1 00:00:00,096 --> 00:00:03,825 And when a grandmaster is planning his or her 2 00:00:03,890 --> 00:00:07,667 chess move, it consumes close to 20 3 00:00:07,764 --> 00:00:11,268 watts of power. The same level of 4 00:00:11,542 --> 00:00:15,287 sophistication and planning exercise running on an AI 5 00:00:15,578 --> 00:00:19,355 inference cluster, whether it is Claude's AI agent 6 00:00:19,599 --> 00:00:23,096 or ChatGPT's AI agent, it consumes close to 1 7 00:00:23,194 --> 00:00:26,422 megawatt of power. Welcome to 8 00:00:26,454 --> 00:00:27,671 Impact Quantum. 9 00:00:40,157 --> 00:00:43,720 Hello and welcome back to Impact Quantum, the podcast where we explore the 10 00:00:43,720 --> 00:00:47,266 emerging industry that is quantum computing. And you don't 11 00:00:47,412 --> 00:00:51,250 have to have a PhD in physics, you just need to be a little bit 12 00:00:51,250 --> 00:00:54,894 curious. And with that in mind, I have the most quantum-curious person I 13 00:00:54,894 --> 00:00:58,526 know, Candace Cooley. How's it going, Candace? It's great. Thank you for 14 00:00:58,703 --> 00:01:02,280 asking. Today's a beautiful day. Blue, blue sky. I'm very 15 00:01:02,312 --> 00:01:06,018 excited. I'm sorry, go ahead. No, no, it's— I'm— it's a little 16 00:01:06,484 --> 00:01:10,125 gloomy here today, but that's okay because it's been like 17 00:01:10,575 --> 00:01:14,361 well above 95 and humid all week. The— if you're 18 00:01:14,441 --> 00:01:18,051 watching this, you'll notice my background is a little different. I am in Hilton Head 19 00:01:18,051 --> 00:01:21,837 Island, South Carolina on a family vacation, and it's been spectacular. 20 00:01:22,286 --> 00:01:25,064 So a little warm, but spectacular. 21 00:01:26,381 --> 00:01:30,074 And I just told you today, in fact, that now it's like summer gets turned 22 00:01:30,074 --> 00:01:33,703 off in mid-August, usually here in Montreal, Quebec, 23 00:01:33,848 --> 00:01:37,541 and it just happened. So, like, we're having the 70s now. 24 00:01:37,653 --> 00:01:41,218 So I'm in, like, the mid-70s, which is beautiful weather, right? But 25 00:01:41,491 --> 00:01:45,265 you're not swimming anymore, but it's still beautiful. And I'm not complaining. I'm not 26 00:01:45,265 --> 00:01:48,798 complaining at all. So today we have the 27 00:01:48,798 --> 00:01:52,612 co-founder of Escape Photonics, Vivek. 28 00:01:53,741 --> 00:01:57,272 We're really excited to talk with you and find out more about the 29 00:01:57,272 --> 00:02:00,849 company and what you're doing there. How are you today? 30 00:02:01,969 --> 00:02:05,588 I'm good. Thanks for asking, Candace and Frank. Enjoying the weather in 31 00:02:06,073 --> 00:02:09,838 California myself. So— Yeah, you get to enjoy the 32 00:02:09,854 --> 00:02:12,555 weather like, what, 11 months out of the year? 33 00:02:15,209 --> 00:02:19,028 Yeah, it's been hotter than we expected, but today it's 34 00:02:19,077 --> 00:02:22,556 been fairly good weather. Cool. So tell me 35 00:02:22,620 --> 00:02:24,285 about Escape Photonics. 36 00:02:26,550 --> 00:02:29,622 Escape Photonics is a company that is 37 00:02:30,366 --> 00:02:34,069 looking to solve the next frontier of 38 00:02:35,864 --> 00:02:39,439 hardware that is looking to mimic 39 00:02:39,439 --> 00:02:43,208 human intelligence. Really? Oh, so it's 40 00:02:43,256 --> 00:02:46,917 not just photonics is the story. It's It sounds like 41 00:02:46,933 --> 00:02:49,866 there's a little bit of AI in there. It's— 42 00:02:50,780 --> 00:02:53,233 everything today is about solving 43 00:02:54,612 --> 00:02:58,171 the quest for human intelligence and what is the most efficient hardware to 44 00:02:58,235 --> 00:03:01,152 build it. And today's data centers are 45 00:03:01,681 --> 00:03:04,984 glorified computers that are mimicking human brains. 46 00:03:05,753 --> 00:03:09,520 And the way we are thinking about it at Xscape 47 00:03:09,520 --> 00:03:11,765 Photonics is how do we make that 48 00:03:13,209 --> 00:03:16,598 AI brain, which is trying to meet human brain's 49 00:03:16,598 --> 00:03:19,986 efficiency, as efficient as possible using a 50 00:03:20,018 --> 00:03:23,695 multicolor laser platform. Interesting. So not just 51 00:03:23,744 --> 00:03:27,501 photonics, but because usually in 52 00:03:27,501 --> 00:03:30,199 my understanding of photonics, traditionally it's been 53 00:03:31,130 --> 00:03:34,936 some one-color laser with different polarization of it. But now 54 00:03:35,049 --> 00:03:38,854 you're mixing colors into the mix. Sounds like there's a lot more bandwidth. 55 00:03:38,919 --> 00:03:42,709 And if you're able to use color. That's absolutely right. 56 00:03:42,773 --> 00:03:46,595 I think the close— the reason I keep talking 57 00:03:46,628 --> 00:03:50,402 about human brain and human intelligence is partly 58 00:03:50,402 --> 00:03:53,421 because today's human brain consumes 59 00:03:54,176 --> 00:03:57,501 35 times more power in communication than in computation. 60 00:03:58,978 --> 00:04:02,576 And the reason that happens is 61 00:04:02,672 --> 00:04:06,478 because human brains can contextualize a different 62 00:04:07,007 --> 00:04:10,484 inputs in the most efficient manner. And 63 00:04:10,872 --> 00:04:14,333 when a grandmaster is planning his or her chess move, 64 00:04:15,890 --> 00:04:18,338 it consumes close to 20 watts of power. 65 00:04:20,255 --> 00:04:23,864 The same level of sophistication and planning exercise 66 00:04:23,993 --> 00:04:27,257 running on an AI inference cluster today, whether it is 67 00:04:27,899 --> 00:04:31,370 Claude's AI agent or ChatGPT's AI agent, 68 00:04:31,756 --> 00:04:35,438 it consumes close to 1 megawatt of power. Yeah, there 69 00:04:35,486 --> 00:04:39,003 is no efficiency. Yeah, 50,000x in efficiency 70 00:04:39,003 --> 00:04:42,150 difference between the two. And the reason for that is the 71 00:04:42,231 --> 00:04:45,651 communication between the processor is not efficient at all, 72 00:04:46,470 --> 00:04:50,228 and it is not fast enough. And using multicolor lasers to 73 00:04:50,501 --> 00:04:54,275 improve the communication between the processors 74 00:04:54,275 --> 00:04:57,968 and memory is a way to get closer to how 75 00:04:58,048 --> 00:05:01,292 a human brain operates and how the efficiency of AI 76 00:05:01,923 --> 00:05:05,561 inference cluster can improve. Interesting. That could be 77 00:05:05,642 --> 00:05:08,681 big because a lot of the complaints about data centers 78 00:05:09,910 --> 00:05:12,596 is that they need so much power, they need so much cooling, and they need 79 00:05:12,596 --> 00:05:16,078 so much land. And one of the things that I've always said was there's a 80 00:05:16,078 --> 00:05:19,690 lot of room for efficiency improvements in 81 00:05:19,722 --> 00:05:23,555 artificial intelligence. Because if you look at the human 82 00:05:23,588 --> 00:05:27,196 brain, it's something like on a regular baseline, it consumes 83 00:05:27,196 --> 00:05:30,976 about 25 watts of power. Right. And that's because nature had a 84 00:05:30,976 --> 00:05:34,727 lot of constraints, right? Consuming calories to the 85 00:05:34,727 --> 00:05:38,029 point where no, no 86 00:05:38,574 --> 00:05:41,764 biological system that I'm aware of could possibly consume 1 87 00:05:42,021 --> 00:05:45,868 megawatt worth of calorie, right? It's 88 00:05:45,868 --> 00:05:49,267 a luxury really only machines have. And I think also too, 89 00:05:49,668 --> 00:05:52,665 another example I like to use is the crow. Crows are 90 00:05:52,938 --> 00:05:56,419 intelligent, very intelligent. They are frequently 91 00:05:56,612 --> 00:06:00,143 rated close to a 5 or 6-year-old child in terms of intelligence. 92 00:06:01,122 --> 00:06:04,895 But they— I don't know what their calorie consumption is, but they 93 00:06:04,895 --> 00:06:07,543 have to pack all of that in a platform that could fly. 94 00:06:08,571 --> 00:06:12,327 Yeah. Right. So you have a dense platform. Yeah. You have 95 00:06:12,359 --> 00:06:15,762 evolutionary pressure on biological systems that 96 00:06:16,132 --> 00:06:19,872 thus far we've not really had in, in kind of artificial systems. So I 97 00:06:19,904 --> 00:06:23,594 find that interesting, and I'm glad somebody It's thinking outside of the 98 00:06:23,658 --> 00:06:27,151 traditional, I just throw another rack of GPUs at it. 99 00:06:27,920 --> 00:06:31,509 Which is how the industry is doing it because no one has actually thought 100 00:06:31,589 --> 00:06:35,291 outside the box like what you're talking about. And we just 101 00:06:35,291 --> 00:06:39,105 think the answer is staring right at your face. The human 102 00:06:39,217 --> 00:06:42,950 brain is all about communication. And when you focus on 103 00:06:43,046 --> 00:06:46,636 building a very efficient communication platform, then 104 00:06:47,309 --> 00:06:50,933 you end up solving that efficiency problem 105 00:06:51,623 --> 00:06:55,251 as well as you can. And today, all the 106 00:06:55,251 --> 00:06:58,654 problem solving has been focused on processing. How can I 107 00:06:58,862 --> 00:07:02,634 process as fast as possible? That's where the GPU investments have 108 00:07:02,634 --> 00:07:06,310 been coming through. And without realizing the 109 00:07:06,310 --> 00:07:09,954 fact that these GPUs still have to talk to one another 110 00:07:10,323 --> 00:07:13,838 and they still need some memory. And inter-process communication. 