Satya Nadella – How Microsoft thinks about AGI (YouTube Video Transcript)

(00:00:00) Your YouTube transcript will appear here (00:00:00) Maybe after the industrial revolution, (00:00:02) this is the biggest thing. But at the (00:00:04) same time, I'm a little grounded in the (00:00:06) fact that this is still early innings. (00:00:08) If you're a model company, you may have (00:00:10) a winner's curse. You may have done all (00:00:12) the hard work, done unbelievable (00:00:14) innovation, except it's kind of like one (00:00:17) copy [music] away from that being (00:00:19) commoditized. We didn't want to just be (00:00:22) a host for one company and have just a (00:00:26) massive book of business with one (00:00:27) customer. [music] That that's not a (00:00:28) business. You can't build an (00:00:30) infrastructure that's optimized for one (00:00:31) model. If you did that, you're one tweak (00:00:33) away. Some like breakthrough that (00:00:35) [music] happens and your entire network (00:00:37) topology goes out of the window. Then (00:00:39) that's a scary thing. Our business, (00:00:41) which today is an enduser tools (00:00:43) business, will become essentially an (00:00:45) [music] infrastructure business in (00:00:46) support of agents doing work. The thing (00:00:48) that you have to think [music] through (00:00:49) is not what you do in the next 5 years, (00:00:51) but what do you do for the next 50. (00:00:54) >> Today we are interviewing Satya Nadella. (00:00:57) We being me and Dylan Patel who is (00:01:00) founder of semi analysis. Satya, (00:01:02) welcome. (00:01:02) >> Thank you. It's great. Thanks for coming (00:01:04) over at Atlanta. (00:01:05) >> Yeah, thank you for giving us a tour of (00:01:06) uh the new facility. It's been really (00:01:08) cool to see. (00:01:08) >> Absolutely. (00:01:10) >> Satya and Scott Guthrie, Microsoft's EVP (00:01:13) [music] of cloud and AI, give us a tour (00:01:14) of their brand new Fairwater 2 data (00:01:16) center, the current most powerful in the (00:01:18) world. (00:01:19) >> We try to 10x the training [music] (00:01:21) capacity every 18 to 24 months. And so (00:01:24) this would be effectively a 10x increase (00:01:26) 10x from what GPD5 was trained with. And (00:01:28) so to put in perspective, the number of (00:01:30) optics, the network optics in this (00:01:32) building is almost as much as all of (00:01:36) Azure across all our data centers 2 and (00:01:38) a half years ago. (00:01:38) >> It's kind of what 5 million network (00:01:41) connections. (00:01:42) >> You've got all this bandwidth between (00:01:43) different sites in a region and between (00:01:45) the two regions. So is this like a big (00:01:47) bet on scaling in the future that you (00:01:48) anticipate in the future there's going (00:01:50) to be some huge model that needs to (00:01:51) require two whole different regions to (00:01:53) train (00:01:54) >> the goal is to be able to kind of (00:01:56) aggregate these flops for a large (00:01:58) training job and then put these things (00:02:00) together across size (00:02:02) >> right (00:02:02) >> and the reality is you'll use it for uh (00:02:07) training and then you'll use it for data (00:02:09) gen you'll use it for inference in all (00:02:11) sort of ways it's not like it's going to (00:02:13) be used only for one workload forever (00:02:15) >> water four which you're going to see (00:02:16) under construction nearby. (00:02:18) >> Yeah. We'll also be on that one pedits (00:02:20) [music] network. (00:02:21) >> Yep. (00:02:21) >> So that we can actually link the two at (00:02:23) a very high rate. And then basically we (00:02:25) do the IWAN connecting to Milwaukee (00:02:27) where we have multiple other fair waters (00:02:29) being built. (00:02:30) >> Literally you can see the the model (00:02:33) parallelism and the data parallelism. (00:02:35) It's kind of built for um essentially (00:02:39) the training jobs, the pods, the super (00:02:41) pods across this campus. And then with (00:02:46) the van, you can go to the Wisconsin (00:02:48) data center and literally run a training (00:02:52) job with all of them getting aggregated. (00:02:54) >> And what we're seeing right here is this (00:02:56) is a cell with no servers in it yet. No (00:02:58) racks. (00:02:59) >> How many uh racks are in a cell? (00:03:01) >> Let me think about it. We don't (00:03:02) necessarily share that per se, but but (00:03:04) we let me (00:03:05) >> The reason I asked [laughter] (00:03:07) you'll see upstairs you can start (00:03:10) counting. We'll let you start counting. (00:03:11) How many cells are there in this (00:03:11) building? (00:03:12) >> That part also I can't tell you. Well, (00:03:13) division is easy, right? (00:03:18) >> My god, it's kind of loud. (00:03:21) >> Are you looking at this like now I see (00:03:23) where my money is going? [laughter] (00:03:25) >> It's kind of like I run a software (00:03:27) company. Welcome to the software (00:03:28) company. (00:03:30) >> How big is the design space once you've (00:03:31) decided to use GB200's and NVIL? How (00:03:34) many other decisions are there to be (00:03:35) made? There is coupling from the model (00:03:39) architecture (00:03:40) to what is the physical plan that's (00:03:43) optimized (00:03:44) >> and it's also scary in that sense which (00:03:47) is hey there's going to be a new chip (00:03:49) that will come out which obviously I (00:03:50) mean you take Vera Rubin ultra I mean (00:03:53) that's going to have power density (00:03:54) that's going to be so different with (00:03:56) cooling requirements that are going to (00:03:57) be so different right so you kind of (00:04:00) don't want to just build all to one spec (00:04:04) so that goes back a little to I think (00:04:06) the dialogue we'll have which is (00:04:08) >> you want to be scaling in time (00:04:12) >> as opposed to scale once and then be (00:04:14) stuck with it. (00:04:15) >> When you look at all the past (00:04:16) technological transitions whether it be (00:04:19) you know railroads or the internet or (00:04:22) you know replaceable parts and (00:04:23) translization uh the cloud all of these (00:04:26) things each revolution has gotten much (00:04:29) faster in the time it goes from (00:04:30) technology discover to ramp and (00:04:32) pervasiveness through the economy. Many (00:04:34) folks who have been on Darkesh's podcast (00:04:36) believe this is sort of the final uh (00:04:39) technological revolution or transition (00:04:41) and this time is very very different. Um (00:04:43) and at least so far in the markets it's (00:04:45) sort of you know in three years we've (00:04:47) already skyrocketed to you know (00:04:48) hyperscalers are doing $500 billion of (00:04:50) capex next year which is a scale that's (00:04:53) un unmatched to prior revolutions in (00:04:55) terms of speed and the end state seems (00:04:58) to be quite different. How how do you (00:05:00) your your framing of this seems quite (00:05:02) different than sort of the I would say (00:05:04) the AI bro who is who is quite um you (00:05:07) know AGI is coming and you know I'd like (00:05:10) to understand that more. (00:05:11) >> Yeah. I mean look I I I start with the (00:05:14) excitement that I also feel for maybe (00:05:17) after the industrial revolution this is (00:05:19) the biggest thing. Um and so therefore I (00:05:22) I I I I start with that premise. uh but (00:05:25) at the same time I'm a little grounded (00:05:27) in the fact that uh this is still early (00:05:29) innings. Uh we've built some very useful (00:05:32) things. We're seeing some great (00:05:33) properties. These scaling laws seem to (00:05:35) be working. Um and I'm optimistic that (00:05:39) they'll continue to work right. Some of (00:05:41) it is um you know it does require real (00:05:44) science breakthroughs but it's also a (00:05:46) lot of engineering and what have you. (00:05:48) But that said, I also sort of take the (00:05:51) view that you know even what has been (00:05:53) happening in the last 70 years of (00:05:55) computing uh has also been a march uh (00:05:58) that has helped us move um you know with (00:06:03) as I said you know I like one of the (00:06:05) things that Raj ready has as a metaphor (00:06:08) for what AI is right he's a he's a (00:06:10) turing award winner out of CMU um and (00:06:14) he's always I think he had this even (00:06:16) pre- AGI but he had this metaphor of AI (00:06:19) should either be a guardian angel or a (00:06:21) cognitive amplifier. I love that uh it's (00:06:24) a simple way to think about what this is (00:06:27) ultimately what is its human utility? It (00:06:30) is going to be a cognitive amplifier uh (00:06:32) and a guardian angel. And so if I sort (00:06:35) of view it that way, I view it as a (00:06:37) tool. But then you can also go very (00:06:39) mystical about it and say, "Wow, this is (00:06:41) you know more than a tool. It does all (00:06:42) these things which only humans did so (00:06:44) far." But that has been the case with (00:06:46) many technologies in the past. only (00:06:47) humans did a lot of things and then we (00:06:49) add tools that did them. (00:06:51) >> I guess we don't have to get wrapped up (00:06:53) in their definition here, but maybe one (00:06:55) way to think about it is like maybe it (00:06:56) takes 5 years, 10 years, 20 years. At (00:06:58) some point eventually a machine is (00:07:00) producing Satya tokens, right? And the (00:07:02) Microsoft board thinks that Satia tokens (00:07:04) are worth a lot. (00:07:04) >> How much how much are you wasting of his (00:07:06) [laughter] of like economic value by (00:07:08) interviewing Satya? (00:07:10) >> You cannot afford the API cost of Satia (00:07:11) tokens. Um but so you know whatever you (00:07:15) want to call it is that are the SATA (00:07:16) tokens a tool or an agent whatever. Um (00:07:19) right now if you have models that cost (00:07:21) on the order of dollars or cents per (00:07:22) million tokens there's just an enormous (00:07:24) room for expansion uh a margin expansion (00:07:27) there where sad a million tokens of SA (00:07:30) are like worth a lot um and where does (00:07:33) that margin go and what level of that (00:07:36) margin is Microsoft involved in is a (00:07:38) question I have. So I think um in in (00:07:42) some sense this goes back a band to (00:07:44) essentially what's the economic growth (00:07:46) picture going to really look like? Um (00:07:49) what's the firm going to look like? (00:07:51) What's productivity going to look like? (00:07:52) And that to me is where again if the (00:07:54) industrial revolution created after (00:07:57) whatever 70 years of diffusion is when (00:07:59) you started seeing the economic growth, (00:08:01) right? it took that's the other thing to (00:08:03) remember is um even if the tech is (00:08:06) diffusing fast uh this time around for (00:08:10) true economic growth to appear it has to (00:08:13) sort of diffuse to a point where the (00:08:15) work the work artifact and the workflow (00:08:17) has to change and so that's kind of one (00:08:18) place where I think uh the change (00:08:21) management required for a corporation to (00:08:23) truly change I think is something we (00:08:25) shouldn't discount so I think going (00:08:27) forward do humans and the tokens they (00:08:30) produce get higher leverage, right? Uh (00:08:34) whether it's the Dark Cesh or the Dylan (00:08:36) tokens of the future. I mean, think (00:08:38) about the amount of techn would you be (00:08:40) able to run semi analysis or this (00:08:42) podcast without technology? No chance, (00:08:44) right? I mean, the scale that you have (00:08:47) been able to achieve, no chance. So, the (00:08:49) question is what's that scale? Is it (00:08:51) going to be 10xed with something that (00:08:53) comes through? Uh absolutely. uh and (00:08:55) therefore within your ramp to some (00:08:57) revenue number or your ramp to some (00:08:59) audience number or what have you and so (00:09:01) that I think is what's going to happen (00:09:03) right I mean the the point is uh that's (00:09:06) whatever what took 70 years maybe 150 (00:09:09) years for the industrial revolution may (00:09:11) happen in 20 years 25 years that's a (00:09:13) better way to like I would love to (00:09:16) compress what happened in 200 years of (00:09:18) the industrial revolution into 20-year (00:09:20) period if you're lucky (00:09:23) >> so Microsoft historically has been (00:09:25) perhaps you know the greatest software (00:09:27) company the largest software as a (00:09:29) service company you know you've gone (00:09:30) through a transition in the past where (00:09:32) you used to sell Windows licenses and (00:09:34) discs of Windows or Microsoft and now (00:09:36) you sell you know subscriptions to 365 (00:09:39) or um as as we go from sort of you know (00:09:43) that transition to where your business (00:09:45) is today um there's also a transition (00:09:47) going after that right uh software as a (00:09:49) service incredibly low incremental cost (00:09:52) per user uh there's a lot of R&D there's (00:09:54) a lot customer acquisition cost. This is (00:09:56) why not Microsoft but the SAS companies (00:09:58) have underperformed massively in the (00:10:00) markets because the cogs of AI is just (00:10:02) so high and that just completely breaks (00:10:04) how these business models work. (00:10:06) >> How do you as a as as a as perhaps the (00:10:09) greatest software company um software as (00:10:11) a service company transition Microsoft (00:10:14) to this new age where COGS matters a lot (00:10:17) um and and the incremental cost per (00:10:19) users is different right because right (00:10:21) now you're charging hey it's 20 bucks (00:10:22) for a co-pilot. (00:10:23) >> Yeah. So I think that this is a it's a (00:10:25) great question because in some sense the (00:10:27) business models themselves I think the (00:10:29) levers are going to remain similar right (00:10:31) which is if I look at the the if if you (00:10:33) look at the menu of models uh starting (00:10:36) from like say consumer all the way right (00:10:39) there will be some ad unit uh there will (00:10:41) be some transaction there will be some (00:10:43) device gross margin for somebody who (00:10:45) builds an AI device um uh there will be (00:10:48) subscriptions consumer and enterprise uh (00:10:51) and then there'll be consumption right (00:10:53) so I still think that that's kind of how (00:10:55) those are all the meters. To your point, (00:10:58) what is a subscription? Up