WEBVTT 00:00:00.240 --> 00:00:03.757 Yeah, I think this has really been an amazing partnership through every phase. 00:00:03.840 --> 00:00:05.917 Uh we had kind of no idea where it was 00:00:06.000 --> 00:00:11.117 all going to go when we started as Satia said. Uh but I I don't think I think 00:00:11.200 --> 00:00:15.037 this is one of the great tech partnerships uh ever and without 00:00:15.120 --> 00:00:16.317 certainly without Microsoft and 00:00:16.400 --> 00:00:22.520 particularly SA's early conviction uh we would not have been able to do this. 00:00:23.580 --> 00:00:25.600 [music] 00:00:32.160 --> 00:00:33.997 What a week. What a week. Great to see 00:00:34.080 --> 00:00:37.437 you both. Um Sam, how's the baby? 00:00:37.520 --> 00:00:41.677 >> Baby is great. That's the best thing ever, man. Every every cliche is true 00:00:41.760 --> 00:00:43.677 and it is the best thing ever. 00:00:43.760 --> 00:00:47.277 >> Uh hey Sacha, with all your time >> smile on Sam's face whenever he talks 00:00:47.360 --> 00:00:49.757 about uh it's just his his baby is just 00:00:49.840 --> 00:00:53.117 so different. It's dad that and compute 00:00:53.200 --> 00:00:57.597 I guess when he talks about compute and his baby. [laughter] 00:00:57.680 --> 00:01:01.277 >> U well Sachi have you given him any dad tips with all this time you guys have 00:01:01.360 --> 00:01:03.917 spent together? >> I said just enjoy it. I mean it's so 00:01:04.000 --> 00:01:06.317 awesome that uh you know I you know we 00:01:06.400 --> 00:01:08.637 had our babies or what our children so 00:01:08.720 --> 00:01:13.357 young and I wish I could redo it. So in some sense it's just the most precious 00:01:13.440 --> 00:01:18.477 time and as they grow it's just so wonderful. I'm so glad Sam is um 00:01:18.560 --> 00:01:21.757 >> I'm happy to be doing it older, but I do think sometimes, man, I wish I had the 00:01:21.840 --> 00:01:25.357 energy when I was like 25. Uh that part's harder. 00:01:25.440 --> 00:01:28.877 >> No doubt about it. What's the average age at Open AI, Sam? Any idea? It's 00:01:28.960 --> 00:01:32.477 young. >> It's not crazy young. Not Not like Not 00:01:32.560 --> 00:01:35.757 like most Silicon Valley startups. I don't know, maybe low 30s average. 00:01:35.840 --> 00:01:39.197 >> Are babies t is it are babies trending positively or negatively? 00:01:39.280 --> 00:01:40.397 >> Babies trending positively. 00:01:40.480 --> 00:01:41.997 >> Oh, that's good. That's good. Yeah. 00:01:42.080 --> 00:01:45.997 >> Well, you guys, such a big week. You know, I was thinking about I started at 00:01:46.080 --> 00:01:48.877 Nvidia's GTC, you know, just hit $5 00:01:48.960 --> 00:01:53.517 trillion. Google, Meta, Microsoft, Satcha, you had your earnings yesterday, 00:01:53.600 --> 00:01:57.197 you know, and we heard consistently not enough compute, not enough compute, not 00:01:57.280 --> 00:02:02.157 enough compute. We got rate cuts on Wednesday. The GDP's tracking near 4%. 00:02:02.240 --> 00:02:05.597 And then I was just saying to Sam, you know, the president's cut these massive 00:02:05.680 --> 00:02:08.317 deals in Malaysia, South Korea, Japan, 00:02:08.400 --> 00:02:12.797 sounds like with China. you know, deals that really incredibly provide the 00:02:12.880 --> 00:02:17.197 financial firepower to re-industrialize America. 80 billion for new nuclear 00:02:17.280 --> 00:02:21.117 fision, all the things that you guys need to build more compute, but 00:02:21.200 --> 00:02:24.877 certainly wasn't what wasn't lost in all of this was you guys had a big 00:02:24.960 --> 00:02:29.357 announcement on Tuesday that clarified your partnership. Congrats on that. And 00:02:29.440 --> 00:02:32.397 I thought we'd just start there. I really want to just break down the deal 00:02:32.480 --> 00:02:34.797 in really simple plain language to make 00:02:34.880 --> 00:02:37.517 sure I understand it and and and others 00:02:37.600 --> 00:02:39.517 but you know we'll just start with your 00:02:39.600 --> 00:02:42.317 investment Satcha you know Microsoft 00:02:42.400 --> 00:02:45.277 started investing in 2019 has invested 00:02:45.360 --> 00:02:48.077 in the ballpark at 134 billion into open 00:02:48.160 --> 00:02:51.997 AI and for that you get 27% of the 00:02:52.080 --> 00:02:55.757 business ownership in the business on a fully diluted basis I think it was about 00:02:55.840 --> 00:02:59.517 a third and you took some dilution over the course of last year with all the 00:02:59.600 --> 00:03:04.317 investment So, does that sound about right in terms of ownership? 00:03:04.400 --> 00:03:06.477 >> Yeah, it does. But I I would say before 00:03:06.560 --> 00:03:08.877 even our stake in it, Brad, I think 00:03:08.960 --> 00:03:12.237 what's pretty unique about OpenAI is the 00:03:12.320 --> 00:03:15.917 fact that as part of OpenAI's process of 00:03:16.000 --> 00:03:20.317 restructuring, one of the largest nonprofit gets created. I mean, let's 00:03:20.400 --> 00:03:24.397 not forget that, you know, in some sense I say at Microsoft, like I, you know, we 00:03:24.480 --> 00:03:27.277 are very proud of the fact that we were, we're associated with the two of the 00:03:27.360 --> 00:03:31.437 largest nonprofits, the Gates Foundation and now the OpenAI Foundation. So, 00:03:31.520 --> 00:03:35.997 that's I think the big news. Uh, we obviously were, you know, are thrilled. 00:03:36.080 --> 00:03:39.757 It's not what we thought. And as I said to somebody, it's not like when we first 00:03:39.840 --> 00:03:43.277 invested our billion dollars that, oh, this is going to be the 100 bagger that 00:03:43.360 --> 00:03:47.357 I'm going to be talking about to VCs about, but here we are. But we are very 00:03:47.440 --> 00:03:50.237 thrilled to be an investor and an early 00:03:50.320 --> 00:03:54.717 backer. Um and and it's a great and it's a really a testament to what Sam and 00:03:54.800 --> 00:03:58.477 team have done quite frankly. I mean they obviously had the vision early 00:03:58.560 --> 00:04:03.517 about what this technology could do and they ran with it and just executed you 00:04:03.600 --> 00:04:06.877 know in a masterful way. >> Yeah. I think this has really been an 00:04:06.960 --> 00:04:08.877 amazing partnership through every phase. 00:04:08.960 --> 00:04:11.037 Uh we had kind of no idea where it was 00:04:11.120 --> 00:04:16.237 all going to go when we started as Satia said. Uh but I I don't think I think 00:04:16.320 --> 00:04:20.157 this is one of the great tech partnerships uh ever and without 00:04:20.240 --> 00:04:23.357 certainly without Microsoft and particularly Sant's early conviction uh 00:04:23.440 --> 00:04:24.717 we would not have been able to do this. 00:04:24.800 --> 00:04:28.077 I don't think there were a lot of other people that would have uh been willing 00:04:28.160 --> 00:04:32.317 to take that kind of a bet given what the world looked like at the time. Um we 00:04:32.400 --> 00:04:36.157 didn't know exactly how the tech was going to go. Well, not exactly. We 00:04:36.240 --> 00:04:38.237 didn't know at all how the tech was going to go. We just had a lot of 00:04:38.320 --> 00:04:42.077 conviction in this this one idea of pushing on on deep learning and trusting 00:04:42.160 --> 00:04:45.597 that if we could do that, we'd figure out ways to make wonderful products and 00:04:45.680 --> 00:04:49.117 create a lot of value and also, as Satia said, create what we believe will be the 00:04:49.200 --> 00:04:52.397 largest nonprofit ever. And I think it's 00:04:52.480 --> 00:04:56.397 going to do amazingly great things. It it was I I really like the structure 00:04:56.480 --> 00:04:58.477 because it lets the nonprofit grow in 00:04:58.560 --> 00:05:04.077 value while the PBC is able to get the capital that it needs to keep scaling. I 00:05:04.160 --> 00:05:07.037 don't think the nonprofit would be able to be this valuable if we didn't come up 00:05:07.120 --> 00:05:09.757 with the structure and if we didn't have partners around the table that were 00:05:09.840 --> 00:05:13.917 excited for it to work this way. But, you know, I think it's been six more 00:05:14.000 --> 00:05:18.797 than six years since we first started this partnership and uh a pretty crazy 00:05:18.880 --> 00:05:22.077 amount of achievement for six years and I think much much more to come. I hope 00:05:22.160 --> 00:05:24.877 that Sasha makes a trillion dollars on the investment, not hundred billion, you 00:05:24.960 --> 00:05:27.837 know, whatever it is. >> Well, as part of the restructuring, you 00:05:27.920 --> 00:05:31.357 guys talked about it. You have this nonprofit on top and a public benefit 00:05:31.440 --> 00:05:33.517 corp below. It's pretty insane. The 00:05:33.600 --> 00:05:36.477 nonprofit is already capitalized with 00:05:36.560 --> 00:05:41.197 $130 billion. $130 billion of Open AI 00:05:41.280 --> 00:05:44.557 stock. It's one of the largest in the world out of the gates. It could end up 00:05:44.640 --> 00:05:48.157 being much much larger. The California Attorney General said they're not going 00:05:48.240 --> 00:05:50.317 to object to it. You already haveund 00:05:50.400 --> 00:05:55.197 this 130 billion dedicated to making sure that AGI benefits all of humanity. 00:05:55.280 --> 00:05:56.477 You announced that you're going to 00:05:56.560 --> 00:05:58.957 direct the first 25 billion to health 00:05:59.040 --> 00:06:04.157 and AI security and resilience. Sam, first let me just say, you know, as 00:06:04.240 --> 00:06:07.437 somebody who participates in the ecosystem, kudos to you both. It's 00:06:07.520 --> 00:06:09.357 incredible this contribution to the 00:06:09.440 --> 00:06:11.837 future of AI. But Sam, talk to us a bit 00:06:11.920 --> 00:06:14.717 about the importance of the the choice 00:06:14.800 --> 00:06:17.277 around health and and resilience. And 00:06:17.360 --> 00:06:21.837 then help us understand how do we make sure that you get maximal benefit 00:06:21.920 --> 00:06:24.077 without it getting weighted down as 00:06:24.160 --> 00:06:26.237 we've seen with so many nonprofits with 00:06:26.320 --> 00:06:29.037 its own political biases. 00:06:29.120 --> 00:06:34.077 >> Yeah. First of all, the the best way to create a bunch of value for the world is 00:06:34.160 --> 00:06:37.117 hopefully what we're we've already been doing, which is to make these amazing 00:06:37.200 --> 00:06:40.957 tools and just let people use them. And I think capitalism is great. I think 00:06:41.040 --> 00:06:45.037 companies are great. I think people are doing amazing work getting advanced AI 00:06:45.120 --> 00:06:46.557 into the hands of a lot of people and 00:06:46.640 --> 00:06:47.597 companies. They're doing incredible 00:06:47.680 --> 00:06:52.077 things. There are some areas where the I 00:06:52.160 --> 00:06:56.877 think market forces don't quite work for what's in the best interest of people 00:06:56.960 --> 00:07:00.877 and you do need to do things in a different way. Uh there are also some 00:07:00.960 --> 00:07:04.157 new things with this technology that just haven't existed before like the 00:07:04.240 --> 00:07:08.717 potential to use AI to do science at a rapid clip like really truly automated 00:07:08.800 --> 00:07:13.277 discovery. And when we thought about the areas we wanted to first focus on, 00:07:13.360 --> 00:07:17.277 clearly if we can cure a lot of disease and make the data and information for 00:07:17.360 --> 00:07:20.797 that broadly available, that would that'd be a wonderful thing to do for 00:07:20.880 --> 00:07:25.117 the world. And then on this point of AI resilience, I do think some things may 00:07:25.200 --> 00:07:29.837 get a little strange and they won't all be addressed by companies doing their 00:07:29.920 --> 00:07:34.317 thing. So as the world has to navigate through this transition, if we can fund 00:07:34.400 --> 00:07:37.837 some work to help with that, and that could be, you know, cyber defense, that 00:07:37.920 --> 00:07:40.797 could be AI safety research, that could be economic studies, all of these 00:07:40.880 --> 00:07:45.437 things, helping society get through this transition smoothly. We're very 00:07:45.520 --> 00:07:48.637 confident about how great it can be on the other side, but you know, I'm sure 00:07:48.720 --> 00:07:50.557 there will be some choppiness along the way. 00:07:50.640 --> 00:07:52.797 >> Let's keep busting through the the the 00:07:52.880 --> 00:07:55.677 um the deal. So models and exclusivity 00:07:55.760 --> 00:08:00.797 Sam OpenAI can distribute its models uh its leading models on Azure but I don't 00:08:00.880 --> 00:08:04.557 think you can distribute them on any other leading the big clouds for seven 00:08:04.640 --> 00:08:07.277 years until 2032 but that would end 00:08:07.360 --> 00:08:09.917 earlier if AGI is verified. We can come 00:08:10.000 --> 00:08:14.717 back to that but you can distribute your open source models Sora agents codecs 00:08:14.800 --> 00:08:16.477 wearables everything else on other 00:08:16.560 --> 00:08:19.517 platforms. So Sam, I assume this means 00:08:19.600 --> 00:08:23.837 no chat GPT or GPT6 on Amazon or Google. 00:08:23.920 --> 00:08:27.437 >> No. So, so we have a C. First of all, we want to do lots of things together to 00:08:27.520 --> 00:08:30.397 help, you know, create value for Microsoft. We want them to do lots of 00:08:30.480 --> 00:08:33.437 things for to create value for us. And there are many many things that'll 00:08:33.520 --> 00:08:37.037 happen in that category. Um, we are keeping what Satia termed once and I 00:08:37.120 --> 00:08:41.357 think it's a great phrase of stateless APIs on Azure exclusively through 2030. 00:08:41.440 --> 00:08:44.317 And everything else we're going to, you know, distribute elsewhere and that's 00:08:44.400 --> 00:08:46.317 obviously in Microsoft's interest, too. 00:08:46.400 --> 00:08:49.117 So, we'll put lots of products, lots of places, and then this thing we'll we'll 00:08:49.200 --> 00:08:52.957 do on Azure and people can get it there or or via us. And I think that's great. 