111 00:07:14,256 --> 00:07:18,059 Exactly. So the investment in that had been like 112 00:07:18,059 --> 00:07:21,872 an afterthought until recently. Now they just have started 113 00:07:21,872 --> 00:07:25,142 thinking about it. Yeah, it's interesting because like, it's funny you mentioned that because, 114 00:07:25,718 --> 00:07:29,004 you know, on the first watch of a Jensen Huang keynote, 115 00:07:29,645 --> 00:07:33,330 he'll talk about the processors. On the second— in the second 116 00:07:33,330 --> 00:07:37,112 watch of it, about 2-3 minutes in, he'll make a 117 00:07:37,112 --> 00:07:40,606 passing reference to the networking of them together, which, 118 00:07:41,102 --> 00:07:44,887 as I heard you say this, it's like Maybe, maybe we need to 119 00:07:44,887 --> 00:07:48,466 put equal emphasis at least. Yes. Yes. 120 00:07:48,562 --> 00:07:52,269 And now it's actually becoming more important because he talks 121 00:07:52,269 --> 00:07:55,559 about Moore's Law, which is the law that governs how well the 122 00:07:55,559 --> 00:07:59,010 processor can improve over time, is getting saturated, 123 00:07:59,572 --> 00:08:03,247 which means it's no longer about how fast you can 124 00:08:03,343 --> 00:08:06,553 process if you cannot contextualize and 125 00:08:06,553 --> 00:08:10,051 communicate that to the neighboring processor. And 126 00:08:10,163 --> 00:08:13,497 today the bottlenecks have shifted from processing 127 00:08:14,090 --> 00:08:17,825 to memory access and how well you can actually communicate between 128 00:08:17,937 --> 00:08:21,688 multiple processors and multiple memory. And that's why you start 129 00:08:21,688 --> 00:08:25,439 seeing optical technologies and memory companies 130 00:08:25,535 --> 00:08:29,318 getting all the limelight in the last few months, 131 00:08:29,783 --> 00:08:33,005 mainly because the realization that now 132 00:08:33,726 --> 00:08:37,468 I actually need a memory to think like human brain is kind of happening 133 00:08:37,581 --> 00:08:41,288 now. And while the solution has been 134 00:08:41,304 --> 00:08:44,542 staring right at our face for a long time, and the 135 00:08:45,011 --> 00:08:48,229 industry has started embracing that inevitability, 136 00:08:48,892 --> 00:08:52,612 and there is a shift in focus on how we think 137 00:08:52,612 --> 00:08:55,070 about the future of AI as we know it. 138 00:08:57,717 --> 00:09:00,655 So I'm taking all this in and my 139 00:09:01,093 --> 00:09:04,843 mind is blown. And so I want to dial down for a 140 00:09:04,843 --> 00:09:08,210 second and Just go back to the idea of what does 141 00:09:08,467 --> 00:09:12,018 light allow us to do in computing that 142 00:09:12,018 --> 00:09:15,756 electricity simply doesn't do as well? That's a great 143 00:09:15,756 --> 00:09:19,540 question. So today, electricity 144 00:09:19,620 --> 00:09:23,364 is very good in processing information. When you can 145 00:09:23,703 --> 00:09:27,480 switch on and off, that gets in some— that information 146 00:09:27,577 --> 00:09:30,478 of a transistor where you can tie, you can 147 00:09:31,218 --> 00:09:34,970 store bits of information in zeros and ones. That 148 00:09:34,970 --> 00:09:38,496 can be done extremely efficiently when you're using electricity and transistors. 149 00:09:38,931 --> 00:09:42,726 That's what is the basic building block of a processing unit. 150 00:09:43,374 --> 00:09:47,180 What electricity has limitation is transmitting 151 00:09:47,180 --> 00:09:50,942 that information. So, uh, it just 152 00:09:50,942 --> 00:09:54,782 turns out that when you have a lot of information that you are processing, 153 00:09:55,416 --> 00:09:58,925 you can, you can do it instantaneously with 154 00:09:59,103 --> 00:10:02,847 electrons. But then when you have to communicate from one 155 00:10:02,993 --> 00:10:06,731 processor to another processor, the— it becomes 156 00:10:06,748 --> 00:10:10,372 extremely difficult to transmit that information over a copper wire 157 00:10:11,225 --> 00:10:14,927 beyond a certain speed. And the reason for 158 00:10:15,041 --> 00:10:18,048 that is electrons are extremely lossy 159 00:10:19,018 --> 00:10:22,590 if you ask the electrons to communicate at high speed. 160 00:10:23,140 --> 00:10:26,346 So think of it as like it runs out of steam. when it is actually 161 00:10:26,378 --> 00:10:29,888 running at fast speed. So, and 162 00:10:30,098 --> 00:10:33,866 for a long time, the industry has been trying to figure out how 163 00:10:34,107 --> 00:10:37,824 fast I can communicate over a copper wire without 164 00:10:37,824 --> 00:10:41,220 it losing stream, losing its strength. 165 00:10:42,983 --> 00:10:46,788 And it has tried to navigate that problem by actually coming up 166 00:10:46,788 --> 00:10:50,443 with advanced circuit compensation techniques where you 167 00:10:51,086 --> 00:10:54,869 all the information that gets lost when you're transmitting through the copper 168 00:10:54,869 --> 00:10:58,347 wire, you end up recovering that information using 169 00:10:58,539 --> 00:11:02,098 advanced encoding technologies. And this is where 170 00:11:02,162 --> 00:11:05,977 technologies like SerDes technology and 171 00:11:06,954 --> 00:11:10,561 electronic copper wire technology came into picture. It 172 00:11:10,561 --> 00:11:13,510 just turns out that when you're doing that, you can only 173 00:11:13,622 --> 00:11:17,389 transmit— over time, the industry has evolved to a level where it 174 00:11:17,389 --> 00:11:21,030 can transmit, say, information at 16 gigabit per 175 00:11:21,078 --> 00:11:24,735 second to 32 to 100. And today the industry can 176 00:11:24,880 --> 00:11:28,072 transmit information at 100 gigabits per second 177 00:11:29,002 --> 00:11:32,804 over 2-meter copper cable. But when it goes 178 00:11:32,804 --> 00:11:36,173 to 200 gigabit per second on a single copper wire, 179 00:11:36,574 --> 00:11:40,280 it cannot transmit for more than half a meter or even say 180 00:11:40,873 --> 00:11:44,691 less than say 10 meters. And as the speed continues to 181 00:11:44,820 --> 00:11:48,270 go up, The physical limitation is electrons just cannot 182 00:11:48,270 --> 00:11:52,089 travel long enough over a copper wire. And this is the same 183 00:11:52,202 --> 00:11:55,925 problem that happened in the telecom world long time back when 184 00:11:55,973 --> 00:11:59,439 the industry went away from copper wire to fiber 185 00:11:59,487 --> 00:12:02,858 optical cable. The reason is, when, 186 00:12:03,275 --> 00:12:06,886 when you replace copper wire with fiber optical cable, the 187 00:12:06,886 --> 00:12:10,545 photons don't have the same energy loss when it is 188 00:12:10,577 --> 00:12:13,239 transmitting at high speed. that electrons do. 189 00:12:14,248 --> 00:12:17,886 So photons can carry information across like 190 00:12:18,190 --> 00:12:21,795 kilometers and kilometers of fiber optic cable without losing energy. 191 00:12:22,676 --> 00:12:26,233 So then you are not spending more energy on recovering that information anymore. 192 00:12:27,563 --> 00:12:30,175 So you kind of get a double boost in efficiency. 193 00:12:30,928 --> 00:12:34,533 Exactly. Because the original problem goes away. And then because the original 194 00:12:34,533 --> 00:12:38,202 problem goes away, you don't need the fixes. And what you're talking about, I know 195 00:12:38,459 --> 00:12:41,890 my dad worked in electronics and I mean, this goes back to the 196 00:12:41,890 --> 00:12:45,519 1880s when they were wiring telephone poles. Right? Like, exactly. 197 00:12:45,760 --> 00:12:49,132 Every— I forget what the exact number is, but every X number of feet, they 198 00:12:49,132 --> 00:12:52,825 had to amplify the signal. And, you know, the, 199 00:12:52,905 --> 00:12:56,726 the repeat. Exactly. Those, those amplifiers break down. They have to 200 00:12:56,742 --> 00:13:00,307 be replaced. They have to be maintained. Right? So, like, the, the switch to 201 00:13:00,355 --> 00:13:03,823 fiber was a, a no-brainer, so to speak, back in '80s, 202 00:13:04,064 --> 00:13:07,757 '90s, I guess. Yes. Which is exactly— we are 203 00:13:07,757 --> 00:13:10,856 seeing a revival of that in the datacom world 204 00:13:11,209 --> 00:13:14,902 where the bandwidth— because now the processor are 205 00:13:15,047 --> 00:13:18,836 communicating, processing information a lot, but then you are not able to 206 00:13:18,901 --> 00:13:22,449 communicate that to the neighboring processor or a memory. So then what's 207 00:13:22,449 --> 00:13:25,403 happening? You are just waiting on the 208 00:13:25,580 --> 00:13:29,402 communication to happen between multiple processors, 209 00:13:29,915 --> 00:13:33,399 and it's not happening fast enough. So then your tokens are getting 210 00:13:33,832 --> 00:13:37,427 limited by how fast you can communicate back and forth. 211 00:13:38,084 --> 00:13:41,390 And now your entire performance is limited by how fast you can 212 00:13:41,390 --> 00:13:44,648 communicate. And that is hitting a wall with copper wires. 