to now, (00:11:00) people like subscriptions because they (00:11:02) can budget for them, right? They are (00:11:05) essentially entitlements to some (00:11:07) consumption rights that come (00:11:09) encapsulated in a subscription. So that (00:11:11) I think is what in some sense it becomes (00:11:13) a pricing decision. Uh so how much (00:11:16) consumption is in you are entitled to is (00:11:19) if you look at all the coding (00:11:21) subscriptions that's kind of what they (00:11:22) are, right? and they kind of have the (00:11:24) pro tier, the standard tier and what (00:11:26) have you. And so I think that's how the (00:11:29) pricing will uh you know and the margin (00:11:31) structures will get tiered. Um the (00:11:34) interesting thing is having at Microsoft (00:11:37) the good news for us is we kind of are (00:11:39) in that business uh all in across all (00:11:42) those meters. in fact that at a as a (00:11:44) portfolio level uh we pretty much have (00:11:48) consumption subscriptions (00:11:50) uh to all of the other consumer levers (00:11:52) as well. Um and then I think time will (00:11:55) tell which of these models make sense in (00:11:58) what categories. Um, one thing on the (00:12:01) SAS side since you brought up which I (00:12:03) think a lot about is uh take uh Office (00:12:05) 365 or Microsoft 365. I mean man having (00:12:08) a low RPO is great because here here's (00:12:10) an interesting thing right during the (00:12:12) transition from server to cloud one of (00:12:15) the questions we used to ask ourselves (00:12:16) is oh my god if all we did was just (00:12:19) basically move the same users who were (00:12:21) using let's call it our office licenses (00:12:24) and our servers at that time office (00:12:26) servers right to the cloud and we had (00:12:29) cogs this is going to basically not only (00:12:31) shrink our margins uh but we'll be (00:12:34) fundamentally a nonprofitable I mean (00:12:35) less profitable company except What (00:12:38) happened was the move to the cloud (00:12:40) expanded the market like crazy. Uh right (00:12:43) I mean we sold a few servers in India (00:12:45) didn't sell much whereas in the cloud (00:12:47) suddenly everybody in India also could (00:12:49) afford fractionally buying uh servers. (00:12:52) The IT costs I in fact the biggest thing (00:12:54) I had not realized for example was the (00:12:57) amount of money people were spending (00:13:00) buying storage underneath SharePoint. In (00:13:03) fact, EMC's biggest segment may have (00:13:06) been storage servers for SharePoint. All (00:13:10) that sort of dropped in the cloud (00:13:12) because nobody had to go buy in fact it (00:13:14) was working capital. I mean basically it (00:13:16) is cash flow out right and so it (00:13:19) expanded the market massively. So this (00:13:22) AI thing will be that right. So if you (00:13:25) take coding (00:13:26) um what we built with GitHub and VS code (00:13:30) in over whatever decades uh suddenly the (00:13:34) coding assistant is that big in one year (00:13:37) and so that I think is what's going to (00:13:39) happen as well which is the market (00:13:40) expands massively. M I I guess there's a (00:13:43) question of the market will expand. Will (00:13:45) the parts of the revenue that touch (00:13:47) Microsoft expand? So copilot is an (00:13:50) example where if you look uh early this (00:13:52) year I think uh I guess according to (00:13:55) Dylan's numbers um the co-pilot revenue (00:13:58) github co-pilot revenue was like 500 (00:14:00) million or something like that and then (00:14:02) u there were like no close competitors (00:14:04) whereas now you have claude code cursor (00:14:07) and copilot with around similar revenue (00:14:09) around a billion and then codeex is (00:14:11) catching up around 700 800 million and (00:14:13) so the question is across all the (00:14:15) services that Microsoft has access to (00:14:17) what is the advantage that mic (00:14:18) Microsoft's equivalents of Copilot have. (00:14:20) >> Yeah, by the way, I love this chart. You (00:14:22) know, I love this chart for so many (00:14:24) reasons. One is we're still on the top. (00:14:26) [laughter] (00:14:27) >> Um, second is all these companies that (00:14:30) are listed here are all companies that (00:14:32) have been born in the last four or five (00:14:34) years. (00:14:35) >> Yeah, (00:14:35) >> that to me is the best sign, right? (00:14:37) Which is if you have new competitors, (00:14:38) new existential problems, when you say, (00:14:40) man, who's it now? Claude's going to (00:14:42) kill you. Cursor is going to kill you. (00:14:44) It's not Borland, right? So, thank God. (00:14:46) like that means we are in the right (00:14:48) direction but this is it right the fact (00:14:50) that we went from nothing to this scale (00:14:54) is the market expansion so this is like (00:14:56) the cloud-like stuff this fundamentally (00:14:58) this category of coding and AI is (00:15:02) probably going to be one of the biggest (00:15:03) categories right it is a software (00:15:05) factory category in fact it may be (00:15:07) bigger than knowledge work (00:15:09) >> so I kind of want to keep myself (00:15:10) open-minded about I mean we're going to (00:15:12) have tough competition I think that's (00:15:13) your point which I think is a great on (00:15:16) uh but man like I'm glad we have we (00:15:19) parlayed uh what we had into this and (00:15:23) now we have to compete and so in the (00:15:25) compete side uh even in the last quarter (00:15:28) we just we did our quarterly uh (00:15:30) announcement I think we grew from 20 to (00:15:32) 26 million subs right so I feel good (00:15:33) about our sub growth uh and where the (00:15:36) direction of travel on that is but the (00:15:38) more interesting thing that has happened (00:15:40) is guess where all the repos of all (00:15:43) these other guys uh who are generating (00:15:46) lots and lots of code go to they go to (00:15:48) GitHub so it GitHub is at an all-time (00:15:51) high in terms of repo creation PRs (00:15:54) everything so that in some sense we want (00:15:57) to keep that open by the way that means (00:15:59) we want to have that right because we (00:16:01) don't want to conflate that with our own (00:16:02) growth right the interestingly enough (00:16:04) we're getting one developer joining (00:16:06) GitHub a second or something that is the (00:16:08) stat I think and then 80% of them just (00:16:10) fall into some GitHub copilot uh (00:16:12) workflow just because there are and by (00:16:14) the way many of these things will even (00:16:16) use some of our coding code review (00:16:18) agents which are by default on just (00:16:20) because you can use it. So we'll have (00:16:22) many many structural shots at this. The (00:16:25) thing that we're also going to do is (00:16:27) what we did with git g get the (00:16:29) primitives of github whether starting (00:16:31) with git to issues to actions these are (00:16:36) powerful lovely things because they kind (00:16:38) of are all built around your repo. So we (00:16:42) want to extend that last week at GitHub (00:16:44) Universe. That's kind of what we did, (00:16:45) right? So we said agent HQ was the (00:16:49) conceptual thing that we said we're (00:16:50) going to build out. This is where for (00:16:52) example you have a thing called mission (00:16:54) control and you go to mission control (00:16:56) and now I can fire off sometimes I (00:16:59) describe it as the cable TV of all these (00:17:01) AI agents because I'll have essentially (00:17:02) packaged into one subscription (00:17:05) codeex claude um you know cognition (00:17:10) staff anyone's agents gro all of them (00:17:12) will be there so I get one package and (00:17:15) then I can literally go issue a task (00:17:19) steer them so they'll all be working in (00:17:21) their independent branches. Uh I can (00:17:23) monitor them. Uh so I literally have (00:17:26) because I think that's going to be one (00:17:27) of the biggest places of innovation, (00:17:29) right? Because right now I want to be (00:17:31) able to use multiple agents. I want to (00:17:33) be able to then digest the output of the (00:17:35) multiple agents. I want to be able to (00:17:36) then keep a h a handle on my repo. So if (00:17:39) there's some some kind of a heads up (00:17:40) display that needs to be built and then (00:17:43) for me to quickly steer and triage what (00:17:45) the coding agents have generated that to (00:17:48) me between VS code GitHub and all of (00:17:51) these new primitives we'll build uh as (00:17:54) mission control I think uh with a (00:17:56) control plane observability I mean think (00:17:58) about every uh one who is going to (00:18:00) deploy all this will require a whole (00:18:02) host of observability of what agent did (00:18:04) what at what time to what code base so I (00:18:07) feel that's the opport opportunity uh (00:18:10) and at the end of the day your point is (00:18:11) well taken which is we better be (00:18:13) competitive and innovate and if we don't (00:18:15) yes we will get toppled but I like the (00:18:17) chart at least as long as we're on the (00:18:19) top even with competition (00:18:20) >> the key point here is sort of that (00:18:21) GitHub will keep growing irregardless of (00:18:24) whose coding agent wins but that that (00:18:26) market only grows at you know call it 10 (00:18:28) 15 20% which is way above GDP it's a (00:18:31) great compounder but these AI coding (00:18:33) agents have grown from you know call it (00:18:35) $500 million run rate at the end of last (00:18:37) year which was basically ally just (00:18:38) GitHub copilot to now the current run (00:18:41) rate across you know GitHub copilot (00:18:43) cloud code cursor cognition wind surflet (00:18:47) uh codeex open codeex that's that's (00:18:49) that's run rating at 56 billion now um (00:18:53) for the for the Q4 of of this year (00:18:55) that's a 10x right and and when you look (00:18:57) at hey what's the TAM of of software (00:18:59) agents is it is it the $2 trillion of (00:19:01) wages you pay people or is it is it is (00:19:04) it something beyond that uh because (00:19:07) every company in the will now be able (00:19:08) to, you know, develop software more. (00:19:11) >> No question Microsoft takes a slice of (00:19:13) that, but you've gone from near 100% or (00:19:16) certainly way above 50% to, you know, (00:19:19) sub 25% market share in just one year. (00:19:22) What is the sort of confidence that (00:19:23) people can get that Microsoft (00:19:25) >> there's no again it goes back a little (00:19:27) bit D to sort of there's no birthright (00:19:29) here that we should have any confidence (00:19:31) other than to say hey we should go (00:19:33) innovate and knowing the the lucky break (00:19:36) we have in some sense is that uh this (00:19:39) category is going to be a lot bigger (00:19:40) than anything we had high share in let's (00:19:43) let me say it that way right in some (00:19:44) sense you could say man we kind of had (00:19:46) high share in VS code we had high share (00:19:48) in the repos for with GitHub uh uh and (00:19:52) that was a good market but the point is (00:19:54) even having a decent share in what is a (00:19:57) much more expansive market right I mean (00:19:58) you could say we had a high share in (00:20:00) client server server computing we have (00:20:02) much lower share than that in hypers (00:20:05) scale but is it a much bigger business (00:20:08) by orders of magnitude so at least (00:20:11) there's existence proof that Microsoft's (00:20:12) been okay uh even if our share position (00:20:16) has not been as strong as it was uh as (00:20:19) long as the markets we're competing in (00:20:21) creating more value and there are (00:20:23) multiple winners. Uh so I think that's (00:20:25) the stuff but I I I take your point that (00:20:28) ultimately it all means you have to get (00:20:31) competitive. So I watch that every (00:20:32) quarter and so that's why I think what (00:20:34) I'm very optimistic that uh what we're (00:20:36) going to do with GitHub HQ and or agent (00:20:39) HQ turning GitHub into a place where all (00:20:42) these agents come uh as I said we'll (00:20:45) have multiple shots on goal on there (00:20:47) right it it need not be that hey some of (00:20:49) these guys can succeed along with us uh (00:20:51) and so it doesn't need to be just one (00:20:53) winner uh and one subscription (00:20:56) >> I I guess the reason to focus on this (00:20:58) question is that it's not just about (00:20:59) GitHub but fundamentally about office (00:21:02) and all the other software that (00:21:04) Microsoft offers which is that one (00:21:06) vision you could have about how I (00:21:07) proceeds is that look the models are (00:21:11) going to keep being hobbled then you'll (00:21:13) need this direct visible um (00:21:16) observability all the time and another (00:21:18) vision is over time these models can now (00:21:20) they're doing tasks that take two (00:21:21) minutes in the future they'll be doing (00:21:22) tasks that next be tasks that take 10 30 (00:21:24) minutes in the future maybe they're (00:21:26) doing days worth of work autonomously (00:21:28) and then the model companies are (00:21:30) charging thousands of dollars maybe for (00:21:32) access to really a co-orker which could (00:21:35) use any UI to communicate with their (00:21:38) human and so forth and migrate between (00:21:41) platforms. So if we're getting closer to (00:21:43) that, why aren't the model companies (00:21:45) that are just getting more and more (00:21:48) profitable, the ones that are taking all (00:21:49) the margin, why is the the place where (00:21:52) the scaffolding happens, which becomes (00:21:53) less and less relevant as a as become (00:21:54) more cap capable going to be that (00:21:56) important? And that goes to, you know, (00:21:58) office as it exists now versus (00:22:00) co-workers that are just doing knowledge (00:22:01) work. (00:22:02) >> Great point. I mean I think that's a I (00:22:03) mean for example I mean this is where (00:22:05) you know does all the value migrate just (00:22:07) to uh the model um and uh or does the (00:22:11) you know the does it get split between (00:22:13) the scaffolding um and the model and (00:22:17) what have you I think that uh time will (00:22:19) tell but my my fundamental point also is (00:22:21) the incentive structure gets clear right (00:22:23) which is if you take um let's take uh (00:22:26) let's take information work or take even (00:22:28) coding u already in fact one of favorite (00:22:32) settings I have uh in GitHub copilot is (00:22:35) called auto um right which will just (00:22:38) optimize in fact I buy a subscription (00:22:41) the auto one will start picking and (00:22:44) optimizing for what I am asking it to do (00:22:47) uh and it could even be fully autonomous (00:22:49) and it could sort of arbitrage the (00:22:51) tokens available across multiple models (00:22:53) to go get a task done so if that is the (00:22:56) that means that if you take that (00:22:58) argument the commodity there will be (00:23:00) models (00:23:01) uh and especially with open source (00:23:03) models you can pick a checkpoint and you (00:23:05) can take a bunch of your data and you're (00:23:07) seeing it right I think all of us will (00:23:09) start you whether it's from cursor or (00:23:10) from Microsoft you'll start seeing some (00:23:13) in-house models even uh which will and (00:23:15) then you'll offload most of your uh task (00:23:18) to it so I think that one argument is if (00:23:21) you win the scaffolding uh which today (00:23:25) is dealing with all the hobbling (00:23:27) problems or the uh the jaggedness of (00:23:30) this intelligence problems which you (00:23:32) kind of have to um if you win that then (00:23:35) you will vertically integrate yourself (00:23:37) into the model just because you will (00:23:39) have the liquidity of the data and what (00:23:40) have you and there are enough and more (00:23:42) checkpoints that are going to be (00:23:43) available. uh that's the other thing (00:23:45) right so structurally I think there will (00:23:47) always be an open- source model uh that (00:23:50) will be fairly capable in the world that (00:23:53) you could then use as long as you have (00:23:56) something that you can use that uh with (00:23:58) which is data uh and a scaffolding right (00:24:01) so I can make the argument that oh my (00:24:03) god uh if you're a model company you may (00:24:05) be you may have a winner's curse you may (00:24:07) have done all the hard work done (00:24:10) unbelievable innovation except it's kind (00:24:12) of like one copy uh away from that being (00:24:16) commoditized and then the person who has (00:24:19) the data for grounding and context (00:24:21) engineering um and the liquidity of data (00:24:25) can then go take that checkpoint and (00:24:27) train it. So I think the argument can be (00:24:29) made both ways. (00:24:30) >> Unpacking sort of what you said, there's (00:24:32) two views of the world, right? One is (00:24:33) that models, there's so many different (00:24:35) models out there. Open source exists. (00:24:38) There will be differences between the (00:24:39) models that will drive some level of, (00:24:41) you know, who wins and who doesn't, but (00:24:43) the scaffolding is what enables you to (00:24:45) win. The other view is that actually (00:24:48) models are the key IP and yes, we're in (00:24:50) a very everyone's in a tight race and (00:24:52) there's some, you know, hey, I can use (00:24:54) anthropic or open AAI and you can see (00:24:56) this in the revenue charts, right? like (00:24:57) OpenAI's revenue started skyrocketing (00:24:59) once they finally had a code model (00:25:01) similar capabilities to anthropic (00:25:02) although in different ways. Um (00:25:05) there's a view that like the model (00:25:07) companies are actually the ones that (00:25:08) garner all the margin right because you (00:25:11) know if you look across this year at (00:25:12) least on entropic their gross margins on (00:25:14) inference went from you know well below (00:25:16) 40% to north of 60 right by the end of (00:25:19) the year um the these the margins are (00:25:21) expanding [clears throat] there despite (00:25:23) hey more Chinese open source models than (00:25:24) ever hey open's competitive hey Google's (00:25:26) competitive hey x Grock is now (00:25:28) competitive right all these all these (00:25:30) companies are now competitive and yet (00:25:31) despite this the margins have expanded (00:25:33) at the model layer significantly. Um h (00:25:37) how do you think about the (00:25:38) >> it's a it's a great question. I I think (00:25:40) that the one thing is perhaps a few (00:25:43) years ago people were saying oh I can (00:25:45) just wrap a model and build a successful (00:25:47) company. Uh and that I think is probably (00:25:51) gotten debunked just because the model (00:25:52) capabilities um and with tools use in (00:25:55) particular. (00:25:57) But the interesting thing is there's no (00:25:58) like when I look at Office 365. Let's (00:26:00) take even this little thing we built (00:26:02) called Excel agent. It's interesting (00:26:04) right? Excel agent is not a UI level (00:26:06) wrapper. It's actually a model that is (00:26:10) in the middle tier. Uh in this case (00:26:13) because we have all the IP from the the (00:26:15) GPT family. uh we are taking that and (00:26:19) putting it into the core middle tier of (00:26:22) the office system to both teach it what (00:26:26) it means to natively understand Excel (00:26:30) everything in it. So it's not just hey I (00:26:32) just have a pixel level understanding I (00:26:34) have an full understanding of all the (00:26:36) native artifacts of Excel uh both when I (00:26:39) see it like because if you think about (00:26:40) it if I'm going to give it some (00:26:42) reasoning task right I need to even fix (00:26:44) the reasoning mistakes I make and so (00:26:46) that means I need to both not just see (00:26:48) the pixels I need to be able to see oh I (00:26:50) got that formula wrong and I need to (00:26:52) understand that and then so to some (00:26:54) degree that's all being done not at the (00:26:56) UI wrapper level with some prompt but (00:26:58) it's being done in the middle tier by (00:27:00) teaching it all the tools of Excel, (00:27:02) right? So, I'm giving it even (00:27:03) essentially a markdown to teach it the (00:27:06) skills of what it means to be a (00:27:07) sophisticated Excel user. So, it's a (00:27:09) weird thing that it it goes back a (00:27:11) little bit to AI brain, right? Which is (00:27:13) you're building not just Excel. You are (00:27:16) now business logic in its traditional (00:27:19) sense. You're taking the Excel business (00:27:21) logic in the traditional sense and (00:27:23) wrapping essentially a cognitive layer (00:27:25) to it using this model which knows how (00:27:28) to use the tool. So in some sense, Excel (00:27:31) will come with an analyst bundled in and (00:27:34) with all the tools used. (00:27:35) >> That's the type of stuff that'll get (00:27:38) built by everybody. So even for the (00:27:40) model companies, they'll have to (00:27:41) compete, right? So if they price stuff (00:27:43) high, uh guess what? If I'm a builder of (00:27:46) a tool like this, I'll substitute you. I (00:27:49) may use you for a while. And so as long (00:27:51) as there's competition, there's always a (00:27:53) winner take all thing, right? If there's (00:27:54) going to be one model that is better (00:27:56) than everybody else with massive (00:27:57) distance, yes, that's a winner take all. (00:27:59) As long as there's going to be (00:28:00) competition where there's multiple (00:28:02) models just like hypers scale (00:28:04) competition and there's an open- source (00:28:05) check, uh there is enough room here uh (00:28:09) to go build value on top of models. (00:28:12) >> Uh but at Microsoft, the way I look at (00:28:13) it and say is uh we are going to be in (00:28:16) the hypers scale business which will (00:28:18) support multiple models. we will have (00:28:21) access to open AI models for uh you know (00:28:24) seven more years which we will innovate (00:28:26) on top of so royalty I mean essentially (00:28:28) I think of ourselves as having a (00:28:29) frontier class model uh that we can use (00:28:32) and innovate on with full uh flexibility (00:28:35) and we'll build our own models uh with (00:28:37) Mai um and and so we will always have a (00:28:41) model level and then we'll build these (00:28:43) whether it's in security whether it's in (00:28:45) knowledge work whether it's in coding or (00:28:47) in science we will build our own (00:28:49) applications scaffolding which will be (00:28:51) model forward right it won't be a (00:28:53) wrapper on a model but the model will be (00:28:56) wrapped into uh the application I have (00:28:59) so many questions about the other things (00:29:01) you mentioned but before we move on to (00:29:02) those topics um I still wonder whether (00:29:05) this is like not forwardlooking on AI (00:29:08) capabilities where you're imagining (00:29:10) models like they exist today where yeah (00:29:12) I can you have to like it takes a (00:29:14) screenshot of your screen but it can't (00:29:16) like look inside each cell and what the (00:29:17) formula is and I think the better mental (00:29:19) audited here is like look a human just (00:29:21) imagine that these models actually will (00:29:22) be able to actually use a computer as (00:29:24) well as a human and a human knowledge (00:29:26) worker who is using Excel can look into (00:29:28) the formulas can you know use (00:29:30) alternative software can migrate data (00:29:32) between office 365 and another piece of (00:29:35) software if the migration is necessary (00:29:36) etc so that's kind of what I'm saying (00:29:39) but if that's the case then the (00:29:41) integration with Excel doesn't matter (00:29:42) that much don't worry about the Excel (00:29:45) integration (00:29:46) after all Excel was built as a tool for (00:29:48) anal analysts. Great. So, whoever is (00:29:52) this AI that is an analyst should have (00:29:55) tools that they can (00:29:57) >> computer, right? Just the way a human (00:29:58) can use a computer, that's their tool. (00:30:00) >> The the tool is the computer. Right. (00:30:02) >> Right. So, that so all I'm saying is I'm (00:30:04) building an analyst as as essentially an (00:30:06) AI agent uh which happens to come with (00:30:10) an a priority knowledge of how to use (00:30:12) all of these analytical tools. But is it (00:30:15) is it something maybe just just to make (00:30:17) sure we're talking about the same thing. (00:30:18) Um is it a thing that a hum like me (00:30:21) using Excel as a podcast (00:30:25) completely autonomous? So just imagine I (00:30:27) work like so we should now maybe sort of (00:30:29) lay out how I think the future of the (00:30:31) company is right. uh the future of the (00:30:33) company would be the tools business (00:30:35) which I have a computer I use Excel and (00:30:38) in fact in the future I'll even have a (00:30:40) co-pilot and that co-pilot will also (00:30:42) have agents right that's still I am I (00:30:44) you know it's still me steering (00:30:46) everything (00:30:47) >> and everything is coming back so that's (00:30:48) kind of one world (00:30:50) >> then the second world is the company (00:30:52) just literally provisions a computing (00:30:55) resource for an AI agent (00:30:58) >> and that is working fully autonomously (00:31:00) >> that fully autonomous agent will have (00:31:03) essentially embodied set of those same (00:31:06) tools, (00:31:06) >> right? (00:31:07) >> Uh that are available to it, right? So (00:31:09) this AI tool that comes in also has not (00:31:13) just a raw computer uh because it's (00:31:15) going to be more token efficient to use (00:31:17) tools to get stuff done. In fact, I kind (00:31:20) of look at it and say our business which (00:31:22) today is an enduser tools business will (00:31:24) become essentially an infrastructure (00:31:26) business in support of agents doing (00:31:28) work. Is there another way to think (00:31:30) about it? Right? So if one of the things (00:31:32) that you'll see us do in in in fact like (00:31:35) all the stuff we built underneath M365 (00:31:39) still is going to be very relevant uh (00:31:42) you need someplace to store it someplace (00:31:44) to do archival someplace to do discovery (00:31:47) someplace to manage all of these (00:31:49) activities even if you're an AI agent. (00:31:52) >> So that's so it's kind of a new (00:31:53) infrastructure. So ju just to make sure (00:31:55) I understand you're saying like look (00:31:57) theoretically a future AI that has (00:32:00) actual computer use which is all these (00:32:02) companies are working on model companies (00:32:03) are working right now could use even if (00:32:05) it's not partnered with Microsoft or (00:32:07) under our umbrella could use Microsoft (00:32:09) software but you're saying we're going (00:32:11) to give them if if you're working with (00:32:13) our infrastructure we're going to give (00:32:14) you like lower level access that makes (00:32:17) it more efficient for you to do the same (00:32:18) things you could have otherwise done (00:32:19) anyways. (00:32:20) >> 100%. I mean so the entire thing in in (00:32:22) fact the way the you know like what (00:32:25) happened is we had servers then there (00:32:27) was virtualization and they had many (00:32:29) more servers. So that's another way to (00:32:32) think about this which is hey don't (00:32:33) think of this the tool as the end thing (00:32:36) what is the entire substrate underneath (00:32:39) that tool that humans use and that (00:32:41) entire substrate is the bootstrap for (00:32:44) the AI agent as well because the AI (00:32:46) agent needs a computer that's kind of (00:32:47) one like you know so in fact one of the (00:32:50) fascinating things we're seeing a (00:32:51) significant amount of growth is all (00:32:53) these guys who are doing these office (00:32:55) artifacts and and what have you as (00:32:57) autonomous agents and so on want to (00:32:59) provision Windows 365 right? They really (00:33:01) want to be able to provision a computer (00:33:04) for these agents. Uh and so absolutely (00:33:07) and that's where I think we're going to (00:33:08) have essentially an enduser computing (00:33:11) infrastructure business which I think is (00:33:14) going to just keep growing because guess (00:33:15) what it's going to grow faster than the (00:33:17) number of users. So in fact that's kind (00:33:19) of one of the other questions people ask (00:33:20) me is hey what happens to the per user (00:33:22) business at least the early signs may be (00:33:24) the way to think about the per user (00:33:26) business is not just per user it's per (00:33:28) agent and if you sort of say it's per (00:33:30) user and per agent the key is what's the (00:33:33) stuff