00:08:53.040 --> 00:08:56.397 >> And then the rev share, there's still a rev share that gets paid by OpenAI to 00:08:56.480 --> 00:08:58.797 Microsoft on all your revenues that also 00:08:58.880 --> 00:09:03.917 runs until 2032 or until AGI is verified. So, let's just assume for the 00:09:04.000 --> 00:09:07.277 sake of argument, I know this is pedestrian, but it's important that the 00:09:07.360 --> 00:09:11.437 rev share is 15%. So that would mean if you had 20 billion in revenue that 00:09:11.520 --> 00:09:15.997 you're paying three billion to Microsoft and that counts as revenue to Azure. 00:09:16.080 --> 00:09:18.237 Satcha, is that does that sound about right? 00:09:18.320 --> 00:09:20.077 >> Yeah, we have a rev share and I think as 00:09:20.160 --> 00:09:22.317 you characterized it is either going to 00:09:22.400 --> 00:09:27.117 AGI or till the end of the term. Uh and I actually don't know exactly where we 00:09:27.200 --> 00:09:30.237 count it quite honestly whether it goes into Azure or somewhere else. That's a 00:09:30.320 --> 00:09:31.677 good question. It's a good question for 00:09:31.760 --> 00:09:34.557 Amy. Given that both exclusivity and the 00:09:34.640 --> 00:09:37.677 revshare end early in the case AGI is 00:09:37.760 --> 00:09:39.917 verified, it seems to make AGI a pretty 00:09:40.000 --> 00:09:44.877 big deal. And as I understand it, you know, if if OpenAI claimed AGI, it 00:09:44.960 --> 00:09:46.797 sounds like it goes to an expert panel. 00:09:46.880 --> 00:09:50.557 And you guys basically select a jury who's got to make a relatively quick 00:09:50.640 --> 00:09:54.077 decision whether or not AGI has been reached. Satcha, you said on yesterday's 00:09:54.160 --> 00:09:58.077 earning call that nobody's even close to getting to AGI and you don't expect it 00:09:58.160 --> 00:10:02.557 to happen anytime soon. You talked about this spiky and jagged intelligence. Sam, 00:10:02.640 --> 00:10:06.637 I've heard you perhaps sound a little bit more bullish on, you know, when we 00:10:06.720 --> 00:10:10.797 might get to AGI. So, I guess the question is to you both. Do you worry 00:10:10.880 --> 00:10:14.397 that over the next two or three years we're going to end up having to call in 00:10:14.480 --> 00:10:18.877 the jury to effectively make a uh a call on whether or not we've hit AGI? 00:10:18.960 --> 00:10:20.477 >> I I realize you got to try to make some 00:10:20.560 --> 00:10:24.557 drama between us here. I [laughter] 00:10:24.640 --> 00:10:26.797 >> you know, I think putting a process in 00:10:26.880 --> 00:10:30.077 place for this is a good thing to do. I 00:10:30.160 --> 00:10:34.877 expect that the technology will take several surprising twists and turns and 00:10:34.960 --> 00:10:37.677 we will continue to be good partners to each other and figure out what makes 00:10:37.760 --> 00:10:42.397 sense. >> That's well said. I think uh and that's one of the reasons why I think this 00:10:42.480 --> 00:10:46.317 process we put in place is a good one and at the end of the day I'm a big 00:10:46.400 --> 00:10:49.357 believer in the fact that intelligence 00:10:49.440 --> 00:10:53.517 uh capability wise is going to continue to improve and our real goal quite 00:10:53.600 --> 00:10:56.877 frankly is that which is how do you put that in the hands of people and 00:10:56.960 --> 00:11:00.717 organizations so that they can get the maximum benefits and that was the 00:11:00.800 --> 00:11:05.037 original mission of open AI that attracted me to open AAI and Sam and 00:11:05.120 --> 00:11:08.237 team and that's kind of what we plan to continue on 00:11:08.320 --> 00:11:11.597 >> Brad to say the obvious if we had super intelligence tomorrow, we would still 00:11:11.680 --> 00:11:15.357 want Microsoft's help getting this product out into people's hands and we 00:11:15.440 --> 00:11:19.917 want them like Yeah, >> of course. Of course. Yeah. No, it again 00:11:20.000 --> 00:11:21.917 I'm asking the questions I know that are 00:11:22.000 --> 00:11:26.957 on people's minds and that makes a ton of sense to me. Obviously s Microsoft is 00:11:27.040 --> 00:11:29.517 one of the largest distribution platforms in the world. You guys have 00:11:29.600 --> 00:11:32.557 been great partners for a long time. But I think it dispels some of the myths 00:11:32.640 --> 00:11:35.597 that are out there. But let's shift gears a little bit. You know, obviously 00:11:35.680 --> 00:11:37.277 OpenAI is one of the fastest growing 00:11:37.360 --> 00:11:39.517 companies in history. Satcha, you said 00:11:39.600 --> 00:11:42.637 on the pod a year ago, this pod, that 00:11:42.720 --> 00:11:47.037 every new phase shift creates a new Google and the Google of this phase 00:11:47.120 --> 00:11:49.517 shift is already known and it's open AI. 00:11:49.600 --> 00:11:52.397 And none of this would have been possible had you guys not made these 00:11:52.480 --> 00:11:55.597 these huge bets. With all that said, you 00:11:55.680 --> 00:11:57.597 know, OpenAI's revenues are still a 00:11:57.680 --> 00:12:01.197 reported 13 billion in 2025. And Sam, on 00:12:01.280 --> 00:12:03.277 your live stream this week, you talked 00:12:03.360 --> 00:12:05.197 about this massive commitment to 00:12:05.280 --> 00:12:08.477 compute, right? 1.4 4 trillion over the 00:12:08.560 --> 00:12:11.357 next four or five years with you know 00:12:11.440 --> 00:12:13.757 big commitments 500 million to Nvidia 00:12:13.840 --> 00:12:16.477 300 million to AMD and Oracle 250 00:12:16.560 --> 00:12:19.517 billion to Azure. So I think the single 00:12:19.600 --> 00:12:21.597 biggest question I've heard all week and 00:12:21.680 --> 00:12:24.637 and hanging over the market is how you 00:12:24.720 --> 00:12:26.637 know how can a company with 13 billion 00:12:26.720 --> 00:12:30.637 in revenues make 1.4 4 trillion of spend 00:12:30.720 --> 00:12:34.237 commitments, you know, and and and you've heard the criticism, Sam. 00:12:34.320 --> 00:12:37.677 >> First of all, we're doing well more revenue than that. Second of all, Brad, 00:12:37.760 --> 00:12:38.797 if you want to sell your shares, I'll 00:12:38.880 --> 00:12:41.357 find you a buyer. [laughter] 00:12:41.440 --> 00:12:46.317 >> I just enough like, you know, people are I I think there's a lot of people who 00:12:46.400 --> 00:12:48.957 would love to buy OpenAI shares. I don't I don't think you 00:12:49.040 --> 00:12:53.917 >> including myself, including myself, >> people who talk with a lot of like 00:12:54.000 --> 00:12:57.117 breathless concern about our comput stuff or whatever that would be thrilled 00:12:57.200 --> 00:13:00.557 to buy shares. So I think we we could sell you know your shares or anybody 00:13:00.640 --> 00:13:03.117 else's to some of the people who are making the most noise on Twitter 00:13:03.200 --> 00:13:07.277 whatever about this very quickly. We do plan for revenue to grow steeply. 00:13:07.360 --> 00:13:10.557 Revenue is growing steeply. We are taking a forward bet that it's going to 00:13:10.640 --> 00:13:13.357 continue to grow grow and that not only 00:13:13.440 --> 00:13:17.917 will Chhatabt keep growing but we will be able to become one of the important 00:13:18.000 --> 00:13:20.157 AI clouds that our consumer device 00:13:20.240 --> 00:13:23.117 business will be a significant and 00:13:23.200 --> 00:13:25.037 important thing that AI that can 00:13:25.120 --> 00:13:27.997 automate science will create huge value. 00:13:28.080 --> 00:13:31.197 So, you know, there are not many times 00:13:31.280 --> 00:13:35.437 that I want to be a public company, but one of the rare times it's appealing is 00:13:35.520 --> 00:13:38.477 when those people are writing these ridiculous OpenAI is about to go out of 00:13:38.560 --> 00:13:41.357 business and, you know, whatever. I would love to tell them they could just 00:13:41.440 --> 00:13:42.637 short the stock and I would love to see 00:13:42.720 --> 00:13:46.637 them get burned on that. Um, but 00:13:46.720 --> 00:13:49.437 you know, I we carefully plan, we 00:13:49.520 --> 00:13:54.477 understand where the technology, where the capability is going to grow, go and 00:13:54.560 --> 00:13:57.677 and how the products we can build around that and the revenue we can generate. we 00:13:57.760 --> 00:14:01.277 might screw it up like this is the bet that we're making and we're taking a 00:14:01.360 --> 00:14:05.277 risk along with that. A certain risk is if we don't have the compute, we will 00:14:05.360 --> 00:14:08.717 not be able to generate the revenue or make the models at these at this kind of 00:14:08.800 --> 00:14:10.317 scale. >> Exactly. And 00:14:10.400 --> 00:14:13.437 >> let me just say one thing uh Brad as 00:14:13.520 --> 00:14:17.197 both a partner and um an investor there 00:14:17.280 --> 00:14:20.077 is not been a single business plan that 00:14:20.160 --> 00:14:22.397 I've seen from OpenAI that they have put 00:14:22.480 --> 00:14:25.997 in and not beaten it. So in some sense 00:14:26.080 --> 00:14:30.877 this is the one place where you know in terms of their growth and just even the 00:14:30.960 --> 00:14:34.717 business it's been unbelievable execution quite frankly I mean obviously 00:14:34.800 --> 00:14:38.637 openai everyone talks about all the success in the usage and what have you 00:14:38.720 --> 00:14:41.117 but even um I would say all up uh the 00:14:41.200 --> 00:14:42.957 business execution has been just pretty 00:14:43.040 --> 00:14:45.757 unbelievable. I heard Greg Brockman say 00:14:45.840 --> 00:14:48.477 on C CBC a couple weeks ago, right? If 00:14:48.560 --> 00:14:50.717 we could 10x our compute, we might not 00:14:50.800 --> 00:14:52.957 have 10x more revenue, but we'd 00:14:53.040 --> 00:14:55.517 certainly have a lot more revenue 00:14:55.600 --> 00:14:57.997 >> simply because of lack of compute power. 00:14:58.080 --> 00:15:00.477 Things like, yeah, it's just it's really 00:15:00.560 --> 00:15:05.117 wild when I just look at how much we are held back. And in many ways, we have, 00:15:05.200 --> 00:15:08.317 you know, we've scaled our compute probably 10x over the past year, but if 00:15:08.400 --> 00:15:11.917 we had 10x more compute, I don't know if we'd have 10x more revenue, but I don't 00:15:12.000 --> 00:15:16.717 think it'd be that far. And we heard this from you as well last night Satcha 00:15:16.800 --> 00:15:19.917 that you were compute constrained and growth would have been higher even if if 00:15:20.000 --> 00:15:21.997 you had more compute. So help us 00:15:22.080 --> 00:15:27.517 contextualize Sam maybe like how compute constrained do you feel today and do you 00:15:27.600 --> 00:15:30.717 when you look at the buildout over the course of the next two to three years do 00:15:30.800 --> 00:15:34.877 you think you'll ever get to the point where you're not compute constrained? 00:15:34.960 --> 00:15:39.197 >> We talk about this question of is there ever enough compute a lot. I I think the 00:15:39.280 --> 00:15:41.277 answer is 00:15:41.360 --> 00:15:46.237 the only the best way to think about this is like a 00:15:46.320 --> 00:15:49.837 energy or something. You can talk about demand for energy at a certain price 00:15:49.920 --> 00:15:54.877 point, but you can't talk about demand for energy without talking about at 00:15:54.960 --> 00:15:58.237 different you know different demand at different 00:15:58.320 --> 00:16:03.277 price levels. If the price of compute per like unit of intelligence or 00:16:03.360 --> 00:16:06.157 whatever, however you want to think about it, fell by a factor of a 100 00:16:06.240 --> 00:16:10.877 tomorrow, you would see usage go up by much more than 100 and there'd be a lot 00:16:10.960 --> 00:16:13.437 of things that people would love to do with that compute that just make no 00:16:13.520 --> 00:16:17.357 economic sense at the current cost, but there would be new kind of demand. So I 00:16:17.440 --> 00:16:21.757 think the the now on the other hand as the models get 00:16:21.840 --> 00:16:25.837 even smarter and you can use these models to cure cancer or discover novel 00:16:25.920 --> 00:16:29.277 physics or drive a bunch of humanoid robots to construct a space station or 00:16:29.360 --> 00:16:32.157 whatever crazy thing you want then maybe 00:16:32.240 --> 00:16:36.717 there's huge willingness to pay a much higher rate cost per unit of 00:16:36.800 --> 00:16:39.997 intelligence for a much higher level of intelligence that we don't know yet but 00:16:40.080 --> 00:16:43.437 I would bet there will be. So I I think 00:16:43.520 --> 00:16:47.757 when you talk about capacity it's it's like a you know cost per unit and you 00:16:47.840 --> 00:16:49.917 know capability per unit and you have to 00:16:50.000 --> 00:16:55.597 kind of without those curves it's sort of a madeup it's not a super well 00:16:55.680 --> 00:16:58.397 specified problem. >> Yeah. I mean I think the one thing that 00:16:58.480 --> 00:17:01.357 you know Sam you've talked about which I think is the right way is to think about 00:17:01.440 --> 00:17:03.517 is that if intelligence is what a log of 00:17:03.600 --> 00:17:08.477 compute then you try and really make sure you keep getting efficient and so 00:17:08.560 --> 00:17:10.157 that means the tokens per dollar per 00:17:10.240 --> 00:17:13.517 watt uh and the economic value that the 00:17:13.600 --> 00:17:17.597 society gets out of it is what we should maximize and reduce the costs and so 00:17:17.680 --> 00:17:21.357 that's where if you sort of where like the Jevans paradox point is that right 00:17:21.440 --> 00:17:23.677 which is you keep reducing it 00:17:23.760 --> 00:17:26.557 commoditizing in some sense intelligence 00:17:26.640 --> 00:17:31.037 uh so that it becomes the real driver of GDP growth all around. 