he sees it as the way people will manage many agents.
Jason says revenue rose by about $90 billion and income doubled over four years with similar headcount. Nadella credits structural change, such as merging product, design and front-end roles into full-stack builders who work in a new AI workflow.
Nadella says he still believes in college recruiting. AI works like a mentor that helps new hires learn a code base faster, and Microsoft is trying apprenticeships that pair a senior developer with a cohort of new graduates.
Diffusion means AI spreading into real use across healthcare, finance, government and small businesses, and across countries. Nadella says benefits come only from intense use, and the US wins when others build on its technology.
Stated by the speakers. Each quote matches the transcript, which shows what was said, not that it is true.
Satya Nadella says he gave up his US green card and got an H-1 visa so his wife could join him in the United States.
“I went to the American Embassy in Delhi, and I said, where's the line to give up my green card? And they said, there is no such line.”
Satya Nadella says his belief in this generation of AI formed when next-edit code suggestions began working accurately with a Codex model that predated GPT-3.5.
“there's a codex model back in the day, it was pre-GPT 3.5. Uh, that's when next edit suggestions started working with some real accuracy.”
Satya Nadella describes AI coding tools as progressing from edit suggestions to chat, actions and autonomous agents that run in the foreground, background, cloud or locally.
“Then we went to chat, then we went to actions, and now to full autonomous agents, and then the autonomous agents can be both, uh, foreground, background, in the cloud or local, right?”
Jason says Microsoft had about the same number of employees as four years earlier but added $90 billion in revenue and doubled its income.
“You had the same number of employees you have at Microsoft now, but you put $90 billion onto the top line of the revenue in that time, and you doubled your income during that time.”
Satya Nadella says LinkedIn combined product manager, designer, front-end engineer and back-end engineer roles into full-stack builders with broader scope.
“So what we did is we sort of took those first four roles and combined them, in fact, increased scope and said, let's they're all full stack builders.”
Satya Nadella says building an AI product starts with evaluations, which feed the science and then the infrastructure.
“It starts with Evals, right? So basically, there's this Eval to science to infrastructure, and so Evals are done by these full stack builders”
Satya Nadella cites economist Diego Comin's study of the Industrial Revolution, which found that countries got ahead by adopting the latest technology and building on it.
“one of my favorite studies has always been this work that, uh, an economist, I think, Auro Dartmouth did, uh, his name is Diego Comin, where he studied, uh, basically, what happened during the Industrial Revolution.”
Satya Nadella says public sector spending is roughly 40% to 50% of GDP in most Global South countries, which he sees as a big opportunity for AI efficiency gains.
“let's say, you know, 40%, 50% of the GDP of most Global South countries is public sector.”
David Sacks says the US should judge its success in the AI race by market share, and that roughly 80% share for American technology in five years would mean it succeeded.
“if we look around the world in five years and we see that American companies, American technology has, say, 80% market share, it means we did a good job.”
Satya Nadella says that on country visits he first studied the total employment Microsoft's partner channel created in that country.
“whenever I did a country visit, the data I would first study is in, let's say, in the UK or in Switzerland or what have you, is what is the total employment created in Switzerland, uh, in our channel?”
David Sacks says his company Yammer was acquired by Microsoft about a decade before this conversation and became part of the SharePoint group.
“about a decade ago when my company Yammer was acquired by Microsoft. We were part of, uh, the SharePoint group”
David Sacks recalls that revenue earned by non-Microsoft implementers of SharePoint was about seven times Microsoft's own SharePoint software revenue.
“the revenue from the SharePoint ecosystem, meaning non-Microsoft, the, um, the consulting community, the implementers who would go into companies and implement SharePoint. I think their revenue was something like seven times greater than Microsoft's own software revenue.”
Satya Nadella says he worked on Microsoft's database products with SAP in the 1990s, and that SQL Server together with SAP R/3 was successful for both companies.
“I worked on our database products, uh, in the nineties, uh, you know, with SAP. In fact, the combination of, uh, SQL Server and R three were successful on both sides.”
Satya Nadella says Microsoft has the intellectual property rights to OpenAI's models in response to the question of where its own foundation model is.
