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

00:00:00Yeah, I think this has really been anamazing partnership through every phase.Uh we had kind of no idea where it wasall going to go when we started as Satiasaid. Uh but I I don't think I thinkthis is one of the great techpartnerships uh ever and withoutcertainly without Microsoft andparticularly SA's early conviction uh wewould not have been able to do this.

00:00:23[music]What a week. What a week. Great to seeyou both. Um Sam, how's the baby?>> Baby is great. That's the best thingever, man. Every every cliche is trueand it is the best thing ever.>> Uh hey Sacha, with all your time>> smile on Sam's face whenever he talksabout uh it's just his his baby is justso different. It's dad that and computeI guess when he talks about compute andhis baby. [laughter]>> U well Sachi have you given him any dadtips with all this time you guys havespent together?>> I said just enjoy it. I mean it's soawesome that uh you know I you know wehad our babies or what our children soyoung and I wish I could redo it. So insome sense it's just the most precioustime and as they grow it's just sowonderful. I'm so glad Sam is um>> I'm happy to be doing it older, but I dothink sometimes, man, I wish I had theenergy when I was like 25. Uh thatpart's harder.>> No doubt about it. What's the averageage at Open AI, Sam? Any idea? It'syoung.>> It's not crazy young. Not Not like Notlike most Silicon Valley startups. Idon't know, maybe low 30s average.>> Are babies t is it are babies trendingpositively or negatively?

00:01:39>> Babies trending positively.>> Oh, that's good. That's good. Yeah.>> Well, you guys, such a big week. Youknow, I was thinking about I started atNvidia's GTC, you know, just hit $5trillion. Google, Meta, Microsoft,Satcha, you had your earnings yesterday,you know, and we heard consistently notenough compute, not enough compute, notenough compute. We got rate cuts onWednesday. The GDP's tracking near 4%.And then I was just saying to Sam, youknow, the president's cut these massivedeals in Malaysia, South Korea, Japan,sounds like with China. you know, dealsthat really incredibly provide thefinancial firepower to re-industrializeAmerica. 80 billion for new nuclearfision, all the things that you guysneed to build more compute, butcertainly wasn't what wasn't lost in allof this was you guys had a bigannouncement on Tuesday that clarifiedyour partnership. Congrats on that. AndI thought we'd just start there. Ireally want to just break down the dealin really simple plain language to makesure I understand it and and and othersbut you know we'll just start with yourinvestment Satcha you know Microsoftstarted investing in 2019 has investedin the ballpark at 134 billion into openAI and for that you get 27% of thebusiness ownership in the business on afully

00:02:54diluted basis I think it was abouta third and you took some dilution overthe course of last year with all theinvestmentSo, does that sound about right in termsof ownership?>> Yeah, it does. But I I would say beforeeven our stake in it, Brad, I thinkwhat's pretty unique about OpenAI is thefact that as part of OpenAI's process ofrestructuring, one of the largestnonprofit gets created. I mean, let'snot forget that, you know, in some senseI say at Microsoft, like I, you know, weare very proud of the fact that we were,we're associated with the two of thelargest nonprofits, the Gates Foundationand now the OpenAI Foundation. So,that's I think the big news. Uh, weobviously were, you know, are thrilled.It's not what we thought. And as I saidto somebody, it's not like when we firstinvested our billion dollars that, oh,this is going to be the 100 bagger thatI'm going to be talking about to VCsabout, but here we are. But we are verythrilled to be an investor and an earlybacker. Um and and it's a great and it'sa really a testament to what Sam andteam have done quite frankly. I meanthey obviously had the vision earlyabout what this technology could do andthey ran with it and just executed youknow in a masterful way.>> Yeah. I think this has really been anamazing partnership through every phase.Uh we had

00:04:09kind of no idea where it wasall going to go when we started as Satiasaid. Uh but I I don't think I thinkthis is one of the great techpartnerships uh ever and withoutcertainly without Microsoft andparticularly Sant's early conviction uhwe would not have been able to do this.I don't think there were a lot of otherpeople that would have uh been willingto take that kind of a bet given whatthe world looked like at the time. Um wedidn't know exactly how the tech wasgoing to go. Well, not exactly. Wedidn't know at all how the tech wasgoing to go. We just had a lot ofconviction in this this one idea ofpushing on on deep learning and trustingthat if we could do that, we'd figureout ways to make wonderful products andcreate a lot of value and also, as Satiasaid, create what we believe will be thelargest nonprofit ever. And I think it'sgoing to do amazingly great things. Itit was I I really like the structurebecause it lets the nonprofit grow invalue while the PBC is able to get thecapital that it needs to keep scaling. Idon't think the nonprofit would be ableto be this valuable if we didn't come upwith the structure and if we didn't havepartners around the table that wereexcited for it to work this way. But,you know, I think it's been six morethan six years since we first startedthis partnership and uh a pretty crazyamount of achievement for six years andI think much much more to come. I hopethat Sasha makes a trillion dollars onthe investment, not hundred billion,

00:05:24youknow, whatever it is.>> Well, as part of the restructuring, youguys talked about it. You have thisnonprofit on top and a public benefitcorp below. It's pretty insane. Thenonprofit is already capitalized with$130 billion. $130 billion of Open AIstock. It's one of the largest in theworld out of the gates. It could end upbeing much much larger. The CaliforniaAttorney General said they're not goingto object to it. You already haveundthis 130 billion dedicated to makingsure that AGI benefits all of humanity.You announced that you're going todirect the first 25 billion to healthand AI security and resilience. Sam,first let me just say, you know, assomebody who participates in theecosystem, kudos to you both. It'sincredible this contribution to thefuture of AI. But Sam, talk to us a bitabout the importance of the the choicearound health and and resilience. Andthen help us understand how do we makesure that you get maximal benefitwithout it getting weighted down aswe've seen with so many nonprofits withits own political biases.>> Yeah. First of all, the the best way tocreate a bunch of value for the world ishopefully what we're we've already beendoing, which is to make these amazingtools and just let people use them. AndI think

00:06:39capitalism is great. I thinkcompanies are great. I think people aredoing amazing work getting advanced AIinto the hands of a lot of people andcompanies. They're doing incrediblethings. There are some areas where the Ithink market forces don't quite work forwhat's in the best interest of peopleand you do need to do things in adifferent way. Uh there are also somenew things with this technology thatjust haven't existed before like thepotential to use AI to do science at arapid clip like really truly automateddiscovery. And when we thought about theareas we wanted to first focus on,clearly if we can cure a lot of diseaseand make the data and information forthat broadly available, that wouldthat'd be a wonderful thing to do forthe world. And then on this point of AIresilience, I do think some things mayget a little strange and they won't allbe addressed by companies doing theirthing. So as the world has to navigatethrough this transition, if we can fundsome work to help with that, and thatcould be, you know, cyber defense, thatcould be AI safety research, that couldbe economic studies, all of thesethings, helping society get through thistransition smoothly. We're veryconfident about how great it can be onthe other side, but you know, I'm surethere will be some choppiness along theway.>> Let's keep busting through the the theum the deal. So models and

00:07:54exclusivitySam OpenAI can distribute its models uhits leading models on Azure but I don'tthink you can distribute them on anyother leading the big clouds for sevenyears until 2032 but that would endearlier if AGI is verified. We can comeback to that but you can distribute youropen source models Sora agents codecswearables everything else on otherplatforms. So Sam, I assume this meansno chat GPT or GPT6 on Amazon or Google.>> No. So, so we have a C. First of all, wewant to do lots of things together tohelp, you know, create value forMicrosoft. We want them to do lots ofthings for to create value for us. Andthere are many many things that'llhappen in that category. Um, we arekeeping what Satia termed once and Ithink it's a great phrase of statelessAPIs on Azure exclusively through 2030.And everything else we're going to, youknow, distribute elsewhere and that'sobviously in Microsoft's interest, too.So, we'll put lots of products, lots ofplaces, and then this thing we'll we'lldo on Azure and people can get it thereor or via us. And I think that's great.>> And then the rev share, there's still arev share that gets paid by OpenAI toMicrosoft on all your revenues that alsoruns until 2032 or until AGI isverified. So, let's just assume for thesake of argument, I know this ispedestrian, but it's important that therev share is 15%. So that would mean ifyou had

