1 00:00:10,040 --> 00:00:13,017 Welcome. Today, we have Satya Nadella. 2 00:00:13,100 --> 00:00:15,637 Thank you so much for coming. Absolutely. 3 00:00:15,720 --> 00:00:19,097 It's my pleasure. And, realizing it's a finals week. 4 00:00:19,180 --> 00:00:22,877 So it's a different time of the day whatnot, 5 00:00:22,960 --> 00:00:27,557 but I appreciate the students that you all made it here. 6 00:00:27,640 --> 00:00:30,957 I was thinking this morning how it's kind of fitting 7 00:00:31,040 --> 00:00:34,817 that you're the last person that we're having to this class. 8 00:00:34,900 --> 00:00:37,417 And I say that because if I think back, 9 00:00:37,500 --> 00:00:42,357 your bet of putting a billion dollars into OpenAI in 2019 10 00:00:42,440 --> 00:00:47,357 feels like that really set the stage for this Cambrian 11 00:00:47,440 --> 00:00:49,337 explosion, if you will, around AI. 12 00:00:49,420 --> 00:00:52,797 So I'm curious to just kick it off like, 13 00:00:52,880 --> 00:00:56,177 what was the thought process to make that bet? 14 00:00:56,260 --> 00:00:59,727 Yeah. I mean, I think it is fascinating to look back 15 00:00:59,810 --> 00:01:04,307 now, six, seven, eight years in some sense. 16 00:01:04,390 --> 00:01:11,390 I think the thing that I feel at least 17 00:01:11,750 --> 00:01:14,987 got me convinced that this was the right thing 18 00:01:15,070 --> 00:01:20,007 to go try at that time is quite frankly, 19 00:01:20,090 --> 00:01:22,387 what I describe as a prepared mind, which 20 00:01:22,470 --> 00:01:28,347 Microsoft's obsession has always been in natural language. 21 00:01:28,430 --> 00:01:32,627 And of course, we were mostly focused 22 00:01:32,710 --> 00:01:36,467 on trying to get to natural language 23 00:01:36,550 --> 00:01:41,547 with some machine learning, some NLP, 24 00:01:41,630 --> 00:01:46,087 but fundamentally, if you had asked us even in 2017, 2018, 25 00:01:46,170 --> 00:01:52,467 it would be some combination of some symbolic logic plus machine learning. 26 00:01:52,550 --> 00:01:56,027 So we were perhaps at that stage not the truest 27 00:01:56,110 --> 00:02:02,587 believers that deep learning can even get you NLP breakthroughs. 28 00:02:02,670 --> 00:02:06,027 But that was something we wanted to have happen. 29 00:02:06,110 --> 00:02:10,487 So I would take shots. In fact, most people talk about just the OpenAI bet. 30 00:02:10,570 --> 00:02:15,627 But we bought a bunch of companies, have invested in a whole lot of others, 31 00:02:15,710 --> 00:02:18,427 because the fundamental thing we were conditioned to do 32 00:02:18,510 --> 00:02:24,247 was anyone who had an ambitious angle on natural language 33 00:02:24,330 --> 00:02:27,247 irrespective of what sort of lineage they came from, 34 00:02:27,330 --> 00:02:32,427 we would always take it, whether it was organically inside the company or outside. 35 00:02:32,510 --> 00:02:37,867 And that is when Sam-- and in fact, we were one happy family at that time. 36 00:02:37,950 --> 00:02:40,487 So the idea was there. That's what I'm saying. 37 00:02:40,570 --> 00:02:43,267 You really did kick off so many things. 38 00:02:43,350 --> 00:02:45,087 Everyone was in the same place. 39 00:02:45,170 --> 00:02:50,387 And so to some degree, the scaling laws paper came out 40 00:02:50,470 --> 00:02:53,987 and their ambition on pushing the transformer with more 41 00:02:54,070 --> 00:02:59,677 compute and data was an appealing thing for us to take a shot at. 42 00:02:59,760 --> 00:03:01,457 And of course, what's happened. 43 00:03:01,540 --> 00:03:04,057 It's pretty stunning that-- the fact that the capability 44 00:03:04,140 --> 00:03:09,657 graph has just stayed at that scaling law Pareto 45 00:03:09,740 --> 00:03:11,817 is just pretty amazing. 46 00:03:11,900 --> 00:03:16,857 I mean, as someone who worked at Microsoft, I guess 20 years ago, 47 00:03:16,940 --> 00:03:18,997 you've changed the culture like so much. 48 00:03:19,080 --> 00:03:23,897 And I think it's striking to me that when you took that bet, I mean, was there 49 00:03:23,980 --> 00:03:27,517 an uprising within Microsoft saying, hey, we can do this ourselves? 50 00:03:27,600 --> 00:03:31,997 Yeah, I mean, I think Microsoft over the years, 51 00:03:32,080 --> 00:03:35,857 I always say that at the end of the day, the core bet has to be the organic bet 52 00:03:35,940 --> 00:03:40,017 and what one does inside. And then there are partnerships, there's 53 00:03:40,100 --> 00:03:44,397 M&A. And I think any company like-- in some sense when 54 00:03:44,480 --> 00:03:48,577 you grow up in Microsoft, you learn that you can create a lot of enterprise value 55 00:03:48,660 --> 00:03:51,557 by building, by partnering. 56 00:03:51,640 --> 00:03:55,607 If I look back, the PC revolution 57 00:03:55,690 --> 00:03:59,407 wouldn't have been possible, but for what people describe 58 00:03:59,490 --> 00:04:02,767 as the Gates-Grove model, which is Intel, Microsoft coming 59 00:04:02,850 --> 00:04:08,407 together to create essentially what was a PC ecosystem. 60 00:04:08,490 --> 00:04:13,167 I worked on SQL Server, and so what we did with SAP to build our database business 61 00:04:13,250 --> 00:04:15,627 and for them to build the ERP application. 62 00:04:15,710 --> 00:04:18,587 So we are conditioned quite frankly, 63 00:04:18,670 --> 00:04:25,387 for these type of ecosystem partnerships as well as organically build. 64 00:04:25,470 --> 00:04:28,707 And so I would say there was not an uprising in some sense. 65 00:04:28,790 --> 00:04:30,227 They would have always been like, 66 00:04:30,310 --> 00:04:33,807 hey, whenever you allocate your scarce resource of whether it's 67 00:04:33,890 --> 00:04:37,127 capital or more-- in this case, it's more than capital, 68 00:04:37,210 --> 00:04:42,947 the biggest decision was about compute concentration on a particular effort. 69 00:04:43,030 --> 00:04:45,587 I mean, that was more the big bet. 70 00:04:45,670 --> 00:04:51,387 And that's why we made it because this was the group that wanted to go drive it. 71 00:04:51,470 --> 00:04:54,327 And we benefited obviously immensely from it. 72 00:04:54,410 --> 00:04:58,127 And now one of the reasons why he's in town is that there was a big developer 73 00:04:58,210 --> 00:04:59,647 conference called Build. 74 00:04:59,730 --> 00:05:03,307 And yesterday you announced this frontier intelligence ecosystem, 75 00:05:03,390 --> 00:05:05,267 which kind of right in line what you're saying. 76 00:05:05,350 --> 00:05:08,087 And you had a bunch of other pretty huge announcements. 77 00:05:08,170 --> 00:05:09,827 And I'd love to talk about some of those. 78 00:05:09,910 --> 00:05:12,237 So you launched seven new models. 79 00:05:12,320 --> 00:05:14,567 And I thought what was really interesting about at least 80 00:05:14,650 --> 00:05:19,547 Mustafa's description of how you did those models is that all the data is very clean. 81 00:05:19,630 --> 00:05:23,907 There's a lot of focus on, let's say, not breaching any copyright things. 82 00:05:23,990 --> 00:05:26,927 I'm just interested to hear from you why seven? 83 00:05:27,010 --> 00:05:28,307 What was the thought behind that? 84 00:05:28,390 --> 00:05:31,087 Obviously you want your own models, which makes a sense. 85 00:05:31,170 --> 00:05:34,827 Yeah, I think if you step back-- in fact, 86 00:05:34,910 --> 00:05:38,467 I think you call this class the frontiers class. 87 00:05:38,550 --> 00:05:45,247 And so I think one of the challenges of this conceptualizing, how does 88 00:05:45,330 --> 00:05:48,607 anyone, any individual entrepreneur or developer 89 00:05:48,690 --> 00:05:51,517 company, participate at the frontier? 90 00:05:51,600 --> 00:05:53,077 There are frontier models, but how 91 00:05:53,160 --> 00:06:00,160 does one have real agency to add value, derive value, and protect 92 00:06:00,320 --> 00:06:02,077 value? Because that's the question. 93 00:06:02,160 --> 00:06:06,077 If you have a model that basically learns from data, 94 00:06:06,160 --> 00:06:08,117 what's the future of the firm even? 95 00:06:08,200 --> 00:06:11,277 Which is, the firm today is about tacit knowledge 96 00:06:11,360 --> 00:06:14,077 inside the company that comes about 97 00:06:14,160 --> 00:06:17,117 because of its operations and human capital. 98 00:06:17,200 --> 00:06:19,237 And in a world where there is going 99 00:06:19,320 --> 00:06:25,777 to be tokens and humans collaborating together, what's the future of the firm? 100 00:06:25,860 --> 00:06:28,217 So there are some substantially big questions. 101 00:06:28,300 --> 00:06:30,957 And so our vision for this is simple, 102 00:06:31,040 --> 00:06:35,157 which is we believe a frontier ecosystem is 103 00:06:35,240 --> 00:06:40,317 one where every company can actually operate 104 00:06:40,400 --> 00:06:45,537 at the frontier with their own IP compounding over time, 105 00:06:45,620 --> 00:06:48,867 not just the human capital, but even that token capital. 106 00:06:48,950 --> 00:06:52,427 So that is the motivation we have. 107 00:06:52,510 --> 00:06:57,847 So for example, when the models we built-- I'll come back to the lineage. 