All things AI w @altcap @sama & @satyanadella. A Halloween Special. 🎃🔥BG2 w/ Brad Gerstner

00:00:00.240 | I think this is really an amazing partnership through every phase.

00:00:04.000 | We had kind of no idea where it was all going to go when we started, as Satya said,

00:00:07.600 | but I don't think, I think this is one of the great tech partnerships ever.

00:00:13.840 | And without, certainly without Microsoft and particularly Satya's early conviction,

00:00:17.600 | we would not have been able to do that.

00:00:19.520 | What a week, what a week. Great to see you both. Sam, how's the baby?

00:00:36.560 | Baby is great. That's the best thing ever. Every cliche is true and it is the best thing ever.

00:00:42.640 | Hey Satya, with all your time-

00:00:45.120 | Smile on Sam's face whenever he talks about his baby. It's just so different.

00:00:50.160 | And compute, I guess, when he talks about compute and his baby.

00:00:55.280 | Well Satya, have you given any dad tips with all this time you guys have spent together?

00:01:01.680 | I said, just enjoy it. I mean, it's so awesome that, you know, we had our babies or our children

00:01:08.320 | so young and I wish I could redo it. So in some sense, it's just the most precious time. And as they

00:01:14.240 | grow, it's just so wonderful. I'm so glad Sam is-

00:01:17.200 | I'm happy to be doing it older, but I do think sometimes, man,

00:01:21.040 | I wish I had the energy of when I was like 25. That part's harder.

00:01:24.320 | No doubt about it. What's the average age at OpenAI, Sam? Any idea? It's young.

00:01:29.040 | It's not crazy young. Not like most Silicon Valley startups. I don't know, maybe low 30s average.

00:01:35.600 | Are babies trending positively or negatively?

00:01:39.120 | Babies trending positively.

00:01:40.240 | Oh, that's good. That's good.

00:01:41.440 | Yeah. Well, you guys, such a big week. You know, I was thinking about I started at NVIDIA's GTC,

00:01:47.280 | you know, just hit $5 trillion. Google, Meta, Microsoft, Satya, you had your earnings yesterday.

00:01:53.440 | You know, and we heard consistently not enough compute, not enough compute, not enough compute.

00:01:58.240 | We got rate cuts on Wednesday. The GDP is tracking near 4%. And then I was just saying to Sam,

00:02:03.840 | you know, the president's cut these massive deals in Malaysia, South Korea, Japan, sounds like with

00:02:08.880 | China. You know, deals that really incredibly provide the financial firepower to re-industrialize

00:02:14.640 | America. 80 billion for new nuclear fission, all the things that you guys need to build more compute.

00:02:20.720 | But certainly wasn't, what wasn't lost in all of this was you guys had a big announcement on Tuesday

00:02:26.160 | that clarified your partnership. Congrats on that. And I thought we'd just start there. I really want

00:02:30.880 | to just break down the deal in really simple, plain language to make sure I understand it and others.

00:02:37.440 | But, you know, we'll just start with your investment, Satya. You know, Microsoft started investing in 2019,

00:02:44.720 | is invested in the ballpark at $13, $14 billion into OpenAI. And for that, you get 27% of the business

00:02:52.640 | ownership in the business on a fully diluted basis. I think it was about a third. And you took some

00:02:57.280 | dilution over the course of last year with all the investment. So does that sound about right in terms

00:03:02.800 | of ownership? Yeah, it does. But I would say before even our stake in it, Brad, I think what's pretty

00:03:09.360 | unique about OpenAI is the fact that as part of OpenAI's process of restructuring, one of the largest

00:03:18.000 | nonprofit gets created. I mean, let's not forget that, you know, in some sense, I say at Microsoft,

00:03:23.520 | like, you know, we are very proud of the fact that we were associated with the two of the largest

00:03:27.520 | nonprofits, the Gates Foundation and now the OpenAI Foundation. So that's, I think, the big news. We

00:03:33.680 | obviously are thrilled. It's not what we thought. And as I said to somebody, it's not like when we

00:03:39.280 | first invested our billion dollars that, oh, this is going to be the hundred bagger that I'm going to

00:03:43.440 | be talking about to VCs about. But here we are. But we are very thrilled to be an investor and an early

00:03:50.160 | backer. And it's a great, and it's really a testament to what Sam and team have done, quite frankly. I mean,

00:03:56.560 | they obviously had the vision early about what this technology could do. And they ran with it and just

00:04:02.480 | executed, you know, in a masterful way. I think this has really been an amazing partnership through

00:04:08.160 | every phase. We had kind of no idea where it was all going to go when we started, as Satya said.

00:04:12.560 | But I don't think, I think this is one of the great tech partnerships ever. And without,

00:04:19.440 | certainly without Microsoft, and particularly Satya's early conviction, we would not have been

00:04:23.920 | able to do this. I don't think there were a lot of other people that would have been willing to take

00:04:28.240 | that kind of a bet given what the world looked like at the time. We didn't know exactly how the tech was

00:04:34.400 | going to go. Well, not exactly. We didn't know at all how the tech was going to go. We just had a lot of

00:04:38.160 | conviction in this one idea of pushing on deep learning and trusting that if we could do that,

00:04:42.800 | we'd figure out ways to make wonderful products and create a lot of value. And also, as Satya said,

00:04:47.360 | create what we believe will be the largest nonprofit ever. And I think it's going to do amazingly great

00:04:53.520 | things. It was, I really like the structure because it lets the nonprofit grow in value while the PBC is

00:05:01.040 | able to get the capital that it needs to keep scaling. I don't think the nonprofit would be able to be

00:05:05.440 | this valuable if we didn't come up with the structure and if we didn't have partners around the table that

00:05:09.440 | we're excited for it to work this way. But, you know, I think it's been six, more than six years since

00:05:14.640 | we first started this partnership. And a pretty crazy amount of achievement for six years. And I think much,

00:05:20.720 | much, much more to come. I hope that Satya makes a trillion dollars on the investment, not a hundred

00:05:24.240 | billion, you know, whatever it is. Well, as part of the restructuring,

00:05:27.520 | you guys talked about it. You have this nonprofit on top and a public benefit corp below. It's pretty

00:05:32.720 | insane. The nonprofit is already capitalized with $130 billion, $130 billion of open AI stock. It's one of

00:05:42.000 | the largest in the world out of the gates. It could end up being much, much larger. The California attorney

00:05:46.960 | general said they're not going to object to it. You already have this $130 billion dedicated to making

00:05:52.720 | sure that AGI benefits all of humanity. You announced that you're going to direct the first $25 billion

00:05:58.400 | to health and AI security and resilience, Sam. First, let me just say, you know, as somebody who

00:06:04.480 | participates in the ecosystem, kudos to you both. It's incredible, this contribution to the future of AI.

00:06:10.240 | But Sam, talk to us a bit about the importance of the choice around health and resilience, and then help us

00:06:18.160 | understand how do we make sure that you get maximal benefit without it getting weighted down, as we've

00:06:24.240 | seen with so many nonprofits with its own political biases?

00:06:27.760 | Yeah, first of all, the best way to create a bunch of value for the world is hopefully what we've already been

00:06:35.680 | doing, which is to make these amazing tools and just let people use them. And I think capitalism is

00:06:40.320 | great. I think companies are great. I think people are doing amazing work getting advanced AI into the

00:06:45.280 | hands of a lot of people and companies that are doing incredible things. There are some areas where the,

00:06:51.200 | I think, market forces don't quite work for what's in the best interest of people, and you do need to do

00:06:58.000 | things in a different way. There are also some new things with this technology that just haven't existed

00:07:03.280 | before, like the potential to use AI to do science at a rapid clip, like really truly automated discovery.

00:07:09.040 | And when we thought about the areas we wanted to first focus on, clearly, if we can cure a lot of

00:07:14.720 | disease and make the data and information for that broadly available, that would be a wonderful thing

00:07:20.160 | to do for the world. And then on this point of AI resilience, I do think some things may get

00:07:25.600 | a little strange and they won't all be addressed by companies doing their thing. So as the world has to

00:07:31.040 | navigate through this transition, if we can fund some work to help with that, and that could be,

00:07:36.640 | you know, cyber defense, that could be AI safety research, that could be economic studies, all of

00:07:40.480 | these things, helping society get through this transition smoothly. We're very confident about how

00:07:46.080 | great it can be on the other side. But, you know, I'm sure there will be some choppiness along the way.

00:07:49.920 | Let's keep busting through the deal. So models and exclusivity, Sam, OpenAI can distribute its models,

00:07:58.400 | its leading models on Azure, but I don't think you can distribute them on any other leading of the big

00:08:03.600 | clouds for seven years until 2032. But that would end earlier if AGI is verified, we can come back to that.

00:08:10.800 | But you can distribute your open source models, Sora, agents, codecs, wearables, everything else on other

00:08:16.320 | platforms. So Sam, I assume this means no chat GPT or GPT-6 on Amazon or Google.

00:08:23.440 | No, so we have a cat. First of all, we want to do lots of things together to help, you know,

00:08:28.080 | create value for Microsoft. We want them to do lots of things to create value for us. And there are many,

00:08:32.640 | many things that will happen in that category. We are keeping what Satya termed once, and I think it's a great

00:08:37.360 | phrase of stateless APIs on Azure exclusively through 2030 and everything else we're going to, you know,

00:08:42.880 | distribute elsewhere. And that's obviously in Microsoft's interest too. So we'll put lots of

00:08:47.120 | products, lots of places, and then this thing we'll do on Azure and people can get it there or via us.