00:09:1020 billion in revenue thatyou're paying three billion to Microsoftand that counts as revenue to Azure.Satcha, is that does that sound aboutright?>> Yeah, we have a rev share and I think asyou characterized it is either going toAGI or till the end of the term. Uh andI actually don't know exactly where wecount it quite honestly whether it goesinto Azure or somewhere else. That's agood question. It's a good question forAmy. Given that both exclusivity and therevshare end early in the case AGI isverified, it seems to make AGI a prettybig deal. And as I understand it, youknow, if if OpenAI claimed AGI, itsounds like it goes to an expert panel.And you guys basically select a jurywho's got to make a relatively quickdecision whether or not AGI has beenreached. Satcha, you said on yesterday'searning call that nobody's even close togetting to AGI and you don't expect itto happen anytime soon. You talked aboutthis spiky and jagged intelligence. Sam,I've heard you perhaps sound a littlebit more bullish on, you know, when wemight get to AGI. So, I guess thequestion is to you both. Do you worrythat over the next two or three yearswe're going to end up having to call inthe jury to effectively make a uh a callon whether or not we've hit AGI?>> I I realize you got to try to make somedrama between us here. I [laughter]>> you know, I

00:10:25think putting a process inplace for this is a good thing to do. Iexpect that the technology will takeseveral surprising twists and turns andwe will continue to be good partners toeach other and figure out what makessense.>> That's well said. I think uh and that'sone of the reasons why I think thisprocess we put in place is a good oneand at the end of the day I'm a bigbeliever in the fact that intelligenceuh capability wise is going to continueto improve and our real goal quitefrankly is that which is how do you putthat in the hands of people andorganizations so that they can get themaximum benefits and that was theoriginal mission of open AI thatattracted me to open AAI and Sam andteam and that's kind of what we plan tocontinue on>> Brad to say the obvious if we had superintelligence tomorrow, we would stillwant Microsoft's help getting thisproduct out into people's hands and wewant them like Yeah,>> of course. Of course. Yeah. No, it againI'm asking the questions I know that areon people's minds and that makes a tonof sense to me. Obviously s Microsoft isone of the largest distributionplatforms in the world. You guys havebeen great partners for a long time. ButI think it dispels some of the mythsthat are out there. But let's shiftgears a little bit. You know, obviouslyOpenAI is one of the fastest growingcompanies in history. Satcha, you saidon the pod

00:11:40a year ago, this pod, thatevery new phase shift creates a newGoogle and the Google of this phaseshift is already known and it's open AI.And none of this would have beenpossible had you guys not made thesethese huge bets. With all that said, youknow, OpenAI's revenues are still areported 13 billion in 2025. And Sam, onyour live stream this week, you talkedabout this massive commitment tocompute, right? 1.4 4 trillion over thenext four or five years with you knowbig commitments 500 million to Nvidia300 million to AMD and Oracle 250billion to Azure. So I think the singlebiggest question I've heard all week andand hanging over the market is how youknow how can a company with 13 billionin revenues make 1.4 4 trillion of spendcommitments, you know, and and andyou've heard the criticism, Sam.>> First of all, we're doing well morerevenue than that. Second of all, Brad,if you want to sell your shares, I'llfind you a buyer. [laughter]>> I just enough like, you know, people areI I think there's a lot of people whowould love to buy OpenAI shares. I don'tI don't think you>> including myself, including myself,>> people who talk with a lot of likebreathless concern about our comput

00:12:55stuff or whatever that would be thrilledto buy shares. So I think we we couldsell you know your shares or anybodyelse's to some of the people who aremaking the most noise on Twitterwhatever about this very quickly. We doplan for revenue to grow steeply.Revenue is growing steeply. We aretaking a forward bet that it's going tocontinue to grow grow and that not onlywill Chhatabt keep growing but we willbe able to become one of the importantAI clouds that our consumer devicebusiness will be a significant andimportant thing that AI that canautomate science will create huge value.So, you know, there are not many timesthat I want to be a public company, butone of the rare times it's appealing iswhen those people are writing theseridiculous OpenAI is about to go out ofbusiness and, you know, whatever. Iwould love to tell them they could justshort the stock and I would love to seethem get burned on that. Um, butyou know, I we carefully plan, weunderstand where the technology, wherethe capability is going to grow, go andand how the products we can build aroundthat and the revenue we can generate. wemight screw it up like this is the betthat we're making and we're taking arisk along with that. A certain risk isif we don't have the compute, we willnot be able to generate the revenue ormake the models at these at this kind ofscale.>> Exactly. And>> let me

00:14:10just say one thing uh Brad asboth a partner and um an investor thereis not been a single business plan thatI've seen from OpenAI that they have putin and not beaten it. So in some sensethis is the one place where you know interms of their growth and just even thebusiness it's been unbelievableexecution quite frankly I mean obviouslyopenai everyone talks about all thesuccess in the usage and what have youbut even um I would say all up uh thebusiness execution has been just prettyunbelievable. I heard Greg Brockman sayon C CBC a couple weeks ago, right? Ifwe could 10x our compute, we might nothave 10x more revenue, but we'dcertainly have a lot more revenue>> simply because of lack of compute power.Things like, yeah, it's just it's reallywild when I just look at how much we areheld back. And in many ways, we have,you know, we've scaled our computeprobably 10x over the past year, but ifwe had 10x more compute, I don't know ifwe'd have 10x more revenue, but I don'tthink it'd be that far. And we heardthis from you as well last night Satchathat you were compute constrained andgrowth would have been higher even if ifyou had more compute. So help uscontextualize Sam maybe like how computeconstrained

00:15:25do you feel today and do youwhen you look at the buildout over thecourse of the next two to three years doyou think you'll ever get to the pointwhere you're not compute constrained?>> We talk about this question of is thereever enough compute a lot. I I think theanswer isthe only the best way to think aboutthis is like aenergy or something. You can talk aboutdemand for energy at a certain pricepoint, but you can't talk about demandfor energy without talking about atdifferentyou know different demand at differentprice levels. If the price of computeper like unit of intelligence orwhatever, however you want to thinkabout it, fell by a factor of a 100tomorrow, you would see usage go up bymuch more than 100 and there'd be a lotof things that people would love to dowith that compute that just make noeconomic sense at the current cost, butthere would be new kind of demand. So Ithink the thenow on the other hand as the models geteven smarter and you can use thesemodels to cure cancer or discover novelphysics or drive a bunch of humanoidrobots to construct a space station orwhatever crazy thing you want then maybethere's huge willingness to pay a muchhigher rate cost per unit ofintelligence for a much higher level ofintelligence that we don't know yet butI would bet there will

00:16:40be. So I I thinkwhen you talk about capacity it's it'slike a you know cost per unit and youknow capability per unit and you have tokind of without those curves it's sortof a madeup it's not a super wellspecified problem.>> Yeah. I mean I think the one thing thatyou know Sam you've talked about which Ithink is the right way is to think aboutis that if intelligence is what a log ofcompute then you try and really makesure you keep getting efficient and sothat means the tokens per dollar perwatt uh and the economic value that thesociety gets out of it is what we shouldmaximize and reduce the costs and sothat's where if you sort of where likethe Jevans paradox point is that rightwhich is you keep reducing itcommoditizing in some sense intelligenceuh so that it becomes the real driver ofGDP growth all around.>> Unfortunately, it's something closer touh log of intelligence equals log ofcompute. But we may figure out betterscaling laws and we may figure out howto beat this. Yeah,>> we heard from both Microsoft and Googleyesterday. Both said their cloudbusinesses would have been growingfaster if they have more GPUs. You know,I asked Jensen on this pod if there wasany chance over the course of the next 5years we would have a compute glut. andhe said it's virtually non-existentchance in the next 2 to

