Well, Satya, thank you for joining me and thank you for hosting me at your beautiful offices here at Microsoft HQ. I'm super grateful to be here, and I'm looking forward to talking to you about what the future of AI in the enterprise actually looks like. Our mission is to help companies own their intelligence by being able to train their own models and improve those the more they use them. So when you recently wrote this article that got a lot of traction on X, I think it was titled "A frontier without an ecosystem is not stable," it really resonated with a lot of the things that we think about at AC. And you know, I wanted to maybe start the conversation talking a little bit about your core argument in that piece, which is the stable equilibrium is really companies building on top of frontier capabilities rather than being attached to one or two frontier model labs. I was curious if you could maybe talk a little about what that looks like in practice, how Microsoft is approaching this ecosystem play, and what are maybe the practical things that are being built on the infrastructure side to make it—
Yeah, no, first of all, yes, it's great to have you here on the Microsoft campus, and also it's great to be partnered with you as you're building out AC and we're building out essentially this frontier ecosystem.
To me, the motivation for both that article and what we're doing stems from understanding, I would say, two things or answering two questions perhaps. One is what's the future of a firm in the ecosystem or in the economy in the age of AI, as well as what's the nature of the ecosystem and platforms with AI, right? So let's start with the first one, which is AI is a learning system, right? So it's sort of not a traditional digital technology or it's not a technology like all the other digital tools we've built. And so therefore, one has to sort of think about in a world where a model can learn anything—
What are the limits to that? Or in other words, how should the firm itself think of itself going forward? And so my simple thing is there should be as many models in the world as firms in the world, right? Because after all, what is a firm? A firm is a learning system that today is mostly about human capital with digital tools.
Every day compounding the tacit knowledge that the firm has and how it produces a good or a service that's valuable in the economy.
Right. So if that is the definition of a firm, my question is then in the age of AI, it should compound because of both its ability to create the token capital that then is being used in conjunction with the human capital, and then the compounding loop between the two, right? So that's sort of one part of it. In order for that to be true, then you do need to have a platform that is really enabling every firm in the economy to be able to create its own token capital. It can't be, "Hey, look, I have two frontier models or three frontier models or whatever, some finite set that have learned everything that is differentiated today in the economy," because then it collapses, right? So there is no stability to even the economy, and definitely the political economy won't tolerate it because you kind of need a world that is—you know, there may be very successful companies. This is not about sort of any one company's success. It is about an economy being stable because there is a lot of economic growth that is more broadspread. And so that's sort of the two-part, and that's why I wrote that, and then we can unpack that in terms of each of the layers.
Totally. Yeah. And I guess maybe drilling in on that point about the stability of the economy, what do you think just breaks down when you have all the value accruing to maybe a small set of model companies?
Yeah. I mean, if you think about it, a definition of these tech platforms at least that I subscribe to is more value has to be created above the platform for the platform to be stable, because—
It's like the definition of—
Yeah, yeah. Otherwise, you need customers in order to be successful, and so that means customers have to be successful. And that is a little bit of the simple logic on it. But more than that, right, I kind of look at this and say today, first of all, I think we've got some very cool attributes to these models. And we now know that what fuels that is, of course, compute, some great talent, but also a lot of data.
And so I sort of look at it and say, hey, we've now created a lot of value in these models using let's call it public data and great talent and lots of compute, but where is the vast amount of data still? It's all in the enterprises or in the companies. And so where the unlock, at least if you sort of—it's not zero-sum, right? It's about what is the positive-sum way for, let's say, to achieve what I think has to be the goal here collectively: can the GDP inflect to 10%, 15%, or what have you?
And I think that'll come when every firm is able to take even a frontier model or an open-weight model or whatever and then hill climb using their RL, right, which is the company itself, and in that process even build maybe even their own model or two which is much more fine-tuned to their needs and doesn't have to spend all the tokens on a frontier model, right? So that I think is where the direction of travel is.
That makes total sense, and it's a perfect lead into what I was going to ask next, which is how do you make this decision about when to own versus rent? Because I think, you know, you've said this before: people will use other people's AI, they will build their own AI, but when do you decide to do what? And from our experience at AC, we've seen four sort of different tailwinds that are kind of driving this, all kind of recent actually. One is just the token apocalypse, things are getting really expensive. Two is, you know, with maybe this most recent Claude release and the stipulations around it, data protections, government stepping in, maybe the models on the frontier APIs won't always be there. The third kind of being like labs are moving more into the application layer, and people are like, "Oh, I want to differentiate on multiple levels than just depend on frontier models." And the fourth being open-source models are starting to get really, really good.
I'm curious, and this data point was a really, really good point. Are there any other major reasons why you think companies might choose to own their intelligence?
I mean, I think you hit on most of the criteria. But the way I would think about it is maybe a firm should say, "Okay, here is, I don't know, an agentic system I want to build for my supply chain—"
Right, and optimizing the supply chain planning, let's call it. After all, supply chains for an industrial company or for anyone or a data center company like us is pretty critical stuff, right? Which is it's also idiosyncratic based on what is unique about us and how we operate and what have you. So I would say one of the things that is least talked about, although it's very critical for any agentic or AI system, is the outcome or the eval that matters for any such business process, right? And so I would start there, which is what's the outcome and my ability to grade that outcome, and start with that and then say, "Okay, I welcome all models."
