WEBVTT 1 00:00:04.240 --> 00:00:07.080 With a round segle and from London and Gerard read 2 00:00:07.160 --> 00:00:10.080 from Berlin, this is redefining energy. 3 00:00:10.800 --> 00:00:15.560 Today. On redefining Energy, Child, we're going to talk about 4 00:00:15.960 --> 00:00:18.760 European sovereign neo cloud. 5 00:00:18.960 --> 00:00:23.280 Yeah, which is obviously a critical topic in this AI 6 00:00:23.600 --> 00:00:25.760 revolution that we're currently at the beginnings of. 7 00:00:26.199 --> 00:00:29.879 Yes, job, because when you see our investment right now, 8 00:00:30.359 --> 00:00:34.960 it is the biggest capital expenditure cycle in history. We 9 00:00:35.159 --> 00:00:38.799 are seeing statistics that the kapex is going to reach 10 00:00:39.159 --> 00:00:43.000 nine percent of global GDP, never seen before the peak 11 00:00:43.000 --> 00:00:46.280 of the railroad boom in the eighteen eighties or six 12 00:00:46.280 --> 00:00:50.799 percent of GDP. And the current AI data center built 13 00:00:51.079 --> 00:00:55.039 is bigger in the US than the rest of the 14 00:00:55.159 --> 00:00:59.920 construction company. So we are really in extremely out sector. 15 00:01:00.479 --> 00:01:03.520 But it's very US and when you're in Europe you 16 00:01:03.560 --> 00:01:04.920 have a different point of view. 17 00:01:05.280 --> 00:01:08.280 Well, listen, I think we be really honest with ourselves. 18 00:01:08.319 --> 00:01:12.959 What's happened in Europe is AI regulations have blocked the 19 00:01:13.040 --> 00:01:16.159 development of AI companies across Europe. So we don't have 20 00:01:16.280 --> 00:01:22.200 any foundational model companies like the Chatchbts of this world. Yeah, 21 00:01:22.239 --> 00:01:25.120 and we're playing catch up, and I think there's a 22 00:01:25.159 --> 00:01:28.439 growing realization that actually, you know, one of the advantages 23 00:01:28.439 --> 00:01:31.879 that Europe has in particular businesses is they have data 24 00:01:31.920 --> 00:01:35.239 and do you certainly want that data sitting in another country, 25 00:01:35.840 --> 00:01:37.840 And even if it's not in another country, you want 26 00:01:37.840 --> 00:01:40.239 to make sure that that data is not leaking out 27 00:01:40.280 --> 00:01:43.159 to the rest of the world. There's also growing concern 28 00:01:43.359 --> 00:01:45.400 that you don't want to be just sticking with all 29 00:01:45.480 --> 00:01:49.439 your data stuck on AWS or Google or any of 30 00:01:49.439 --> 00:01:53.439 these other big American players. We need to really make sure, 31 00:01:53.560 --> 00:01:56.159 as I said, that your data is your data. 32 00:01:56.599 --> 00:02:01.640 Absolutely so. We have a supper guest, Michelle Butwelle is 33 00:02:01.680 --> 00:02:06.040 the co founder and CEO of a German complete called Polarize. 34 00:02:06.480 --> 00:02:10.719 Polarize is one of the few European what we call 35 00:02:10.840 --> 00:02:16.800 neo cloud and is a partner of Envigad and his 36 00:02:17.080 --> 00:02:22.360 mission is to build sovereign AI factories high performance across 37 00:02:22.400 --> 00:02:23.560 Europe and. 38 00:02:23.560 --> 00:02:25.520 Ron I think what we also want to do is 39 00:02:25.639 --> 00:02:29.919 we want to thank the BMW Foundation Herbert Kun for 40 00:02:30.000 --> 00:02:32.840 their support and those of you who don't know the 41 00:02:32.879 --> 00:02:36.599 foundation what they're all about as uniting leaders across sectors 42 00:02:36.639 --> 00:02:41.599 to develop solutions that foster innovative economy and a future 43 00:02:41.639 --> 00:02:44.919 proof society. A key focus is the energy transition and 44 00:02:44.919 --> 00:02:50.000 climate change, where the Foundation drives international collaboration to accelerate 45 00:02:50.039 --> 00:02:54.759 the energy transition. Obviously, with rising energy demand from AI 46 00:02:54.919 --> 00:02:59.039 and data centers, new partnerships, effective collaboration, exchange, a science 47 00:02:59.080 --> 00:03:04.199 based solutions and strategies are not just the samsual but 48 00:03:04.599 --> 00:03:07.280 like they're critical really going forward. 