00:17:553 years and Iassume you guys would both agree withJensen that while we can't see out 5 6 7years certainly over the course of thenext 2 to three years for the for thereasons we just discussed that it'salmost a non-existent chance that youhave excess compute well I mean I thinkthe the cycles of demand and supply inthis particular case you can't reallypredict right I mean even the the pointis What's the secular trend? The seculartrend is what Sam said, which is at theend of the day, because quite frankly,the the biggest issue we are now havingis not a compute glut, but it's a powerand it's sort of the ability to get thebuilds done fast enough close to power.So, if you can't do that, you mayactually have a bunch of chips sittingin inventory that I can't plug in. Infact, that is my problem today, right?It's not a supply issue of chips. It'sactually uh the fact that I don't havewarm shells to plug into. And so howsome supply chain constraints emergetough to predict uh because the demandis just going you know is tough topredict right I mean I wouldn't you it'snot like Sam and I would want to besitting here saying oh my god we're lessshort on compute it's because we justwere not that good at being able toproject out what the demand would

00:19:11reallylook like. So I think that that's and bythe way the worldwide side right oneit's one thing to sort of talk about onesegment in one country but it's aboutyou know really getting it out toeverywhere in the world and so therewill be constraints and how we workthrough them is going to be the mostimportant thing it won't be a linearpath for sure there there will come aglut for sure and whether that's like intwo to three years or five to six Ican't tell you but uh like it's going tohappen at some point probably severalpoints along the way like this isthere's something deep about humanpsychology here and bubbles and also asSatia said like there's it's such acomplex supply chain weird stuff getsbuilt the technological landscape shiftsin big ways so you know ifa very cheap form of energy comes onlinesoon at mass scale then a lot of peopleare going to be extremely burned withexisting contracts they've signed it Iif if we can continue this unbelievablereduction in cost per unit ofintelligence let's say it's beenaveraging like 40x X for a given levelper year. You know, that's like a veryscary exponentfrom an infrastructure buildoutstandpoint. Now, again, we're taking thebet that there will be a lot more demandas that gets cheaper, but I have somefear that it's just

00:20:26like, man, we keepgoing with these breakthroughs andeverybody can run like a personal AGI ontheir laptop and we just did an insanething here. Some people are going to getreally burned like has happened in everyother tech infrastructure cycle at somepoints along the way.>> I think that's really well said and youhave to hold those two simultaneoustruths. We had that happen in 20201 andyet the internet became much bigger andproduced much greater outcomes forsociety than anybody estimated in thatperiod of time.>> Yeah. But I think that the one thingthat Sam said is not talked about enoughwhich is the current for example theoptimizations that OpenAI has done onthe inference stack for a given GPU. Imean I it's kind of like it's you knowwe talk about the MOS law improvement onone end but the software improvementsare much more exponential than that.Someday we will make a incredibleconsumer device that can run a GPT5 orGPD6 capable model completely locally ata low power draw. And this is like sohard to wrap my head around.>> That will be incredible. And you knowthat's the type of thing I think thatscares some of the people who arebuilding obviously these largecentralized compute uh stacks. AndSatcha you've talked a lot about thedistribution both to the edge as well ashaving

00:21:41inference capability distributedaround the world. Yeah, I mean the wayat least I've thought about it is moreabout really building a fungeable fleet.I mean when I look at sort of in thecloud infrastructure business, one ofthe key things you have to do is havetwo things. One is an effic like in thiscontext in a very efficient tokenfactory and then high utilization.That's that's it. There are two simplethings that you need to achieve and inorder to have high utilization you haveto have multiple workloads that can bescheduled even on the training. I mean,if you look at the AI pipelines, there'spre-training, there's mid-training,there's post- training, there's RL. Youwant to be able to do all of thosethings. So, thinking about fungeibilityof the fleet is everything for a cloudprovider.>> Okay. So, Sam, you referenced, you know,and and Reuters was reporting yesterdaythat OpenAI may be planning to go publiclate 26 or in 27.>> No, no, no. We we don't we don't haveanything that specific. I I'm a realist.I assume it will happen someday, butthat was uh I don't know why peoplewrite these reports. We don't have likedate in mind decision to do this oranything like that. I just assume it'swhere things will eventually go.>> But it does seem to me if you guys were,you know, are are doing in excess ofhundred billion dollars of revenue in 28or 29 that you at least would be in pos

00:22:56>> what?>> How about 27?>> Yeah, 27 even better. You are inposition to do an IPO and the rumoredtrillion dollars. Again, just tocontextualize for listeners, if you guyswent public at 10 times 100 billion inrevenue, right, which would be, I think,a lower multiple than Facebook wentpublic at, a lower multiple than a lotof other uh big consumer companies wentpublic at, that would put you at atrillion dollars. If you floated 10 to20% of the company, that raises ahundred to$200 billion, which seems likethat would be a good path to fund a lotof the growth and a lot of the stuffthat we just talked about. So, you'reyou're you're not opposed to it. You'renot But you guys are making fund thecompany with revenue growth, which iswhat I would like us to do.>> But no doubt about it.Well, I've also said I think that thisis such an important company and youknow there are so many people includingmy kids who like to trade their littleaccounts and they use chat GPT and Ithink having retail investors have anopportunity to buy one of the mostimportant and largest>> honestly that that is probably thesingle most appealing thing about it tome. Um that would be really nice.One

00:24:11of the things I've talked to youboth about um shifting gears again ispart of the big beautiful bill, youknow, Senator Cruz had included federalpreeemption so that we wouldn't havethis state patchwork 50 different lawsthat mireers the industry down in kindof needless compliance and regulation.unfortunately got killed at the lastsecond by Senator Blackburn becausefrankly I think AI is pretty poorlyunderstood in Washington and there's alot of dumerism I think that has gainedtraction in Washington. So now we havestate laws like the Colorado AI act thatgoes into full effect in February Ibelieve that creates this whole newclass of litigants anybody who claimsany unfair impact from an algorithmicdiscrimination in a chatbot. So somebodycould claim harm for countless reasons.Sam, how worried are you that, you know,having this state patchwork of AI, youknow, poses real challenges to, youknow, our ability to continue toaccelerate and compete around the world.>> I don't know how we're supposed tocomply with that California, sorry,Colorado law. I would love them to tellus uh and, you know, we'd like to beable to do it, but that's just from whatI've read of that. That's like a Iliterally don't know what we're supposedto do. I'm very worried about a 50-state

00:25:27patchwork. I think it's a big mistake. Ithink it's there's a reason we don'tusually do that for these sorts ofthings. I think it'd be bad.>> Yeah. I mean, I think the thefundamental problem of um you know, thispatchwork approach is quite frankly, Imean, between OpenAI and Microsoft,we'll figure out a way to navigate this,right? I mean, uh we can figure thisout. The problem is anyone starting astartup and trying to kind this it'ssort of it just goes to the exactopposite of I think what the intent hereis which obviously safety is veryimportant making sure that thefundamental um you know concerns peoplehave are addressed but there's a way todo that at the federal level so I thinkthe U if we don't do this again you knowEU will do it and then that'll cause itsown issues so I think if US leads it'sbetter uh as you as one regulatoryframework>> for sure.>> And to be clear, it's not that one isadvocating for no regulation. It'ssimply saying let's have, you know,agreed upon regulation at the federallevel as opposed to 50 competing statelaws which certainly uh firebombs thethe AI startup industry and I think itmakes it makes it super challenging evenfor companies like yours who can affordto defend all these cases.>> Yeah. And I would just say quite franklymy hope is