“where's your foundation model? I mean, at the end of the day, we do have the IP.”
Satya Nadella says Azure is Microsoft's biggest business and that Microsoft needs to excel at building 'token factories', meaning large-scale AI computing infrastructure.
“our biggest business today is Azure business, and the Azure business, the TAM given what's going to happen, is is so huge that we now need to be fantastic, uh, at building these token factories.”
Satya Nadella says Microsoft's healthcare decision orchestrator, which assigns roles such as investigator, data analyst and domain expert to multiple models, outperformed any single frontier model.
“What it proves is that by assigning roles, right? So, investigator, data analyst, domain expert, just giving even prompted roles to models, and then orchestrating them gets better results than any one single frontier model.”
Satya Nadella predicts a debate over firms embedding their tacit knowledge in the weights of models they control.
“A firm should be able to take the tacit knowledge it has, and embed it inside a weights in a model that they control, right?”
Satya Nadella says Microsoft has a locally resident model called Phi Silica that runs on a PC's NPUs and GPUs.
“there's a Phi Silica model, which is completely resident, using NPUs and, of course, using, uh, GPUs.”
Satya Nadella says Microsoft works with about 500 fiber operators around the world for Azure's network.
“we manage something like 500 odd, uh, fiber operators around the world, and and in Azure today.”
Satya Nadella says Microsoft is experimenting with an apprenticeship model in which a cohort of college hires works alongside a senior individual-contributor developer.
“one of the things we're experimenting with is a different type of apprenticeship, right? Which is you take somebody who's an IC senior dev, have, like, a cohort of college, uh, hires working with them”
Nadella explains how he gave up a green card so his wife could join him. The hosts then turn to how knowledge workers will use AI.
“I had to give up my green card.”
Nadella uses coding to show how AI moved from suggestions to autonomous agents. He offers 'macro delegate and micro steer' as the way to manage them.
“we macro delegate and micro steer”
He explains why agents need identities and permissions so organizations can trace the work to a source.
“who did what to whom is sort of the most important query in an organization”
Nadella describes merging roles into full-stack builders and a workflow that starts with evaluations. He ties it to Microsoft's growth on flat headcount.
“we sort of took those first four roles and combined them”
He says new rivals each decade keep Microsoft sharp. He expects tech to be a larger share of GDP.
“a complete new set of competitors every decade because that keeps you fit”
Nadella and Sacks argue that AI benefits come from broad use. They say US success is measured by market share and by the ecosystems built around US technology.
“you create the technology, but really the benefits come only by intense use”
Nadella explains Microsoft's token-factory and app-server strategy and why customers will use many models. He also describes plans for local models on the PC.
“anyone building any application or any company is gonna use not one model”
He says adoption will be both top down and bottom up, with employees building agents to remove drudgery. Skills come from using the tools.
“I think it's both the top down, bottom up”
Nadella argues AI steepens the learning curve for new hires, and Microsoft is trying apprenticeships. He says the company remains committed to hiring graduates.
“of a college hire is gonna be much steeper than it ever before”
Alright everybody, we're thrilled to have the one, the only Satya Nadella here, the third CEO of Microsoft for a, uh, impromptu fireside chat with David Sacks, our czar of AI and crypto. Satya, third CEO of Microsoft, born in India, what an incredible story, came here right after college, and, uh, you had a little round trip to pick up your wife and your book to to bring her here. Tell everybody briefly, uh, how that occurred.
Well, um, uh, you know, so that's a that's a great story of, uh, the, uh, the labyrinth that is the immigration policies of the United States, I think. Um, I, uh, but my wife and I went to college together in India. I came here for grad school, we then got married. Uh, I got my green card, um, and she couldn't come join because we got married. So the the story goes, basically, I had to give up my green card.
So the funny thing is, I went to the American Embassy in Delhi, and I said, where's the line to give up my green card? And they said, there is no such line.
That would be a crazy thing to do in the nineties.
So it was a strange thing to give up your green card, get an H-1 so that she could join, but it all worked out. So, um, you know, it's, uh, long lost memory, but it was, you know, a way to work around it.