213 00:13:45,226 --> 00:13:49,045 And now with optical cables, all we are reviving 214 00:13:49,494 --> 00:13:53,233 is the telecom industry's invention, which has 215 00:13:53,265 --> 00:13:56,860 replaced copper wire with optical cable. And with optical 216 00:13:56,876 --> 00:14:00,212 cable, photons can travel longer. And now I can, um, 217 00:14:00,709 --> 00:14:04,266 and it can travel faster. So you can travel longer and 218 00:14:04,266 --> 00:14:07,936 faster without needing additional energy. And now you can 219 00:14:07,936 --> 00:14:11,509 increase the speed of communication by increasing number of 220 00:14:11,621 --> 00:14:15,307 colors, because every color can transmit one, like, high-speed data. 221 00:14:16,028 --> 00:14:19,730 So now not only can I send 200 gigabit per 222 00:14:19,730 --> 00:14:23,288 second on a single color, I can also send 223 00:14:23,288 --> 00:14:27,058 100 such colors. So that means I 224 00:14:27,090 --> 00:14:30,349 end up getting a 100x boost in the speed of 225 00:14:30,381 --> 00:14:34,074 communication, and I also get 100x boost in the 226 00:14:34,074 --> 00:14:37,365 distance of communication, which means the 227 00:14:37,445 --> 00:14:41,154 entire data center today can be wired 228 00:14:41,202 --> 00:14:45,007 up with copper wire, with optical cables that 229 00:14:45,055 --> 00:14:48,683 can make it behave as one big processor 230 00:14:48,683 --> 00:14:52,182 unit rather than Chunks of small 231 00:14:52,214 --> 00:14:55,067 processing units waiting on one another for communication. 232 00:14:56,045 --> 00:14:59,876 I remember reading something about this. I don't know, maybe like a couple of years 233 00:14:59,908 --> 00:15:03,323 ago, they were, they were talking like, if you have an optical cable 234 00:15:03,419 --> 00:15:06,705 connecting CPU to CPU or GPU to GPU, 235 00:15:07,202 --> 00:15:10,408 right? Because you don't have the speed limits and 236 00:15:10,408 --> 00:15:13,903 limitations, like they could be across the room, right? Now, again, 237 00:15:14,175 --> 00:15:17,847 there is a latency cost because it's the speed of light. But honestly, Whether 238 00:15:17,847 --> 00:15:21,521 it's— for most applications, I'm sure someone will— you can 239 00:15:21,537 --> 00:15:24,955 correct me— whether it's across the room or like a football field away, 240 00:15:25,725 --> 00:15:28,725 the speed of light is fast enough that it's going to be 241 00:15:29,174 --> 00:15:32,833 mostly negligible. That's right. It's actually— 242 00:15:34,148 --> 00:15:37,357 you cannot even realize that because it's a 243 00:15:37,357 --> 00:15:40,614 nanosecond. So the speed of light is such that 244 00:15:41,369 --> 00:15:44,979 it takes only 5 nanoseconds over a meter. So if you're talking about a 245 00:15:44,979 --> 00:15:48,556 kilometer, You're talking about a microsecond. You might not even realize 246 00:15:48,909 --> 00:15:52,694 that it's taking time anymore. Right. And this 247 00:15:52,694 --> 00:15:56,512 is not new. The data center already realized that this is what 248 00:15:56,512 --> 00:16:00,025 they need as the distance increases. So they have been 249 00:16:00,843 --> 00:16:04,356 using optical cables within the data center too, when they're connecting 250 00:16:04,468 --> 00:16:07,885 multiple racks for a long distance, like 2-kilometer cable and 251 00:16:07,917 --> 00:16:11,654 3-kilometer cable. What has changed with AI is 252 00:16:11,670 --> 00:16:15,295 the fact that even within a rack, which was all copper cable, 253 00:16:15,936 --> 00:16:19,159 that is becoming all optical cable, which means 254 00:16:19,689 --> 00:16:23,297 optics is becoming a core part of the 255 00:16:23,297 --> 00:16:26,889 compute. It's getting closer and closer to the processor. And 256 00:16:26,921 --> 00:16:30,448 when it gets closer and closer to the processor, that division of boundary 257 00:16:30,513 --> 00:16:34,281 between processor memory and optical cable is 258 00:16:34,938 --> 00:16:38,739 diluting to a point where you need to treat each, like all 259 00:16:38,787 --> 00:16:42,156 of them together as a single building block. So the definition 260 00:16:42,541 --> 00:16:46,058 of compute is no longer just processing. The definition of 261 00:16:46,090 --> 00:16:49,911 compute becomes processing, communication, 262 00:16:50,168 --> 00:16:53,797 and memory. I have a bunch of questions. They need to be designed together. 263 00:16:54,295 --> 00:16:57,105 Yes. I'm sorry, Candace. I don't mean to— No, no, no. Go ahead. I have 264 00:16:57,105 --> 00:17:00,846 a bunch of questions. Great. First of all is, so if you take out a 265 00:17:00,846 --> 00:17:04,651 typical motherboard, I build— if you build PCs, right? Or you open 266 00:17:04,651 --> 00:17:08,070 up a laptop, right? You see a motherboard. you have each of the chips and 267 00:17:08,086 --> 00:17:11,695 there's some kind of copper and some kind of breadboard. You're talking about replacing the 268 00:17:11,695 --> 00:17:15,064 copper there with optics? That's exactly right. 269 00:17:15,626 --> 00:17:18,930 That is the difference between what has been happening in the data center till now, 270 00:17:19,299 --> 00:17:22,973 where there are optical cables, but they are transmitting much lower speed, 271 00:17:23,406 --> 00:17:27,224 let's say 100 gigabit per second. But when you keep going closer 272 00:17:27,256 --> 00:17:30,946 and closer to the processor, and if you're replacing that motherboard, 273 00:17:31,026 --> 00:17:34,716 the copper wires on a motherboard with optical cable, now you do 274 00:17:35,020 --> 00:17:38,776 optical cable becomes a new motherboard. So, I mean, 275 00:17:39,370 --> 00:17:43,127 then, uh, then what you need is, uh, think of it 276 00:17:43,127 --> 00:17:46,820 as, um, you have, uh, 2 chips right 277 00:17:46,820 --> 00:17:50,289 next to one another with an optical cable that becomes your unit. 278 00:17:50,835 --> 00:17:54,656 And you can think of this as like a mesh, like an optical mesh 279 00:17:54,849 --> 00:17:58,076 of like a fabric of processor 280 00:17:58,140 --> 00:18:01,592 units and memory units that is all tied 281 00:18:01,592 --> 00:18:05,184 together. with like a multicolor fabric where you 282 00:18:05,216 --> 00:18:08,759 can essentially communicate any processor to any 283 00:18:08,759 --> 00:18:12,173 memory at speed of light. Effectively, 284 00:18:12,189 --> 00:18:15,603 instantaneously. Instantaneously, 285 00:18:15,715 --> 00:18:19,290 effectively with a pretty wide bandwidth. Exactly. 286 00:18:19,626 --> 00:18:23,185 Which means the speed of communication within an internal 287 00:18:23,249 --> 00:18:26,583 processor today happens at petabits of 288 00:18:27,497 --> 00:18:31,250 bandwidth. Petabits is 10 to the power of 12. It sometimes— And 289 00:18:31,330 --> 00:18:34,800 also over a wider distance. So you could have, you know, let's just, 290 00:18:34,881 --> 00:18:38,592 let's come up with a ridiculous scenario, right? I can get like 291 00:18:38,656 --> 00:18:42,336 3 or 4 old PCs and then wire them together with 292 00:18:42,336 --> 00:18:46,063 optics. Obviously there's a lot of practical engineering that would, but let's put that 293 00:18:46,063 --> 00:18:49,502 aside for now. I could have that as one massive 294 00:18:49,502 --> 00:18:52,570 supercomputer 'cause they'd all work together. Interesting. Exactly. 295 00:18:53,117 --> 00:18:56,829 Because now I'm just like weaving them together and now 296 00:18:56,975 --> 00:19:00,697 the speed at which they communicate within a chip and speed at which they 297 00:19:00,762 --> 00:19:04,581 communicate outside the chip is exactly the same. So then the 298 00:19:04,581 --> 00:19:08,381 boundaries don't matter anymore. Right. So that's what optics bring to table is the 299 00:19:08,381 --> 00:19:11,629 distance and speed that copper cannot 300 00:19:12,098 --> 00:19:15,396 achieve. So now distance is a 301 00:19:15,412 --> 00:19:18,644 non-factor, whether there are 2 chips at the 2 ends of the 302 00:19:19,048 --> 00:19:22,360 football field or whether they're inside the same device, 303 00:19:22,832 --> 00:19:26,649 they are talking at exactly the same speed. So then it views 304 00:19:26,681 --> 00:19:30,209 them as you are essentially building the internet of the 305 00:19:30,754 --> 00:19:34,507 GPU world where you are— we are able 306 00:19:34,539 --> 00:19:38,035 to have a real-time conversation with Candace in Montreal and you in 307 00:19:38,773 --> 00:19:42,092 North Carolina. And there is no, absolutely no latency. 