to provision for every agent a (00:33:36) computer u a set of security things (00:33:39) around it an identity around it uh and (00:33:42) all those things observability and so on (00:33:45) are the management layers and that's I (00:33:47) think all going to get baked into that (00:33:49) >> the way to frame it at least the way I (00:33:51) currently think about it and I'd like to (00:33:52) hear your you know your view is that (00:33:54) >> uh these model companies are all (00:33:55) building environments to train their (00:33:57) models to use Excel or Amazon shopping (00:34:00) or whatever whatever it is book flights (00:34:03) um but at the same time they're also (00:34:06) training these models to do migration (00:34:08) from because that that is probably the (00:34:10) most immediate uh valuable thing right (00:34:12) converting mainframe based systems to (00:34:15) standard cloud systems converting um (00:34:18) Excel databases into real databases uh (00:34:20) with SQL, right? Or or converting um you (00:34:24) know what is done in Word and Excel to (00:34:27) something that is more programmatic and (00:34:28) more efficient in a classical sense that (00:34:31) can actually be done by humans as well. (00:34:33) It's just not cost-ffective for the (00:34:34) software developer to do that. That (00:34:36) seems to be what everyone is going to do (00:34:37) with AI for the next, you know, few (00:34:39) years at least to massively drive value. (00:34:41) Um h how does Microsoft fit into that? (00:34:44) if the models can utilize the tools (00:34:47) themselves to migrate to something and (00:34:49) yes Microsoft has you know a leadership (00:34:51) position in databases and in storage and (00:34:54) and in all these other categories but (00:34:57) the use of say a office ecosystem is (00:35:01) going to be significantly less just like (00:35:02) potentially the use of a mainframe (00:35:04) ecosystem could be potentially less now (00:35:06) mainframes have grown for the last two (00:35:07) decades actually even though no one (00:35:08) talks about them anymore they've still (00:35:10) grown 100% I [laughter] agree with that (00:35:12) how does how does that flow forward (00:35:13) >> I mean at the end of the day This is not (00:35:15) about sort of hey u there is going to be (00:35:17) a significant amount of time where (00:35:19) there's going to be a hybrid world right (00:35:20) because people are going to be using the (00:35:22) tools that are going to be working with (00:35:23) agents that have to use tools and by the (00:35:25) way they have to communicate with each (00:35:27) other what's the artifact I generate (00:35:29) that then a human needs to see so like (00:35:32) all of these things will be real (00:35:33) considerations in any place so the (00:35:35) outputs input so I don't think it'll (00:35:36) just be about oh I migrate it off right (00:35:38) but the bottom line is I have to live in (00:35:39) this hybrid world so let's but that (00:35:41) doesn't fully answer your question (00:35:43) because there can be a real new (00:35:44) efficient frontier where I it's just (00:35:47) agents working with agents uh and (00:35:49) completely optimizing and even when (00:35:51) agents are working with agents what are (00:35:53) the primitives that are needed uh do you (00:35:56) need a storage system (00:35:57) >> uh does that storage system need to have (00:35:59) eiscocovery does that eiscocovery do you (00:36:02) need to have observability do you need (00:36:04) to have an identity system that is going (00:36:06) to use multiple models with all having (00:36:08) one identity system so these are all the (00:36:10) core underlying rails we have today for (00:36:14) what are office systems or what have (00:36:16) you. Uh and that's what I think we will (00:36:18) have in the future as well. You talked (00:36:19) about databases, right? I mean take you (00:36:22) know man I would love all of Excel to (00:36:23) have a database backend, right? In fact (00:36:25) I would love for all that to happen (00:36:27) immediately. Uh and that database is a (00:36:30) good database. I mean databases in fact (00:36:31) will be a big thing that'll grow. uh in (00:36:34) fact if I think about all of the office (00:36:36) artifacts uh being structured better the (00:36:39) ability to do the joins between (00:36:41) structured and unstructured better (00:36:43) because of the agenting what that'll (00:36:44) grow the underlying what is (00:36:46) infrastructure business it happens the (00:36:48) consumption of that is all being driven (00:36:50) by agents you could say all that is just (00:36:52) in time generated software by a model (00:36:54) company that could also be true if we we (00:36:57) will be one such model company too uh (00:36:59) and so we will build in so the (00:37:01) competition could be uh that we will (00:37:04) build a model plus all the (00:37:05) infrastructure and provision it and then (00:37:08) there will be competition between a (00:37:09) bunch of those folks who can do that. H (00:37:12) um I guess speaking of model companies (00:37:14) you say okay we will also be one of the (00:37:16) not only will we have the infrastructure (00:37:17) we'll have the model itself right now (00:37:19) Microsoft AI's most recent model that (00:37:21) was released 2 months ago is 36 and (00:37:22) Shabbat arena and there's a I mean you (00:37:26) obviously have the IP rights to open so (00:37:28) there's a question of first to the (00:37:29) extent you agree with that it seems to (00:37:31) be behind why is that the case (00:37:33) especially given the fact that you could (00:37:35) um you theoretically have the right to (00:37:36) just like fork open's monor repo or (00:37:39) distill on their models Um yeah, (00:37:42) especially if it's a big part of your (00:37:43) strategy that we need to have a leading (00:37:44) model company. (00:37:45) >> Yeah. I mean, so first of all, we are (00:37:48) absolutely going to use the OpenAI (00:37:50) models uh to the maximum uh across all (00:37:54) of our products, right? I mean, that's I (00:37:55) think the core thing that we're going to (00:37:57) continue to do all the way for the next (00:37:59) seven years. Uh and not just use it uh (00:38:02) but then add value to it. That's kind of (00:38:04) where the analyst in this Excel agent (00:38:07) and these are all things that we will do (00:38:08) where you know we'll do I'll you know RL (00:38:11) fine-tuning we'll do some mid-training (00:38:13) runs on top of a GPT family where we (00:38:15) have unique data assets and build (00:38:17) capability (00:38:19) um the MI model the way I think we're (00:38:22) going to think about it is the the good (00:38:24) news here in fact with the new agreement (00:38:26) is even we can be very very clear that (00:38:28) we're going to build a worldclass super (00:38:30) intelligence team and go after it with (00:38:32) high ambition but that at the same time. (00:38:33) We're also going to use this time to be (00:38:35) smart about how to use both these (00:38:37) things. So that means we will on one end (00:38:40) be very product focused on on the other (00:38:43) end be very research focused. In other (00:38:45) words, uh because we have access uh to (00:38:47) the GPT family. The last thing I don't (00:38:49) want to do is use my flops in a way that (00:38:52) is just duplicative and doesn't add much (00:38:54) value. So I want to be able to take uh (00:38:57) the flops that we use to generate a GPT (00:39:01) family [snorts] and maximize its value (00:39:03) while my MAI flops are being used for (00:39:06) let's take the image model that we (00:39:07) launched which I think just launched uh (00:39:09) it's a number nine in the uh image arena (00:39:12) you know we're using it you know both (00:39:14) for cost optimization it's on copilot (00:39:17) it's in Bing and we're going to use that (00:39:18) we have a audio model in copilot which (00:39:21) it's got personality and what have you (00:39:23) optimized it for our product So we will (00:39:25) do those even on the LM Marina we (00:39:27) started on the text one I think it was (00:39:29) it debuted at night 13 and by the way it (00:39:32) was it was done only on whatever 15,000 (00:39:35) uh H100s and so it was a very small (00:39:37) model and uh so it was again to prove (00:39:40) out uh the core capability the (00:39:42) instruction following and everything (00:39:44) else which but you know we wanted to (00:39:45) make sure we can match what was (00:39:47) state-of-the-art and so that shows us (00:39:49) given scaling laws what we are capable (00:39:51) of doing if you gave more flops to it (00:39:53) right so the next thing we will is an (00:39:55) omni model where we will take sort of (00:39:57) the work we've done in audio, what we've (00:39:59) done in image and what we've done in (00:40:01) text, that'll be the next pit stop on (00:40:03) the MAI side. So when I think about the (00:40:05) MAI road map, we're going to build a (00:40:07) first class super intelligence team. (00:40:09) We're going to continue to drop and do (00:40:10) on in the open some of these models, (00:40:13) they will either be in our products (00:40:15) being used because they're going to be (00:40:16) latency friendly, cogs friendly, or what (00:40:18) have you, or they'll have some special (00:40:20) capability. and we will do real research (00:40:23) in order to be ready for some next five, (00:40:26) six, seven, eight break breakthroughs uh (00:40:28) that are all needed on this march (00:40:30) towards super intelligence. So I think (00:40:31) that's and while exploiting (00:40:34) the advantage we have of having the GPT (00:40:37) family that we can work on top of as (00:40:39) well. (00:40:40) >> Say we roll forward seven years uh you (00:40:42) no longer have access to open AI models. (00:40:44) what does one get confidence or what (00:40:46) does Microsoft do to make sure they are (00:40:49) leading or have a leading AI lab right (00:40:52) today you know it's it's all open has (00:40:54) developed many of the breakthroughs (00:40:55) whether it be scaling or reasoning or (00:40:57) Google's developed all the breakthroughs (00:40:58) like transformers uh but but it it is (00:41:01) also a big talent game right you know (00:41:03) you've seen meta spend you know north (00:41:05) of20 billion on talent right uh you've (00:41:07) seen anthropic uh poach the entire blue (00:41:10) shift reasoning team from Google last (00:41:12) year you've seen meta poach a large (00:41:14) reasoning and post training team from (00:41:16) Google more recently. These these sorts (00:41:18) of talent wars are very capital (00:41:20) intensive. They're the ones that, you (00:41:21) know, arguably, you know, if you're (00:41:23) spending hundred billion dollars on (00:41:24) infrastructure, you should also spend, (00:41:26) you know, x amount of money on on the (00:41:28) people using the infrastructure so that (00:41:30) they're more efficiently making these (00:41:31) new breakthroughs. what what confidence (00:41:33) can one get that you know hey Microsoft (00:41:35) will have a team that's world class that (00:41:37) can make these breakthroughs and you (00:41:39) know once you decide to turn on the (00:41:41) money faucet you know you're being a bit (00:41:42) capital efficient right now which is (00:41:44) which is smart it seems uh to not waste (00:41:46) money doing duplicative work but once (00:41:48) you decide you need to you know how how (00:41:50) can one say oh yeah now you can shoot up (00:41:52) to where the top five model (00:41:54) >> well look I mean at the at the end of (00:41:56) the day we're going to build a (00:41:57) world-class team and we already have a (00:41:59) world-class team that's beginning to be (00:42:01) sort of assembled right Mustafa coming (00:42:03) in. We have Karen. We have Amar (00:42:04) Subramanyan who did a lot of the post (00:42:06) training at Gemini. Tufi who is at (00:42:08) Microsoft. Nando who did a lot of the (00:42:10) multimedia work at Deep Mind is there. (00:42:12) And so we're going to build a worldclass (00:42:15) team. And in fact I think later this (00:42:17) week even Mustafa published some you (00:42:19) know a little more clarity on what our (00:42:20) lab is going to go do. Um I think the (00:42:23) thing that I want uh the world to know (00:42:27) perhaps uh is we are going to build the (00:42:30) infrastructure that'll support multiple (00:42:32) models. Uh you know uh we because from a (00:42:36) hypers scale perspective we want to (00:42:37) build the most scaled infrastructure (00:42:40) fleet that's capable of supporting all (00:42:43) the models the world needs whether it's (00:42:44) from open source or whether it's (00:42:46) obviously from open AI and others and so (00:42:48) that's kind of one job. Second is in our (00:42:50) own model capability. We will absolutely (00:42:53) use the open AI model in our products (00:42:55) and we will start building our own (00:42:56) models and we may like in in GitHub (00:42:58) copilot anthropic is used. So we will (00:43:00) even have other frontier models that are (00:43:02) going to be wrapped into our products as (00:43:04) well. So I think that that's kind of how (00:43:06) at least each time at the end of the day (00:43:08) the eval of the product as it meets a (00:43:11) particular task or a job is what matters (00:43:14) and we'll sort of back from there into (00:43:16) the vertical integration needed. uh (00:43:19) knowing that as long as you're service (00:43:20) you know you're serving the market well (00:43:22) with the product you can always cost (00:43:24) optimize (00:43:26) >> there there's a question going forward (00:43:27) so right now we have models that have (00:43:29) this distinction between training and (00:43:30) inference and one could argue that (00:43:33) there's like a smaller and smaller (00:43:35) difference between the different models (00:43:37) um going forward if you're really (00:43:38) expecting something like human level (00:43:39) intelligence humans learn on the job you (00:43:42) know if you think about your last 30 (00:43:43) years what makes s token so valuable (00:43:45) it's the last 30 years of wisdom and (00:43:47) experience you've gained at Microsoft (00:43:49) Um and we will eventually have models if (00:43:51) they get to human level which will have (00:43:52) this ability to continuously learn on (00:43:54) the job and that will drive so much (00:43:56) value to the model company that is ahead (00:43:58) at least in my view