108 00:06:57,930 --> 00:07:01,587 There's a nice technical report that I would encourage folks 109 00:07:01,670 --> 00:07:03,867 to go read, because I think it's probably 110 00:07:03,950 --> 00:07:07,987 one of the most transparent, good, detailed document 111 00:07:08,070 --> 00:07:12,027 on the entire pipeline that's been written lately 112 00:07:12,110 --> 00:07:16,307 from a model of this size, and I think you'll learn a lot from it. 113 00:07:16,390 --> 00:07:20,227 But the purpose of both, let's take out thinking in coding models, was to be 114 00:07:20,310 --> 00:07:24,387 able to do this in such a way that we can license it along 115 00:07:24,470 --> 00:07:28,867 with the weights and really allow every company 116 00:07:28,950 --> 00:07:30,687 to build their own hill-climbing machine. 117 00:07:30,770 --> 00:07:35,847 So we ourselves climbed our hill using, as you said, 118 00:07:35,930 --> 00:07:38,947 very clean lineage of data, making sure 119 00:07:39,030 --> 00:07:43,347 we were not adding a bunch of synth data in the mix. 120 00:07:43,430 --> 00:07:45,827 So all that was very much true so 121 00:07:45,910 --> 00:07:51,867 that we could truly have a model where reasoning emerged. 122 00:07:51,950 --> 00:07:57,187 And so we now have a fantastic, good, efficient model, 123 00:07:57,270 --> 00:08:02,827 but inside of a hill-climbing machine that any company sets up, it can go 124 00:08:02,910 --> 00:08:07,627 learn using the traces of that company and those tasks. 125 00:08:07,710 --> 00:08:10,427 So our goal is every company starts 126 00:08:10,510 --> 00:08:17,247 thinking strategically about what's the RLE environment that they set up? 127 00:08:17,330 --> 00:08:20,827 What is the private evals that they have? 128 00:08:20,910 --> 00:08:25,407 How do they then welcome any model into that gym, 129 00:08:25,490 --> 00:08:31,907 so to speak, and then allow them to retain the IP and not leak value? 130 00:08:31,990 --> 00:08:35,807 So to me, that's what I think every company will need to start doing. 131 00:08:35,890 --> 00:08:40,587 Because if you're just a consumer of a foundation model, 132 00:08:40,670 --> 00:08:46,877 then I'm not sure how you can retain enterprise value, let alone create. 133 00:08:46,960 --> 00:08:51,677 So the only way I see this ecosystem quite frankly 134 00:08:51,760 --> 00:08:56,697 being non-zero sum or positive sum where lots of participants 135 00:08:56,780 --> 00:09:00,337 can all be at the frontier, is they're 136 00:09:00,420 --> 00:09:03,917 able to take frontier models, take open-weight models, 137 00:09:04,000 --> 00:09:08,857 take a model like ours which is a licensed IP, and then 138 00:09:08,940 --> 00:09:13,557 hill climb on their own environment, and then build out their own IP. 139 00:09:13,640 --> 00:09:15,117 So that's the core premise. 140 00:09:15,200 --> 00:09:19,437 And we unpacked that in a lot of detail and all the tooling around it. 141 00:09:19,520 --> 00:09:21,017 For example, one of the coolest things 142 00:09:21,100 --> 00:09:25,577 is if you're a Microsoft 365 customer, we can bootstrap even. 143 00:09:25,660 --> 00:09:27,417 Because after all, what is Microsoft 365? 144 00:09:27,500 --> 00:09:30,937 Today you use it to run your business. 145 00:09:31,020 --> 00:09:35,417 People communicate with other people related to a business process. 146 00:09:35,500 --> 00:09:40,457 So you can imagine we can bootstrap the RLE. 147 00:09:40,540 --> 00:09:44,477 In fact, we can even generate the evals for, let's say, 148 00:09:44,560 --> 00:09:48,907 an HR onboarding process based on the observation of what you're doing. 149 00:09:48,990 --> 00:09:52,937 That's unique to the company. And first of all, it's their data. 150 00:09:53,020 --> 00:09:56,257 Think of we built a multi-tenant SaaS application. 151 00:09:56,340 --> 00:10:00,337 We now can turn that into a multi-tenant hill-climbing 152 00:10:00,420 --> 00:10:04,457 service where the data and the environment and the models 153 00:10:04,540 --> 00:10:07,937 and the traces and the outcomes are owned by the company. 154 00:10:08,020 --> 00:10:11,547 Do you think that most companies, though, have the right talent to be able to build 155 00:10:11,630 --> 00:10:13,557 those hill-climbing machines? Yeah, it's a great one. 156 00:10:13,640 --> 00:10:15,777 So that's why I think this is the easy button on it. 157 00:10:15,860 --> 00:10:18,237 So we are now not saying you need to build. 158 00:10:18,320 --> 00:10:24,497 So you have the hill-climbing machine that has been instantiated for you. 159 00:10:24,580 --> 00:10:27,377 All you need is a bit of strategic discipline 160 00:10:27,460 --> 00:10:30,677 in making sure that these models, the harness, 161 00:10:30,760 --> 00:10:36,777 the context, the evals are all artifacts and constructs that you understand 162 00:10:36,860 --> 00:10:40,417 and you manage them as assets. 163 00:10:40,500 --> 00:10:45,467 Just like how you have historically done where you cared about privacy, 164 00:10:45,550 --> 00:10:49,207 you cared about confidentiality, you cared about security, 165 00:10:49,290 --> 00:10:53,527 I think in a world where AI comes into your company, 166 00:10:53,610 --> 00:10:57,207 these things will become as important architectural and 167 00:10:57,290 --> 00:11:00,927 strategic considerations. And one of the other products you announced 168 00:11:01,010 --> 00:11:04,147 was Scout around enterprise cloud. 169 00:11:04,230 --> 00:11:07,687 I'm kind of interested to hear the vision thinking behind that. 170 00:11:07,770 --> 00:11:10,327 One of the things that we're very excited about 171 00:11:10,410 --> 00:11:15,927 is when I look at Copilot and its evolution, it started with chat. 172 00:11:16,010 --> 00:11:18,927 And chat became very powerful, especially with reasoning 173 00:11:19,010 --> 00:11:24,727 models, because you could not only just essentially 174 00:11:24,810 --> 00:11:30,227 use it more like search, but you could use it as a thinking assistant essentially. 175 00:11:30,310 --> 00:11:34,907 And so that became powerful. Then Cowork was the next form factor. 176 00:11:34,990 --> 00:11:38,967 And Cowork is pretty neat as a way to delegate tasks. 177 00:11:39,050 --> 00:11:44,517 It's a multi-step reasoning tool-calling agentic loop, 178 00:11:44,600 --> 00:11:50,497 and so therefore you're able to do a short TAS assignment. 179 00:11:50,580 --> 00:11:54,157 It's very much like what we were doing with GitHub Copilot, let's 180 00:11:54,240 --> 00:11:56,697 say, even two years ago when the agent loop started coming. 181 00:11:56,780 --> 00:11:58,917 So we're now doing it for knowledge work. 182 00:11:59,000 --> 00:12:04,817 But now with Scout, you essentially have the third form factor, which is autopilot. 183 00:12:04,900 --> 00:12:11,517 So now you have the long-running agent where it continuously is operating. 184 00:12:11,600 --> 00:12:14,417 It's monitoring. It's got a heartbeat. 185 00:12:14,500 --> 00:12:21,317 It's dreaming. All the things that you expect from a Claw you can now have. 186 00:12:21,400 --> 00:12:24,697 And you can create it. You can have one with your identity. 187 00:12:24,780 --> 00:12:27,977 So essentially, if I have an Entra ID, 188 00:12:28,060 --> 00:12:31,257 I can give Scout my Entra ID as a delegated ID, 189 00:12:31,340 --> 00:12:36,717 and it's essentially my digital twin that's working on my behalf 190 00:12:36,800 --> 00:12:41,397 continuously. But not just that, we can also allow 191 00:12:41,480 --> 00:12:47,197 you to mint more autopilots. And those things can have their own identity 192 00:12:47,280 --> 00:12:51,257 and their own sandboxes. And so it's a pretty complete system. 193 00:12:51,340 --> 00:12:55,397 So I think of it as an enterprise OpenClaw 194 00:12:55,480 --> 00:12:59,377 and a UI that fits in nicely with the rest of the Copilot system. 195 00:12:59,460 --> 00:13:01,657 And it makes sense because you have those identities, 196 00:13:01,740 --> 00:13:04,157 you can address the security question. 197 00:13:04,240 --> 00:13:07,037 I mean, obviously, I don't know how many of you have set up 198 00:13:07,120 --> 00:13:12,257 OpenClaw, but I struggled-- YOLO and get all my credentials because I'm like, 199 00:13:12,340 --> 00:13:16,777 I don't trust it. Yeah, we even announced. 200 00:13:16,860 --> 00:13:19,877 In fact, Peter was on stage with us at Build 201 00:13:19,960 --> 00:13:21,797 as well, because one of the other things 202 00:13:21,880 --> 00:13:24,157 is we were even working with the OpenClaw Foundation 203 00:13:24,240 --> 00:13:28,377 to make sure that it can be run securely. 204 00:13:28,460 --> 00:13:31,797 We will have, in fact, on Windows an out-of-the-box 205 00:13:31,880 --> 00:13:38,547 experience where you can install OpenClaw and have it secured 206 00:13:38,630 --> 00:13:45,387 or contained in what is this new, essentially a container called MXC, 207 00:13:45,470 --> 00:13:48,907 which is essentially a way to sandbox the environment. 208 00:13:48,990 --> 00:13:51,307 And so I think containment is key. 209 00:13:51,390 --> 00:13:54,107 Because after all, you now have these long-running agents 210 00:13:54,190 --> 00:13:57,427 that are able to generate code and execute code. 211 00:13:57,510 --> 00:14:01,207 And so therefore it will become very important to govern the execution. 212 00:14:01,290 --> 00:14:06,047 And so we have a container that then you can set policy and isolation boundaries. 213 00:14:06,130 --> 00:14:09,967 It can be process level isolation, session level isolation. 214 00:14:10,050 --> 00:14:12,467 You can even have a VM boundary if you wanted. 