00:26:42that this time around evenacross EU and the United States likethat'll be the dream right quite franklyfor any European startup.>> I don't think that's going to happen.>> What is that?>> That would be great. I don't I wouldn'thold your breath for that one. Thatwould be great. No, but I I I reallythink that if you think about it right,if you sort of if anyone in Europe isthinking about their you know what howcan they participate in this AI uheconomy with their companies uh thisshould be the main concern there aswell. So therefore uh that's I hopethere is some enlightened approach to itbut I agree with you that you know todayI wouldn't bet on that.I do think that with Sachs as the AISare, you at least have a president thatI think might fight for that in terms ofcoordination of of AI policy, usingtrade as a lever to make sure that, youknow, we don't end up with overlyrestricted European policy. But we shallsee. I think first things first, federalpreeemption in the United States ispretty critical. You know, we've beendown in the weeds a little bit here,Sam. So, I want to telescope out alittle bit. You know, I've heard peopleon your team talk about all the greatthings coming up and and as you startthinking about much more unlimitedcompute chat GPT6 and beyond robotics,physical devices,

00:27:58scientific research as you as you lookforward to 2026, what do you thinksurprises us the most? What what whatwhat are you most excited about in termsof what's on the drawing board? you Imean you just hit on a lot of the keypoints there. I I thinkcodeex has been a very cool thing towatch this year and as these go frommulti-our tasks to multi-day tasks whichI expect to happen next year what peoplebe able to do to createsoftware at an unprecedented rate andand really in fundamentally new ways.I'm very excited for that. I think we'llsee that in other industries too. I havelike a bias towards coding. I understandthat one better. I think we'll see thatreally start to transform what peopleare capable of. I I I hope for verysmall scientific discoveries in 2026,but if we can get those very small ones,we'll get bigger ones in future years.That's a really crazy thing to say isthat like AI is going to make a novelscientific discovery in 2026. Even avery small one. This is like this is awildly important thing to be talkingabout. So, I'm excited for that.Certainly, robotics and computer and newkind of computers in future years.That'll be that'll be uh very important.Butyeah, my personal bias is if we canreally get AI to do science here, thatis

00:29:14I mean that is super intelligence insome sense. Like if if this is expandingthe total sum of human knowledge that isa crazy big deal.>> Yeah. I mean I think one of the thingsto use your codeex example I think thecombination of the model capability Imean if you think about the magicalmoment that happened with chat GPT wasthe UI that met intelligence that justtook off right there it's just you knowunbelievable right form fact and some ofit was also the instruction followingpiece of model capability was ready forchat I think that that's what the codeexand the you know these coding agents areabout to uh help us which is what's thatyou know coding agent goes off for along period of time comes back and thenI'm then dropped into what I shouldsteer like one of the metaphors I thinkwe're all sort of working towards is Ido this macro delegation and microsteering what is that UI meets this newintelligence capability and you can seethe beginnings of that with codeex rightthe way at least I use it inside aGitHub copilot is I you know it's Now,it's just a it's a just a different waythan the chat interface. And I thinkthat that I think would be a new way forthe human computer interface. Quitefrankly, it's

00:30:29probably bigger than>> uh that that might be the departure.>> That's one reason I'm very excited thatwe're doing new form factors ofcomputing devices cuz computers were notbuilt for that kind of workflow verywell. Certainly, a UI like Chacht iswrong for it. But this idea that you canhave a device that is sort of alwayswith you but able to go off and dothings and get micro steer from you whenit needs and have like really goodcontextual awareness of your whole lifeand flow. And I think that'll be cool.>> And what neither of you have talkedabout is the consumer use case. I thinka lot about, you know, again, we gounder this device and we have to huntand peck through a hundred differentapplications and fill out little webforms, things that really haven'tchanged in 20 years. But to just have,you know, a personal assistant that wetake for granted perhaps that weactually have a personal assistant, butto give a personal assistant forvirtually free to billions of peoplearound the world to improve their lives,whether it's, you know, ordering diapersfor their kid or whether it's, you know,booking their hotel or or or makingchanges in their calendar. I thinksometimes it's the pedestrian that'sthat's the most impactful. And as wemove from answers to memory and actionsand then the ability to interface withthat through an earbud or some otherdevice that doesn't require me toconstantly be st staring at thisrectangular piece

00:31:44of glass. I think it'spretty extraordinary.>> I think that that's what Sam wasteasing.>> Yeah. Yeah.>> Hope we get it right. I got to drop offunfortunately.>> Sam, it was great to see you. Thanks forjoining us. Congrats again on this bigstep forward and we'll talk soon.>> Thanks for letting me crash.>> See you Sam. Take care. See you.>> As Samwell knows, we're certainly abuyer, not a seller. Um, but but butsometimes, you know, I think it'simportant because the world, you know,we're a pretty small, we spend all daylong thinking about this stuff, right?And so conviction, it comes from the10,000 hours we've spent thinking aboutit. But the reality is we have to bringalong the rest of the world. And therest of the world doesn't spend 10,000hours thinking about this. Um, andfrankly they look at some things thatappear overly ambitious, right, and getworried about whether or not we can pullthose things off. You took this idea tothe board in 2019 to invest a billiondollars into open AI. Was it ano-brainer in the boardroom? You know,did you have to expend any politicalcapital to get it done? dish dish for mea little bit like what that moment waswas like because I think it was such apivotal moment not just for Microsoftnot just for the country but I really dothink for the world.

00:32:59Yeah, I mean it'sit's interesting when you look back thethe journey when I look at it it's beena you know we were involved even in 2016uh when initially open AI uh started infact Azure was even the first sponsor Ithink and then they were doing a lotmore reinforcement learning at that timeI remember the Dota 2 competition Ithink happened on Azure and then uh theymoved on to other things and you know Iwas interested in RL but quite franklyyou know it speaks a little bit to your10,000 hours or the prepared had mind.Uh Microsoft since 1995 was obsessed. Imean, Bill's obsession for the companywas natural language. Natural language.I mean, after all, we're a codingcompany. We're information work company.>> So, it's when Sam in 2019 startedtalking about text and natural languageand transformers and scaling laws.>> Uh that's when I said, "Wow, like thisis an interesting I mean he, you know,this is a team that was going in thedirection or the direction of travel wasnow clear. it had a lot more overlapwith our interest. So in that sense itwas a no-brainer. Obviously you go tothe board and say hey I have an idea oftaking a billion dollars and giving itto this crazy structure which we don'teven kind of understand what is it. It'sa nonprofit

00:34:15blah blah blah and andsaying go for it. Uh there was a debate.Uh Bill was kind of rightfully soskeptical because and then he becamelike once he saw the GPD4 demo like thatwas like the thing that Bill's talkedabout publicly where uh when he saw ithe said it's the best demo he saw afteryou know what Charles Simony showed himat Xerox Park and but you know quitehonestly none of us could uh so themoment for me was that you know let's gogive it a shot then seeing the earlycodeex inside of uh copilot inside of uhGitHub copilot and seeing just the codecompletions and seeing it work. That'swhen I would say we I I felt like I cango from 1 to 10 because that was the bigcall quite frankly. One wascontroversial.>> Uh but the 1 to 10 was what really madethis entire era possible and thenobviously uh the great execution by theteam and the productization on theirpart, our part. I mean if I think aboutit right the collective monetizationreach of GitHub copilot chat GPTMicrosoft 365 copilot and co-pilot youadd those four things that is it rightthat's the biggest sort of AI set ofproducts uh out