because you have (00:44:00) copies of one model broadly deployed (00:44:02) through the economy learning how to do (00:44:04) every single job and unlike humans they (00:44:06) can amalgamate their learnings to that (00:44:08) model. So there's this sort of (00:44:10) continuous learning sort of exponential (00:44:12) feedback loop um which almost looks like (00:44:14) a sort of intelligence explosion. uh if (00:44:17) that happens and Microsoft isn't the (00:44:19) leading model company by that time (00:44:23) doesn't then this uh you know you're (00:44:25) saying well we substitute one model for (00:44:27) another etc matter less because it's (00:44:28) just like this one model knows how to do (00:44:30) every single job of the economy the (00:44:32) other long tale of don't (00:44:33) >> yeah no I think your point about if (00:44:35) there's one model that is the only model (00:44:37) that is most broadly deployed in the (00:44:39) world and it sees all the data and it (00:44:41) does continuous learning that's game set (00:44:43) match and you know is shut right I mean (00:44:46) the the reality at least I see um is the (00:44:51) world even today for all the dominance (00:44:55) of any one model it's not the case um (00:44:59) it's like take any take coding there's (00:45:02) multiple models in fact every day it's (00:45:04) less the case where there is not one (00:45:07) model that is getting deployed broadly (00:45:09) in fact there's multiple models that are (00:45:11) getting deployed it's kind of like (00:45:12) databases right it's always the thing (00:45:14) it's like hey can one database be the (00:45:16) one that just is used everywhere except (00:45:18) it's not uh there are multiple types of (00:45:20) databases that are getting deployed uh (00:45:22) for different use cases. So I think that (00:45:25) there is going to be some network (00:45:27) effects of continual learning or data (00:45:29) you you know I'll call liquidity that (00:45:31) any one model has. Uh is it going to (00:45:34) happen in all domains? I don't think so. (00:45:36) Is it going to happen in all geos? I (00:45:38) don't think so. Is it going to happen in (00:45:39) all segments? I don't think so. It'll (00:45:41) happen in all categories at the same (00:45:42) time. I don't think so. So therefore I (00:45:44) feel like the design space is so large (00:45:47) uh that there's plenty of opportunity (00:45:49) but your fundamental point is having a (00:45:52) capability which is at the (00:45:53) infrastructure layer, model layer and at (00:45:56) the scaffolding layer and then to be (00:45:59) able to compose these things not just as (00:46:01) a vertical stack but to be able to (00:46:03) compose each thing for what its purpose (00:46:05) is. Right? You can't build an (00:46:06) infrastructure that's optimized for one (00:46:08) model. If you do that what if you go (00:46:10) fall behind? In fact, all the (00:46:12) infrastructure you built will be a (00:46:14) waste, right? You kind of need to build (00:46:16) an infrastructure that's capable of (00:46:18) supporting multiple sort of families and (00:46:21) lineages of models. Otherwise, the (00:46:22) capital you put in which is optimized (00:46:24) for one model architecture. That means (00:46:26) you're one tweak away from some like (00:46:29) breakthrough that happens for somebody (00:46:30) else and your entire network topology (00:46:32) goes out of the window. Then that's a (00:46:34) scary thing, right? So therefore, you (00:46:36) kind of want the infrastructure to (00:46:38) support whatever may come. in fact in (00:46:40) your own model family and other model (00:46:42) families and you got to be open if you (00:46:44) if you're serious about the hypers scale (00:46:45) business you got to be serious about (00:46:46) that right um if you're serious about (00:46:49) being a model company you've got to (00:46:51) basically say hey what are the ways (00:46:53) people can actually do things on top of (00:46:55) the model so that I can have an ISV (00:46:58) ecosystem unless I'm thinking I'll own (00:46:59) every category that just can't be that (00:47:01) then you won't have an API business and (00:47:03) that by definition will mean you'll (00:47:05) never be uh a platform company that's (00:47:07) going to be successfully deployed (00:47:09) everywhere right So therefore the (00:47:11) industry structure is so such that it (00:47:14) will (00:47:16) really force people to specialize and (00:47:19) [snorts] that in that specialization (00:47:22) a company like Microsoft should compete (00:47:24) in each layer by its merits uh but not (00:47:28) think that this is all about all a road (00:47:31) to game set match where I just compose (00:47:33) vertically all these layers. That's that (00:47:35) just doesn't happen. So according to (00:47:37) Dylan's numbers, there's going to be (00:47:39) half a trillion in AI capex next year (00:47:41) alone, and labs are already spending (00:47:43) billions of dollars to snag top (00:47:45) researcher talent. But none of that (00:47:47) matters if there's not enough (00:47:48) highquality data to train on. Without (00:47:50) the right data, even the most advanced (00:47:52) infrastructure and world-class talent (00:47:54) won't translate into end value for the (00:47:57) user. That's where Libox comes in. Libox (00:48:00) produces highquality data at massive (00:48:03) scale, powering any capability that you (00:48:05) want your model to have. It doesn't (00:48:07) matter whether you need a coding agent (00:48:08) that needs detailed feedback on (00:48:10) multi-our trajectories or a robotics (00:48:12) model that needs thousands of samples on (00:48:14) everyday tasks or a voice agent that can (00:48:17) also perform real world actions for the (00:48:19) user like booking them a flight. To be (00:48:21) clear, this isn't just off-the-shelf (00:48:22) data. Labelbox can design and launch a (00:48:26) custom production scale data pipeline in (00:48:29) 48 hours and they can get you tens of (00:48:31) thousands of targeted examples in weeks. (00:48:34) Reach out at labelbox.com/darkh. (00:48:39) All right, back to Satia. (00:48:42) >> So last year Microsoft was on path to be (00:48:45) the largest infrastructure provider uh (00:48:47) by far. You were the earliest in 23. So (00:48:49) you you went out there, you acquired all (00:48:50) the resources in terms of leasing data (00:48:52) center, starting construction, securing (00:48:54) power, everything. You guys were on pace (00:48:56) to beat Amazon in 26 or 27. Um but (00:49:00) certainly by 28, you were going to beat (00:49:01) them. Um since then, you you know, in (00:49:04) let's call it the second half of last (00:49:05) year, Microsoft did this big pause, (00:49:07) right, where they let go of a bunch of (00:49:10) leasing sites that they were going to (00:49:11) take, which then Google, Meta, um (00:49:14) Amazon, in some cases, Oracle, uh took (00:49:17) these sites. We're sitting in one of the (00:49:19) largest data centers in the world. So (00:49:20) obviously it's not everything. You guys (00:49:21) are expanding like crazy. Uh but there (00:49:23) are sites that you just stopped working (00:49:24) on. (00:49:25) >> Why why did you do this? Right. (00:49:27) >> Yeah. I mean the fundamental thing we (00:49:30) this goes back a little bit to what is (00:49:32) the hypers scale business all about (00:49:34) right which is one of the key decisions (00:49:36) we made was that if you're going to (00:49:39) build out Azure to be fantastic for all (00:49:44) sort of stages of AI uh from training to (00:49:48) mid-training to data genen to inference (00:49:51) we just need fungibility uh of the fleet (00:49:55) um and and so that entire thing caused (00:49:59) us not to basically go build a a whole (00:50:02) lot of capacity with a particular set of (00:50:05) generations. Uh because the other thing (00:50:08) that you got to realize is having (00:50:10) actually for up to now 10xed every 18 (00:50:13) months enough training capacity for the (00:50:15) various open AI models. uh we realize (00:50:18) that um the key is to stay on that path (00:50:22) but the more important thing is to (00:50:25) actually have a balance to not just (00:50:27) train but to be able to serve these (00:50:29) models all around the world because at (00:50:31) the end of the day the rate of (00:50:32) monetization is what then will allow us (00:50:34) to even keep uh funding and then the (00:50:36) infrastructure was going to need us to (00:50:39) support as I said multiple models and (00:50:41) what have you. So once we said that (00:50:43) that's the case since then we just (00:50:45) course corrected to where the path we're (00:50:48) on right if I look at the path we're on (00:50:49) is we're doing lot more starts now uh (00:50:52) we're also buying up as much capacity as (00:50:55) we can whether it's to build whether (00:50:57) it's to lease or even GPUs as a service (00:50:59) but we're building it for where we see (00:51:01) the demand uh and the serving needs and (00:51:04) our training needs and we didn't want to (00:51:07) just be a host for one company uh and (00:51:11) have just a massive book of business (00:51:13) with one customer that that's not a (00:51:15) business right that is sort of uh you (00:51:17) know you should be vertically integrated (00:51:18) with that company uh and so given the (00:51:22) the thing that openai was going to be a (00:51:24) successful independent company which is (00:51:25) fantastic right I think it's makes sense (00:51:27) right and even meta may use third party (00:51:30) capacity but ultimately they're all (00:51:32) going to be first party uh for anyone (00:51:34) who has large scale they'll be you know (00:51:37) they'll be a hyperscaler on their own (00:51:39) and so to me was to build out a hypers (00:51:42) scale fleet and our own research (00:51:44) compute. Uh and that's what the (00:51:46) adjustment was. Um you know and then and (00:51:49) so I feel very very good. Oh by the way (00:51:50) the other thing is I didn't want to get (00:51:53) stuck with massive scale of one (00:51:56) generation. I mean we just saw the the (00:51:57) GB200s. I mean the GB300's are coming (00:52:00) right and by the time I get to Vera (00:52:02) Rubin Ver Rubin ultra guess what the (00:52:04) data center is going to look very (00:52:07) different because the power per rack (00:52:09) power per row is going to be so (00:52:10) different uh the cooling requirements (00:52:12) are going to be so different and that (00:52:14) that means I don't want to just go build (00:52:16) out like a whole number of gigawatts (00:52:18) that are only for one generation one (00:52:21) family and so I think the pacing matters (00:52:25) and the funibility and the location (00:52:27) matters (00:52:28) uh the workload diversity matters, (00:52:30) customer diversity matters and that's (00:52:32) what we're building towards. The other (00:52:34) thing that we've learned a lot is um (00:52:36) every AI workload does require not only (00:52:39) the AI accelerator but it requires a (00:52:41) whole lot of other things right and in (00:52:43) fact a lot of the margin structure for (00:52:44) us will be in those other things and so (00:52:46) therefore we want to build out Azure as (00:52:50) being fantastic for the long tail of the (00:52:53) workloads because that's the hypers (00:52:55) scale business while knowing that we've (00:52:57) got to be super competitive starting (00:52:59) with the bare metal for the highest end (00:53:02) training And but that can't crowd out (00:53:04) the rest of the business, right? Because (00:53:06) we're not in the business of just doing (00:53:08) five contracts with five customers being (00:53:11) their bare metal service. That's not a a (00:53:14) Microsoft business. That may be a (00:53:15) business for someone else and that's a (00:53:17) good thing. What we have said is we are (00:53:18) in the hypers scale business which is at (00:53:20) the end of the day a longtail business (00:53:23) uh for AI workloads and in order to do (00:53:26) that we will have some leading bare (00:53:29) metal as a service capabilities for a (00:53:32) set of models including our own uh and (00:53:34) that I think is the balance you see the (00:53:36) another sort of question that comes (00:53:37) around this whole fungeibility topic is (00:53:40) okay it's not where you want it right (00:53:42) you would rather have it in a good (00:53:43) population center like Atlanta we're (00:53:45) here um there there's there's There's (00:53:47) also the question of like well how much (00:53:49) does that matter if as the horizon of AI (00:53:51) tasks grows well actually you know 30 (00:53:54) seconds for a reasoning prompt or you (00:53:57) know 30 minutes for a deep research or (00:53:59) you know it's going to be hours for (00:54:00) software agents at some point um and (00:54:03) days and so on and so forth the time to (00:54:04) human interaction why does it matter if (00:54:06) it's if it's a great it's a great (00:54:08) question (00:54:08) >> a location A B or C (00:54:10) >> that's exactly right so in fact that's (00:54:11) one of the other reasons why we want to (00:54:13) think about like hey what is an Azure (00:54:14) region look like and what is the in fact (00:54:16) the networking between Azure regions. So (00:54:18) this is where uh I think as the model (00:54:20) capabilities evolve and I think the (00:54:23) usage of these tokens whether it's (00:54:25) synchronously or asynchronously evolves (00:54:27) and in fact you don't want to be out of (00:54:29) position right then on top of that by (00:54:31) the way what are the data residency laws (00:54:34) right where do I like I mean the entire (00:54:36) EU thing uh for us where we literally (00:54:39) had to create an EU data boundary (00:54:41) basically meant that you can't just (00:54:43) round trip a call to wherever even if (00:54:45) it's asynchronous and so therefore you (00:54:47) need to have maybe regional things that (00:54:49) are high density and then the power (00:54:51) costs and so on. But you're 100% right (00:54:53) in bringing up that the topology as we (00:54:57) build out uh will have to evolve one for (00:55:02) tokens per dollar per watt uh what are (00:55:04) the economics (00:55:06) overlay that with what is the usage (00:55:08) pattern um usage pattern in terms of (00:55:11) synchronous asynchronous but also what (00:55:13) is the compute storage because the (00:55:14) latencies may matter for certain things (00:55:17) uh the