215 00:14:12,550 --> 00:14:16,347 For me for example, if anything I wanted to ever run even, 216 00:14:16,430 --> 00:14:19,427 I'll just run it on Windows 365, which is my cloud instance. 217 00:14:19,510 --> 00:14:25,087 So a complete cloud instance that's fully isolated for long-running agents. 218 00:14:25,170 --> 00:14:30,627 So I think we are all going to learn how to work with many agents. 219 00:14:30,710 --> 00:14:35,677 And we're also going to learn how to isolate the environments for these agents, 220 00:14:35,760 --> 00:14:39,207 just like how back in the day, we thought about processes. 221 00:14:39,290 --> 00:14:41,537 We're going to think about the process boundary, session 222 00:14:41,620 --> 00:14:45,697 boundaries, and container boundaries for agents. 223 00:14:45,780 --> 00:14:48,537 One of the other things that you also announced 224 00:14:48,620 --> 00:14:55,457 was around bringing, we'll say, AI to consumers. 225 00:14:55,540 --> 00:14:59,777 And I'm kind of curious what does that mean. 226 00:14:59,860 --> 00:15:03,897 I mean, there's a lot of big announcements with NVIDIA. 227 00:15:03,980 --> 00:15:05,977 Yeah, I know there are a couple of things on that. 228 00:15:06,060 --> 00:15:12,777 One is we're very excited about this concept of unmetered intelligence. 229 00:15:12,860 --> 00:15:18,857 So if you think about it, every PC historically the install base 230 00:15:18,940 --> 00:15:22,717 had a lot of GPUs. If you count the number of PCs with GPUs, 231 00:15:22,800 --> 00:15:26,007 it's pretty substantial-- the dGPU install base. 232 00:15:26,090 --> 00:15:28,177 So one of the things that we are trying to make sure 233 00:15:28,260 --> 00:15:32,437 is that in a world where these models are there, 234 00:15:32,520 --> 00:15:34,877 there is applications that are being written, 235 00:15:34,960 --> 00:15:37,217 the tokens are in short supply, we 236 00:15:37,300 --> 00:15:42,057 want to tap into essentially the edge compute silicon. 237 00:15:42,140 --> 00:15:45,537 And in that context obviously NVIDIA 238 00:15:45,620 --> 00:15:48,957 announced a new SoC which we are very excited about their RTX. 239 00:15:49,040 --> 00:15:50,617 So we have a Surface Laptop which 240 00:15:50,700 --> 00:15:54,397 is going to come out in the fall, which is built on it. 241 00:15:54,480 --> 00:15:58,697 In fact, all our OEMs will have fantastic designs for it. 242 00:15:58,780 --> 00:16:01,077 We also announced a dev box. I mean, think about it. 243 00:16:01,160 --> 00:16:03,877 It's going to have a petaflop of AI compute. 244 00:16:03,960 --> 00:16:05,777 It's going to have 20 CPU cores, 128 245 00:16:05,860 --> 00:16:12,497 gigabytes of unified memory for both the CPU and the AI compute. 246 00:16:12,580 --> 00:16:19,057 And it's going to run something like a trillion parameter model locally. 247 00:16:19,140 --> 00:16:25,237 And by the way, and we also worked with Jensen to get Windows working on a GB300. 248 00:16:25,320 --> 00:16:27,357 So we even have a DGX workstation. 249 00:16:27,440 --> 00:16:32,247 So I think of it as a data center desktop. 250 00:16:32,330 --> 00:16:35,407 And so I think that there's going to be real demand for all of this 251 00:16:35,490 --> 00:16:38,487 because people will want-- especially when you install 252 00:16:38,570 --> 00:16:41,467 something like Scout or Claw or what have you, 253 00:16:41,550 --> 00:16:46,887 and I want it to just keep working 24 by 7 254 00:16:46,970 --> 00:16:49,407 and I don't want to get billed for it, the best way 255 00:16:49,490 --> 00:16:53,167 to do that is to run it on your laptop or on your desktop. 256 00:16:53,250 --> 00:16:58,607 So we are very excited about just even the rebirth of the existing PC form 257 00:16:58,690 --> 00:17:03,447 factor with this new unbelievable functionality 258 00:17:03,530 --> 00:17:07,207 brought forth because of both the silicon innovation and the model capabilities 259 00:17:07,290 --> 00:17:12,187 that now we can have locally. So that was a lot of what we talked about. 260 00:17:12,270 --> 00:17:14,847 But the other thing that we also said 261 00:17:14,930 --> 00:17:19,207 is just as there's new functionality in the old form 262 00:17:19,290 --> 00:17:22,647 factors, I think there's a real opportunity 263 00:17:22,730 --> 00:17:26,247 to create new form factors for the agent era. 264 00:17:26,330 --> 00:17:28,567 So that's where Project Solara comes in. 265 00:17:28,650 --> 00:17:34,137 And our goal there is to say, we showed two reference designs, 266 00:17:34,220 --> 00:17:40,917 one is a badge and the other one was a desk companion, if you will. 267 00:17:41,000 --> 00:17:42,797 But the badge is pretty interesting. 268 00:17:42,880 --> 00:17:47,857 So you can imagine an agent that has a fingerprint reader 269 00:17:47,940 --> 00:17:51,237 and a badge that has a fingerprint reader as well 270 00:17:51,320 --> 00:17:57,557 as a camera, and has enough onboard compute-- it's a MediaTek processor-- 271 00:17:57,640 --> 00:18:00,717 to be able to wake up something like Copilot. 272 00:18:00,800 --> 00:18:04,797 And I can literally get notified. 273 00:18:04,880 --> 00:18:08,277 Like I can even give it, say, a coding task or whatever, 274 00:18:08,360 --> 00:18:12,197 I can dictate to it, it will take the input 275 00:18:12,280 --> 00:18:16,877 and then go execute it in the cloud, notify me back. 276 00:18:16,960 --> 00:18:22,757 You can imagine in healthcare if I was a nurse, I was moving station to station, I 277 00:18:22,840 --> 00:18:26,857 could use that to badge in the data versus the phone. 278 00:18:26,940 --> 00:18:29,837 Like right now we're conditioned either we're entering in the PC 279 00:18:29,920 --> 00:18:34,677 or we are using the phone. In an agent era where you really have ambient intelligence 280 00:18:34,760 --> 00:18:38,517 and ubiquitous computing, you can imagine these form factors 281 00:18:38,600 --> 00:18:43,077 now that are just endpoints for long-running agents 282 00:18:43,160 --> 00:18:49,797 that wake up, notify, and help you get both output input that's 283 00:18:49,880 --> 00:18:51,177 right there in the real world. 284 00:18:51,260 --> 00:18:55,177 And so we're very excited about bringing even a platform for it. 285 00:18:55,260 --> 00:18:58,477 So we will build some. But the goal here is also to have even by the way 286 00:18:58,560 --> 00:19:02,857 new platform rules. So Windows has always been-- it's 287 00:19:02,940 --> 00:19:06,997 fascinating that we are the only open platform out there. 288 00:19:07,080 --> 00:19:09,417 You can go through our App Store or not. 289 00:19:09,500 --> 00:19:12,017 You can install anything on Windows. 290 00:19:12,100 --> 00:19:15,877 It's always had that ethos of being not something that only 291 00:19:15,960 --> 00:19:18,477 Microsoft-- you don't need to call Microsoft to build 292 00:19:18,560 --> 00:19:20,897 applications for Windows. How about that? 293 00:19:20,980 --> 00:19:25,237 So that's the openness we want even in this new agent platform 294 00:19:25,320 --> 00:19:28,347 so that we don't have the carryover of these platform 295 00:19:28,430 --> 00:19:32,467 rules that were written for the previous era. 296 00:19:32,550 --> 00:19:34,087 I'm going to switch gears a little bit. 297 00:19:34,170 --> 00:19:38,247 So here we're at Stanford University, probably 298 00:19:38,330 --> 00:19:41,747 the center of the world in terms of AI-pilled people. 299 00:19:41,830 --> 00:19:47,707 And when you get outside of the Bay Area, Seattle, 300 00:19:47,790 --> 00:19:50,927 people are looking at AI and saying, like, what's good for me? 301 00:19:51,010 --> 00:19:53,067 And I think there was a prior speaker that 302 00:19:53,150 --> 00:19:55,207 used the metaphor, which I found quite powerful, 303 00:19:55,290 --> 00:20:01,227 which was as electricity came about, we didn't sell electricity, we sold light. 304 00:20:01,310 --> 00:20:03,367 And what do you think is that equivalent for AI? 305 00:20:03,450 --> 00:20:06,587 Because right now there's not a lot of good messaging around AI 306 00:20:06,670 --> 00:20:08,627 of how it's going to benefit people. 307 00:20:08,710 --> 00:20:10,987 Yeah, I think that's right in the sense 308 00:20:11,070 --> 00:20:16,987 that we have perhaps gone too into-- the bubble 309 00:20:17,070 --> 00:20:20,427 that I guess we all live in is more about 310 00:20:20,510 --> 00:20:25,537 hyping the tech and the tech progress for its sake. 311 00:20:25,620 --> 00:20:28,177 We live in it and it's great to be impressed by it 312 00:20:28,260 --> 00:20:30,397 and push the frontiers of it and what have you. 313 00:20:30,480 --> 00:20:35,577 But at the end of the day, the world will evaluate us in what was the value we 314 00:20:35,660 --> 00:20:39,217 created for the world, one community at a time. 315 00:20:39,300 --> 00:20:42,717 I mean, it should always be the case. 316 00:20:42,800 --> 00:20:48,137 And so unless I can see the true benefits of this technology 317 00:20:48,220 --> 00:20:52,897 be broad spread. We talked about healthcare. 318 00:20:52,980 --> 00:20:55,597 When we suddenly start seeing AI in healthcare 319 00:20:55,680 --> 00:21:02,317 change the cost equation, the care one can get on, not in an abstract sense, 320 00:21:02,400 --> 00:21:04,877 but when it happens to someone in our community, 321 00:21:04,960 --> 00:21:09,197 in our family when-- or even take economic opportunity. 