00:17:31.120 --> 00:17:34.717 >> Unfortunately, it's something closer to uh log of intelligence equals log of 00:17:34.800 --> 00:17:37.277 compute. But we may figure out better scaling laws and we may figure out how 00:17:37.360 --> 00:17:40.717 to beat this. Yeah, >> we heard from both Microsoft and Google 00:17:40.800 --> 00:17:43.277 yesterday. Both said their cloud businesses would have been growing 00:17:43.360 --> 00:17:47.837 faster if they have more GPUs. You know, I asked Jensen on this pod if there was 00:17:47.920 --> 00:17:52.717 any chance over the course of the next 5 years we would have a compute glut. and 00:17:52.800 --> 00:17:57.517 he said it's virtually non-existent chance in the next 2 to 3 years and I 00:17:57.600 --> 00:18:02.557 assume you guys would both agree with Jensen that while we can't see out 5 6 7 00:18:02.640 --> 00:18:06.157 years certainly over the course of the next 2 to three years for the for the 00:18:06.240 --> 00:18:09.917 reasons we just discussed that it's almost a non-existent chance that you 00:18:10.000 --> 00:18:13.037 have excess compute well I mean I think 00:18:13.120 --> 00:18:17.197 the the cycles of demand and supply in 00:18:17.280 --> 00:18:19.917 this particular case you can't really 00:18:20.000 --> 00:18:24.477 predict right I mean even the the point is What's the secular trend? The secular 00:18:24.560 --> 00:18:27.837 trend is what Sam said, which is at the end of the day, because quite frankly, 00:18:27.920 --> 00:18:31.757 the the biggest issue we are now having is not a compute glut, but it's a power 00:18:31.840 --> 00:18:34.157 and it's sort of the ability to get the 00:18:34.240 --> 00:18:37.517 builds done fast enough close to power. 00:18:37.600 --> 00:18:41.037 So, if you can't do that, you may actually have a bunch of chips sitting 00:18:41.120 --> 00:18:45.037 in inventory that I can't plug in. In fact, that is my problem today, right? 00:18:45.120 --> 00:18:47.917 It's not a supply issue of chips. It's 00:18:48.000 --> 00:18:49.917 actually uh the fact that I don't have 00:18:50.000 --> 00:18:52.957 warm shells to plug into. And so how 00:18:53.040 --> 00:18:57.757 some supply chain constraints emerge tough to predict uh because the demand 00:18:57.840 --> 00:19:01.597 is just going you know is tough to predict right I mean I wouldn't you it's 00:19:01.680 --> 00:19:05.117 not like Sam and I would want to be sitting here saying oh my god we're less 00:19:05.200 --> 00:19:09.037 short on compute it's because we just were not that good at being able to 00:19:09.120 --> 00:19:13.357 project out what the demand would really look like. So I think that that's and by 00:19:13.440 --> 00:19:17.277 the way the worldwide side right one it's one thing to sort of talk about one 00:19:17.360 --> 00:19:20.717 segment in one country but it's about you know really getting it out to 00:19:20.800 --> 00:19:23.997 everywhere in the world and so there will be constraints and how we work 00:19:24.080 --> 00:19:27.117 through them is going to be the most important thing it won't be a linear 00:19:27.200 --> 00:19:31.517 path for sure there there will come a glut for sure and whether that's like in 00:19:31.600 --> 00:19:35.837 two to three years or five to six I can't tell you but uh like it's going to 00:19:35.920 --> 00:19:39.277 happen at some point probably several points along the way like this is 00:19:39.360 --> 00:19:41.517 there's something deep about human 00:19:41.600 --> 00:19:45.357 psychology here and bubbles and also as 00:19:45.440 --> 00:19:49.837 Satia said like there's it's such a complex supply chain weird stuff gets 00:19:49.920 --> 00:19:52.077 built the technological landscape shifts 00:19:52.160 --> 00:19:57.597 in big ways so you know if a very cheap form of energy comes online 00:19:57.680 --> 00:20:00.477 soon at mass scale then a lot of people are going to be extremely burned with 00:20:00.560 --> 00:20:03.677 existing contracts they've signed it I 00:20:03.760 --> 00:20:08.397 if if we can continue this unbelievable reduction in cost per unit of 00:20:08.480 --> 00:20:12.237 intelligence let's say it's been averaging like 40x X for a given level 00:20:12.320 --> 00:20:17.117 per year. You know, that's like a very scary exponent 00:20:17.200 --> 00:20:20.557 from an infrastructure buildout standpoint. Now, again, we're taking the 00:20:20.640 --> 00:20:25.117 bet that there will be a lot more demand as that gets cheaper, but I have some 00:20:25.200 --> 00:20:28.397 fear that it's just like, man, we keep going with these breakthroughs and 00:20:28.480 --> 00:20:31.757 everybody can run like a personal AGI on their laptop and we just did an insane 00:20:31.840 --> 00:20:35.277 thing here. Some people are going to get 00:20:35.360 --> 00:20:39.517 really burned like has happened in every other tech infrastructure cycle at some 00:20:39.600 --> 00:20:42.637 points along the way. >> I think that's really well said and you 00:20:42.720 --> 00:20:44.237 have to hold those two simultaneous 00:20:44.320 --> 00:20:47.357 truths. We had that happen in 20201 and 00:20:47.440 --> 00:20:51.597 yet the internet became much bigger and produced much greater outcomes for 00:20:51.680 --> 00:20:54.477 society than anybody estimated in that period of time. 00:20:54.560 --> 00:20:58.077 >> Yeah. But I think that the one thing that Sam said is not talked about enough 00:20:58.160 --> 00:21:00.397 which is the current for example the 00:21:00.480 --> 00:21:05.197 optimizations that OpenAI has done on the inference stack for a given GPU. I 00:21:05.280 --> 00:21:08.957 mean I it's kind of like it's you know we talk about the MOS law improvement on 00:21:09.040 --> 00:21:11.197 one end but the software improvements 00:21:11.280 --> 00:21:14.797 are much more exponential than that. 00:21:14.880 --> 00:21:17.037 Someday we will make a incredible 00:21:17.120 --> 00:21:20.237 consumer device that can run a GPT5 or 00:21:20.320 --> 00:21:23.757 GPD6 capable model completely locally at 00:21:23.840 --> 00:21:27.837 a low power draw. And this is like so hard to wrap my head around. 00:21:27.920 --> 00:21:30.957 >> That will be incredible. And you know that's the type of thing I think that 00:21:31.040 --> 00:21:34.317 scares some of the people who are building obviously these large 00:21:34.400 --> 00:21:38.557 centralized compute uh stacks. And Satcha you've talked a lot about the 00:21:38.640 --> 00:21:42.797 distribution both to the edge as well as having inference capability distributed 00:21:42.880 --> 00:21:46.477 around the world. Yeah, I mean the way at least I've thought about it is more 00:21:46.560 --> 00:21:49.517 about really building a fungeable fleet. 00:21:49.600 --> 00:21:52.957 I mean when I look at sort of in the cloud infrastructure business, one of 00:21:53.040 --> 00:21:57.757 the key things you have to do is have two things. One is an effic like in this 00:21:57.840 --> 00:22:02.237 context in a very efficient token factory and then high utilization. 00:22:02.320 --> 00:22:05.917 That's that's it. There are two simple things that you need to achieve and in 00:22:06.000 --> 00:22:09.197 order to have high utilization you have to have multiple workloads that can be 00:22:09.280 --> 00:22:12.557 scheduled even on the training. I mean, if you look at the AI pipelines, there's 00:22:12.640 --> 00:22:15.517 pre-training, there's mid-training, there's post- training, there's RL. You 00:22:15.600 --> 00:22:19.037 want to be able to do all of those things. So, thinking about fungeibility 00:22:19.120 --> 00:22:22.877 of the fleet is everything for a cloud provider. 00:22:22.960 --> 00:22:24.717 >> Okay. So, Sam, you referenced, you know, 00:22:24.800 --> 00:22:29.357 and and Reuters was reporting yesterday that OpenAI may be planning to go public 00:22:29.440 --> 00:22:32.317 late 26 or in 27. >> No, no, no. We we don't we don't have 00:22:32.400 --> 00:22:34.077 anything that specific. I I'm a realist. 00:22:34.160 --> 00:22:38.877 I assume it will happen someday, but that was uh I don't know why people 00:22:38.960 --> 00:22:42.877 write these reports. We don't have like date in mind decision to do this or 00:22:42.960 --> 00:22:45.597 anything like that. I just assume it's where things will eventually go. 00:22:45.680 --> 00:22:48.877 >> But it does seem to me if you guys were, 00:22:48.960 --> 00:22:50.717 you know, are are doing in excess of 00:22:50.800 --> 00:22:56.477 hundred billion dollars of revenue in 28 or 29 that you at least would be in pos 00:22:56.560 --> 00:22:58.237 >> what? >> How about 27? 00:22:58.320 --> 00:23:00.877 >> Yeah, 27 even better. You are in 00:23:00.960 --> 00:23:05.117 position to do an IPO and the rumored 00:23:05.200 --> 00:23:06.557 trillion dollars. Again, just to 00:23:06.640 --> 00:23:09.117 contextualize for listeners, if you guys 00:23:09.200 --> 00:23:12.237 went public at 10 times 100 billion in 00:23:12.320 --> 00:23:16.477 revenue, right, which would be, I think, a lower multiple than Facebook went 00:23:16.560 --> 00:23:18.957 public at, a lower multiple than a lot 00:23:19.040 --> 00:23:23.837 of other uh big consumer companies went public at, that would put you at a 00:23:23.920 --> 00:23:25.917 trillion dollars. If you floated 10 to 00:23:26.000 --> 00:23:28.077 20% of the company, that raises a 00:23:28.160 --> 00:23:31.677 hundred to$200 billion, which seems like 00:23:31.760 --> 00:23:35.597 that would be a good path to fund a lot of the growth and a lot of the stuff 00:23:35.680 --> 00:23:39.757 that we just talked about. So, you're you're you're not opposed to it. You're 00:23:39.840 --> 00:23:43.677 not But you guys are making fund the company with revenue growth, which is 00:23:43.760 --> 00:23:47.037 what I would like us to do. >> But no doubt about it. 00:23:47.120 --> 00:23:51.437 Well, I've also said I think that this is such an important company and you 00:23:51.520 --> 00:23:53.277 know there are so many people including 00:23:53.360 --> 00:23:56.237 my kids who like to trade their little 00:23:56.320 --> 00:24:01.037 accounts and they use chat GPT and I think having retail investors have an 00:24:01.120 --> 00:24:04.237 opportunity to buy one of the most important and largest 00:24:04.320 --> 00:24:06.637 >> honestly that that is probably the 00:24:06.720 --> 00:24:11.437 single most appealing thing about it to me. Um that would be really nice. 00:24:11.520 --> 00:24:15.677 One of the things I've talked to you both about um shifting gears again is 00:24:15.760 --> 00:24:17.917 part of the big beautiful bill, you 00:24:18.000 --> 00:24:20.637 know, Senator Cruz had included federal 00:24:20.720 --> 00:24:23.517 preeemption so that we wouldn't have 00:24:23.600 --> 00:24:26.717 this state patchwork 50 different laws 00:24:26.800 --> 00:24:31.517 that mireers the industry down in kind of needless compliance and regulation. 00:24:31.600 --> 00:24:35.437 unfortunately got killed at the last second by Senator Blackburn because 00:24:35.520 --> 00:24:39.357 frankly I think AI is pretty poorly understood in Washington and there's a 00:24:39.440 --> 00:24:41.597 lot of dumerism I think that has gained 00:24:41.680 --> 00:24:46.637 traction in Washington. So now we have state laws like the Colorado AI act that 00:24:46.720 --> 00:24:50.237 goes into full effect in February I believe that creates this whole new 00:24:50.320 --> 00:24:54.717 class of litigants anybody who claims any unfair impact from an algorithmic 00:24:54.800 --> 00:24:56.957 discrimination in a chatbot. So somebody 00:24:57.040 --> 00:24:59.917 could claim harm for countless reasons. 00:25:00.000 --> 00:25:03.117 Sam, how worried are you that, you know, 00:25:03.200 --> 00:25:08.317 having this state patchwork of AI, you know, poses real challenges to, you 00:25:08.400 --> 00:25:12.157 know, our ability to continue to accelerate and compete around the world. 00:25:12.240 --> 00:25:15.517 >> I don't know how we're supposed to comply with that California, sorry, 00:25:15.600 --> 00:25:19.357 Colorado law. I would love them to tell us uh and, you know, we'd like to be 00:25:19.440 --> 00:25:23.677 able to do it, but that's just from what I've read of that. That's like a I 00:25:23.760 --> 00:25:27.277 literally don't know what we're supposed to do. I'm very worried about a 50-state 00:25:27.360 --> 00:25:30.717 patchwork. I think it's a big mistake. I think it's there's a reason we don't 00:25:30.800 --> 00:25:33.997 usually do that for these sorts of things. I think it'd be bad. 00:25:34.080 --> 00:25:37.757 >> Yeah. I mean, I think the the fundamental problem of um you know, this 00:25:37.840 --> 00:25:42.317 patchwork approach is quite frankly, I mean, between OpenAI and Microsoft, 00:25:42.400 --> 00:25:46.237 we'll figure out a way to navigate this, right? I mean, uh we can figure this 00:25:46.320 --> 00:25:48.397 out. The problem is anyone starting a 00:25:48.480 --> 00:25:53.517 startup and trying to kind this it's sort of it just goes to the exact 00:25:53.600 --> 00:25:58.157 opposite of I think what the intent here is which obviously safety is very 00:25:58.240 --> 00:26:02.397 important making sure that the fundamental um you know concerns people 00:26:02.480 --> 00:26:05.677 have are addressed but there's a way to do that at the federal level so I think 00:26:05.760 --> 00:26:09.037 the U if we don't do this again you know 00:26:09.120 --> 00:26:11.437 EU will do it and then that'll cause its 00:26:11.520 --> 00:26:17.357 own issues so I think if US leads it's better uh as you as one regulatory 00:26:17.440 --> 00:26:21.757 framework >> for sure. >> And to be clear, it's not that one is 00:26:21.840 --> 00:26:25.357 advocating for no regulation. It's simply saying let's have, you know, 00:26:25.440 --> 00:26:26.957 agreed upon regulation at the federal 00:26:27.040 --> 00:26:29.277 level as opposed to 50 competing state 00:26:29.360 --> 00:26:32.637 laws which certainly uh firebombs the 00:26:32.720 --> 00:26:36.717 the AI startup industry and I think it makes it makes it super challenging even 00:26:36.800 --> 00:26:39.677 for companies like yours who can afford to defend all these cases. 00:26:39.760 --> 00:26:41.517 >> Yeah. And I would just say quite frankly 00:26:41.600 --> 00:26:45.997 my hope is that this time around even across EU and the United States like 00:26:46.080 --> 00:26:49.757 that'll be the dream right quite frankly for any European startup. 00:26:49.840 --> 00:26:51.597 >> I don't think that's going to happen. 