storage better be there if I have (00:55:19) a Cosmos DB close to this for session (00:55:21) data or even for an autonomous thing (00:55:23) then that also has to be somewhere close (00:55:25) to it and so on. So I think that all of (00:55:27) those considerations is what will shape (00:55:30) uh the hypers scale business. M (00:55:32) >> you know prior to the pause you were you (00:55:34) were you you know versus you know what (00:55:36) we had forecasted for you by 28 you're (00:55:38) going to be like 12 13 gawatt and now (00:55:41) we're at you know 9 and a half or so (00:55:43) right but you know something that's even (00:55:44) more relevant right and it's it's you (00:55:46) know I just want you to like more (00:55:47) concretely state that this is the (00:55:49) business you don't want to be in but (00:55:50) like Oracle is going from like 1/5if (00:55:52) your size to bigger than you by end of (00:55:54) 2027 and while it's not a Microsoft (00:55:57) level quality of return on invested (00:56:00) capital right they're still making 35% (00:56:02) gross margins, right? sort of the (00:56:04) question is like does it is it isn't it (00:56:06) is it is it you know hey it's not (00:56:08) Microsoft's business to maybe do this (00:56:10) but you've created a hyperscaler now by (00:56:12) refusing this business by giving away (00:56:14) the right of first refusal whatever I'm (00:56:16) not first of all I don't I don't want to (00:56:18) take away any thing from the success (00:56:21) Oracle has had in building their (00:56:22) business and I wish them well and so the (00:56:25) thing that I think I've answered for you (00:56:26) is it didn't make sense for us uh to go (00:56:30) be a host for one model company uh with (00:56:35) limited time horizon RPO let's let's (00:56:38) just put it that way right the thing (00:56:40) that you have to think through is not (00:56:41) what you do in the next 5 years but what (00:56:43) do you do for the next 50 uh because (00:56:46) that's kind of what I we made our set of (00:56:48) decisions um I feel very good about our (00:56:51) open AI partnership and what we're doing (00:56:53) we have a decent book a book book of (00:56:55) business we wish them a lot of success (00:56:57) in fact we are buyers even of Oracle (00:56:59) capacity we wish them success but you (00:57:02) know at this point. I think the (00:57:04) industrial logic for what we're trying (00:57:06) to do is pretty clear which is it's not (00:57:08) about like chasing I first of all I (00:57:10) track by the way your uh things whether (00:57:12) it's the AWS or the Google and ours (00:57:14) which I think is super useful uh but (00:57:17) doesn't mean I got to chase those uh I (00:57:21) have to chase them for not just uh the (00:57:23) gross margin that they may represent in (00:57:25) a period of time you know does m what (00:57:27) what is this book of business that (00:57:29) Microsoft uniquely can go clear which (00:57:32) makes sense for us to clear and that's (00:57:34) what we'll do. I I guess I have a (00:57:36) question even stepping back from this of (00:57:37) okay I I take your point that it's a (00:57:40) better business to be in all else equal (00:57:41) to have a long tale of customers you can (00:57:44) have higher margin from rather than just (00:57:46) serving bare metal to a few labs but (00:57:49) then there's a question of okay which (00:57:50) way is the industry evolving and so if (00:57:52) we believe we're on the path to smarter (00:57:54) and smarter AIs then why isn't the shape (00:57:57) of the industry that the open AIs and (00:58:00) anthropics and deep minds are the (00:58:02) platform which the long tale of (00:58:05) enterprises are actually doing business (00:58:07) with where they need bare metal but like (00:58:09) they are the platform. What is the (00:58:10) longtail that is directly using Azure um (00:58:14) because you know you you want to use the (00:58:16) general (00:58:17) >> going to be available on Azure right? So (00:58:19) any workload that says hey I want to use (00:58:22) um you know some open source model and (00:58:24) an open AI model like I mean if you go (00:58:26) to Azure foundry today you have all (00:58:28) these models that you can provision buy (00:58:30) PTUs get a cosmos DB get a SQL DB get (00:58:34) some storage get some compute that's (00:58:35) what a real workload looks like a real (00:58:37) workload is not just hey I did an API (00:58:39) call to a model a real workload needs (00:58:42) all of these things to go build an app (00:58:46) or instantiate an application in fact (00:58:48) the model companies need that right to (00:58:50) build anything it's just not like I have (00:58:52) a token factory I have to have all of (00:58:54) these things that's the hypers scale (00:58:56) business uh and it's not any one model (00:58:58) but all these models and so if you want (00:59:01) grock plus let's say uh open AI plus an (00:59:04) open source model come to Azure foundry (00:59:06) provision them build your application (00:59:09) here is a database that's kind of what (00:59:11) the business is uh you there is a (00:59:14) separate business called just selling (00:59:15) raw bare metal services to model (00:59:17) companies and that's the argument about (00:59:19) how much of that business you want to be (00:59:21) in and not be in and what is that it's a (00:59:24) very different segment of the business (00:59:25) which we are in and we also have limits (00:59:28) to how much of it is going to crowd out (00:59:30) the rest of it. Uh but that's kind of at (00:59:32) least the way I look at it. So, so (00:59:35) there's there's sort of two questions (00:59:36) here, right? Like why why couldn't you (00:59:37) just do both is one and then the other (00:59:39) one is um given, you know, our our (00:59:42) estimates on what your capacity is in (00:59:43) 2028 is 3 and a half gigawatts lower. (00:59:46) Sure, you could have dedicated that to (00:59:48) OpenAI training and inference capacity, (00:59:51) but you could have also dedicated that (00:59:53) to hey the this three and a half (00:59:55) gigawatts is actually just running Azure (00:59:57) is running Microsoft 365 that's running (00:59:59) GitHub copilot. it doesn't actually I (01:00:01) could have built it and not given it to (01:00:02) open AAI (01:00:03) >> or I may want to build it in a different (01:00:05) location. I may want to build it in UAE. (01:00:07) I may want to build it in India. I may (01:00:08) want to build it in Europe. Right? So (01:00:09) one of the other things is as I said (01:00:11) like where we have real capacity (01:00:13) constraints right now are given the (01:00:14) regulatory needs and the data (01:00:16) sovereignty needs. We got to build all (01:00:17) over the world. Uh it's first of all (01:00:19) state side capacity is super important (01:00:21) and we're going to build everything. But (01:00:22) one of the things is when I look out to (01:00:24) 2030 uh I have a sort of a global view (01:00:27) of what does Microsoft shape of business (01:00:29) by first party and third party third (01:00:31) party segmented by the frontier collabs (01:00:34) and their how much they want versus the (01:00:36) inference capacity we want to build for (01:00:39) multiple models um and our own research (01:00:42) compute needs right so that's all what's (01:00:45) going into my calculus versus saying hey (01:00:48) um I think you're rightfully pointing (01:00:49) out the pause but the pause was not done (01:00:53) because we said oh my god we don't want (01:00:55) to build that we realized that oh we (01:00:58) want to build what we want to build (01:01:00) slightly differently uh by both workload (01:01:04) type as well as geo type and timing as (01:01:07) well like we'll keep ramping up our (01:01:09) gigawatts and the question is at what (01:01:12) pace and in what location and in what (01:01:15) sort of how do I write even the mos's (01:01:16) law on it right which is do I really (01:01:18) want to overbuild 3 and a half in 27 or (01:01:21) do I want to spread that in 2728 knowing (01:01:24) even one of the biggest learnings we had (01:01:26) even with Nvidia is their pace increased (01:01:29) uh in terms of their model I mean their (01:01:31) migrations so that was a big factor I (01:01:33) didn't want to go get stuck for four (01:01:35) years 5 years of depreciation on one uh (01:01:38) generation and I wanted to just (01:01:40) basically buy like in fact Jensen's (01:01:42) advice to me was two things one is hey (01:01:44) get on the speed of light execution (01:01:46) that's why I think even the execution in (01:01:48) this Atlanta data center I mean like in (01:01:49) 90 days right between when we get it and (01:01:52) to hand off to a real workload. that's (01:01:54) sort of real speed of light execution on (01:01:56) their front and so I wanted to get good (01:01:58) at that and then that way then I'm (01:02:00) building this each generation and (01:02:03) scaling uh and then every 5 years then (01:02:06) you have a much more balanced so it (01:02:08) becomes really literally like a flow uh (01:02:11) for a large scale industrial operation (01:02:13) like this where you suddenly are not (01:02:15) lopsided where you built up a lot in one (01:02:17) time and then you take a ma massive (01:02:19) hiatus because you're stuck with all (01:02:20) this to your point in one location which (01:02:23) may be great for training, may not be (01:02:24) great for infants because I can't serve (01:02:26) even if it's like it's all asynchronous, (01:02:28) but Europe ain't going to let me round (01:02:30) trip to Texas. So, that's all of the (01:02:32) things. (01:02:33) >> How do I rationalize this statement with (01:02:35) what you've done over the last few (01:02:36) weeks? You've announced deals with Iris (01:02:38) Energy, um with Nebius, um and Lambda (01:02:42) Labs, and there's a few more coming as (01:02:43) well. Uh you're you're going out there (01:02:45) and securing capacity that you're (01:02:47) renting from the Neoclouds, um rather (01:02:50) than having built it yourself. What was (01:02:51) the what was (01:02:52) >> I think it's it's fine for us because we (01:02:54) now have you know when you have line of (01:02:56) sight to demand which can be served (01:02:58) where people are building it it's great (01:02:59) in fact we'll even have I would say you (01:03:02) know we will take leases we will take (01:03:05) build to suite we'll take even GPUs as a (01:03:07) service where we don't have capacity but (01:03:09) we need capacity and someone else has (01:03:11) that uh and by the way I would even sort (01:03:14) of welcome every Neocloud to just be (01:03:16) part of our marketplace uh because again (01:03:19) guess what if if they go bring their (01:03:21) capacity into our marketplace. That (01:03:23) customer who comes through Azure will (01:03:24) use the Neocloud which is a great win (01:03:26) for them and we'll use compute, storage, (01:03:29) databases, all the rest from Azure. So (01:03:31) I'm not at all thinking of this as just (01:03:34) a you know hey I should just gobble up (01:03:36) all of that myself. M um so you (01:03:39) mentioned the how the you know you're (01:03:42) depreciating this asset that's 5 six (01:03:44) years and this is the majority of the (01:03:46) you know 75% of the TCO of a data center (01:03:49) and Jensen is taking a 75% margin on (01:03:52) that so what all the hyperscalers are (01:03:55) trying to do is develop their own (01:03:56) accelerator so that they can reduce this (01:03:59) overwhelming cost for um uh equipment to (01:04:03) increase their margins. (01:04:04) >> Yeah. And then and then like you know (01:04:05) when you look at where they are right (01:04:07) Google's way ahead of everyone else (01:04:08) right they've been doing it for the (01:04:09) longest they're going to make something (01:04:10) like 5 to 7 million chips right of their (01:04:13) own TPUs you look at Amazon they're (01:04:15) trying to make 3 to 5 million but when (01:04:16) we look at what you know Microsoft is is (01:04:19) ordering of their own chips it's it's (01:04:20) it's way below that number um you've had (01:04:23) a program for just as long what's going (01:04:25) on with your internal (01:04:27) >> good question so so the couple of things (01:04:29) one is the thing that is the biggest (01:04:32) competitor for any new accelerator is (01:04:34) kind of even the previous generation of (01:04:36) Nvidia right I mean in a fleet what I'm (01:04:38) going to look at is the overall TCO so (01:04:40) the bar I have even for our own and (01:04:42) which by the way you know I was just (01:04:43) looking at the data for Maya 200 which (01:04:46) looks great um a except that one of the (01:04:49) things that we learned even on the (01:04:51) compute side right which is we had a lot (01:04:52) of Intel then we introduced AMD and then (01:04:55) we introduced cobalt and so that's kind (01:04:57) of how we scaled it and so we have good (01:05:00) um sort of existence proof of at least (01:05:02) in core compute on how to build your own (01:05:04) silicon and then manage a fleet where (01:05:06) all three are at play in some balance. (01:05:08) Uh because by the way even Google's (01:05:10) buying Nvidia and so is uh Amazon. It (01:05:12) makes sense because Nvidia is innovating (01:05:14) and it's the general purpose thing. All (01:05:16) models run on it. Uh and customer demand (01:05:19) is there because if you build your own (01:05:21) vertical thing, you better have your own (01:05:24) model uh which is you know either going (01:05:26) to use it for training or inference and (01:05:28) you have to generate your own demand for (01:05:29) it or subsidize the demand for it. So (01:05:31) therefore you want to uh make sure um (01:05:34) you scale it appropriately. So the way (01:05:36) we're going to go do it is um have a (01:05:39) closed loop between our own MAI models (01:05:43) and our silicon because I feel like (01:05:45) that's the that's what gives you the (01:05:47) birthright to really do your own silicon (01:05:49) right where you literally have uh (01:05:52) designed the micro architecture with (01:05:55) what you're doing and then you keep pace (01:05:57) with your own models. Um in our case the (01:06:00) the good news here is OpenAI has a (01:06:02) program uh which we have access to. Um (01:06:05) and so therefore to think that Microsoft (01:06:07) is not going to have something that's (01:06:09) >> what level of access do you have to that (01:06:10) >> all of