Hello. Every technical leader in the enterprise is going to face a pretty critical question. As you deploy AI at scale, are you accumulating institutional intelligence or merely consuming model capability? Every expert correction, every agent trace, and every eval that you run creates an opportunity for your company to learn. But if you don't capture that learning somewhere you control and govern, it doesn't automatically become a lasting advantage for your business. In some cases, you might even run the risk that your secret sauce simply gets absorbed into a common intelligence layer that everybody has access to. So today, we're going to explore how enterprises can build a sovereign AI stack that keeps that intelligence under their control. And there's no one I'd rather explore this with than the leader of one of the largest enterprise AI companies on the planet. Someone who's been a leader in the discussion around the importance of model diversity, company sovereignty, and AI safety. Please welcome chairman and CEO of Microsoft, Satya Nadella.
Satya, thanks for joining us. As models become more capable, the context of the business becomes even more important. A model might understand accounting, but it doesn't automatically recognize how your company recognizes revenue or decides who can approve a transaction. And so to preserve that business context and the learning from AI interactions, which part of the AI stack should an enterprise own or control?
First of all, it's so wonderful to be with you and thank you for what is an amazing partnership. And I think you're getting to the crux of what I think enterprises need to do, which is what is an enterprise? At the end of the day, any company, any firm is about the knowledge they continuously create, right? It's that context that is situational, that is highly nuanced, that is accumulating. It's that tacit knowledge. So if anything, I would say the future of the firm will be always the same, which is their ability to continue to build that context layer, which manifests both as human capital as well as the digital or token capital, if you will, right? And so to me, owning the learning loop is what is most important. And to own the learning loop, you want to be very, very sort of principled in terms of even the enterprise architecture. Where is my context layer? Where is my intelligence layer? What is driving what, right? Because one of the fundamental points I sort of put out in a blog where I talked about this reverse information paradox is this is the first time where the network effects of intelligence are such that if the enterprise has not really got a deep understanding of controlling their learning and not outsourcing their thinking, they could lose control of the firm itself. So that's why I think the work you and we are doing together to ensure that the enterprises can thrive continuously even in the intelligence era is the most important thing.
Couldn't be said better. You know, every company is the data and the context that builds into the institutional knowledge that it's all about. And preserving that, enhancing that is the most important job a company has. And of course, Microsoft has one of the deepest frontier lab partnerships in the industry and amazing teams within. And yet you've built a multi-model strategy both with your own models and models from other labs in your products. And so how did you make that call for Microsoft and what advice would you have for enterprises so they too, like you, extract maximum value from the frontier models while ensuring that they're not locked in or lose their knowledge?
My early thinking was always that the model is like a database, right? And that I use to put my data in. And then I query it and I write business logic. In fact, that was, I would say, even my wrong assumption on the intelligence layer. Then I came to realize that wow, there is a network effect where the intelligence literally... In fact, I think you even, one of our sort of chats even said it, you know, intelligence eats everything it sees, right? I mean, it's sort of... And so one of the things you have to adjust is, as a firm, forget Microsoft or Snowflake, as a company. This is where I think most people are all intrigued by, oh, what's happening in the tech sector? This has got nothing to do with the tech sector. It's got to do with whether you're in healthcare, whether you're in financial services, whether you're in any firm. You have to then really say, well, this is no longer just a our database decision. This is a decision on who owns intelligence. And so once you get to that, the multi-model becomes an obvious thing. That means, in fact, you start with the outcome or the eval that matters to you, which is firm specific, not some silly public benchmark which is, you know, which all is all saturated and it's a waste of time. But a real benchmark is your firm's output that is differentiated in the marketplace. And then, you know, what you do is you, in fact, bring all the models in and have the models compete daily for a place in your rotation to create better outcomes for you, right? And then you should be able to substitute any model. In fact, the true sovereignty test for me is whether you can substitute any model and still hill climb on an outcome that matters to you. And if you can't, that means you're not sovereign. And so that's, once I, you know, we all realized it, we felt like, okay, if we are enterprise trusted vendor, you can't go in and say, oh, I'm going to lock you in. I mean, that's just not a, no enterprise will accept that. And so that's kind of where we made a principal decision to say we, even though we will, as you said, we have, you know, we have a royalty-free access to an OpenAI model for a long time, right? So we have a lot of financial incentive for that. And by the way, we're building fantastic MAI models, which is whatever just off frontier. So we have like our own code, code in-house model. But still, I don't think just hawking that as the only thing would work for any enterprise. And so that's why we said whether it's in Copilot, whether it's in Foundry, whether it's in every layer of the stack, we just wanted to be multi-model. Same thing with GitHub as well.
That's right. You touched on something in your answer just now that I think is super critical. You touched on evals and the flexibility to be able to go to a different model than a company wants. But of course, not every enterprise in the world has the same tech capabilities as a Microsoft or a Snowflake. And companies shouldn't have to start over every time a better model comes along. And whichever models they choose, the context and controls need to carry over. So in that context, how should enterprise leaders rethink their AI architecture as models give them more choice, but also make the problem of having their processes work just as effectively with the right models for their business becomes even more complicated?