00:26:51.680 --> 00:26:53.357 >> What is that? >> That would be great. I don't I wouldn't 00:26:53.440 --> 00:26:57.117 hold your breath for that one. That would be great. No, but I I I really 00:26:57.200 --> 00:27:00.717 think that if you think about it right, if you sort of if anyone in Europe is 00:27:00.800 --> 00:27:02.557 thinking about their you know what how 00:27:02.640 --> 00:27:05.677 can they participate in this AI uh 00:27:05.760 --> 00:27:10.157 economy with their companies uh this should be the main concern there as 00:27:10.240 --> 00:27:14.397 well. So therefore uh that's I hope there is some enlightened approach to it 00:27:14.480 --> 00:27:17.917 but I agree with you that you know today I wouldn't bet on that. 00:27:18.000 --> 00:27:20.877 I do think that with Sachs as the AIS 00:27:20.960 --> 00:27:23.037 are, you at least have a president that 00:27:23.120 --> 00:27:25.597 I think might fight for that in terms of 00:27:25.680 --> 00:27:30.877 coordination of of AI policy, using trade as a lever to make sure that, you 00:27:30.960 --> 00:27:33.917 know, we don't end up with overly restricted European policy. But we shall 00:27:34.000 --> 00:27:37.277 see. I think first things first, federal preeemption in the United States is 00:27:37.360 --> 00:27:40.237 pretty critical. You know, we've been down in the weeds a little bit here, 00:27:40.320 --> 00:27:42.157 Sam. So, I want to telescope out a 00:27:42.240 --> 00:27:45.117 little bit. You know, I've heard people 00:27:45.200 --> 00:27:47.917 on your team talk about all the great 00:27:48.000 --> 00:27:50.717 things coming up and and as you start 00:27:50.800 --> 00:27:52.557 thinking about much more unlimited 00:27:52.640 --> 00:27:56.877 compute chat GPT6 and beyond robotics, 00:27:56.960 --> 00:27:58.877 physical devices, 00:27:58.960 --> 00:28:04.077 scientific research as you as you look forward to 2026, what do you think 00:28:04.160 --> 00:28:08.477 surprises us the most? What what what what are you most excited about in terms 00:28:08.560 --> 00:28:12.717 of what's on the drawing board? you I mean you just hit on a lot of the key 00:28:12.800 --> 00:28:17.517 points there. I I think codeex has been a very cool thing to 00:28:17.600 --> 00:28:21.437 watch this year and as these go from multi-our tasks to multi-day tasks which 00:28:21.520 --> 00:28:23.837 I expect to happen next year what people 00:28:23.920 --> 00:28:29.197 be able to do to create software at an unprecedented rate and 00:28:29.280 --> 00:28:30.637 and really in fundamentally new ways. 00:28:30.720 --> 00:28:33.117 I'm very excited for that. I think we'll see that in other industries too. I have 00:28:33.200 --> 00:28:36.877 like a bias towards coding. I understand that one better. I think we'll see that 00:28:36.960 --> 00:28:39.117 really start to transform what people 00:28:39.200 --> 00:28:44.317 are capable of. I I I hope for very small scientific discoveries in 2026, 00:28:44.400 --> 00:28:47.517 but if we can get those very small ones, we'll get bigger ones in future years. 00:28:47.600 --> 00:28:50.477 That's a really crazy thing to say is that like AI is going to make a novel 00:28:50.560 --> 00:28:52.637 scientific discovery in 2026. Even a 00:28:52.720 --> 00:28:57.677 very small one. This is like this is a wildly important thing to be talking 00:28:57.760 --> 00:28:59.517 about. So, I'm excited for that. 00:28:59.600 --> 00:29:03.597 Certainly, robotics and computer and new kind of computers in future years. 00:29:03.680 --> 00:29:06.397 That'll be that'll be uh very important. 00:29:06.480 --> 00:29:10.797 But yeah, my personal bias is if we can 00:29:10.880 --> 00:29:15.437 really get AI to do science here, that is I mean that is super intelligence in 00:29:15.520 --> 00:29:19.357 some sense. Like if if this is expanding the total sum of human knowledge that is 00:29:19.440 --> 00:29:22.397 a crazy big deal. >> Yeah. I mean I think one of the things 00:29:22.480 --> 00:29:25.757 to use your codeex example I think the 00:29:25.840 --> 00:29:29.517 combination of the model capability I mean if you think about the magical 00:29:29.600 --> 00:29:31.997 moment that happened with chat GPT was 00:29:32.080 --> 00:29:35.197 the UI that met intelligence that just 00:29:35.280 --> 00:29:39.837 took off right there it's just you know unbelievable right form fact and some of 00:29:39.920 --> 00:29:41.677 it was also the instruction following 00:29:41.760 --> 00:29:44.717 piece of model capability was ready for 00:29:44.800 --> 00:29:47.597 chat I think that that's what the codeex 00:29:47.680 --> 00:29:49.997 and the you know these coding agents are 00:29:50.080 --> 00:29:52.877 about to uh help us which is what's that 00:29:52.960 --> 00:29:54.637 you know coding agent goes off for a 00:29:54.720 --> 00:29:57.277 long period of time comes back and then 00:29:57.360 --> 00:30:01.837 I'm then dropped into what I should steer like one of the metaphors I think 00:30:01.920 --> 00:30:04.157 we're all sort of working towards is I 00:30:04.240 --> 00:30:07.437 do this macro delegation and micro 00:30:07.520 --> 00:30:11.117 steering what is that UI meets this new 00:30:11.200 --> 00:30:15.597 intelligence capability and you can see the beginnings of that with codeex right 00:30:15.680 --> 00:30:19.917 the way at least I use it inside a GitHub copilot is I you know it's Now, 00:30:20.000 --> 00:30:22.317 it's just a it's a just a different way 00:30:22.400 --> 00:30:26.957 than the chat interface. And I think that that I think would be a new way for 00:30:27.040 --> 00:30:30.557 the human computer interface. Quite frankly, it's probably bigger than 00:30:30.640 --> 00:30:33.037 >> uh that that might be the departure. 00:30:33.120 --> 00:30:36.397 >> That's one reason I'm very excited that we're doing new form factors of 00:30:36.480 --> 00:30:39.677 computing devices cuz computers were not built for that kind of workflow very 00:30:39.760 --> 00:30:42.477 well. Certainly, a UI like Chacht is 00:30:42.560 --> 00:30:47.277 wrong for it. But this idea that you can have a device that is sort of always 00:30:47.360 --> 00:30:50.557 with you but able to go off and do things and get micro steer from you when 00:30:50.640 --> 00:30:53.997 it needs and have like really good contextual awareness of your whole life 00:30:54.080 --> 00:30:55.757 and flow. And I think that'll be cool. 00:30:55.840 --> 00:30:59.437 >> And what neither of you have talked about is the consumer use case. I think 00:30:59.520 --> 00:31:02.797 a lot about, you know, again, we go under this device and we have to hunt 00:31:02.880 --> 00:31:05.837 and peck through a hundred different applications and fill out little web 00:31:05.920 --> 00:31:09.917 forms, things that really haven't changed in 20 years. But to just have, 00:31:10.000 --> 00:31:12.877 you know, a personal assistant that we take for granted perhaps that we 00:31:12.960 --> 00:31:16.637 actually have a personal assistant, but to give a personal assistant for 00:31:16.720 --> 00:31:21.437 virtually free to billions of people around the world to improve their lives, 00:31:21.520 --> 00:31:25.117 whether it's, you know, ordering diapers for their kid or whether it's, you know, 00:31:25.200 --> 00:31:29.277 booking their hotel or or or making changes in their calendar. I think 00:31:29.360 --> 00:31:31.357 sometimes it's the pedestrian that's 00:31:31.440 --> 00:31:36.477 that's the most impactful. And as we move from answers to memory and actions 00:31:36.560 --> 00:31:39.917 and then the ability to interface with that through an earbud or some other 00:31:40.000 --> 00:31:43.517 device that doesn't require me to constantly be st staring at this 00:31:43.600 --> 00:31:46.877 rectangular piece of glass. I think it's pretty extraordinary. 00:31:46.960 --> 00:31:49.357 >> I think that that's what Sam was teasing. 00:31:49.440 --> 00:31:53.197 >> Yeah. Yeah. >> Hope we get it right. I got to drop off unfortunately. 00:31:53.280 --> 00:31:56.957 >> Sam, it was great to see you. Thanks for joining us. Congrats again on this big 00:31:57.040 --> 00:31:58.077 step forward and we'll talk soon. 00:31:58.160 --> 00:31:59.757 >> Thanks for letting me crash. 00:31:59.840 --> 00:32:02.397 >> See you Sam. Take care. See you. 00:32:02.480 --> 00:32:07.437 >> As Samwell knows, we're certainly a buyer, not a seller. Um, but but but 00:32:07.520 --> 00:32:11.997 sometimes, you know, I think it's important because the world, you know, 00:32:12.080 --> 00:32:16.237 we're a pretty small, we spend all day long thinking about this stuff, right? 00:32:16.320 --> 00:32:18.957 And so conviction, it comes from the 00:32:19.040 --> 00:32:23.837 10,000 hours we've spent thinking about it. But the reality is we have to bring 00:32:23.920 --> 00:32:27.757 along the rest of the world. And the rest of the world doesn't spend 10,000 00:32:27.840 --> 00:32:31.437 hours thinking about this. Um, and frankly they look at some things that 00:32:31.520 --> 00:32:34.317 appear overly ambitious, right, and get 00:32:34.400 --> 00:32:36.077 worried about whether or not we can pull 00:32:36.160 --> 00:32:38.717 those things off. You took this idea to 00:32:38.800 --> 00:32:41.917 the board in 2019 to invest a billion 00:32:42.000 --> 00:32:46.077 dollars into open AI. Was it a no-brainer in the boardroom? You know, 00:32:46.160 --> 00:32:50.477 did you have to expend any political capital to get it done? dish dish for me 00:32:50.560 --> 00:32:54.557 a little bit like what that moment was was like because I think it was such a 00:32:54.640 --> 00:32:58.477 pivotal moment not just for Microsoft not just for the country but I really do 00:32:58.560 --> 00:33:02.317 think for the world. Yeah, I mean it's it's interesting when you look back the 00:33:02.400 --> 00:33:04.157 the journey when I look at it it's been 00:33:04.240 --> 00:33:07.597 a you know we were involved even in 2016 00:33:07.680 --> 00:33:10.557 uh when initially open AI uh started in 00:33:10.640 --> 00:33:14.717 fact Azure was even the first sponsor I think and then they were doing a lot 00:33:14.800 --> 00:33:18.957 more reinforcement learning at that time I remember the Dota 2 competition I 00:33:19.040 --> 00:33:23.597 think happened on Azure and then uh they moved on to other things and you know I 00:33:23.680 --> 00:33:27.517 was interested in RL but quite frankly you know it speaks a little bit to your 00:33:27.600 --> 00:33:30.477 10,000 hours or the prepared had mind. 00:33:30.560 --> 00:33:35.597 Uh Microsoft since 1995 was obsessed. I mean, Bill's obsession for the company 00:33:35.680 --> 00:33:37.277 was natural language. Natural language. 00:33:37.360 --> 00:33:41.197 I mean, after all, we're a coding company. We're information work company. 00:33:41.280 --> 00:33:43.597 >> So, it's when Sam in 2019 started 00:33:43.680 --> 00:33:45.997 talking about text and natural language 00:33:46.080 --> 00:33:48.877 and transformers and scaling laws. 00:33:48.960 --> 00:33:53.197 >> Uh that's when I said, "Wow, like this is an interesting I mean he, you know, 00:33:53.280 --> 00:33:57.277 this is a team that was going in the direction or the direction of travel was 00:33:57.360 --> 00:34:00.157 now clear. it had a lot more overlap 00:34:00.240 --> 00:34:02.957 with our interest. So in that sense it 00:34:03.040 --> 00:34:07.997 was a no-brainer. Obviously you go to the board and say hey I have an idea of 00:34:08.080 --> 00:34:12.237 taking a billion dollars and giving it to this crazy structure which we don't 00:34:12.320 --> 00:34:14.157 even kind of understand what is it. It's 00:34:14.240 --> 00:34:16.877 a nonprofit blah blah blah and and 00:34:16.960 --> 00:34:19.677 saying go for it. Uh there was a debate. 00:34:19.760 --> 00:34:22.157 Uh Bill was kind of rightfully so 00:34:22.240 --> 00:34:27.117 skeptical because and then he became like once he saw the GPD4 demo like that 00:34:27.200 --> 00:34:31.037 was like the thing that Bill's talked about publicly where uh when he saw it 00:34:31.120 --> 00:34:33.357 he said it's the best demo he saw after 00:34:33.440 --> 00:34:35.037 you know what Charles Simony showed him 00:34:35.120 --> 00:34:38.317 at Xerox Park and but you know quite 00:34:38.400 --> 00:34:40.637 honestly none of us could uh so the 00:34:40.720 --> 00:34:43.037 moment for me was that you know let's go 00:34:43.120 --> 00:34:46.157 give it a shot then seeing the early 00:34:46.240 --> 00:34:50.157 codeex inside of uh copilot inside of uh 00:34:50.240 --> 00:34:52.717 GitHub copilot and seeing just the code 00:34:52.800 --> 00:34:55.357 completions and seeing it work. That's 00:34:55.440 --> 00:35:00.317 when I would say we I I felt like I can go from 1 to 10 because that was the big 00:35:00.400 --> 00:35:03.277 call quite frankly. One was controversial. 00:35:03.360 --> 00:35:06.877 >> Uh but the 1 to 10 was what really made 00:35:06.960 --> 00:35:09.517 this entire era possible and then 00:35:09.600 --> 00:35:14.237 obviously uh the great execution by the team and the productization on their 00:35:14.320 --> 00:35:15.997 part, our part. I mean if I think about 00:35:16.080 --> 00:35:18.797 it right the collective monetization 00:35:18.880 --> 00:35:22.237 reach of GitHub copilot chat GPT 00:35:22.320 --> 00:35:26.477 Microsoft 365 copilot and co-pilot you add those four things that is it right 00:35:26.560 --> 00:35:28.717 that's the biggest sort of AI set of 00:35:28.800 --> 00:35:33.517 products uh out there on the planet and that's um you know what obviously has 00:35:33.600 --> 00:35:37.917 let us sustain all of this and I think not many people know that your CTO Kevin 00:35:38.000 --> 00:35:42.077 Scott you know an ex googler lives down here in Silicon Valley and to 00:35:42.160 --> 00:35:44.077 contextualize it right Microsoft had 00:35:44.160 --> 00:35:46.397 missed out on search had missed out on 00:35:46.480 --> 00:35:49.197 mobile. You become CEO, almost had 00:35:49.280 --> 00:35:51.677 missed out on the cloud, right? You 00:35:51.760 --> 00:35:53.837 you've described it, caught the last 00:35:53.920 --> 00:35:56.957 train out of town to capture the cloud. 00:35:57.040 --> 00:36:00.717 And I think you were pretty determined to have eyes and ears down here so you 00:36:00.800 --> 00:36:05.277 didn't miss the next big thing. So I assume that Kevin played a good role for 00:36:05.360 --> 00:36:08.797 you as well. >> Absolutely. >> Deep Seek and Open AI. 00:36:08.880 --> 00:36:11.437 >> Yeah. I mean I mean if uh it's in fact I 00:36:11.520 --> 00:36:14.237 would say Kevin's conviction uh and 00:36:14.320 --> 00:36:16.157 Kevin was also skeptical like that was 00:36:16.240 --> 00:36:18.797 the thing I I I always watch for people 00:36:18.880 --> 00:36:22.077 who are skeptical who change uh their 00:36:22.160 --> 00:36:26.797 opinion because to me that's a signal so I'm always looking for someone who's a 00:36:26.880 --> 00:36:30.397 non-believer in something and then suddenly changes and then they get 00:36:30.480 --> 00:36:32.397 excited about it that I have all the 00:36:32.480 --> 00:36:37.597 time for that because I'm then curious why what and so Kevin started with all 00:36:37.680 --> 00:36:41.997 of us were kind of skeptical Right. No, I mean in some sense it defies the the 00:36:42.080 --> 00:36:45.837 you know we're all having gone to school and said god you know there must be an 00:36:45.920 --> 00:36:50.317 algorithm to crack this versus just let's scaling laws and throw compute. 