1 00:00:00,800 --> 00:00:03,677 Especially in technology business, we 2 00:00:03,760 --> 00:00:06,237 collectively as an industry have to just 3 00:00:06,320 --> 00:00:08,717 mature and mature at a very fast rate 4 00:00:08,800 --> 00:00:10,797 because the impact of digital technology 5 00:00:10,880 --> 00:00:14,317 in our lives and in our society is so 6 00:00:14,400 --> 00:00:16,877 deep. Now the first thing that we have 7 00:00:16,960 --> 00:00:19,757 been doing is building the system all 8 00:00:19,840 --> 00:00:22,957 the way from power to the DC to the rack 9 00:00:23,040 --> 00:00:29,240 to the accelerators to the network. You have to create magic. 10 00:00:35,600 --> 00:00:38,477 Welcome to the stage Satia Nadella. 11 00:00:38,560 --> 00:00:41,597 >> Hey Satya, thank you so much for doing this. I know last minute schedule 12 00:00:41,680 --> 00:00:44,077 changes. You can be here in person. 13 00:00:44,160 --> 00:00:45,837 >> No, not at all. It's my pleasure Anel. 14 00:00:45,920 --> 00:00:50,317 It's great to be with you. >> Yeah. Right before this, we uh had an 15 00:00:50,400 --> 00:00:54,237 opportunity to hear from Bob Mechaf about buildouts throughout the 80s and 16 00:00:54,320 --> 00:00:56,637 90s. So you've got a tall order. I know 17 00:00:56,720 --> 00:00:59,837 you admire Bob as much as I do. 18 00:00:59,920 --> 00:01:03,997 U so so may maybe one place to start is 19 00:01:04,080 --> 00:01:07,597 you know there's so much buildout happening now. What do you think will 20 00:01:07,680 --> 00:01:10,957 make or break the next decade of this infrastructure buildout that's happening 21 00:01:11,040 --> 00:01:13,277 now? 22 00:01:13,360 --> 00:01:17,197 >> Yeah. I mean the way I think um analis 23 00:01:17,280 --> 00:01:20,077 each of these buildouts um of 24 00:01:20,160 --> 00:01:22,397 infrastructure always comes down to two 25 00:01:22,480 --> 00:01:25,997 sets of things right. how efficient is 26 00:01:26,080 --> 00:01:28,397 the buildout for what is the new 27 00:01:28,480 --> 00:01:33,117 platform uh and then what's the way it gets to 28 00:01:33,200 --> 00:01:36,557 ubiquity in some sense it's the app server right I've always thought about 29 00:01:36,640 --> 00:01:39,917 at least from a a platform perspective 30 00:01:40,000 --> 00:01:41,997 what's the core primitives of the 31 00:01:42,080 --> 00:01:46,957 platform and then how does the platform have a developer interface so that 32 00:01:47,040 --> 00:01:50,157 developers can use it right I mean at the end of the day you could say that's 33 00:01:50,240 --> 00:01:54,157 what operating systems are right so the operating system book at least I read in 34 00:01:54,240 --> 00:01:57,917 school was that tenant bomb thing which is you abstract the hardware and expose 35 00:01:58,000 --> 00:02:01,997 an app model and that's a good way to I think think about the infrastructure so 36 00:02:02,080 --> 00:02:05,757 in that context the first thing is how do you make things efficient this time 37 00:02:05,840 --> 00:02:07,757 around it's that tokens per dollar per 38 00:02:07,840 --> 00:02:11,197 watt right so the token factory 39 00:02:11,280 --> 00:02:13,197 um has to be super efficient so there's 40 00:02:13,280 --> 00:02:15,197 a lot of work that goes on I think is 41 00:02:15,280 --> 00:02:18,557 going on is at the systems level uh all 42 00:02:18,640 --> 00:02:22,077 the way to power uh right uh and so 43 00:02:22,160 --> 00:02:26,477 that's one side of it. Then on the app server side, it's more like an agent 44 00:02:26,560 --> 00:02:30,397 factory. So if you have a token factory, so think of it just as throughput's got 45 00:02:30,480 --> 00:02:33,437 to be most efficient. Uh but then you 46 00:02:33,520 --> 00:02:36,637 got to harness that to build very 47 00:02:36,720 --> 00:02:40,477 efficient effective agents and that is 48 00:02:40,560 --> 00:02:42,557 an entire different ball of wax, right? 49 00:02:42,640 --> 00:02:45,117 which is you got to have a set of 50 00:02:45,200 --> 00:02:47,597 application services so that any agent 51 00:02:47,680 --> 00:02:51,117 that you're building uh delivers on some 52 00:02:51,200 --> 00:02:53,917 eval metric that matters in the world 53 00:02:54,000 --> 00:02:56,317 and so those are the two at least I 54 00:02:56,400 --> 00:02:59,037 think as the infrastructure pieces. 