308 00:19:42,557 --> 00:19:45,219 Today, that world in the 309 00:19:46,005 --> 00:19:49,773 processor world is still happening through postal lines where 310 00:19:49,773 --> 00:19:53,538 there is a lot of, I'm still sending you a return letter and I'm 311 00:19:53,602 --> 00:19:55,926 waiting for a post box and 312 00:19:56,567 --> 00:20:00,268 USPS to like figure out where it needs to go. And I'm 313 00:20:00,268 --> 00:20:03,921 still waiting for Candace and hoping that she's actually, she has 314 00:20:04,033 --> 00:20:07,847 time to read that letter and then respond. And I'm still 315 00:20:07,863 --> 00:20:11,644 waiting and twiddling my thumbs until that happens. There's a lot of idle time. 316 00:20:12,028 --> 00:20:15,521 And this is basically, I think the easiest analogy is I'm 317 00:20:15,650 --> 00:20:19,495 introducing, um, the real-time video conferencing, and that 318 00:20:19,495 --> 00:20:23,100 cannot happen with copper wires. It has to happen with optical 319 00:20:23,100 --> 00:20:26,913 cables. So then the next question is, I suppose since you're dealing 320 00:20:26,946 --> 00:20:30,759 with photons, you can also encode quantum-level information 321 00:20:31,592 --> 00:20:34,957 in how the thing spins. Do you do that too, or is for now color 322 00:20:34,957 --> 00:20:38,610 wavelength is enough? Yeah, we are currently focusing on 323 00:20:38,706 --> 00:20:42,327 let's just make the industry navigate the 324 00:20:42,407 --> 00:20:46,173 postal era to a video conferencing era first. And then we 325 00:20:46,237 --> 00:20:50,007 can really talk about augmented reality version of it, which 326 00:20:50,007 --> 00:20:53,359 is the figuring out how can I encode 327 00:20:53,711 --> 00:20:57,480 qubits with these multicolor lasers. And our 328 00:20:57,545 --> 00:21:01,330 co-founders like Professor Alex Gater and Yoshi, they 329 00:21:01,330 --> 00:21:04,345 have been doing this comb laser technology 330 00:21:04,922 --> 00:21:08,691 for quantum applications for the longest of time. So they 331 00:21:08,772 --> 00:21:12,444 are very well aware of the application as it relates to 332 00:21:12,476 --> 00:21:16,135 quantum computing. And they do believe that the underlying 333 00:21:16,312 --> 00:21:19,732 platform that we are currently building at Xscape can be 334 00:21:20,567 --> 00:21:23,489 easily applicable there in terms of multi-qubit 335 00:21:23,875 --> 00:21:27,295 encoding and being able to control multiple 336 00:21:27,311 --> 00:21:30,281 qubits at the same time and being able to 337 00:21:31,341 --> 00:21:35,002 use that as a pump for low noise 338 00:21:35,018 --> 00:21:37,427 qubit generation. And I 339 00:21:37,668 --> 00:21:41,088 wouldn't claim to know anything 340 00:21:41,200 --> 00:21:44,988 about quantum, well enough to comment on what the 341 00:21:44,988 --> 00:21:48,775 technical implications would look like, except the fact that the 342 00:21:48,823 --> 00:21:52,353 application of comb lasers in quantum has been a deep 343 00:21:52,369 --> 00:21:55,787 field of research for my co-founders for over a decade. 344 00:21:56,284 --> 00:21:59,782 And there is a lot of interest in that field when that 345 00:21:59,911 --> 00:22:02,975 becomes a reality. And once the fiber is there, 346 00:22:04,404 --> 00:22:08,158 it can carry— if it can carry the same photon with or 347 00:22:08,158 --> 00:22:11,962 without extra quantum information thrown into it. That's exactly right. So 348 00:22:11,962 --> 00:22:14,980 you would— you're building the infrastructures, and then whatever you do on top of that, 349 00:22:14,996 --> 00:22:18,802 which presumably would be maybe another order of magnitude more 350 00:22:19,492 --> 00:22:23,104 information across from point to point. That's interesting. Sorry, Candace, I'll 351 00:22:23,104 --> 00:22:26,059 stop hogging the mic now. No, no, this is fantastic. So 352 00:22:27,199 --> 00:22:30,940 does photonics change what is computationally 353 00:22:31,052 --> 00:22:34,568 possible, or does it mainly let us do the same 354 00:22:34,745 --> 00:22:38,446 things faster and more efficiently? In today's data center, 355 00:22:40,124 --> 00:22:43,673 it is mainly focused on efficiency and 356 00:22:43,673 --> 00:22:47,349 speed. And with the applications like quantum 357 00:22:47,414 --> 00:22:49,983 computing, it changes the way how computing is done. 358 00:22:51,556 --> 00:22:55,366 And photonics and the platform that we are building for that is 359 00:22:55,366 --> 00:22:58,482 certainly the base infrastructure on which 360 00:22:59,063 --> 00:23:02,753 people can reimagine the way the computing can be done. 361 00:23:03,562 --> 00:23:06,718 I still think that's going to be the next wave of innovation that is going 362 00:23:06,718 --> 00:23:10,227 to ride on top of the existing AI wave. 363 00:23:11,157 --> 00:23:14,698 Would the— would it run cooler 364 00:23:14,874 --> 00:23:17,998 slightly? Because obviously CPU heat is going to be CPU heat. 365 00:23:18,767 --> 00:23:22,549 But the fact that it seems like it would run slightly cooler, right? If 366 00:23:22,549 --> 00:23:26,250 not chips themselves, the interconnecting parts would not heat up as much. 367 00:23:26,955 --> 00:23:30,560 100%, because you are— you no longer need that many overhead to 368 00:23:30,560 --> 00:23:32,949 communicate anymore. Right. You don't need 369 00:23:32,949 --> 00:23:36,718 additional overhead on figuring out 370 00:23:36,718 --> 00:23:40,263 where the loss of the signal comes from, recovering that loss, 371 00:23:40,889 --> 00:23:44,643 and spending additional power on that. So, 372 00:23:45,156 --> 00:23:48,412 and you don't have to necessarily worry about errors 373 00:23:48,508 --> 00:23:52,342 associated with sending it through copper wire. So when you're using 374 00:23:52,374 --> 00:23:56,112 optical cables, you end up like removing 375 00:23:57,010 --> 00:24:00,408 a lot of additional signal recovery 376 00:24:00,408 --> 00:24:03,732 circuits that consume some significant portion of the 377 00:24:03,796 --> 00:24:07,342 communication power. So you're looking at roughly 10 to 378 00:24:07,535 --> 00:24:11,375 20x improvement in overall energy efficiency of the 379 00:24:11,375 --> 00:24:15,183 system. And in the future, you can even 380 00:24:15,794 --> 00:24:19,257 unlock an additional vector with multicolor lasers as well. 381 00:24:19,853 --> 00:24:23,285 Interesting. So what 382 00:24:23,366 --> 00:24:26,998 happens when you combine photonics and electronics? 383 00:24:28,043 --> 00:24:30,707 Rather than treating them as competing technologies? 384 00:24:32,893 --> 00:24:36,356 The short answer is they are not competing technologies. I still think you need 385 00:24:36,356 --> 00:24:38,418 electronics for processing the information. 386 00:24:40,012 --> 00:24:43,395 And my view is today the most 387 00:24:43,395 --> 00:24:46,632 efficient way of doing it is to marry electronics and 388 00:24:46,632 --> 00:24:50,418 photonics and use electronics wherever you 389 00:24:50,515 --> 00:24:54,155 can and use photonics wherever you must. 390 00:24:54,552 --> 00:24:58,079 which is in the communication of multiple electronic 391 00:24:58,079 --> 00:25:01,863 circuits at high speed and still 392 00:25:01,927 --> 00:25:05,246 rely on electrons for processing information. 393 00:25:06,176 --> 00:25:09,591 Even when you're looking at quantum, you still need quantum error correction 394 00:25:10,024 --> 00:25:12,461 and additional circuits that are still 395 00:25:12,894 --> 00:25:16,164 electronics-driven. And electronics are something that is 396 00:25:16,197 --> 00:25:20,044 extremely efficient when it comes to doing math 397 00:25:20,044 --> 00:25:22,898 operations and doing complicated 398 00:25:23,668 --> 00:25:26,285 signal processing techniques. So 399 00:25:27,409 --> 00:25:31,086 that's where the strength of electronics lies, and I will continue 400 00:25:31,359 --> 00:25:35,180 leveraging electronics for that, which is information 401 00:25:35,276 --> 00:25:38,792 creation, and using photonics for information 402 00:25:39,692 --> 00:25:43,159 communication transfer, and using memory for 403 00:25:43,159 --> 00:25:46,820 information storage. So I do think each of them have 404 00:25:46,868 --> 00:25:50,689 their own strengths, and we just— the ideal platform should 405 00:25:50,689 --> 00:25:52,311 use the best of all 3 of them. 406 00:25:54,742 --> 00:25:55,760 That makes a lot of sense. 407 00:25:58,580 --> 00:26:02,270 So if AI keeps demanding more computing power, 408 00:26:03,833 --> 00:26:07,462 where do you think light becomes essential rather than simply 409 00:26:07,835 --> 00:26:08,663 advantageous? 410 00:26:12,079 --> 00:26:15,762 Today, at this point of time, the inflection point of 411 00:26:17,081 --> 00:26:20,866 light becoming essential is happening And we are living through that 412 00:26:20,866 --> 00:26:23,521 transition phase as we speak today. 413 00:26:26,545 --> 00:26:29,750 The— it turns out when you are processing large 414 00:26:29,928 --> 00:26:33,630 context information and when you are trying 415 00:26:33,727 --> 00:26:37,065 to do reasoning applications where 416 00:26:37,677 --> 00:26:41,206 the model has to think through different scenarios, 417 00:26:42,132 --> 00:26:45,817 contextualize different scenarios in its head, and think 418 00:26:45,817 --> 00:26:49,506 about what is the right optimum solution for a 419 00:26:49,506 --> 00:26:52,857 particular query. What the AI model is 420 00:26:52,969 --> 00:26:56,706 actually doing is mimicking the way humans 421 00:26:56,738 --> 00:26:59,768 are thinking through different scenarios and 422 00:27:00,266 --> 00:27:03,120 thinking about what is the right answer and what should I really 423 00:27:04,130 --> 00:27:07,401 respond and how I should respond to a particular question and why. 424 00:27:08,267 --> 00:27:11,779 And that becomes more and more complicated when you are getting a lot of 425 00:27:12,132 --> 00:27:15,714 multimodal inputs. What I mean by that is I'm, I'm watching you 426 00:27:16,388 --> 00:27:19,087 speak. I'm also hearing 427 00:27:20,468 --> 00:27:24,066 Frank's feedback about certain technology. 