00:08:51.920 | And I think that's great. And then the rev share, there's still a rev share that gets paid by OpenAI

00:08:56.080 | to Microsoft on all your revenues that also runs until 2032, more until AGI is verified. So let's just

00:09:03.200 | assume for the sake of argument, I know this is pedestrian, but it's important that the

00:09:07.200 | rev share is 15%. So that would mean if you had $20 billion in revenue that you're paying $3 billion

00:09:12.720 | to Microsoft and that counts as revenue to Azure. Satya, does that sound about right?

00:09:17.920 | Yeah, we have a rev share. And I think as you characterized it, it is either going to AGI

00:09:22.960 | or till the end of the term. And I actually don't know exactly where we count it, quite honestly,

00:09:28.000 | whether it goes into Azure or somewhere else. That's a good question. It's a good question for Amy.

00:09:31.920 | Given that both exclusivity and the rev share end early in the case AGI is verified,

00:09:38.240 | it seems to make AGI a pretty big deal. And as I understand it, you know, if OpenAI claimed AGI,

00:09:44.560 | it sounds like it goes to an expert panel and you guys basically select a jury who's got to make a

00:09:49.600 | relatively quick decision whether or not AGI has been reached. Satya, you said on yesterday's earning call

00:09:54.880 | that nobody's even close to getting to AGI and you don't expect it to happen anytime soon. You talked about

00:10:00.000 | this spiky and jagged intelligence. Sam, I've heard you perhaps sound a little bit more bullish on,

00:10:05.760 | you know, when we might get to AGI. So I guess the question is to you both, do you worry that over the

00:10:11.200 | next two or three years, we're going to end up having to call in the jury to effectively make a call on

00:10:17.200 | whether or not we've hit AGI? I realize you've got to try to make some drama between us here.

00:10:21.280 | I think putting a process in place for this is a good thing to do. I expect that the technology

00:10:32.400 | will take several surprising twists and turns and we will continue to be good partners to each other and

00:10:36.880 | figure out what makes sense. That's well said, I think. And that's one of the reasons why I think

00:10:42.000 | this process we put in place is a good one. And at the end of the day, I'm a big believer in the fact that

00:10:47.760 | intelligence capability-wise is going to continue to improve. And our real goal, quite frankly,

00:10:53.760 | is that, which is how do you put that in the hands of people and organizations so that they can get the

00:10:58.400 | maximum benefits? And that was the original mission of OpenAI that attracted me to OpenAI and Sam and

00:11:04.880 | team. And that's kind of what we plan to continue on. Brad, to say the obvious, if we had super

00:11:09.920 | intelligence tomorrow, we would still want Microsoft's help getting this product out into people's hands.

00:11:14.960 | And we want them, like, yeah. Of course, of course. Yeah, no, again, I'm asking the questions I know that

00:11:21.680 | are on people's minds and that makes a ton of sense to me. Obviously, Microsoft is one of the largest

00:11:27.520 | distribution platforms in the world. You guys have been great partners for a long time, but I think it

00:11:31.440 | dispels some of the myths that are out there. But let's shift gears a little bit. You know, obviously,

00:11:35.520 | OpenAI is one of the fastest growing companies in history. Satya, you said on the pod a year ago,

00:11:41.120 | this pod, that every new phase shift creates a new Google and the Google of this phase shift is already

00:11:47.760 | known and it's OpenAI. And none of this would have been possible had you guys not made these huge bets.

00:11:53.440 | With all that said, you know, OpenAI's revenues are still a reported $13 billion in 2025. And Sam,

00:12:00.880 | on your live stream this week, you talked about this massive commitment to compute, right? 1.4 trillion

00:12:07.920 | over the next four or five years with, you know, big commitments, 500 million to Nvidia, 300 million to

00:12:14.560 | AMD and Oracle, 250 billion to Azure. So I think the single biggest question I've heard all week and,

00:12:21.680 | and hanging over the market is how, you know, how can the company with 13 billion in revenues make 1.4

00:12:29.120 | trillion of spend commitments, you know, and, and, and you've heard the criticism.

00:12:33.920 | Sam, we're doing well more revenue than that. Second of all, Brad, if you want to sell your shares,

00:12:38.560 | I'll find you a buyer. I just, enough, like, you know, people are, I think there's a lot of people

00:12:46.080 | who would love to buy OpenAI shares. I don't, I don't think you would. Including myself. Including myself.

00:12:51.200 | People who talk with a lot of like breathless concern about our compute stuff or whatever,

00:12:56.320 | that would be thrilled to buy shares. So I think we could sell, you know, your shares or anybody else's to

00:13:00.960 | some of the people who are making the most noise on Twitter, whatever about this very quickly. We do plan for

00:13:05.760 | revenue to grow steeply. Revenue is growing steeply. We are taking a forward bet that it's going to

00:13:10.400 | continue to grow. And that not only will Chow Chow Chow Tee keep growing, but we will be able to become

00:13:16.880 | one of the important AI clouds. That our consumer device business will be a significant and important

00:13:23.360 | thing. That AI that can automate science will create huge value. So, you know, there are not many times that I

00:13:31.920 | want to be a public company, but one of the rare times it's appealing is when those people that are

00:13:36.160 | writing these ridiculous open AI is about to go out of business and, you know, whatever. I would love

00:13:40.400 | to tell them they could just short the stock and I would love to see them get burned on that. But,

00:13:45.120 | you know, I, we carefully plan, we understand where the technology, where the capability is going to

00:13:51.920 | grow, go and, and how the products we can build around that and the revenue we can generate.

00:13:57.280 | We might screw it up. Like this is the bet that we're making and we're taking a risk along with

00:14:01.840 | that. A certain risk is if we don't have the compute, we will not be able to generate the revenue or make

00:14:06.800 | the models at these, at this kind of scale. Exactly. And let me just say one thing, Brad,

00:14:12.400 | as both a partner and an investor, there is not been a single business plan that I've seen from OpenAI

00:14:21.600 | that they've put in and not beaten it. So in some sense, this is the one place where, you know, in

00:14:28.320 | terms of their growth and just even the business, it's been unbelievable execution, quite frankly. I mean,

00:14:34.240 | obviously, OpenAI, everyone talks about all the success and the usage and what have you. But even, I'd say all

00:14:40.240 | up, the business execution has been just pretty unbelievable. I heard Greg Brockman say on CNBC a couple of weeks ago,

00:14:48.000 | right, if we could 10x our compute, we might not have 10x more revenue, but we'd certainly have a lot

00:14:54.080 | more revenue. Simply because of lack of compute power, things like, yeah, it's just, it's really

00:15:00.400 | wild when I just look at how much we are held back. And in many ways we have, you know, we've scaled our

00:15:05.840 | compute probably 10x over the past year. But if we had 10x more compute, I don't know if we'd have 10x more

00:15:10.800 | revenue, but I don't think it'd be that far. And we heard this from you as well last night, Satya,

00:15:16.640 | that you were compute constrained and growth would have been higher even if you add more compute.

00:15:21.120 | So help us contextualize Sam, maybe like how compute constrained do you feel today? And do you when

00:15:27.600 | you look at the build out over the course of the next two to three years, do you think you'll ever get to

00:15:31.760 | the point where you're not compute constrained? We talked about this question of is there ever enough

00:15:37.200 | compute a lot? I think the answer is the only, the best way to think about this is like energy or

00:15:46.720 | something. You can talk about demand for energy at a certain price point, but you can't talk about demand

00:15:51.600 | for energy without talking about at different, you know, different demand at different price levels.

00:15:58.800 | If the price of compute per like unit of intelligence or whatever, however you want to think about it,

00:16:04.800 | fell by a factor of a hundred tomorrow, you would see usage go up by much more than a hundred.

00:16:10.080 | And there'd be a lot of things that people would love to do with that compute that just make no economic

00:16:13.760 | sense at the current cost, but there would be new kind of demand. So I think the, now on the other

00:16:20.720 | hand, as the models get even smarter and you can use these models to cure cancer or discover novel

00:16:25.680 | physics or drive a bunch of humanoid robots to construct a space station or whatever crazy thing

00:16:29.840 | you want, then maybe there's huge willingness to pay a much higher rate cost per unit of intelligence

00:16:37.040 | for a much higher level of intelligence that we don't know yet, but I would bet there will be.

00:16:41.200 | So I, I think when you talk about capacity, it's, it's like a, you know, cost per unit and, you know,

00:16:48.080 | capability per unit. And you have to kind of, without those curves, it's sort of a made up note. It's

00:16:53.120 | not a super well-specified problem. Yeah. I mean, I think the one thing that, you know, Sam, you've

00:16:59.040 | talked about, which I think is the right ways to think about is that if intelligence is whatever log of

00:17:03.360 | compute, then you try and really make sure you keep getting efficient. And so that means the tokens per dollar

00:17:09.840 | per watt, uh, and the economic value that the society gets out of it is what we should maximize and reduce

00:17:16.640 | the costs. And so that's where, if you sort of where like the Jevons paradox point is that, which is

00:17:22.080 | you keep reducing it, commoditizing in some sense intelligence, uh, so that it becomes the real driver

00:17:28.880 | of GDP growth all around. Unfortunately, it's something closer to, uh, log of intelligence.

00:17:34.080 | It was log of compute, but we may figure out better scaling laws and we may figure out how to do this.