55 00:02:59,120 --> 00:03:02,557 >> Yeah, I think you have such a engineer point of view of this and I I remember 56 00:03:02,640 --> 00:03:04,797 you mentioning the sort of complete 57 00:03:04,880 --> 00:03:06,877 thought from silicon to operating 58 00:03:06,960 --> 00:03:09,837 systems to applications. um how do you 59 00:03:09,920 --> 00:03:11,997 think that cascades down to the core 60 00:03:12,080 --> 00:03:14,397 part of how networks are designed in all 61 00:03:14,480 --> 00:03:19,277 this buildout and across the WAN and the edge and how do you think about that? 62 00:03:19,360 --> 00:03:22,637 >> Yeah, I mean it's it's fascinating. I mean obviously what you all are doing uh 63 00:03:22,720 --> 00:03:26,557 in fact the first time I remember you and I chatted when you said you're going 64 00:03:26,640 --> 00:03:30,717 to vertically integrate build this network as a service platform. It's 65 00:03:30,800 --> 00:03:34,477 pretty impressive to sort of say that's I think a little bit of what we are also 66 00:03:34,560 --> 00:03:37,997 doing right when people say wow there's a lot of capex being spent. I mean the 67 00:03:38,080 --> 00:03:42,717 reality is the data center is the computer is becoming more real today 68 00:03:42,800 --> 00:03:44,877 right I mean that's kind of where you 69 00:03:44,960 --> 00:03:47,677 want to really think about uh uh the 70 00:03:47,760 --> 00:03:52,237 complete thought is how do you really go deliver on that tokens per dollar per 71 00:03:52,320 --> 00:03:54,637 watt uh but just take one of the new 72 00:03:54,720 --> 00:03:57,757 data centers we just brought live uh 73 00:03:57,840 --> 00:04:00,077 last week or a week before uh called 74 00:04:00,160 --> 00:04:03,197 fairwater uh I mean interesting enough 75 00:04:03,280 --> 00:04:07,357 we built it as a double-decka uh data center this is the first time we did 76 00:04:07,440 --> 00:04:09,597 that uh and we did that because to in 77 00:04:09,680 --> 00:04:11,757 fact for the network which is can you 78 00:04:11,840 --> 00:04:14,877 have the 5 million network connections 79 00:04:14,960 --> 00:04:17,037 uh that are there in that particular 80 00:04:17,120 --> 00:04:20,637 data center optimized for cable length 81 00:04:20,720 --> 00:04:22,637 uh right because we wanted the most 82 00:04:22,720 --> 00:04:25,677 dense GPU environment uh and so in order 83 00:04:25,760 --> 00:04:27,837 to do that uh you kind of had to even 84 00:04:27,920 --> 00:04:30,717 think rethink the construction uh if you 85 00:04:30,800 --> 00:04:33,517 will um and so to me and then not only 86 00:04:33,600 --> 00:04:37,757 did we do that in one data center but we also built an AI van And I was sort of 87 00:04:37,840 --> 00:04:41,917 very pleased to see you all working with Lumen and building your own sort of 88 00:04:42,000 --> 00:04:45,037 connection between what you're doing and what they're doing. In fact, that's kind 89 00:04:45,120 --> 00:04:49,517 of a similar thing we did. We now have uh we have a Wisconsin Fairwater and 90 00:04:49,600 --> 00:04:51,997 Atlanta Fairwater and both of them are 91 00:04:52,080 --> 00:04:54,157 sort oforked through this AI van. So you 92 00:04:54,240 --> 00:04:56,877 can now run distributed training across 93 00:04:56,960 --> 00:05:01,437 these. So to me that's I think what you kind of take the workload. I worked for 94 00:05:01,520 --> 00:05:05,757 this gentleman who at Microsoft uh called Paul Merittz long time ago. He 95 00:05:05,840 --> 00:05:09,597 used to always teach us as you know systems folks growing up saying look at 96 00:05:09,680 --> 00:05:13,277 the end of the day uh you really want to fall in love with sort of the what the 97 00:05:13,360 --> 00:05:17,357 new workload is and then design infrastructure uh and application 98 00:05:17,440 --> 00:05:20,477 servers. So that's what we're doing. You look at these new data parallel 99 00:05:20,560 --> 00:05:24,157 workloads which kind of are unlike anything we've seen before a training 100 00:05:24,240 --> 00:05:26,637 inference data gen data pipelines for 101 00:05:26,720 --> 00:05:29,357 all these evalu rubrics what have you 102 00:05:29,440 --> 00:05:32,397 you go build for that and that's where I 103 00:05:32,480 --> 00:05:34,157 think the infrastructure is being 104 00:05:34,240 --> 00:05:39,437 rethought end to end yeah is here somewhere in the audience but I I 105 00:05:39,520 --> 00:05:41,037 watched the interview you did with him 106 00:05:41,120 --> 00:05:45,037 and the scale of the buildout um how do 107 00:05:45,120 --> 00:05:49,117 you feel as an engineer like these are masscale projects X how do you think 108 00:05:49,200 --> 00:05:53,997 about this and you know as an engineer I know like you understand these systems 109 00:05:54,080 --> 00:05:57,917 are you in awe just as much as the rest of us are to look at the scale of this 110 00:05:58,000 --> 00:06:00,317 >> yeah I mean I must say we're in awe of 111 00:06:00,400 --> 00:06:04,397 it but I'm also quite frankly humbled by 112 00:06:04,480 --> 00:06:09,837 one fact Anel which is uh the amount of 113 00:06:09,920 --> 00:06:11,757 expertise that is required to pull 114 00:06:11,840 --> 00:06:15,197 something like this off right um if you 115 00:06:15,280 --> 00:06:19,277 think about the electrical engineers the mechanical engineers, the construction 116 00:06:19,360 --> 00:06:23,837 civil engineers, the obviously the computer engineers, all of them coming 117 00:06:23,920 --> 00:06:27,757 together, composing on a project timeline, making something like this 118 00:06:27,840 --> 00:06:30,717 happen, delivering on an SLA. Uh so to 119 00:06:30,800 --> 00:06:34,317 some degree uh it is what I'm in awe of 120 00:06:34,400 --> 00:06:37,997 and I I also feel um that this is one of 121 00:06:38,080 --> 00:06:40,477 those places where we've got to get it 122 00:06:40,560 --> 00:06:43,757 right that not only do we have to build 123 00:06:43,840 --> 00:06:46,157 stuff that works and works super 124 00:06:46,240 --> 00:06:50,717 efficiently. Uh but it's also to do what 125 00:06:50,800 --> 00:06:52,797 uh I think that's the main thing as 126 00:06:52,880 --> 00:06:55,677 engineers. uh because for the first time 127 00:06:55,760 --> 00:06:58,477 as I mean when I say uh tokens per 128 00:06:58,560 --> 00:07:00,957 dollar per watt if you're really going 129 00:07:01,040 --> 00:07:03,917 to consume a lot of energy uh one thing 130 00:07:04,000 --> 00:07:08,397 that I think we all in this era have to be more grounded on is the social 131 00:07:08,480 --> 00:07:12,397 permission uh to go take that type of energy 132 00:07:12,480 --> 00:07:17,037 convert it into tokens is not going to be about celebrating like some eal that 133 00:07:17,120 --> 00:07:19,277 you and I can get hyped up about. It has 134 00:07:19,360 --> 00:07:24,557 to be something that the world actually benefits from. you guys are building out 135 00:07:24,640 --> 00:07:28,477 the entire stack like you mentioned which is amazing to watch. How do you 136 00:07:28,560 --> 00:07:32,797 think you maintain that while also thinking about reliability and trust 137 00:07:32,880 --> 00:07:36,237 because so many companies depend on what you guys are doing and how much you're 138 00:07:36,320 --> 00:07:40,317 building up? >> I mean that's a you know sort of when I 139 00:07:40,400 --> 00:07:42,477 think about my you know priorities for 140 00:07:42,560 --> 00:07:47,437 us quite frankly there are three uh at the company. One is security 141 00:07:47,520 --> 00:07:51,757 uh because the cyber landscape just becomes even more complex every day. uh 142 00:07:51,840 --> 00:07:56,157 and then quality and reliability and if you do a good job on those two you get 143 00:07:56,240 --> 00:08:00,637 sort of permission to even think about AI that's kind of I'm pretty clear that 144 00:08:00,720 --> 00:08:04,717 that's the sequence the order of operations even for us because uh you 145 00:08:04,800 --> 00:08:07,357 know if I look at uh Azure today in the 146 00:08:07,440 --> 00:08:11,997 60 plus regions around the world uh people count on it uh as critical 147 00:08:12,080 --> 00:08:15,677 infrastructure and so therefore we don't take it for granted and so you have to 148 00:08:15,760 --> 00:08:20,717 think about uh operating it uh with those qualities and the fundamentals and 149 00:08:20,800 --> 00:08:22,477 that I think is the key right which is 150 00:08:22,560 --> 00:08:25,277 these are um at scale nothing is a 151 00:08:25,360 --> 00:08:27,837 commodity it's sort of very complex and 152 00:08:27,920 --> 00:08:30,477 so therefore what and in fact if if you 153 00:08:30,560 --> 00:08:33,837 think about it the primitives have to be 154 00:08:33,920 --> 00:08:36,157 rethought right for the next 10x 155 00:08:36,240 --> 00:08:39,277 everything right um uh and so what you 156 00:08:39,360 --> 00:08:41,917 built for like we added more data center 157 00:08:42,000 --> 00:08:45,517 capacity in the last 18 months uh than 