322 00:21:09,280 --> 00:21:14,697 Talking about this as something where it takes away jobs, 323 00:21:14,780 --> 00:21:21,077 it's clear that any technology that's disruptive will have real displacement. 324 00:21:21,160 --> 00:21:22,577 But at the same time, there is going 325 00:21:22,660 --> 00:21:26,437 to be new economic activity where humans will have agency, 326 00:21:26,520 --> 00:21:29,177 which will have wages, which in fact, if you think about it, 327 00:21:29,260 --> 00:21:32,497 if what is current intelligence gets commoditized, 328 00:21:32,580 --> 00:21:35,677 humans are the one species that are most adaptive, 329 00:21:35,760 --> 00:21:41,857 and the sense of creating new value on top of what's the new commodity. 330 00:21:41,940 --> 00:21:46,577 And that has to not be abstract, but it has to be real. 331 00:21:46,660 --> 00:21:51,957 And it will happen. But until that transition happens, to your point, 332 00:21:52,040 --> 00:21:54,157 as we go from electricity to light, 333 00:21:54,240 --> 00:21:57,297 and the light is not seen only by the AGI-pilled people 334 00:21:57,380 --> 00:22:01,697 in the zip code, but it's seen by the world as something 335 00:22:01,780 --> 00:22:06,337 that they can thrive in. And even my point about that frontier ecosystem, 336 00:22:06,420 --> 00:22:09,597 when every company is not sitting there thinking that, 337 00:22:09,680 --> 00:22:13,577 oh my God, if I let any one of these frontier models 338 00:22:13,660 --> 00:22:17,057 into my organization, it's just going to run over 339 00:22:17,140 --> 00:22:23,967 all the IP I've created. Why would they welcome that? 340 00:22:24,050 --> 00:22:25,907 By definition they should not. 341 00:22:25,990 --> 00:22:31,547 And so I think that's why as entrepreneurs, as students, 342 00:22:31,630 --> 00:22:35,927 and as incumbents, we have to shape this to an ecosystem which 343 00:22:36,010 --> 00:22:38,067 is positive-sum by definition. 344 00:22:38,150 --> 00:22:40,927 If we are not and it's about a few firms that 345 00:22:41,010 --> 00:22:43,587 have all the returns and everybody else 346 00:22:43,670 --> 00:22:50,167 it's all in bad shape, you will absolutely 347 00:22:50,250 --> 00:22:53,347 lose social permission or we will lose social permission. 348 00:22:53,430 --> 00:22:56,647 OK. I'm going to switch over to questions. 349 00:22:56,730 --> 00:23:01,127 So I'm generally curious about your custom silicon program 350 00:23:01,210 --> 00:23:05,247 and how you've learned from other hyperscalers like Google 351 00:23:05,330 --> 00:23:09,647 and Amazon, who have made some progress there. 352 00:23:09,730 --> 00:23:13,727 It seems like at the hardware side, their chips 353 00:23:13,810 --> 00:23:16,887 that they offer are actually pretty bifurcated. 354 00:23:16,970 --> 00:23:19,087 They have training chips and inference chips 355 00:23:19,170 --> 00:23:23,527 versus in AMD you've kept it unified too. 356 00:23:23,610 --> 00:23:25,967 On the networking side, we had unembodied 357 00:23:26,050 --> 00:23:30,227 class talk about the optical mouse system that they built. 358 00:23:30,310 --> 00:23:33,047 That was really interesting. And at the software side, it seems 359 00:23:33,130 --> 00:23:37,167 like they built their own versions of CUDA 360 00:23:37,250 --> 00:23:40,607 with Neuron and XLA, respectively, 361 00:23:40,690 --> 00:23:45,727 whereas you guys are building on something based off of Triton for mining. 362 00:23:45,810 --> 00:23:49,207 So I'm curious given these differing design decisions 363 00:23:49,290 --> 00:23:51,527 that you don't agree, what have you learned 364 00:23:51,610 --> 00:23:53,947 and where are you taking your own customers? 365 00:23:54,030 --> 00:23:59,367 Yeah, I think the key thing is to recognize 366 00:23:59,450 --> 00:24:00,967 what are the new workloads. 367 00:24:01,050 --> 00:24:06,487 Whenever you think about any new system, 368 00:24:06,570 --> 00:24:10,547 you want to be motivated by what's the new software or what's the new workload. 369 00:24:10,630 --> 00:24:15,397 And the good news here is that there are these three dominant new workloads. 370 00:24:15,480 --> 00:24:17,977 There's the training workload, there's the inference workload, 371 00:24:18,060 --> 00:24:22,757 and now we can say there is the long-running agent that uses 372 00:24:22,840 --> 00:24:26,417 inference and regular compute. 373 00:24:26,500 --> 00:24:28,897 So if you said that's what you have, 374 00:24:28,980 --> 00:24:33,157 then you can start from a first principles looking at it. 375 00:24:33,240 --> 00:24:35,457 And these are interesting type of workloads. 376 00:24:35,540 --> 00:24:38,697 They're not like the previous scale-out workloads. 377 00:24:38,780 --> 00:24:45,780 These are synchronous data parallel workloads where 378 00:24:46,100 --> 00:24:50,477 to a means, I guess point, which is you got to even think about the scale-up part. 379 00:24:50,560 --> 00:24:54,947 Some of the tricks that worked for us for scale-out in the past won't work. 380 00:24:55,030 --> 00:24:58,557 So therefore you now need to even innovate on the scale-up and the scale-out 381 00:24:58,640 --> 00:25:05,640 to really keep things coherent and the MFUs on a training run 382 00:25:06,160 --> 00:25:09,477 are maximized and so on. So therefore, the way we come out of this 383 00:25:09,560 --> 00:25:13,477 and say, OK, even just yesterday we announced there's 384 00:25:13,560 --> 00:25:17,187 Maia 200 that's essentially being codesigned 385 00:25:17,270 --> 00:25:23,567 with our own models, plus the OpenAI models, because we have that IP. 386 00:25:23,650 --> 00:25:29,707 And right now, in fact, Maia 200 is running GPT 5.5 in multiple data centers 387 00:25:29,790 --> 00:25:33,527 powering Copilot, and giving us total TCO advantage. 388 00:25:33,610 --> 00:25:35,867 So that's a great way to round trip 389 00:25:35,950 --> 00:25:39,267 for an inference workload what's the advantage of that is. 390 00:25:39,350 --> 00:25:44,667 We not only did that but we also built 391 00:25:44,750 --> 00:25:50,427 Cobalt which is our ARM processor for compute. 392 00:25:50,510 --> 00:25:52,507 And we're benchmarking it to improve 393 00:25:52,590 --> 00:25:57,047 both for latency performance when it comes to, for example, 394 00:25:57,130 --> 00:25:58,867 the agentic loop because the place where 395 00:25:58,950 --> 00:26:02,627 you need great cores are for these agent loops. 396 00:26:02,710 --> 00:26:08,087 So we're using all the GitHub Copilot traces to optimize our ARM processor even. 397 00:26:08,170 --> 00:26:11,987 And bringing all this together with even the networking stack. 398 00:26:12,070 --> 00:26:16,087 And so our approach would be to not-- and at the same time, 399 00:26:16,170 --> 00:26:19,667 we love to have the GPUs because they're general purpose to your point, which 400 00:26:19,750 --> 00:26:21,367 is, in fact, we're using GPUs. 401 00:26:21,450 --> 00:26:27,907 In fact, we're using the old GPUs in our fleet to accelerate our data warehouse. 402 00:26:27,990 --> 00:26:34,287 So Fabric is seeing 7x plus performance gains because of GPU acceleration. 403 00:26:34,370 --> 00:26:38,567 So we think of our fleet as a heterogeneous fleet, 404 00:26:38,650 --> 00:26:42,067 where we will use software to get the maximum benefit out 405 00:26:42,150 --> 00:26:45,867 of it and do smart workload placement. 406 00:26:45,950 --> 00:26:49,547 At the same time, optimize for the high-volume workloads 407 00:26:49,630 --> 00:26:53,887 like inference and training and agent loop with our own ground-up system. 408 00:26:53,970 --> 00:26:55,307 And there's lots of design points. 409 00:26:55,390 --> 00:26:57,447 Most people get fixated on the AI accelerator. 410 00:26:57,530 --> 00:26:59,267 But the AI accelerator is one, the CPU 411 00:26:59,350 --> 00:27:03,967 is one, the network accelerator, the storage accelerator, 412 00:27:04,050 --> 00:27:06,427 the AI WAN is another one. 413 00:27:06,510 --> 00:27:10,987 You want to be able to really do multi-data center hops even. 414 00:27:11,070 --> 00:27:17,097 So, lots of stuff. It's a great time, by the way, to be in computer architecture. 415 00:27:17,180 --> 00:27:22,777 I think when I started in the industries when the Patterson book first came out 416 00:27:22,860 --> 00:27:28,657 and that was the RISK versus CISC debate, I feel like we're back at a time like this 417 00:27:28,740 --> 00:27:33,777 where you can really rethink from the physical design 418 00:27:33,860 --> 00:27:38,247 of a data center to-- by the way, the electrons. 419 00:27:38,330 --> 00:27:40,497 I mean, one of the places where I'm very excited about 420 00:27:40,580 --> 00:27:43,657 is the efficiency with which we can bring the electrons all 421 00:27:43,740 --> 00:27:48,337 the way to the CPU so that the tokens are that much more 422 00:27:48,420 --> 00:27:50,757 efficient without all the losses in between. 423 00:27:50,840 --> 00:27:54,177 So I think there's just a tremendous design [INAUDIBLE]. 424 00:27:54,260 --> 00:27:56,257 And another announcement you had yesterday 425 00:27:56,340 --> 00:27:58,997 was around quantum, which is kind of adjacent to what you talked about. 426 00:27:59,080 --> 00:28:01,367 So I'm kind of curious what was the announcement and what 427 00:28:01,450 --> 00:28:03,597 was the recent advances there? 