49 00:03:07.840 --> 00:03:11.280 Yeah, for our listeners, we had to learn what the 50 00:03:11.560 --> 00:03:14.800 neo cloud company is because most of the time we're 51 00:03:14.879 --> 00:03:18.400 outside the data centers, so I'm more like energy in 52 00:03:18.439 --> 00:03:21.520 real estate, but here we are going inside the data center. 53 00:03:22.319 --> 00:03:26.080 So a real definition I just found a neo cloud 54 00:03:26.159 --> 00:03:31.199 company is a vertically integrated AI era infrastructure provider that 55 00:03:31.360 --> 00:03:34.919 controls the land power, data center build out and GPU 56 00:03:34.960 --> 00:03:39.560 cluster end to end offering hyperscal a great compute without 57 00:03:39.599 --> 00:03:44.199 being a traditional cloud like a WS or Azure. So 58 00:03:44.360 --> 00:03:47.520 the type of company which are now clouds in the US, 59 00:03:47.599 --> 00:03:52.039 you have Core with Amnhibius, Cruso, and so they are 60 00:03:52.120 --> 00:03:55.319 different from what they call the powerkshell the whole one, 61 00:03:55.439 --> 00:03:59.400 the aliks the advantage the cyrus one aligned data centers, 62 00:04:00.000 --> 00:04:03.000 you know, which provide the infrastator from the outside, and 63 00:04:03.080 --> 00:04:06.840 of course you've got another layer the original one, the 64 00:04:07.000 --> 00:04:12.639 traditional collocation reads Equinique Digital Reality, which are multi tenant. 65 00:04:13.199 --> 00:04:18.120 It's more retail wholesale focus. And then you've got the 66 00:04:18.160 --> 00:04:22.000 eper scalers and on the top of that a lot 67 00:04:22.079 --> 00:04:26.079 of money. The Blackstone, the Kickr, the Brookfield, Stone Peak 68 00:04:26.160 --> 00:04:32.839 Digital Bridge, they're all financing this absolute crazy build. We'll 69 00:04:32.879 --> 00:04:36.079 see what you're up as to reply to that US push. 70 00:04:36.720 --> 00:04:38.360 So let's bring Michelle on the show. 71 00:04:39.120 --> 00:04:40.439 Michelle, Welcome to the show. 72 00:04:41.160 --> 00:04:41.879 Thanks for having me. 73 00:04:42.639 --> 00:04:45.079 Well, Michelle, maybe i'd like to kick off here now, 74 00:04:45.120 --> 00:04:48.600 and I'm going out with I spose a thesis that 75 00:04:48.959 --> 00:04:51.399 really were in the fifth Industrial Revolution and this is 76 00:04:51.439 --> 00:04:54.839 the biggest industrial revolution the history of mankind. And there's 77 00:04:54.879 --> 00:04:59.839 two things. One is it's AI and AI needs alecturously 78 00:05:00.040 --> 00:05:04.439 and actually a funny way, electricity needs AI. I'd really 79 00:05:04.600 --> 00:05:07.199 love to firstly ask whether you agree with that view, 80 00:05:07.759 --> 00:05:10.759 and then secondly just talk about Europe's role in this 81 00:05:10.920 --> 00:05:12.560 revolution going forward and how you see it. 82 00:05:13.480 --> 00:05:15.600 Yeah, I totally agree with that, and I would even 83 00:05:15.639 --> 00:05:18.639 go a little bit further. It's not just an industrial revolution, 84 00:05:19.079 --> 00:05:22.040 it's also like a revolution of the society because we 85 00:05:22.199 --> 00:05:25.319 have to adapt to all these things which AI brings 86 00:05:25.360 --> 00:05:27.319 to us. For example, it changed the way how we 87 00:05:27.439 --> 00:05:29.920 work in the future. It changed the way how we 88 00:05:30.160 --> 00:05:33.439 it's gonna spend time, you know, always work in silence 89 00:05:33.439 --> 00:05:35.560 of forth. So that means yeah, I'm totally agree to that. 90 00:05:36.120 --> 00:05:38.240 My second part to the question then really is with 91 00:05:38.399 --> 00:05:41.839 in and around Europe. What's Europe's role now, because again 92 00:05:42.120 --> 00:05:44.600 stand back and look at it, I see us doing 93 00:05:44.639 --> 00:05:48.879 a huge amount, particularly in and around learning for national models. 94 00:05:49.199 --> 00:05:52.120 China going a slightly different approach, which is a more 95 00:05:52.160 --> 00:05:54.319 open source, and then I sort of see Europe and 96 00:05:54.439 --> 00:05:56.160 I wonder what Europe is actually doing. 