Um, I wanted to ask you, uh, having launched a co-pilot first with GitHub, then having a co-pilot on the desktop, you made a very bold move for Microsoft to put that in the Windows product, which I use every day on the desktop. But you did that before it really could recognize the file system and interact with applications. Got a little bit of a lukewarm reception, but now you've been doubling down, doubling down, and there seems to be, in my estimation, three modalities for knowledge workers.
Elon's building at XAI, what they're calling a human emulator, if you saw that leak this week, yeah, uh, where they're just building employees and just putting them into their chat rooms and email. Then you have Claude came out with co-work this week, incredibly powerful, people are kind of losing their minds over it. I've been playing with it for the last forty hours, truly impressive. What's your vision for Microsoft and how knowledge workers will actually put this to use?
Cause there seems to be a gap between, you know, playing around with ChatGPT and getting some interesting results and getting business results.
Yeah, so I think it one of the most, uh, perhaps illustrative examples, um, of trying to understand these various form factors, is looking at coding, which is obviously, uh, a form of knowledge work, or, uh, probably the best example of knowledge work. And if you think about the journey coding has been, it started with, uh, essentially, uh, uh, the next edit suggest, right? That was the first time, in fact, my own belief in this entire, uh, generation of tech really sort of got formulated when I started seeing.
I think it was G, you know, there's a codex model back in the day, it was pre-GPT 3.5. Uh, that's when next edit suggestions started working with some real accuracy. Then we went to chat, then we went to actions, and now to full autonomous agents, and then the autonomous agents can be both, uh, foreground, background, in the cloud or local, right? So that's all the form factors that exist today when you're coding. And interestingly, you know, if you look at it, you use all of them, right?
It's not like there's only one form factor, so that's, I think, probably one of the other lessons. So, for example, when I'm in a CLI, I can you go a foreground, uh, agent, background agent, and then just literally go edit in VS Code right there, all happening in parallel, right? So that sort of shows how these form factors even compose. So then you bring that to knowledge work, to your point, we started with chat, chat with reasoning, sort of goes beyond just request response because you now have that chain of thought, uh, where you can see it work.
Now, they're actions, right, essentially either through computer use or through, uh, API, you know, basically skills, uh, and agent calls, so you can do actions. So that's kind of the state of the co-pilot today. Now, there is a way to think about, you know, the the theory of the mind evolution, right? Because you need, like, if you remember, you know, Jobs had the best line, I would say, for PCs or computers was to say, if you, it's a bicycle for the mind.
Uh, Bill had a line, which I liked as well, which was it's information at your fingertips. We kind of need now a new concept metaphor for how we use computers in the AI age. And
You have one?
A and the one I like actually came from the CEO of Notion, which I lo, you know, that manager of
Incredible product. You haven't bought it yet.
I've not bought that. Uh, but the uh, it's both, you know, management, uh, you know, basically, um, uh, a manager of infinite minds. That's a nice way to think about it, right? When you sort of really look at all the agents that you are working with, you kind of need to understand what I, in fact, the other term I like is we macro delegate and micro steer. In fact, you kind of need that in in coding, you kind of have it, right?
So you do a macro delegation, and then I can, in parallel, give it instructions while it is doing work. So that's sort of the state, even today, of co-pilot or what have you. You bring up a little bit of one of one of the form factors I'm very excited about, and you'll see us even in the next week even, uh, do things, is while I'm sitting in GitHub co-pilot. Wi i i it's not as if software developers sit in isolation, right?
It's not like the only thing I work on is my repo. I attend meetings, um, I write specs, or others have written specs that I'm implementing, uh, I need to have my repo be consistent with that. So that means using either a straightforward MCP server or a skill, I want to be able to call into my work IQ, which is the co-pilot, bring that in. That's the type of composition, uh, of knowledge work that'll happen. Same thing with security.
Say you're a security professional, you have lots of logs, uh, how do you sort of really analyze them? You drop them into a file system, then write code on top of it, create a dashboard, what have you? Those are the types of knowledge work that we can enable. There, I think you bring up one more thing, which is, can you create quote unquote digital employees, digital coworkers, or what have you? And it's all about credentials, right?
