WEBVTT 1 00:00:03.399 --> 00:00:07.719 Welcome to Bedtime Astronomy. Explore the wonders of the cosmos 2 00:00:07.759 --> 00:00:12.279 with our soothing Bedtime Astronomie podcast. Each episode offers a 3 00:00:12.359 --> 00:00:16.320 gentle journey through the stars, planets, and beyond, perfect for 4 00:00:16.399 --> 00:00:20.239 unwinding after a long day. Let's travel through the mysteries 5 00:00:20.239 --> 00:00:22.440 of the universe as you drift off into a peaceful 6 00:00:22.480 --> 00:00:23.839 slumber under the night sky. 7 00:00:26.879 --> 00:00:29.879 Okay, I want you to start by picturing something well, 8 00:00:29.920 --> 00:00:33.880 truly immense, the Sun. Our sources are telling us that 9 00:00:33.920 --> 00:00:36.079 the Sun, just doing its thing, pumps out more power 10 00:00:36.079 --> 00:00:40.200 than me. Get this, one hundred trillion times all the 11 00:00:40.280 --> 00:00:44.320 electricity humanity generates across the globe. I mean that number. 12 00:00:44.600 --> 00:00:46.759 It's not just big, it's almost impossible to really wrap 13 00:00:46.799 --> 00:00:48.960 your head around, right, Yeah, that scale of energy. And 14 00:00:49.039 --> 00:00:51.719 yet down here on Earth or a modern world, especially 15 00:00:52.280 --> 00:00:56.000 our push for bigger and bigger AI artificial intelligence, it's 16 00:00:56.039 --> 00:00:58.240 kind of hitting a ceiling, a real physical limit based 17 00:00:58.280 --> 00:01:00.479 on how much energy we can actually feed these things 18 00:01:00.640 --> 00:01:03.520 now we cool them. We're burning through massive amounts of 19 00:01:03.520 --> 00:01:06.120 power just to train and run the really big machine 20 00:01:06.159 --> 00:01:08.560 learning systems, and that demand is just going up and up. 21 00:01:08.799 --> 00:01:11.159 So what happens when a company like Google, I mean, 22 00:01:11.200 --> 00:01:13.280 a company that's all in on scaling AI, looks up 23 00:01:13.319 --> 00:01:15.319 and asks, well, what if the best place, maybe the 24 00:01:15.319 --> 00:01:18.560 only sustainable place to really scale this stuff isn't here. 25 00:01:18.599 --> 00:01:21.799 What if it's you know, up there in space, tap 26 00:01:21.840 --> 00:01:24.640 and right of that massive power source. Today we're doing 27 00:01:24.640 --> 00:01:28.280 a deep dive into a really bold idea, Google's project Suncatcher. 28 00:01:28.319 --> 00:01:32.000 We've got this fascinating, super detailed research paper that outlines 29 00:01:32.040 --> 00:01:35.799 their technical plan for actually moving entire data centers, huge 30 00:01:35.799 --> 00:01:38.480 stacks of computers running their own custom chips in the 31 00:01:38.519 --> 00:01:41.400 low Earth orbit. It sounds like something straight out of 32 00:01:41.400 --> 00:01:44.640 sci fi, honestly, but when you look at the engineering details, 33 00:01:44.680 --> 00:01:47.879 they're surprisingly well grounded in tech. It's either here now 34 00:01:48.079 --> 00:01:48.799 or very close. 35 00:01:49.079 --> 00:01:52.359 Yeah, And that grounding, that sort of technical seriousness, is 36 00:01:52.400 --> 00:01:55.760 what makes this more than just a cool thought experiment. 