it. (01:06:11) >> You just get the IP for all of that. So (01:06:13) the only IP you don't have is a consumer (01:06:14) hardware. (01:06:15) >> That's it. (01:06:16) >> Oh wow. Okay. (01:06:17) >> Yeah. (01:06:19) Interesting. [laughter] (01:06:20) >> Yeah. And by the way we gave them a (01:06:23) bunch of IP as well to bootstrap them. (01:06:25) Right. So this is one of the reasons why (01:06:27) they had a mass because we built all (01:06:29) these supercomputers together uh or we (01:06:31) built it for them and they uh benefited (01:06:34) from it rightfully so and uh and now as (01:06:36) they innovate even at the system level (01:06:38) we get access to all of it uh and uh we (01:06:42) first wants to want to instantiate what (01:06:44) they build uh for them uh but then we'll (01:06:47) extend it and so to think that we don't (01:06:49) have and so if anything the way I I (01:06:51) think about to your question is (01:06:53) Microsoft wants wants to be a fantastic (01:06:57) I'll call it speed of light execution (01:07:00) partner for Nvidia because quite frankly (01:07:02) that fleet uh is life itself. I'm not (01:07:05) worried about I mean obviously Jensen is (01:07:07) doing super well with his margins but (01:07:09) the TCO has many dimensions to it and I (01:07:11) want to be great at that TCO. Uh on top (01:07:14) of that, I want to be able to sort of (01:07:16) really work with the OpenAI lineage uh (01:07:19) and the MAI lineage and the system (01:07:22) design knowing that we have the IP (01:07:25) rights on both ends. (01:07:26) >> Uh speaking of rights, one thing you (01:07:28) know, you had an interview a couple days (01:07:30) ago uh where you said that we have (01:07:33) rights to the the new agreement you made (01:07:36) with OpenAI. have right the exclusivity (01:07:38) to the stateless API calls that OpenAI (01:07:42) makes and we were sort of confused about (01:07:46) if there's any state whatsoever. I mean (01:07:47) you were just mentioning a second ago (01:07:48) that all these complicated workloads (01:07:49) that are coming up are going to require (01:07:51) memory and databases and storage and so (01:07:53) forth and is that now not stateless of (01:07:56) chat GPT is storing stuff on (01:07:58) >> that's the reason why so the the thing (01:07:59) the business the strategic decision we (01:08:02) made and also accommodating for the (01:08:05) flexibility open AI needed in order to (01:08:07) be able to procure compute for (01:08:09) essentially think of open AI having um a (01:08:12) pass business and a SAS business SAS (01:08:15) business is chat GPT (01:08:17) that pass business is their API. (01:08:19) >> That API is Azure exclusive. (01:08:22) >> The SAS business, they can run it (01:08:25) anywhere (01:08:25) >> and they can partner with anyone they (01:08:27) want to to build SAS products. (01:08:29) >> So if they want to partner and the (01:08:31) partner wants to use a a stateless API, (01:08:34) then Azure is the place where they can (01:08:36) get the stateless API. (01:08:37) >> It seems like there's a way for them to (01:08:38) make you you know build the product (01:08:41) together and and it's a state. (01:08:42) >> No, even that they'll have to come to (01:08:44) Azure. Okay. So if it is any partner and (01:08:46) so so fundamentally you know so again (01:08:49) this is done in the spirit of what is it (01:08:51) that we valued as part of our (01:08:53) partnership and we made sure while at (01:08:56) the same time we were good partners to (01:08:57) open AAI given all the flexibility they (01:08:59) need. So for example, Salesforce wants (01:09:01) to integrate OpenAI. It's not through an (01:09:02) API. They actually work together, train (01:09:04) a model together, deploy it on let's say (01:09:07) Amazon. Now is that is that allowed or (01:09:09) uh or do they have to use (01:09:10) >> for any custom agreement like that? They (01:09:13) will have to come run it. There are some (01:09:15) few exceptions to US government and so (01:09:17) on that we made, but other than that, (01:09:18) they'll have to come to Azure. (01:09:20) >> So as s explained, as AI agents get more (01:09:22) capable, you're going to need more and (01:09:24) more observability into what they're (01:09:26) doing. You're going to need to catch (01:09:27) them when they're making mistakes. (01:09:28) You're going to need highle summaries of (01:09:30) what they're doing and you're going to (01:09:31) need a picture of how everything that (01:09:33) they're doing fits together. This is (01:09:35) exactly what Code Rabbit provides. You (01:09:37) just make a normal pull request and Code (01:09:40) Rabbit automatically reviews the PR. It (01:09:42) generates a summary of changes so you (01:09:43) can understand exactly what the PR's (01:09:45) author was intending and it uses the (01:09:47) context from your full code base to (01:09:49) provide line by line feedback on how (01:09:51) things could be improved. This is (01:09:53) helpful whether you're reviewing a PR (01:09:55) from a co-orker or an agent. In either (01:09:57) case, Code Rabbit will write up its (01:09:59) thoughts and flag any issues so that (01:10:01) your teammate or your agent can [music] (01:10:03) go fix them. I've noticed that when I'm (01:10:05) coding with agents, Code Rabbit catches (01:10:08) a lot of mistakes that the models make (01:10:10) by default. [music] For example, the (01:10:11) models have a bad habit of using old (01:10:14) versions of libraries. So, in one (01:10:16) session, I watch Code Rabbit [music] (01:10:18) cache a call to an old model, figure out (01:10:21) what the new version was, and then (01:10:23) suggest that improvement. Go to code (01:10:26) rabbit.a a/4 (01:10:28) to learn more. (01:10:30) >> Stepping back, a question I have is, you (01:10:32) know, when we were walking back and (01:10:33) forth with the factory, one of the (01:10:35) things we're talking about is, you know, (01:10:38) Microsoft, you can think of it as a (01:10:39) software business, but now it's really (01:10:40) becoming an industrial business. Uh, (01:10:42) there's all this capex, there's all this (01:10:44) construction, and if you just look over (01:10:46) the last two um two years, your sort of (01:10:49) capex is like tripled. And maybe you (01:10:51) extrapolate that forward. It just (01:10:53) actually just becomes this huge (01:10:55) industrial uh explosion. (01:10:56) >> Well, their hyperscalers are taking (01:10:58) loans, right? Meta's Meta's done a $20 (01:11:00) billion loan at Louisiana. They've take (01:11:01) they've done a corporate loan. It seems (01:11:03) clear everyone's free cash flow is going (01:11:04) to zero. Um which is which I'm sure Amy (01:11:08) is like going to beat you up for for (01:11:09) even if you even try to do that. But (01:11:11) like uh what what what's happening? I (01:11:13) mean I think I think the structural (01:11:16) change um is what you're referencing (01:11:20) which I think is massive right which is (01:11:22) I I describe it as we are now a capital (01:11:25) intensive business and a knowledge (01:11:27) intensive business and in fact we have (01:11:28) to use our knowledge to increase the (01:11:30) ROIC on the capital spend right because (01:11:33) that's kind you know look the hardware (01:11:34) guys have done a great job uh of (01:11:36) marketing the morals law which I think (01:11:38) is unbelievable and it's great but if (01:11:40) you even look I think some of the stats (01:11:42) I even did in my earnings call which is (01:11:44) for a given GPT family right uh the (01:11:47) improvement software improvements of (01:11:49) really throughput in terms of tokens per (01:11:52) dollar per watt that we're able to get (01:11:54) uh you know quarter over quarter year (01:11:57) over year is massive uh right so it's 5x (01:12:00) 10x maybe 40x in some of these cases (01:12:02) right just because uh how you can (01:12:04) optimize that's sort of knowledge intens (01:12:08) intensity coming to bring out capital (01:12:10) efficiency (01:12:11) >> so that at at some level the that's what (01:12:15) we have to master. What does it mean? (01:12:16) Like somebody people ask me what was the (01:12:18) difference between uh you know a classic (01:12:20) oldtime host and a hyperscaler it was (01:12:23) software. So yes it is capital intensive (01:12:27) but as long as you have systems knowhow (01:12:30) software capability to optimize by (01:12:33) workload by fleet that's why I think (01:12:35) when when we say fungeibility (01:12:37) there's so much software in it. It's (01:12:39) just not about the fleet, right? It's (01:12:40) kind of the ability to evict a workload, (01:12:43) you know, and then schedule another (01:12:44) workload. Can I like manage the that (01:12:48) algorithm of scheduling around u that is (01:12:51) the type of stuff that we have to be (01:12:53) world class at? And so yes, so I think (01:12:55) we'll still remain a software company. (01:12:57) >> Uh but yes, this is a different (01:12:59) business. Um and we're going to manage. (01:13:01) Look, I think at the end of the day, uh (01:13:02) the cash flow that Microsoft has allows (01:13:05) us to have um both these arms firing, (01:13:10) you know, uh well, (01:13:12) >> it seems like in the short term you have (01:13:14) more sort of um credence on things (01:13:17) taking a while, being more jagged, but (01:13:18) in the maybe in the long term, you think (01:13:20) like the people who say talk about AGI (01:13:22) and ASI are correct like Sam Sam will be (01:13:24) right, but eventually. Um and I I have a (01:13:27) broader question about what makes sense (01:13:29) for a hyperscaler to do given that you (01:13:32) have to invest massively in this thing (01:13:34) which depreciates over 5 years. So so (01:13:37) you if you have 2040 timelines to the (01:13:39) kind of thing that somebody like Sam (01:13:41) anticipates in three years um you know (01:13:43) what what is a reasonable thing for you (01:13:45) to do in that world? there needs to be (01:13:48) an allocation (01:13:49) uh to I'll call it research compute (01:13:53) >> that needs to be done like you did R&D (01:13:57) >> right so that's the best way to even (01:13:58) account for it quite frankly we should (01:14:00) think of it as just R&D expense and you (01:14:02) should say hey what's the research (01:14:04) computer and know how do you want to (01:14:05) scale it (01:14:06) >> um and let's [clears throat] even say (01:14:08) it's an order of magnitude scale um in (01:14:12) some period pick your thing is it two (01:14:14) years is it 16 months what have you (01:14:16) Right. So that's sort of one piece which (01:14:19) is kind of that's kind of table stakes. (01:14:21) That's R&D expenses and the rest is all (01:14:24) demand driven. Right. I mean ultimately (01:14:26) you can you have to build ahead of (01:14:27) demand but you better have a demand uh (01:14:29) uh plan uh that doesn't go completely (01:14:32) offkilter. Do you buy so these labs are (01:14:35) now projecting revenues of 100 billion (01:14:38) in 2728 (01:14:40) uh and they're projecting you know (01:14:41) revenue keeps growing at this rate of (01:14:42) like 3x 2x a year (01:14:44) >> a lot in the marketplace right there's (01:14:46) all kinds of incentives right now and (01:14:48) and rightfully so right I mean what what (01:14:51) do you expect an independent lab that is (01:14:53) sort of trying to raise money to do (01:14:55) right they have to put some numbers out (01:14:57) there such that they can actually go (01:14:59) raise money so that they can pay their (01:15:00) bills for compute and what have you and (01:15:02) it's And it's good thing I someone's (01:15:04) going to take some risk and put it in (01:15:06) there and they've shown traction. It's (01:15:08) not like it's all risk without seeing (01:15:11) the fact that they've been performing (01:15:13) whether it's open AAI, whether it's (01:15:14) anthropic. So I feel great about what (01:15:16) they've done. (01:15:17) >> Uh and we have massive book of business (01:15:19) with these Japs. So therefore uh that's (01:15:21) all good. (01:15:22) >> But overall (01:15:24) ultimately there's two simple things. (01:15:26) One is you got to allocate for R&D. You (01:15:29) brought up even talent. You got to like (01:15:31) the talent for AI is at a premium. You (01:15:33) got to spend there. You got to spend on (01:15:35) compute. So in some sense researcher to (01:15:38) GPU ratios have to be high. Uh that is (01:15:41) sort of what it takes to be a leading (01:15:43) R&D company in this world. Uh and that's (01:15:47) something that needs to scale. Um and (01:15:49) you have to have a balance sheet that (01:15:51) allows you to scale that long before (01:15:52) it's conventional wisdom and so on. So (01:15:54) that's kind of one thing. But the other (01:15:58) is all about sort of knowing how to (01:16:00) forecast (01:16:01) >> as we look across the world right (01:16:03) America has dominated many tech stacks (01:16:06) right um the US owns Windows right (01:16:09) through Microsoft which is deployed even (01:16:11) in China right that's the main operating (01:16:12) system um of course there's Linux which (01:16:14) is open source but you know Windows is (01:16:16) deployed everywhere in China on personal (01:16:17) computers um you look at you look at (01:16:20) word it's it's deployed everywhere you (01:16:21) look at all these various technologies (01:16:22) it's deployed everywhere the thing that (01:16:25) is quite unique and and and Microsoft (01:16:27) and other companies have grown (01:16:28) elsewhere, right? they've built they're (01:16:29) building data centers in Europe and in (01:16:31) India and in and in all these other you (01:16:33) know in Southeast Asia and in Latam in (01:16:35) Africa right all of these different (01:16:37) places you're building capacity but this (01:16:39) seems quite different right you know to (01:16:41) today the political aspect of technology (01:16:46) of compute you know you know the US (01:16:48) administration didn't care about the (01:16:49) dotcom bubble right um it seems like the (01:16:52) US administration as well as every other (01:16:54) administration around the world cares a (01:16:55) lot about AI and the question is you (01:16:58) know we we're in sort of a bipolar world (01:17:00) at least with US and China but Europe (01:17:02) and and India and all these other (01:17:04) countries are are saying no actually (01:17:06) we're going to have sovereign AI as (01:17:07) well. How does