1 00:00:00,560 --> 00:00:03,239 Speaker 1: This is Tom orlans Reese and you're listening to Switched 2 00:00:03,279 --> 00:00:06,640 Speaker 1: on the podcast brought to you by bloomberg Enif as 3 00:00:06,720 --> 00:00:09,920 Speaker 1: AI drives a new wave of data center expansion, attention 4 00:00:10,039 --> 00:00:12,400 Speaker 1: is shifting to the supply chains that deliver the physical 5 00:00:12,400 --> 00:00:16,439 Speaker 1: infrastructure behind them. Higher density computing is pushing new demands 6 00:00:16,480 --> 00:00:20,400 Speaker 1: across equipment markets, from transformers in switch gear to advanced 7 00:00:20,400 --> 00:00:24,120 Speaker 1: cooling systems and on site energy storage, straining supply chains 8 00:00:24,120 --> 00:00:28,080 Speaker 1: that were never designed for this speed or scale. Manufacturers, developers, 9 00:00:28,080 --> 00:00:32,280 Speaker 1: and utilities are now reworking how data center projects are designed, sourced, 10 00:00:32,400 --> 00:00:35,360 Speaker 1: and built to keep pace with demand. So how are 11 00:00:35,440 --> 00:00:38,159 Speaker 1: data center supply chains evolving and where are the biggest 12 00:00:38,159 --> 00:00:41,560 Speaker 1: constraints and opportunities emerging. On today's show, we bring you 13 00:00:41,600 --> 00:00:44,559 Speaker 1: a discussion on data center supply chain innovation from the 14 00:00:44,600 --> 00:00:47,600 Speaker 1: b and EF Summit San Francisco, where Mark Daily, b 15 00:00:47,720 --> 00:00:51,240 Speaker 1: and EF's head of Technology and Innovation, spoke with Steve Carlini, 16 00:00:51,400 --> 00:00:54,360 Speaker 1: chief advocate for Data Centers and the AI Energy Management 17 00:00:54,400 --> 00:00:57,920 Speaker 1: business Unit at Schneider Electric. Harun Inam, the co founder 18 00:00:57,920 --> 00:01:01,600 Speaker 1: and chief executive officer of DG Matrix, Jill mccari, the 19 00:01:01,680 --> 00:01:05,280 Speaker 1: chief financial Officer of VERE and Angela Taylor, the chief 20 00:01:05,280 --> 00:01:07,920 Speaker 1: of Staff and head of Strategy at Liquid Stack. B 21 00:01:08,080 --> 00:01:11,560 Speaker 1: ANDIF summits bring together experts from across finance, energy and 22 00:01:11,640 --> 00:01:16,039 Speaker 1: technology to share insights, build connections, and explore how innovations 23 00:01:16,040 --> 00:01:20,160 Speaker 1: in clean energy technologies are reshaping global markets. Held in 24 00:01:20,240 --> 00:01:22,880 Speaker 1: cities around the world, they offer the ideas and perspective 25 00:01:23,000 --> 00:01:25,759 Speaker 1: needed to navigate the transition to a cleaner future. If 26 00:01:25,760 --> 00:01:28,400 Speaker 1: you'd like to join us at an upcoming event, including 27 00:01:28,440 --> 00:01:31,760 Speaker 1: the BNEF Summit Beijing on March twelfth, head to BONF 28 00:01:31,840 --> 00:01:34,399 Speaker 1: dot com and click on events, where you can also 29 00:01:34,440 --> 00:01:37,280 Speaker 1: watch highlights and videos from past summits. And now to 30 00:01:37,360 --> 00:01:40,280 Speaker 1: Mark's panel from the bn EF Summit San Francisco titled 31 00:01:40,560 --> 00:01:49,360 Speaker 1: Seeing Beyond the Chips Innovations in Data Center supply Chains. 32 00:01:54,880 --> 00:01:57,800 Speaker 2: We ended that last panel with actually Raj, I mean, 33 00:01:57,880 --> 00:02:00,200 Speaker 2: in case anyone wasn't here, Raj from a line data 34 00:02:00,240 --> 00:02:03,080 Speaker 2: centers kind of said, oh, we're developing these shells that 35 00:02:03,120 --> 00:02:05,720 Speaker 2: you can plug servers into, and really that's the outlet 36 00:02:06,320 --> 00:02:06,880 Speaker 2: and what's. 37 00:02:06,760 --> 00:02:08,040 Speaker 3: Changing is the server. 38 00:02:08,600 --> 00:02:10,400 Speaker 2: But there's actually a little bit more nuanced to it 39 00:02:10,480 --> 00:02:12,200 Speaker 2: than that, and he's kind of perfectly set us up 40 00:02:12,200 --> 00:02:14,520 Speaker 2: for this panel, which is really about how the power 41 00:02:14,560 --> 00:02:17,520 Speaker 2: outlet and all the other ancillary equipment in a data 42 00:02:17,520 --> 00:02:20,880 Speaker 2: center is changing, because obviously, when there's so much capital 43 00:02:21,040 --> 00:02:23,240 Speaker 2: going into a business like this and there's there's always 44 00:02:23,280 --> 00:02:26,920 Speaker 2: opportunity for improving the efficiency and the types of products 45 00:02:26,919 --> 00:02:30,000 Speaker 2: that go into it. So thank you very much for 46 00:02:30,360 --> 00:02:34,440 Speaker 2: joining me here today. So let's start with you, Angela. 47 00:02:35,480 --> 00:02:37,399 Speaker 2: If you read about data centers and kind of what's 48 00:02:37,480 --> 00:02:39,400 Speaker 2: what's changing in data centers right now, one of the 49 00:02:40,080 --> 00:02:43,760 Speaker 2: forgive me hot topics is how the chips are actually cooled. 50 00:02:43,800 --> 00:02:46,040 Speaker 2: And this phrase liquid cooling is what you see in 51 00:02:46,080 --> 00:02:47,600 Speaker 2: all the headlines now, So can you give us a 52 00:02:47,600 --> 00:02:49,200 Speaker 2: bit of a primary like what exactly that means and 53 00:02:49,200 --> 00:02:50,600 Speaker 2: how it's different than what came before. 54 00:02:51,000 --> 00:02:55,960 Speaker 4: Yeah, sure, happy to. So liquid cooling is as it sounds. 55 00:02:56,080 --> 00:03:01,200 Speaker 4: It is cooling the compute within the rack using liquid 56 00:03:01,440 --> 00:03:05,040 Speaker 4: versus air, which is traditionally how it's been done in 57 00:03:05,080 --> 00:03:07,600 Speaker 4: the data center, and they're doing it for a number 58 00:03:07,639 --> 00:03:11,760 Speaker 4: of reasons. Liquid has higher cooling capacity, it has a 59 00:03:11,840 --> 00:03:13,359 Speaker 4: higher thermal connectivity. 60 00:03:13,760 --> 00:03:15,680 Speaker 5: So the reason that you're hearing a. 61 00:03:15,639 --> 00:03:18,919 Speaker 4: Lot about it is because as we start to deploy 62 00:03:19,440 --> 00:03:23,040 Speaker 4: racks that are more dense, moving from forty kilowatts a 63 00:03:23,160 --> 00:03:27,040 Speaker 4: rack up to one hundred and forty right now, next 64 00:03:27,160 --> 00:03:32,120 Speaker 4: eight hundred than perhaps one megawat. You really the physics 65 00:03:32,440 --> 00:03:36,400 Speaker 4: and just economics of air no longer exist. 66 00:03:36,640 --> 00:03:40,040 Speaker 5: So yeah, liquid cooling is here. 67 00:03:40,520 --> 00:03:40,760 Speaker 3: Yeah. 68 00:03:40,960 --> 00:03:43,400 Speaker 2: Great, And just to kind of clarify, it's like when 69 00:03:43,400 --> 00:03:45,560 Speaker 2: you talk about a rack's it's kind of that big 70 00:03:45,600 --> 00:03:47,360 Speaker 2: metal shelf where like it used to be a HPE 71 00:03:47,480 --> 00:03:51,200 Speaker 2: server now it's in Nvidio GPU that costs millions of 72 00:03:51,240 --> 00:03:53,200 Speaker 2: dollars exactly. 73 00:03:53,400 --> 00:03:53,800 Speaker 6: On the. 74 00:03:55,520 --> 00:03:58,240 Speaker 2: Oftentimes, something that I'm asked by my friends is, oh, 75 00:03:58,280 --> 00:04:01,480 Speaker 2: like liquid cooling. It's I'm kind of worried about the 76 00:04:01,520 --> 00:04:04,680 Speaker 2: water consumption of data centers. Is this going to make 77 00:04:04,720 --> 00:04:07,440 Speaker 2: it worse? Is AI just like making data centers even 78 00:04:07,480 --> 00:04:08,280 Speaker 2: more water hungry? 