Then the important thing is the "all" part, which is I don't want to be locked into any one model.
Right. I'm happy to have all the frontier models. I want to use them to see where they are on the eval.
Then I want to be able to use my own context, my own data, in fact my own traces to maybe even take a much more open-weight, cost-efficient model or a fine-tuned model which is not using general intelligence but is using the intelligence I need in order to basically hit the same eval that a frontier model had. That is the empirical way I think companies are going to go about getting value, because what that achieves is, one, you're grounded on what the result is. Second is you are also ensuring that you're not locked in to any one model.
And that means you're in control of your token economics long term. That is how I think companies can decide, which is really be outcome-oriented and continuously optimizing. There's no such thing as, "Hey, this model and this model family is the only thing that I am locked into." If anything, you're locked into one thing, which is the monotonic improvement of your outputs.
Yeah, I think that's really, really well said. And actually it hadn't occurred to me before is by simply having the benchmark, you get all of this control because you can see how things relate to one another. In the article you were talking about this idea of like a learning loop outside of your model, and I think this kind of touches on that: once you have these private evals which are very unique to the company, you can start to make your systems better. I'm curious, out of the companies that Microsoft works with, what are the sort of initial foundations of learning loops that you're seeing people start to build around their AI systems?
Yeah. In fact, one of the biggest things that we're trying to get done is, after all, we built Office 365, Microsoft 365 as this multi-tenant SaaS service in which some of the most important data/context of a given enterprise is present, right? Which is, if you think about it, all their communications data and information, their projects, their timelines, their calendars, their documents, all of that.
Is the unstructured database where all that tacit knowledge—in fact, not only the unstructured data is there, but it's the pointers to all the other of the business process, right? Because humans are communicating, creating documents, all this unstructured data is in the context of even structured processes, right? And so to me, taking that and really—in fact, we describe it as Work IQ. We want that to be the context layer along with any model, right? So when we even built Copilot, that's kind of how we built it, right? Which is Copilot takes the Work IQ context, mixes up with any model, and then helps you with the outcomes you want and the intents you want. And we are even exposing that. So for example, one of the things that we did even at Microsoft was to take the HR onboarding process.
The HR onboarding process has a system of record somewhere, but a lot of it also happens in email, in calendars, and what have you. So if we can take that context and then create an agentic system which streamlines the onboarding process and it's token-efficient, right, because the context is being used for both input tokens being input properly, and then you're not using frontier models to generate and regenerate output tokens that don't need to be regenerated. And that to me is a great way to, in some sense, create a learning loop that continuously gets better before we even have come to fine-tuning a model or what have you.
Then on top of it, you can also have an—
Exactly, right. And you can always—that's sort of where I think people will go: is have evals, have context, then have models all hill climb. And as the models are hill climbed and you say, "Oh, but I've got all this new trace data, let me take all that and tune a more compute-efficient model for this particular outcome and then keep doing the cycle continuously."
And I think one of the points you made there is actually kind of a centerpiece of the whole original article, is this idea of human capital and token capital, and that they're very intertwined and interlaced. And when I was looking at some of the reactions to your post, one of them that caught my eye was like, this is an incredibly positive-sum take on AI, which is actually not what the narrative is usually around AI—that tokens are going to eat into what humans do, and it's a little bit fatalistic that AI is just going to do everything. But your view is that humans are actually essential to making the token economy grow and these models actually do effective work. So evals is one way, but when you're thinking about these learning loops that are being built, how is Microsoft thinking about designing learning loops that are very human-centric so that they—
It's a great thing, right? Because at some level, I go back a little bit to—first of all, I think we should be clear-eyed about the change to the knowledge economy and the knowledge work, and so therefore even the displacement with something as disruptive as this.
But if I look back, let's say in the early '80s, long before you were born, if somebody had come to me and said, "Hey, three billion people are going to wake up and they're all going to be typing every day," we would have said, "What are we talking about? Like, why would the world need three billion typers?" But we did invent a complete thing called knowledge work. And it turns out that all of us in different forms, right, whether you're an accountant or a supply chain analyst or a coder, we're all doing knowledge work, right? So we kind of figured out how to use these computers to accelerate what was work that was being done in different ways. I think we are at that—we now need that same new reinvention of knowledge work. And I mean take even coding, right? Coding for the longest time, even for you and for me, it was code as input.
Suddenly code has become output, right? So in some sense, we now—but at the same time, I do think that, you know, what does a coder do today? One of the things that one of my colleagues taught me, and he's in fact written a skill, is called cognitive coverage, right? Which is the human is still in the loop. Just like test coverage in the past, humans are needed to have cognitive coverage over machine output so that you know what the machine produced, you're able to interrogate it.
Totally. That we will in fact value, that you will then need skills to do, that you will go to school for, and that'll have wage support. It may be different than what a software developer's wages in the past were; it may be more or less the same, who knows? But there is a job to be done.
And so one of my things is, hey, let's now get into this thing about where are all the human capital interventions to create more token capital.