428 00:28:03,680 --> 00:28:07,577 Yeah, so look, this quantum, we have 429 00:28:07,660 --> 00:28:10,747 been on at this for now the last 20 plus years, 430 00:28:10,830 --> 00:28:14,707 and it's really exciting to see the progress. 431 00:28:14,790 --> 00:28:21,007 I'll just say one of the things is, even independent of the quantum program, 432 00:28:21,090 --> 00:28:25,407 even with what we were able to achieve in the last couple of years, with even 433 00:28:25,490 --> 00:28:28,507 the natural atom-based quantum computers with our stack-- 434 00:28:28,590 --> 00:28:30,107 we worked with partners on it. 435 00:28:30,190 --> 00:28:35,867 We're able to generate now these very good traces, which those traces-- 436 00:28:35,950 --> 00:28:39,287 basically, if you think about what's the purpose of a quantum computer, 437 00:28:39,370 --> 00:28:42,167 a quantum computer can simulate nature. 438 00:28:42,250 --> 00:28:45,787 I mean, since nature is quantum. 439 00:28:45,870 --> 00:28:49,767 And so instead of relying on DFTs or what have you, 440 00:28:49,850 --> 00:28:56,247 you can now have a lot better fidelity of say, chemistry 441 00:28:56,330 --> 00:28:59,947 or molecular dynamics or what have you. 442 00:29:00,030 --> 00:29:04,307 And those traces then can be taken back and you can train a model. 443 00:29:04,390 --> 00:29:09,167 In fact, we are doing that with our material science models, 444 00:29:09,250 --> 00:29:11,947 where you can take the traces from even a whatever, 445 00:29:12,030 --> 00:29:17,207 an early stage quantum computer, to improve the data on which you 446 00:29:17,290 --> 00:29:22,327 train a model for something like material science or chemistry. 447 00:29:22,410 --> 00:29:27,667 Now, our quantum program itself is, as I said, there's a software side to it, 448 00:29:27,750 --> 00:29:31,427 which we will put on ion trap machines, which we're putting with partners. 449 00:29:31,510 --> 00:29:34,607 We're putting it on a photonics-based machine. 450 00:29:34,690 --> 00:29:36,427 We're also putting it on natural atoms. 451 00:29:36,510 --> 00:29:40,967 We have a partnership with in Denmark called QuNorth, where we will even 452 00:29:41,050 --> 00:29:45,807 have a quantum computer powered by Atom Computing with our stack 453 00:29:45,890 --> 00:29:49,127 within the year and so on. So that's one side of it. 454 00:29:49,210 --> 00:29:52,367 The second side is ultimately in order 455 00:29:52,450 --> 00:29:57,367 to build a quantum computer at scale, at utility-scale, 456 00:29:57,450 --> 00:30:00,727 you need fault tolerance. 457 00:30:00,810 --> 00:30:04,687 Our bet on that has been that we-- 458 00:30:04,770 --> 00:30:11,770 there was a theoretical physicist who theorized 459 00:30:12,000 --> 00:30:19,000 essentially a state of matter called a Majorana in the 1930s. 460 00:30:20,000 --> 00:30:24,517 And so one of the things that we felt that that was the state of 461 00:30:24,600 --> 00:30:29,677 matter that we needed to essentially fabricate and make real. 462 00:30:29,760 --> 00:30:35,037 So we launched our first QPU, which was Majorana 1 a year ago, 463 00:30:35,120 --> 00:30:38,197 which essentially proved out the fundamental physics 464 00:30:38,280 --> 00:30:44,197 breakthrough, that you can actually have this and then instantiate it. 465 00:30:44,280 --> 00:30:47,077 And now we've got Majorana 2, which 466 00:30:47,160 --> 00:30:52,637 allows this to be built at industrial scale. 467 00:30:52,720 --> 00:30:57,997 And so there's a lot of detail in terms of how long these qubits can be stable for. 468 00:30:58,080 --> 00:31:00,517 And by the way, one of the other things 469 00:31:00,600 --> 00:31:03,837 is we have perfected the digital control of this quantum computer 470 00:31:03,920 --> 00:31:06,067 because that's going to be super important. 471 00:31:06,150 --> 00:31:09,867 So overall, we feel that the quantum program at Microsoft's 472 00:31:09,950 --> 00:31:11,527 progressing on two dimensions. 473 00:31:11,610 --> 00:31:16,507 One is in the near-term with even what are the quantum computers that I think 474 00:31:16,590 --> 00:31:20,977 are most easy to fabricate and build today with these things like natural atoms. 475 00:31:21,060 --> 00:31:22,347 And then in the long-run, we want 476 00:31:22,430 --> 00:31:28,787 to build out what is needed in order for true quantum computers 477 00:31:28,870 --> 00:31:32,167 to act like utility-scale quantum computers. 478 00:31:32,250 --> 00:31:35,627 On that ladder, if you had a guess at timeline? 479 00:31:35,710 --> 00:31:40,067 I'm the third CEO at Microsoft to keep 480 00:31:40,150 --> 00:31:42,667 going on the quantum journey. 481 00:31:42,750 --> 00:31:47,247 I would say that what I'm now a lot more bullish is it may not-- 482 00:31:47,330 --> 00:31:50,447 it's kind of like the previous discussion. 483 00:31:50,530 --> 00:31:54,267 I think of quantum as the new accelerator. 484 00:31:54,350 --> 00:31:57,767 And remember, by the way, quantum is not going to replace classical. 485 00:31:57,850 --> 00:32:00,547 Quantum is not going to be great at storage and memory and so on. 486 00:32:00,630 --> 00:32:02,127 It's going to be great at computation. 487 00:32:02,210 --> 00:32:06,647 And so you have to marry classical plus quantum in order to do things. 488 00:32:06,730 --> 00:32:10,987 And so therefore, I think of this as maybe a lot more staged even. 489 00:32:11,070 --> 00:32:13,467 So if you have 100 logical qubits with good error 490 00:32:13,550 --> 00:32:19,787 correction, we can start using it to generate synth data for science models. 491 00:32:19,870 --> 00:32:22,267 Like that'll be a pretty important milestone that 492 00:32:22,350 --> 00:32:25,467 may be even more achievable in the short-run. 493 00:32:25,550 --> 00:32:30,467 So we'll see. We I think made the claim even yesterday 494 00:32:30,550 --> 00:32:32,707 that by the end of the decade, we 495 00:32:32,790 --> 00:32:35,707 believe we will be able to build a quantum computer that starts 496 00:32:35,790 --> 00:32:37,857 solving some real challenges. Real problems? 497 00:32:37,940 --> 00:32:39,810 Yeah. Amazing. Next question. 498 00:32:46,590 --> 00:32:50,907 Thanks so much for coming. I spent nine years at Microsoft. 499 00:32:50,990 --> 00:32:57,187 I joined a year after you became a CEO as a market employee. 500 00:32:57,270 --> 00:33:03,617 So I went through cloud transformation and then I go into AI fascination. 501 00:33:03,700 --> 00:33:08,477 And I think my program was such an amazing experience for me. 502 00:33:08,560 --> 00:33:11,417 And it shaped the younger person. 503 00:33:11,500 --> 00:33:14,417 How do you think it contributed and is it 504 00:33:14,500 --> 00:33:18,177 still contributing to the culture and success of mine? 505 00:33:18,260 --> 00:33:20,677 Yeah, I mean, first of all, we're very, very thrilled about, 506 00:33:20,760 --> 00:33:25,037 obviously students coming in and joining and having essentially the Mac program. 507 00:33:25,120 --> 00:33:26,897 There are a couple of programs like that at Microsoft we 508 00:33:26,980 --> 00:33:31,857 created where people can even rotate through various functions. 509 00:33:31,940 --> 00:33:37,717 At the end of the day, any company for it to be at the frontier, so to speak, 510 00:33:37,800 --> 00:33:41,477 has to be able to get people coming with fresh ideas, 511 00:33:41,560 --> 00:33:43,837 fresh energy, and reshaping it. 512 00:33:43,920 --> 00:33:47,857 I always say to anyone joining Microsoft, of course you want to come in and learn 513 00:33:47,940 --> 00:33:52,697 about how Microsoft works, but we also want Microsoft to learn from you. 514 00:33:52,780 --> 00:33:55,897 And more importantly, for you to have the agency to reshape 515 00:33:55,980 --> 00:33:59,967 what is Microsoft's culture. It's not a static thing. 516 00:34:00,050 --> 00:34:04,107 It's an organic thing that gets shaped by the behaviors, 517 00:34:04,190 --> 00:34:07,627 the decisions of people at the company. 518 00:34:07,710 --> 00:34:12,047 And so we always would welcome students coming in, 519 00:34:12,130 --> 00:34:15,327 building their career at Microsoft. 520 00:34:15,410 --> 00:34:18,947 One of the things as a 50-year-old company, I mean, 521 00:34:19,030 --> 00:34:22,767 Mike's an alum at Microsoft and still engaged with us. 522 00:34:22,850 --> 00:34:27,467 We have people who've come, had a tour of duty, gone out, come back. 523 00:34:27,550 --> 00:34:32,287 And so at this point, the way to think 524 00:34:32,370 --> 00:34:36,687 about it is the uniqueness of the Microsoft 525 00:34:36,770 --> 00:34:40,747 is our core DNA has remained. 526 00:34:40,830 --> 00:34:46,907 We are a developer tools platform, knowledge worker tools company. 527 00:34:46,990 --> 00:34:49,767 That's kind of what we've done for 50 years. 528 00:34:49,850 --> 00:34:55,567 But the interesting thing about us is that we have been able to reinterpret 529 00:34:55,650 --> 00:35:01,567 that with every new platform. In fact, I joined the company back in the '90s when 530 00:35:01,650 --> 00:35:04,967 my existential competition was Novell. 