79 00:04:09,840 --> 00:04:16,560 Speaker 4: It depends now, well, so you have to separate when 80 00:04:16,600 --> 00:04:18,760 Speaker 4: you think about data centers. You have to separate legacy 81 00:04:18,839 --> 00:04:22,080 Speaker 4: data centers in AI data centers. So when you think 82 00:04:22,240 --> 00:04:27,160 Speaker 4: about a water consumption of an AI data center, it 83 00:04:27,200 --> 00:04:30,320 Speaker 4: can be very minimal. Because when you implement liquid cooling, 84 00:04:30,400 --> 00:04:34,160 Speaker 4: especially if it's a completely liquid cooled environment over air, 85 00:04:34,360 --> 00:04:39,479 Speaker 4: you can save somewhere between thirty to fifty percent in energy, 86 00:04:39,640 --> 00:04:42,800 Speaker 4: and you can I'm sorry, no, you can save thirty 87 00:04:43,160 --> 00:04:46,520 Speaker 4: to eighty percent in water, not energy. 88 00:04:46,880 --> 00:04:48,840 Speaker 5: So yeah, in the water side. 89 00:04:48,600 --> 00:04:51,440 Speaker 4: Of it, you do have some substantial savings that you 90 00:04:51,480 --> 00:04:55,960 Speaker 4: can realize with liquid cooling. Again, it's going to be 91 00:04:55,960 --> 00:04:58,560 Speaker 4: a foundational technology for the AI side of it. So 92 00:04:58,920 --> 00:05:02,320 Speaker 4: I think your opportunity here is can we implemented in 93 00:05:02,520 --> 00:05:05,640 Speaker 4: the non AI side, in which case you would start 94 00:05:05,680 --> 00:05:08,120 Speaker 4: to see substantial savings in the water side. 95 00:05:08,160 --> 00:05:09,960 Speaker 2: And is that kind of like one of the questions 96 00:05:10,000 --> 00:05:12,800 Speaker 2: for you as a product development officer, like, are you 97 00:05:12,880 --> 00:05:14,680 Speaker 2: thinking about how do I expand this to more types 98 00:05:15,240 --> 00:05:15,960 Speaker 2: data entries? 99 00:05:16,400 --> 00:05:18,400 Speaker 4: Well, you have to you have to have the need, right, 100 00:05:18,440 --> 00:05:22,120 Speaker 4: There's lots of legacy infrastructure that exists there, so there 101 00:05:22,160 --> 00:05:25,320 Speaker 4: has to be some motivation to move into that direction. 102 00:05:25,640 --> 00:05:29,000 Speaker 4: And yeah, of course we do think about that. And 103 00:05:29,080 --> 00:05:32,880 Speaker 4: now that data centers will have more familiarity with liquid 104 00:05:32,880 --> 00:05:35,960 Speaker 4: cooling than perhaps the barrier to entry becomes a little lower. 105 00:05:37,160 --> 00:05:39,400 Speaker 2: And then on the energy efficiency you kind of mentioned that, 106 00:05:39,560 --> 00:05:40,920 Speaker 2: is it is it a positive negative? 107 00:05:40,920 --> 00:05:42,200 Speaker 3: How does liquid cooling impact that? 108 00:05:43,200 --> 00:05:47,479 Speaker 4: Yeah, it depends on such a such an engineering answer. 109 00:05:47,520 --> 00:05:50,880 Speaker 4: It depends it depends on the type of liquid cooling 110 00:05:51,120 --> 00:05:52,360 Speaker 4: that you are implementing. 111 00:05:52,440 --> 00:05:58,000 Speaker 5: So there are or sort of generally accepted types of 112 00:05:58,040 --> 00:05:58,799 Speaker 5: liquid cooling. 113 00:05:59,279 --> 00:06:03,080 Speaker 4: You have two methods, directed hip and immersion, and within 114 00:06:03,200 --> 00:06:06,040 Speaker 4: each of those you have single phase and two phase, 115 00:06:06,480 --> 00:06:10,680 Speaker 4: and the phase just is the liquid phase throughout the solution. 116 00:06:11,320 --> 00:06:15,800 Speaker 4: So when you're thinking about energy in a direct to 117 00:06:15,920 --> 00:06:20,279 Speaker 4: chip solution right now, what's happening is you have a 118 00:06:20,320 --> 00:06:25,400 Speaker 4: cold plate on the hottest components within the board, the 119 00:06:25,400 --> 00:06:30,039 Speaker 4: compute board, and they're circulating water and picking up that 120 00:06:30,080 --> 00:06:33,039 Speaker 4: heat and dispersing it in other parts of the system. 121 00:06:33,839 --> 00:06:38,719 Speaker 4: But you're only really cooling seventy percent, so it's not 122 00:06:38,880 --> 00:06:41,839 Speaker 4: on everything right, and you still need a fan in 123 00:06:41,960 --> 00:06:44,240 Speaker 4: order to be able to cool the other parts of 124 00:06:44,240 --> 00:06:48,080 Speaker 4: it and then pumping around. So that's a little less 125 00:06:48,200 --> 00:06:53,039 Speaker 4: energy efficient, but still more energy efficient than air. In immersion, 126 00:06:53,440 --> 00:06:56,720 Speaker 4: you are dunking it in a dielectric fluid, and that 127 00:06:56,800 --> 00:06:59,440 Speaker 4: has much much greater potential for energy savings. 128 00:07:00,080 --> 00:07:02,640 Speaker 2: Okay, and I mean, Stephen, can you talk a little 129 00:07:02,640 --> 00:07:05,680 Speaker 2: bit about the trade offs between this kind of immersion 130 00:07:05,800 --> 00:07:08,919 Speaker 2: versus directed chip. I mean, Scheider's active in the cooling 131 00:07:08,960 --> 00:07:10,280 Speaker 2: space for data centers. 132 00:07:10,240 --> 00:07:14,240 Speaker 7: Yeah, it's liquid cooling is not new. It's been around 133 00:07:14,320 --> 00:07:19,240 Speaker 7: now for computing over sixty years. So it's not something 134 00:07:19,360 --> 00:07:25,480 Speaker 7: that you know, all of a sudden has become mainstream 135 00:07:25,720 --> 00:07:29,720 Speaker 7: because you know, it uses less water or it's more efficient. 136 00:07:29,840 --> 00:07:32,920 Speaker 7: It's become mainstream because it's impossible to cool the new 137 00:07:33,000 --> 00:07:38,000 Speaker 7: GPUs without without liquid cooling. But we've you know, through 138 00:07:38,040 --> 00:07:42,080 Speaker 7: our acquisitions some of the companies that that we used 139 00:07:42,080 --> 00:07:47,160 Speaker 7: to do crazy supercomputers back in twenty eighteen at four 140 00:07:47,240 --> 00:07:50,520 Speaker 7: hundred kilowatch per rack. So this new high density and 141 00:07:50,560 --> 00:07:54,000 Speaker 7: the challenges you know, as Angelas said, liquid cooling is 142 00:07:54,040 --> 00:07:56,520 Speaker 7: an architecture and if your architecture right, it could use 143 00:07:56,760 --> 00:07:59,440 Speaker 7: minimal use of water. It may it may use a 144 00:07:59,440 --> 00:08:01,440 Speaker 7: little bit more electricity at that point. 145 00:08:01,200 --> 00:08:03,960 Speaker 6: But it will use much less water. 146 00:08:04,480 --> 00:08:09,360 Speaker 7: The different types of cooling that are available today, you know, 147 00:08:09,480 --> 00:08:14,280 Speaker 7: the you know, it depends on you know, where who 148 00:08:14,320 --> 00:08:17,520 Speaker 7: the data center operator is. All of the new data 149 00:08:17,560 --> 00:08:20,440 Speaker 7: centers that are deploying, the Blackwell Ultras and now the 150 00:08:20,520 --> 00:08:24,440 Speaker 7: Vera Rubens, they are all being built from scratch and 151 00:08:24,520 --> 00:08:27,720 Speaker 7: they are going to use director chip liquid cooling. Some 152 00:08:27,800 --> 00:08:31,800 Speaker 7: of the older data centers. It's very, very very very 153 00:08:31,960 --> 00:08:36,080 Speaker 7: tough and expensive to overhaul a ten kilo Why Paract 154 00:08:36,120 --> 00:08:38,480 Speaker 7: data center into one hundred and forty kilo Why Paract 155 00:08:38,559 --> 00:08:41,040 Speaker 7: Data Center? So what a lot of those data centers 156 00:08:41,080 --> 00:08:44,520 Speaker 7: are doing are doing some other types of technologies that 157 00:08:44,600 --> 00:08:46,960 Speaker 7: some people call liquid cooling, some don't like rear door 158 00:08:46,960 --> 00:08:49,480 Speaker 7: heat e changers with red or heating changers. It's you 159 00:08:49,520 --> 00:08:51,360 Speaker 7: could bolt them onto the back of the rack. You 160 00:08:51,400 --> 00:08:54,360 Speaker 7: could use your existor chilling technology, and you can get 161 00:08:54,360 --> 00:08:55,880 Speaker 7: seventy five kilo watch peract. 