00:36:50.400 --> 00:36:52.717 But quite frankly uh Kevin's conviction 00:36:52.800 --> 00:36:55.757 that this is worth going after is one of 00:36:55.840 --> 00:36:59.917 the big things that drove this. Well, we talk about, you know, that that 00:37:00.000 --> 00:37:03.997 investment that that's now worth 130 billion, I suppose, could be worth a 00:37:04.080 --> 00:37:06.397 trillion someday, as Sam says, but it 00:37:06.480 --> 00:37:08.077 really in many ways understates the 00:37:08.160 --> 00:37:11.517 value of the partnership, right? So, you 00:37:11.600 --> 00:37:13.677 have the value in the revshare, billions 00:37:13.760 --> 00:37:16.797 per year going to Microsoft. You have 00:37:16.880 --> 00:37:19.997 the profit you make off the $250 billion 00:37:20.080 --> 00:37:22.317 of the Azure compute commitment from 00:37:22.400 --> 00:37:25.837 OpenAI. And of course you get huge sales 00:37:25.920 --> 00:37:28.637 from the exclusive distribution of the 00:37:28.720 --> 00:37:31.277 API. So talk to us how you think about 00:37:31.360 --> 00:37:36.637 the value across those domains especially how this exclusivity has 00:37:36.720 --> 00:37:40.317 brought a lot of customers who may have been on AWS to Azure. 00:37:40.400 --> 00:37:43.117 >> Yeah. No absolutely. I mean so to us um 00:37:43.200 --> 00:37:45.917 if I look at it um you know aside from 00:37:46.000 --> 00:37:48.477 all the uh the equity parts the real 00:37:48.560 --> 00:37:51.117 strategic thing that comes together and 00:37:51.200 --> 00:37:53.917 that remains going forward uh is that 00:37:54.000 --> 00:37:56.877 stateless API exclusivity on Azure that 00:37:56.960 --> 00:37:59.437 helps quite frankly both open AAI and us 00:37:59.520 --> 00:38:01.677 and our customers uh because when 00:38:01.760 --> 00:38:03.997 somebody in the enterprise uh is trying 00:38:04.080 --> 00:38:06.637 to build an application they want an API 00:38:06.720 --> 00:38:09.357 that's stateless they want to mix it up 00:38:09.440 --> 00:38:12.477 with uh compute in storage, put a 00:38:12.560 --> 00:38:14.717 database underneath it to capture state 00:38:14.800 --> 00:38:16.797 and build a full workload and that's 00:38:16.880 --> 00:38:19.917 where uh you know Azure coming together 00:38:20.000 --> 00:38:24.877 with this API and so what we're doing with even uh Azure foundry right because 00:38:24.960 --> 00:38:28.317 in some sense you let's say you want to build an AI application but the key 00:38:28.400 --> 00:38:31.117 thing is uh how do you make sure that 00:38:31.200 --> 00:38:33.677 the eval 00:38:33.760 --> 00:38:35.677 are great so that's where you need even 00:38:35.760 --> 00:38:39.117 a full app server in Foundry that's what 00:38:39.200 --> 00:38:41.597 we've done and so therefore I feel that 00:38:41.680 --> 00:38:43.757 that is the way we will go to market in 00:38:43.840 --> 00:38:46.397 our infrastructure business. The other 00:38:46.480 --> 00:38:48.637 side of the value capture for us is 00:38:48.720 --> 00:38:51.277 going to be incorporating all this IP. 00:38:51.360 --> 00:38:53.917 Not only we have the exclusivity of the 00:38:54.000 --> 00:38:58.877 model in uh Azure but we have access to the IP. I mean having a royaltyfree 00:38:58.960 --> 00:39:03.037 let's even forgetting all the the knowhow and the knowledge side of it but 00:39:03.120 --> 00:39:07.677 having royalty-free access all the way till seven more years gives us a lot of 00:39:07.760 --> 00:39:11.837 flexibility business model wise. It's kind of like having a frontier model for 00:39:11.920 --> 00:39:15.997 free uh in some sense if you're an MSFT shareholder. That's kind of where you 00:39:16.080 --> 00:39:20.077 should start from is to think about we have a frontier model that we can then 00:39:20.160 --> 00:39:23.677 deploy whether it's in GitHub, whether it's in M365, whether it's in our 00:39:23.760 --> 00:39:26.877 consumer copilot, then add to it our own 00:39:26.960 --> 00:39:31.837 data, post train it. Uh so that means we can have it embedded in the weights 00:39:31.920 --> 00:39:33.597 there. And so therefore we're excited 00:39:33.680 --> 00:39:36.237 about the value creation on both the 00:39:36.320 --> 00:39:38.877 Azure and the infrastructure side as 00:39:38.960 --> 00:39:43.437 well as in our high value domains uh whether it is in health whether it's in 00:39:43.520 --> 00:39:46.077 knowledge work whether it's in coding or security 00:39:46.160 --> 00:39:49.997 >> you've been consolidating the losses from open AI you know I think you you 00:39:50.080 --> 00:39:54.077 just reported earnings yesterday I think you consolidated 4 billion of losses in 00:39:54.160 --> 00:39:58.957 the quarter do you think that investors are I mean they may even be attributing 00:39:59.040 --> 00:40:02.397 negative value right because of the losses you know as they apply their 00:40:02.480 --> 00:40:06.877 multiple of earnings. Satcha, whereas I hear this and I think about all of those 00:40:06.960 --> 00:40:11.437 benefits we just described, not to mention the look through equity value 00:40:11.520 --> 00:40:15.757 that you own in a company that could be worth a trillion unto itself. You know, 00:40:15.840 --> 00:40:17.917 do you think that the market is is is 00:40:18.000 --> 00:40:22.957 kind of misunderstanding the value of open AI as a component of Microsoft? 00:40:23.040 --> 00:40:26.797 >> Yeah, that's a good one. So, I think the the approach that Amy is going to take 00:40:26.880 --> 00:40:31.597 is full transparency because at some level I'm no accounting expert. So 00:40:31.680 --> 00:40:36.397 therefore the best thing to do is to give uh all of the transparency I think 00:40:36.480 --> 00:40:40.077 this time around as well. I think that's why the non-GAAP gap so that at least 00:40:40.160 --> 00:40:44.877 people can see the EPS numbers because the the the common sense way I look at 00:40:44.960 --> 00:40:46.797 it Brad is simple. If you've invested 00:40:46.880 --> 00:40:49.517 let's call it 13.5 billion. You can of 00:40:49.600 --> 00:40:52.397 course lose 13.5 billion but you can't 00:40:52.480 --> 00:40:56.797 lose more than 13.5 billion. At least the last time I checked that's what you 00:40:56.880 --> 00:40:58.957 have at risk. You could also say hey the 00:40:59.040 --> 00:41:02.397 $135 billion that has you know today our 00:41:02.480 --> 00:41:05.357 equity stake you know is sort of illquid 00:41:05.440 --> 00:41:09.837 what have you we don't plan to sell it so therefore it's got risk associated 00:41:09.920 --> 00:41:14.717 with it but the real story I think you were pulling is all the other things uh 00:41:14.800 --> 00:41:18.317 that are happening what's happening with Azure growth right would Azure be 00:41:18.400 --> 00:41:22.637 growing if we had not sort of had the openi partnership to your point the 00:41:22.720 --> 00:41:25.837 number of customers who came from other clouds 00:41:25.920 --> 00:41:29.997 for the first time right this is the thing that really we benefited from 00:41:30.080 --> 00:41:33.757 what's happening with Microsoft 365. In fact, one of the things about Microsoft 00:41:33.840 --> 00:41:35.837 365 was what was the next big thing 00:41:35.920 --> 00:41:38.637 after E5? Guess what? We found it in 00:41:38.720 --> 00:41:41.997 copilot. It's bigger than any suite. 00:41:42.080 --> 00:41:44.637 Like you know, we talk about penetration 00:41:44.720 --> 00:41:49.837 and usage uh and the pace. It's bigger than anything we've done in our 00:41:49.920 --> 00:41:53.997 information work which we've been added for decades. And so so we pretty feel 00:41:54.080 --> 00:41:58.317 very very good about the opportunity to create value for our shareholders. Uh 00:41:58.400 --> 00:42:01.757 and then at the same time be fully transparent so that people can look 00:42:01.840 --> 00:42:05.197 through the what are the losses. I mean who knows what the accounting rules are 00:42:05.280 --> 00:42:09.037 but we will do whatever is needed and people will then be able to see what's 00:42:09.120 --> 00:42:13.357 happening. But a year ago, Satcha, there were a bunch of headlines that Microsoft 00:42:13.440 --> 00:42:17.517 was pulling back on AI infrastructure, right? Fair or unfair, they're they were 00:42:17.600 --> 00:42:20.797 out there, you know, and and and perhaps you guys were a little more 00:42:20.880 --> 00:42:25.837 conservative, a little more skeptical of what was going on. Amy said on the call 00:42:25.920 --> 00:42:29.277 last night, though, that you've been short power and infrastructure for many 00:42:29.360 --> 00:42:33.277 quarters, and she thought that you would catch up, but you haven't c caught up 00:42:33.360 --> 00:42:36.957 because demand keeps increasing. So I guess the question is were you too 00:42:37.040 --> 00:42:40.957 conservative you know knowing what you know now and and and what's the road map 00:42:41.040 --> 00:42:43.437 from here? >> Yeah it's a great question because see 00:42:43.520 --> 00:42:46.877 the the thing that we realized and I'm 00:42:46.960 --> 00:42:50.477 glad we did uh is that the concept of 00:42:50.560 --> 00:42:54.237 building a fleet that truly was funible 00:42:54.320 --> 00:42:56.317 fungeible for all the parts of the life 00:42:56.400 --> 00:42:59.997 cycle of AI funible across geographies 00:43:00.080 --> 00:43:03.037 and fungeible across generations. Right? 00:43:03.120 --> 00:43:07.117 So because one of the key things is when you have let's take even uh what Jensen 00:43:07.200 --> 00:43:10.877 and team are doing right I mean they're at a pace in fact one of the things I 00:43:10.960 --> 00:43:14.797 like is the speed of light right we now have GB300's bringing you know that 00:43:14.880 --> 00:43:16.477 we're bringing up so you don't want to 00:43:16.560 --> 00:43:20.077 have ordered a bunch of GB200's that are 00:43:20.160 --> 00:43:22.397 getting plugged in only to find that 00:43:22.480 --> 00:43:25.197 GB2300s are in full production. So you 00:43:25.280 --> 00:43:27.197 kind of have to make sure you're 00:43:27.280 --> 00:43:29.757 continuously modernizing, you're 00:43:29.840 --> 00:43:32.397 spreading the fleet all over, you are 00:43:32.480 --> 00:43:35.357 really truly funible by workload uh and 00:43:35.440 --> 00:43:39.917 you're adding to that the software optimizations we talked about. So to me 00:43:40.000 --> 00:43:44.957 that is the decision we made and we said look sometimes you may have to say no to 00:43:45.040 --> 00:43:48.717 some of the demand including some of the open AI demand right because sometimes 00:43:48.800 --> 00:43:50.717 you know Sam may say hey we build me a 00:43:50.800 --> 00:43:53.437 dedicated you know big you know whatever 00:43:53.520 --> 00:43:55.997 multi- gigawatt data center in one 00:43:56.080 --> 00:43:58.637 location for training makes sense from 00:43:58.720 --> 00:44:03.757 an open AI perspective doesn't make sense from a long-term infrastructure 00:44:03.840 --> 00:44:07.277 buildout for Azure and that's where I thought they did the right thing to give 00:44:07.360 --> 00:44:09.677 them flexibility to go procure that from 00:44:09.760 --> 00:44:14.477 others while m maintaining uh again a significant book of business from open 00:44:14.560 --> 00:44:16.077 AAI but more importantly giving 00:44:16.160 --> 00:44:21.677 ourselves the flexibility with other customers our own one P remember like 00:44:21.760 --> 00:44:25.837 one of the things that we don't want to do is be short on uh is you know we talk 00:44:25.920 --> 00:44:29.517 about Azure in fact some of times our investors are overly fixated on the 00:44:29.600 --> 00:44:34.397 Azure number but remember for me the high margin business for me is co-pilot 00:44:34.480 --> 00:44:36.317 it is security co-pilot it's GitHub 00:44:36.400 --> 00:44:39.197 co-pilot it's the healthcare co-pilot So 00:44:39.280 --> 00:44:43.597 we want to make sure we have a balanced way to approach the returns that the 00:44:43.680 --> 00:44:47.837 investors have. And so that's kind of one of the other misunderstood perhaps 00:44:47.920 --> 00:44:50.957 in our investor base in particular, which I find pretty strange and funny 00:44:51.040 --> 00:44:55.437 because I think they they want to hold Microsoft because of the portfolio we 00:44:55.520 --> 00:44:57.837 have. But man are they fixated on the 00:44:57.920 --> 00:44:59.517 growth number of one little thing called 00:44:59.600 --> 00:45:04.477 Azure. On that point, Azure grew 39% in 00:45:04.560 --> 00:45:07.837 the quarter on a staggering $93 billion 00:45:07.920 --> 00:45:09.677 run rate. And you know, I think that 00:45:09.760 --> 00:45:13.597 compares to GCP that grew at 32% and AWS 00:45:13.680 --> 00:45:17.197 closer to 20%. But could Azure because 00:45:17.280 --> 00:45:20.477 you did give compute to 1P and because 00:45:20.560 --> 00:45:23.037 you did give compute to research, it 00:45:23.120 --> 00:45:25.277 sounds like Azure could have grown 41 00:45:25.360 --> 00:45:28.157 42% had you had more compute to offer. 