428 00:27:24,563 --> 00:27:28,274 And at the same time, I'm also thinking about what does the AI 429 00:27:28,354 --> 00:27:31,662 model would look like under these scenarios. 430 00:27:32,128 --> 00:27:35,919 And then I'm coming up with my response. And when the 431 00:27:35,983 --> 00:27:39,596 level of sophistication and the context continues to grow. 432 00:27:40,222 --> 00:27:43,466 In other words, as the AI model is going to grow from being a 433 00:27:43,610 --> 00:27:47,143 toddler to being an educated scholar, the 434 00:27:47,223 --> 00:27:50,852 experience that it is going to get and the context it's going to 435 00:27:51,061 --> 00:27:54,176 continue to have when it is responding is going to evolve. 436 00:27:54,994 --> 00:27:58,720 And when that is happening, it just turns out that the 437 00:27:58,880 --> 00:28:02,284 efficiency with which it is able to recover information 438 00:28:02,702 --> 00:28:06,102 is becoming the key factor that 439 00:28:06,183 --> 00:28:09,694 makes light essential rather than a nice-to-have. 440 00:28:10,480 --> 00:28:14,312 And the right figure of merit of how to 441 00:28:14,344 --> 00:28:17,744 actually evaluate it comes from thinking about 442 00:28:18,658 --> 00:28:22,362 how many tokens per second, which is the way 443 00:28:22,506 --> 00:28:25,970 AI models respond, am I able to generate? Like, how many 444 00:28:26,050 --> 00:28:29,898 tokens per second am I generating and how much of calories am I 445 00:28:29,898 --> 00:28:33,725 consuming, which is the megawatts? So that tokens per second 446 00:28:33,757 --> 00:28:37,285 per megawatt today is, say, 447 00:28:39,254 --> 00:28:42,779 10,000x when you are thinking about very short 448 00:28:42,779 --> 00:28:46,384 context length. If you are just asking what is an apple, then it can do 449 00:28:46,400 --> 00:28:49,826 it at like a speed of like 10,000 tokens per second per megawatt. 450 00:28:50,535 --> 00:28:54,293 But when you are asking it to process a certain slide deck and asking 451 00:28:54,326 --> 00:28:57,922 you this, asking AI this question, then it is generating 452 00:28:57,955 --> 00:29:01,529 tokens at 10 tokens per second per 453 00:29:01,593 --> 00:29:05,042 megawatt. So as the context length is going 454 00:29:06,912 --> 00:29:10,607 larger and larger, the agents are not able 455 00:29:10,639 --> 00:29:14,108 to reason out in a timely manner that is economically 456 00:29:14,108 --> 00:29:17,709 viable because it cannot continue to consume that much 457 00:29:18,454 --> 00:29:22,258 power when it is providing that response. So the 458 00:29:22,436 --> 00:29:26,264 key is for this large context application, when you want 459 00:29:26,264 --> 00:29:30,022 the AI robots to behave like humans, light 460 00:29:30,022 --> 00:29:33,796 becomes essential. It's no longer necessary. Like, it 461 00:29:33,796 --> 00:29:37,538 is a fundamental building block that needs to be factored in 462 00:29:37,650 --> 00:29:41,424 when you're designing these systems. But if you are really just focused 463 00:29:41,424 --> 00:29:45,166 on one-off, like, questions on what is an apple, what is a 464 00:29:45,278 --> 00:29:48,892 mango, and describe the weather today, those are all, 465 00:29:48,908 --> 00:29:52,698 like, much smaller context, simple applications where you 466 00:29:52,762 --> 00:29:56,324 probably don't need light. So it's really a matter of how 467 00:29:56,452 --> 00:30:00,062 smart you want the AI to be. Light makes it way smarter 468 00:30:00,497 --> 00:30:04,334 and way more intelligent. If you just want the models 469 00:30:04,366 --> 00:30:07,590 to be like a toddler, then you don't need light anymore. 470 00:30:10,793 --> 00:30:14,177 Interesting. Yeah, because I think that we're— as we go push the frontier models and 471 00:30:14,209 --> 00:30:18,057 things like that, we're really running— we're outrunning our 472 00:30:18,057 --> 00:30:21,674 ability to improve infrastructure. in terms of the chips. And this 473 00:30:21,722 --> 00:30:25,507 seems like would be a good solution to that 474 00:30:25,539 --> 00:30:28,875 problem. Yes. Fundamentally today, 475 00:30:29,965 --> 00:30:32,660 that's why the optics industry is 476 00:30:33,959 --> 00:30:37,407 becoming a hot topic of conversation across the 477 00:30:37,487 --> 00:30:41,256 entire industry, mainly because it 478 00:30:42,058 --> 00:30:45,907 is supply constrained and the demand 479 00:30:45,907 --> 00:30:49,066 is outgrowing supply at a pace that the industry is not able to 480 00:30:49,131 --> 00:30:52,978 meet And that's one of the reasons why the data center 481 00:30:52,978 --> 00:30:56,777 power consumption is becoming unsustainable. So 482 00:30:56,825 --> 00:31:00,673 there is no smoke without fire. And the way the industry 483 00:31:00,673 --> 00:31:04,247 is handling it is just throwing more data centers and inefficient 484 00:31:04,344 --> 00:31:08,191 cores because at the end of the day, I can only process 10 tokens per 485 00:31:08,191 --> 00:31:11,926 second, but I still need to process 10,000 to get to a revenue. How 486 00:31:11,974 --> 00:31:15,597 do I do it? I will just buy 1,000 more GPUs. 487 00:31:16,078 --> 00:31:19,815 If I'm buying 1,000 more GPUs, now I don't have supply. Like, who's 488 00:31:19,831 --> 00:31:23,664 gonna make 1,000 more GPUs? I need to install more 489 00:31:23,664 --> 00:31:27,304 foundry capacity. So now it becomes a 490 00:31:27,304 --> 00:31:30,143 supply-constrained world where to meet an 491 00:31:30,191 --> 00:31:33,671 inefficient ecosystem design 492 00:31:33,783 --> 00:31:37,632 constraints, you have to throw money and supply at it. And then you 493 00:31:37,632 --> 00:31:41,417 start hitting the supply-constrained world that the industry is currently navigating. 494 00:31:43,607 --> 00:31:47,063 So what advances in optics have made photonic 495 00:31:47,208 --> 00:31:49,744 computing more practical today than 496 00:31:51,407 --> 00:31:54,815 10 years ago? The ability to 497 00:31:54,976 --> 00:31:58,820 process light-based devices, the devices 498 00:31:58,852 --> 00:32:02,490 that can process light encoding and convert 499 00:32:02,490 --> 00:32:06,260 electrical data to optical data. All these devices 500 00:32:06,438 --> 00:32:09,531 traditionally were manufactured in a very 501 00:32:10,751 --> 00:32:13,864 discrete, small volume manufacturing line. 502 00:32:14,473 --> 00:32:18,228 Today, the industry can process and manufacture all of them in a 503 00:32:18,389 --> 00:32:22,160 12-inch wafer, which is a silicon CMOS wafer that 504 00:32:22,160 --> 00:32:25,385 the industry typically uses for electronic 505 00:32:26,075 --> 00:32:29,878 consumer industry. So if you take an iPhone or 506 00:32:30,488 --> 00:32:34,179 an iPad or even a Pixel phone, the chips that go inside it are 507 00:32:34,179 --> 00:32:37,214 manufactured on a 12-inch CMOS wafer. And 508 00:32:37,327 --> 00:32:40,633 that's why they were able to bring down the cost 509 00:32:41,681 --> 00:32:44,406 and increase the supply 510 00:32:45,132 --> 00:32:48,766 and being able to meet the demand of the 511 00:32:48,766 --> 00:32:52,429 entire world. The photonics industry was never 512 00:32:53,136 --> 00:32:56,414 able to do that until like a few years ago when 513 00:32:57,651 --> 00:33:00,924 we demonstrated the ability to manufacture these things 514 00:33:01,773 --> 00:33:05,309 on a 12-inch wafer. Since then, I think there has been a 515 00:33:05,470 --> 00:33:09,303 clear inflection point in the ability to drive unit economics down 516 00:33:09,416 --> 00:33:13,200 and drive volume up to a point where photonics has become 517 00:33:13,200 --> 00:33:16,621 more viable at scale, and it 518 00:33:16,637 --> 00:33:19,114 unlocks new applications like photonics computing as well. 519 00:33:26,127 --> 00:33:29,937 Interesting. How do you actually use the properties 520 00:33:30,049 --> 00:33:33,459 of light to perform computation 521 00:33:34,531 --> 00:33:36,935 rather than simply transmit information? 