162 00:08:56,200 --> 00:08:58,439 Speaker 6: So that's much much less disruptive. 163 00:08:58,840 --> 00:09:01,199 Speaker 7: But as angelet's say, you know, we're at one hundred 164 00:09:01,200 --> 00:09:03,960 Speaker 7: and forty kilo watch in six months, there'll be two 165 00:09:04,040 --> 00:09:07,240 Speaker 7: hundred killowats and then before you know it, we're back 166 00:09:07,240 --> 00:09:09,199 Speaker 7: to four hundred, six hundred killowatch. 167 00:09:08,880 --> 00:09:10,559 Speaker 3: For rackt Okay, great. 168 00:09:10,640 --> 00:09:13,080 Speaker 2: I mean you mentioned actually that the Schneider made some 169 00:09:13,120 --> 00:09:16,480 Speaker 2: acquisitions in the space. I think that's this verymuch theme 170 00:09:16,520 --> 00:09:19,160 Speaker 2: we've seen in the heating and cooling space more broadly 171 00:09:19,400 --> 00:09:21,400 Speaker 2: in the economy has been a lot of heat pump acquisitions, 172 00:09:21,520 --> 00:09:24,360 Speaker 2: but specifically relating to data centers, and there's been a 173 00:09:24,400 --> 00:09:25,040 Speaker 2: lot of activity. 174 00:09:25,120 --> 00:09:26,520 Speaker 6: So Schneider, Yeah. 175 00:09:26,400 --> 00:09:30,720 Speaker 7: Yeah, Schneider made an acquisition. We again had a long 176 00:09:30,800 --> 00:09:33,480 Speaker 7: history of cooling. We needed more of a portfolio, we 177 00:09:33,559 --> 00:09:37,520 Speaker 7: needed more production capacity. We just you know, when when 178 00:09:37,880 --> 00:09:40,880 Speaker 7: the Blackstone guy said, oh, it's seven trillion dollars, so 179 00:09:40,920 --> 00:09:43,439 Speaker 7: they you know this, this number seems small. We just 180 00:09:43,600 --> 00:09:46,280 Speaker 7: signed a contract with Switch Data Centers one point nine 181 00:09:46,320 --> 00:09:51,200 Speaker 7: billion day dollars four chillers for liquid cooling, four data centers. 182 00:09:51,679 --> 00:09:56,480 Speaker 7: So you know, that's that's an incredible, incredible number just 183 00:09:56,480 --> 00:09:59,680 Speaker 7: just for cooling. But one of our competitors who is 184 00:09:59,760 --> 00:10:02,679 Speaker 7: not yet very very strong in power but not in cooling, 185 00:10:02,720 --> 00:10:06,040 Speaker 7: which Eaton Corporation, and they just spent over nine billion 186 00:10:06,080 --> 00:10:10,520 Speaker 7: dollars on Boyd. So it's a company that's very close 187 00:10:10,559 --> 00:10:12,079 Speaker 7: to my house in Florida. 188 00:10:12,679 --> 00:10:13,680 Speaker 6: There, my davor. 189 00:10:13,920 --> 00:10:17,120 Speaker 7: But that's a huge nine billion dollar acquisition for for 190 00:10:17,280 --> 00:10:18,959 Speaker 7: mainly for a liquid cooling company. 191 00:10:19,240 --> 00:10:20,959 Speaker 2: So it's a lot of electrical companies kind of building 192 00:10:20,960 --> 00:10:24,440 Speaker 2: this full value stack of Yeah, okay, so I got 193 00:10:24,520 --> 00:10:28,800 Speaker 2: a bit more vertical integration. So using that to kind 194 00:10:28,800 --> 00:10:30,840 Speaker 2: of spin off into having a bit more of a 195 00:10:30,840 --> 00:10:33,800 Speaker 2: conversation about electrical equipment and the innovations that are happening there. 196 00:10:34,280 --> 00:10:36,480 Speaker 3: And Jill, can you talk us through. 197 00:10:37,080 --> 00:10:40,599 Speaker 2: You've explained what FEAR does, can you explain how I 198 00:10:40,640 --> 00:10:43,520 Speaker 2: mean it didn't start out as a data center electrical company, 199 00:10:43,520 --> 00:10:45,199 Speaker 2: and that's that's not just what you do now, but 200 00:10:45,440 --> 00:10:48,040 Speaker 2: it's gotten from where it's founded to where it is now. 201 00:10:48,320 --> 00:10:53,040 Speaker 8: Yeah, So Viers actually founded to solve all the power 202 00:10:53,120 --> 00:10:55,920 Speaker 8: density problems that exists actually at the high voltage level 203 00:10:55,960 --> 00:11:01,320 Speaker 8: and transmission and distribution and really using evaporative liquid nitrogen 204 00:11:01,400 --> 00:11:05,640 Speaker 8: cooling technology combined with superconducting which has been around for 205 00:11:05,920 --> 00:11:09,199 Speaker 8: twenty five twenty five plus years and lots of different 206 00:11:09,240 --> 00:11:12,200 Speaker 8: grid connected projects. But really what we've found is an 207 00:11:12,200 --> 00:11:15,559 Speaker 8: opportunity to take that idea and commercialize low voltage and 208 00:11:15,640 --> 00:11:18,920 Speaker 8: medium voltage products to help solve some of the behind 209 00:11:18,960 --> 00:11:22,120 Speaker 8: the meter challenges that we see in data centers today. 210 00:11:23,360 --> 00:11:27,600 Speaker 8: And for us, that really means on the medium voltage side, 211 00:11:27,600 --> 00:11:31,440 Speaker 8: helping to move power around the data center campus. So 212 00:11:31,480 --> 00:11:34,360 Speaker 8: this is from the substation into the buildings and really 213 00:11:34,400 --> 00:11:37,080 Speaker 8: this is otherwise done through copper cables and a huge 214 00:11:37,120 --> 00:11:41,680 Speaker 8: amount of cement and labor. Our products actually reduced that 215 00:11:41,880 --> 00:11:45,320 Speaker 8: labor and cement cost by about forty percent inside the building. 216 00:11:45,320 --> 00:11:49,960 Speaker 8: We're actually able to offer a different alternative to traditional 217 00:11:50,000 --> 00:11:53,800 Speaker 8: bus bar copper bus bar solutions, and we are able 218 00:11:53,800 --> 00:11:58,120 Speaker 8: to move about ten x the power with the same space, 219 00:11:58,320 --> 00:12:02,640 Speaker 8: so that I'll offer more weight savings, space savings and 220 00:12:02,760 --> 00:12:07,680 Speaker 8: performance savings as GPUs require increased proximity to one another 221 00:12:07,760 --> 00:12:09,599 Speaker 8: to really maximize that performance. 222 00:12:10,280 --> 00:12:12,760 Speaker 2: Yeah, I mean, I've heard some really funny statistics about 223 00:12:12,800 --> 00:12:14,840 Speaker 2: like if you want to power like a one megawatt 224 00:12:15,160 --> 00:12:18,160 Speaker 2: GPU rack with a copper cable and something like that 225 00:12:18,200 --> 00:12:20,480 Speaker 2: thick and then you have to just reinforce it so 226 00:12:20,600 --> 00:12:23,920 Speaker 2: much so there's a lot of engineering challenges associated with 227 00:12:23,960 --> 00:12:26,239 Speaker 2: this and the heat. 228 00:12:26,559 --> 00:12:26,959 Speaker 3: And the heat. 229 00:12:27,040 --> 00:12:31,320 Speaker 2: Yeah, I heard walk us through. You said a lot 230 00:12:31,360 --> 00:12:34,400 Speaker 2: of words when you introduced your products. Can you explain 231 00:12:34,440 --> 00:12:37,120 Speaker 2: what exactly it is that it does and again how 232 00:12:37,160 --> 00:12:39,160 Speaker 2: it relates to this data center conversation. 