So the I today you could, like, you can literally assign
Are you working on that as well?
Yeah, so in fact, we introduced something called Agent Three Six Five as a way to give identities, in fact, extending the identities we have for humans today, uh, and the endpoint protection we have for their compute devices to agents. So
So you might clone me working in the HR department or working in the marketing department and have a virtual Correct. version of me inside of Office.
That's correct. So so there are two sort of modalities there. One is you give every knowledge worker infinite minds, that's kind of one, and then you create even infinite minds independent of the your identity, because the identity is one of the key things you gotta get right, even for it to work, right? So, For permissions for for, uh,
and decision making.
permissions, decision making, and like, one of the key in things is who did what to whom is sort of the most important query in an organization, right? At the end of the day, the organization needs to understand what work got done, ho, and what's the provenance of that work, uh, and how do you trace it back, right? So therefore, you kind of want either, uh, if it's a human with a lot of agents, then it's really macro delegation, micro steering by the human whose identity was passed on.
So it's delegation versus a separate identity.
And that was done by a level of management, product management, that you've eliminated, that alphabets eliminated, meta has started to eliminate in their organization four years ago. You had the same number of employees you have at Microsoft now, but you put $90 billion onto the top line of the revenue in that time, and you doubled your income during that time. So, how did that happen? Is that automation of those jobs? Is it you were a little bit overstaffed?
Unpack that.
I think it's, uh, it's actually you, you're pulling on a very interesting thread, which is, at some level, what's the big structural change that needs to happen? In fact, I would say this is probably the biggest change in knowledge work since PCs. I mean, I always, you know, think about, like, how did work happen pre-PCs, right? I mean, think about a multinational company like ours trying to do a forecast, uh, right? Faxes went around, inter office memos got sent, and then you kind of created a f, you know, a a forecast.
Then suddenly, you know, PCs became standard issue, uh, you put an Excel spreadsheet, put some numbers, sent it in email, everybody entered numbers, and you had a forecast. So the work, the work artifact, and the work flow all changed. That's what's happening. So, for example, I'll give you, uh, at LinkedIn, uh, we used to have product managers, we had designers, uh, we had front-end engineers, and then we had sys back-end engineers and so on. So what we did is we sort of took those first four roles and combined them, in fact, increased scope and said, let's they're all full stack builders.
So I like that because that's a structural change that allows for us to increase the change, both the work and the workflow between these functions.
And I would assume the velocity because you don't have four people communicating and that throughput of ideas, it's just one person in vibe coding.
Exactly. And there's a new workflow. So, while at the same time, as you can imagine, if to build an AI product today, there is a complete new workflow, right? It starts with Evals, right? So basically, there's this Eval to science to infrastructure, and so Evals are done by these full stack builders and what have you, and product managers in the new form. The infrastructure is built by the systems engineers at the back end because they support the science that supports the product.
So, in some sense, there's a new loop, uh, and you have to structurally change. And so a lot of what is happening inside of tech is that change, uh, which is, I think, going to be pretty massive. Uh, and at the same time, a company like ours, I have to do everything. It's not like I can just so go live in the future. I have to make sure we're doing a fantastic job of doing hot patching on Windows is done with quality, uh, while at the same time building the Evals that are improving Copilot quality, right?
And so both of those have to be first class.
I assume this is the most challenging moment of your career because Microsoft was so dominant, duopoly in some spaces, um, but you really weren't up against the competition level you're up against now. I was talking to Elon, you know, and he was sort of saying, well, building cars was pretty easy, uh, because I was up against the legacy car makers, and now I'm up against, just look at the set you're up against.
Yeah, it's a it's a pretty intense time. I mean, so the way I I I always think is it's always helpful, uh, when you have a complete new set of competitors every decade because that keeps you fit. Uh, if you Elon think about it, I joined, uh, Microsoft Scaramella. in ninety-two when I had Novell as the big existential competitor we had, um, uh, and here we are in 2026, uh, and it you're absolutely right, it's a pretty intense time.