531 00:35:05,050 --> 00:35:10,127 And now, it's some foundation lab I'd not even heard of five years ago. 532 00:35:10,210 --> 00:35:12,847 But that is the thing that I think 533 00:35:12,930 --> 00:35:16,587 keeps us vibrant, which is our existential challenge, 534 00:35:16,670 --> 00:35:23,107 or what we need to compete with is new and fresh versus it's the same old. 535 00:35:23,190 --> 00:35:25,607 And I think that that's an attractive part. 536 00:35:25,690 --> 00:35:28,207 When you come to Microsoft, you will 537 00:35:28,290 --> 00:35:32,887 be able to go at that mission of being able to empower people 538 00:35:32,970 --> 00:35:38,167 and organizations all over the planet, which means a lot to us. 539 00:35:38,250 --> 00:35:43,607 But to do so, recognizing that we as a company 540 00:35:43,690 --> 00:35:46,967 can bring a lot to that mission. 541 00:35:47,050 --> 00:35:49,607 I have a follow-up question. So one of the many attributes I've 542 00:35:49,690 --> 00:35:52,927 admired about you is you have a growth mindset, 543 00:35:53,010 --> 00:35:55,597 and you really look at your leadership team and drive it. 544 00:35:55,680 --> 00:35:58,477 How have you instilled that across the company? 545 00:35:58,560 --> 00:36:00,717 Because you clearly have. You just pointed that out, that you've 546 00:36:00,800 --> 00:36:03,257 been able to deal with these platform transformations. 547 00:36:03,340 --> 00:36:10,197 Yeah, I mean, at some level, I think it's not something you instill per se, 548 00:36:10,280 --> 00:36:14,877 Mike, you invoke what is innate in all of us, I think. 549 00:36:14,960 --> 00:36:20,637 I mean, you have to do it more out of practice. 550 00:36:20,720 --> 00:36:23,957 Mostly what I have to exhibit more 551 00:36:24,040 --> 00:36:29,617 than anything else is my ability to confront my fixed mindset. 552 00:36:29,700 --> 00:36:33,417 Because at the end of the day, all of us, it's easy to talk about growth mindset, 553 00:36:33,500 --> 00:36:36,777 but it's very difficult to exercise it individually. 554 00:36:36,860 --> 00:36:38,357 As somebody said to me, which I always 555 00:36:38,440 --> 00:36:45,440 liked as a sort of a nice quip, is everybody likes change, 556 00:36:45,860 --> 00:36:48,577 except they want the other person to change, not themselves. 557 00:36:48,660 --> 00:36:53,257 And that, I think, is the challenge of growth mindset. 558 00:36:53,340 --> 00:36:55,617 So it's not about talking about growth mindset. 559 00:36:55,700 --> 00:36:59,037 It's about having the courage to confront one's own fixed mindset. 560 00:36:59,120 --> 00:37:01,957 So it can become corporate dogma to your point. 561 00:37:02,040 --> 00:37:06,037 So one of the keys, the reason why it's worked at Microsoft 562 00:37:06,120 --> 00:37:10,277 is we never made it like, oh-- --some mandate. 563 00:37:10,360 --> 00:37:15,697 It started with you. And also it's not like trademarked to Microsoft. 564 00:37:15,780 --> 00:37:20,957 I mean, if you exercise growth mindset or you confront a fixed mindset, you'll 565 00:37:21,040 --> 00:37:23,257 be a better human being first. 566 00:37:23,340 --> 00:37:25,997 You'll be a better colleague, a better friend, a better 567 00:37:26,080 --> 00:37:29,137 neighbor, a better parent, a better student, everything. 568 00:37:29,220 --> 00:37:32,337 So you're not even doing this for Microsoft's sake. 569 00:37:32,420 --> 00:37:34,017 You're doing this for yourself. 570 00:37:34,100 --> 00:37:38,877 And I think giving that oxygen, leaving that at that, as opposed 571 00:37:38,960 --> 00:37:42,697 to some new corporate thing, has been very, very helpful. 572 00:37:42,780 --> 00:37:48,037 So I'm an advocate of it, not just at Microsoft, anywhere. 573 00:37:48,120 --> 00:37:51,987 And more importantly, it's that practice of-- 574 00:37:52,070 --> 00:37:55,547 there are two things that I feel that were pretty influential 575 00:37:55,630 --> 00:38:01,127 for me, which I learned through my wife's readings, quite frankly. 576 00:38:01,210 --> 00:38:04,747 One was this thing called nonviolent communications, which 577 00:38:04,830 --> 00:38:09,567 is also another form of developing a sense of empathy, 578 00:38:09,650 --> 00:38:12,287 understanding where the other person is coming from, 579 00:38:12,370 --> 00:38:16,067 not having your amygdala always triggered and what have you. 580 00:38:16,150 --> 00:38:20,307 So that's one which I think is it's a great read if you've not read it. 581 00:38:20,390 --> 00:38:24,487 And then, of course, Carol Dweck's work on growth mindset. 582 00:38:24,570 --> 00:38:26,527 These are two things that are I think, 583 00:38:26,610 --> 00:38:30,227 relevant for children and students and child psychology. 584 00:38:30,310 --> 00:38:33,307 But I think they apply to corporate cultures 585 00:38:33,390 --> 00:38:35,787 because I think one of the fascinating things 586 00:38:35,870 --> 00:38:40,127 is what Herbert Simon described as the bounded rationality. 587 00:38:40,210 --> 00:38:45,690 I think humans are great, but we have this unfortunate-- 588 00:38:50,700 --> 00:38:55,897 we don't see what's in our interest all the time. 589 00:38:55,980 --> 00:39:01,217 We get hijacked often without being 590 00:39:01,300 --> 00:39:06,297 able to do the simple calculus to what 591 00:39:06,380 --> 00:39:10,137 does it mean to be at the frontier of our own behavior. 592 00:39:10,220 --> 00:39:14,117 And I think that these are nice practices that gets us and pushes us. 593 00:39:14,200 --> 00:39:18,457 So think of it as your training run that you need. 594 00:39:18,540 --> 00:39:21,877 That's great guidance. Thank you so much for the answer. 595 00:39:21,960 --> 00:39:26,557 And I had my dissertation 10 minutes ago, but I really wanted to meet you. 596 00:39:26,640 --> 00:39:28,077 Thank you so much. All right. Good luck. 597 00:39:28,160 --> 00:39:33,997 We realize it's finals week. Next question. 598 00:39:34,080 --> 00:39:35,677 Yeah. Thanks so much for coming. 599 00:39:35,760 --> 00:39:39,497 I was wondering, how did you become such a good public speaker and what-- 600 00:39:39,580 --> 00:39:42,880 [LAUGHTER] 601 00:39:44,780 --> 00:39:51,637 I don't know, man. I mean, I'm glad you think I'm a good public speaker. 602 00:39:51,720 --> 00:39:57,937 Let's leave it at that. I mean, look, I think, like anything else, 603 00:39:58,020 --> 00:40:01,497 it's not that I think about public speaking as a key thing 604 00:40:01,580 --> 00:40:03,817 that I'm trying to develop or what have you. 605 00:40:03,900 --> 00:40:10,900 But the lucky part that I find myself is in particular, even 606 00:40:12,020 --> 00:40:18,097 I think one of the things is when I became CEO, you had to talk about things, perhaps 607 00:40:18,180 --> 00:40:21,037 that you didn't get the opportunity to talk about previously. 608 00:40:21,120 --> 00:40:23,137 Maybe that's a better way to characterize it. 609 00:40:23,220 --> 00:40:26,057 But the good news is, it's not as if I was not thinking 610 00:40:26,140 --> 00:40:28,817 about those things previously. 611 00:40:28,900 --> 00:40:32,937 And I reflected on it. Is that why is it that I was thinking about those things 612 00:40:33,020 --> 00:40:38,377 previously? And I think that comes out of just natural interest, 613 00:40:38,460 --> 00:40:41,477 for example, thinking about technology, 614 00:40:41,560 --> 00:40:45,457 but its impact, what does it mean? 615 00:40:45,540 --> 00:40:48,347 I have many pet passions. 616 00:40:48,430 --> 00:40:53,727 Like what does any technological progress like AI 617 00:40:53,810 --> 00:40:59,367 mean to the Global South? What does it mean to even what has been a dream-- 618 00:40:59,450 --> 00:41:01,627 and I grew up as a son of a development economist, 619 00:41:01,710 --> 00:41:04,447 so he instilled in me that, hey, this convergence growth 620 00:41:04,530 --> 00:41:07,107 is going to happen and it's going to be great and so on. 621 00:41:07,190 --> 00:41:09,247 And so I'm obsessed about it. 622 00:41:09,330 --> 00:41:13,367 And so as long as I think you have these passions that 623 00:41:13,450 --> 00:41:19,927 allow you to think broadly, and then for you to be able to talk freely about it, 624 00:41:20,010 --> 00:41:26,887 is the other one. And so I'm not particularly an expert on public speaking, 625 00:41:26,970 --> 00:41:33,507 but I think the more any of us can have broad interests that we can articulate. 626 00:41:33,590 --> 00:41:35,607 And in today's day and age, the media 627 00:41:35,690 --> 00:41:40,907 allows us to be able to have our own outlets. 628 00:41:40,990 --> 00:41:44,767 And so I think that this is a great time to both build that interest and then 629 00:41:44,850 --> 00:41:47,257 to be able to have different medium, whether it's speaking, 630 00:41:47,340 --> 00:41:50,317 whether it's writing, whether it's podcasts, what have you. 631 00:41:50,400 --> 00:41:51,877 There are a variety of ways I think 632 00:41:51,960 --> 00:41:56,420 we can express, reach, debate, which I think is fantastic. 633 00:41:59,820 --> 00:42:03,477 Thank you for coming. I was wondering if your undergraduate self was 634 00:42:03,560 --> 00:42:06,797 sitting in the audience right now, what advice would you give him? 635 00:42:06,880 --> 00:42:08,417 And knowing what you know now, I guess, 636 00:42:08,500 --> 00:42:14,317 and what would you tell him to put his energy into and to maybe avoid? 637 00:42:14,400 --> 00:42:19,217 Yeah, it's a great question. I mean, it's such a privilege in some sense. 638 00:42:19,300 --> 00:42:20,240 I wish I could. 639 00:42:23,000 --> 00:42:24,697 Because everything is in front of you. 640 00:42:24,780 --> 00:42:29,157 You're risk on 100% of the time. 641 00:42:29,240 --> 00:42:33,277 Maybe that's what it is, which is-- 642 00:42:33,360 --> 00:42:37,257 let me just say two interesting things. 