233 00:12:39,440 --> 00:12:44,200 Speaker 9: So our technology basically can take any AC or DC 234 00:12:44,400 --> 00:12:48,480 Speaker 9: source of power, multiple ones at the same time, and 235 00:12:48,679 --> 00:12:54,400 Speaker 9: aggregate them using a solid state transformer. It's heavily patented 236 00:12:54,440 --> 00:12:58,800 Speaker 9: by us. It's nobody else has this technology. And what 237 00:12:58,840 --> 00:13:01,160 Speaker 9: we can do is we can add for example, solar 238 00:13:02,000 --> 00:13:06,600 Speaker 9: or fuel cell or batteries or natural gas gen sets 239 00:13:07,880 --> 00:13:10,280 Speaker 9: or the grid, and we can add them in any 240 00:13:10,400 --> 00:13:13,560 Speaker 9: ratio we want down to the millionth of a second, 241 00:13:14,120 --> 00:13:17,240 Speaker 9: and we balance the power in order to extract the 242 00:13:17,240 --> 00:13:21,040 Speaker 9: best transients from the best transient source. Otherwise, we lower 243 00:13:21,080 --> 00:13:24,640 Speaker 9: the cost of the energy that we transfer and we 244 00:13:24,679 --> 00:13:27,280 Speaker 9: look for the highest efficiency. So we have our own 245 00:13:27,400 --> 00:13:31,920 Speaker 9: energy management system to optimize this, and simultaneously we can 246 00:13:31,960 --> 00:13:36,920 Speaker 9: deliver multiple sources of power to multiple megawatts of Nvidia 247 00:13:37,360 --> 00:13:42,040 Speaker 9: eight hundred volt racks for the GENAI servers. That's our technology. 248 00:13:42,040 --> 00:13:46,479 Speaker 9: So it's consolidating eleven to seventeen different pieces of equipment 249 00:13:46,760 --> 00:13:48,120 Speaker 9: into one piece of equipment. 250 00:13:48,400 --> 00:13:50,839 Speaker 2: Adam I right in describing its like it's kind of 251 00:13:50,880 --> 00:13:53,360 Speaker 2: the entry point from the grid to the data center, 252 00:13:53,400 --> 00:13:54,760 Speaker 2: and now everything's just being put in. 253 00:13:55,080 --> 00:13:58,320 Speaker 9: Yeah, for entry point from the grid to the data center, 254 00:13:58,400 --> 00:14:01,400 Speaker 9: but also for all these other ener sources. The way 255 00:14:01,440 --> 00:14:04,680 Speaker 9: it's done today, in a legacy system, you have a 256 00:14:04,800 --> 00:14:10,400 Speaker 9: number of transformers, rectifiers, inverters, maximum PowerPoint tracking fuel cell 257 00:14:10,960 --> 00:14:13,920 Speaker 9: DC to DC converters. We eliminate all of that and 258 00:14:13,960 --> 00:14:17,320 Speaker 9: we take raw sources of energy into our multi port 259 00:14:17,400 --> 00:14:21,360 Speaker 9: solid state transformer, each port devoted to one source and 260 00:14:21,400 --> 00:14:25,200 Speaker 9: one load, and then we provide consolidated output power we 261 00:14:25,360 --> 00:14:27,600 Speaker 9: clean up the harmonics, we clean up the surges. 262 00:14:28,360 --> 00:14:28,600 Speaker 3: Yeah. 263 00:14:28,680 --> 00:14:31,880 Speaker 9: I mean it took us six one hundred thousand engineering 264 00:14:31,920 --> 00:14:36,040 Speaker 9: hours to deliver the first first pilot, so it wasn't 265 00:14:36,400 --> 00:14:37,360 Speaker 9: an easy thing. 266 00:14:37,200 --> 00:14:37,720 Speaker 5: To go through. 267 00:14:38,040 --> 00:14:38,240 Speaker 3: Yeah. 268 00:14:38,320 --> 00:14:40,320 Speaker 2: Yeah, And I mean you didn't start as a data 269 00:14:40,360 --> 00:14:43,080 Speaker 2: center company either. This was something that was built for 270 00:14:43,200 --> 00:14:44,560 Speaker 2: general great applications, right. 271 00:14:44,800 --> 00:14:49,000 Speaker 9: We actually started for fleet charging originally because fleets couldn't 272 00:14:49,000 --> 00:14:51,680 Speaker 9: get enough power. And when we'd put our first pilot 273 00:14:51,720 --> 00:14:53,760 Speaker 9: in the ground to show what we could do, we 274 00:14:53,840 --> 00:14:56,360 Speaker 9: got a call from one of these GPU companies and 275 00:14:56,440 --> 00:14:59,120 Speaker 9: say is this real? We go absolutely, and then they 276 00:14:59,160 --> 00:15:02,080 Speaker 9: started giving us becks and helping us out, and that's 277 00:15:02,120 --> 00:15:06,120 Speaker 9: how we got to where we are. We're looking at 278 00:15:05,360 --> 00:15:08,640 Speaker 9: a massive pipeline over the next two three years as 279 00:15:08,680 --> 00:15:09,880 Speaker 9: we start delivering. 280 00:15:09,520 --> 00:15:10,280 Speaker 6: Product this year. 281 00:15:11,960 --> 00:15:13,800 Speaker 2: Up actually kind of interested because like, neither of the 282 00:15:13,880 --> 00:15:16,160 Speaker 2: companies are founded with this in mind. But how much 283 00:15:16,920 --> 00:15:19,040 Speaker 2: effort would you say you spend now and your kind 284 00:15:19,080 --> 00:15:22,400 Speaker 2: of utility customers and your grid projects versus like what's 285 00:15:22,440 --> 00:15:24,160 Speaker 2: going into data centers? 286 00:15:24,280 --> 00:15:25,400 Speaker 3: Can you share anything about that? 287 00:15:25,560 --> 00:15:25,800 Speaker 5: Sure? 288 00:15:25,840 --> 00:15:29,560 Speaker 9: I think I think utilities are a conservative market. It 289 00:15:29,600 --> 00:15:34,560 Speaker 9: takes time to break into utilities it's it's a massive market, 290 00:15:34,560 --> 00:15:37,400 Speaker 9: but it's going to take time. The commercial and industrial 291 00:15:37,480 --> 00:15:40,840 Speaker 9: market is where we started deliberately in order to overcome 292 00:15:40,880 --> 00:15:44,840 Speaker 9: those issues, Okay, and then the data centers came along, 293 00:15:44,880 --> 00:15:47,360 Speaker 9: which was quite frankly, just pure luck. We were there 294 00:15:47,360 --> 00:15:50,120 Speaker 9: at the right time and they need power and they 295 00:15:50,160 --> 00:15:50,720 Speaker 9: need it now. 296 00:15:52,280 --> 00:15:52,520 Speaker 5: Yeah. 297 00:15:52,520 --> 00:15:56,720 Speaker 8: Similarly, we you know, we have lots of utility customers 298 00:15:56,720 --> 00:15:59,520 Speaker 8: that we continue to have conversations with, but really we 299 00:15:59,600 --> 00:16:03,440 Speaker 8: have the opportunity to show how superconducting works and how 300 00:16:03,440 --> 00:16:07,000 Speaker 8: it solves a lot of the medium voltage challenges that 301 00:16:07,520 --> 00:16:09,280 Speaker 8: many of the utilities face. But we get to do 302 00:16:09,360 --> 00:16:12,760 Speaker 8: that sort of with the data center customers, build that 303 00:16:12,840 --> 00:16:14,680 Speaker 8: traction and then scale it to utilities. 304 00:16:15,600 --> 00:16:16,200 Speaker 3: Okay, great. 305 00:16:17,040 --> 00:16:19,320 Speaker 2: A question like I think everyone always asks a lot 306 00:16:19,320 --> 00:16:21,760 Speaker 2: about data centers because it had a huge influence historically. 307 00:16:21,880 --> 00:16:23,920 Speaker 2: Was that people thought the power amount was going to 308 00:16:23,920 --> 00:16:25,760 Speaker 2: blow up since like two thousand and eight, and it 309 00:16:25,800 --> 00:16:28,240 Speaker 2: never really did because things always got more and more efficient. 310 00:16:28,400 --> 00:16:32,800 Speaker 2: So on the electrical equipment side, like innovations are focused 311 00:16:32,840 --> 00:16:34,120 Speaker 2: on getting the power. 312 00:16:34,000 --> 00:16:36,120 Speaker 3: Into a smaller form factor, making it cleaner. 313 00:16:36,440 --> 00:16:40,280 Speaker 2: Are they impacting efficiency on a meaningful, meaningful level? 314 00:16:42,240 --> 00:16:46,400 Speaker 9: I think Yeah, Well, first of all, it's a radical 315 00:16:46,520 --> 00:16:49,440 Speaker 9: change when you go from a six kilowatt rack or 316 00:16:49,440 --> 00:16:52,880 Speaker 9: an eight kilo WoT rack to a metal walt class rack, right, 317 00:16:53,240 --> 00:16:56,120 Speaker 9: and that's where we're headed by twenty twenty seven, early 318 00:16:56,200 --> 00:16:59,520 Speaker 9: to middle of twenty seven. That's one hundred x increase. 