I'm glad there's the competition. Uh, it's it's quite honestly, at the end of the day, when I look at it, right? As a percentage of GDP, five years from now, where will tech be, right? Uh, it will be higher. So we're blessed to be in this industry. It's a lot of intense competition, but it's not so zero-sum as some people make it out.
Pie is getting much bigger.
Much mo- the TAM and the, you know, just the impact of this tech is gonna be so massive. Um, the question then, of course, is, what is, like, I I'm always go back to, what's the brand identity Microsoft has, brand permission we have, what do customers expect from us? It's sometimes we kind of oo- w- overthink somehow that every customer wants the same thing from all of the competitors, and finding that out, right? It's kind of a different take on the Peter Thiel thing, which is you gotta avoid competition by really understanding what customers really want from you, uh, versus thinking everybody's a competitor.
David?
Yeah, so there are a lot of heads of state here, obviously, at Davos, as well as CEOs of Fortune 500 companies. And I think you got asked a question last night at at the dinner about how they should think about AI and how to be successful. And I recall that you used the word diffusion, and I was wondering if you could expand on those remarks because that really resonated with some of the policy work I've been doing.
No, absolutely. In fact, uh, what you all have been doing to make sure, in in this context of the American tech stack, um, is broadly used around the world and is trusted around the world. Because I think, uh, when I look back, David, to me, um, uh, at the end of the day, you create the technology, but really the benefits come only by intense use. In fact, one of my favorite studies has always been this work that, uh, an economist, I think, Auro Dartmouth did, uh, his name is Diego Comin, where he studied, uh, basically, what happened during the Industrial Revolution.
Um, how did countries get ahead? Um, and the simple sort of takeaway from that was any country, uh, that brought the latest technology into their, uh, country and then did value-add technology on top of it, right? So it's like, don't reinvent the wheel, bring the latest, and then build on top of it. That's to me, uh, what happens, you know, when you have diffusion. So, especially with general-purpose technologies like AI, it needs to spread, like, right in our, you know, in our own country, in the United States, we now need, we have the tech.
The question is, is it being used in healthcare? Is it being used, um, in financial services? Is it being used in every sector of the economy by large businesses, small business, public sector? Um, so to me, under unless and until we see that diffusion and intense use, uh, we're not gonna have the success. Um, uh, and so that's the phase we are in. It's fa- you know, it's diffusing faster. Um, and so some of the work, policy work you have done, um, and in general, all you know, the good news here is the technology is there, the rails around cloud and mobile that were laid out make it possible for this thing to spread, right?
It's not, you know, impossible to get the tokens. The question is, what are the use cases? How do, and how do you manage the change and all of that? Um, you know, like, one of the questions, at least in Davos, is it's one thing for the West and the developed nations, what about the global South? Um, I think Global South has a huge opportunity too, quite frankly, because to me, like, let's say, you know, 40%, 50% of the GDP of most Global South countries is public sector.
So just imagine this tech making a difference in how the governments, uh, really parlay their taxpayer money into services for citizens. And there's, if there's efficiency gains, that's probably couple of points of GDP growth right there. And so I'm very optimistic that there's gonna be a pull, uh, and that we should, as the United States, given the technology stack we have, uh, in Europe, in Asia, in, you know, in South America, in Africa, and everywhere else, uh, get it to be broadly deployed.
One of the questions I get asked a lot about the AI race is, how do you know if you're winning? Or how do you know if the United States is ahead of its global competitors? And the answer I give is market share. You know, if we look around the world in five years and we see that American companies, American technology has, say, 80% market share, it means we did a good job. If we look around the world in five years and see that it's, say, Chinese chips and Chinese models that are being used all over the world, well, it means we probably lost.
So, you know, ultimately, usage is the proof of the pudding is in the eating of it. I mean, the in this case, the way that you know that you're succeeding is through market shares, through usage.