97 00:05:58.399 --> 00:06:00.399 This is the problem because Europe is not doing much 98 00:06:00.480 --> 00:06:02.720 at the moment. The biggest issue is if we are 99 00:06:02.839 --> 00:06:05.759 not starting now, we are missing the train. Because you 100 00:06:05.839 --> 00:06:08.079 see what is happening in America, you see what is 101 00:06:08.120 --> 00:06:10.639 happening in China. They're taking a major role in this 102 00:06:10.759 --> 00:06:13.360 AI business. And it means as soon as the European 103 00:06:13.519 --> 00:06:17.399 economy starts to adapt to AI, it will it needs 104 00:06:17.399 --> 00:06:21.160 a choice, and if there's no choice from a European perspective, 105 00:06:21.240 --> 00:06:23.519 then they will go to an hyperskiller, to an American one, 106 00:06:23.639 --> 00:06:25.759 or to a Chinese one, and they get sticky and 107 00:06:25.879 --> 00:06:28.160 then they can't go back anymore or not so easy 108 00:06:28.240 --> 00:06:31.800 back anymore. And it's not just about sovereignty. Is also 109 00:06:32.000 --> 00:06:35.160 just it's about how we're gonna give out our datas 110 00:06:35.279 --> 00:06:38.040 for free to these models to learn to replace us 111 00:06:38.040 --> 00:06:38.560 in the future. 112 00:06:38.600 --> 00:06:39.560 And this is a real threat. 113 00:06:40.199 --> 00:06:42.600 And in my opinion, we're doing a big mistake in 114 00:06:42.680 --> 00:06:47.079 Europe because we are trying to replicate things from America 115 00:06:47.279 --> 00:06:49.560 like we did in the European Union. There's a tender. 116 00:06:49.959 --> 00:06:52.519 The tender says we want to have one hundred thousand 117 00:06:52.720 --> 00:06:56.040 chips in one giga factory. And to be honest, I 118 00:06:56.199 --> 00:06:58.480 think this paper was made by a lot of lawyers, 119 00:06:58.519 --> 00:07:02.160 but no technical guys sitting there and explaining what is 120 00:07:02.279 --> 00:07:03.480 really needed in Europe. 121 00:07:03.839 --> 00:07:07.360 So Michelle jumping on that, it's tough because let's go 122 00:07:07.439 --> 00:07:09.360 back to the fact that it is a revolution that 123 00:07:09.399 --> 00:07:12.040 we're going through. And I be critical of the European 124 00:07:12.120 --> 00:07:14.839 Union too, because when I look at their AI Directive 125 00:07:14.959 --> 00:07:17.399 and all this type of stuff, it just slows down. 126 00:07:17.480 --> 00:07:22.399 You're trying to regulate something that is changing too quickly 127 00:07:23.279 --> 00:07:25.560 and we're not going about it in the right way. 128 00:07:25.639 --> 00:07:28.040 So I'm totally witcher and I suppose the question then 129 00:07:28.120 --> 00:07:29.439 to ask is well, what should we do? 130 00:07:30.480 --> 00:07:33.079 First of all, we as polarized trying to build something 131 00:07:33.120 --> 00:07:35.439 which is not existing at the moment. In Europe, it's 132 00:07:35.480 --> 00:07:38.399 called the neo cloud, so means a cloud who's specialized 133 00:07:38.439 --> 00:07:42.040 in AYE for a eyepurpose like AI infrastructure. So we 134 00:07:42.199 --> 00:07:46.199 need to enable European companies, startups and solence of force 135 00:07:46.519 --> 00:07:48.800 so that they are starting to build something that they're 136 00:07:48.839 --> 00:07:53.040 going to invest heavily into data center, into AI. If 137 00:07:53.160 --> 00:07:55.439 without that, we can't catch up anymore in my opinion, 138 00:07:55.879 --> 00:07:59.040 So we don't have to regulate everything beforehand, which we 139 00:07:59.160 --> 00:08:01.360 need to have some form of sandboxes where we can 140 00:08:01.439 --> 00:08:05.439 start to work with. Startups needs to be supported, scale 141 00:08:05.519 --> 00:08:07.439 ups needs to be supported. We don't have to look 142 00:08:07.600 --> 00:08:09.279 just at the big carps. We have to look like 143 00:08:09.399 --> 00:08:12.720 a mixture of everything. And if we're not starting now 144 00:08:12.839 --> 00:08:15.720 to trust in ourselves means that we're going to lose. 145 00:08:15.600 --> 00:08:16.279 On the long end. 146 00:08:16.480 --> 00:08:20.160 In my opinion, do I understand it, and that's coming 147 00:08:20.240 --> 00:08:23.360 from somebody who doesn't understand much, is that you've got 148 00:08:23.399 --> 00:08:27.079 to type of AI work. You've got the work where 149 00:08:27.079 --> 00:08:30.319 basically you run those big models. I have no idea 150 00:08:30.360 --> 00:08:32.399 how it works, but it looks like it consumes a 151 00:08:32.480 --> 00:08:34.919 lot of power. And then you've got what they call 152 00:08:35.000 --> 00:08:38.799 the inference, which I finally understood what it means. Basically, 153 00:08:39.000 --> 00:08:42.279 it's the asset management. Once the models are running, so 154 00:08:42.600 --> 00:08:44.960 do you concentrate on the first and on the second. 