319 00:17:00,200 --> 00:17:03,560 Speaker 9: So you cannot possibly have big fit, long runs of 320 00:17:03,600 --> 00:17:06,520 Speaker 9: copper that are going to carry one hundred times more current. 321 00:17:06,560 --> 00:17:10,040 Speaker 9: You've got to rethink it out. In addition, if data 322 00:17:10,119 --> 00:17:13,120 Speaker 9: centers have a real problem in getting. 323 00:17:12,800 --> 00:17:14,960 Speaker 6: Power to the facility, it. 324 00:17:14,960 --> 00:17:18,680 Speaker 9: Isn't about efficiency as much as it is about stranded power. 325 00:17:19,160 --> 00:17:23,840 Speaker 9: Power that's designed in legacy architectures that leaves ten, twenty percent, 326 00:17:23,920 --> 00:17:27,000 Speaker 9: thirty percent of power that can't be delivered to the chip. 327 00:17:27,600 --> 00:17:31,640 Speaker 9: So our focus is on eliminating that stranded power. Why well, 328 00:17:31,720 --> 00:17:35,080 Speaker 9: let's take one hundred megawat data center. Let's say it's 329 00:17:35,200 --> 00:17:40,040 Speaker 9: just fifteen percent. That's fifteen megawots. Fifteen megawats will generate 330 00:17:40,080 --> 00:17:43,440 Speaker 9: one hundred and fifty million dollars of AI revenue a year. 331 00:17:44,600 --> 00:17:48,280 Speaker 9: It's not efficiency, it's about eliminating stranded power under all 332 00:17:48,320 --> 00:17:51,639 Speaker 9: sorts of failure modes and the right architecture to blend 333 00:17:51,640 --> 00:17:55,320 Speaker 9: those sources and deliver it, and that's what we're absolutely 334 00:17:55,400 --> 00:17:59,520 Speaker 9: laser focused on. But efficiency, sorry if I couldn't, also matters, 335 00:17:59,560 --> 00:18:02,879 Speaker 9: because when you aggregate all this equipment into a small footprint, 336 00:18:03,160 --> 00:18:06,040 Speaker 9: you lose less copper, you have far less losses, so 337 00:18:06,119 --> 00:18:09,280 Speaker 9: you gain a few percentage points, but the stranded power 338 00:18:09,280 --> 00:18:10,359 Speaker 9: is where the real money is. 339 00:18:11,200 --> 00:18:13,480 Speaker 8: I would just also add I agree with everything you've said. 340 00:18:13,480 --> 00:18:15,439 Speaker 8: I would also just add, well, we hear a lot 341 00:18:15,440 --> 00:18:17,919 Speaker 8: from our customers about speed to power. So this is 342 00:18:17,960 --> 00:18:22,720 Speaker 8: about those greenfield projects that Stephen was referring to, how quickly. 343 00:18:22,440 --> 00:18:23,480 Speaker 5: Can they get built? 344 00:18:23,600 --> 00:18:26,040 Speaker 8: And so for us, this is really about how helping 345 00:18:26,080 --> 00:18:28,800 Speaker 8: to solve that challenge by eliminating some of the labor 346 00:18:28,840 --> 00:18:31,840 Speaker 8: costs and some of the cement costs and some of 347 00:18:31,880 --> 00:18:37,399 Speaker 8: the install time really through a more simplified system. And 348 00:18:38,000 --> 00:18:41,080 Speaker 8: in our case, it's just digging less trenches. 349 00:18:41,600 --> 00:18:43,639 Speaker 3: Yeah. Sometimes innovation really is that simple. 350 00:18:43,800 --> 00:18:46,280 Speaker 8: Yeah. 351 00:18:46,320 --> 00:18:49,439 Speaker 2: Cool, So we I mean I didn't really mention this 352 00:18:49,480 --> 00:18:52,240 Speaker 2: in the intro, but it's kind of become apparent threat 353 00:18:52,280 --> 00:18:55,400 Speaker 2: the conversation that so much of what's going on at 354 00:18:55,400 --> 00:18:58,399 Speaker 2: all of your companies is really defined by what's happening 355 00:18:58,400 --> 00:19:02,840 Speaker 2: in another company and the fact that they fundamentally redesigned 356 00:19:03,119 --> 00:19:05,800 Speaker 2: what's going into data centers and what data centers can 357 00:19:05,800 --> 00:19:08,960 Speaker 2: even be used for. But they kind of have a 358 00:19:08,960 --> 00:19:11,080 Speaker 2: funny influence because then a lot of the sector is 359 00:19:11,160 --> 00:19:13,440 Speaker 2: kind of designing around what they use. 360 00:19:13,920 --> 00:19:14,960 Speaker 3: And an interesting story. 361 00:19:14,960 --> 00:19:17,800 Speaker 2: Maybe some people in the audience saw Jensen Lang, the 362 00:19:17,880 --> 00:19:20,560 Speaker 2: CEO of Nvidia, was at a conference recently and kind 363 00:19:20,560 --> 00:19:22,919 Speaker 2: of spoke about the fact that the new GPU, the 364 00:19:22,960 --> 00:19:25,639 Speaker 2: Vera Ruben, is going to be running so hot that 365 00:19:25,720 --> 00:19:27,879 Speaker 2: you can use entirely liquid cooling and actually won't need 366 00:19:27,920 --> 00:19:30,160 Speaker 2: to use chillers at all, so that one point nine 367 00:19:30,160 --> 00:19:35,720 Speaker 2: billion order. Stephen, you were talking about what like, do 368 00:19:35,760 --> 00:19:39,320 Speaker 2: you think that this is realistic? Do you think what's 369 00:19:39,359 --> 00:19:42,400 Speaker 2: the kind of relationship between how you develop your equipment 370 00:19:42,880 --> 00:19:45,080 Speaker 2: on the cooling side and what a Vidia is saying 371 00:19:45,320 --> 00:19:45,800 Speaker 2: on that front. 372 00:19:45,840 --> 00:19:48,919 Speaker 7: Yeah, So I spent I spent all morning with Nvidia 373 00:19:49,160 --> 00:19:52,879 Speaker 7: and today, Yes, and we actually spent an hour on 374 00:19:52,920 --> 00:19:58,000 Speaker 7: this on this specific subject because we were preparing kind 375 00:19:58,040 --> 00:20:03,240 Speaker 7: of a I would call it a position paper on 376 00:20:03,400 --> 00:20:07,600 Speaker 7: exactly what Gentsen meant, because he really didn't mean that 377 00:20:07,720 --> 00:20:10,640 Speaker 7: anyone was going to run their data centers without chillers. 378 00:20:11,359 --> 00:20:14,000 Speaker 7: What he meant was it's possible at forty degrees c 379 00:20:14,720 --> 00:20:19,199 Speaker 7: in certain climates that they could possibly run part of 380 00:20:19,200 --> 00:20:24,480 Speaker 7: the time, you know, without without mechanical chillers. But that's 381 00:20:24,600 --> 00:20:28,879 Speaker 7: really not how it came across. And Nvidia had a 382 00:20:28,880 --> 00:20:32,000 Speaker 7: lot of negative reactions from a lot of their cooling suppliers, 383 00:20:32,560 --> 00:20:35,919 Speaker 7: but we were not, as I guess, aggressive, So what 384 00:20:36,040 --> 00:20:38,359 Speaker 7: happened was things calmed down and they decided not to 385 00:20:38,400 --> 00:20:42,359 Speaker 7: issue this this paper. But the reality is nobody is 386 00:20:42,359 --> 00:20:45,959 Speaker 7: going to run their data centers at forty five degrees 387 00:20:45,960 --> 00:20:48,800 Speaker 7: celsius because you don't know how long the chips are 388 00:20:48,800 --> 00:20:52,200 Speaker 7: going to last, and you don't know if they're going 389 00:20:52,240 --> 00:20:54,640 Speaker 7: to overheat and you're going to lose control and not 390 00:20:54,720 --> 00:20:58,960 Speaker 7: have you know, functionality. So so most of our customers, 391 00:20:58,960 --> 00:21:02,760 Speaker 7: and this happened happened with Blackwell too, the same exact thing. 392 00:21:03,560 --> 00:21:05,800 Speaker 7: They were saying, you can run them at forty So 393 00:21:05,840 --> 00:21:08,480 Speaker 7: we started testing them and they weren't really operating right, 394 00:21:08,480 --> 00:21:11,119 Speaker 7: they kept them so we had a lower the minimum 395 00:21:11,160 --> 00:21:14,760 Speaker 7: to thirty or the maximum to thirty. 396 00:21:15,080 --> 00:21:17,879 Speaker 6: So you know, it's not. 397 00:21:17,840 --> 00:21:21,080 Speaker 7: That that all of a sudden there's going to be 398 00:21:21,080 --> 00:21:22,880 Speaker 7: no need for chillers and data centers. 