00:17:37.600 | Yeah. We heard from both Microsoft and Google yesterday, both said their cloud businesses

00:17:42.400 | would have been growing faster if they have more GPUs. You know, I asked Jensen on this pod,

00:17:47.200 | if there was any chance over the course of the next five years, we would have a compute glut. And he said,

00:17:52.880 | it's virtually non-existent chance in the next two to three years. And I assume you guys would both

00:17:58.960 | agree with Jensen that while we can't see out five, six, seven years, certainly over the course of the

00:18:04.320 | next two to three years for the, for the reasons we just discussed, that it's almost a non-existent

00:18:09.120 | chance that you have excess compute. Well, I mean, I think the, the cycles of demand and supply

00:18:16.880 | in this particular case, you can't really predict, right? I mean, even the point is what's the secular trend?

00:18:23.360 | The secular trend is what Sam said, which is at the end of the day, because quite frankly, the biggest issue

00:18:28.880 | we are now having is not a compute glut, but it's a power and it's sort of the ability to get the bills done

00:18:34.960 | fast enough close to power. So if you can't do that, you may actually have a bunch of chips sitting

00:18:41.040 | in inventory that I can't plug in. And in fact, that is my problem today, right? It's not a supply

00:18:46.160 | issue of chips. It's actually the fact that I don't have warm shells to plug into. And so

00:18:51.920 | how some supply chain constraints emerge, tough to predict because the demand is just going, you know,

00:18:59.040 | is tough to predict, right? I mean, I wouldn't, it's not like Sam and I would want to be sitting

00:19:03.520 | here saying, oh my God, we're less short on compute. It's because we just were not that good at being able

00:19:08.720 | to project out what the demand would really look like. So I think that that's, and by the way, the worldwide

00:19:14.480 | side, right? It's one thing to sort of talk about one segment in one country, but it's about, you know,

00:19:19.520 | really getting it out to everywhere in the world. And so there will be constraints and how we work

00:19:23.920 | through them is going to be the most important thing. It won't be a linear path for sure.

00:19:27.760 | There will come a glut for sure. And whether that's like in two to three years or five to six, Satya,

00:19:33.440 | and I can't tell you, but like, it's going to happen at some point, probably several points along

00:19:37.920 | the way. Like this is, there is something deep about human psychology here and bubbles. And also,

00:19:44.880 | as Satya said, like there's, it's such a complex supply chain, weird stuff gets built. The technological

00:19:51.120 | landscape shifts in big ways. So, you know, if a very cheap form of energy comes online soon at mass

00:19:58.080 | scale, and a lot of people are going to be extremely burned with existing contracts they've signed,

00:20:01.760 | if, if we can continue this unbelievable reduction in cost per unit of intelligence, let's say it's been

00:20:09.280 | averaging like 40 X for a given level per year, you know, that's like a very scary exponent

00:20:17.040 | I think that's really well said. And you have to hold those two simultaneous truths. We had that

00:20:45.360 | happen in 2000, 2001, and yet the internet became much bigger and produced much greater outcomes for

00:20:51.440 | society than anybody estimated in that period of time. Yeah, but I think that the one thing that Sam

00:20:56.320 | said is not talked about enough, which is the, for example, the optimizations that OpenAI has done on the

00:21:02.560 | inference stack for a given GPU. I mean, it's kind of like, it's, you know, we talk about the Moore's law

00:21:08.320 | improvement on one end, but the software improvements are much more exponential than that.

00:21:13.520 | I mean, someday we will make an incredible consumer device that can run a GPT-5 or GPT-6

00:21:20.800 | capable model completely locally at a low power draw. And this is like, so hard to wrap my head around.

00:21:27.600 | That will be incredible. And, you know, that's the type of thing I think that scares some of the people

00:21:32.240 | who are building, obviously, these large centralized compute stacks. And Satya, you've talked a lot about the

00:21:38.320 | distribution, both to the edge, as well as having inference capability distributed around the world.

00:21:43.280 | Yeah, I mean, at least I've thought about it is more about really building a fungible fleet. I mean, when I look at sort of in the

00:21:51.040 | cloud infrastructure business, one of the key things you have to do is have two things. One is an efficient, like in this context, in a very efficient token

00:21:59.360 | factory, and then high utilization. That's it. There are two simple things that you need to achieve.

00:22:05.440 | And in order to have a high utilization, you have to have multiple workloads that can be scheduled,

00:22:09.840 | even on the training. I mean, if you look at the AI pipelines, there's pre-training, there's mid-training,

00:22:13.920 | there's post-training, there's RL, you want to be able to do all of those things. So thinking about

00:22:18.160 | fungibility of the fleet is everything for a cloud provider.

00:22:21.920 | Okay, so Sam, you referenced, you know, and Reuters was reporting yesterday that OpenAI may be planning

00:22:28.240 | to go public late 26 or in 27. No, no, no, we don't have anything that specific. I'm a realist. I assume

00:22:34.320 | it will happen someday. But that was, I don't know why people write these reports. We don't have like a date in mind.

00:22:40.640 | Great to know. I don't have a decision to do this or anything like that. I just assume it's where

00:22:44.240 | things will eventually go. But it does seem to me, if you guys were, you know, are doing in excess of

00:22:50.640 | $100 billion of revenue in 28 or 29, that you at least would be in position. What? How about 27?

00:22:58.160 | Yeah, 27, even better. You are in position to do an IPO and the rumored trillion dollars, again, just to

00:23:06.400 | contextualize for listeners. If you guys went public at 10 times $100 billion in revenue, right, which

00:23:13.520 | would be, I think, a lower multiple than Facebook went public at, a lower multiple than a lot of other

00:23:19.440 | big consumer companies went public at, that would put you at a trillion dollars. If you floated 10 to 20%

00:23:26.320 | of the company, that raises $100 to $200 billion, which seems like that would be a good path to fund

00:23:33.440 | a lot of the growth and a lot of the stuff that we just talked about. So you're, you're not opposed

00:23:38.800 | to it. You're not, but you guys are making the company with revenue growth, which is what I would

00:23:43.920 | like us to do. But no doubt about it. Well, I've also said, I think that this is such an important

00:23:50.080 | company. And, you know, there are so many people, including my kids who like to trade their little

00:23:56.080 | accounts and they use ChatGPT. And I think having retail investors have an opportunity to buy one of

00:24:02.240 | the most important and largest companies. That would be nice. Honestly, that is probably

00:24:05.440 | the single most appealing thing about it to me. That would be really nice.

00:24:10.560 | So one of the things I've talked to you both about shifting gears again is part of the big,

00:24:16.640 | beautiful bill, you know, Senator Cruz had included federal preemption so that we wouldn't have this

00:24:24.000 | state patchwork, 50 different laws that mires the industry down and kind of needless compliance and

00:24:30.160 | regulation. Unfortunately, got killed at the last second by Senator Blackburn, because frankly, I think

00:24:36.320 | AI is pretty poorly understood in Washington. And there's a lot of doomerism, I think, that has gained

00:24:41.520 | traction in Washington. So now we have state laws like the Colorado AI Act that goes into full effect

00:24:47.600 | in February, I believe, that creates this whole new class of litigants. Anybody who claims any unfair

00:24:52.720 | impact from an algorithmic discrimination in a chatbot. So somebody could claim harm for countless reasons.

00:24:59.840 | Sam, how worried are you that, you know, having this state patchwork of AI, you know,

00:25:06.000 | poses real challenges to, you know, our ability to continue to accelerate and compete around the world?

00:25:11.360 | I don't know how we're supposed to comply with that California, sorry, Colorado law, I would love

00:25:16.400 | them to tell us, you know, we'd like to be able to do it. But that's just from what I've read of that,

00:25:22.560 | that's like, I literally don't know what we're supposed to do. I'm very worried about a 50 state

00:25:27.280 | patchwork. I think it's a big mistake. I think it's, there's a reason we don't usually do that for

00:25:31.840 | these sorts of things. I think it'd be bad.

00:25:33.360 | Yeah, I mean, I think the fundamental problem of, you know, this patchwork approach is,

00:25:38.880 | quite frankly, I mean, between OpenAI and Microsoft, we'll figure out a way to navigate this, right? I

00:25:44.160 | mean, we can figure this out. The problem is anyone starting a startup and trying to kind of,

00:25:50.720 | this is sort of, it just goes to the exact opposite, or I think what the intent here is,

00:25:55.440 | which obviously safety is very important, making sure that the fundamental, you know, concerns people

00:26:02.240 | have are addressed, but there's a way to do that at the federal level. So, I think if we don't do this,

00:26:07.680 | again, you know, EU will do it, and then that'll cause its own issues. So, I think if US leads,

00:26:13.680 | it's better as, you know, as one regulatory framework. For sure.

00:26:18.400 | And to be clear, it's not that one is advocating for no regulation. It's simply saying, let's have,

00:26:24.400 | you know, agreed upon regulation at the federal level, as opposed to 50 competing state laws,

00:26:29.840 | which certainly firebombs the AI startup industry. And I think it makes it super challenging, even for

00:26:36.800 | companies like yours, who can afford to defend all these cases.

00:26:39.440 | Yeah. And I would just say, quite frankly, my hope is that this time around, even across EU and the

00:26:45.200 | United States, like that'll be the dream, right? Quite frankly, for any European startup.

00:26:49.360 | I don't think that's going to happen, Satu.

00:26:51.120 | What is that?