155 00:08:46.080 --> 00:08:49.000 So we are concentrating on the second because maybe I'm 156 00:08:49.000 --> 00:08:53.360 gonna make it a little more understandable. Large language models 157 00:08:53.440 --> 00:08:55.519 needs to be trained first, so that means we have 158 00:08:55.759 --> 00:08:58.399 to raise the child. We have to give them the 159 00:08:58.480 --> 00:09:01.080 ability to speak and to think. This is let's say 160 00:09:01.240 --> 00:09:04.639 training in the world, and we don't have so much 161 00:09:04.720 --> 00:09:07.480 foundation model, so we don't have so much training in Europe. 162 00:09:07.559 --> 00:09:09.600 So this is done by OpenAI and by all the 163 00:09:09.639 --> 00:09:13.320 big guys. And the inferencing is like AI in operation, 164 00:09:13.480 --> 00:09:16.200 So that means as soon as you use AI, you 165 00:09:16.320 --> 00:09:18.919 are implementing into your workflows and it starts to do 166 00:09:19.120 --> 00:09:22.879 something for you. Then it's called inferencing because like consuming 167 00:09:23.399 --> 00:09:26.840 constantly tokens, and tokens is they your form of energy? 168 00:09:26.879 --> 00:09:30.600 So our computers are transforming energy into tokens. The tokens 169 00:09:30.600 --> 00:09:33.679 are consumed by the models, and this is how it operates, 170 00:09:34.080 --> 00:09:36.960 and we are specialized in inferencing, and because what we 171 00:09:37.159 --> 00:09:40.240 think is that we don't need to have so much 172 00:09:40.279 --> 00:09:42.559 foundation mode, because there are a lot of open source 173 00:09:42.639 --> 00:09:45.600 models out there. We can use these open source models, 174 00:09:45.639 --> 00:09:48.360 we can fine tune them, we can put them on 175 00:09:48.440 --> 00:09:51.759 an infrastructure which is sovereign, and then we can start 176 00:09:51.840 --> 00:09:54.360 to use it in Europe or European companies can use 177 00:09:54.360 --> 00:09:58.320 it because we need it, because otherwise we can't compete 178 00:09:58.480 --> 00:10:00.960 with other companies in the world who have this advantage 179 00:10:01.000 --> 00:10:03.519 of the eye. So we have two choices. One choice 180 00:10:03.639 --> 00:10:07.320 is we go with foundation model, which are proparatary models 181 00:10:07.399 --> 00:10:10.480 from the Googles of the world, and then that means 182 00:10:10.519 --> 00:10:13.320 we're giving the datas into this these models are learning 183 00:10:13.399 --> 00:10:15.600 from it, and then it could be in the future 184 00:10:15.639 --> 00:10:17.960 it could be possible that they're going to replace work 185 00:10:18.600 --> 00:10:22.000 which was done here at European companies. Or we can 186 00:10:22.200 --> 00:10:25.360 use open source model, we put them on a European 187 00:10:25.399 --> 00:10:29.279 based infrastructure with European companies and then it has the 188 00:10:29.320 --> 00:10:31.399 same effect and then the companies can use the I 189 00:10:31.559 --> 00:10:31.879 as well. 190 00:10:32.559 --> 00:10:35.399 And this is what we're doing in Polarized Michelle, I 191 00:10:35.799 --> 00:10:37.559 just want to go back to open source here again 192 00:10:37.639 --> 00:10:40.120 because if I look and I see all the tests 193 00:10:40.320 --> 00:10:42.440 and studies that have been done. Now, if we went 194 00:10:42.559 --> 00:10:45.879 back three years ago, there was like a year's difference 195 00:10:46.360 --> 00:10:50.639 in terms of the quality of an open source one 196 00:10:50.759 --> 00:10:53.440 versus a proprietary system, but now it's like one month. 197 00:10:54.120 --> 00:10:56.200 And so am I wrong to look at it and say, 198 00:10:56.799 --> 00:10:59.440 who cares if it's proprietary or not? If it's open source? 