399 00:21:22,920 --> 00:21:25,439 Speaker 6: I mean, it would be great for the environment and 400 00:21:25,480 --> 00:21:25,760 Speaker 6: for the. 401 00:21:25,760 --> 00:21:29,840 Speaker 7: Efficiency if it was possible, But things are actually trending 402 00:21:29,880 --> 00:21:31,760 Speaker 7: in the other direction. We're going to need more and 403 00:21:31,800 --> 00:21:36,399 Speaker 7: more chillers. But the beautiful thing about liquid cooling is 404 00:21:36,880 --> 00:21:39,639 Speaker 7: you can run it at a higher temperature than air cooling, 405 00:21:39,640 --> 00:21:42,879 Speaker 7: and the chillers are our higher temperature chiller than the 406 00:21:42,920 --> 00:21:45,520 Speaker 7: air cool chiller, so there is more efficiency. Used to 407 00:21:45,560 --> 00:21:50,359 Speaker 7: be had hoppers, the one before Blackwell, A lot of 408 00:21:50,400 --> 00:21:53,480 Speaker 7: those were run on pre cooling a lot of the 409 00:21:53,520 --> 00:21:56,200 Speaker 7: time and only in rare cases, and Blackwell a little 410 00:21:56,240 --> 00:21:58,520 Speaker 7: bit more. And if vera Reuben, it's going to be more. 411 00:22:00,080 --> 00:22:02,560 Speaker 7: I don't think he's going to retract that statement. But 412 00:22:03,480 --> 00:22:06,480 Speaker 7: it wasn't it wasn't exactly accurate. 413 00:22:06,720 --> 00:22:09,080 Speaker 2: Yeah, I mean, Angela, can you I mean you read 414 00:22:09,119 --> 00:22:10,440 Speaker 2: a lot of stuff in the news all the time 415 00:22:10,560 --> 00:22:12,840 Speaker 2: about everything that happens in the media, but I mean, 416 00:22:12,880 --> 00:22:16,400 Speaker 2: can you, from just a technical perspective, explain why people 417 00:22:16,440 --> 00:22:19,520 Speaker 2: aren't going to run their chips at forty degrees? I mean, 418 00:22:19,560 --> 00:22:22,120 Speaker 2: there's nothing sounds wrong on no. 419 00:22:22,359 --> 00:22:24,320 Speaker 4: Actually, if you can run it at forty five s, 420 00:22:24,480 --> 00:22:27,400 Speaker 4: which is what he said, that's one hundred and thirteen fahrenheit. 421 00:22:28,640 --> 00:22:32,720 Speaker 4: That'd be fantastic because, like Stephen said, you could reduce 422 00:22:32,760 --> 00:22:36,880 Speaker 4: the amount of reduce the need for the checks, right, 423 00:22:36,960 --> 00:22:39,439 Speaker 4: you could use a dry cooler and you could just 424 00:22:40,200 --> 00:22:43,320 Speaker 4: cool with the air outside. The problem with it is, 425 00:22:43,520 --> 00:22:46,439 Speaker 4: as Stephen said, not in like in Arizona, you're not 426 00:22:46,440 --> 00:22:47,880 Speaker 4: going to be able to do that one hundred percent 427 00:22:47,920 --> 00:22:51,080 Speaker 4: of the time. In most places around the world you 428 00:22:51,080 --> 00:22:52,720 Speaker 4: will not be able to do it one hundred percent 429 00:22:52,760 --> 00:22:55,920 Speaker 4: of the time. So don't We're not going to get 430 00:22:55,960 --> 00:23:01,760 Speaker 4: away from having that for sure, But what will And 431 00:23:01,840 --> 00:23:06,320 Speaker 4: I like to think that Jensen was painting this, you know, 432 00:23:06,400 --> 00:23:09,600 Speaker 4: this idea of what could be, And I think that's 433 00:23:09,680 --> 00:23:13,879 Speaker 4: useful because when you put out a vision of what 434 00:23:14,000 --> 00:23:17,800 Speaker 4: could be and you start making small steps towards it, then. 435 00:23:17,560 --> 00:23:18,880 Speaker 5: You know it becomes a reality. 436 00:23:19,320 --> 00:23:21,479 Speaker 4: And so in order for something like that to happen, 437 00:23:22,560 --> 00:23:25,679 Speaker 4: as Stephen mentioned, you'd have to verify all the components 438 00:23:25,720 --> 00:23:27,600 Speaker 4: that are in the system to that. 439 00:23:27,560 --> 00:23:29,640 Speaker 5: Level for reliability. 440 00:23:30,920 --> 00:23:34,840 Speaker 4: You would also likely have to have a dedicated loop 441 00:23:35,320 --> 00:23:42,080 Speaker 4: for that particular it, so you'd likely have I'm not 442 00:23:42,119 --> 00:23:45,280 Speaker 4: an infrastructure person, maybe Stephen could do this, but you'd 443 00:23:45,400 --> 00:23:48,080 Speaker 4: likely have a hot loop and a cold loop because 444 00:23:48,280 --> 00:23:51,639 Speaker 4: right now the facility water loop usually has multiple systems 445 00:23:51,680 --> 00:23:54,520 Speaker 4: on it, So there need to be some architecture changes 446 00:23:54,560 --> 00:23:58,280 Speaker 4: and some validation and warranties that would happen in order 447 00:23:58,560 --> 00:24:01,959 Speaker 4: for that to be true. But if it could happen, 448 00:24:02,520 --> 00:24:04,480 Speaker 4: it would save a lot of water and a lot 449 00:24:04,480 --> 00:24:05,000 Speaker 4: of energy. 450 00:24:05,280 --> 00:24:07,879 Speaker 7: Yeah, and if you're running AI training, you know you 451 00:24:07,880 --> 00:24:10,800 Speaker 7: could you could stand some disruptions because you'll just pick 452 00:24:10,880 --> 00:24:12,679 Speaker 7: up from where you're left off if you'll read. But 453 00:24:12,720 --> 00:24:17,480 Speaker 7: if AI an AI is becoming an increasingly embedded and 454 00:24:17,560 --> 00:24:21,080 Speaker 7: critical business processes, data center operators are not going to 455 00:24:21,160 --> 00:24:22,920 Speaker 7: take the chance of going down. 456 00:24:23,000 --> 00:24:23,879 Speaker 6: A lot of data. 457 00:24:23,600 --> 00:24:26,879 Speaker 7: Centers are being considered in the UK critical facilities, in 458 00:24:26,920 --> 00:24:30,199 Speaker 7: the US critical facilities, and there's certain rules now that 459 00:24:30,280 --> 00:24:32,760 Speaker 7: you can't like take these chances you need to. 460 00:24:33,200 --> 00:24:35,760 Speaker 2: It was the Chicago Metal Exchange, was it that went 461 00:24:35,920 --> 00:24:38,920 Speaker 2: down because it was exolent distribution unit broke inside the 462 00:24:39,000 --> 00:24:41,879 Speaker 2: data center and that brought it offline for a while. 463 00:24:43,080 --> 00:24:44,760 Speaker 2: I mean, I also think it's interesting because I think 464 00:24:44,840 --> 00:24:47,280 Speaker 2: like the hotter you run the chip, the more degrades, 465 00:24:47,320 --> 00:24:49,359 Speaker 2: and then on the previous pannel where you're discussing the 466 00:24:49,400 --> 00:24:51,640 Speaker 2: role of depreciation schedules. 467 00:24:51,160 --> 00:24:52,680 Speaker 3: Kind of long term chip performance. 468 00:24:53,160 --> 00:24:56,159 Speaker 2: And yeah, so even even like the details of how 469 00:24:56,200 --> 00:24:59,760 Speaker 2: the CDU runs and the temperature it runs us influencing 470 00:25:00,240 --> 00:25:01,280 Speaker 2: markets these days. 471 00:25:01,720 --> 00:25:03,600 Speaker 4: Well, I do think that if you're going to run it, 472 00:25:03,600 --> 00:25:05,560 Speaker 4: if you're going to overclock it when you run it 473 00:25:05,680 --> 00:25:08,199 Speaker 4: hotter like that, immersion is a better solution. 474 00:25:08,480 --> 00:25:10,200 Speaker 3: Immersions better. So okay, interesting. 475 00:25:11,280 --> 00:25:13,960 Speaker 2: So, I mean there was that conversation from a video 476 00:25:14,000 --> 00:25:16,800 Speaker 2: about the chillers. I think that drew a lot of headlines. 477 00:25:17,359 --> 00:25:19,919 Speaker 2: I think on the electrical side of things. I mean, 478 00:25:20,320 --> 00:25:22,880 Speaker 2: you mentioned this a couple of times that you want 479 00:25:22,920 --> 00:25:27,080 Speaker 2: to move to an electrical architecture that really reduces the 480 00:25:27,080 --> 00:25:29,879 Speaker 2: amount of lost power that kind of enters the data center. 