00:45:28.240 --> 00:45:31.277 >> Absolutely. Absolutely. There's no question. There is no question. So 00:45:31.360 --> 00:45:35.837 that's why I think the internal thing is to balance out what we think again is in 00:45:35.920 --> 00:45:39.597 the long-term interests of our shareholders and uh and also to serve 00:45:39.680 --> 00:45:43.677 our customers well and also not to kind of you know one of the other things was 00:45:43.760 --> 00:45:47.437 you know people talk about concentration risk right we obviously want a lot of 00:45:47.520 --> 00:45:52.717 open AI but we also want other customer and so we're shaping the demand here you 00:45:52.800 --> 00:45:56.077 know we are in a supply you know you know we're not demand constraint we're 00:45:56.160 --> 00:46:01.197 supply constraint so we are shaping the demand such that it matches is the 00:46:01.280 --> 00:46:04.637 supply in the optimal way with the long-term uh view. 00:46:04.720 --> 00:46:08.957 >> To that point, Satcha, you you talked about 400 billion. It's incredible 00:46:09.040 --> 00:46:13.997 number of remaining performance obligations. Last night, you said that, 00:46:14.080 --> 00:46:17.997 you know, that's your booked business today. It'll surely go up tomorrow as 00:46:18.080 --> 00:46:21.997 sales continue to come in. And you said you're going to, you know, your need to 00:46:22.080 --> 00:46:24.237 build out capacity just to serve that 00:46:24.320 --> 00:46:26.877 backlog is very high. You know, how 00:46:26.960 --> 00:46:29.357 diversified is that backlog to your to 00:46:29.440 --> 00:46:34.157 your point? And how confident are you that that 400 billion does turn into 00:46:34.240 --> 00:46:37.037 revenue over the course of the next couple years? 00:46:37.120 --> 00:46:39.917 >> Yeah, that that 400 billion uh has a 00:46:40.000 --> 00:46:41.997 very short duration as Amy explained. 00:46:42.080 --> 00:46:45.197 It's the 2-year uh duration on average. 00:46:45.280 --> 00:46:47.117 So that's definitely uh our intent. 00:46:47.200 --> 00:46:51.117 That's one of the reasons why uh we're spending the capital outcllay with high 00:46:51.200 --> 00:46:54.797 certainty that we just need to clear this backlog. And to your point, it's 00:46:54.880 --> 00:46:57.117 pretty diversified both on the 1 P and 00:46:57.200 --> 00:46:59.677 the 3P. our own demand is quite frankly 00:46:59.760 --> 00:47:02.157 pretty high for our one first party uh 00:47:02.240 --> 00:47:04.157 and even amongst third party one of the 00:47:04.240 --> 00:47:07.677 things we now are seeing is the the rise 00:47:07.760 --> 00:47:12.557 of all the other companies building real workloads uh that are scaling uh and so 00:47:12.640 --> 00:47:16.797 given that I think we feel very good I mean obviously it's uh that's one of the 00:47:16.880 --> 00:47:20.717 best things about RPO is you can be planful quite frankly and so therefore 00:47:20.800 --> 00:47:24.077 we feel very very good about building and then this doesn't include obviously 00:47:24.160 --> 00:47:29.037 the additional demand that we're already going to start seeing including the 250 00:47:29.120 --> 00:47:32.877 uh you know which will have a longer duration and we'll build accordingly 00:47:32.960 --> 00:47:35.837 >> right so there are a lot of new entrance 00:47:35.920 --> 00:47:38.237 right uh in this race to build out 00:47:38.320 --> 00:47:41.757 compute Oracle coreweave cruso etc and 00:47:41.840 --> 00:47:46.237 normally we think that will compete away margins but you've somehow managed to 00:47:46.320 --> 00:47:48.317 build all this out while maintaining 00:47:48.400 --> 00:47:50.717 healthy operating margins at Azure so I 00:47:50.800 --> 00:47:53.437 guess the question is for Microsoft how 00:47:53.520 --> 00:47:56.797 do you compete in this world that is uh 00:47:56.880 --> 00:48:01.277 where people are levering up, taking lower margins while balancing that 00:48:01.360 --> 00:48:05.997 profit and and and risk. And do you see any of those competitors doing deals 00:48:06.080 --> 00:48:09.357 that cause you to scratch your head and say, "Oh, we're just setting ourselves 00:48:09.440 --> 00:48:11.757 up for another boom and bust cycle." 00:48:11.840 --> 00:48:13.757 >> I mean, I'd say at some level the the 00:48:13.840 --> 00:48:19.037 good news for us has been competing even as a hyperscaler every day. You know, 00:48:19.120 --> 00:48:23.197 there's a lot of competition, right, between us and Amazon and Google on all 00:48:23.280 --> 00:48:26.477 of these, right? I mean it's sort of one of those interesting things which is 00:48:26.560 --> 00:48:30.397 everything is a commodity right compute storage I remember everybody saying wow 00:48:30.480 --> 00:48:32.957 how can there be a margin except at 00:48:33.040 --> 00:48:37.837 scale nothing is a commodity um and so therefore yes so we have to have our 00:48:37.920 --> 00:48:39.597 cost structure our supply chain 00:48:39.680 --> 00:48:43.037 efficiency our software efficiencies all 00:48:43.120 --> 00:48:46.077 have to kind of continue to compound in 00:48:46.160 --> 00:48:50.717 order to make sure that there's margins uh but scale and to your point one of 00:48:50.800 --> 00:48:52.957 the things that I really love about the 00:48:53.040 --> 00:48:55.277 OpenAI partnership is it's gotten us to 00:48:55.360 --> 00:48:58.157 scale, right? This is a scale game. When 00:48:58.240 --> 00:49:00.877 you have uh the biggest workload there 00:49:00.960 --> 00:49:05.357 is running on your cloud, that means not only are we going to learn faster on 00:49:05.440 --> 00:49:08.397 what it means to operate with scale, that means your cost structure is going 00:49:08.480 --> 00:49:10.477 to come down faster than anything else. 00:49:10.560 --> 00:49:14.237 And guess what? That'll make us price competitive. And so I feel pretty 00:49:14.320 --> 00:49:18.957 confident about our ability to, you know, have margins. And and that this is 00:49:19.040 --> 00:49:21.677 where the portfolio helps. I've always said 00:49:21.760 --> 00:49:26.317 >> you know you know I've been forced into giving the Azure numbers right because 00:49:26.400 --> 00:49:28.157 at some level I never thought of 00:49:28.240 --> 00:49:30.797 allocating I mean my capital allocation 00:49:30.880 --> 00:49:34.237 is for the cloud from whether it is Xbox 00:49:34.320 --> 00:49:38.317 cloud gaming or Microsoft 365 or for 00:49:38.400 --> 00:49:41.117 Azure it's one capital outlay uh and 00:49:41.200 --> 00:49:45.917 then everything is a meter as far as I'm concerned from an MSF perspective it's a 00:49:46.000 --> 00:49:48.077 question of hey the blended average of 00:49:48.160 --> 00:49:50.877 that should match the operating margins 00:49:50.960 --> 00:49:55.197 we need as a company because after all otherwise why we're not a conglomerate 00:49:55.280 --> 00:49:59.357 we're one company with one platform logic it's not running five six 00:49:59.440 --> 00:50:03.197 different businesses we're in these five six different businesses only to 00:50:03.280 --> 00:50:05.997 compound the returns on the cloud and AI 00:50:06.080 --> 00:50:10.317 investment >> yeah I I love that line uh nothing is a 00:50:10.400 --> 00:50:14.957 commodity at scale you know there's been a lot of ink and time spent even on this 00:50:15.040 --> 00:50:19.037 podcast with my partner Bill Gurley talking about circular revenues 00:50:19.120 --> 00:50:23.117 including including Microsoft Stasher credits right to OpenAI that were booked 00:50:23.200 --> 00:50:25.997 as revenue. Do you see anything going on 00:50:26.080 --> 00:50:28.317 like the AMD deal, you know, where they 00:50:28.400 --> 00:50:30.477 traded 10% of their equity and, you 00:50:30.560 --> 00:50:33.437 know, for a deal or the Nvidia deal? 00:50:33.520 --> 00:50:37.277 Again, I don't want to be overly fixated on concern, but I do want to address 00:50:37.360 --> 00:50:41.917 headon what is uh being talked about every day on CNBC and Bloomberg and 00:50:42.000 --> 00:50:45.997 there are a lot of these overlapping deals that are going on out there. Do 00:50:46.080 --> 00:50:50.157 you do you when you think about that in the context of Microsoft does any of 00:50:50.240 --> 00:50:52.477 that worry you again as to the 00:50:52.560 --> 00:50:56.157 sustainability or durability of uh the 00:50:56.240 --> 00:50:58.157 AI revenues that we see in the world? 00:50:58.240 --> 00:51:00.717 >> Yeah. I mean first of all our investment 00:51:00.800 --> 00:51:05.597 of uh let's say that 13 and a half which was all the training investment that was 00:51:05.680 --> 00:51:10.077 not booked as revenue. That is the that is the reason why we have the equity 00:51:10.160 --> 00:51:12.557 percentage. That's the reason why we 00:51:12.640 --> 00:51:15.517 have the 27% or 135 billion. So that was 00:51:15.600 --> 00:51:19.757 not something some that somehow that made it into Azure revenue. In fact, if 00:51:19.840 --> 00:51:21.997 anything, the Azure revenue was purely 00:51:22.080 --> 00:51:25.757 the consumption revenue of chat GPT and 00:51:25.840 --> 00:51:28.157 anything else and the APIs they put out 00:51:28.240 --> 00:51:30.637 that they monetized and we monetized 00:51:30.720 --> 00:51:33.677 >> to your aspect of others. You know, to 00:51:33.760 --> 00:51:37.837 some degree, it's always been there in terms of vendor financing, right? So 00:51:37.920 --> 00:51:42.077 it's not like a new concept that when someone's building something and they 00:51:42.160 --> 00:51:44.637 have a customer who is also building 00:51:44.720 --> 00:51:49.357 something but they need financing you know for whether it is in you know it's 00:51:49.440 --> 00:51:53.517 it's sort of some they're taking some exotic forms uh which obviously need to 00:51:53.600 --> 00:51:55.277 be scrutinized by the investment 00:51:55.360 --> 00:51:57.997 community but that said you know vendor 00:51:58.080 --> 00:52:02.477 financing is not a new concept interestingly enough we have not had to 00:52:02.560 --> 00:52:04.557 do any of that right I mean we may have 00:52:04.640 --> 00:52:07.357 you know really uh either invested in 00:52:07.440 --> 00:52:10.477 OpenAI and essentially got an equity uh 00:52:10.560 --> 00:52:13.357 stake in it for return for compute or 00:52:13.440 --> 00:52:17.997 essentially sold them great pricing of compute in order to be able to sort of 00:52:18.080 --> 00:52:22.397 bootstrap them. But you know others choose to do so differently and uh and I 00:52:22.480 --> 00:52:24.557 think circularity ultimately will be 00:52:24.640 --> 00:52:27.037 tested by demand because all this will 00:52:27.120 --> 00:52:29.917 work uh as long as there is demand for 00:52:30.000 --> 00:52:33.117 the final out output of it and up to now 00:52:33.200 --> 00:52:37.117 that has been the case. Certainly, certainly. Well, I want to shift uh you 00:52:37.200 --> 00:52:41.117 know, as you said, over half your business is software uh applications. 00:52:41.200 --> 00:52:44.477 You know, I want to think about software and agents. You know, last year on this 00:52:44.560 --> 00:52:49.437 pod, you made a bit of a stir by saying that much of application software, you 00:52:49.520 --> 00:52:51.277 know, was this thin layer that sat sat 00:52:51.360 --> 00:52:54.877 on top of a CRUD database. The notion 00:52:54.960 --> 00:52:59.757 that business applications exist, 00:52:59.840 --> 00:53:03.437 that's probably where they'll all collapse, right, in the agent era. 00:53:03.520 --> 00:53:04.797 Because if you think about it right, 00:53:04.880 --> 00:53:07.277 they are essentially 00:53:07.360 --> 00:53:10.717 crowd databases with a bunch of business 00:53:10.800 --> 00:53:12.237 logic. 00:53:12.320 --> 00:53:15.997 The business logic is all going to these 00:53:16.080 --> 00:53:18.637 agents. Public software companies are 00:53:18.720 --> 00:53:21.677 now trading at about 5.2 times forward 00:53:21.760 --> 00:53:25.917 revenue. So that's below their 10-year average of seven times despite the 00:53:26.000 --> 00:53:29.757 markets being at all-time highs. And there's lots of concern that SAS 00:53:29.840 --> 00:53:32.157 subscriptions and margins may be put at 00:53:32.240 --> 00:53:37.357 risk by AI. So how today is AI affecting 00:53:37.440 --> 00:53:41.517 the growth rates of your software products of you know those core products 00:53:41.600 --> 00:53:43.277 and specifically as you think about 00:53:43.360 --> 00:53:47.357 database fabric security office 360 and 00:53:47.440 --> 00:53:52.237 then second question I guess is what are you doing to make sure that software is 00:53:52.320 --> 00:53:56.957 not disrupted but is instead superpowered by AI? Yeah, I think that's 00:53:57.040 --> 00:54:01.437 a Yeah, that's right. So, the last time we talked about this, my my point really 00:54:01.520 --> 00:54:03.757 there was the architecture of SAS 00:54:03.840 --> 00:54:08.717 applications is changing because this agent tier is replacing the old business 00:54:08.800 --> 00:54:11.917 logic tier. And so, because if you think about it, the way we built SAS 00:54:12.000 --> 00:54:15.997 applications in the past was you had the data, the logic tier, and the UI all 00:54:16.080 --> 00:54:20.557 tightly coupled. Uh, and AI quite frankly doesn't respect that coupling 00:54:20.640 --> 00:54:25.437 because it requires you to be able to decouple. And yet the context 00:54:25.520 --> 00:54:28.237 engineering is going to be very important. I mean take you know 00:54:28.320 --> 00:54:30.477 something like uh office 365. One of the 00:54:30.560 --> 00:54:33.277 things I love about uh our Microsoft 365 00:54:33.360 --> 00:54:36.717 offering is it's low arpoo 00:54:36.800 --> 00:54:40.957 uh high usage right I mean if you think about it right outlook or teams or 00:54:41.040 --> 00:54:45.517 sharepoint you pick word or excel like people are using it all the time 00:54:45.600 --> 00:54:47.437 creating lots and lots of data which is 00:54:47.520 --> 00:54:50.077 going into the graph and our arpoo is 00:54:50.160 --> 00:54:52.717 low. So that's sort of what gives me 00:54:52.800 --> 00:54:56.237 real confidence that this AI tier with I 00:54:56.320 --> 00:54:59.597 can meet it by exposing all my data. In 00:54:59.680 --> 00:55:01.037 fact, one of the fascinating things 00:55:01.120 --> 00:55:03.917 that's happened uh Brad with both GitHub 00:55:04.000 --> 00:55:07.597 and Microsoft 365 is thanks to AI, we 00:55:07.680 --> 00:55:12.397 are seeing alltime highs in terms of data that's going into the graph or the 00:55:12.480 --> 00:55:14.797 repo. >> I mean think about it. The more code 00:55:14.880 --> 00:55:16.477 that gets generated, whether it is 00:55:16.560 --> 00:55:19.437 codeex or cloud or wherever, where is it 00:55:19.520 --> 00:55:24.477 going? GitHub, more PowerPoints that get created, Excel models that get created, 00:55:24.560 --> 00:55:26.317 all these artifacts and chat 00:55:26.400 --> 00:55:28.237 