37 00:01:55.799 --> 00:01:59.000 It's real engineering because when we talk about AI right now, 38 00:01:59.040 --> 00:02:01.560 we're basically talking about an energy crisis that's just around 39 00:02:01.599 --> 00:02:04.799 the corner. Training the biggest models today, they already use 40 00:02:04.840 --> 00:02:08.240 more power than small cities. So projects on Ketcher it's 41 00:02:08.240 --> 00:02:13.000 a serious engineering first attempt to tackle what's fast becoming 42 00:02:13.080 --> 00:02:15.680 a huge bottleneck for the whole tech industry. It's an 43 00:02:15.680 --> 00:02:19.120 infrastructure problem fundamentally, So for you listening in, we're going 44 00:02:19.159 --> 00:02:22.039 to try and get past that initial wow factor and 45 00:02:22.120 --> 00:02:25.439 really unpack the nuts and bolts of how Google actually 46 00:02:25.439 --> 00:02:28.080 thinks they can make this work. We've got three main 47 00:02:28.080 --> 00:02:31.199 areas we're going to break down today. First, the really huge, 48 00:02:31.400 --> 00:02:34.560 undeniable energy advantage you get in orbit and the specific 49 00:02:34.639 --> 00:02:38.120 orbital mechanics they need. Second, the let's call it extreme 50 00:02:38.159 --> 00:02:41.039 engineering that's required for the communication and for keeping a 51 00:02:41.039 --> 00:02:44.479 whole data center flying together in a tight formation. And finally, 52 00:02:44.759 --> 00:02:48.719 something pretty surprising how well the hardware might actually hold up, 53 00:02:49.120 --> 00:02:52.599 and the razor thin economics that really decide if this 54 00:02:52.639 --> 00:02:55.159 whole thing makes sense in the next say decade or two. 55 00:02:55.360 --> 00:02:58.719 Okay, let's start right there with energy, because that's the 56 00:02:58.759 --> 00:03:02.080 whole point, isn't it. That's the core problem projects suncatchers 57 00:03:02.080 --> 00:03:04.840 aiming at. Like you said, the power demands for large 58 00:03:04.840 --> 00:03:09.599 scale machine learning are just enormous, So why mess around 59 00:03:09.599 --> 00:03:14.560 building ever bigger solar farms down here where you've got clouds, nighttime, 60 00:03:14.680 --> 00:03:17.840 the atmosphere blocking sunlight land issues. Yeah, when you could 61 00:03:17.879 --> 00:03:20.479 just put the computers right next door to the power 62 00:03:20.479 --> 00:03:21.159 source itself. 63 00:03:21.199 --> 00:03:23.960 It really boils down to efficiency, doesn't it. If energy 64 00:03:24.000 --> 00:03:26.319 is your biggest cost, you go where the energy is 65 00:03:26.439 --> 00:03:29.879 cheapest and most abundant, and the suncatch paper or their 66 00:03:29.879 --> 00:03:32.919 analysis it suggests that solar panels in orbit could be 67 00:03:33.039 --> 00:03:35.560 up to eight times more productive than the best ones 68 00:03:35.599 --> 00:03:38.479 we have down here on Earth. Wow, that eight x factor, 69 00:03:38.719 --> 00:03:41.479 that's the fundamental thing driving this whole project. It just 70 00:03:41.599 --> 00:03:44.039 completely changes the math before you even start talking about 71 00:03:44.080 --> 00:03:45.520 chips or rockets or anything else. 72 00:03:45.599 --> 00:03:50.000 Eight times more productive, that's a staggering difference. Why why 73 00:03:50.039 --> 00:03:51.960 such a huge gap? Is it just because there's no 74 00:03:52.039 --> 00:03:53.039 atmosphere up there? 75 00:03:53.199 --> 00:03:56.680 That's a big slice of it. Yeah, our atmosphere, you know, 76 00:03:56.719 --> 00:03:59.560 great for breathing, but it absorbs or reflects something like 77 00:03:59.599 --> 00:04:01.520 thirty percent of the Sun's energy right off the bat 78 00:04:01.560 --> 00:04:03.759 and that's on a clear day, not counting clouds or 79 00:04:03.800 --> 00:04:06.800 dust or anything. Up in LEO low Earth orbit, you're 80 00:04:06.840 --> 00:04:10.479 getting the raw deal, full spectrum, no filtering, pure sunlight. 