481 00:25:30,240 --> 00:25:33,919 Speaker 2: And I think the phrase that defines this movement is 482 00:25:33,920 --> 00:25:37,920 Speaker 2: the kind of eight hundred volt direct current And could 483 00:25:37,920 --> 00:25:40,359 Speaker 2: you give a primer for the audience on like what 484 00:25:40,520 --> 00:25:43,960 Speaker 2: that means in the context of a data center, And yeah, 485 00:25:44,040 --> 00:25:45,720 Speaker 2: it just give a primer on that. 486 00:25:45,840 --> 00:25:47,879 Speaker 9: So I think I think one of the reasons to 487 00:25:48,400 --> 00:25:51,600 Speaker 9: raise the voltage from from two forty or two twenty 488 00:25:51,720 --> 00:25:56,399 Speaker 9: or two seventy seven single phase line to neutral almost 489 00:25:56,440 --> 00:25:59,560 Speaker 9: take it up three times to DC is you're going 490 00:25:59,600 --> 00:26:02,679 Speaker 9: to use the amount of current, and by reducing the 491 00:26:02,680 --> 00:26:05,480 Speaker 9: amount of current, your bus bars, your copper bus bars 492 00:26:05,520 --> 00:26:09,640 Speaker 9: can carry more power. The voltage generally can be free. 493 00:26:10,040 --> 00:26:12,880 Speaker 9: But the issue with a higher voltage or the safety 494 00:26:12,920 --> 00:26:16,359 Speaker 9: aspects where people are worried about arc flash what happens 495 00:26:16,359 --> 00:26:20,320 Speaker 9: if there are shorts and this this flash happens that 496 00:26:20,400 --> 00:26:24,200 Speaker 9: can melt a human body in a fraction right of time. 497 00:26:24,520 --> 00:26:28,480 Speaker 9: So it's it's mitigating those concerns. That's the big that's 498 00:26:28,520 --> 00:26:31,040 Speaker 9: that's one of the big topics. But eight hundred volts 499 00:26:31,119 --> 00:26:35,720 Speaker 9: DC also eliminates a number of power conversion stages, getting 500 00:26:35,760 --> 00:26:38,760 Speaker 9: to where you can get higher efficiency, less losses, and 501 00:26:38,880 --> 00:26:41,880 Speaker 9: more delivery of power. And it allows you to take 502 00:26:42,040 --> 00:26:46,600 Speaker 9: take those gen Ai servers straight in and mitigate those 503 00:26:46,640 --> 00:26:51,359 Speaker 9: pulses that that the Genai servers are requiring. Also with 504 00:26:51,440 --> 00:26:54,880 Speaker 9: a similar battery or supercapacitor. And that's where this eight 505 00:26:54,920 --> 00:26:59,960 Speaker 9: hundred volt DC architecture came from. High lower currents, higher efficiency, 506 00:27:00,920 --> 00:27:03,640 Speaker 9: less uh less number of convergent stages. 507 00:27:04,040 --> 00:27:08,240 Speaker 7: Yeah, so there's there's not one eight hundred volt DC architectures. 508 00:27:08,560 --> 00:27:10,600 Speaker 2: So I mean This is what I's gonna ask you, Stephen, 509 00:27:10,640 --> 00:27:13,959 Speaker 2: can't splain the kind of the fork that's happened between 510 00:27:14,040 --> 00:27:16,240 Speaker 2: what what some companies want and what other companies want. 511 00:27:16,440 --> 00:27:21,480 Speaker 7: Yeah, it's it's it's interesting because we you know, the 512 00:27:21,560 --> 00:27:24,560 Speaker 7: industry landed on eight inter volts DC because they wanted 513 00:27:24,880 --> 00:27:27,639 Speaker 7: you know, there was there was connectors, and there was breakers, 514 00:27:28,080 --> 00:27:31,560 Speaker 7: and there's s S T S i'll say transformers from 515 00:27:31,600 --> 00:27:34,520 Speaker 7: the car industry, the automobile industry for I voltage charging, 516 00:27:35,240 --> 00:27:38,159 Speaker 7: so all the safety things were in place, so they 517 00:27:38,200 --> 00:27:41,320 Speaker 7: decided to do that. The first iteration of eight hundred 518 00:27:41,359 --> 00:27:45,280 Speaker 7: volt DC is what the industry is calling sidecar. So, 519 00:27:45,480 --> 00:27:49,320 Speaker 7: because you want to run your AI, the accelerated compute 520 00:27:49,440 --> 00:27:53,800 Speaker 7: as one monolithic GPU, the GPUs have to be as 521 00:27:53,880 --> 00:27:57,600 Speaker 7: close together as possible. The way to get the GPUs 522 00:27:57,720 --> 00:28:00,680 Speaker 7: or more GPUs closer is to take the power supplies 523 00:28:00,840 --> 00:28:03,680 Speaker 7: out of the server. So we're taking the power supplies 524 00:28:03,760 --> 00:28:05,800 Speaker 7: out of the server and we're putting it into this 525 00:28:05,960 --> 00:28:09,280 Speaker 7: thing the industry is calling a sidecar, which is sitting 526 00:28:09,359 --> 00:28:12,440 Speaker 7: next to the rack. And the sidecar has the power supplies, 527 00:28:12,520 --> 00:28:17,359 Speaker 7: has the capacitors, has you know, the battery backup units, 528 00:28:17,720 --> 00:28:21,160 Speaker 7: and the breakers and fuses in it, So it's it's 529 00:28:21,520 --> 00:28:24,919 Speaker 7: you're disaggregating all of that, You're you're you're freeing up 530 00:28:24,960 --> 00:28:27,240 Speaker 7: the ability, and you're taking eight hundred volts from that 531 00:28:27,359 --> 00:28:30,040 Speaker 7: side car to the. 532 00:28:31,560 --> 00:28:32,400 Speaker 3: GPU rack. 533 00:28:32,520 --> 00:28:36,720 Speaker 7: And in Vidia terms, it's the NBL seventy twos, they're 534 00:28:36,760 --> 00:28:39,200 Speaker 7: going to be NBL one forty four's, they're eventually going 535 00:28:39,280 --> 00:28:41,200 Speaker 7: to be NBL five to seventy six, which is the 536 00:28:41,280 --> 00:28:43,240 Speaker 7: number of GPUs that they're going to try. 537 00:28:43,160 --> 00:28:48,360 Speaker 6: To stuff into these boxes. So basically they're just getting stuffing. 538 00:28:48,080 --> 00:28:51,440 Speaker 7: More and more in. But AMD has their own spec 539 00:28:51,560 --> 00:28:54,200 Speaker 7: and their own version of this which has different capacitors 540 00:28:54,280 --> 00:28:57,720 Speaker 7: and then different ride through. The Nvidia chips actually have 541 00:28:57,880 --> 00:28:59,800 Speaker 7: a higher degree of disruption back. 542 00:28:59,720 --> 00:29:02,440 Speaker 6: Onto the utility than some of the others. The AMD. 543 00:29:04,160 --> 00:29:08,320 Speaker 7: Meta Meta has a plus and minus four intervol DC 544 00:29:08,520 --> 00:29:11,440 Speaker 7: that they call mont Diablo, and then Google has another one. 545 00:29:11,600 --> 00:29:16,160 Speaker 6: So we have to develop four of those different. 546 00:29:15,880 --> 00:29:20,240 Speaker 7: Sidecars, one for each of these different applications. But overwhelmingly, 547 00:29:20,800 --> 00:29:25,880 Speaker 7: you know, the market is in Nvidia overwhelming, Okay. 548 00:29:26,440 --> 00:29:29,480 Speaker 2: So there's a lot of i mean, from across the 549 00:29:29,520 --> 00:29:32,080 Speaker 2: whole spectrum of equipment that's going to the data centers. 550 00:29:32,120 --> 00:29:36,800 Speaker 2: There's a lot of innovation and ship changes. To what 551 00:29:36,960 --> 00:29:40,360 Speaker 2: extent is there any kind of bottlenecks in terms of 552 00:29:40,440 --> 00:29:43,000 Speaker 2: manufacturing to kind of obviously this enormous demand to keep 553 00:29:43,000 --> 00:29:45,520 Speaker 2: building out these data centers, And are. 554 00:29:45,440 --> 00:29:46,320 Speaker 3: There any bottlenecks? 555 00:29:46,440 --> 00:29:50,880 Speaker 2: Are you investing in additional manufacturing capacity? I mean, Angela, 556 00:29:50,960 --> 00:29:51,800 Speaker 2: go to you first. 