00:35:30there on the planet andthat's um you know what obviously haslet us sustain all of this and I thinknot many people know that your CTO KevinScott you know an ex googler lives downhere in Silicon Valley and tocontextualize it right Microsoft hadmissed out on search had missed out onmobile. You become CEO, almost hadmissed out on the cloud, right? Youyou've described it, caught the lasttrain out of town to capture the cloud.And I think you were pretty determinedto have eyes and ears down here so youdidn't miss the next big thing. So Iassume that Kevin played a good role foryou as well.>> Absolutely.>> Deep Seek and Open AI.>> Yeah. I mean I mean if uh it's in fact Iwould say Kevin's conviction uh andKevin was also skeptical like that wasthe thing I I I always watch for peoplewho are skeptical who change uh theiropinion because to me that's a signal soI'm always looking for someone who's anon-believer in something and thensuddenly changes and then they getexcited about it that I have all thetime for that because I'm then curiouswhy what and so Kevin started with allof us were kind of skeptical Right. No,I mean in some sense it defies the theyou know we're all having gone to schooland said god you know there

00:36:45must be analgorithm to crack this versus justlet's scaling laws and throw compute.But quite frankly uh Kevin's convictionthat this is worth going after is one ofthe big things that drove this. Well, wetalk about, you know, that thatinvestment that that's now worth 130billion, I suppose, could be worth atrillion someday, as Sam says, but itreally in many ways understates thevalue of the partnership, right? So, youhave the value in the revshare, billionsper year going to Microsoft. You havethe profit you make off the $250 billionof the Azure compute commitment fromOpenAI. And of course you get huge salesfrom the exclusive distribution of theAPI. So talk to us how you think aboutthe value across those domainsespecially how this exclusivity hasbrought a lot of customers who may havebeen on AWS to Azure.>> Yeah. No absolutely. I mean so to us umif I look at it um you know aside fromall the uh the equity parts the realstrategic thing that comes together andthat remains going forward uh is thatstateless API exclusivity on Azure thathelps quite frankly both open AAI and usand our customers

00:38:01uh because whensomebody in the enterprise uh is tryingto build an application they want an APIthat's stateless they want to mix it upwith uh compute in storage, put adatabase underneath it to capture stateand build a full workload and that'swhere uh you know Azure coming togetherwith this API and so what we're doingwith even uh Azure foundry right becausein some sense you let's say you want tobuild an AI application but the keything is uh how do you make sure thatthe evalare great so that's where you need evena full app server in Foundry that's whatwe've done and so therefore I feel thatthat is the way we will go to market inour infrastructure business. The otherside of the value capture for us isgoing to be incorporating all this IP.Not only we have the exclusivity of themodel in uh Azure but we have access tothe IP. I mean having a royaltyfreelet's even forgetting all the theknowhow and the knowledge side of it buthaving royalty-free access all the waytill seven more years gives us a lot offlexibility business model wise. It'skind of like having a frontier model forfree uh in some sense if you're an MSFTshareholder. That's kind of where youshould

00:39:16start from is to think about wehave a frontier model that we can thendeploy whether it's in GitHub, whetherit's in M365, whether it's in ourconsumer copilot, then add to it our owndata, post train it. Uh so that means wecan have it embedded in the weightsthere. And so therefore we're excitedabout the value creation on both theAzure and the infrastructure side aswell as in our high value domains uhwhether it is in health whether it's inknowledge work whether it's in coding orsecurity>> you've been consolidating the lossesfrom open AI you know I think you youjust reported earnings yesterday I thinkyou consolidated 4 billion of losses inthe quarter do you think that investorsare I mean they may even be attributingnegative value right because of thelosses you know as they apply theirmultiple of earnings. Satcha, whereas Ihear this and I think about all of thosebenefits we just described, not tomention the look through equity valuethat you own in a company that could beworth a trillion unto itself. You know,do you think that the market is is iskind of misunderstanding the value ofopen AI as a component of Microsoft?>> Yeah, that's a good one. So, I think thethe approach that Amy is going to takeis full transparency because at somelevel I'm no accounting expert.

00:40:31Sotherefore the best thing to do is togive uh all of the transparency I thinkthis time around as well. I think that'swhy the non-GAAP gap so that at leastpeople can see the EPS numbers becausethe the the common sense way I look atit Brad is simple. If you've investedlet's call it 13.5 billion. You can ofcourse lose 13.5 billion but you can'tlose more than 13.5 billion. At leastthe last time I checked that's what youhave at risk. You could also say hey the$135 billion that has you know today ourequity stake you know is sort of illquidwhat have you we don't plan to sell itso therefore it's got risk associatedwith it but the real story I think youwere pulling is all the other things uhthat are happening what's happening withAzure growth right would Azure begrowing if we had not sort of had theopeni partnership to your point thenumber of customers who came from othercloudsfor the first time right this is thething that really we benefited fromwhat's happening with Microsoft 365. Infact, one of the things about Microsoft365 was what was the next big thingafter E5? Guess what? We found it incopilot. It's bigger than any suite.Like you know, we talk about penetrationand usage uh and

00:41:46the pace. It's biggerthan anything we've done in ourinformation work which we've been addedfor decades. And so so we pretty feelvery very good about the opportunity tocreate value for our shareholders. Uhand then at the same time be fullytransparent so that people can lookthrough the what are the losses. I meanwho knows what the accounting rules arebut we will do whatever is needed andpeople will then be able to see what'shappening. But a year ago, Satcha, therewere a bunch of headlines that Microsoftwas pulling back on AI infrastructure,right? Fair or unfair, they're they wereout there, you know, and and and perhapsyou guys were a little moreconservative, a little more skeptical ofwhat was going on. Amy said on the calllast night, though, that you've beenshort power and infrastructure for manyquarters, and she thought that you wouldcatch up, but you haven't c caught upbecause demand keeps increasing. So Iguess the question is were you tooconservative you know knowing what youknow now and and and what's the road mapfrom here?>> Yeah it's a great question because seethe the thing that we realized and I'mglad we did uh is that the concept ofbuilding a fleet that truly was funiblefungeible for all the parts of the lifecycle of AI funible across geographiesand fungeible across

00:43:01generations. Right?So because one of the key things is whenyou have let's take even uh what Jensenand team are doing right I mean they'reat a pace in fact one of the things Ilike is the speed of light right we nowhave GB300's bringing you know thatwe're bringing up so you don't want tohave ordered a bunch of GB200's that aregetting plugged in only to find thatGB2300s are in full production. So youkind of have to make sure you'recontinuously modernizing, you'respreading the fleet all over, you arereally truly funible by workload uh andyou're adding to that the softwareoptimizations we talked about. So to methat is the decision we made and we saidlook sometimes you may have to say no tosome of the demand including some of theopen AI demand right because sometimesyou know Sam may say hey we build me adedicated you know big you know whatevermulti- gigawatt data center in onelocation for training makes sense froman open AI perspective doesn't makesense from a long-term infrastructurebuildout for Azure and that's where Ithought they did the right thing to givethem flexibility to go procure that fromothers while m maintaining uh again asignificant book of business from openAAI but more importantly givingourselves the

00:44:16flexibility with othercustomers our own one P remember likeone of the things that we don't want todo is be short on uh is you know we talkabout Azure in fact some of times ourinvestors are overly fixated on theAzure number but remember for me thehigh margin business for me is co-pilotit is security co-pilot it's GitHubco-pilot it's the healthcare co-pilot Sowe want to make sure we have a balancedway to approach the returns that theinvestors have. And so that's kind ofone of the other misunderstood perhapsin our investor base in particular,which I find pretty strange and funnybecause I think they they want to holdMicrosoft because of the portfolio wehave. But man are they fixated on thegrowth number of one little thing calledAzure. On that point, Azure grew 39% inthe quarter on a staggering $93 billionrun rate. And you know, I think thatcompares to GCP that grew at 32% and AWScloser to 20%. But could Azure becauseyou did give compute to 1P and becauseyou did give compute to research, itsounds like Azure could have grown 4142% had you had more compute to offer.>> Absolutely. Absolutely. There's noquestion. There is no question. Sothat's why I think