643 00:42:37,340 --> 00:42:41,097 Yesterday on Hacker News, I came across, I forget, one of the CS classes here. 644 00:42:41,180 --> 00:42:46,797 It had the guidelines on how to use coding agents, which I thought was well done. 645 00:42:46,880 --> 00:42:53,477 They had the do's and don'ts. And so the fascinating thing I find 646 00:42:53,560 --> 00:43:00,560 right now is the ability to learn new things 647 00:43:01,000 --> 00:43:03,837 has become so much easier. 648 00:43:03,920 --> 00:43:09,117 Because you have this very accessible, personalized tutor 649 00:43:09,200 --> 00:43:14,357 that's deep that you can go and work with. 650 00:43:14,440 --> 00:43:20,997 And so I would say more so than any assignment anxiety 651 00:43:21,080 --> 00:43:27,557 or I don't grade anxiety or what have you, you can have-- 652 00:43:27,640 --> 00:43:32,857 one of the terms one of my colleagues uses-- real cognitive coverage. 653 00:43:32,940 --> 00:43:37,237 Like test coverage, you can now have 654 00:43:37,320 --> 00:43:42,587 cognitive coverage that really follow you through your curiosity. 655 00:43:42,670 --> 00:43:48,107 If I were back as an undergrad, I would be trying to-- it's kind of what I 656 00:43:48,190 --> 00:43:49,687 do with GitHub sessions today. 657 00:43:49,770 --> 00:43:54,847 GitHub app, which is what are all the coding agents and what are all they doing? 658 00:43:54,930 --> 00:43:56,407 So that's what I would do. 659 00:43:56,490 --> 00:44:03,490 I would be sitting with, what are the 10, 100 agents 660 00:44:03,710 --> 00:44:09,607 and me at Stanford learning? But I need to have cognitive coverage. 661 00:44:09,690 --> 00:44:11,447 It's not I've offloaded to the 100. 662 00:44:11,530 --> 00:44:17,087 The key is, what am I instructing them? 663 00:44:17,170 --> 00:44:19,227 And then when they get something done, 664 00:44:19,310 --> 00:44:24,427 can I understand what they did in order to have learned? 665 00:44:24,510 --> 00:44:27,947 It's kind of like it's 100 classes. 666 00:44:28,030 --> 00:44:30,747 It's so fascinating. And that I think, is what I think will happen. 667 00:44:30,830 --> 00:44:37,747 One of these days somebody's going to break a new pedagogy that goes with-- 668 00:44:37,830 --> 00:44:39,637 the tools, for example, are evolving. 669 00:44:39,720 --> 00:44:41,717 Like think about what happened in developer tools. 670 00:44:41,800 --> 00:44:46,657 We went from saying, hey, we have 100 CLIs to now 671 00:44:46,740 --> 00:44:50,377 we need a thing that to manage our CLI complexity, which 672 00:44:50,460 --> 00:44:53,537 is the new AD, which is like for example, the GitHub app is 673 00:44:53,620 --> 00:44:55,457 fantastic in that context because it's 674 00:44:55,540 --> 00:44:58,657 like the new inbox for managing my sessions. 675 00:44:58,740 --> 00:45:02,857 What's the moral equivalent of that, that allows a student 676 00:45:02,940 --> 00:45:07,057 to navigate through their learning experience 677 00:45:07,140 --> 00:45:14,140 and be max curious, but really getting deeper, faster on things 678 00:45:14,580 --> 00:45:17,217 that you're trying to cover? 679 00:45:17,300 --> 00:45:20,997 And I think that that's what I would do and not have anxiety. 680 00:45:21,080 --> 00:45:24,637 Because you can always push a button to get an assignment done. 681 00:45:24,720 --> 00:45:28,477 So that's no longer the case. And the grades may or may not matter. 682 00:45:28,560 --> 00:45:34,907 And so therefore there's a lot of relitigation on the things we valued. 683 00:45:34,990 --> 00:45:38,457 That's a really good answer. Next question. 684 00:45:38,540 --> 00:45:41,017 Hey, thank you so much for coming. 685 00:45:41,100 --> 00:45:44,877 So, like, whenever we're interacting with a computer, 686 00:45:44,960 --> 00:45:49,977 pretty much we always interact with a GUI interface. 687 00:45:50,060 --> 00:45:55,297 But agents, they just so happen to be good at coding, 688 00:45:55,380 --> 00:46:00,077 happen to be not that good at interacting with GUI interfaces. 689 00:46:00,160 --> 00:46:03,097 So for example, if I want to design a poster 690 00:46:03,180 --> 00:46:06,497 and I want to be able to do it, it is easier for me 691 00:46:06,580 --> 00:46:09,457 to get it to generate my HTML/CSS code 692 00:46:09,540 --> 00:46:14,617 rather than to use a GUI design product. 693 00:46:14,700 --> 00:46:20,337 In that case, what do you think importation 694 00:46:20,420 --> 00:46:27,420 of [INAUDIBLE] for BUI interfaces versus a seal. 695 00:46:27,750 --> 00:46:30,547 Yeah, I mean, I think you're bringing up a couple of different things. 696 00:46:30,630 --> 00:46:34,737 One is essentially codegen is powerful, 697 00:46:34,820 --> 00:46:39,287 and therefore HTML and other web UI, 698 00:46:39,370 --> 00:46:45,887 as an artifact creation process, I think is going to really proliferate. 699 00:46:45,970 --> 00:46:49,187 So basically, we've gone from-- we in fact, 700 00:46:49,270 --> 00:46:51,927 always Bill used to have this thing where, what's 701 00:46:52,010 --> 00:46:55,987 the difference between building an app, writing a document, 702 00:46:56,070 --> 00:47:01,147 or creating a website? At this point, there's none. 703 00:47:01,230 --> 00:47:04,767 You can just basically do all three by using code. 704 00:47:04,850 --> 00:47:11,167 So that's one side of it. But I think that the direct manipulation is the challenge. 705 00:47:11,250 --> 00:47:14,527 I think in the intermediate time frame, what's going to happen is you're going 706 00:47:14,610 --> 00:47:16,167 to have an intermediate format. 707 00:47:16,250 --> 00:47:20,547 So you're going to do the HTML and then you can convert into Excel, PowerPoint, 708 00:47:20,630 --> 00:47:23,187 PowerPoint into intermediate format and then have agents. 709 00:47:23,270 --> 00:47:25,447 So I think that that's what you see in Copilot 710 00:47:25,530 --> 00:47:29,007 and elsewhere when you think about artifact creation. 711 00:47:29,090 --> 00:47:32,367 But the ultimate thing is, can you truly 712 00:47:32,450 --> 00:47:37,937 teach even the agent on the model, the Canvas, 713 00:47:38,020 --> 00:47:40,557 and the direct manipulation of the Canvas, 714 00:47:40,640 --> 00:47:47,557 which has to be done fundamentally by teaching it the semantics of that Canvas? 715 00:47:47,640 --> 00:47:50,067 And so it has to be exposed to whether it's through APIs 716 00:47:50,150 --> 00:47:54,557 or whether it is through a protocol or what have you. 717 00:47:54,640 --> 00:47:57,057 And so therefore, I think you will see innovation like that. 718 00:47:57,140 --> 00:47:59,297 But it is true that direct manipulation-- by the way, 719 00:47:59,380 --> 00:48:00,757 you talked about, one of the other things that 720 00:48:00,840 --> 00:48:02,717 struck me is one of the nice little features 721 00:48:02,800 --> 00:48:07,437 we added to GitHub yesterday was a thing called Canvas, 722 00:48:07,520 --> 00:48:12,777 and the reason was not because the agents need UI, 723 00:48:12,860 --> 00:48:15,037 but we need UI, because it's now become 724 00:48:15,120 --> 00:48:19,077 too dense to just keep tracking my CLI session 725 00:48:19,160 --> 00:48:23,077 or the chat session because, first of all, it's linear 726 00:48:23,160 --> 00:48:26,937 and it's painful you're trying to scroll through it. 727 00:48:27,020 --> 00:48:29,237 And so one of the things that we said 728 00:48:29,320 --> 00:48:31,997 is now we can-- for example, I can have a Kanban board 729 00:48:32,080 --> 00:48:37,277 as a visualization, which both the agent and I are working on. 730 00:48:37,360 --> 00:48:42,117 So I think that this idea of generated UI becoming 731 00:48:42,200 --> 00:48:45,997 the new way for human agent interaction 732 00:48:46,080 --> 00:48:51,237 might be one of the coolest things that will happen across all product lines. 733 00:48:51,320 --> 00:48:56,157 Next question. Hi, Santhya. Thank you so much for coming. 734 00:48:56,240 --> 00:48:58,437 If you were at our age like college freshman 735 00:48:58,520 --> 00:49:00,217 and you had the world at your fingertips, 736 00:49:00,300 --> 00:49:04,357 what's the problem that you would encourage us in [INAUDIBLE]? 737 00:49:04,440 --> 00:49:09,417 Oh man, I was that, but I don't know. 738 00:49:09,500 --> 00:49:15,357 I mean, it's always interesting to look back and say, 739 00:49:15,440 --> 00:49:18,377 what would I pick? I don't know. 740 00:49:18,460 --> 00:49:24,197 I mean, in an interesting world like ours right now, 741 00:49:24,280 --> 00:49:26,677 I think you have to go and say, what 742 00:49:26,760 --> 00:49:31,547 is the thing that you have inherent interests in 743 00:49:31,630 --> 00:49:36,807 and the world will value? I think always whenever people are making choices, 744 00:49:36,890 --> 00:49:39,367 I think they have two things they're trying to intersect. 745 00:49:39,450 --> 00:49:45,067 They're trying to intersect something that they believe they have a real passion for. 746 00:49:45,150 --> 00:49:50,907 But they're also doing, quite frankly, the calculus on, what does the world value? 747 00:49:50,990 --> 00:49:53,407 They always have some destination in mind. 748 00:49:53,490 --> 00:49:55,727 I want that career, I want that job, 749 00:49:55,810 --> 00:49:58,807 I want to start that company or what have you. 750 00:49:58,890 --> 00:50:00,667 And so I would focus on that. 751 00:50:00,750 --> 00:50:05,527 Answering those two questions is what I think will lead. 752 00:50:05,610 --> 00:50:08,647 And in my case, I would probably go step back and look at that. 753 00:50:08,730 --> 00:50:11,787 It may, may be even outside. 