conversations. Chat conversations are 00:55:28.320 --> 00:55:31.597 new docs, they're all going in to the 00:55:31.680 --> 00:55:34.557 graph and and all that is needed again 00:55:34.640 --> 00:55:37.357 >> for grounding. Uh so that's what you 00:55:37.440 --> 00:55:39.517 know you turn it into a forward index 00:55:39.600 --> 00:55:42.077 into an embedding and basically that 00:55:42.160 --> 00:55:45.277 semantics is what you really go ground 00:55:45.360 --> 00:55:48.397 any agent request. And so I think the 00:55:48.480 --> 00:55:50.637 next generation of SAS applications will 00:55:50.720 --> 00:55:53.517 have to sort of if you are high RPO low 00:55:53.600 --> 00:55:55.837 usage then you have a little bit of a 00:55:55.920 --> 00:55:58.237 problem. But if you are we are the exact 00:55:58.320 --> 00:56:01.197 opposite. we are low RPO, high usage and 00:56:01.280 --> 00:56:05.917 I think that anyone who can structure that and then use this AI as in fact an 00:56:06.000 --> 00:56:10.637 accelerant because I mean like if you look at the M365 copilot price I mean 00:56:10.720 --> 00:56:12.637 it's higher than any other thing that we 00:56:12.720 --> 00:56:14.717 sell and yet it's getting deployed 00:56:14.800 --> 00:56:17.917 faster and with more usage and so I feel 00:56:18.000 --> 00:56:20.157 very good oh or coding right who would 00:56:20.240 --> 00:56:25.277 have thought in fact take GitHub right what GitHub did in first I don't know 15 00:56:25.360 --> 00:56:29.197 years of its existence or 10 years of its existence it was basically done in 00:56:29.280 --> 00:56:31.277 the last year just because coding is no 00:56:31.360 --> 00:56:34.717 longer a tool. It's more a substitute 00:56:34.800 --> 00:56:39.517 for wages and so it's a very different type of business model even kind of 00:56:39.600 --> 00:56:43.277 thinking about the stack and where value gets distributed. So until very 00:56:43.360 --> 00:56:48.317 recently, right, clouds largely ran pre-ompiled software. You didn't need a 00:56:48.400 --> 00:56:51.677 lot of GPUs and most of the value acrewed to the software layer to the 00:56:51.760 --> 00:56:56.077 database to the applications like CRM and Excel. But it does seem in the 00:56:56.160 --> 00:57:00.557 future that these interfaces will only be valuable, right? If they're if 00:57:00.640 --> 00:57:03.837 they're uh intelligent, right? If they're pre-ompiled, they're kind of 00:57:03.920 --> 00:57:05.917 dumb. The software's got to be able to 00:57:06.000 --> 00:57:09.117 think and to act and to advise. And that 00:57:09.200 --> 00:57:13.277 requires you know the production of these tokens you know dealing with the 00:57:13.360 --> 00:57:17.677 everchanging context. And so in that world it does seem like much more of the 00:57:17.760 --> 00:57:20.237 value will acrue to the AI factory if 00:57:20.320 --> 00:57:23.917 you will to you know Jensen producing 00:57:24.000 --> 00:57:25.837 you know uh helping to produce these 00:57:25.920 --> 00:57:29.037 tokens at uh uh the lowest cost and to 00:57:29.120 --> 00:57:33.197 the models and maybe that the agents or the software will acrue a little bit 00:57:33.280 --> 00:57:36.957 less of the value in the future than they've accured in the in the past. 00:57:37.040 --> 00:57:39.677 Well, steelman for me. Why that's wrong? 00:57:39.760 --> 00:57:45.037 >> Yeah. So, I think there are two things that are necessary to try and to drive 00:57:45.120 --> 00:57:48.077 the value of AI. One is what you described first, which is the token 00:57:48.160 --> 00:57:51.917 factory. And even [clears throat] if you unpack the token factory, uh it's the 00:57:52.000 --> 00:57:54.877 hardware silicon system, but then it is 00:57:54.960 --> 00:57:57.837 about running it most efficiently with 00:57:57.920 --> 00:57:59.837 the system software with all the 00:57:59.920 --> 00:58:03.277 fungibility, max utilization. That's 00:58:03.360 --> 00:58:05.277 where the hyperscaler's role is, right? 00:58:05.360 --> 00:58:09.357 What is a hyperscaler? Is hyperscaler like everybody says if you sort of said 00:58:09.440 --> 00:58:12.957 hey I want to run a hyperscaler. Yeah you could say oh it's simple. I'll buy a 00:58:13.040 --> 00:58:16.237 bunch of servers and wire them up and run it. It's not that right. I mean it 00:58:16.320 --> 00:58:19.197 was that simple then there would have been more than three hyperscalers by 00:58:19.280 --> 00:58:22.397 now. So the hyperscaler is the knowhow 00:58:22.480 --> 00:58:27.357 of running that max util and the token factories. And it's not and by the way 00:58:27.440 --> 00:58:28.477 it's going to be heterogeneous. 00:58:28.560 --> 00:58:30.397 Obviously Jensen's super competitive. 00:58:30.480 --> 00:58:34.637 Lisa is going to come, you know, Hawk's going to produce things uh from 00:58:34.720 --> 00:58:38.557 Broadcom. We will all do our own. So there's going to be a combination. So 00:58:38.640 --> 00:58:40.237 you want to run ultimately a 00:58:40.320 --> 00:58:43.357 heterogeneous fleet that is maximized 00:58:43.440 --> 00:58:47.837 for token throughput and efficiency and so on. So that's kind of one job. The 00:58:47.920 --> 00:58:49.997 next thing is what I call the agent 00:58:50.080 --> 00:58:52.717 factory. Remember that a SAS application 00:58:52.800 --> 00:58:54.877 in the modern world is driving a 00:58:54.960 --> 00:58:58.077 business outcome. it knows how to most 00:58:58.160 --> 00:59:01.517 efficiently use the tokens to create 00:59:01.600 --> 00:59:06.077 some business value. Uh in fact, GitHub copilot is a great example of it, right? 00:59:06.160 --> 00:59:10.317 Which is, you know, if you think about it, it the auto mode of GitHub copilot 00:59:10.400 --> 00:59:12.157 is the smartest thing we've done, right? 00:59:12.240 --> 00:59:15.197 So, it chooses based on the prompt which 00:59:15.280 --> 00:59:18.317 model to use for a code completion or a 00:59:18.400 --> 00:59:23.037 task handoff, right? That's what you and you do that not just by, you know, 00:59:23.120 --> 00:59:27.117 choosing in some roundrobin fashion. You do it because of the feedback cycle. You 00:59:27.200 --> 00:59:29.597 have you have the eval, the data loops 00:59:29.680 --> 00:59:34.237 and so on. So the new SAS applications as you rightfully said are intelligent 00:59:34.320 --> 00:59:36.477 applications that are optimized for a 00:59:36.560 --> 00:59:39.357 set of evals and a set of outcomes that 00:59:39.440 --> 00:59:41.837 then know how to use the token facto's 00:59:41.920 --> 00:59:44.797 output most efficiently. Sometimes 00:59:44.880 --> 00:59:47.437 latency matters, sometimes uh 00:59:47.520 --> 00:59:49.437 performance matters and knowing how to 00:59:49.520 --> 00:59:52.557 do that trade uh in a smart way is where 00:59:52.640 --> 00:59:57.037 the SAS application value is. But overall it is going to be true that 00:59:57.120 --> 01:00:01.197 there is a real marginal cost to software this time around. It was there 01:00:01.280 --> 01:00:03.277 in the cloud era too when we were doing 01:00:03.360 --> 01:00:08.397 you know CDROMs there wasn't much of a marginal cost you know with the cloud 01:00:08.480 --> 01:00:11.757 there was and this time around it's a lot more and so therefore the business 01:00:11.840 --> 01:00:16.077 models have to adjust and you have to do these optimizations for the agent 01:00:16.160 --> 01:00:19.517 factory and the token factory separately. you have a big search 01:00:19.600 --> 01:00:23.357 business that most people don't know about, you know, but it turns out that 01:00:23.440 --> 01:00:26.237 that's probably one of the most profitable businesses in the history of 01:00:26.320 --> 01:00:28.637 the world because people are running 01:00:28.720 --> 01:00:30.477 lots of searches, billions of searches, 01:00:30.560 --> 01:00:33.277 and the cost of completing a search if 01:00:33.360 --> 01:00:37.837 you're Microsoft is many fractions of a penny, right? Doesn't cost very much to 01:00:37.920 --> 01:00:40.237 complete a search, but the comparable 01:00:40.320 --> 01:00:43.277 query or prompt stack today when when 01:00:43.360 --> 01:00:45.597 you use a chatbot looks different, 01:00:45.680 --> 01:00:49.037 right? So I guess the question is 01:00:49.120 --> 01:00:51.357 assume similar levels of revenue in the 01:00:51.440 --> 01:00:53.997 future for those two businesses, right? 01:00:54.080 --> 01:00:56.077 Do you ever get to a point where kind of 01:00:56.160 --> 01:00:59.117 that chat interaction has unit economics 01:00:59.200 --> 01:01:01.917 that are as profitable as search? I 01:01:02.000 --> 01:01:03.837 think that's a great point because see 01:01:03.920 --> 01:01:07.837 search was pretty magical uh in terms of 01:01:07.920 --> 01:01:11.437 its ad unit uh and its cost economics 01:01:11.520 --> 01:01:13.997 because there was the index which was a 01:01:14.080 --> 01:01:16.477 fixed cost that you could then amortize 01:01:16.560 --> 01:01:19.037 in a much more efficient way 01:01:19.120 --> 01:01:21.517 >> uh whereas this one you know each chat 01:01:21.600 --> 01:01:26.157 uh to your point you have to burn a lot more GPU cycles uh both with the intent 01:01:26.240 --> 01:01:30.157 and the retrieval so the economics are different so I think you do that's why I 01:01:30.240 --> 01:01:34.077 think a lot of the early sort of economics of chat have been the premium 01:01:34.160 --> 01:01:38.397 model and subscription on the even on the consumer side. So we are yet to 01:01:38.480 --> 01:01:42.957 discover whether it's agentic commerce or whatever is the ad unit how it's 01:01:43.040 --> 01:01:45.117 going to be litigated but at the same 01:01:45.200 --> 01:01:47.837 time the fact that at this point you 01:01:47.920 --> 01:01:51.357 know I kind of know in fact I use search 01:01:51.440 --> 01:01:53.757 uh for very very specific navigational 01:01:53.840 --> 01:01:56.477 queries I used to say I use it a lot for 01:01:56.560 --> 01:01:59.197 commerce but that's also shifting to my 01:01:59.280 --> 01:02:01.277 you know co-pilot like I look at the 01:02:01.360 --> 01:02:05.757 co-pilot mode in edge and bing uh or 01:02:05.840 --> 01:02:09.597 copilot now they're blend ending in. So I think that yes, I think there is going 01:02:09.680 --> 01:02:13.837 to be a relitigation just like that we talked about the SAS disruption. We're 01:02:13.920 --> 01:02:18.557 in the beginning of the cheese being a little moved in consumer economics of 01:02:18.640 --> 01:02:22.077 that category, >> right? I I mean and given that it's the 01:02:22.160 --> 01:02:25.757 multi-trillion dollar this this is the thing that's driven all the economics of 01:02:25.840 --> 01:02:27.997 the internet, right? when you move the 01:02:28.080 --> 01:02:30.877 economics of search for both you and 01:02:30.960 --> 01:02:32.797 Google and it converges on something 01:02:32.880 --> 01:02:35.037 that looks more like a personal agent, a 01:02:35.120 --> 01:02:39.917 personal assistant chat. Um, you know, that could end up being much much bigger 01:02:40.000 --> 01:02:43.757 in terms of the total value delivered to humanity, but the unit economics, you're 01:02:43.840 --> 01:02:46.877 not just advertising this one time fixed index. 01:02:46.960 --> 01:02:49.517 >> That's right. >> And so, that's right. I think that the 01:02:49.600 --> 01:02:52.477 consumer could be worse. Yeah. the consumer category because you are 01:02:52.560 --> 01:02:55.517 pulling a thread on something that I think a lot about right which is what 01:02:55.600 --> 01:02:57.757 during these disruptions 01:02:57.840 --> 01:02:59.757 you you kind of have to have a real 01:02:59.840 --> 01:03:02.637 sense of where is is the what is the 01:03:02.720 --> 01:03:05.517 category economics uh is it winner take 01:03:05.600 --> 01:03:09.677 all um uh and both matter uh right the 01:03:09.760 --> 01:03:13.277 the problem on consumer space always is 01:03:13.360 --> 01:03:16.717 that there's finite amount of time uh 01:03:16.800 --> 01:03:20.557 and so if I'm not doing one thing uh I'm 01:03:20.640 --> 01:03:24.877 doing something else and if your monetization is predicated on some human 01:03:24.960 --> 01:03:28.717 interaction in particular if there was truly agentic stuff even on consumer 01:03:28.800 --> 01:03:32.637 that could be different. Uh whereas in the enterprise one is it's not winner 01:03:32.720 --> 01:03:35.357 take all and two it is going to be a lot 01:03:35.440 --> 01:03:38.157 more friendly for agentic interaction. 01:03:38.240 --> 01:03:42.477 So it's not like for example the per seat versus consumption. The reality is 01:03:42.560 --> 01:03:44.957 agents are the new seats. 01:03:45.040 --> 01:03:47.437 >> And so you can think of it as uh the 01:03:47.520 --> 01:03:50.157 enterprise monetization is much clearer. 01:03:50.240 --> 01:03:53.997 The consumer monetization I think is a little more murky. You know, we've seen 01:03:54.080 --> 01:03:58.157 a spade of layoffs recently with Amazon announcing some big big layoffs this 01:03:58.240 --> 01:04:02.317 week. You know, the Mag 7 has had little job growth over the last three years 01:04:02.400 --> 01:04:06.317 despite really robust top lines. You know, you didn't grow your headcount 01:04:06.400 --> 01:04:10.557 really from 24 to 25. It's around 225,000. 01:04:10.640 --> 01:04:14.317 You know, many attribute this to normal getting fit, you know, just getting more 01:04:14.400 --> 01:04:17.997 efficient coming out of co and I think there's a lot of truth to that. But do 01:04:18.080 --> 01:04:20.717 you think part of this is due to AI? Do 01:04:20.800 --> 01:04:25.757 you think that AI is going to be a net job creator? And do you see this being a 01:04:25.840 --> 01:04:27.517 long-term positive for Microsoft 01:04:27.600 --> 01:04:30.077 productivity? Like it feels to me like 01:04:30.160 --> 01:04:32.957 the pie grows, but you can do all these 01:04:33.040 --> 01:04:35.117 things much more efficiently, which 01:04:35.200 --> 01:04:38.477 either means you your margins expand or 01:04:38.560 --> 01:04:42.557 it means you reinvest those margin dollars and you grow faster for longer. 