I and I I would agree with that, but David, since you even worked at Microsoft for a few years, um, you, you re, you know, one of the things that I'm very grounded on is always, uh, that Bill Gates line of a platform, right? So one of the things that I always think about is it's market share, but it's also ecosystem effects, right? See, what the United States always has done is not just about our market share or even, um, the revenues to U.S.
companies. In fact, one of the things I learned at Microsoft is whenever I did a country visit, the data I would first study is in, let's say, in the UK or in Switzerland or what have you, is what is the total employment created in Switzerland, uh, in our channel? That used to be like the number one thing, uh, in our country reports, right? And the total number
Exactly like the number of IT workers, the number
IT. W- Exactly. of office We used to have, see, workers. channel, so channel partners, we used ISVs, so the number of ISVs, uh, who were there. So we used to have a complete marker of how did the ecosystem around the platform get built, one country at a time. And that is what the United States has always done. In fact, the U.S. tech stack, including in China, got built because others built around our tech stack. The same thing is gonna happen, so that's why I think the work you're doing around diffusion is about really increasing the size of the pie, the trust in the platform, so that there is true economic opportunity, quite frankly.
Well, you're right, and and I remember actually you're, you brought back some memories from this is about a decade ago when my company Yammer was acquired by Microsoft. We were part of, uh, the SharePoint group, and I remember that the, um, the product managers there were very proud of the fact that the revenue from the SharePoint ecosystem, meaning non-Microsoft, the, um, the consulting community, the implementers who would go into companies and implement SharePoint. I think their revenue was something like seven times greater than Microsoft's own software revenue.
And I think
In aggregate.
it in aggregate, and I think, uh, and I think Bill had a line about you're not an ecosystem or a platform until the revenue on top of your platform is
That's right.
some, you know, factor of your own revenue.
That's right.
And and I, and I think that I think what's really important about this is when we talk about diffusion, and obviously we want the United States to have this leading position. It doesn't mean as bad for the rest of the world because they're able to build on top of those platforms
100%.
and create even more value.
100%. In fact, that's sort of the most important point, right? So this is not that this is not about, uh, American tech, um, and Ameri- revenues to the United States. It's actually creating opportunity using a new platform everywhere. And, in fact, you know, the W- you know, like, I remember I worked on our database products, uh, in the nineties, uh, you know, with SAP. In fact, the combination of, uh, SQL Server and R three were successful on both sides.
A lot there's a lot talked about Intel and Microsoft, but one of the other things, uh, that I grew up in, which has sort of been foundational in how I look at the world, is what we did with a European software company that is still, uh, you know, a giant. And so that, you know, who knows what the next big AI app will be and where and what will happen. But, uh, I I sort of go in with the attitude that there will be tech companies, uh, maybe even top five tech companies that could emerge everywhere with even the American tech stack.
You have, um, done some amazing acquisitions, and you're quite a dealmaker on top of being a technologist. That's probably the least reported aspect of your spectacular tenure and the massive growth you've had. But you did a deal with OpenAI, and probably one of the most savvy slash controversial dealmakers of all time, Sam Altman. That deal was looked at as you you you're you're set up to get a windfall in cash, which you don't need as Microsoft, always nice, I'm guess guessing if they IPO.
But did you create potentially, and this was the criticism of it, an ultimate competitor to Microsoft? And how do you think about that? And how can Microsoft, which missed Steve Ballmer's biggest regret, missing the mobile revolution, how can you not have a Gemini, an XAI, a Claude that is your own, or in your mind, do you have that, cause you have the source code of OpenAI?
Yeah, I think that that that's right. So, when when people say, uh, where's your foundation model? I mean, at the end of the day, we do have the IP. But that said, I think you bring up a couple of different things, right? One is to us, the most important thing when I look at what is Microsoft's, uh, strategy today. One is we wanna build token factories, right? So, our biggest business today is Azure business, and the Azure business, the TAM given what's going to happen, is is so huge that we now need to be fantastic, uh, at building these token factories.
And, um, that's means a heterogeneous fleet of infrastructure, and that every hyperscaler has always done, which is use software to make maximum use of it, and for TCO and utilization. So that's one side of it. Then there's the app server business, right? Which is, if everybody, we you talked about, like, if everyone's gonna be building agents, have infinite minds, have these RL gyms, have Evals, what have you? There's an entire just like every platform has had an app server, this one has an app server.