199 00:10:59.519 --> 00:11:01.279 Is a way to go going forward? Which would mean 200 00:11:01.919 --> 00:11:03.919 if I'm a European and I want to catch up, 201 00:11:04.440 --> 00:11:07.159 I really want to be the fast follower. I don't 202 00:11:07.159 --> 00:11:08.799 want to be the leader the US who's going to 203 00:11:08.840 --> 00:11:10.200 do and all this spend, all this money and all 204 00:11:10.240 --> 00:11:12.120 this foundation. Might I just go and take these open 205 00:11:12.159 --> 00:11:15.480 source models and I go directly into inference. Am I 206 00:11:15.759 --> 00:11:18.399 thinking about this right? Wrong? How do you think about it? 207 00:11:19.440 --> 00:11:21.559 You're absolutely thinking in the right way. So I think 208 00:11:21.679 --> 00:11:23.960 that the open source models has catched up so fast, 209 00:11:24.039 --> 00:11:26.600 and they are pretty good, and you have so much varieties, 210 00:11:26.720 --> 00:11:29.559 and you also have to think how much performance do 211 00:11:29.679 --> 00:11:32.720 you need? How much parameters are really necessary? If you 212 00:11:32.799 --> 00:11:35.480 look at the biggest foundation models in the market and 213 00:11:35.600 --> 00:11:38.080 you see the parameters, I would say in OPO's four 214 00:11:38.120 --> 00:11:41.159 point zero is not necessary for ninety nine percent of 215 00:11:41.240 --> 00:11:43.399 the people who are using EYE at the moment, So 216 00:11:43.480 --> 00:11:46.279 that means you don't need this super super super high 217 00:11:46.279 --> 00:11:49.720 intelligent parameters. But you can go with open source. You 218 00:11:49.799 --> 00:11:51.720 can save a lot of costs and you can also 219 00:11:51.799 --> 00:11:54.480 save your datas and it will be at the end 220 00:11:54.480 --> 00:11:57.159 of the day, the open source model will your personally 221 00:11:57.240 --> 00:12:00.000 EYE model in my opinion, and that is super interesting. 222 00:12:00.240 --> 00:12:02.039 So that means, yes, I'm totally following what you. 223 00:12:02.080 --> 00:12:05.559 Say if I understand. Well, you say you're a Neil cloud, 224 00:12:05.639 --> 00:12:09.480 so I know because that's kind of funny. Michael Intrator 225 00:12:09.639 --> 00:12:12.039 used to be my carbon broker twenty five years ago. 226 00:12:12.320 --> 00:12:15.080 Well it's not very well the sea of core weave now, 227 00:12:15.720 --> 00:12:20.759 so you kind of a European core weave. Correct, correct, Yes, Okay. 228 00:12:21.000 --> 00:12:23.399 What we do know from the energy guys is basically 229 00:12:23.519 --> 00:12:27.799 what Rowan is doing created by Quinn Brooke, fantastic company, 230 00:12:27.879 --> 00:12:31.519 but the others it's basically all outside the data center. 231 00:12:31.720 --> 00:12:34.679 So you know, they arrived, they put a terrain, they 232 00:12:34.759 --> 00:12:39.440 build the all infrastructure, and basically they deliver you or 233 00:12:39.519 --> 00:12:43.240 they deliver somebody a building which is ready to be 234 00:12:43.639 --> 00:12:46.320 powered and cool and everything. And that's where you come 235 00:12:46.360 --> 00:12:48.440 in explain a bit what you're doing. 236 00:12:49.200 --> 00:12:50.919 No, it's in fact, it's not where we are coming. 237 00:12:50.960 --> 00:12:53.679 We'recoming a stage earlier. Because if you look at the 238 00:12:53.759 --> 00:12:56.399 actual data center industry in the world, and it's mainly 239 00:12:56.600 --> 00:12:58.639 let's say cloud business like storage and so on and 240 00:12:58.759 --> 00:13:02.600 so force, it was like that the infrastructure, so the 241 00:13:02.759 --> 00:13:06.919 real estate building was thought from real estate guys, and 242 00:13:07.039 --> 00:13:10.639 then the computing which was implemented there was staying in 243 00:13:10.759 --> 00:13:13.679 the same shape for years, for maybe ten fifteen years. 244 00:13:14.080 --> 00:13:17.159 So inside the data center you had like a real 245 00:13:17.279 --> 00:13:19.679 estate part, and then in the in the real estate part, 246 00:13:19.759 --> 00:13:22.720 we had like a standard of some form of it 247 00:13:22.879 --> 00:13:25.879 called rex. It's like shelfs and then these shelvels had 248 00:13:26.200 --> 00:13:30.200 around fifteen to twenty kilowat each shelf today and we 249 00:13:30.360 --> 00:13:32.320 are seeing it the other way around because what we 250 00:13:32.440 --> 00:13:34.080 have understand and this is why we are building the 251 00:13:34.159 --> 00:13:36.720 data center ourselves, because at the moment, there are no 252 00:13:37.039 --> 00:13:40.679 co location data center which are suitable for the II infrastructure, 253 00:13:40.759 --> 00:13:43.399 so to say, because you have to think backwards, you 254 00:13:43.519 --> 00:13:46.639 have to think from the it back to the real estate. 