00:45:32the internal thing isto balance out what we think again is inthe long-term interests of ourshareholders and uh and also to serveour customers well and also not to kindof you know one of the other things wasyou know people talk about concentrationrisk right we obviously want a lot ofopen AI but we also want other customerand so we're shaping the demand here youknow we are in a supply you know youknow we're not demand constraint we'resupply constraint so we are shaping thedemand such that it matches is thesupply in the optimal way with thelong-term uh view.>> To that point, Satcha, you you talkedabout 400 billion. It's incrediblenumber of remaining performanceobligations. Last night, you said that,you know, that's your booked businesstoday. It'll surely go up tomorrow assales continue to come in. And you saidyou're going to, you know, your need tobuild out capacity just to serve thatbacklog is very high. You know, howdiversified is that backlog to your toyour point? And how confident are youthat that 400 billion does turn intorevenue over the course of the nextcouple years?>> Yeah, that that 400 billion uh has avery short duration as Amy explained.It's the 2-year uh duration on average.So that's definitely uh our intent.

00:46:47That's one of the reasons why uh we'respending the capital outcllay with highcertainty that we just need to clearthis backlog. And to your point, it'spretty diversified both on the 1 P andthe 3P. our own demand is quite franklypretty high for our one first party uhand even amongst third party one of thethings we now are seeing is the the riseof all the other companies building realworkloads uh that are scaling uh and sogiven that I think we feel very good Imean obviously it's uh that's one of thebest things about RPO is you can beplanful quite frankly and so thereforewe feel very very good about buildingand then this doesn't include obviouslythe additional demand that we're alreadygoing to start seeing including the 250uh you know which will have a longerduration and we'll build accordingly>> right so there are a lot of new entranceright uh in this race to build outcompute Oracle coreweave cruso etc andnormally we think that will compete awaymargins but you've somehow managed tobuild all this out while maintaininghealthy operating margins at Azure so Iguess the question is for Microsoft howdo you compete in this world that is uhwhere people are levering up, takinglower margins while balancing thatprofit and and and

00:48:02risk. And do you seeany of those competitors doing dealsthat cause you to scratch your head andsay, "Oh, we're just setting ourselvesup for another boom and bust cycle.">> I mean, I'd say at some level the thegood news for us has been competing evenas a hyperscaler every day. You know,there's a lot of competition, right,between us and Amazon and Google on allof these, right? I mean it's sort of oneof those interesting things which iseverything is a commodity right computestorage I remember everybody saying wowhow can there be a margin except atscale nothing is a commodity um and sotherefore yes so we have to have ourcost structure our supply chainefficiency our software efficiencies allhave to kind of continue to compound inorder to make sure that there's marginsuh but scale and to your point one ofthe things that I really love about theOpenAI partnership is it's gotten us toscale, right? This is a scale game. Whenyou have uh the biggest workload thereis running on your cloud, that means notonly are we going to learn faster onwhat it means to operate with scale,that means your cost structure is goingto come down faster than anything else.And guess what? That'll make us pricecompetitive. And so I feel prettyconfident about our ability to, youknow, have margins.

00:49:18And and that this iswhere the portfolio helps. I've alwayssaid>> you know you know I've been forced intogiving the Azure numbers right becauseat some level I never thought ofallocating I mean my capital allocationis for the cloud from whether it is Xboxcloud gaming or Microsoft 365 or forAzure it's one capital outlay uh andthen everything is a meter as far as I'mconcerned from an MSF perspective it's aquestion of hey the blended average ofthat should match the operating marginswe need as a company because after allotherwise why we're not a conglomeratewe're one company with one platformlogic it's not running five sixdifferent businesses we're in these fivesix different businesses only tocompound the returns on the cloud and AIinvestment>> yeah I I love that line uh nothing is acommodity at scale you know there's beena lot of ink and time spent even on thispodcast with my partner Bill Gurleytalking about circular revenuesincluding including Microsoft Stashercredits right to OpenAI that were bookedas revenue. Do you see anything going onlike the AMD deal, you know, where theytraded 10% of their equity and, youknow, for a deal or the Nvidia

00:50:33deal?Again, I don't want to be overly fixatedon concern, but I do want to addressheadon what is uh being talked aboutevery day on CNBC and Bloomberg andthere are a lot of these overlappingdeals that are going on out there. Doyou do you when you think about that inthe context of Microsoft does any ofthat worry you again as to thesustainability or durability of uh theAI revenues that we see in the world?>> Yeah. I mean first of all our investmentof uh let's say that 13 and a half whichwas all the training investment that wasnot booked as revenue. That is the thatis the reason why we have the equitypercentage. That's the reason why wehave the 27% or 135 billion. So that wasnot something some that somehow thatmade it into Azure revenue. In fact, ifanything, the Azure revenue was purelythe consumption revenue of chat GPT andanything else and the APIs they put outthat they monetized and we monetized>> to your aspect of others. You know, tosome degree, it's always been there interms of vendor financing, right? Soit's not like a new concept that whensomeone's building something and theyhave a customer who is also buildingsomething but they need financing youknow for whether it

00:51:48is in you know it'sit's sort of some they're taking someexotic forms uh which obviously need tobe scrutinized by the investmentcommunity but that said you know vendorfinancing is not a new conceptinterestingly enough we have not had todo any of that right I mean we may haveyou know really uh either invested inOpenAI and essentially got an equity uhstake in it for return for compute oressentially sold them great pricing ofcompute in order to be able to sort ofbootstrap them. But you know otherschoose to do so differently and uh and Ithink circularity ultimately will betested by demand because all this willwork uh as long as there is demand forthe final out output of it and up to nowthat has been the case. Certainly,certainly. Well, I want to shift uh youknow, as you said, over half yourbusiness is software uh applications.You know, I want to think about softwareand agents. You know, last year on thispod, you made a bit of a stir by sayingthat much of application software, youknow, was this thin layer that sat saton top of a CRUD database. The notionthat business applications exist,that's probably where they'll allcollapse, right, in the agent

00:53:03era.Because if you think about it right,they are essentiallycrowd databases with a bunch of businesslogic.The business logic is all going to theseagents. Public software companies arenow trading at about 5.2 times forwardrevenue. So that's below their 10-yearaverage of seven times despite themarkets being at all-time highs. Andthere's lots of concern that SASsubscriptions and margins may be put atrisk by AI. So how today is AI affectingthe growth rates of your softwareproducts of you know those core productsand specifically as you think aboutdatabase fabric security office 360 andthen second question I guess is what areyou doing to make sure that software isnot disrupted but is insteadsuperpowered by AI? Yeah, I think that'sa Yeah, that's right. So, the last timewe talked about this, my my point reallythere was the architecture of SASapplications is changing because thisagent tier is replacing the old businesslogic tier. And so, because if you thinkabout it, the way we built SASapplications in the past was you had thedata, the logic tier, and the UI alltightly coupled. Uh, and AI