158 00:08:45,600 --> 00:08:49,997 perhaps in the first 10 years of Azure right so if you sort of say that's the 159 00:08:50,080 --> 00:08:53,517 scale uh then some of the core primitives uh 160 00:08:53,600 --> 00:08:58,157 have to be rebuilt. Uh but then on top of it, there's one other thing that you 161 00:08:58,240 --> 00:09:02,237 brought up which I think is super important which is operating around the 162 00:09:02,320 --> 00:09:04,957 world requires us to respect what I 163 00:09:05,040 --> 00:09:07,597 think of as the legitimate sovereignty 164 00:09:07,680 --> 00:09:11,997 needs uh people have right because in Europe they have their own data 165 00:09:12,080 --> 00:09:15,597 protection laws. So therefore, for example, one of the hardest projects we 166 00:09:15,680 --> 00:09:18,477 did over the last five years is the EU 167 00:09:18,560 --> 00:09:21,117 data boundary. uh right which is how do 168 00:09:21,200 --> 00:09:25,837 you make sure that all the data that uh is European remains in Europe right that 169 00:09:25,920 --> 00:09:29,677 so that's the data plane now we have a control plane that goes with it that is 170 00:09:29,760 --> 00:09:32,477 resident um and so we've kind of 171 00:09:32,560 --> 00:09:35,517 continuously built for what I think is a 172 00:09:35,600 --> 00:09:38,957 distributed infrastructure but with real 173 00:09:39,040 --> 00:09:41,917 sovereignty boundaries uh and now with 174 00:09:42,000 --> 00:09:44,637 AI we're even getting to an another new 175 00:09:44,720 --> 00:09:49,277 sovereignty boundary which is at the end of the day every firm wants to know that 176 00:09:49,360 --> 00:09:53,277 their data and their unique understanding of data is being parlayed 177 00:09:53,360 --> 00:09:55,837 into their competitive differentiation 178 00:09:55,920 --> 00:09:58,157 versus just fodder for a model company. 179 00:09:58,240 --> 00:10:02,557 And so that I think is one of the more interesting sovereignty debates right 180 00:10:02,640 --> 00:10:04,477 what's the future of a firm right if the 181 00:10:04,560 --> 00:10:06,797 firm is all about knowledge you have 182 00:10:06,880 --> 00:10:10,557 which is tacet inside uh but if a model 183 00:10:10,640 --> 00:10:13,197 you know can see like you know 200 great 184 00:10:13,280 --> 00:10:17,917 samples of data and pick it up then that's a problem so therefore uh you 185 00:10:18,000 --> 00:10:22,557 really need to think about even sovereignty uh at the firm level very 186 00:10:22,640 --> 00:10:25,357 differently >> I remember you saying like the hardest 187 00:10:25,440 --> 00:10:29,917 part is not the models but changing how people think and work with the stuff. 188 00:10:30,000 --> 00:10:33,917 What have you guys seen work best? What are you seeing from your customers and 189 00:10:34,000 --> 00:10:36,957 how does that work? Because you know, everybody has a different perception of 190 00:10:37,040 --> 00:10:40,237 how well these things work, how they don't or whether they're creative or 191 00:10:40,320 --> 00:10:44,557 not. How what have you seen in that change of in how people think and work 192 00:10:44,640 --> 00:10:46,557 with these models? 193 00:10:46,640 --> 00:10:50,157 >> Yeah. I mean I think the 194 00:10:50,240 --> 00:10:54,237 the first thing that is interesting uh that I see when somebody you know 195 00:10:54,320 --> 00:10:58,557 besides sort of let's say the hit app right whether it is chat GPT or copilot 196 00:10:58,640 --> 00:11:00,717 that you can deploy use and I mean you 197 00:11:00,800 --> 00:11:03,757 know if you go get beyond any firstparty 198 00:11:03,840 --> 00:11:08,797 stuff that we're doing right even GitHub copilot or what have you uh for an 199 00:11:08,880 --> 00:11:10,317 enterprise 200 00:11:10,400 --> 00:11:12,877 they need to start with I think a little 201 00:11:12,960 --> 00:11:16,637 bit of a change in mindset which is um 202 00:11:16,720 --> 00:11:19,917 we built all the workflows and uh at 203 00:11:20,000 --> 00:11:25,117 work uh based on a particular regime of digital infrastructure. I mean the 204 00:11:25,200 --> 00:11:28,717 reality is nothing much has changed since the be you know when I started in 205 00:11:28,800 --> 00:11:31,597 computing in terms of the workflow right 206 00:11:31,680 --> 00:11:35,757 I mean we had PCs uh and then talking to 207 00:11:35,840 --> 00:11:40,397 minis and then PCs started talking to servers and it became client servers 