255 00:13:47.159 --> 00:13:51.279 And there's a lot of differences between the old data 256 00:13:51.360 --> 00:13:53.600 center economy and the old data center economy. 257 00:13:53.679 --> 00:13:56.440 It will not be ending. It will stay as it is. 258 00:13:56.519 --> 00:13:58.519 We're still going to consume data is We're still going 259 00:13:58.600 --> 00:14:01.360 to have Internet, We're still going to have storage, and 260 00:14:01.440 --> 00:14:04.399 then we'll also have a gross every year. And then 261 00:14:04.440 --> 00:14:06.600 there is a new form of data center. These are 262 00:14:06.639 --> 00:14:08.919 the AI factories, So we are not calling them any 263 00:14:09.000 --> 00:14:11.960 more data centers. We are calling them AI factories, and 264 00:14:12.080 --> 00:14:15.720 these AI factory are totally different constructed. First of all, 265 00:14:15.799 --> 00:14:18.919 we are not going with fifteen kilowatt uperrac. We are 266 00:14:19.000 --> 00:14:21.720 going with up to one hundred and fifteen more. And 267 00:14:21.840 --> 00:14:24.080 we have to think not just about the generation of 268 00:14:24.200 --> 00:14:27.159 compute today, we also have to think about the generation 269 00:14:27.279 --> 00:14:30.960 of compute of tomorrow because the iteration of compute is 270 00:14:31.000 --> 00:14:34.080 going so fast. So every GPU model it will be 271 00:14:34.200 --> 00:14:36.240 replaced in the next three to five years. So we're 272 00:14:36.240 --> 00:14:38.960 going to have an efficiency every three to five years 273 00:14:39.000 --> 00:14:42.600 in GPUs, and we're going to consume more energy per 274 00:14:42.759 --> 00:14:45.919 GPU in the same amount of time. It always doubles 275 00:14:45.960 --> 00:14:48.639 the town sometimes even triples, and that needs to be 276 00:14:48.759 --> 00:14:50.799 sought as well. So if you're going in a classical 277 00:14:50.879 --> 00:14:53.559 data center, and if you don't think about all these 278 00:14:53.960 --> 00:14:56.919 different elements you're going to be outdated super fast in 279 00:14:57.000 --> 00:14:59.519 the data center world. So that means you need to 280 00:14:59.639 --> 00:15:02.480 think differently in THEI factories than you did in the 281 00:15:02.600 --> 00:15:05.799 past with normal data centers. Normal data center are not 282 00:15:05.960 --> 00:15:09.799 suitable for AI, and that means there's a totally new 283 00:15:09.879 --> 00:15:13.000 classification of data center coming into the market with a 284 00:15:13.200 --> 00:15:17.279 big hunger of energy, but with a lot less space. 285 00:15:18.000 --> 00:15:21.200 So for example, we are putting per square meter a 286 00:15:21.279 --> 00:15:24.639 lot more energy than a normal classical data center, So 287 00:15:24.679 --> 00:15:27.679 it means we can build in two thousand square meters, 288 00:15:27.759 --> 00:15:30.519 we can build twenty twenty five megawats. 289 00:15:31.080 --> 00:15:33.360 Very interesting you're saying about energy there. But I'm going 290 00:15:33.440 --> 00:15:35.879 to ask now, just coming from the energy side, when 291 00:15:35.919 --> 00:15:37.919 you look at data centers going forward, are we going 292 00:15:37.960 --> 00:15:40.039 to build big, huge ones or are they going to 293 00:15:40.080 --> 00:15:42.200 be small? In other words, are we going grade edge? 294 00:15:42.279 --> 00:15:46.120 And if it is grid edge, jesus fifty megawats, twenty 295 00:15:46.159 --> 00:15:49.159 megawards five one? How do you see the future in 296 00:15:49.240 --> 00:15:53.039 Europe of building these data centers or sorry AI factory, 297 00:15:53.120 --> 00:15:55.519 sorry AI factories, no whorries. 