That's what we're doing with Foundry and what have you, right? So there's an app server business. In that app server, one of the things that structurally now is pretty clear is anyone building any application or any company is gonna use not one model, but all the models, right? Why would I not? Right? Which is, in fact, I will orchestrate for any given task, even multiple models, right? There's this one nice thing that we came out in our healthcare practice called the decision orchestrator.
What it proves is that by assigning roles, right? So, investigator, data analyst, domain expert, just giving even prompted roles to models, and then orchestrating them gets better results than any one single frontier model.
Am I right to read into that then that you're bullish on the open source models and think large language models will largely be commoditized, and that's not where the value will accrue?
In fact, the way I think about it is that, just like I think Tim what happened in Cook and Apple the.
thinks that too, by the way.
By the way, w-what the way you think about what happened in the database market, right? You know, I used to be like, Oh, everything is just a SQL database until it was not, right? There was, I mean, think about it, there, NoSQL databases, there is NoSQL databases, the proliferation of databases, right? Who would have thought that the database market would have such a richness to it?
Or that it could ever be open source, that was
That's, that's true, too.
my love.
I mean, talk about Postgres, or what has happened even with Mongo, which is open, but there are even companies that are backed. And so so, to me, that's what's gonna happen. I to me, a model is like the database market. You know, it's it's got, it's going to differences, but I sort of somehow think that, uh, it's not, there are definitely going to be frontier models that are closed source. You know, there're going to be open source models that are going to be, uh, uh, frontier class.
In fact, if anything, I think in this next year, what'll be probably a big part of the discussion is what's the future of a firm? A firm should be able to take the tacit knowledge it has, and embed it inside a weights in a model that they control, right? So, when somebody asks me how many models should be there, I'll say as many models as firms in the world, right? That's sort of the an extreme way, uh, because because to me, that's how I think this, you know, this knowledge economy becomes an AI economy.
Are you secretly, and you can say it here since we're on all in, working on an LLM to exist on the Windows desktop because that you are?
We do have it. Like, today, there's a Phi Silica model, which is completely resident, using NPUs and, of course, using, uh, GPUs. In fact, the largest installation, um, of high power, in fact, it's one of the fascinating, the workstation is back. I'm one of the most, if you went to Which is great CES, it's so for Microsoft awesome. Yeah,
because you you have a nice desktop business.
uh, absolutely. And so we and, in fact, we think that that form factor, especially, I mean, I, I, I always say this, which is, uh, you know, I started my career on a command line. Who knows, I may just end it in a command line. Well, you But, you know,
started at Sun, which was the original five ten thousand dollar workstation. Do you see a time where you'll be meeting with your customers here and advocating a ten twenty thousand dollar desktop machine that has an LLM and the hardware I mean, to support you can, like, it?
you can put a DGX card and you can have, like, just a fantastic machine. And the models I, and, by the way, you know, we are one architecture tweak away from even having some kind of a distributed model architecture, right? Even an MOE architecture that shows knows how to really distribute itself, right? That's the type of breakthrough that can completely change, uh, what hybrid AI may look like. But we're absolutely committed and focused on making the PC a great place for local models, uh, and local models that then do even a lot of the prompt processing and call into the cloud, right?
So, there's a whole lot of work that can happen, and that's sort of definitely something that's on the way.
Yeah, I think that the cloud co-work has kind of shown the power of tapping into the local file drive and be able to use that. That that brings up another point, you you got me thinking about Yammer, and for people who don't know, um, you know, Yammer's claim to fame, this is about fifteen years ago, was that it pioneered a lot of, um, well, it used a lot of consumer growth tactics to attack enterprise software. I'm wondering, as you think about enterprise adoption of AI, how do you think it's gonna spread over the next year?
It feels like we're at sort of a a critical point. Do you think it's gonna be top down? Is it gonna come from the CEO directing a team, giving them a strategic transformation project, and they're gonna do an RFP? Or do you think it's gonna spread bottom up in the enterprise through AI native employees who are adaptable, who are using the tools in their own lives, and they start to bring these things to work and, uh, c- start accomplishing amazing things.