208 00:11:40,480 --> 00:11:44,557 right that's sort of what uh was the last time people said oh we'll have 209 00:11:44,640 --> 00:11:46,877 standard issues of office productivity 210 00:11:46,960 --> 00:11:49,437 we will have email and it'll replace fax 211 00:11:49,520 --> 00:11:51,917 machines and inter office memos and we 212 00:11:52,000 --> 00:11:54,557 will now be able to do a forecast in 213 00:11:54,640 --> 00:11:57,197 Excel, put in an email, send it around, 214 00:11:57,280 --> 00:11:59,437 people can enter numbers and you know, 215 00:11:59,520 --> 00:12:02,237 so there it went and oh, we'll buy an 216 00:12:02,320 --> 00:12:06,477 ERP system, a CRM system. We'll have a system of record. So, we kind of build a 217 00:12:06,560 --> 00:12:08,237 pretty sophisticated, I'll call a 218 00:12:08,320 --> 00:12:12,077 digital information back plane and now 219 00:12:12,160 --> 00:12:15,837 it's time to relitigate the entirety of it, right? because you kind of say, 220 00:12:15,920 --> 00:12:17,837 okay, this is how work happens today, 221 00:12:17,920 --> 00:12:21,517 but I need a mindset shift on if a mo a 222 00:12:21,600 --> 00:12:23,837 I have multiple agents 223 00:12:23,920 --> 00:12:26,317 working alongside all of the human 224 00:12:26,400 --> 00:12:29,277 workflow with digital systems. Uh how do 225 00:12:29,360 --> 00:12:31,917 I change the process uh of something 226 00:12:32,000 --> 00:12:36,797 like even forecasting? Um and I think that that's the first one. The second 227 00:12:36,880 --> 00:12:40,637 thing is you need a new tool chain, right? So you kind of need uh just like 228 00:12:40,720 --> 00:12:42,957 how the tools of the past were or I had 229 00:12:43,040 --> 00:12:45,197 Word or I had Excel. what are the new 230 00:12:45,280 --> 00:12:48,397 tools to be able to build agents uh and 231 00:12:48,480 --> 00:12:52,797 you know deploy agents, monitor agents, steer agents in fact so there's a 232 00:12:52,880 --> 00:12:54,717 complete new tool chain uh that is 233 00:12:54,800 --> 00:12:57,917 required um then that that requires us 234 00:12:58,000 --> 00:13:02,077 to develop skills not like the elite sort of AI scientist skill like it's 235 00:13:02,160 --> 00:13:04,237 sort of yeah we may have taken a class 236 00:13:04,320 --> 00:13:07,197 or two of Excel macros uh but we all 237 00:13:07,280 --> 00:13:09,757 could sort of use Excel though so what's 238 00:13:09,840 --> 00:13:14,317 I think everybody's going to be building agents like they did a spreadsheet need 239 00:13:14,400 --> 00:13:18,557 uh and so that's the skill set that needs to be broadly diffused and then 240 00:13:18,640 --> 00:13:21,037 the last thing is data because if you 241 00:13:21,120 --> 00:13:25,677 think about it the most interesting thing that we do is to do this context 242 00:13:25,760 --> 00:13:29,837 engineering which is how how do I make sure that the data meets the 243 00:13:29,920 --> 00:13:33,597 intelligence that meets the process change I want to affect right it's not 244 00:13:33,680 --> 00:13:37,917 like data defined in the silos I have because that's kind of like you have a 245 00:13:38,000 --> 00:13:41,677 very powerful intelligence in a reasoning engine but it's only as good 246 00:13:41,760 --> 00:13:44,477 as the data you feed it uh in terms of 247 00:13:44,560 --> 00:13:46,637 what it should reason and its action 248 00:13:46,720 --> 00:13:49,437 space and so I think these four or five 249 00:13:49,520 --> 00:13:52,477 things have to fall in place and that's 250 00:13:52,560 --> 00:13:57,277 the hard change management uh that we have and it's sort of we have 251 00:13:57,360 --> 00:14:01,517 to compress what hap think about it right all through the 90s and the 2000s 252 00:14:01,600 --> 00:14:03,677 you could say even in the cloud like 253 00:14:03,760 --> 00:14:05,597 what did we do mostly we moved the same 254 00:14:05,680 --> 00:14:09,037 systems to the cloud so quite frankly we 255 00:14:09,120 --> 00:14:12,877 had a 30-year run which was pretty stable 256 00:14:12,960 --> 00:14:17,117 in terms of how we thought about work getting done and now we suddenly have to 257 00:14:17,200 --> 00:14:19,437 take that 30 years and maybe compress it 258 00:14:19,520 --> 00:14:21,277 you know in the next three four years 259 00:14:21,360 --> 00:14:24,157 five years into something that yields 260 00:14:24,240 --> 00:14:28,797 massive productivity gains. >> You all started