298 00:15:55.799 --> 00:15:57.360 So I see it in the same way like we're 299 00:15:57.360 --> 00:16:00.600 producing energy today, so we have like big new power 300 00:16:00.679 --> 00:16:03.919 plants in Europe. We have smaller coal plants, we have 301 00:16:04.080 --> 00:16:06.879 gas plants, we have windmills, We have a lot of 302 00:16:07.039 --> 00:16:10.840 different power production all over Europe, which is necessary because 303 00:16:10.879 --> 00:16:14.039 you need sometimes a really big power plant to produce 304 00:16:14.039 --> 00:16:15.879 a lot of energy because you have a region, we 305 00:16:16.039 --> 00:16:18.799 have big cities and so on, and the force whereas necessary. 306 00:16:19.240 --> 00:16:21.519 I see the same in AI. So of course we're 307 00:16:21.519 --> 00:16:23.879 going to have one hundred megawatter one hundred and fifty 308 00:16:23.960 --> 00:16:26.799 megawatt big AI factories, but we're going to also have 309 00:16:26.960 --> 00:16:30.200 like twenty megawat factories. We're going to have fifty megawatt factories. 310 00:16:30.519 --> 00:16:32.000 And what we think, we. 311 00:16:32.120 --> 00:16:34.240 See it as a and we call it giga grit 312 00:16:34.320 --> 00:16:38.200 in polarized. So we don't need these big centralized single 313 00:16:38.279 --> 00:16:40.919 point of failures in my opinion. So what we need 314 00:16:41.120 --> 00:16:44.480 is all the mixture of different form of AI factory 315 00:16:44.480 --> 00:16:48.000 which fits into the region and into the market. For example, 316 00:16:48.039 --> 00:16:50.120 if you're living in Munich and we have one data 317 00:16:50.200 --> 00:16:52.360 center or one AI factory in Munich where we are 318 00:16:52.759 --> 00:16:56.120 running ten thousand GPUs, it has around twenty five megawatts, 319 00:16:56.559 --> 00:16:59.320 so it's it's in terms of data center as standard 320 00:16:59.360 --> 00:17:02.919 data centers quite huge, but in terms of AI factories 321 00:17:03.080 --> 00:17:06.200 is pretty small and with this ten thousand GPUs, we 322 00:17:06.279 --> 00:17:10.160 have doubled the AI capacity in Germany at the moment. 323 00:17:10.920 --> 00:17:13.079 That means at the same time there is a lot 324 00:17:13.200 --> 00:17:15.920 more hunger for GPUs and AI factory, so we need 325 00:17:16.000 --> 00:17:18.519 to build them fast. And in my opinion, there's a 326 00:17:18.559 --> 00:17:20.440 lot of options in the market if you want to 327 00:17:20.480 --> 00:17:23.279 go fast, because if you build a two hundred megawat 328 00:17:23.400 --> 00:17:26.279 or even bigger AI factory, it takes time because the 329 00:17:26.359 --> 00:17:28.759 grid connection is not so easy. You need to find 330 00:17:28.799 --> 00:17:31.440 the right point to connect, then you need to build it. 331 00:17:31.920 --> 00:17:34.079 Then you need to have all this residancy. You need 332 00:17:34.119 --> 00:17:36.519 to have a lot of things to build these huge factories. 333 00:17:36.839 --> 00:17:40.720 But if you find an old production fabric which already 334 00:17:40.799 --> 00:17:43.559 has twenty or ten or fifteen mega what's in place, 335 00:17:44.160 --> 00:17:47.960 and you are capable to retrofit it, and you have 336 00:17:48.119 --> 00:17:52.680 the creativity how to transform a former whatever factory it was, 337 00:17:53.119 --> 00:17:57.200 into an AI factory and using the same infrastructure which 338 00:17:57.359 --> 00:17:59.799 was already there, and you just upgraded, you just put 339 00:17:59.839 --> 00:18:02.160 it in a way that is suitable for an AI factory, 340 00:18:02.200 --> 00:18:04.880 and then you have a very fast and a very 341 00:18:04.920 --> 00:18:08.240 efficient AI factory. And it's also in my opinion, so 342 00:18:08.359 --> 00:18:11.160 that you don't have to build everything from Squattionally you're 343 00:18:11.200 --> 00:18:13.200 saving a lot of cou two for example, and so 344 00:18:13.400 --> 00:18:15.759 and so forth. And we have in Europe and especially 345 00:18:15.799 --> 00:18:19.359 in Germany, we have so much potential sites and we 346 00:18:19.480 --> 00:18:22.359 see that every day which could be transformed into an 347 00:18:22.400 --> 00:18:25.599 AI factory and being a super efficient and a super 348 00:18:25.640 --> 00:18:29.160 fast way forward. So there is enough options, but the 349 00:18:29.279 --> 00:18:32.880 option needs some creativity from companies like US or others. 