00:26:51.920 | That would be great. I don't, I wouldn't hold your breath for that one. That would be great.

00:26:54.720 | No, but I really think that if you think about it, right, if you sort of, if anyone in Europe is

00:27:00.480 | thinking about their, you know, how can they participate in this AI economy with their companies,

00:27:07.440 | this should be the main concern there as well. So therefore that's, I hope there is some enlightened

00:27:13.600 | approach to it, but I agree with you that, you know, today I wouldn't bet on that.

00:27:16.960 | I do think that with SACS, as the AIs are, you at least have a president that I think might fight for

00:27:24.320 | that in terms of coordination of AI policy, using trade as a lever to make sure that, you know,

00:27:30.880 | we don't end up with overly restricted European policy, but we shall see. I think first things

00:27:35.040 | first, federal preemption in the United States is pretty critical. You know, we've been down in the

00:27:38.960 | weeds a little bit here, Sam. So I want to telescope out a little bit. You know, I've heard people on your

00:27:45.600 | team talk about all the great things coming up. And as you start thinking about much more unlimited

00:27:52.480 | compute chat GPT six and beyond robotics, physical devices, scientific research, as you, as you look

00:28:02.080 | forward to 2026, what do you think surprises us the most? What, what, what are you most excited about

00:28:08.000 | in terms of what's on the drawing board?

00:28:09.600 | You, I mean, you just hit on a lot of the key points there. I think

00:28:13.520 | Codex has been a very cool thing to watch this year. And as these go from multi-hour tasks to multi-day

00:28:20.720 | tasks, which I expect to happen next year, what people be able to do to create

00:28:24.720 | software at an unprecedented rate and really in fundamentally new ways. I'm very excited for that.

00:28:31.280 | I think we'll see that in other industries too. I have like a bias towards coding. I understand that

00:28:35.040 | one better, but I think we'll see that really start to transform what people are capable of. I, I hope for

00:28:42.000 | very small scientific discoveries in 2026, but if we can get those very small ones, we'll get bigger ones in

00:28:46.400 | future years. That's a really crazy thing to say is that like AI is going to make a novel scientific

00:28:50.720 | discovery in 2026, even a very small one. This is like, this is a wildly important thing to be talking

00:28:57.600 | about. So I'm excited for that. Certainly robotics and computer and new kind of computers in future

00:29:03.280 | years. That'll be, that'll be very important. But yeah, my personal bias is if we can really get AI to do

00:29:11.680 | science here, that is, I mean, that is super intelligence in some sense. Like if this is expanding

00:29:17.520 | the total sum of human knowledge, that is a crazy big deal.

00:29:20.000 | Yeah. I mean, I think one of the things to use your codex example, I think the combination of the model

00:29:27.040 | capability, I mean, if you think about the magical moment that happened with ChatGPT was the UI that

00:29:33.200 | met intelligence that just took off, right? It's just, you know, unbelievable right form fact. And some

00:29:39.680 | of it was also the instruction following piece of model capability was ready for chat. I think that that's

00:29:46.320 | what the codex and the, you know, these coding agents are about to help us, which is what's that,

00:29:52.080 | you know, coding agent goes off for a long period of time, comes back, and then I'm then dropped into

00:29:59.040 | what I should steer. Like one of the metaphors I think we're all sort of working towards is I do this

00:30:04.560 | macro delegation and micro steering. What is that UI meets this new intelligence capability? And you can see

00:30:13.520 | the beginnings of that with codex, right? The way at least I use it inside a GitHub Copilot is like,

00:30:18.960 | you know, it's now, it's just a, it's just a different way than the chat interface. And I think

00:30:24.160 | that, that I think would be a new way for the human computer interface, quite frankly, it's probably

00:30:29.280 | bigger than, that might be the departure. That's one reason I'm very excited that we're doing new form

00:30:35.840 | factors of computing devices, because computers were not built for that kind of workflow very well.

00:30:40.240 | Certainly, a UI like ChatGPT is wrong for it. But this idea that you can have a device that is sort of

00:30:45.520 | always with you, but able to go off and do things and get micro steer from you when it needs and have

00:30:51.200 | like really good contextual awareness of your whole life and flow. And I think that'd be cool.

00:30:55.680 | And what neither of you have talked about is the consumer use case, I think a lot about,

00:31:00.160 | you know, again, we go onto this device, and we have to hunt and pack through 100 different

00:31:04.080 | applications and fill out little web forms, things that really haven't changed in 20 years. But to

00:31:08.880 | just have, you know, a personal assistant that we take for granted, perhaps that we actually have a

00:31:13.360 | personal assistant, but to give a personal assistant for virtually free to billions of people around the

00:31:19.120 | world to improve their lives, whether it's, you know, ordering diapers for their kid, or whether it's,

00:31:24.720 | you know, booking their hotel or making changes in their calendar, I think sometimes it's the

00:31:30.160 | pedestrian that's the most impactful. And as we move from answers to memory and actions, and then the

00:31:36.880 | ability to interface with that through an earbud or some other device that doesn't require me to

00:31:41.600 | constantly be staring at this rectangular piece of glass, I think it's pretty extraordinary.

00:31:46.320 | I think that that's what Sam was teasing.

00:31:48.560 | Yeah, yeah.

00:31:50.480 | All right. I got to drop off, unfortunately.

00:31:52.720 | Sam, it was great to see you. Thanks for joining us. Congrats again on this big step forward and we'll

00:31:57.680 | talk soon.

00:31:58.080 | Thanks for letting me crash.

00:31:58.800 | See you, Sam.

00:32:00.080 | Take care.

00:32:00.640 | See ya.

00:32:01.040 | As Sam well knows, we're certainly a buyer, not a seller. But sometimes, you know, I think it's important

00:32:09.520 | because the world, you know, we're pretty small. We spend all day long thinking about this stuff,

00:32:15.680 | right? And so conviction, it comes from the 10,000 hours we've spent thinking about it. But the reality

00:32:23.040 | is we have to bring along the rest of the world and the rest of the world doesn't spend 10,000 hours

00:32:28.080 | thinking about this. And frankly, they look at some things that appear overly ambitious, right? And get

00:32:34.160 | worried about whether or not we can pull those things off. So you took this idea to the board in 2019

00:32:40.720 | to invest a billion dollars into OpenAI. Was it a no-brainer in the boardroom?

00:32:45.040 | You know, did you have to expend any political capital to get it done? Dish for me a little bit,

00:32:51.280 | like, what that moment was like. Because I think it was such a pivotal moment, not just for Microsoft,

00:32:56.640 | not just for the country, but I really do think for the world.

00:32:59.040 | It's interesting when you look back. The journey, when I look at it, it's been, you know, we were

00:33:05.200 | involved even in 2016 when initially OpenAI started. In fact, Azure was even the first sponsor, I think.

00:33:13.520 | And then they were doing a lot more reinforcement learning at that time. I remember the Dota 2

00:33:18.080 | competition, I think, happened on Azure. And then they moved on to other things. And, you know, I was

00:33:23.520 | interested in RL, but quite frankly, you know, it speaks a little bit to your 10,000 hours or the

00:33:28.880 | prepared mind. Microsoft, since 1995, was obsessed. I mean, Bill's obsession for the company was natural

00:33:36.160 | language, natural language. I mean, after all, we had a coding company, information work company.

00:33:40.640 | So it's when Sam in 2019 started talking about text and natural language and transformers and scaling

00:33:47.840 | laws. That's when I said, wow, like, this is an interesting, I mean, he, you know, this was a team

00:33:54.000 | that was going in the direction or the direction of travel was now clear. It had a lot more overlap

00:34:00.160 | with our interest. So in that sense, it was a no brainer. Obviously, you go to the board and say,

00:34:06.160 | hey, I have an idea of taking a billion dollars and giving it to this crazy structure, which we

00:34:11.920 | don't even kind of understand. What is it? It's a nonprofit, blah, blah, blah. And saying, go for it.

00:34:17.840 | There was a debate. Bill was kind of rightfully so skeptical because, and then he became like,

00:34:24.800 | once he saw the GPT-4 demo, like that was like the thing that Bill's talked about publicly, where

00:34:29.520 | when he saw it, he said, it's the best demo he saw after, you know, what Charles Simone showed him at

00:34:35.200 | Xerox PARC. And, but, you know, quite honestly, none of us could. So the moment for me was that,

00:34:42.160 | you know, let's go give it a shot. Then seeing the early codecs inside of Copilot, inside of

00:34:50.080 | GitHub Copilot and seeing just the code completions and seeing it work, that's when I would say we,

00:34:56.320 | I felt like I can go from one to 10 because that was the big call, quite frankly. One was controversial,

00:35:03.440 | but the one to 10 was what really made this entire era possible. And then obviously,

00:35:09.920 | the great execution by the team and the productization on their part, our part. I mean,

00:35:15.280 | if I think about it, right, the collective monetization reach of GitHub Copilot, ChatGPT,

00:35:22.160 | Microsoft 365 Copilot and Copilot, you add those four things, that is it, right? That's the biggest sort of AI

00:35:28.160 | set of products out there on the planet. And that's, you know, what obviously has let us sustain all of this.