754 00:50:11,870 --> 00:50:16,387 One of the things if I go back to the computer industry-- 755 00:50:16,470 --> 00:50:21,427 I was an electrical engineer and I then drifted into software. 756 00:50:21,510 --> 00:50:27,547 But if I went back now, I may go back into hardware, 757 00:50:27,630 --> 00:50:33,967 just because there is just such an unbelievable time in-- 758 00:50:34,050 --> 00:50:38,187 There are a lot of things that I would love to go deep on understanding what 759 00:50:38,270 --> 00:50:43,647 the optical side of networking would look like, some of the system design. 760 00:50:43,730 --> 00:50:47,487 So I think that's how I think things will pan out. 761 00:50:47,570 --> 00:50:50,427 People will pick the thing that they're good 762 00:50:50,510 --> 00:50:54,047 at, then they see the trajectory of what they're good at and say, 763 00:50:54,130 --> 00:50:57,187 wow, I'm going to bet on myself to get good at this 764 00:50:57,270 --> 00:51:03,147 and start something and/or policy. 765 00:51:03,230 --> 00:51:08,227 And people talk about safety engineering, but I was thinking, wow, there are so 766 00:51:08,310 --> 00:51:11,787 many aspects of what does it mean 767 00:51:11,870 --> 00:51:17,867 to have safety around AI that require people to think through deeply. 768 00:51:17,950 --> 00:51:20,307 And so anyway, so there are lots of choices out there. 769 00:51:20,390 --> 00:51:23,227 Along those lines, have you guys hired like philosophers 770 00:51:23,310 --> 00:51:26,087 to help you with the AI of guidance? 771 00:51:26,170 --> 00:51:28,557 I think Mustafa is a philosopher dropout or something. 772 00:51:28,640 --> 00:51:30,160 So he's-- Sung. 773 00:51:31,460 --> 00:51:36,377 So we have a quasi wannabe philosopher. 774 00:51:36,460 --> 00:51:41,017 But he thinks clearly from that. 775 00:51:41,100 --> 00:51:42,737 I mean, he's been, obviously, since being 776 00:51:42,820 --> 00:51:47,937 a founder of DeepMind to now, he's always thought about it, 777 00:51:48,020 --> 00:51:51,497 and we have always had folks in MSR 778 00:51:51,580 --> 00:51:55,697 who are brought real, deep, multidisciplinary approach 779 00:51:55,780 --> 00:52:02,577 to it, whether it's the economist, the moral philosophers, the sociologists. 780 00:52:02,660 --> 00:52:05,857 And I think we'll always have that. 781 00:52:05,940 --> 00:52:08,817 Next question. Hi, Satya. 782 00:52:08,900 --> 00:52:11,997 I was curious, what do you think about space data centers. 783 00:52:12,080 --> 00:52:16,437 Because I've had a lot of startup CEOs that come here and talk and just try 784 00:52:16,520 --> 00:52:18,477 to convince me that we should be looking at that. 785 00:52:18,560 --> 00:52:21,297 And the research I'm doing is showing that Elon is probably 786 00:52:21,380 --> 00:52:24,137 the only one that's being able to do it profitably just 787 00:52:24,220 --> 00:52:25,957 from the large costs per kilo. 788 00:52:26,040 --> 00:52:27,287 I was just curious where you're. 789 00:52:27,370 --> 00:52:34,370 I'm not an expert on any of that in both the supply side of it 790 00:52:34,690 --> 00:52:39,447 and the economic side of it, but it's plausible, at least, 791 00:52:39,530 --> 00:52:42,967 what I've read and what I've talked to people who are experts in it. 792 00:52:43,050 --> 00:52:49,887 Seems like it makes sense. The question really is you now need 793 00:52:49,970 --> 00:52:54,127 to not only solve both, how do you get there, 794 00:52:54,210 --> 00:52:58,447 but how do you build the stack that operates there 795 00:52:58,530 --> 00:53:02,807 and then solve all the practical issues of RMA and others. 796 00:53:02,890 --> 00:53:06,967 Because therefore, I think there's a whole supply chain. 797 00:53:07,050 --> 00:53:13,087 When you think about the data center, most people-- it's a complex project. 798 00:53:13,170 --> 00:53:18,527 We're built on the shoulders of unbelievable engineering depth. 799 00:53:18,610 --> 00:53:21,607 Starting from civil engineering, on 800 00:53:21,690 --> 00:53:24,407 and to electrical engineering, to mechanical engineering, 801 00:53:24,490 --> 00:53:31,207 to then ultimately have it meet the needs of computing. 802 00:53:31,290 --> 00:53:34,727 And so that level of sophistication 803 00:53:34,810 --> 00:53:39,687 for this new payload in space has to get built, and it could get built. 804 00:53:39,770 --> 00:53:42,817 And as far as I'm concerned, from a Microsoft standpoint, I would love to. 805 00:53:42,900 --> 00:53:49,747 I mean, I think we have a few instances where we have already had some Azure SKUs. 806 00:53:49,830 --> 00:53:54,167 We had a program where we even put Azure SKUs in space 807 00:53:54,250 --> 00:53:57,447 and what have you, but they were more like edge. 808 00:53:57,530 --> 00:54:01,087 And so to the degree to which if somebody says to me 809 00:54:01,170 --> 00:54:06,507 that there's gigawatts available or megawatts available, 810 00:54:06,590 --> 00:54:12,247 I'm happy to plug myself in. Next question. 811 00:54:12,330 --> 00:54:19,167 Yeah. So it seems that Meta is now pulling back from building frontier open models. 812 00:54:19,250 --> 00:54:21,687 And I understand that Google [INAUDIBLE] 813 00:54:21,770 --> 00:54:26,717 doing some work with open models or flash and [INAUDIBLE], 814 00:54:26,800 --> 00:54:31,717 but they tend to be pretty small for writing like one [INAUDIBLE]. 815 00:54:31,800 --> 00:54:36,957 Given how you personally embrace open source, the biggest example being Linux. 816 00:54:37,040 --> 00:54:39,957 I'm curious, do you see Microsoft and AI, 817 00:54:40,040 --> 00:54:44,517 OpenAI building frontier open models like at the parameter 818 00:54:44,600 --> 00:54:50,177 [INAUDIBLE]. Yeah, I think the thing that we're focused on, 819 00:54:50,260 --> 00:54:52,777 we definitely will always have open rate models. 820 00:54:52,860 --> 00:54:56,297 And to your point, they will be more for what we will ship. 821 00:54:56,380 --> 00:54:59,337 In fact, we launched two even yesterday, 822 00:54:59,420 --> 00:55:04,897 both an instruct model and a plan model for local agent loop and what have you. 823 00:55:04,980 --> 00:55:07,997 So they are derivatives of what we have done with Phi Silica 824 00:55:08,080 --> 00:55:11,697 before called Ion Instruct and Ion Plan. 825 00:55:11,780 --> 00:55:14,997 And they'll run on Windows and they'll be open rate. 826 00:55:15,080 --> 00:55:18,597 The thing that we have focused on with the MAI 827 00:55:18,680 --> 00:55:22,727 lineage of models is, again, think of them as licensed, 828 00:55:22,810 --> 00:55:25,347 but we are going to license them pretty broadly. 829 00:55:25,430 --> 00:55:26,927 So for example, you can go to base 10. 830 00:55:27,010 --> 00:55:29,547 You can go to fireworks and you can then fine tune 831 00:55:29,630 --> 00:55:32,187 even using their inference stack. 832 00:55:32,270 --> 00:55:37,487 And so on. But the reason why we're doing that is because we want quite frankly, 833 00:55:37,570 --> 00:55:40,027 every company, whether it's a SaaS company, an AI 834 00:55:40,110 --> 00:55:42,107 native or an enterprise company to have 835 00:55:42,190 --> 00:55:48,217 their own model that they can then post-train, they can RL, 836 00:55:48,300 --> 00:55:51,307 and what have you. So therefore, that'll be our goal 837 00:55:51,390 --> 00:55:57,907 is to build an ecosystem around the MAI lineage of models. 838 00:55:57,990 --> 00:56:02,707 And the reason why we want to make sure they're still licensed 839 00:56:02,790 --> 00:56:09,027 is because at this point, there's going to be real need for inspection, safety, 840 00:56:09,110 --> 00:56:12,897 so there's the accounts at which we are. 841 00:56:12,980 --> 00:56:14,687 And even if you look at the Chinese models, 842 00:56:14,770 --> 00:56:17,247 they're also quickly becoming closed source and so on. 843 00:56:17,330 --> 00:56:22,687 So I think there will be. And I know Jensen's working on some open rate models, 844 00:56:22,770 --> 00:56:24,847 so we're definitely supporters of it. 845 00:56:24,930 --> 00:56:29,227 But we want to make sure that we are all leading. 846 00:56:29,310 --> 00:56:31,787 The ultimate goal here is to have everyone 847 00:56:31,870 --> 00:56:36,667 have real agency in being able to take some model 848 00:56:36,750 --> 00:56:41,027 and to be able to then add to it and then protect it 849 00:56:41,110 --> 00:56:47,347 from having it go back. So I think that they are [INAUDIBLE]. 850 00:56:47,430 --> 00:56:49,047 It won't be open. It will be licensed. 851 00:56:49,130 --> 00:56:52,307 What does that? So we'll license the weights. 852 00:56:52,390 --> 00:56:56,067 OK. And then how does, if someone's using it on fireworks 853 00:56:56,150 --> 00:57:00,527 or together or something like that, how does that help y'all at Microsoft? 854 00:57:00,610 --> 00:57:04,307 It will be licensed. And so therefore, we will have an economic model 855 00:57:04,390 --> 00:57:07,367 in all of this. Well, I think we're out of time. 856 00:57:07,450 --> 00:57:09,367 But thank you so much for coming. 857 00:57:09,450 --> 00:57:13,260 Thank you so much. Fantastic and--