00:54:18quitefrankly doesn't respect that couplingbecause it requires you to be able todecouple. And yet the contextengineering is going to be veryimportant. I mean take you knowsomething like uh office 365. One of thethings I love about uh our Microsoft 365offering is it's low arpoouh high usage right I mean if you thinkabout it right outlook or teams orsharepoint you pick word or excel likepeople are using it all the timecreating lots and lots of data which isgoing into the graph and our arpoo islow. So that's sort of what gives mereal confidence that this AI tier with Ican meet it by exposing all my data. Infact, one of the fascinating thingsthat's happened uh Brad with both GitHuband Microsoft 365 is thanks to AI, weare seeing alltime highs in terms ofdata that's going into the graph or therepo.>> I mean think about it. The more codethat gets generated, whether it iscodeex or cloud or wherever, where is itgoing? GitHub, more PowerPoints that getcreated, Excel models that get created,all these artifacts and chatconversations. Chat conversations arenew docs, they're all going in to thegraph and and all that is needed

00:55:33again>> for grounding. Uh so that's what youknow you turn it into a forward indexinto an embedding and basically thatsemantics is what you really go groundany agent request. And so I think thenext generation of SAS applications willhave to sort of if you are high RPO lowusage then you have a little bit of aproblem. But if you are we are the exactopposite. we are low RPO, high usage andI think that anyone who can structurethat and then use this AI as in fact anaccelerant because I mean like if youlook at the M365 copilot price I meanit's higher than any other thing that wesell and yet it's getting deployedfaster and with more usage and so I feelvery good oh or coding right who wouldhave thought in fact take GitHub rightwhat GitHub did in first I don't know 15years of its existence or 10 years ofits existence it was basically done inthe last year just because coding is nolonger a tool. It's more a substitutefor wages and so it's a very differenttype of business model even kind ofthinking about the stack and where valuegets distributed. So until veryrecently, right, clouds largely ranpre-ompiled software. You didn't need alot of GPUs

00:56:49and most of the valueacrewed to the software layer to thedatabase to the applications like CRMand Excel. But it does seem in thefuture that these interfaces will onlybe valuable, right? If they're ifthey're uh intelligent, right? Ifthey're pre-ompiled, they're kind ofdumb. The software's got to be able tothink and to act and to advise. And thatrequires you know the production ofthese tokens you know dealing with theeverchanging context. And so in thatworld it does seem like much more of thevalue will acrue to the AI factory ifyou will to you know Jensen producingyou know uh helping to produce thesetokens at uh uh the lowest cost and tothe models and maybe that the agents orthe software will acrue a little bitless of the value in the future thanthey've accured in the in the past.Well, steelman for me. Why that's wrong?>> Yeah. So, I think there are two thingsthat are necessary to try and to drivethe value of AI. One is what youdescribed first, which is the tokenfactory. And even [clears throat] if youunpack the token factory, uh it's thehardware silicon system, but then it isabout running it most efficiently withthe system software with all thefungibility, max utilization. That'swhere the hyperscaler's

00:58:04role is, right?What is a hyperscaler? Is hyperscalerlike everybody says if you sort of saidhey I want to run a hyperscaler. Yeahyou could say oh it's simple. I'll buy abunch of servers and wire them up andrun it. It's not that right. I mean itwas that simple then there would havebeen more than three hyperscalers bynow. So the hyperscaler is the knowhowof running that max util and the tokenfactories. And it's not and by the wayit's going to be heterogeneous.Obviously Jensen's super competitive.Lisa is going to come, you know, Hawk'sgoing to produce things uh fromBroadcom. We will all do our own. Sothere's going to be a combination. Soyou want to run ultimately aheterogeneous fleet that is maximizedfor token throughput and efficiency andso on. So that's kind of one job. Thenext thing is what I call the agentfactory. Remember that a SAS applicationin the modern world is driving abusiness outcome. it knows how to mostefficiently use the tokens to createsome business value. Uh in fact, GitHubcopilot is a great example of it, right?Which is, you know, if you think aboutit, it the auto mode of GitHub copilotis the smartest thing we've done, right?So, it chooses based on the prompt whichmodel to use for a code completion or atask handoff,

00:59:19right? That's what you andyou do that not just by, you know,choosing in some roundrobin fashion. Youdo it because of the feedback cycle. Youhave you have the eval, the data loopsand so on. So the new SAS applicationsas you rightfully said are intelligentapplications that are optimized for aset of evals and a set of outcomes thatthen know how to use the token facto'soutput most efficiently. Sometimeslatency matters, sometimes uhperformance matters and knowing how todo that trade uh in a smart way is wherethe SAS application value is. Butoverall it is going to be true thatthere is a real marginal cost tosoftware this time around. It was therein the cloud era too when we were doingyou know CDROMs there wasn't much of amarginal cost you know with the cloudthere was and this time around it's alot more and so therefore the businessmodels have to adjust and you have to dothese optimizations for the agentfactory and the token factoryseparately. you have a big searchbusiness that most people don't knowabout, you know, but it turns out thatthat's probably one of the mostprofitable businesses in the history ofthe world because people are runninglots of searches, billions of searches,and the cost of completing a search ifyou're Microsoft is many

01:00:34fractions of apenny, right? Doesn't cost very much tocomplete a search, but the comparablequery or prompt stack today when whenyou use a chatbot looks different,right? So I guess the question isassume similar levels of revenue in thefuture for those two businesses, right?Do you ever get to a point where kind ofthat chat interaction has unit economicsthat are as profitable as search? Ithink that's a great point because seesearch was pretty magical uh in terms ofits ad unit uh and its cost economicsbecause there was the index which was afixed cost that you could then amortizein a much more efficient way>> uh whereas this one you know each chatuh to your point you have to burn a lotmore GPU cycles uh both with the intentand the retrieval so the economics aredifferent so I think you do that's why Ithink a lot of the early sort ofeconomics of chat have been the premiummodel and subscription on the even onthe consumer side. So we are yet todiscover whether it's agentic commerceor whatever is the ad unit how it'sgoing to be litigated but at the sametime the fact that at this point youknow I kind of know in fact

01:01:50I use searchuh for very very specific navigationalqueries I used to say I use it a lot forcommerce but that's also shifting to myyou know co-pilot like I look at theco-pilot mode in edge and bing uh orcopilot now they're blend ending in. SoI think that yes, I think there is goingto be a relitigation just like that wetalked about the SAS disruption. We'rein the beginning of the cheese being alittle moved in consumer economics ofthat category,>> right? I I mean and given that it's themulti-trillion dollar this this is thething that's driven all the economics ofthe internet, right? when you move theeconomics of search for both you andGoogle and it converges on somethingthat looks more like a personal agent, apersonal assistant chat. Um, you know,that could end up being much much biggerin terms of the total value delivered tohumanity, but the unit economics, you'renot just advertising this one time fixedindex.>> That's right.>> And so, that's right. I think that theconsumer could be worse. Yeah. theconsumer category because you arepulling a thread on something that Ithink a lot about right which is whatduring these disruptionsyou you kind of have to have a realsense of where is is the what is thecategory economics uh is it winner

01:03:05takeall um uh and both matter uh right thethe problem on consumer space always isthat there's finite amount of time uhand so if I'm not doing one thing uh I'mdoing something else and if yourmonetization is predicated on some humaninteraction in particular if there wastruly agentic stuff even on consumerthat could be different. Uh whereas inthe enterprise one is it's not winnertake all and two it is going to be a lotmore friendly for agentic interaction.So it's not like for example the perseat versus consumption. The reality isagents are the new seats.>> And so you can think of it as uh theenterprise monetization is much clearer.The consumer monetization I think is alittle more murky. You know, we've seena spade of layoffs recently with Amazonannouncing some big big layoffs thisweek. You know, the Mag 7 has had littlejob growth over the last three yearsdespite really robust top lines. Youknow, you didn't grow your headcountreally from 24 to 25. It's around225,000.You know, many attribute this to normalgetting fit, you know, just getting moreefficient coming out of co and I thinkthere's a lot of truth to that. But doyou think part of this is due to AI?