working on Azure 10 15 261 00:14:28,880 --> 00:14:33,037 years ago and a lot of the fundamentals that were set that led to being able to 262 00:14:33,120 --> 00:14:37,197 build these large models. If you knew that this is how much buildout was going 263 00:14:37,280 --> 00:14:40,157 to happen, is there some things you would change on how you would do it 264 00:14:40,240 --> 00:14:44,397 differently? Like if you knew, you know, in 2025, this is the mass scale buildout 265 00:14:44,480 --> 00:14:47,277 that's going to happen to the level of, you know, national projects that hadn't 266 00:14:47,360 --> 00:14:51,037 been done before. Are there things you would do differently with that benefit 267 00:14:51,120 --> 00:14:53,117 of hindsight? 268 00:14:53,200 --> 00:14:54,637 >> Yeah, I mean the so the interesting 269 00:14:54,720 --> 00:14:58,477 thing is we did have, you know, 270 00:14:58,560 --> 00:15:00,397 basically the the cloud buildout was 271 00:15:00,480 --> 00:15:03,917 about making servers, right? because um 272 00:15:04,000 --> 00:15:08,397 much more you know it was much more elastic than any of us thought right I 273 00:15:08,480 --> 00:15:12,957 mean virtual machines uh there was insatiable demand for it right people 274 00:15:13,040 --> 00:15:17,517 wanted to just spin up a lot more compute uh one of the other things I had 275 00:15:17,600 --> 00:15:19,757 not realized quite frankly is the amount 276 00:15:19,840 --> 00:15:22,717 of I'll call it uh capital that was 277 00:15:22,800 --> 00:15:25,997 spent on say storage systems underneath 278 00:15:26,080 --> 00:15:31,037 virtual machines uh which suddenly now you could deploy a lot more efficiently 279 00:15:31,120 --> 00:15:34,557 and so that I think created And then of course the new workloads 280 00:15:34,640 --> 00:15:38,877 like the Hadoops and the you know the the data parallel workloads of that day 281 00:15:38,960 --> 00:15:41,277 more batch processing stuff that work 282 00:15:41,360 --> 00:15:44,397 has done at scale this time around I 283 00:15:44,480 --> 00:15:46,397 think it's kind of a new infrastructure 284 00:15:46,480 --> 00:15:49,517 right which is I feel this we've not 285 00:15:49,600 --> 00:15:52,477 seen anything as synchronous data 286 00:15:52,560 --> 00:15:56,877 parallel stuff like this and so therefore that's why I think even the 287 00:15:56,960 --> 00:15:58,557 shape of what we're building we've 288 00:15:58,640 --> 00:16:02,477 adjusted what we were doing to build for 289 00:16:02,560 --> 00:16:06,717 AI. But now when I look at even the way 290 00:16:06,800 --> 00:16:09,357 I think Azure will look five years from 291 00:16:09,440 --> 00:16:11,517 now, it'll look unlike anything we've 292 00:16:11,600 --> 00:16:14,797 did in the first 20 years of Azure. Um 293 00:16:14,880 --> 00:16:19,037 and so the you kind of reorient your 294 00:16:19,120 --> 00:16:24,157 core compute and storage and the network um or even by the way the memory 295 00:16:24,240 --> 00:16:27,517 interface. I think one of the more interesting things that may even happen 296 00:16:27,600 --> 00:16:30,557 is what's going to be uh the model 297 00:16:30,640 --> 00:16:33,117 architecture, the micro architecture in 298 00:16:33,200 --> 00:16:36,957 the AI accelerator and the memory. Uh 299 00:16:37,040 --> 00:16:39,437 that's the thing that perhaps is going 300 00:16:39,520 --> 00:16:43,197 to completely change. Um and so yes, so 301 00:16:43,280 --> 00:16:46,477 I think my short answer is we have to 302 00:16:46,560 --> 00:16:47,997 rethink 303 00:16:48,080 --> 00:16:51,037 uh the core footprint of what one 304 00:16:51,120 --> 00:16:54,717 considers a cloud. um knowing at one end 305 00:16:54,800 --> 00:16:58,237 I have to run an SAP workload at 59s uh 306 00:16:58,320 --> 00:17:00,717 and on the other end I have to run a 307 00:17:00,800 --> 00:17:03,277 highly efficient 308 00:17:03,360 --> 00:17:06,397 uh data parallel training job u right 309 00:17:06,480 --> 00:17:10,877 which gets has the maximum token throughput and so balancing that out in 310 00:17:10,960 --> 00:17:14,797 the right way I think is going to be one of the fundamental challenges for us 311 00:17:14,880 --> 00:17:17,837 >> in the last few buildouts one of the interesting things that happened is 312 00:17:17,920 --> 00:17:20,157 after this buildout starts maybe in the 313 00:17:20,240 --> 00:17:23,437 80s then in the 90s um things start to 314 00:17:23,520 --> 