00:35:34.640 | And I think not many people know that your CTO, Kevin Scott, you know, an ex Googler lives down

00:35:40.320 | here in Silicon Valley. And to contextualize it, right, Microsoft had missed out on search,

00:35:44.880 | had missed out on mobile, you become CEO, almost had missed out on the cloud, right? You've described it,

00:35:52.480 | caught the last train out of town to capture the cloud. And I think you were pretty determined to have

00:35:59.200 | eyes and ears down here. So you didn't miss the next big thing. So I assume that Kevin played a good

00:36:04.720 | role for you as well. Absolutely. I find deep seek and open AI. Yeah, I mean, I mean, if it's in fact,

00:36:11.200 | I would say Kevin's conviction, and Kevin was also skeptical, like that was the thing. I always watch

00:36:18.240 | for people who are skeptical, who change their opinion, because to me, that's a signal. So I'm always

00:36:25.040 | looking for someone who's a non-believer in something, and then suddenly changes. And then

00:36:29.840 | they get excited about it, that I have all the time for that, because I'm then curious, why? And so,

00:36:35.680 | Kevin started with thinking that all of us were kind of skeptical, right? No, I mean, in some sense,

00:36:40.080 | it defies the, you know, we're all having gone to school and said, God, you know, there must be an

00:36:45.680 | algorithm to crack this versus just scaling laws and throw compute. But quite frankly, Kevin's conviction

00:36:52.720 | that this is worth going after is one of the big things that drove this.

00:36:57.040 | Well, we talk about, you know, that investment that it's now worth 130 billion, I suppose, could be

00:37:03.200 | worth a trillion someday, as Sam says. But it really, in many ways, understates the value of the partnership,

00:37:10.080 | right? So you have the value in the rev share billions per year going to Microsoft. You have the

00:37:16.880 | profits you make off the $250 billion of the Azure compute commitment from OpenAI. And of course,

00:37:24.160 | you get huge sales from the exclusive distribution of the API. So talk to us how you think about the

00:37:31.440 | value across those domains, especially how this exclusivity has brought a lot of customers who may

00:37:37.760 | have been on AWS to Azure. Yeah, no, absolutely. I mean, so to us, if I look at it, you know,

00:37:45.360 | aside from all the equity parts, the real strategic thing that comes together and that remains going

00:37:52.240 | forward, is that stateless API exclusivity on Azure that helps quite frankly, both OpenAI and us and our

00:37:59.840 | customers. Because when somebody in the enterprise is trying to build an application, they want an API

00:38:06.640 | that's stateless, they want to mix it up with compute and storage, put a database underneath it to capture

00:38:14.240 | your state and build a full workload. And that's where, you know, Azure coming together with this API.

00:38:21.040 | And so what we're doing with even Azure Foundry, right? Because in some sense, you let's say you want

00:38:26.160 | to build an AI application. But the key thing is, how do you make sure that the evals of what you're doing

00:38:32.880 | with AI are great. So that's where you need even a full app server in Foundry. That's what we have done.

00:38:39.600 | And so therefore, I feel that that is the way we will go to market in our infrastructure business.

00:38:45.280 | The other side of the value capture for us is going to be incorporating all this IP. Not only we have the

00:38:52.320 | exclusivity of the model in Azure, but we have access to the IP. I mean, having a royalty free,

00:38:58.880 | let's even forgetting all the know how and the knowledge side of it. But having royalty free

00:39:04.080 | access all the way till seven more years gives us a lot of flexibility business model wise. It's kind

00:39:09.920 | of like having a frontier model for free. In some sense, if you're an MSFT shareholder, that's kind of

00:39:15.440 | where you should start from is to think about, we have a frontier model that we can then deploy,

00:39:20.480 | whether it's in GitHub, whether it's in M365, whether it's in our consumer copilot, then add to it our own

00:39:26.720 | data, post train it. So that means we can have it embedded in the weights there. And so therefore,

00:39:32.640 | we are excited about the value creation on both the Azure and the infrastructure side,

00:39:38.480 | as well as in our high value domains, whether it is in health, whether it's in knowledge work,

00:39:44.000 | whether it's in coding or security.

00:39:45.920 | You've been consolidating the losses from OpenAI. You know, I think you just reported earnings

00:39:50.800 | yesterday. I think you consolidated 4 billion of losses in the quarter. Do you think that investors

00:39:56.160 | are, I mean, they may even be attributing negative value, right? Because of the losses, you know, as they

00:40:01.760 | apply their multiple of earnings, Satya. Whereas I hear this, and I think about all of those benefits we

00:40:07.440 | just described, not to mention the look through equity value that you own in a company that could

00:40:12.800 | be worth a trillion unto itself. You know, do you think that the market is kind of misunderstanding

00:40:19.040 | the value of OpenAI as a component of Microsoft?

00:40:22.240 | Dr. Abdul- Yeah, that's a good one. So I think the approach that Amy is going to take

00:40:26.800 | is full transparency, because at some level, I'm no accounting expert. So therefore, the best thing

00:40:32.320 | to do is to give all the transparency. I think this time around as well, I think that's why the non-GAAP

00:40:38.800 | gap, so that at least people can see the EPS numbers. Because the common sense way I look at it,

00:40:44.960 | Brad, is simple. If you've invested, let's call it $13.5 billion, you can, of course, lose $13.5 billion.

00:40:51.040 | Dr. Abdul- Right, right.

00:40:51.680 | Dr. Abdul- But you can't lose more than $13.5 billion. At least the last time I checked,

00:40:56.000 | that's what you have at risk. You could also say, hey, the $135 billion that is, you know, today,

00:41:02.080 | our equity stake, you know, is sort of illiquid, what have you. We don't plan to sell it. So therefore,

00:41:08.320 | it's got risk associated with it. But the real story I think you are pulling is all the other things

00:41:14.720 | that are happening. What's happening with Azure growth, right? Would Azure be growing if we had not sort of

00:41:19.840 | had the OpenAI partnership. To your point, the number of customers who came from other

00:41:24.160 | clouds for the first time, right? This is the thing that really we benefited from. What's happening with

00:41:31.040 | Microsoft 365? In fact, one of the things about Microsoft 365 was, what was the next big thing

00:41:35.840 | after E5? Guess what? We found it in Copilot. It's bigger than any suite. Like, you know, we talk about

00:41:43.280 | penetration and usage and the pace. It's bigger than anything we have done in our information work,

00:41:50.560 | which we've been at it for decades. And so we feel very, very good about the opportunity to create

00:41:56.240 | value for our shareholders. And then at the same time, be fully transparent so that people can look

00:42:01.680 | through the, what are the losses? I mean, who knows what the accounting rules are, but we will do whatever

00:42:06.240 | is needed and people will then be able to see what's happening. But a year ago, Satya, there were a bunch

00:42:12.160 | of headlines that Microsoft was pulling back on AI infrastructure, right? Fair or unfair, they were out

00:42:17.600 | there. You know, and perhaps you guys were a little more conservative, a little more skeptical

00:42:22.640 | of what was going on. Amy said on the call last night, though, that you've been short power and

00:42:28.000 | infrastructure for many quarters, and she thought that you would catch up, but you haven't caught up

00:42:33.200 | because demand keeps increasing. So I guess the question is, were you too conservative, you know,

00:42:38.000 | knowing what you know now and what's the roadmap from here? Yeah, it's a great question because,

00:42:42.960 | see, the thing that we realized, and I'm glad we did, is that the concept of building a fleet

00:42:51.440 | that truly was fungible, fungible for all the parts of the life cycle of AI, fungible across geographies,

00:42:59.920 | and fungible across generations, right? So because one of the key things is when you have, let's take even

00:43:06.320 | what Jensen and team are doing, right? I mean, they're at a pace. In fact, one of the things I like

00:43:11.120 | is the speed of light, right? We now have GB300s bringing, you know, that we're bringing up. So you

00:43:15.920 | don't want to have ordered a bunch of GB200s that are getting plugged in only to find the GB300s are in

00:43:24.000 | full production. So you kind of have to make sure you're continuously modernizing, you're spreading the

00:43:30.320 | fleet all over, you are really truly fungible by workload, and you're adding to that the software

00:43:37.200 | optimizations we talked about. So to me, that is the decision we made. And we said, look, sometimes you

00:43:42.880 | may have to say no to some of the demand, including some of the open AI demand, right? Because sometimes,

00:43:48.560 | you know, Sam may say, hey, build me a dedicated, you know, big, you know, whatever, multi-gigawatt data

00:43:54.640 | center in one location for training. Makes sense from an open AI perspective. Doesn't make sense from a long-term

00:44:02.800 | infrastructure build out for Azure. And that's where I thought they did the right thing to give them

00:44:07.440 | flexibility to go procure that from others, while maintaining, again, a significant book of business

00:44:13.840 | from open AI. But more importantly, giving ourselves the flexibility with other customers, our own 1P.

00:44:20.960 | Remember, like one of the things that we don't want to do is be short on is, you know, we talk about

00:44:25.920 | Azure. In fact, some of the times our investors are overly fixated on the Azure number. But remember, for me,

00:44:31.280 | the high margin business for me is co-pilot. It is security co-pilot. It's GitHub co-pilot. It's

00:44:37.200 | the healthcare co-pilot. So we want to make sure we have a balanced way to approach the returns that

00:44:43.280 | the investors have. And so that's kind of one of the other misunderstood, perhaps, in our investor

00:44:48.560 | base in particular, which I find pretty strange and funny because I think they want to hold Microsoft

00:44:53.920 | because of the portfolio we have. But man, are they fixated on the growth number of one little thing

00:44:59.120 | called Azure. On that point, Azure grew 39% in the quarter on a staggering 93

00:45:06.960 | billion-dollar run rate. And, you know, I think that compares to GCP that grew at 32% and AWS closer to

00:45:14.080 | 20%. But could Azure, because you did give compute to 1P and because you did give compute to research,

00:45:22.720 | it sounds like Azure could have grown 41%, 42% had you had more compute to offer.