81 00:04:10.879 --> 00:04:13.520 But there's actually another factor that's maybe even more powerful, 82 00:04:13.639 --> 00:04:16.360 especially when you think about cost, and that's the idea 83 00:04:16.399 --> 00:04:19.680 of continuous collection. See down here, even in the sunniest 84 00:04:19.720 --> 00:04:22.079 desert on Earth, you lose half your operating time just 85 00:04:22.120 --> 00:04:25.439 because of night plus you've got seasons, bad weather. But 86 00:04:25.560 --> 00:04:28.519 in orbit, especially the orbit they're planning, these satellites are 87 00:04:28.560 --> 00:04:30.160 in almost constant sunlight. 88 00:04:30.279 --> 00:04:32.560 Almost constant sunlight. Okay, And if you get rid of 89 00:04:32.600 --> 00:04:35.000 the night cycle, what big engineering headache just. 90 00:04:35.839 --> 00:04:40.759 Vanishes storage batteries. It basically eliminates the biggest, heaviest, and 91 00:04:40.759 --> 00:04:44.399 frankly most expensive part of large scale solar power on Earth, 92 00:04:44.839 --> 00:04:47.120 which is storing enough energy to last through the night. 93 00:04:47.399 --> 00:04:49.720 If you're generating power pretty much twenty four to seven, 94 00:04:49.800 --> 00:04:53.720 you just don't need those massive, heavy, costly battery banks 95 00:04:53.759 --> 00:04:57.079 that are absolutely vital for any serious solar project down here. 96 00:04:57.439 --> 00:04:59.399 Ah right, that makes sense. 97 00:05:00.240 --> 00:05:03.839 Heavy, incredibly heavy. Think about the mass budget for launch. 98 00:05:04.360 --> 00:05:07.519 Lithium ion batteries add a lot of weight. To power 99 00:05:07.560 --> 00:05:10.160 a data center on Earth through the night, you'd need 100 00:05:10.399 --> 00:05:15.240 enormous battery arrays adds mass, adds volume, adds cost in space. 101 00:05:15.319 --> 00:05:18.480 If you need far less battery mass, your satellite is lighter, 102 00:05:18.759 --> 00:05:22.759 and that directly lowers the complexity and crucially the cost 103 00:05:22.959 --> 00:05:25.879 to launch the whole thing, So that AIGHTX efficiency isn't 104 00:05:25.920 --> 00:05:28.519 just about getting more power out. It's also about needing 105 00:05:28.600 --> 00:05:30.759 left heavy stuff like batteries sent up there in the 106 00:05:30.759 --> 00:05:33.439 first place. It's minimizing that dead weight. 107 00:05:33.639 --> 00:05:36.759 Okay, so Project Suncatcher isn't just about grabbing more sunlight. 108 00:05:36.839 --> 00:05:39.920 It's using a really specific orbital trick to get that 109 00:05:39.959 --> 00:05:42.680 continuous power. So let's talk about what this thing actually 110 00:05:42.680 --> 00:05:45.279 looks like. What's the architecture. Is it one giant space station? 111 00:05:45.720 --> 00:05:48.560 No, not monolithic at all. The concept is actually a 112 00:05:48.600 --> 00:05:53.120 distributed constellation. Think lots and lots of separate satellites. Each 113 00:05:53.120 --> 00:05:55.720 one would have its own solar panels, its own processors, 114 00:05:55.720 --> 00:05:58.959 specifically Google's own ships, their TPUs, and they'd all be 115 00:05:58.959 --> 00:06:03.079 connected together with super fast laser links, optical links. And 116 00:06:03.120 --> 00:06:05.560 the orbit they chose that's really the key piece of 117 00:06:05.600 --> 00:06:10.120 engineering here. They've zero in on a sun synchronous. 118 00:06:09.560 --> 00:06:12.279 Low Earth orbit l e O Sun synchronous ALIO. 119 00:06:12.199 --> 00:06:15.759 Yeah, sometimes called SSO, and choosing that specific orbit. It's 120 00:06:15.879 --> 00:06:18.920 not random at all. It's this really clever bit of 121 00:06:19.279 --> 00:06:23.160 orbital mechanics, a kind of geometric trick designed purely to 122 00:06:23.279 --> 00:06:25.920 keep the sunlight hitting those panels almost all the time. 123 00:06:26.079 --> 00:06:28.519 Okay, for those of us listening who hear Sun synchronous 124 00:06:28.519 --> 00:06:30.639 and maybe you just picture it following the Sun around. 