00:17:27,357 standardize. Today, if you go look at data centers and other places, every 315 00:17:27,440 --> 00:17:32,157 company is doing interesting things that they hold true themselves, but others 316 00:17:32,240 --> 00:17:35,517 might be doing different things than them. When do you think it'll be more 317 00:17:35,600 --> 00:17:39,517 standard on how to build this type of infrastructure where we have to 318 00:17:39,600 --> 00:17:43,117 standardize and then maybe follow up to that is what do you think should be 319 00:17:43,200 --> 00:17:45,197 standardized or is it the way we're 320 00:17:45,280 --> 00:17:47,117 doing it now where everybody does their 321 00:17:47,200 --> 00:17:50,237 own proprietary thing is the way to go? 322 00:17:50,320 --> 00:17:52,637 Yeah, I mean it it's it's a good point 323 00:17:52,720 --> 00:17:55,117 because I think the specialization will 324 00:17:55,200 --> 00:17:58,157 come as the supply chain and the value 325 00:17:58,240 --> 00:18:02,877 chain here sort of really achieves scale right I mean if I look at it um it's 326 00:18:02,960 --> 00:18:07,117 sort you know what I mean if you take the the the classic argument right which 327 00:18:07,200 --> 00:18:09,197 is if you have scale you vertically 328 00:18:09,280 --> 00:18:12,717 integrate it can be very t you know very 329 00:18:12,800 --> 00:18:16,477 efficient very tight to sort of the idiosyncrasies of sort of how you think 330 00:18:16,560 --> 00:18:20,477 about whether it's your model or your application scaffolding and its 331 00:18:20,560 --> 00:18:24,797 requirements. Uh but for us as a platform provider, I mean like at this 332 00:18:24,880 --> 00:18:27,357 point we have to make sure that any new 333 00:18:27,440 --> 00:18:30,237 thing that gets written runs on Azure. 334 00:18:30,320 --> 00:18:34,797 Um and so therefore uh even at the silicon accelerator level we have to be 335 00:18:34,880 --> 00:18:38,237 heterogeneous, right? I mean at the end of the day the workloads are going to be 336 00:18:38,320 --> 00:18:40,157 uh heterogeneous. This the system 337 00:18:40,240 --> 00:18:43,037 architecture underneath those has to 338 00:18:43,120 --> 00:18:47,277 support right whether it's Nvidia, AMD and our own accelerators or what have 339 00:18:47,360 --> 00:18:49,677 you. uh and so I think that that's how 340 00:18:49,760 --> 00:18:51,277 we are going to have to standardize 341 00:18:51,360 --> 00:18:54,477 which is to basically build out uh what 342 00:18:54,560 --> 00:18:57,757 is a way to meet the real world needs um 343 00:18:57,840 --> 00:19:00,557 which by definition is going to be 344 00:19:00,640 --> 00:19:02,717 heterogeneous and therefore what's the 345 00:19:02,800 --> 00:19:05,277 set of standards that emerge I think the 346 00:19:05,360 --> 00:19:08,557 first place is networking uh if I look 347 00:19:08,640 --> 00:19:13,597 at the scale up and scale out um network 348 00:19:13,680 --> 00:19:16,477 for these new workloads 349 00:19:16,560 --> 00:19:19,677 uh all coming together may be the place 350 00:19:19,760 --> 00:19:21,197 right because I think that you know the 351 00:19:21,280 --> 00:19:24,157 model has still and the model evolution 352 00:19:24,240 --> 00:19:27,437 has still a lot more um influence on 353 00:19:27,520 --> 00:19:31,597 I'll call it the micro architecture of the one chiplet in the AI you know 354 00:19:31,680 --> 00:19:34,157 accelerator for sure uh and its 355 00:19:34,240 --> 00:19:35,997 connection to memory so some of that 356 00:19:36,080 --> 00:19:39,757 stuff may still be more fastm moving but 357 00:19:39,840 --> 00:19:42,077 everything else quite frankly you can 358 00:19:42,160 --> 00:19:45,757 standardize um u and so that I think is 359 00:19:45,840 --> 00:19:50,317 where some of the exciting uh work by the way and it has to go all 360 00:19:50,400 --> 00:19:53,757 the way to the edge as well and I think that this is an obviously a massive 361 00:19:53,840 --> 00:19:58,157 thing for you all to participate in and we're excited about sort of being able 362 00:19:58,240 --> 00:19:59,677 to sort of say at the end of the day 363 00:19:59,760 --> 00:20:01,597 what's this managed service around 364 00:20:01,680 --> 00:20:05,437 networking uh that spans uh the entirety 365 00:20:05,520 --> 00:20:09,917 of the distributed computing fabric and that requires uh what you are doing as 366 00:20:10,000 --> 00:20:14,957 well >> and there's so many