01:04:20Doyou think that AI is going to be a netjob creator? And do you see this being along-term positive for Microsoftproductivity? Like it feels to me likethe pie grows, but you can do all thesethings much more efficiently, whicheither means you your margins expand orit means you reinvest those margindollars and you grow faster for longer.I call it the golden age of marginexpansion. I'm a firm believer that thethe productivity curve does uh and willbend in the sense that we will startseeing some of what is the work and theworkflow in particular change, right?there's going to be more agency for youat a task level to get to job completebecause of the power of these tools uhin your hand and that I think is goingto be the case. So that's why I think weare even internally for example when youtalked about even our allocation oftokens we want to make sure thateverybody at Microsoft standard issueright all of them have Microsoft 365 tothe tilt in the sort of most un uhlimited way and have GitHub copilot sothat they can really be more productivebut here

01:05:35is the other interesting thingBrad we're learning is there is a newway to even learn right which is youknow how to work with agents Right? Sothat's kind of like when the first whenword, excel, powerpoint all showed up inoffice, you kind of we learned how torethink let's say how we did a forecast,right? Right? I mean, think about it,right? In the 80s, the forecasts wereinter office memos and faxes and whathave you. And then suddenly somebodysaid, "Oh, here's an Excel spreadsheet.Let's put it an email. Send it around.People enter numbers and there was aforecast.">> Similarly, right now, any planning, anyexecution starts with AI. You researchwith AI. You think with AI, you sharewith your colleagues and what have you.So, there's a new artifact being createdand a new workflow being created. Andthat is the rate of the pace of changeof the business process that matches thecapability of AI. That's where theproductivity efficiencies come. And soorganizations that can master that aregoing to be the biggest beneficiarieswhether it's in our industry or quitefrankly in the real world.>> And so is Microsoft benefiting fromthat? You know, so let's let's thinkabout a couple years from now. Fiveyears from now at the current

01:06:50growthrate will be sooner, but let's call itfive years from now, your top line istwice as big as what it is today.Satcha, how many more employees will youhave if you're if you're if you growrevenue by>> like one of the best things right now isthese examples that I'm hit with everyday from the employees of Microsoft.There was this person who leads ournetwork operations, right? I mean if youthink about the amount of uh fiber wehave had to put uh for like this youknow this 2 gawatt data center we justbuilt out uh in fair water right and theamount of fiber there the AI and whathave you it's just crazy right so>> and it turns out this is a real worldasset there are I think 400 differentfiber operators we're dealing withworldwide every time something happenswe are literally going and dealing withall these DevOps pipelines the personwho leads it she basically said to meyou what I there's no way I'll ever getthe headcount to go do all this. Notforget even if I even approve thebudget. I can't hire all these folks. Soshe she did the next best thing. Shejust built herself a whole bunch ofagents to automate the DevOps pipelineof how to deal with the maintenance.That is an example of you to your pointa team>> with AI tools being able to get moreproductivity. So in if you are questionI will say we

01:08:05will grow a headcount butthe way I look at it is that headcountwe grow will grow with a lot moreleverage than the headcount we had preAI>> and that's the adjustment I thinkstructurally you're seeing first rightwhich is one you called it getting fit Ithink of it as more getting to a placewhere everybody is really not learninghow to rethink how they work and it'sthe how not even the what even If thewhat remains the constant, how you goabout it has to be relearned. And it'sthe unlearning and learning process thatI think will take the next year or so,then the headcount growth will come withmax leverage.Yeah. No, it's a I think we're on theverge of incredible economicproductivity growth. It does feel likewhen I talk to you or Michael Dell thatmost companies aren't even really in thefirst inning, maybe the the first batterin the first inning in reworking thoseworkflows to get maximum leverage fromthese agents. But it sure feels likeover the course of the next two to threeyears, that's where a lot of gains aregoing to start coming from. And again, Iyou know, I I certainly am an optimist.I think we're going to have net jobgains from all of this. But I think forthose companies, they'll just be able togrow their bottom line,

01:09:20their number ofemployees slower than their top line.That is the productivity gain to thecompany. Aggregate all that up. That'sthe productivity gain to the economy.And then we'll just take that consumersurplus and invest it in creating a lotof things that didn't exist before.>> 100%. 100%. Even in softwaredevelopment, right? One of the things Ilook at it is no one would say we'regoing to have a challenge in having, youknow, more software engineers contributeto our sort of society because thereality is you look at the IT backlog inany organization. And so the question isall these software agents are hopefullygoing to help us go and take a whack atall of the IT backlog we have and thinkof that dream of evergreen software.That's going to be true. and then thinkabout the demand for software. So Ithink that to your point it's the levelsof abstraction at which knowledge workhappens will change. We will adjust tothat the work and the workflow that willthen adjust itself even in terms of thedemand for the products of thisindustry.>> I'm going to end on this which is reallyaround the reindustrialization ofAmerica. I've said if you add up the $4trillion of capex that you and these andand so many of of the big large US techcompanies are investing over

01:10:35the courseof the next four or five years, it'sabout 10 times the size of the Manhattanproject on an inflation adjusted or GDGDP adjusted basis. So it's a massiveundertaking for America. The presidenthas made it a real priority of hisadministration to recut the trade dealsand it looks like we now have trillionsof dollars. South Koreans committed $350billion dollars of investments uh justtoday into the United States. And whenyou think about, you know, what you seegoing on in power in the United States,both production, the grid, etc., whatyou see going on in terms of thisre-industrialization,how how do you think this is all going?and uh you you know maybe just reflecton where we're landing the plane hereand your level of optimism for the thethe few years ahead.>> Yeah. No, I I I feel very veryoptimistic because in some sense, youknow, Brad Smith was telling me aboutsort of the economy around a Wisconsindata center. It's fascinating. Mostpeople think, oh, data center that issort of like, yeah, uh it's going to beone big warehouse and there's, you know,fully automated. Uh a lot of it is true.uh but first of all what went into theconstruction uh of that data center andthe local supply chain

01:11:51of the datacenter uh that is in some sense thereindustrialization of the United Statesas well uh even before you get to whatis happening in Arizona with the TSMCplants or what was happening with Micronand their investments in memory or Inteland their fabs and what have you rightthere's a lot of stuff that we will wantto start building doesn't mean we won'thave trade deals that make sense for theUnited States with other countries. Butto your point, the reindustrializationfor the new economy and take making surethat all the skills and all thatcapacity from power on down, I think issort of very important right for us. Andin and the other thing that I also say,Brad, it's important and this issomething that I've had a chance to talkto President Trump as well as uhSecretary Lutnik and others is it'simportant to recognize that we ashyperscalers of the United States arealso investing around the world. So inother words, the United States is thebiggest investor of compute factories ortoken factories uh around the world. Butnot only are we attracting foreigncapital to invest in our country so thatwe can re-industrialize, we are helpingwhether it's in Europe or in Asia

01:13:06orelsewhere in Latin America and in Africawith our capital investments, bringingthe best American tech uh to the worldthat they can then innovate on andtrust. And so both of those I think arereally bode well for the United Stateslong term.>> I'm grateful for your leadership Sam isis is really helping lead the charge atopen AI for America. I think this is amoment where I look ahead, you know, youcan see 4% GDP growth on the horizon.We'll have our challenges. We'll haveour ups and downs. These tend to bestairs, you know, stairs up rather thana line straight up and to the right. ButI for one see a level of coordinationgoing on between Washington and SiliconValley between big tech and there-industrialization of America thatgives me cause for incredible hope.Watching what happened this week in Asiauh led by the president and his team andthen watching what's happening here uhis is super exciting. So thanks formaking the time. We're big fans. Thanks.Thanks Satcha.>> Thanks so much Brad. Thank you.[music]

01:14:16As a reminder to everybody, just ouropinions, not investment advice.