Microsoft and OpenAI compete while sharing key infrastructure

Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

1. 25:56

   Satya, we spoke about Sam and yourself. I'd be interested to hear from you. How is the relationship between OpenAI and Microsoft really? What has happened over the last years and what is the current status of it? Yeah, first of all, I would say, you know, one of the things that gives, you know, it's fantastic is from where we started. We didn't, you know, we basically. But there is competitive tension at the same time.

2. 26:21

   It's a partnership. We started as, you know, it was a research lab, which was a nonprofit. And we were able to sponsor their work and then benefit, you know, help them scale. And obviously we benefited. And so we feel very good. Obviously, there's competitive tension at the edges, but at the core of it, right, They have a very successful company at this point in which we are big investors. We have IP rights, which we then can use to innovate on top, which is what we are doing.

3. 26:49

   We have our own ambition to build our own models on top of it. We can do that. And so I feel very, very good. And by the way, in all of this, I'm glad that we created one of the world's largest nonprofits as well. Now, the OpenAI Foundation next to the Gates Foundation are two, you know, I'll say nonprofits that are associated with Microsoft. Is it actually true that Bill Gates warned you when you were about to do the first one billion investment not to invest in OpenAI?

4. 27:19

   And if so, did he apologize meanwhile? No, but, you know, Bill, you know, like all of us, like it is a crazy bet. I mean, in all honesty, to go in and say, hey, we're going to spend a billion dollars. At that time, it sounds like, oh, yeah, obviously, every CEO of a large tech company should be taking these bits. It's not that easy, right? I mean, to go and say, because it's not, it is absolutely an unknown.

5. 27:43

   And so he, like everyone else, was skeptical that the scaling laws on sort of something like text would work. And quite frankly, even in the first instances when it started working, he was pretty skeptical until I would say GPT-40. There's this very famous demo of GPT-40 that we did at his house where he saw it and he had his own tests. I think they were all the MCAT medical tests and so on that he wanted to see. And that's when I think he changed because Bill was a much more, I would say Bill's core was always more symbolic logic. He was not a neural network person, right? He just didn't believe that neural nets can just learn. And he just wanted a symbolic logic to learn the relationships between people, places, and things. And so this was the first time when I saw him cross over. He'll still

6. 28:42

   talk, by the way, if you talk to him ever, he'll still say this thing. You know, it goes back a little bit to my point about that cognitive core. So I'm sure Bill would still say, yeah, it's great, except it needs a cognitive core, which is symbolic in some sense. And so I think he's right, quite honestly, because you can't just say, hey, there's a machine that kind of does something that we are very thrilled about, but accept it's unpredictable. It has all this jagged intelligence because it's very problematic to deploy AI in mission-critical systems if it makes a mistake sometimes and you don't know when. I mean, from your perspective, it looks a little bit like, I assume, like you have two running horses. Yeah. Old man on the one end and on the other hand, and Mustafa Suleyman, with this kind of little detail that one horse has to help the other.

7. 29:34

   I have a hard time to understand how that could, in the long run, really work well. I think that the better way to maybe the metaphor is to think about it as we have a set of products, they have a set of products, and that's where we compete. But we share, I think, and the best way to think about it is the common engines, right? So it's going to take, you know, two car manufacturers that will share a chassis design, right?

8. 29:56

   Right. And there is scale economics because of the chassis design. We get to share the, you know, for example, even at the system level, we are cooperating a lot. Right. Even what we build out as the plans for these AI factories. And so, therefore, it's kind of like, hey, we'll have two brands. We'll have two products. We have, by the way, in our products, we'll use both Mustafa's models and MAI. We have a fantastic team. We're going to grow that.

9. 30:21

   But we may even use other models from other frontier models, too, because we're not beholden. We love OpenAI. We obviously have royalty-free access to OpenAI. So we want to maximize the use of that and building on top of the technology. But at the end of the day, we also have a lot of good ways to share, I'll call it, the core engine parts and the supply chain of it. Are you occasionally having dinner with both of them or is that not a good idea?

10. 30:49

    Yeah, we do. I mean, I think, interestingly enough, the company, you know, we've grown up and I've grown up in all of this, whether it's the partnership with SAP, which, you know, I was very much part of from Haasa's time and Bill and Haasa got together and said, hey, let's build SAP on top of SQL Server. That was a bit, so it was kind of like SAP built a very successful business. We built a very successful business, the Intel-Microsoft relationship.

11. 31:14

    So I'm a student of, and a cultivator or positive sum partnerships. So I think of the OpenAI thing for all of its issues which we have overlap, but think about it this way, right? OpenAI competes with everybody today, right? It competes with Google, it competes with Meta, with Sora, it competes with, they're gonna build a device or they're gonna compete with Apple. So at some level, I'm glad that OpenAI, to your point, as an investor, has got great ambition, is showing progress, but at the end of the day, Microsoft very much will be able to sort of go do what we need to do but is that really a fundamental change in an ai economy that partnerships play a different role that almost everybody is partnering with everybody and there is a completely kind of new definition sort of competition in that context?

12. 32:02

    I think that what is more important, perhaps, is what I said before, which is I think when we partner, at the end of the day, you kind of want to know, because why do people partner? People partner to get leverage, right? Because in some sense, it's always the question of vertical integration, right? Everyone can start by saying, let me produce my own electricity to the output, right? But at the end of the day, hey, specialization has a role and you could say, wow, as long as I can bring together partnerships and specialization to compose so that I can do something that's unique and I get leverage, right? That's the core thesis and it's positive sum. And so I feel like that's what needs to be found. So sometimes I think in whether it's in the tech, a lot is written about the tech industry and its partnerships, but foundationally tech industry knows how to differentiate in a

13. 32:58

    partnership. What I worry the most, quite frankly, is the next drug company, the next media company, the next auto company, right? How do they partner in smart ways with tech companies so that they remain sovereign in their capability? My biggest fear for sort of anyone who is not in the tech industry, because in the tech industry, we kind of have grown up with that DNA of how to partner and specialize and differentiate.

14. 33:30

    I fear that it's not just even about data privacy, right? It's about being able to actually orchestrate multiple of these partners to build ultimately what I described earlier, your own AI capability, your own AI factory, because that I think is going to be the key, right? Just like how we don't have an electricity industry, everybody uses electricity. AI will be, everybody uses AI to create value-added output.