522 00:33:39,232 --> 00:33:42,776 Yeah, that's a good question. Today we use it for only 523 00:33:42,857 --> 00:33:46,514 transmitting information, but the way to think about how do you, uh, 524 00:33:46,804 --> 00:33:50,187 how to convert that for computation is to 525 00:33:51,073 --> 00:33:54,875 think about how, um, what are the building 526 00:33:54,939 --> 00:33:58,716 blocks of computation would look like. The first building block is 527 00:33:58,925 --> 00:34:01,607 zeros and ones. Can you actually process zeros and ones? 528 00:34:02,426 --> 00:34:06,200 And the— in the analog world, 529 00:34:06,281 --> 00:34:09,991 like when you're talking about light, the way to think about the 530 00:34:09,991 --> 00:34:13,653 equivalent of it is whether when there is light, there is 531 00:34:13,830 --> 00:34:17,459 1, and when there is no light, it is 0. So 532 00:34:17,459 --> 00:34:21,234 the equivalent of zeros and ones is literally light and 533 00:34:21,234 --> 00:34:24,655 no light. So you can make a device that can 534 00:34:24,767 --> 00:34:28,586 transmit light and that can transmit no light. And 535 00:34:28,619 --> 00:34:32,324 depending upon that, you can actually process and convert that as a compute 536 00:34:32,324 --> 00:34:35,259 element. Interesting. So, 537 00:34:36,094 --> 00:34:39,655 there are devices that you can make where you can actually 538 00:34:40,104 --> 00:34:43,601 send in signals, and depending upon the signals that you 539 00:34:43,601 --> 00:34:47,194 send, either electrical signals or just optical signals, 540 00:34:47,627 --> 00:34:50,033 when you're sending in optical signals, depending upon the 541 00:34:50,755 --> 00:34:54,211 interference between these optical signals, you either get a 542 00:34:54,211 --> 00:34:56,508 response out, which can either be a 0 or a 1. 543 00:34:57,592 --> 00:35:00,989 And you can use that property of interference 544 00:35:01,636 --> 00:35:05,273 between 2 optical signals as a compute element. 545 00:35:06,462 --> 00:35:10,189 Interesting. And this is very similar to 546 00:35:10,574 --> 00:35:14,219 how, at the end of the day, everything is a wave. 547 00:35:14,283 --> 00:35:17,325 Photonics is nothing but a wave. So if you are essentially 548 00:35:18,231 --> 00:35:21,982 taking 2 streams of information and sending it through a through 549 00:35:21,982 --> 00:35:25,572 a device that can take both of it in, the 550 00:35:25,572 --> 00:35:28,937 output of it is essentially a compute output. And 551 00:35:29,498 --> 00:35:32,831 the industry has been looking at that for quite some time. Those are like some 552 00:35:32,879 --> 00:35:35,539 of the fundamental building blocks of what is being used in 553 00:35:36,084 --> 00:35:39,802 communication as well, and also what is being used in quantum 554 00:35:39,850 --> 00:35:43,648 as well. Just with the right constraints and right inputs, you 555 00:35:43,696 --> 00:35:47,206 can also make it as a compute element. I 556 00:35:47,206 --> 00:35:50,779 remember reading a science fiction short story in Omni magazine. So 557 00:35:50,827 --> 00:35:54,513 shout out for anyone who remembers Omni magazine. as a kid, 558 00:35:54,529 --> 00:35:58,056 and they were talking about— it was set, you know, some cyberpunk type 559 00:35:58,104 --> 00:36:01,838 future on it— was they were basically saying like, you know, well, 560 00:36:01,886 --> 00:36:04,339 if you use light instead of like electrons, 561 00:36:05,396 --> 00:36:09,067 interference is different. You have a lot more efficiency. You can pack things 562 00:36:09,099 --> 00:36:12,914 more densely. And it was an interesting concept. And 563 00:36:13,250 --> 00:36:16,103 I wasn't sure if it was still science fiction, but it sounds like it's less 564 00:36:16,151 --> 00:36:19,646 science fiction today than it was when I was a kid. 100%. Yeah. I mean, 565 00:36:19,758 --> 00:36:23,414 things are moving in that direction. And it's 566 00:36:23,766 --> 00:36:27,424 just the max linear equation, Maxwell's equations 567 00:36:27,520 --> 00:36:31,129 have been around for a long, long time. And the 568 00:36:31,129 --> 00:36:34,321 industry has just not figured out how to make it 569 00:36:34,433 --> 00:36:37,978 efficiently at scale and how to leverage the properties of those. 570 00:36:38,828 --> 00:36:42,582 And now we actually have a way to unlock that. 571 00:36:42,919 --> 00:36:46,111 That's always the research and the development, right? Like some 572 00:36:46,367 --> 00:36:49,994 physicist comes up with some crazy, figures something out crazy. And 573 00:36:49,994 --> 00:36:53,440 then it could take like 50 years or 100 years to figure out how to, 574 00:36:54,165 --> 00:36:57,804 how to engineer that in a practical way. Yeah. Yeah. Yeah. 575 00:36:57,868 --> 00:37:00,503 Interesting. And it's really just figuring out, 576 00:37:02,120 --> 00:37:05,919 realizing at the end of the day that all these are just zeros and 577 00:37:05,951 --> 00:37:09,718 ones. Once you realize that the way to think about zeros and 578 00:37:09,751 --> 00:37:13,496 ones is just, and you add colors to it and you add 579 00:37:15,482 --> 00:37:19,194 circuits to it and you, you do math around that. But the 580 00:37:19,210 --> 00:37:22,921 way computers do math is very different than the way we do math because 581 00:37:22,986 --> 00:37:26,296 everything is zeros and ones. So you just have to line them up in a 582 00:37:26,296 --> 00:37:29,932 way where it meets the same outcome as we do. So 583 00:37:30,094 --> 00:37:33,180 do you need new network protocols defined? Like, is there gonna be like an 584 00:37:33,308 --> 00:37:37,014 802.11 something? Like, or 585 00:37:37,014 --> 00:37:40,592 does it already exist? Like, how much of this— so obviously the hardware 586 00:37:40,592 --> 00:37:44,297 stack changes, But how much of the other 587 00:37:44,297 --> 00:37:48,119 higher-up stack? I'm thinking transport layer stuff, like 588 00:37:48,119 --> 00:37:51,651 what else has to change because of this? Yeah, I mean, every— 589 00:37:52,631 --> 00:37:56,292 today the electronic world has like a certain stack, and then there's a quantum 590 00:37:56,405 --> 00:37:59,793 world that has a different stack. And I think the photonic computing world will 591 00:37:59,841 --> 00:38:03,663 essentially be an intermediate between them. So there will be 592 00:38:04,530 --> 00:38:08,336 certain level of changes, like you might end up reusing similar hardware in 593 00:38:08,352 --> 00:38:11,880 the electronic world. But then you will end up like bringing in photonics 594 00:38:11,880 --> 00:38:15,679 closer to the electronics and make it 595 00:38:15,727 --> 00:38:19,541 look like a photonic computing hardware, which means 596 00:38:19,830 --> 00:38:23,244 there are devices that is— that can do both processing and 597 00:38:23,821 --> 00:38:26,610 communication and memory 598 00:38:27,187 --> 00:38:30,937 storage, all with optics. And 599 00:38:31,851 --> 00:38:35,617 in the future, that will get morphed into all optical, which is for 600 00:38:35,617 --> 00:38:39,432 quantum computing. So I do view The transition from 601 00:38:39,944 --> 00:38:43,630 the current world to the quantum computing world is where the 602 00:38:43,630 --> 00:38:47,156 photonics computing is going to act as that intermediate step 603 00:38:47,717 --> 00:38:51,290 to get us there. So in terms of the stack that has to change, 604 00:38:51,643 --> 00:38:55,248 it's likely layers that are on the software protocol 605 00:38:55,248 --> 00:38:58,710 level, like how you write the software, how you write the 606 00:38:58,774 --> 00:39:02,620 communication between them. There's a chance that the communication transport layer might not 607 00:39:02,684 --> 00:39:05,874 change, But there is likely 608 00:39:06,371 --> 00:39:10,122 that the coding, the control logic, like which information 609 00:39:10,298 --> 00:39:14,129 goes where, those need to be like rewired because 610 00:39:14,145 --> 00:39:17,175 you are physically training your brain 611 00:39:17,993 --> 00:39:21,824 to communicate. You are putting the structure in place saying that when you 612 00:39:21,840 --> 00:39:25,463 have this calculation, go there. When you have this calculation, go 613 00:39:25,463 --> 00:39:28,316 there. Which today everything is done 614 00:39:28,990 --> 00:39:32,581 electronically. by encoding a packet 615 00:39:32,598 --> 00:39:36,099 header. What I mean by that, think of it as, um, 616 00:39:36,971 --> 00:39:40,580 an address that you write on every data saying that this is 617 00:39:40,694 --> 00:39:43,854 the address you need to send this particular data to. 618 00:39:45,520 --> 00:39:49,112 And an electronic circuit reads that address and says, okay, 619 00:39:49,582 --> 00:39:53,304 if this came from GPU 1, I'm reading this address, it 620 00:39:53,304 --> 00:39:57,088 says I have to send it to GPU 1 million. So let me send it 621 00:39:57,088 --> 00:39:59,895 out, let me figure out what is the path I need to take. 622 00:40:00,938 --> 00:40:04,532 Do I need to catch a flight to this particular switch? So it's 623 00:40:04,644 --> 00:40:07,821 usually not the same. It might not connect into the same 624 00:40:08,158 --> 00:40:11,896 country, right? It could be like the analogy that I'm 625 00:40:12,024 --> 00:40:15,842 drawing is a million GPUs within a cluster can 626 00:40:15,874 --> 00:40:19,644 be connected via multiple switching layers. And you 627 00:40:19,644 --> 00:40:22,966 can think of those switching layers as terminals. that an 628 00:40:23,351 --> 00:40:26,899 airport would take. So the message has to first go to one 629 00:40:26,963 --> 00:40:30,719 switch, one terminal, it processes it, then it says, okay, I'm 630 00:40:30,799 --> 00:40:34,379 currently in Newark. Now I need a way to go to, 631 00:40:35,358 --> 00:40:39,130 say, Belgium. I could either take a 632 00:40:39,130 --> 00:40:42,790 direct flight, but then I have to wait for one more 633 00:40:42,838 --> 00:40:46,530 day to take the direct flight, or I could take one stop to a 634 00:40:46,530 --> 00:40:49,693 different flight. So it optimizes based on the header 635 00:40:50,287 --> 00:40:54,124 And, um, and constraints in the address also tells 636 00:40:54,301 --> 00:40:57,978 you which, um, what is the latency constraint you have 637 00:40:58,090 --> 00:41:01,912 to follow. So you have to follow the protocol of that communication. 