00:45:28.000 | Absolutely. Absolutely. There's no question. There's no question. So that's why I think the internal

00:45:33.040 | thing is to balance out what we think, again, is in the long-term interests of our shareholders

00:45:38.160 | and also to serve our customers well. And also not to kind of, you know, one of the other things was,

00:45:43.440 | you know, people talk about concentration risk, right? We obviously want a lot of OpenAI,

00:45:47.920 | but we also want other customers. And so we're shaping the demand here. You know, we are in a supply,

00:45:53.600 | you know, we're not demand-constrained, we're supply-constrained. So we are shaping the demand

00:45:59.200 | such that it matches the supply in the optimal way with a long-term view.

00:46:04.320 | So to that point, Sacha, you talked about $400 billion, it's an incredible number,

00:46:09.520 | of remaining performance obligations last night. You said that, you know, that's your booked business

00:46:15.440 | today. It'll surely go up tomorrow as sales continue to come in. And you said you're going to,

00:46:21.040 | you know, your need to build out capacity just to serve, that backlog is very high.

00:46:25.440 | You know, how diversified is that backlog to your point? And how confident are you that that $400

00:46:32.720 | billion does turn into revenue over the course of the next couple of years?

00:46:36.960 | Yeah, that $400 billion, it has a very short duration as Amy explained. It's the two-year

00:46:42.720 | duration on average. So that's definitely our intent. That's one of the reasons why we're spending the

00:46:49.760 | capital out clear with high certainty that we just need to clear the backlog. And to your point,

00:46:54.480 | it's pretty diversified both on the 1P and the 3P. Our own demand is quite frankly pretty high for our

00:47:00.480 | one first party. And even amongst third party, one of the things we now are seeing is the rise of all

00:47:08.000 | the other companies building real workloads that are scaling. And so given that, I think we feel very

00:47:14.080 | good. I mean, obviously, that's one of the best things about RPO is you can be planful, quite frankly.

00:47:19.680 | And so therefore, we feel very, very good about building. And then this doesn't include, obviously,

00:47:24.000 | the additional demand that we're already going to start seeing, including the 250,

00:47:28.480 | you know, which will have a longer duration and will build accordingly.

00:47:32.080 | Right. So there are a lot of new entrants, right, in this race to build out compute Oracle, CoreWeave,

00:47:39.840 | Crusoe, et cetera. And normally we think that will compete away margins, but you've somehow managed to

00:47:46.160 | build all this out while maintaining healthy operating margins at Azure. So I guess the question is for

00:47:51.920 | Microsoft, how do you compete in this world that is where people are levering up, taking lower margins

00:47:59.920 | while balancing that profit and risk? And do you see any of those competitors doing deals that cause

00:48:06.480 | you to scratch your head and say, oh, we're just setting ourselves up for another boom and bust cycle?

00:48:11.200 | I mean, at some level, the good news for us has been competing even as a hyperscaler every day.

00:48:18.480 | You know, there's a lot of competition, right, between us and Amazon and Google on all of these,

00:48:23.440 | right? I mean, it's sort of one of those interesting things, which is everything is a commodity, right?

00:48:27.840 | Compute, storage. I remember everybody saying, wow, how can there be a margin? Except at scale,

00:48:33.280 | nothing is a commodity. And so therefore, yes, so we have to have a cost structure, our supply chain

00:48:39.440 | efficiency, our software efficiencies all have to kind of continue to compound in order to make sure

00:48:46.800 | that there's margins, but scale. And to your point, one of the things that I really love about the OpenAI

00:48:53.440 | partnership is it's gotten us to scale, right? This is a scale game. When you have the biggest workload

00:49:00.560 | there is running on your cloud, that means not only are we going to learn faster on what it means to

00:49:06.000 | operate with scale, that means your cost structure is going to come down faster than anything else.

00:49:10.320 | And guess what? That'll make us price competitive. And so I feel pretty confident about our ability to,

00:49:16.480 | you know, have margins. And this is where the portfolio helps. I've always said, you know,

00:49:21.840 | I've been forced into giving the Azure numbers, right? Because at some level, I've never thought of

00:49:28.000 | allocating compute. I mean, my capital allocation is for the cloud, from whether it is Xbox cloud

00:49:34.480 | gaming, or Microsoft 365, or for Azure, it's one capital outlay. And then everything is a meter,

00:49:42.160 | as far as I'm concerned, from an MSFT perspective. It's a question of, hey, the blended average of that

00:49:48.240 | should match the operating margins we need as a company. Because after all, otherwise, we're not a

00:49:54.400 | conglomerate. We're one company with one platform logic. It's not running five, six different businesses.

00:50:00.320 | We're in these five, six different businesses only to compound the returns on the cloud and AI investment.

00:50:06.720 | Yeah, I love that line. Nothing is a commodity at scale. You know, there's been a lot of ink and time

00:50:13.920 | spent even on this podcast with my partner, Bill Gurley, talking about circular revenues,

00:50:18.400 | including Microsoft Stasher credits, right to open AI that were booked as revenue. Do you see anything

00:50:25.360 | going on like the AMD deal, you know, where they traded 10% of their equity and, you know, for a deal

00:50:31.920 | or the Nvidia deal? Again, I don't want to be overly fixated on concern, but I do want to address head

00:50:37.440 | on what is being talked about every day on CNBC and Bloomberg. And there are a lot of these overlapping

00:50:43.760 | deals that are going on out there. Do you, do you, when you think about that in the context of Microsoft,

00:50:49.440 | does any of that worry you again, as to the sustainability or durability of the AI revenues

00:50:56.880 | that we see in the world? Yeah. I mean, first of all, our investment of let's say that 13 and a half,

00:51:03.040 | which was all the training investment that was not booked as revenue. That is the, that is the reason

00:51:08.080 | why we have the equity percentage. That's the reason why we have the 27% or 135 billion. So that was not

00:51:15.680 | something some that somehow that made it into Azure revenue. In fact, if anything, the Azure revenue was

00:51:21.440 | purely the consumption revenue of chat GPT and anything else, and the APIs they put out that they

00:51:28.560 | monetized and we monetized. To your aspect of others, you know, to some degree, it's always been there in

00:51:36.000 | terms of vendor financing, right? So it's not like a new concept that when someone's building something and

00:51:41.760 | they have a customer who's also building something, but they need financing, you know, for whether it is,

00:51:48.240 | you know, it's, it's sort of some, they're taking some exotic forms, which obviously need to be

00:51:53.520 | scrutinized by the investment community. But that said, you know, vendor financing is not a new concept.

00:52:00.480 | Interestingly enough, we have not had to do any of that, right? I mean, we may have, you know, really

00:52:06.080 | either invested in OpenAI and essentially got an equity stake in it for return for compute,

00:52:12.640 | or essentially sold them great pricing of compute in order to be able to sort of bootstrap them. But,

00:52:19.040 | you know, others choose to do so differently. And, and I think circularity ultimately will be tested by

00:52:25.120 | demand because all this will work as long as there is demand for the final output of it. And up to now,

00:52:33.040 | that has been the case. Certainly, certainly. Well, I want to shift, you know, as you said,

00:52:37.760 | over half your business's software applications, you know, I want to think about software and agents,

00:52:43.360 | you know, last year on this pod, you made a bit of a stir by saying that much of the application software,

00:52:49.120 | you know, was this thin layer that sat on top of a CRUD database.

00:52:52.560 | So the notion that business applications exist, that's probably where they will all collapse,

00:53:02.080 | right? In the agent era. Because if you think about it, right, they are essentially CRUD databases

00:53:08.880 | with a bunch of business logic. The business logic is all going to these agents.

00:53:16.320 | public software companies are now trading at about 5.2 times forward revenue. So that's below their

00:53:23.280 | 10 year average of seven times, despite the markets being at all time highs. And there's lots of concern

00:53:29.040 | that SaaS subscriptions and margins may be put at risk by AI. So how today is AI affecting the growth

00:53:37.840 | rates of your software products of, you know, those core products? And specifically, as you think about

00:53:43.200 | database, fabric security, Office 360. And then second question, I guess, is what are you doing

00:53:50.480 | to make sure that software is not disrupted, but is instead superpowered by AI?

00:53:56.240 | Yeah, I think that's right. So the last time we talked about this, my point really there was the

00:54:02.000 | architecture of SaaS applications is changing because this agent tier is replacing the old business logic

00:54:09.120 | tier. And so because if you think about it, the way we built SaaS applications in the past was

00:54:13.120 | you had the data, the logic tier and the UI all tightly coupled. And AI, quite frankly, doesn't

00:54:19.120 | respect that coupling because it requires you to be able to decouple. And yet the context engineering is

00:54:26.240 | going to be very important. I mean, take, you know, something like Office 365. One of the things I love

00:54:30.800 | about our Microsoft 365 offering is it's low ARPU, high usage, right? I mean, if you think about it,

00:54:39.120 | right, Outlook or Teams or SharePoint, you pick Word or Excel, like people are using it all the time,

00:54:45.440 | creating lots and lots of data, which is going into the graph, and our ARPU is low. So that's sort of

00:54:51.120 | what gives me real confidence that this AI tier, I can meet it by exposing all my data. In fact, one of

00:54:59.920 | the fascinating things that's happened, Brad, with both GitHub and Microsoft 365 is thanks to AI, we are

00:55:07.840 | seeing all time highs in terms of data that's going into the graph or the repo. I mean, think about it,

00:55:14.000 | the more code that gets generated, whether it is Codex or Cloud or wherever, where is it going?