people here in the audience that are CIOS CTO's that are 367 00:20:15,040 --> 00:20:18,717 figuring out how to think about all these new tools that are coming out. 368 00:20:18,800 --> 00:20:23,357 What where have you seen the biggest ROI in Microsoft? Um because you guys are 369 00:20:23,440 --> 00:20:26,797 obviously at the forefront. You have access to everything. How have those 370 00:20:26,880 --> 00:20:29,517 workflows changed the most? Yeah, I mean 371 00:20:29,600 --> 00:20:31,437 I I mean one of the fascinating things 372 00:20:31,520 --> 00:20:34,877 is the best news now for me Anil is when 373 00:20:34,960 --> 00:20:39,757 I start you know hearing of the ground up use of the tools right whether it's 374 00:20:39,840 --> 00:20:43,837 copilot whether it's foundry whether it's GitHub copilot people using them 375 00:20:43,920 --> 00:20:48,557 right so this is not uh you know some top- down mandate of hey because like 376 00:20:48,640 --> 00:20:52,237 the obvious use cases I think we're all well on our way right to customer 377 00:20:52,320 --> 00:20:54,877 service or what have you but the thing 378 00:20:54,960 --> 00:20:57,277 for example on networking I had myself 379 00:20:57,360 --> 00:20:59,357 like lost track of the number of for 380 00:20:59,440 --> 00:21:02,877 example fiber operators uh that power uh 381 00:21:02,960 --> 00:21:05,997 Azure like there are 500 plus uh 382 00:21:06,080 --> 00:21:10,477 operators that we have contracts with all over the world uh on our fiber 383 00:21:10,560 --> 00:21:12,397 operations and I mean at the end of the 384 00:21:12,480 --> 00:21:14,637 day it's a physical asset things happen 385 00:21:14,720 --> 00:21:17,677 there are fiber cuts there's issues and 386 00:21:17,760 --> 00:21:21,037 it's fascinating the devops pipeline 387 00:21:21,120 --> 00:21:25,517 that is there I mean I mean sometimes oh devops seems sophisticated it's not like 388 00:21:25,600 --> 00:21:27,677 GitHub actions I mean the thing is 389 00:21:27,760 --> 00:21:31,997 literally emails going on and and so one 390 00:21:32,080 --> 00:21:34,157 of our engineers what she did was she 391 00:21:34,240 --> 00:21:37,037 took Foundry built herself a bunch of 392 00:21:37,120 --> 00:21:41,357 agents. She's given them all names uh because she she had to there was no 393 00:21:41,440 --> 00:21:43,357 headcount. There was no way even if she 394 00:21:43,440 --> 00:21:45,917 had headcount she could grow a team to 395 00:21:46,000 --> 00:21:48,637 manage what was becoming the most manual 396 00:21:48,720 --> 00:21:51,037 of processes. And so she fundamentally 397 00:21:51,120 --> 00:21:54,717 identified the entire process of dealing 398 00:21:54,800 --> 00:21:58,157 with triages around fiber operations and 399 00:21:58,240 --> 00:22:01,117 maintenance and cuts and what have you. 400 00:22:01,200 --> 00:22:05,277 And to me that was like a great firsthand example. Same thing you know 401 00:22:05,360 --> 00:22:09,677 to your point about supply chain and the data you know these massive buildouts of 402 00:22:09,760 --> 00:22:11,837 our data center. You should see like the 403 00:22:11,920 --> 00:22:14,637 I thought man maybe somebody in uh you 404 00:22:14,720 --> 00:22:19,117 know in finance or in legal will figure out how to do contracts. the people who 405 00:22:19,200 --> 00:22:21,517 needed to really figure out the agentic 406 00:22:21,600 --> 00:22:26,397 way to go get contracts done uh because I did one data center contract here with 407 00:22:26,480 --> 00:22:30,317 this one supplier and replicated everywhere and so on. That's a place 408 00:22:30,400 --> 00:22:35,117 where completely the supply chain operations was completely inverted. I 409 00:22:35,200 --> 00:22:38,877 mean if you think about supply chain the basic way I I've learned is at the end 410 00:22:38,960 --> 00:22:42,317 of the quarter you go to your CFO you get beaten by the CFO for all the 411 00:22:42,400 --> 00:22:45,197 mistakes you made. Uh and the idea now 412 00:22:45,280 --> 00:22:47,517 is to actually have these agents help 413 00:22:47,600 --> 00:22:49,677 with the daily beating so to speak and 414 00:22:49,760 --> 00:22:55,277 so that is uh the fundamental way you can completely transform a business process. 415 00:22:55,360 --> 00:22:57,837 >> Amazing. Thank you so much for your time Sati. I really appreciate you doing 416 00:22:57,920 --> 00:23:03,960 this. >> My pleasure Anel. Thank you so much.