01:04:42.640 --> 01:04:44.157 I call it the golden age of margin 01:04:44.240 --> 01:04:49.117 expansion. I'm a firm believer that the 01:04:49.200 --> 01:04:52.637 the productivity curve does uh and will 01:04:52.720 --> 01:04:55.677 bend in the sense that we will start 01:04:55.760 --> 01:04:59.357 seeing some of what is the work and the 01:04:59.440 --> 01:05:01.997 workflow in particular change, right? 01:05:02.080 --> 01:05:05.917 there's going to be more agency for you 01:05:06.000 --> 01:05:09.357 at a task level to get to job complete 01:05:09.440 --> 01:05:12.157 because of the power of these tools uh 01:05:12.240 --> 01:05:16.477 in your hand and that I think is going to be the case. So that's why I think we 01:05:16.560 --> 01:05:20.557 are even internally for example when you talked about even our allocation of 01:05:20.640 --> 01:05:22.717 tokens we want to make sure that 01:05:22.800 --> 01:05:24.797 everybody at Microsoft standard issue 01:05:24.880 --> 01:05:27.757 right all of them have Microsoft 365 to 01:05:27.840 --> 01:05:30.397 the tilt in the sort of most un uh 01:05:30.480 --> 01:05:35.037 limited way and have GitHub copilot so that they can really be more productive 01:05:35.120 --> 01:05:36.637 but here is the other interesting thing 01:05:36.720 --> 01:05:39.197 Brad we're learning is there is a new 01:05:39.280 --> 01:05:42.157 way to even learn right which is you 01:05:42.240 --> 01:05:46.637 know how to work with agents Right? So that's kind of like when the first when 01:05:46.720 --> 01:05:48.557 word, excel, powerpoint all showed up in 01:05:48.640 --> 01:05:54.077 office, you kind of we learned how to rethink let's say how we did a forecast, 01:05:54.160 --> 01:05:57.277 right? Right? I mean, think about it, right? In the 80s, the forecasts were 01:05:57.360 --> 01:06:01.677 inter office memos and faxes and what have you. And then suddenly somebody 01:06:01.760 --> 01:06:03.277 said, "Oh, here's an Excel spreadsheet. 01:06:03.360 --> 01:06:04.877 Let's put it an email. Send it around. 01:06:04.960 --> 01:06:07.357 People enter numbers and there was a forecast." 01:06:07.440 --> 01:06:10.237 >> Similarly, right now, any planning, any 01:06:10.320 --> 01:06:13.117 execution starts with AI. You research 01:06:13.200 --> 01:06:17.437 with AI. You think with AI, you share with your colleagues and what have you. 01:06:17.520 --> 01:06:19.437 So, there's a new artifact being created 01:06:19.520 --> 01:06:22.157 and a new workflow being created. And 01:06:22.240 --> 01:06:25.437 that is the rate of the pace of change 01:06:25.520 --> 01:06:29.757 of the business process that matches the 01:06:29.840 --> 01:06:32.077 capability of AI. That's where the 01:06:32.160 --> 01:06:37.357 productivity efficiencies come. And so organizations that can master that are 01:06:37.440 --> 01:06:40.717 going to be the biggest beneficiaries whether it's in our industry or quite 01:06:40.800 --> 01:06:44.877 frankly in the real world. >> And so is Microsoft benefiting from 01:06:44.960 --> 01:06:48.397 that? You know, so let's let's think about a couple years from now. Five 01:06:48.480 --> 01:06:52.237 years from now at the current growth rate will be sooner, but let's call it 01:06:52.320 --> 01:06:56.077 five years from now, your top line is twice as big as what it is today. 01:06:56.160 --> 01:06:59.677 Satcha, how many more employees will you have if you're if you're if you grow 01:06:59.760 --> 01:07:02.717 revenue by >> like one of the best things right now is 01:07:02.800 --> 01:07:06.557 these examples that I'm hit with every day from the employees of Microsoft. 01:07:06.640 --> 01:07:10.317 There was this person who leads our network operations, right? I mean if you 01:07:10.400 --> 01:07:12.637 think about the amount of uh fiber we 01:07:12.720 --> 01:07:15.597 have had to put uh for like this you 01:07:15.680 --> 01:07:17.677 know this 2 gawatt data center we just 01:07:17.760 --> 01:07:20.717 built out uh in fair water right and the 01:07:20.800 --> 01:07:25.517 amount of fiber there the AI and what have you it's just crazy right so 01:07:25.600 --> 01:07:29.677 >> and it turns out this is a real world asset there are I think 400 different 01:07:29.760 --> 01:07:33.277 fiber operators we're dealing with worldwide every time something happens 01:07:33.360 --> 01:07:37.277 we are literally going and dealing with all these DevOps pipelines the person 01:07:37.360 --> 01:07:41.357 who leads it she basically said to me you what I there's no way I'll ever get 01:07:41.440 --> 01:07:44.957 the headcount to go do all this. Not forget even if I even approve the 01:07:45.040 --> 01:07:49.117 budget. I can't hire all these folks. So she she did the next best thing. She 01:07:49.200 --> 01:07:53.037 just built herself a whole bunch of agents to automate the DevOps pipeline 01:07:53.120 --> 01:07:55.597 of how to deal with the maintenance. 01:07:55.680 --> 01:07:59.357 That is an example of you to your point a team 01:07:59.440 --> 01:08:02.397 >> with AI tools being able to get more 01:08:02.480 --> 01:08:06.957 productivity. So in if you are question I will say we will grow a headcount but 01:08:07.040 --> 01:08:11.277 the way I look at it is that headcount we grow will grow with a lot more 01:08:11.360 --> 01:08:14.717 leverage than the headcount we had pre AI 01:08:14.800 --> 01:08:18.317 >> and that's the adjustment I think structurally you're seeing first right 01:08:18.400 --> 01:08:21.277 which is one you called it getting fit I 01:08:21.360 --> 01:08:24.077 think of it as more getting to a place 01:08:24.160 --> 01:08:27.757 where everybody is really not learning 01:08:27.840 --> 01:08:30.317 how to rethink how they work and it's 01:08:30.400 --> 01:08:33.037 the how not even the what even If the 01:08:33.120 --> 01:08:35.437 what remains the constant, how you go 01:08:35.520 --> 01:08:37.757 about it has to be relearned. And it's 01:08:37.840 --> 01:08:40.077 the unlearning and learning process that 01:08:40.160 --> 01:08:44.877 I think will take the next year or so, then the headcount growth will come with 01:08:44.960 --> 01:08:49.357 max leverage. Yeah. No, it's a I think we're on the 01:08:49.440 --> 01:08:53.597 verge of incredible economic productivity growth. It does feel like 01:08:53.680 --> 01:08:57.837 when I talk to you or Michael Dell that most companies aren't even really in the 01:08:57.920 --> 01:09:02.317 first inning, maybe the the first batter in the first inning in reworking those 01:09:02.400 --> 01:09:06.397 workflows to get maximum leverage from these agents. But it sure feels like 01:09:06.480 --> 01:09:09.357 over the course of the next two to three years, that's where a lot of gains are 01:09:09.440 --> 01:09:13.757 going to start coming from. And again, I you know, I I certainly am an optimist. 01:09:13.840 --> 01:09:17.677 I think we're going to have net job gains from all of this. But I think for 01:09:17.760 --> 01:09:21.437 those companies, they'll just be able to grow their bottom line, their number of 01:09:21.520 --> 01:09:23.517 employees slower than their top line. 01:09:23.600 --> 01:09:27.357 That is the productivity gain to the company. Aggregate all that up. That's 01:09:27.440 --> 01:09:29.517 the productivity gain to the economy. 01:09:29.600 --> 01:09:33.277 And then we'll just take that consumer surplus and invest it in creating a lot 01:09:33.360 --> 01:09:35.197 of things that didn't exist before. 01:09:35.280 --> 01:09:39.037 >> 100%. 100%. Even in software development, right? One of the things I 01:09:39.120 --> 01:09:42.157 look at it is no one would say we're 01:09:42.240 --> 01:09:46.877 going to have a challenge in having, you know, more software engineers contribute 01:09:46.960 --> 01:09:51.117 to our sort of society because the reality is you look at the IT backlog in 01:09:51.200 --> 01:09:53.677 any organization. And so the question is 01:09:53.760 --> 01:09:56.477 all these software agents are hopefully 01:09:56.560 --> 01:09:59.357 going to help us go and take a whack at 01:09:59.440 --> 01:10:02.557 all of the IT backlog we have and think 01:10:02.640 --> 01:10:05.277 of that dream of evergreen software. 01:10:05.360 --> 01:10:08.637 That's going to be true. and then think about the demand for software. So I 01:10:08.720 --> 01:10:12.637 think that to your point it's the levels of abstraction at which knowledge work 01:10:12.720 --> 01:10:14.717 happens will change. We will adjust to 01:10:14.800 --> 01:10:19.757 that the work and the workflow that will then adjust itself even in terms of the 01:10:19.840 --> 01:10:22.237 demand for the products of this industry. 01:10:22.320 --> 01:10:26.237 >> I'm going to end on this which is really around the reindustrialization of 01:10:26.320 --> 01:10:28.477 America. I've said if you add up the $4 01:10:28.560 --> 01:10:31.837 trillion of capex that you and these and 01:10:31.920 --> 01:10:36.157 and so many of of the big large US tech companies are investing over the course 01:10:36.240 --> 01:10:39.837 of the next four or five years, it's about 10 times the size of the Manhattan 01:10:39.920 --> 01:10:42.397 project on an inflation adjusted or GD 01:10:42.480 --> 01:10:47.117 GDP adjusted basis. So it's a massive undertaking for America. The president 01:10:47.200 --> 01:10:51.997 has made it a real priority of his administration to recut the trade deals 01:10:52.080 --> 01:10:56.477 and it looks like we now have trillions of dollars. South Koreans committed $350 01:10:56.560 --> 01:10:58.717 billion dollars of investments uh just 01:10:58.800 --> 01:11:02.237 today into the United States. And when 01:11:02.320 --> 01:11:04.717 you think about, you know, what you see 01:11:04.800 --> 01:11:09.517 going on in power in the United States, both production, the grid, etc., what 01:11:09.600 --> 01:11:13.357 you see going on in terms of this re-industrialization, 01:11:13.440 --> 01:11:16.317 how how do you think this is all going? 01:11:16.400 --> 01:11:20.957 and uh you you know maybe just reflect on where we're landing the plane here 01:11:21.040 --> 01:11:24.717 and your level of optimism for the the the few years ahead. 01:11:24.800 --> 01:11:28.077 >> Yeah. No, I I I feel very very optimistic because in some sense, you 01:11:28.160 --> 01:11:30.157 know, Brad Smith was telling me about 01:11:30.240 --> 01:11:35.197 sort of the economy around a Wisconsin data center. It's fascinating. Most 01:11:35.280 --> 01:11:39.117 people think, oh, data center that is sort of like, yeah, uh it's going to be 01:11:39.200 --> 01:11:41.117 one big warehouse and there's, you know, 01:11:41.200 --> 01:11:43.837 fully automated. Uh a lot of it is true. 01:11:43.920 --> 01:11:46.477 uh but first of all what went into the 01:11:46.560 --> 01:11:49.517 construction uh of that data center and 01:11:49.600 --> 01:11:54.477 the local supply chain of the data center uh that is in some sense the 01:11:54.560 --> 01:11:59.117 reindustrialization of the United States as well uh even before you get to what 01:11:59.200 --> 01:12:01.357 is happening in Arizona with the TSMC 01:12:01.440 --> 01:12:03.997 plants or what was happening with Micron 01:12:04.080 --> 01:12:06.877 and their investments in memory or Intel 01:12:06.960 --> 01:12:08.797 and their fabs and what have you right 01:12:08.880 --> 01:12:11.597 there's a lot of stuff that we will want 01:12:11.680 --> 01:12:16.317 to start building doesn't mean we won't have trade deals that make sense for the 01:12:16.400 --> 01:12:20.397 United States with other countries. But to your point, the reindustrialization 01:12:20.480 --> 01:12:22.957 for the new economy and take making sure 01:12:23.040 --> 01:12:25.437 that all the skills and all that 01:12:25.520 --> 01:12:29.517 capacity from power on down, I think is 01:12:29.600 --> 01:12:33.197 sort of very important right for us. And in and the other thing that I also say, 01:12:33.280 --> 01:12:37.277 Brad, it's important and this is something that I've had a chance to talk 01:12:37.360 --> 01:12:41.597 to President Trump as well as uh Secretary Lutnik and others is it's 01:12:41.680 --> 01:12:45.517 important to recognize that we as hyperscalers of the United States are 01:12:45.600 --> 01:12:48.477 also investing around the world. So in 01:12:48.560 --> 01:12:51.037 other words, the United States is the 01:12:51.120 --> 01:12:54.477 biggest investor of compute factories or 01:12:54.560 --> 01:12:57.677 token factories uh around the world. But 01:12:57.760 --> 01:13:01.597 not only are we attracting foreign capital to invest in our country so that 01:13:01.680 --> 01:13:04.157 we can re-industrialize, we are helping 01:13:04.240 --> 01:13:06.397 whether it's in Europe or in Asia or 01:13:06.480 --> 01:13:09.197 elsewhere in Latin America and in Africa 01:13:09.280 --> 01:13:11.997 with our capital investments, bringing 01:13:12.080 --> 01:13:14.877 the best American tech uh to the world 01:13:14.960 --> 01:13:19.437 that they can then innovate on and trust. And so both of those I think are 01:13:19.520 --> 01:13:22.957 really bode well for the United States long term. 01:13:23.040 --> 01:13:25.517 >> I'm grateful for your leadership Sam is 01:13:25.600 --> 01:13:30.237 is is really helping lead the charge at open AI for America. I think this is a 01:13:30.320 --> 01:13:34.637 moment where I look ahead, you know, you can see 4% GDP growth on the horizon. 01:13:34.720 --> 01:13:37.437 We'll have our challenges. We'll have our ups and downs. These tend to be 01:13:37.520 --> 01:13:41.677 stairs, you know, stairs up rather than a line straight up and to the right. But 01:13:41.760 --> 01:13:46.317 I for one see a level of coordination going on between Washington and Silicon 01:13:46.400 --> 01:13:50.237 Valley between big tech and the re-industrialization of America that 01:13:50.320 --> 01:13:51.917 gives me cause for incredible hope. 01:13:52.000 --> 01:13:54.317 Watching what happened this week in Asia 01:13:54.400 --> 01:13:58.957 uh led by the president and his team and then watching what's happening here uh 01:13:59.040 --> 01:14:03.677 is is super exciting. So thanks for making the time. We're big fans. Thanks. 01:14:03.760 --> 01:14:05.277 Thanks Satcha. 01:14:05.360 --> 01:14:09.387 >> Thanks so much Brad. Thank you. 01:14:09.470 --> 01:14:11.490 [music] 01:14:16.640 --> 01:14:22.320 As a reminder to everybody, just our opinions, not investment advice.