00:55:19.680 | GitHub. More PowerPoints that get created, Excel models that get created, all these artifacts and chat

00:55:26.160 | conversations. Chat conversations are new docs. They're all going in to the graph and all that is needed,

00:55:33.520 | again, for grounding. So that's what, you know, you turn it into a forward index, into an embedding,

00:55:40.880 | and basically that semantics is what you really go ground any agent request. And so I think the

00:55:48.240 | next generation of SaaS applications will have to sort of, if you are high ARPU, low usage,

00:55:53.920 | then you have a little bit of a problem. But if you are, we are the exact opposite. We are low ARPU,

00:55:59.760 | high usage. And I think that anyone who can structure that and then use this AI as in fact an accelerant,

00:56:06.640 | because I mean, like, if you look at the M365 Copilot price, I mean, it's higher than any other

00:56:11.920 | thing that we sell. And yet it's getting deployed faster and with more usage. And so I feel very

00:56:18.080 | good. Oh, or coding, right? Who would have thought? In fact, take GitHub, right? What GitHub did in the

00:56:24.000 | first, I don't know, 15 years of its existence or 10 years of its existence, it was basically done in the

00:56:29.120 | last year, just because coding is no longer a tool, it's more a substitute for wages. And so it's a very

00:56:36.880 | different type of business model, even. Kind of thinking about the stack and where value gets

00:56:41.360 | distributed. So until very recently, right, clouds largely ran pre-compiled software, you didn't need a

00:56:48.160 | lot of GPUs, and most of the value accrued to the software layer, to the database, to the applications like

00:56:53.440 | CRM and Excel. But it does seem in the future that these interfaces will only be valuable,

00:56:59.120 | right, if they're intelligent, right? If they're pre-compiled, they're kind of dumb. The software's

00:57:05.040 | got to be able to think and to act and to advise. And that requires, you know, the production of these

00:57:11.120 | tokens, you know, dealing with the ever-changing context. And so in that world, it does seem like much

00:57:17.120 | more of the value will accrue to the AI factory, if you will, to, you know, gents in producing,

00:57:23.040 | you know, helping to produce these tokens at the lowest cost, and to the models. And maybe that

00:57:30.560 | the agents or the software will accrue a little bit less of the value in the future than they've accrued

00:57:35.680 | in the past. Steelman, for me, why that's wrong? Yeah. So I think there are two things that are necessary

00:57:43.200 | to try and to drive the value of AI. One is what you described first, which is the token factory. And

00:57:48.720 | even if you unpack the token factory, it's the hardware silicon system. But then it is about running

00:57:55.600 | it most efficiently with the system software, with all the fungibility, max utilization. That's where the

00:58:03.520 | hyperscaler's role is, right? What is a hyperscaler? Is hyperscaler, like everybody says, if you sort of

00:58:09.040 | said, hey, I want to run a hyperscaler. Yeah, you could say, oh, it's simple. I'll buy a bunch

00:58:12.960 | of servers and wire them up and run it. It's not that, right? I mean, if it was that simple, then

00:58:17.040 | there would have been more than three hyperscalers by now. So the hyperscaler is the know-how of running

00:58:23.280 | that max util and the token factories. And it's not, and by the way, it's going to be heterogeneous.

00:58:28.320 | Obviously, Jensen's super competitive. Lisa is going to come, you know, Hawk's going to produce things from

00:58:34.480 | Broadcom, we will all do our own. So there's going to be a combination. So you want to run

00:58:39.440 | ultimately a heterogeneous fleet that is maximized for token throughput and efficiency and so on.

00:58:45.840 | So that's kind of one job. The next thing is what I call the agent factory. Remember that a SaaS

00:58:52.080 | application in the modern world is driving a business outcome. It knows how to most efficiently use the

00:58:59.520 | tokens to create some business value. In fact, GitHub Copilot is a great example of it, right? Which is,

00:59:06.480 | you know, if you think about it, the auto mode of GitHub Copilot is the smartest thing we've done,

00:59:11.760 | right? So it chooses based on the prompt, which model to use for a code completion or a task handoff,

00:59:19.600 | right? That's what you, and you do that not just by, you know, choosing in some round robin fashion,

00:59:24.720 | you do it because of the feedback cycle you have, you have the evals, the data loops and so on. So

00:59:30.080 | the new SaaS applications, as you rightfully said, are intelligent applications that are optimized for

00:59:36.160 | a set of evals and a set of outcomes that then know how to use the token factory's output

00:59:42.320 | most efficiently. Sometimes latency matters, sometimes performance matters. And knowing how to do that trade

00:59:50.960 | in a smart way is where the SaaS application value is. But overall, it is going to be true that there

00:59:57.120 | is a real marginal cost to software this time around. It was there in the cloud era too. When

01:00:02.240 | we were doing, you know, CD-ROMs, there wasn't much of a marginal cost. You know, with the cloud,

01:00:08.240 | there was, and this time around, it's a lot more. And so therefore, the business models have to adjust,

01:00:13.200 | and you have to do these optimizations for the agent factory and the token factory separately.

01:00:18.480 | You have a big search business that most people don't know about, you know, but it turns out

01:00:23.040 | that that's probably one of the most profitable businesses in the history of the world because

01:00:27.040 | people are running lots of searches, billions of searches, and the cost of completing a search,

01:00:33.040 | if you're Microsoft, is many fractions of a penny, right? It doesn't cost very much to complete a search.

01:00:39.040 | But the comparable query or prompt stack today when you use a chatbot looks different, right? So I guess the

01:00:46.960 | question is, assume similar levels of revenue in the future for those two businesses, right? Do you

01:00:54.160 | ever get to a point where kind of that chat interaction has unit economics that are as profitable as search?

01:01:00.880 | I think that's a great point because, see, search was pretty magical in terms of its ad unit

01:01:09.600 | and its cost economics, because there was the index, which was a fixed cost that you could then amortize

01:01:16.400 | in a much more efficient way. Whereas this one, you know, each chat, to your point, you have to burn a

01:01:23.280 | lot more GPU cycles, both with the intent and the retrieval. So the economics are different. So I think

01:01:29.040 | you know, that's why I think a lot of the early sort of economics of chat have been the freemium model and

01:01:34.560 | subscription on the even on the consumer side. So we are yet to discover whether it's agentic commerce or

01:01:40.640 | whatever is the ad unit, how it's going to be litigated. But at the same time, the fact that at this point, you know, I

01:01:48.240 | kind of know, in fact, I use search for very, very specific navigational queries. I used to say I use

01:01:55.520 | it a lot for commerce, but that's also shifting to my, you know, copilot. Look, I look at the copilot mode in

01:02:03.120 | edge and Bing or copilot. Now they're blending in. So I think that, yes, I think there is going to be

01:02:10.000 | a real litigation, just like that we talked about the SaaS disruption, where in the beginning of the cheese

01:02:15.440 | being a little moved in consumer economics of that category. Right. I mean, and given that it's the

01:02:21.920 | multi-trillion dollar, this is the thing that's driven all the economics of the internet, right?

01:02:26.960 | When you move the economics of search for both you and Google, and it converges on something that looks

01:02:33.120 | more like a personal agent, a personal assistant chat, you know, that could end up being much, much

01:02:39.440 | bigger in terms of the total value delivered to humanity. But the unit economics, you're not just

01:02:44.080 | amortizing this one-time fixed index, right? That's right. That's right. I think that the

01:02:49.360 | consumer- Yeah, the consumer category, because you're putting a thread on something that I think a lot

01:02:54.080 | about, right? Which is what, during these disruptions, you kind of have to have a real sense of where is,

01:03:01.200 | what is the category economics? Is it winner take all? And both matter, right? The problem on consumer space

01:03:11.680 | always is that there's a finite amount of time. And so, if I'm not doing one thing,

01:03:19.280 | I'm doing something else. And if your monetization is predicated on some human interaction in particular,

01:03:26.160 | if there was truly agentic stuff even on consumer, that could be different. Whereas in the enterprise,

01:03:31.280 | one is it's not winner take all. And two, it is going to be a lot more friendly for agentic interaction.

01:03:38.000 | So, it's not like, for example, the per seat versus consumption. The reality is agents are the new seats.

01:03:44.160 | And so, you can think of it as the enterprise monetization is much clearer. The consumer

01:03:50.720 | monetization, I think, is a little more murky. You know, we've seen a spate of layoffs recently with

01:03:55.520 | Amazon announcing some big, big layoffs this week. You know, the mag seven has had little job growth

01:04:00.880 | over the last three years, despite really robust top lines. You know, you didn't grow your head

01:04:06.000 | count really from 24 to 25. It's around 225,000. You know, many attribute this to normal getting fit,

01:04:13.200 | you know, just getting more efficient coming out of COVID. And I think there's a lot of truth to that.

01:04:17.120 | But do you think part of this is due to AI? Do you think that AI is going to be a net job creator?