557 00:29:54,400 --> 00:30:00,120 Speaker 4: Sure, yes, yes to all that we are investing in 558 00:30:00,320 --> 00:30:04,160 Speaker 4: additional manufacturing. We just announced a new factory. We have 559 00:30:04,680 --> 00:30:09,520 Speaker 4: two factories in the Dallas area. Cd use our hot commodity. 560 00:30:10,400 --> 00:30:13,880 Speaker 4: We've announced a range of CDUs and Schneider just made 561 00:30:13,920 --> 00:30:18,760 Speaker 4: an announcement recently as well. So yeah, it's growing gang 562 00:30:18,840 --> 00:30:19,520 Speaker 4: Vestors right now. 563 00:30:20,160 --> 00:30:24,040 Speaker 2: And I mean, and Joe, what extent is are the 564 00:30:24,040 --> 00:30:26,440 Speaker 2: products that you're manufacturing like commercial and being sold at 565 00:30:26,440 --> 00:30:29,000 Speaker 2: scale right now? Like what's your timeline for the next 566 00:30:29,080 --> 00:30:30,680 Speaker 2: few years? I'll go to jailers. 567 00:30:30,920 --> 00:30:34,480 Speaker 8: Yeah, So for us, we're just about to commission our 568 00:30:35,560 --> 00:30:38,200 Speaker 8: first commercial scale cable winding machine that will go live 569 00:30:38,320 --> 00:30:41,480 Speaker 8: this year, and that will be our first commercial, commercial 570 00:30:41,480 --> 00:30:46,240 Speaker 8: scale production line. We do have one production line already 571 00:30:46,320 --> 00:30:47,920 Speaker 8: up and running, but that will be sort of a 572 00:30:48,120 --> 00:30:51,880 Speaker 8: broad scale commercial lines. We're busy investing exactly in that 573 00:30:52,080 --> 00:30:55,640 Speaker 8: manufacturing and as well as building out our supply chain. Right, 574 00:30:55,680 --> 00:30:59,680 Speaker 8: we are a system, so we source components from all 575 00:30:59,720 --> 00:31:02,600 Speaker 8: sorts of other vendors and we're kind of bringing those 576 00:31:02,640 --> 00:31:03,400 Speaker 8: together and then. 577 00:31:03,320 --> 00:31:05,000 Speaker 5: Shipping them out to customers. 578 00:31:05,120 --> 00:31:07,360 Speaker 8: That's top priority for twenty twenty six for us. 579 00:31:09,040 --> 00:31:11,120 Speaker 9: When we started, we went to one of the top 580 00:31:11,320 --> 00:31:13,600 Speaker 9: consulting companies in the world and we showed them our 581 00:31:13,640 --> 00:31:16,320 Speaker 9: device concept and we said, hey, look we got this 582 00:31:16,480 --> 00:31:19,960 Speaker 9: ultra small factory to make it. And they said not 583 00:31:20,120 --> 00:31:22,239 Speaker 9: in those words. They said, hey, you're an idiot. If 584 00:31:22,280 --> 00:31:24,440 Speaker 9: you get this product, you won't ever be able to 585 00:31:24,520 --> 00:31:25,440 Speaker 9: keep up with the demand. 586 00:31:26,000 --> 00:31:28,000 Speaker 6: So what we urge you to do is come up 587 00:31:28,040 --> 00:31:28,280 Speaker 6: with a. 588 00:31:28,360 --> 00:31:32,200 Speaker 9: Small work cell that you can expand far faster than 589 00:31:32,200 --> 00:31:35,400 Speaker 9: the orders come in. So we took that advice to heart, 590 00:31:35,440 --> 00:31:37,920 Speaker 9: and we have perfected a five thousand square foot work 591 00:31:38,040 --> 00:31:40,600 Speaker 9: cell in which we can make four hundred megawatts a 592 00:31:40,680 --> 00:31:44,440 Speaker 9: year for two to three million dollars capex, and so 593 00:31:44,680 --> 00:31:46,560 Speaker 9: so you're going to get hundreds of millions of dollars 594 00:31:46,600 --> 00:31:49,320 Speaker 9: on revenue from that. But when we go in there 595 00:31:49,320 --> 00:31:51,840 Speaker 9: and we get one hundred thousand square foot facility, we 596 00:31:51,960 --> 00:31:54,360 Speaker 9: can moderate the capex of two to three million per 597 00:31:54,400 --> 00:31:57,760 Speaker 9: work sell to speed up as the orders come in, 598 00:31:58,240 --> 00:32:01,720 Speaker 9: and so we're putting in about a multi gigawatt facility 599 00:32:01,840 --> 00:32:04,840 Speaker 9: in Raleigh, North Carolina, and we're looking at a similar 600 00:32:04,920 --> 00:32:08,960 Speaker 9: facility in UA and free trade zones to meet the 601 00:32:09,280 --> 00:32:14,360 Speaker 9: unprecedented demand. We're in negotiations for capacity reservations with some 602 00:32:14,520 --> 00:32:18,160 Speaker 9: Neo cloud people and some other data center developers. But 603 00:32:18,280 --> 00:32:21,720 Speaker 9: we're seeing a demand come and had we not engineered 604 00:32:21,760 --> 00:32:24,840 Speaker 9: the quality and the supply chain from day zero, then 605 00:32:24,880 --> 00:32:27,520 Speaker 9: the wheels would be coming off. But I think it 606 00:32:27,720 --> 00:32:31,920 Speaker 9: was that advice from that top consulting company to focus 607 00:32:32,040 --> 00:32:33,680 Speaker 9: on supply chain and quality upfront. 608 00:32:34,880 --> 00:32:37,520 Speaker 2: I understand correctly then that across all the products, it's 609 00:32:37,560 --> 00:32:41,760 Speaker 2: like a relatively on short supply chain already. It's not 610 00:32:41,840 --> 00:32:43,040 Speaker 2: like concentrated abroad. 611 00:32:43,720 --> 00:32:47,000 Speaker 9: I think in our case, our supply chain is multifaceted 612 00:32:47,440 --> 00:32:49,360 Speaker 9: because not all the demand is going to come from 613 00:32:49,400 --> 00:32:51,880 Speaker 9: the United States A big part of it is, but 614 00:32:52,080 --> 00:32:54,960 Speaker 9: we know the country of origin of every single part, 615 00:32:55,560 --> 00:32:59,200 Speaker 9: and so we can shift the supply chain based upon 616 00:32:59,360 --> 00:33:03,120 Speaker 9: what people. If people want BABA compliance, we'll have a 617 00:33:03,200 --> 00:33:07,760 Speaker 9: slightly amended supply chain if the DoD Department of Defense 618 00:33:07,840 --> 00:33:11,480 Speaker 9: wants exclusive US manufacturing, then then the cost is going 619 00:33:11,520 --> 00:33:13,440 Speaker 9: to be a little bit higher, but so it is 620 00:33:13,520 --> 00:33:16,280 Speaker 9: for everybody else. So we have that, and then if 621 00:33:16,360 --> 00:33:19,800 Speaker 9: some are agnostics to the supply chain in the rest 622 00:33:19,840 --> 00:33:22,280 Speaker 9: of the world, then we can alter and compete on 623 00:33:23,080 --> 00:33:24,200 Speaker 9: cost more ruthlessly. 624 00:33:25,200 --> 00:33:26,720 Speaker 3: Please join me in and give you my finals, sir, 625 00:33:26,840 --> 00:33:27,440 Speaker 3: rundbook Box. 626 00:33:37,840 --> 00:33:40,960 Speaker 4: Today's episode of Switched On was produced by Cam Gray 627 00:33:41,200 --> 00:33:43,560 Speaker 4: with production assistance from Kamala Shelling. 628 00:33:43,760 --> 00:33:46,640 Speaker 7: Bloomberg ne EF is a service provided by Bloomberg Finance 629 00:33:46,760 --> 00:33:47,880 Speaker 7: LP and its affiliates. 630 00:33:48,000 --> 00:33:50,680 Speaker 4: This recording does not constitute, nor should it be construed 631 00:33:50,720 --> 00:33:54,480 Speaker 4: as investment a vice, investment recommendations, or a recommendation as 632 00:33:54,560 --> 00:33:57,480 Speaker 4: to an investment or other strategy. Bloomberguinniaf SH not be 633 00:33:57,600 --> 00:34:01,320 Speaker 4: considered as information sufficient upon which to base an investment decision. 634 00:34:01,480 --> 00:34:04,440 Speaker 4: Neither Bloomberg Finance LP nor any of its affiliates makes 635 00:34:04,480 --> 00:34:08,160 Speaker 4: any representation or warranty as to the accuracy or completeness 636 00:34:08,200 --> 00:34:11,200 Speaker 4: of the information contained in this recording, and any liability 637 00:34:11,239 --> 00:34:13,880 Speaker 4: as a result of this recording is expressly disclaimed.