638 00:41:02,409 --> 00:41:05,846 So, um, then it says, okay, I'm going to take single hop, or I will 639 00:41:05,878 --> 00:41:09,474 take 1 stop or 2 stops. So every information that you 640 00:41:10,133 --> 00:41:13,713 do or think through when you're optimizing a route through kayak.com 641 00:41:14,083 --> 00:41:17,357 is what a switch is trying to think through 642 00:41:17,919 --> 00:41:20,951 in terms of where it needs to send the packet and how it needs to 643 00:41:20,951 --> 00:41:24,497 send the packet. Now, it's like IP routing tables, so to speak, 644 00:41:24,738 --> 00:41:27,995 right? Yes. It's going to give you the shortest path 645 00:41:28,734 --> 00:41:32,216 or it is going to optimize for something else. And then eventually it goes there. 646 00:41:32,745 --> 00:41:35,826 Now, if you go all optical, this electronic 647 00:41:36,291 --> 00:41:40,062 packeting layer is going to get replaced with optical circuits. 648 00:41:40,688 --> 00:41:44,395 Now, you still need that intelligence somewhere. And 649 00:41:44,395 --> 00:41:47,831 that intelligence will end up happening on the node level 650 00:41:47,831 --> 00:41:51,475 where the buyer, like, so instead of like 651 00:41:51,475 --> 00:41:55,135 relying on a switch or like an airport to figure out what is the 652 00:41:55,135 --> 00:41:58,972 fastest path, as a sender, I need to figure out what is the fastest 653 00:41:59,085 --> 00:42:02,905 path and pick that route and send it out so that I'm 654 00:42:03,018 --> 00:42:06,469 guaranteed that it is reaching the destination. So 655 00:42:07,031 --> 00:42:10,835 think of this as like 2 different methods of how do you communicate communicate when 656 00:42:10,883 --> 00:42:14,170 you're communicating to optics versus communicating through electronics. 657 00:42:14,876 --> 00:42:18,387 You just, so it's essentially like coming up with communication 658 00:42:18,419 --> 00:42:22,187 protocol is really like figuring out how to connect point A to point 659 00:42:22,187 --> 00:42:25,875 B and who controls it. Yeah. Is it the airport? 660 00:42:26,019 --> 00:42:29,322 Is it the networking layer that is gonna control it? Or is it the sender 661 00:42:29,322 --> 00:42:33,106 that is gonna control it? So then you get into this philosophical discussion 662 00:42:33,106 --> 00:42:36,922 on who should actually own it, right? As an end 663 00:42:36,922 --> 00:42:40,708 consumer, would you rather rely on yourself to come up with the 664 00:42:40,708 --> 00:42:44,399 best strategy to reach point B, or would you rely on someone like 665 00:42:44,399 --> 00:42:48,154 kayak.com to tell you, or would you go to a different one, 666 00:42:48,186 --> 00:42:51,107 right? So then, so this is the same thing 667 00:42:51,300 --> 00:42:54,814 that the analogy to what is 668 00:42:54,830 --> 00:42:58,169 happening in the data center today. And what is the most energy 669 00:42:58,201 --> 00:43:00,849 efficient way of doing that? That's an additional constraint. 670 00:43:01,972 --> 00:43:04,877 And who owns what piece of the stack and what piece of the value chain? 671 00:43:05,727 --> 00:43:09,078 And those are all the discussions that are happening as we speak, which is 672 00:43:09,752 --> 00:43:13,280 the control layers, the transport layer, the 673 00:43:13,344 --> 00:43:16,166 communication between multiple GPOs, the communication 674 00:43:16,166 --> 00:43:19,999 between one device to 675 00:43:19,999 --> 00:43:23,831 another device, multiple vendors. So all these 676 00:43:24,008 --> 00:43:27,856 conversations is what the hyperscalers are actively 677 00:43:27,856 --> 00:43:31,464 having with vendors like us and end users like— 678 00:43:32,042 --> 00:43:35,571 Anthropic and OpenAI before they have finalized the 679 00:43:35,651 --> 00:43:39,389 right solution for every generation. And that's a good point because, 680 00:43:39,646 --> 00:43:43,111 you know, because of the particulars of deployment, 681 00:43:43,913 --> 00:43:47,763 there are probably people who have no experience with AI, so 682 00:43:47,763 --> 00:43:51,389 to speak, working on these types of hardware and routing and switching 683 00:43:51,437 --> 00:43:55,111 problems, which I think traditionally we would have called just networking kind of— 684 00:43:55,384 --> 00:43:58,688 not network engineering, more network design. Or 685 00:43:59,474 --> 00:44:02,922 protocol definitions, right? There's probably a better title for it. Yeah. Like 686 00:44:03,243 --> 00:44:07,060 the people who decided what Ethernet is, right? How Ethernet should look, how it should 687 00:44:07,092 --> 00:44:10,749 work. I suspect that those conversations in the standards bodies 688 00:44:10,797 --> 00:44:14,534 are working on that now. Exactly. And it is happening. And now, to 689 00:44:14,534 --> 00:44:18,351 your point, until now, that has been a different conversation from 690 00:44:18,800 --> 00:44:22,617 who decides to make processors because they view 691 00:44:22,970 --> 00:44:26,354 exactly the way you view it, which is like, oh, it's just a networking problem. 692 00:44:26,418 --> 00:44:30,094 Let's, let's actually like throw it over the wall. Let me just make 693 00:44:30,094 --> 00:44:33,770 the most efficient processor that can process a lot of information. 694 00:44:34,284 --> 00:44:37,383 Now the conversation has shifted from being a device and 695 00:44:37,495 --> 00:44:41,236 networking, like separate buckets and combining all of 696 00:44:41,268 --> 00:44:45,089 them to a cluster-level definition, saying that it's no 697 00:44:45,121 --> 00:44:48,669 longer your problem versus my problem. It's, it's an entire 698 00:44:48,717 --> 00:44:51,591 cluster-level problem where networking is part of this decision. 699 00:44:52,506 --> 00:44:56,219 Because networking owns which processor is going to talk 700 00:44:56,219 --> 00:45:00,002 to who. And today it just happens that networking 701 00:45:00,180 --> 00:45:03,354 choice is going to govern how fast you can think. So now 702 00:45:04,557 --> 00:45:08,392 if tokens per second is determined by networking performance, you can argue 703 00:45:08,392 --> 00:45:10,131 that networking is the new compute. 704 00:45:12,260 --> 00:45:15,912 So that's why the hyper— the conversation is no 705 00:45:15,928 --> 00:45:19,775 longer a separate conversation with a processor, separate conversation with 706 00:45:19,807 --> 00:45:23,205 networking. With the AI world, this entire 707 00:45:23,254 --> 00:45:26,961 conversation with networking, memory, and processing is happening at the 708 00:45:26,961 --> 00:45:30,717 same time. And the winners have been those who have actually treated them as a 709 00:45:30,717 --> 00:45:34,473 single entity, like NVIDIA. Sure. NVIDIA actually designs the 710 00:45:34,489 --> 00:45:38,132 entire cluster as an entire system. And now to your 711 00:45:38,276 --> 00:45:42,080 point, Jensen no longer talks about GPU roadmap alone. 712 00:45:42,433 --> 00:45:46,205 He starts off with NVLink cluster. He talks about 713 00:45:46,221 --> 00:45:50,009 NVLink as a separate business unit that is generating certain 714 00:45:50,090 --> 00:45:53,478 revenue, and that has been growing 200% year over year. 715 00:45:53,992 --> 00:45:57,510 And the networking component of the business is 716 00:45:57,767 --> 00:46:01,364 outgrowing the processor component of the business for, in 717 00:46:01,413 --> 00:46:05,074 certain, in certain companies. So what you start to 718 00:46:05,074 --> 00:46:08,399 realize is when the distance becomes, when you're starting to 719 00:46:09,025 --> 00:46:12,559 design an entire cluster as a unit of compute, it's no 720 00:46:12,559 --> 00:46:15,530 longer chip boundaries, no longer 721 00:46:16,172 --> 00:46:19,882 geared towards what you're designing on silicon. Chip boundary is that entire data center, 722 00:46:20,492 --> 00:46:23,205 and everyone involved is part of the chip design. 723 00:46:24,426 --> 00:46:27,653 And that is creating this huge 724 00:46:28,921 --> 00:46:32,679 complex web of decision-making that can potentially 725 00:46:33,433 --> 00:46:36,596 delay deployment cycles. And that's where 726 00:46:36,950 --> 00:46:40,482 vendors like NVIDIA have a huge advantage because they're fully vertically 727 00:46:40,482 --> 00:46:43,962 integrated. where they're not relying on so many vendors to come 728 00:46:44,027 --> 00:46:47,762 together and arrive upon a decision before they can actually 729 00:46:48,435 --> 00:46:52,058 agree and handshake and start the development. Right. So, 730 00:46:52,202 --> 00:46:55,809 right. So the vertical aspect of integration. So yeah. So it 731 00:46:55,809 --> 00:46:59,416 sounds like the future is a lot more discipline, 732 00:46:59,561 --> 00:47:03,264 multidisciplinary. Yes. Kind of organizations have a significant 733 00:47:03,264 --> 00:47:06,807 advantage on this. Exactly. Yeah. I remember 734 00:47:07,576 --> 00:47:10,960 It must have been the keynote back in March for one of 735 00:47:11,425 --> 00:47:15,098 the GC, the big GTC in San Jose. And he had 736 00:47:15,226 --> 00:47:18,803 said InfiniBand, I think, or something like that, whatever it's called. 