01:04:23.760 | And do you see this being a long-term positive for Microsoft productivity? Like it feels to me

01:04:29.120 | like the pie grows, but you can do all these things much more efficiently, which either means your margins

01:04:37.120 | expand or it means you reinvest those margin dollars and you grow faster for longer. I call it the golden

01:04:43.360 | agent margin expansion. I'm a firm believer that the productivity curve does and will bend or in the

01:04:53.440 | sense that we will start seeing some of what is the work and the workflow in particular change, right?

01:05:01.840 | There's going to be more agency for you at a task level to get to job complete because of the power of

01:05:10.640 | these tools in your hand. And that I think is going to be the case. So that's why I think we are even

01:05:16.800 | internally, for example, when you talked about even our allocation of tokens, we want to make sure that

01:05:22.560 | everybody at Microsoft standard issue, right? All of them have Microsoft 365 to the tilt in the sort of

01:05:29.040 | most unlimited way and have GitHub co-pilot so that they can really be more productive. But here is the other

01:05:35.920 | interesting thing, Brad, we're learning is there's a new way to even learn, right? Which is, you know,

01:05:42.080 | how to work with agents, right? So that's kind of like when the first, when Word, Excel, PowerPoint all

01:05:47.840 | showed up in office, kind of, we learned how to rethink, let's say, how we did a forecast, right? I mean,

01:05:54.400 | think about it, right? In the 80s, the forecasts were inter-office memos and faxes and what have you. And then

01:06:00.640 | suddenly, somebody said, Oh, here's an Excel spreadsheet, let's put in an email, send it around,

01:06:04.720 | people enter numbers, and there was a forecast. Similarly, right now, any planning, any execution

01:06:11.040 | starts with AI, you research with AI, you think with AI, you share with your colleagues, and what have you. So

01:06:17.440 | there's a new artifact being created and a new workflow being created. And that is the rate of,

01:06:23.920 | the pace of change of the business process that matches the capability of AI, that's where the

01:06:31.920 | productivity efficiencies come. And so organizations that can master that are going to be the biggest

01:06:38.000 | beneficiaries, whether it's in our industry, or quite frankly, in the real world.

01:06:41.600 | And so is Microsoft benefiting from that? You know, so let's, let's think about a couple years from now,

01:06:48.000 | five years from now, at the current growth rate will be sooner, but let's call it five years from now,

01:06:53.120 | your top line is twice as big as what it is today, Satya. How many more employees will you have? If you're,

01:06:58.720 | if you're, if you grow revenue by it?

01:07:00.480 | Like one of the best things right now is these examples that I'm hit with every day from the employees

01:07:05.760 | of Microsoft. There is this person who leads our network operations, right? I mean, if you think about the

01:07:10.560 | amount of fiber we have had to put for like this, you know, this two gigawatt data center, we just

01:07:17.600 | built out in Fairwater, right? And the amount of fiber there, the AI ran and what have you, it's just

01:07:23.840 | crazy, right? So, and it turns out, this is a real world asset. There are, I think 400 different fiber

01:07:30.000 | operators we're dealing with worldwide. Every time something happens, we're literally going and dealing

01:07:34.720 | with all these DevOps pipelines. The person who leads it, she basically said to me, you know what,

01:07:39.440 | I, there's no way I'll ever get the headcount to go do all this. Not forget, even if I even approve

01:07:44.640 | the budget, I can't hire all these folks. So she, she did the next best thing. She just built herself

01:07:49.840 | a whole bunch of agents to automate the DevOps pipeline of how to deal with the maintenance. That

01:07:55.760 | is an example of, to your point, a team with AI tools being able to get more productivity. So

01:08:03.440 | if you have a question, I will say we will grow a headcount. But the way I look at it is that headcount

01:08:09.040 | we grow will grow with a lot more leverage than the headcount we had pre AI. And that's the adjustment,

01:08:15.920 | I think structurally you're seeing first, which is one, you called it getting fit. I think of it as

01:08:22.400 | more getting to a place where everybody is really not learning how to rethink how they work. And it's the

01:08:30.640 | how, not even the what. Even if the what remains the constant, how you go about it has to be relearned.

01:08:37.280 | And it's the unlearning and learning process that I think will take the next year or so. Then the

01:08:42.960 | headcount growth will come with max leverage. Yeah, no, it's a, I think we're on the verge of

01:08:49.920 | incredible economic productivity growth. It does feel like when I talk to you or Michael Dell that

01:08:55.440 | most companies aren't even really in the first inning, maybe the first batter in the first inning in

01:09:01.360 | reworking those workflows to get maximum leverage from these agents. But it sure feels like over the

01:09:06.640 | course of the next two to three years, that's where a lot of gains are going to start coming from.

01:09:10.400 | And again, I, you know, I certainly am an optimist. I think we're going to have net job gains from all of

01:09:16.320 | this. But I think for those companies, they'll just be able to grow their bottom line, their number of

01:09:21.360 | employees slower than their top line. That is the productivity gain to the company. Aggregate all that up.

01:09:27.120 | That's the productivity gain to the economy. And then we'll just take that consumer surplus and invest it in

01:09:32.560 | creating a lot of things that didn't exist before.

01:09:34.880 | A hundred percent. A hundred percent. Even in software development, right? One of the things I look at it

01:09:39.360 | is no one would say we're going to have a challenge in having, you know, more software engineers contribute

01:09:46.800 | to our sort of society. Because the reality is you look at the IT backlog in any organization. And so the

01:09:52.640 | question is all these software agents are hopefully going to help us go and take a whack at all of the IT

01:10:00.240 | backlog we have. And think of that dream of evergreen software. That's going to be true. And then think

01:10:06.800 | about the demand for software. So I think that to your point, it's the levels of abstraction at which

01:10:11.840 | knowledge work happens will change. We will adjust to that, the work and the workflow. That will then

01:10:17.760 | adjust itself even in terms of the demand for the products of this industry.

01:10:21.520 | -I'm going to end on this, which is really around the reindustrialization of America. I've said if you

01:10:27.600 | add up the $4 trillion of CapEx that you and these and so many of the big large US tech companies are

01:10:35.120 | investing over the course of the next four or five years, it's about 10 times the size of the Manhattan

01:10:39.680 | Bitcoin project on an inflation adjusted or GDP adjusted basis. So it's a massive undertaking for

01:10:45.840 | America. But the president has made it a real priority of his administration to recut the trade

01:10:51.680 | deals. And it looks like we now have trillions of dollars. South Koreans committed $350 billion of

01:10:57.280 | investments just today into the United States. And when you think about, you know, what you see going

01:11:04.960 | on in power in the United States, both production, the grid, et cetera, what you see going on in terms

01:11:11.040 | of this reindustrialization, how do you think this is all going? And, you know, maybe just reflect on

01:11:19.360 | where we're landing the plane here and your level of optimism for the few years ahead.

01:11:24.320 | -Yeah, no, I feel very, very optimistic because in some sense, you know, Brad Smith was telling me about

01:11:30.080 | sort of the economy around our Wisconsin data center. It's fascinating. Most people think,

01:11:35.520 | oh, data center, that is sort of like, yeah, it's going to be one big warehouse and there is, you know,

01:11:41.040 | fully automated. A lot of it is true. But first of all, what went into the construction of that data center

01:11:49.120 | and the local supply chain of the data center, that is in some sense, the reindustrialization of the United

01:11:55.920 | States as well. Even before you get to what is happening in Arizona with the TSMC plants or what

01:12:02.640 | was happening with Micron and their investments in memory or Intel and their fabs and what have you,

01:12:08.400 | right? There's a lot of stuff that we will want to start building. Doesn't mean we won't have trade

01:12:14.560 | deals that make sense for the United States with other countries. But to your point, the reindustrialization

01:12:20.400 | for the new economy and making sure that all the skills and all that capacity from power on down,

01:12:28.320 | I think, is sort of very important, right, for us. And the other thing that I also say, Brad,

01:12:33.840 | it's important, and this is something that I've had a chance to talk to President Trump as well as

01:12:38.720 | Secretary Lutnik and others, is it's important to recognize that we as hyperscalers of the United States

01:12:45.360 | are also investing around the world. So in other words, the United States is the biggest investor

01:12:52.400 | of compute factories or token factories around the world. But not only are we attracting foreign capital

01:12:59.920 | to invest in our country so that we can reindustrialize, we are helping, whether it's in Europe or in Asia or

01:13:06.320 | elsewhere, in Latin America and in Africa, with our capital investments, bringing the best American tech

01:13:13.520 | to the world that they can then innovate on and trust. And so both of those, I think, are really

01:13:20.000 | bode well for the United States long term. I'm grateful for your leadership.

01:13:24.640 | Sam is really helping lead the charge at OpenAI for America. I think this is a moment where I look ahead,

01:13:31.600 | you know, you can see 4% GDP growth on the horizon. We'll have our challenges, we'll have our ups and

01:13:36.320 | downs. These tend to be stairs, you know, stairs up rather than a line straight up and to the right.

01:13:41.360 | But I, for one, see a level of coordination going on between Washington and Silicon Valley, between big

01:13:47.200 | tech and the reindustrialization of America that gives me cause for incredible hope. Watching what

01:13:52.320 | happened this week in Asia, led by the president and his team, and then watching what's happening here

01:13:58.560 | is super exciting. So thanks for making the time. We're big fans. Thanks, Satya.

01:14:04.320 | Thanks so much, Brad. Thank you.

01:14:14.320 | As a reminder to everybody, just our opinions, not investment advice.

All things AI w @altcap @sama & @satyanadella. A Halloween Special. 🎃🔥BG2 w/ Brad Gerstner

Chapters