Microsoft navigate, you (01:17:09) know, the difference of the 90s where (01:17:11) it's like there's one country in the (01:17:12) world that matters, right? It's America (01:17:14) and we do our companies sell everywhere (01:17:16) and therefore Microsoft benefits (01:17:17) massively to a world where it is bipolar (01:17:20) where hey Microsoft can't just (01:17:21) necessarily have the right to win all of (01:17:23) Europe or India or you know Singapore. (01:17:26) There's actually sovereign AI efforts. (01:17:28) What what is your thought process here (01:17:29) and how do you think about this? It's (01:17:30) it's I think a super you know critical (01:17:34) um piece which is um I think that the (01:17:38) key key priority for the US tech sector (01:17:41) and the US government is to ensure that (01:17:44) we not only do leading innovative work (01:17:48) but we also collectively build trust (01:17:52) around the world on our tech stack right (01:17:56) because I always say the United States (01:17:58) is just an unbelievable place It's just (01:18:00) unique in history, right? It's 4% of the (01:18:03) world's population, 25% of the GDP, and (01:18:06) 50% of the market cap. And I think you (01:18:08) should think about those ratios and uh (01:18:10) really and reflect on it. that 50% (01:18:13) happens because quite frankly the trust (01:18:15) the world has in the United States (01:18:18) whether it's its capital markets or (01:18:20) whether it's its technology and and its (01:18:23) stewardship of what matters at any given (01:18:26) time in terms of leading uh sector. So (01:18:30) if that is broken uh then that's not a (01:18:33) good day for the United States. And so (01:18:35) if we start with that which I think the (01:18:37) you know President Trump gets, the White (01:18:39) House, David Sachs, everyone uh really I (01:18:43) think gets it. Uh and so therefore I (01:18:46) applaud anything that the United States (01:18:49) government and the tech sector jointly (01:18:51) does to quite frankly for example put (01:18:55) our own capital at risk collectively as (01:18:57) an industry in every part of the world. (01:18:59) Right? So I would like in fact the USG (01:19:02) to take credit for foreign direct (01:19:04) investment by American companies all (01:19:07) over the world right it's kind of like (01:19:09) uh least talked about but the best (01:19:11) marketing that the United States should (01:19:12) be doing is it's not just about all the (01:19:15) foreign direct investment coming into (01:19:16) the United States but the most leading (01:19:19) sector which is these AI factories are (01:19:22) all being created all over the world by (01:19:24) whom by America and American companies (01:19:27) and so you start there and Then you even (01:19:30) build other agreements around it which (01:19:32) are around their continuity, their (01:19:35) legitimate sovereignty concerns around (01:19:37) whether it's data residency, whether (01:19:39) it's even what happens um uh for them to (01:19:44) have real agency and guarantees uh on (01:19:48) privacy and so on and so that in fact (01:19:50) our European commitments I think are (01:19:52) worth reading. Right? So we made a (01:19:54) series of commitments to Europe on how (01:19:57) we will really govern our hypers scale (01:20:00) investment there uh such that really (01:20:03) European uh union and the European (01:20:06) countries have sovereignty. We're also (01:20:07) building sovereign clouds in in France (01:20:09) and in Germany. We have something called (01:20:11) sovereign services on Azure which (01:20:14) literally give people key management (01:20:17) services along with confidential (01:20:19) computing including confidential (01:20:21) computing in GPUs which we've done great (01:20:23) innovative work with Nvidia. Um and so I (01:20:26) think I feel very very good about being (01:20:28) able to build both technically and (01:20:32) through policy this trust in the (01:20:35) American tech stack. M and how do you (01:20:37) see the shaking out as you know you do (01:20:39) have this uh network effect with (01:20:41) learning and things on the model level (01:20:43) maybe you have equivalent things at the (01:20:45) hyperscaler level as well and do you (01:20:48) expect that the countries will say look (01:20:50) it's clearly one model or a couple (01:20:51) models are the best and so we're going (01:20:53) to use them but we're going to have some (01:20:54) laws around well the weights have to be (01:20:55) hosted in our country or do you expect (01:20:57) that there will be uh this push to have (01:21:01) it has to be a model trained in our (01:21:02) country maybe an analogy here is like (01:21:04) people would you know the semiconductor (01:21:06) is very important to the economy and (01:21:07) people would like to have their sort of (01:21:08) sovereign semiconductors but like TSMC (01:21:11) is just better and so semiconductors are (01:21:13) so important to the economy that you (01:21:14) will just go to Taiwan and buy the (01:21:16) semiconductors you have to will it be (01:21:18) like that with AI or is there (01:21:20) >> um ultimately I think what matters is (01:21:23) the use of AI in their economy to create (01:21:27) economic value right I mean that's the (01:21:29) uh the diffusion theory which is (01:21:31) ultimately it's not the leading sector (01:21:33) but it's the ability to use the leading (01:21:36) technology to create your own (01:21:37) comparative advantage right so that I (01:21:39) think will fundamentally be the core (01:21:41) driver (01:21:42) >> but that said they will want continuity (01:21:45) of that right so in some sense that's (01:21:46) one of the reasons why I believe there's (01:21:48) always going to be a check a little bit (01:21:50) to sort of some of your points on hey (01:21:53) can this one model have all the runaway (01:21:56) deployment that's why open source is (01:21:58) always going to be there they will be by (01:22:01) definition uh multiple models that'll be (01:22:04) one way like it's kind of the you know (01:22:06) that's one way for people to sort of (01:22:07) demand continuity and not have (01:22:09) concentration risk is another way to say (01:22:11) it is right um and so you say hey I want (01:22:13) multiple models and then I want an open (01:22:15) source so I feel uh as long as that's (01:22:18) there every country will feel like okay (01:22:21) I don't have to worry about deploying (01:22:23) the best model and broadly diffusing (01:22:25) because I can always take uh what is my (01:22:28) data and my liquidity and move it uh to (01:22:31) another model whether it's open source (01:22:33) or from another country or what have (01:22:35) you. So concentration risk (01:22:38) >> um and sovereignty right which is really (01:22:40) agency those are the two things I think (01:22:42) that'll drive the market structure. The (01:22:44) the thing about this is that this (01:22:46) doesn't exist for semiconductors, right? (01:22:47) You know, all refrigerators, cars have (01:22:49) chips made in Taiwan. (01:22:51) >> It didn't exist until now. Until now, (01:22:53) everybody is now like like (01:22:55) >> even even then, right, America, you (01:22:56) know, if Taiwan is cut off, there is (01:22:58) there are no more cars, there are no (01:22:59) more refrigerators. TSMC Arizona is not (01:23:01) replacing any real fraction of the (01:23:04) production. Like there it is. It there (01:23:06) the sovereignty is a bit of like a a (01:23:08) scam, if you will, right? I mean, it's (01:23:09) it's worthwhile having it. It's (01:23:10) important to have it, but it's not a (01:23:12) real it's not real sovereignty, right? (01:23:14) And we're a global economy. We don't we (01:23:15) >> I think it's kind of like Dylan saying, (01:23:17) hey, at this point, we've not learned (01:23:20) anything about sort of uh what (01:23:22) resilience means and what one needs to (01:23:25) do, right? So, it's kind of (01:23:28) any nation state, including the United (01:23:31) States at this point, will do what it (01:23:34) takes to be more self-sufficient on some (01:23:38) of these critical supply chains. (01:23:40) So I as a multinational company have to (01:23:44) think about that as a first class (01:23:46) requirement right if I don't then I'm (01:23:48) not respecting what is in the sort of (01:23:53) policy interests of that country long (01:23:55) term right and I'm not saying they won't (01:23:57) make practical decisions in the short (01:23:59) term right absolutely I mean the (01:24:00) globalization can't just be rewound (01:24:02) right I mean all these capital (01:24:04) investments cannot be made uh in in a (01:24:07) way at the pace at which at the same (01:24:09) time you have to kind of like if I think (01:24:11) about it, right? If somebody showed up (01:24:12) in Washington and said, "Hey, you know, (01:24:13) you know what? We're not going to build (01:24:15) any semiconductor plans, they're going (01:24:17) to be kicked out of the United States." (01:24:19) Um, and and the same thing is going to (01:24:21) be the true in every other country, too. (01:24:23) Uh, and so therefore, I think we have to (01:24:26) as companies respect what the lessons (01:24:29) learned are. Um, you know, whether it's, (01:24:32) you know, you could say the pandemic (01:24:33) woke us up or whatever. But (01:24:35) nevertheless, people are saying, "Look, (01:24:36) globalization was fantastic. (01:24:39) uh it helped the supply chains be (01:24:41) globalized and be super efficient but (01:24:43) there's such a thing called resilience (01:24:44) and we are happy you know we want (01:24:46) resilience and so therefore that feature (01:24:49) will get built at what [snorts] pace I (01:24:51) think you point you're making it can't (01:24:53) be like you can't snap your fingers and (01:24:54) say all the TSMC plants now are all in (01:24:57) Arizona and with all of the capability (01:24:59) they're not going to be but is there a (01:25:01) plan there will be a plan and should we (01:25:03) respect that absolutely and so I I feel (01:25:06) that's the world I want to meet the (01:25:08) world where it is and what it wants to (01:25:13) do going forward as opposed to say hey (01:25:15) we have a point of view that doesn't (01:25:16) respect your view. M so ju just to make (01:25:18) sure I understand the idea here is each (01:25:23) country will want some kind of data (01:25:25) residency privacy etc. And Microsoft is (01:25:27) especially privileged here because you (01:25:29) have relationships with these countries (01:25:31) who have expertise in setting up these (01:25:33) kinds of sovereign data centers and (01:25:36) therefore Microsoft is uniquely fit for (01:25:38) a world with um more sovereignty (01:25:41) requirements. (01:25:42) >> Yeah. I mean I I I don't want to sort of (01:25:44) describe it as somehow we're uniquely (01:25:46) privileged. Uh I would just say I think (01:25:48) of that as a business requirement that (01:25:50) we have been doing all the hard work all (01:25:52) these decades and we plan to and so my (01:25:55) answer to Dylan's previous question was (01:25:57) I take these you know whether it's in (01:26:00) the United States quite frankly uh when (01:26:03) you know when the white house and the (01:26:05) USG says hey we want you to allocate (01:26:08) more of your I don't know wafer starts (01:26:11) to uh uh fabs in the US we take that (01:26:16) seriously. (01:26:16) or whether it is data center and the EU (01:26:19) boundary, we take that seriously. So to (01:26:21) me, (01:26:22) >> um respecting what I think are (01:26:25) legitimate reasons why countries care (01:26:27) about sovereignty and building for it as (01:26:30) a software and a physical plant is what (01:26:32) I I would say we are going to do. And as (01:26:35) we go to like the bipolar world, right, (01:26:37) US, China. Yeah. Um there is there is a (01:26:40) lot around, (01:26:42) >> you know, American tech does not, you (01:26:43) know, it's not just you versus Amazon, (01:26:45) um or you versus, you know, anthropic or (01:26:47) you versus Google. Yeah. There is a (01:26:49) whole host of competit competition. How (01:26:52) does how does America rebuild the trust? (01:26:54) What do you do to rebuild the trust to (01:26:56) say actually no, American companies will (01:26:58) be the main provider for you? Um and how (01:27:01) do you think about competition with up (01:27:02) and cominging Chinese companies whether (01:27:04) it be you know bite dance and Alibaba or (01:27:06) Deepseek and Moonshot (01:27:07) >> and so just add to the question one (01:27:09) concern is look we're talking about how (01:27:10) AI is becoming this sort of industrial (01:27:12) capex race uh where you're just rapidly (01:27:15) having to build quickly across all those (01:27:17) supply chain when you hear that at least (01:27:19) till now you just think about China (01:27:21) right this is like their comparative (01:27:22) advantage and especially if we're not (01:27:25) going to moonshot to ASI next year but (01:27:28) we it's going to be this decades of (01:27:30) buildouts and infrastructure and so (01:27:32) forth, how do you deal with Chinese (01:27:35) competition? Are they privileged in that (01:27:37) world? (01:27:37) >> Yeah. So, it's a great qu. I mean, in (01:27:38) fact, you just made the point of why I (01:27:41) think trust in American tech is probably (01:27:45) the most important feature. It's not (01:27:48) even the model capability. (01:27:50) Maybe it is like can I trust you the (01:27:54) company? Can I trust you? Your country (01:27:58) and its institutions to be a long-term (01:28:01) supplier may be the thing that wins the (01:28:04) world. (01:28:05) >> That's a good note to end on. (01:28:06) >> Yeah, (01:28:07) >> Satia, thank you for doing this. (01:28:08) >> Thank you so much. Thank you. It's such (01:28:10) a pleasure. Such a pleasure. (01:28:12) >> It's awesome. It's like, man, you two (01:28:14) guys are like quite the team. [laughter] (01:28:17) >> Hey everybody, I hope you enjoyed that (01:28:19) episode. If you did, the most helpful (01:28:21) thing you can do is just share it with (01:28:23) other people who you think might enjoy (01:28:24) it. It's also helpful if you leave a (01:28:27) rating or a comment on whatever platform (01:28:29) you're listening on. If you're (01:28:31) interested in sponsoring the podcast, (01:28:33) you can reach out at (01:28:34) dwarcash.com/advertise. (01:28:38) Otherwise, I'll see you on the next one.