737 00:47:18,899 --> 00:47:22,749 But he had said what the bandwidth was 738 00:47:22,749 --> 00:47:26,036 on their networking stack for their cluster. And he said, like, you know, 739 00:47:26,197 --> 00:47:29,966 it can handle all the network traffic for the internet, entire internet 740 00:47:29,966 --> 00:47:33,655 for like 2 minutes or something like that. And I was like, that sounds almost 741 00:47:33,655 --> 00:47:36,445 impossible. But now I can kind of see, like, 742 00:47:37,249 --> 00:47:41,040 Maybe it's not impossible. And I also struggled with why would you need that 743 00:47:41,040 --> 00:47:44,430 much bandwidth on a particular cluster? And then now this makes sense. 744 00:47:45,041 --> 00:47:48,832 Yeah. Today, within not the— they are 745 00:47:48,832 --> 00:47:52,656 absolutely right. The given cluster, one rack unit 746 00:47:53,154 --> 00:47:56,560 communicates as much amount of information that the entire 747 00:47:56,560 --> 00:47:59,998 internet has communicated over the last 20 years. Wow. 748 00:48:00,802 --> 00:48:04,237 Yeah, it sounds unbelievable, but Yes. Yeah. 749 00:48:04,558 --> 00:48:08,099 Yes. That's why it's like, I mean, the volume of 750 00:48:08,099 --> 00:48:11,880 information is just going through the roof to a point where 751 00:48:12,986 --> 00:48:16,671 you have to really rewire these cables to 752 00:48:16,671 --> 00:48:19,523 make it really efficient. Otherwise, you are just going to continuously 753 00:48:20,501 --> 00:48:24,026 dissipate a lot of power that is going to make it 754 00:48:24,026 --> 00:48:27,871 unsustainable. Yeah. Wow. Yeah, I 755 00:48:27,871 --> 00:48:31,674 suppose for hyperscalers in these types of businesses, This is no longer a 756 00:48:31,674 --> 00:48:35,443 nice to have. I mean, this is straight up survival. 100%. Yeah. I 757 00:48:35,539 --> 00:48:39,292 mean, it is— there are multiple 758 00:48:39,292 --> 00:48:42,954 narratives of the data center power consumption story that you 759 00:48:42,954 --> 00:48:45,971 must have read and seen. But there is certainly 760 00:48:47,117 --> 00:48:50,845 one report from McKinsey that I find to be closest to reality, which 761 00:48:51,103 --> 00:48:54,765 is predicting close to 10 to 12% of 762 00:48:54,862 --> 00:48:58,428 the entire US power being consumed by data centers by 763 00:48:58,428 --> 00:49:02,130 2030. And I actually think it's on, it's on target, 764 00:49:02,130 --> 00:49:05,870 like it's actually on course for that. And that's a lot of power. 765 00:49:06,030 --> 00:49:09,590 10 to 12% is a lot of power. I mean, 766 00:49:09,590 --> 00:49:12,330 it's 1 out of 8. Yeah. I mean, yeah, 767 00:49:12,330 --> 00:49:15,490 yeah. It's, um, I think 768 00:49:15,490 --> 00:49:19,050 Bloomberg, somebody on Bloomberg was saying it was going to be 25%. 769 00:49:19,050 --> 00:49:22,850 Like, yeah, I wouldn't be surprised. Like 20, 35%. Yeah. If they 770 00:49:22,850 --> 00:49:26,570 don't fix it soon, power is going to become a premium and, uh, 771 00:49:27,493 --> 00:49:30,905 We all might have to revisit our plan to Mars, 772 00:49:31,258 --> 00:49:34,767 right? Yeah, seriously. So, but I do think 773 00:49:36,000 --> 00:49:39,830 efficiency is the next frontier of innovation. That is a 774 00:49:39,830 --> 00:49:43,643 must-have, not a nice-to-have, where you basically treat 775 00:49:43,675 --> 00:49:46,607 power as a hard constraint and say, I'm not going to give you more than 776 00:49:46,639 --> 00:49:49,411 10% because we have people to feed 777 00:49:50,228 --> 00:49:53,768 and we just cannot afford to give more than 10% of 778 00:49:53,993 --> 00:49:57,824 US power for these data centers. And imposing some 779 00:49:58,017 --> 00:50:01,739 hard constraints as a budget would force 780 00:50:02,028 --> 00:50:05,878 innovations on efficiency rather than— rather 781 00:50:05,878 --> 00:50:09,488 than taking it for granted that we just have to figure out a way to 782 00:50:09,568 --> 00:50:13,258 make more power. Well, it was always, it's somebody else's problem. It's 783 00:50:13,258 --> 00:50:16,787 somebody else's problem. Eventually, you kind of run out of— Until 784 00:50:17,269 --> 00:50:21,039 you start seeing power fluctuations in your house because there is a 785 00:50:21,039 --> 00:50:24,806 data center close by. Well, there's always controversy around 786 00:50:24,806 --> 00:50:28,413 electric bills going up, and sometimes that's not always the data center's fault, 787 00:50:28,573 --> 00:50:32,067 right? Sometimes there's other reasons for that. But, you know, certainly 788 00:50:32,324 --> 00:50:35,930 they are getting a lot of the blame, and ultimately it's going to make it 789 00:50:35,930 --> 00:50:39,697 harder to get a data center approved, right? If the— Yeah, 100%, I think. 790 00:50:40,178 --> 00:50:43,047 And it is also, I think it's unfair to 791 00:50:43,720 --> 00:50:47,422 limit the innovation and 792 00:50:47,567 --> 00:50:51,093 limit AI access because there is no power. So, um, 793 00:50:51,558 --> 00:50:54,861 the, the real solution is you cannot stop people from asking, uh, 794 00:50:55,326 --> 00:50:59,141 um, uh, uh, them to use AI 795 00:50:59,270 --> 00:51:02,428 applications because once you start using them, it's like using internet. 796 00:51:02,909 --> 00:51:06,468 So once you start using internet, you are not gonna go back. You're not going 797 00:51:06,468 --> 00:51:10,316 back. Like I, it, at that point of time, it becomes a part and parcel 798 00:51:10,316 --> 00:51:14,068 of your life. So the question is, uh, tomorrow if 799 00:51:14,180 --> 00:51:17,851 we just said the only solution is stop using your iPhones and stop 800 00:51:17,851 --> 00:51:21,583 using internet, people are not gonna like that either. So I almost think the 801 00:51:21,583 --> 00:51:25,006 real solution is how do you continue innovating and providing 802 00:51:25,119 --> 00:51:28,720 performance improvement and enable these AI applications 803 00:51:29,218 --> 00:51:32,898 without consuming a lot of power? Interesting. While staying within this. 804 00:51:33,075 --> 00:51:36,499 And there are a lot of solutions there. I certainly think to your 805 00:51:36,547 --> 00:51:39,528 point, human brains consume only 25 watts of power. 806 00:51:40,319 --> 00:51:43,917 And we have 300 million 807 00:51:45,901 --> 00:51:49,577 humans in US, if I'm not mistaken. So it's certainly like, so all I'm, 808 00:51:51,996 --> 00:51:55,623 trying to get towards the calorie— the power consumption of these data 809 00:51:55,703 --> 00:51:59,410 centers can be brought down drastically, and optics is a way to do that 810 00:51:59,410 --> 00:52:02,988 for sure. And the industry is starting to recognize that, 811 00:52:03,004 --> 00:52:06,502 and certainly there is a lot of progress that is being 812 00:52:06,615 --> 00:52:10,338 made in that direction. Interesting. I could talk to you for another hour, 813 00:52:10,626 --> 00:52:13,948 but I want to be respectful of your time. Where can folks find out more 814 00:52:13,980 --> 00:52:17,767 about you and about your company? We have a website, 815 00:52:17,832 --> 00:52:21,520 escapephotonics.com. where we talk about our 816 00:52:21,520 --> 00:52:25,063 laser products for data center application. And 817 00:52:25,784 --> 00:52:29,375 they can certainly join us. They can 818 00:52:29,375 --> 00:52:33,014 apply for job openings there or just reach out to us on 819 00:52:33,078 --> 00:52:36,829 any product briefing that they would need or 820 00:52:36,829 --> 00:52:39,955 just reach out if they just wanted to chat. 821 00:52:40,932 --> 00:52:44,299 Excellent. Fantastic. Awesome. Any parting thoughts, Candace? 822 00:52:44,860 --> 00:52:48,707 This was fantastic. I'm telling you, it was a brilliant 823 00:52:48,707 --> 00:52:52,460 conversation. Thank you so much for your brain. It was great. 824 00:52:52,878 --> 00:52:56,651 It was awesome. I won't look at fiber optic cables quite the 825 00:52:56,731 --> 00:53:00,441 same way again. Totally, exactly. We'll let the outro 826 00:53:00,489 --> 00:53:03,475 music play. Awesome. Thank you. 827 00:53:24,819 --> 00:53:28,535 The multiverse is skanking, skanking in time. Black holes are 828 00:53:28,535 --> 00:53:31,913 wailing in a horn line so fine. From Planck scales to planets, they're 829 00:53:31,913 --> 00:53:35,696 connecting the dots. Candace and Frank, they're the cosmic 830 00:53:35,696 --> 00:53:39,177 hotshots! 831 00:53:39,983 --> 00:53:43,686 Quantum hotshots! Turn it up fast. 832 00:53:44,009 --> 00:53:47,328 Candace and Frank blowing my mind at last. Quantum 833 00:53:47,328 --> 00:53:50,876 Podcast, they're breaking the mold. Science has got beats. 834 00:53:51,510 --> 00:53:54,762 It's bold and it's gold.