350 00:18:33.599 --> 00:18:35.799 And if you just go the normal way like we 351 00:18:35.920 --> 00:18:38.160 did it in the normal data center world, for example 352 00:18:38.240 --> 00:18:40.880 in Frankfurt. So we have in Frankfort a lot of 353 00:18:41.039 --> 00:18:43.519 data centers because we have there the International and the 354 00:18:43.559 --> 00:18:47.319 Transatlantic Internet KNOT and there's DCAKES which is one of 355 00:18:47.400 --> 00:18:51.319 the biggest Internet distribution not in the world. This is 356 00:18:51.359 --> 00:18:53.519 why everybody wants to be in Frankfort. But if you 357 00:18:53.599 --> 00:18:55.640 want to build a data center at Frankfoot, it will 358 00:18:55.720 --> 00:18:59.119 take you at least six seven eight years until you 359 00:18:59.240 --> 00:19:02.279 get the grid concert and it will cost you forty 360 00:19:02.319 --> 00:19:05.319 percent more than other regions, and that will be too 361 00:19:05.359 --> 00:19:07.039 late in my opinion. If you're going to go the 362 00:19:07.119 --> 00:19:07.559 same root. 363 00:19:08.519 --> 00:19:13.119 It's a very interesting approach. It's a very German approach decentralization. 364 00:19:13.880 --> 00:19:16.319 But if you cross the border and you are in France, 365 00:19:16.359 --> 00:19:19.680 where everything is stopped down, decided by the president, the 366 00:19:19.759 --> 00:19:23.079 only thing they're interested is to sign seventy five billion 367 00:19:23.319 --> 00:19:25.480 with a soft bank. Who's going to deliver or not? 368 00:19:26.359 --> 00:19:30.160 Is it compatible? Or who's right, who's wrong, everybody's right? 369 00:19:30.640 --> 00:19:31.519 How does it work? 370 00:19:32.440 --> 00:19:35.000 It is compatible because if you look at the reality 371 00:19:35.039 --> 00:19:37.960 in France, the grid connection times is even longer than 372 00:19:38.079 --> 00:19:40.640 sometimes for example, if you look in Paris region or 373 00:19:40.920 --> 00:19:44.319 Bordeaux region or other bigger regions, it's a nightmare how 374 00:19:44.400 --> 00:19:46.759 long it takes until you get a real data center 375 00:19:46.839 --> 00:19:49.000 on the ground at the end of the day. In 376 00:19:49.079 --> 00:19:52.119 a more centralized country like France, they're going to have 377 00:19:52.200 --> 00:19:54.440 the same promise and there are the same options. You 378 00:19:54.519 --> 00:19:57.640 also have a lot of industries which are not performing 379 00:19:58.160 --> 00:20:01.240 and in the same way, which can be retrofitted into 380 00:20:01.519 --> 00:20:04.400 two AI factories. And we see that all over Europe, 381 00:20:04.480 --> 00:20:07.240 so in Spain, in England, and we have projects in 382 00:20:07.319 --> 00:20:09.680 Spain we haven't we have also a project in England. 383 00:20:10.039 --> 00:20:13.359 The grids in England, it's also like a very bureaucratic 384 00:20:13.680 --> 00:20:16.039 and not an easy way to go. So that means 385 00:20:16.079 --> 00:20:18.400 We have that all over Europe, the same problems, the 386 00:20:18.480 --> 00:20:21.480 same kind of grid constraints, and if you look at 387 00:20:21.519 --> 00:20:24.119 Scandinavia where it's a little bit better, but still also 388 00:20:24.720 --> 00:20:28.079 Scandinavia is now catching up in bureaucratic processes and not 389 00:20:28.279 --> 00:20:31.039 being so easy anymore than there was in the past. 390 00:20:31.720 --> 00:20:34.240 Michelle, I like your approach and I think, okay, yeah, 391 00:20:34.359 --> 00:20:36.920 using existing infrastructure makes a lot of sense to me. 392 00:20:37.359 --> 00:20:39.599 Talk a little bit about the client perspective, In other words, 393 00:20:39.720 --> 00:20:41.519 why would a client go to you rather than go 394 00:20:41.599 --> 00:20:44.960 to Google or Microsoft or someone who they trust and 395 00:20:45.079 --> 00:20:47.039 they know for many years. In other words, what is 396 00:20:47.079 --> 00:20:48.960 it they differentiates you from them? 397 00:20:49.960 --> 00:20:51.880 I think there are three reasons. First of all, we 398 00:20:51.960 --> 00:20:55.960 are not more expensive than taking a hyperscilla. Second, we 399 00:20:56.079 --> 00:20:59.400 are sovereign, means all our clients are protected against the 400 00:20:59.599 --> 00:21:02.759 US out So that means that a government can decide 401 00:21:03.200 --> 00:21:05.599