Yeah, no, I think, you know, like all things, David, I think it's both the top down, bottom up, right? Uh, the the reason I say that, uh, top down is if I look at the ROI, uh, of, uh, applying AI in customer service, uh, or in supply chain, or in HR self-service, those are the easy projects where, uh, IT and CXOs can make calls, and that's where you're seeing the first drop of, uh, real AI adoption.
But the bottom up is what ultimately will happen, right? I mean, that even with the PCs, in fact, if you think back at it, the lawyers brought Word in, and then, uh, finance bought Excel in, and then email came, and then it became standard issue. That's what's happening right now. So, for example, these agents, when I sort of talk about everybody's building agents, they're figuring out a way to go create these things that are changing workflow and removing drudgery in their work, right?
That's sort of the beginning of what is a bottom-up transformation. Um, I, you know, you, I was, uh, in fact, the thing that I'm most excited about is this bottom-up change. Even at Microsoft, for example, we manage something like 500 odd, uh, fiber operators around the world, and and in Azure today. And, by the way, I'd not myself realized it, a lot of it, you know, it's called DevOps, but it's a it's a physical asset.
Things get cut, and when you sort of say DevOps, that means you literally are emailing people and saying, Hey, what happened to that fiber cut? How do we repair it? So, there's a lot of back and forth. So, this network, the the person who runs our global network, basically has built, to your point about these, person, they're just digital employees, essentially, that are doing all of that DevOps, uh, and so that's, and there's a completely bottoms up, uh, where you see the tools, it's kind of like, Hey, I have the new way to build agents, it's there, I'm gonna use it to create levels of automation, uh, that remove drudgery, improve efficiency, improve quality.
And that ultimately is a skilling thing, which is sort of the big issue, which is, um, and skilling is not mystical, it's just by doing, right? So, it's not like I go to a class, per se, it's like the diffusion of the tools, uh, and using the tools, and that, I think, is what's really going to be happening.
And and we're in a very interesting moment. Empowering an existing employee with these tools is so much easier than hiring and mentoring and bringing up the next generation. So it feels like we're in a little bit of an indigestion moment. At Microsoft, do you think, who's gonna have my job in thirty or forty years if the company stays the same size? Because given your technology first approach, there's really no reason to ever add another Microsoft employee at the pace this is going, and you haven't for four years, so how, you may have swapped some in and out and changed the texture of it.
So how do you think about maybe this next generation, what advice would you have for these college graduates who maybe don't have an offer from Microsoft right now? And you used to spend a lot of time on that building that group, but maybe you don't have that luxury now. Do you think about it ever?
No, I, I mean, it's a great question. I, you know, there's a little bit of a debate what happens to early in career and how is college recruiting. I still am a big believer in, uh, college recruiting because at the end of the day, um, this is going to change the curve by which anyone can pick up proficiency in a code base, let's just. It takes sort of just regular CS hiring. Uh, what has changed is perhaps for someone who comes in new into a team, and to be able to ramp up thanks to all of, uh, the markdowns, the skills, uh, the fact that I can go ask the agent.
I mean, think about it, right? It's like having an unbelievable mentor who is getting you onboarded onto a code base faster. So, in some sense, the productivity curve, uh, of a college hire is gonna be much steeper than it ever before. So, I think there might be a difference. In fact, one of the things we're experimenting with is a different type of apprenticeship, right? Which is you take somebody who's an IC senior dev, have, like, a cohort of college, uh, hires working with them because it's a new way of working.
It's like, I remember, like, all, you know, everybody who joined Microsoft would say, go, how how did, uh, you know, whatever, um, uh, Cutler implement Malick or what have you, right? He would go try to read, uh, his code to understand, uh, what great craftsmanship looks like. Nowadays, I think that great craftsmanship, uh, comes by looking at even how the ten x hundred x engineers use AI to build great quality products, uh, and that is what these new college grads will learn and learn faster.
And so that's a beneficial thing for a company like ours. Because at the end of the day, you know, until we solve longevity or something, we need people to come into the workforce, be successful at Microsoft. So, we are very committed, but we are also making sure that the scopes of the jobs make sense for what the aspirations of people are going to be, both who are currently in the workforce and people who are entering the workforce.
Okay, on that note, Satya Nadella.
Thank you so much.