125 00:06:30.920 --> 00:06:33.519 Can you break down the orbital magic there? What makes 126 00:06:33.680 --> 00:06:35.040 SSO so good for power? 127 00:06:35.279 --> 00:06:37.959 Sure? So, an SSO is an orbit that goes almost 128 00:06:38.000 --> 00:06:41.319 over the poles, a near polar orbit, and the magic, 129 00:06:41.360 --> 00:06:43.120 as you put it, is how the plane of that 130 00:06:43.279 --> 00:06:47.279 orbit rotates. The orbit itself actually precesses. It sort of 131 00:06:47.279 --> 00:06:49.560 turns around the Earth at the exact same speed that 132 00:06:49.560 --> 00:06:52.000 the Earth goes around the Sun, which is about one 133 00:06:52.040 --> 00:06:55.199 degree per day. Okay, So the geometry works out such 134 00:06:55.279 --> 00:06:57.959 that the satellite always crosses the equator at the same 135 00:06:58.040 --> 00:07:00.160 local solar time every day. 136 00:07:00.199 --> 00:07:02.560 Ah Okay, so the Sun is always hitting it from 137 00:07:02.639 --> 00:07:04.319 roughly the same direction relative to. 138 00:07:04.240 --> 00:07:07.279 Its pass exactly. The satellite keeps a near constant angle 139 00:07:07.319 --> 00:07:09.759 relative to the Sun, so the solar panels can be 140 00:07:09.839 --> 00:07:13.360 oriented to be almost perfectly perpendicular to the sunlight for 141 00:07:13.439 --> 00:07:17.600 most of the orbit, maximizing collection. And the big payoff 142 00:07:17.639 --> 00:07:20.480 is that the satellite hardly ever goes into Earth's shadow. 143 00:07:21.040 --> 00:07:23.120 If it does, the eclipse is super short, maybe just 144 00:07:23.160 --> 00:07:26.319 a few minutes. Compare that to a typical l EO satellite, 145 00:07:26.360 --> 00:07:29.120 which might spend thirty or forty minutes in darkness every 146 00:07:29.160 --> 00:07:33.959 single orbit difference, huge difference. This specific orbit choice guarantees 147 00:07:34.040 --> 00:07:38.279 that almost constant sunlight, which then lets them drastically minimize 148 00:07:38.279 --> 00:07:41.279 how many batteries they need onboard. It's really the sweet 149 00:07:41.279 --> 00:07:45.720 spot maximum sun exposure for power, but still in low orbit, 150 00:07:46.000 --> 00:07:48.959 which helps keep the communication latency back down to Earth 151 00:07:49.160 --> 00:07:50.120 somewhat reasonable. 152 00:07:50.199 --> 00:07:53.800 Okay, that perfect energy supply, that constant sunlight. It leads 153 00:07:53.879 --> 00:07:56.920 us straight into the second, really big challenge, maybe the 154 00:07:56.920 --> 00:08:00.920 hardest one. Taking a static building some data center on 155 00:08:00.959 --> 00:08:04.040 the ground and turning it into a swarm of individual 156 00:08:04.040 --> 00:08:08.199 satellites flying together information that demands coordination on a scale 157 00:08:08.240 --> 00:08:10.600 that just sounds well, incredibly difficult. 158 00:08:10.680 --> 00:08:12.839 Yeah, and the coordination isn't just a nice to have, 159 00:08:13.000 --> 00:08:16.319 It's an absolute requirement, and it's driven entirely by the 160 00:08:16.360 --> 00:08:19.319 need for fast data links between the satellites, getting that 161 00:08:19.439 --> 00:08:21.720 data center level of performance the kind you need for 162 00:08:21.759 --> 00:08:24.560 these giant AI models. It absolutely depends on the satellites 163 00:08:24.600 --> 00:08:27.519 flying and what the paper calls extremely tight formation. 164 00:08:27.879 --> 00:08:30.560 Extremely tight Why why so close as it just so 165 00:08:30.560 --> 00:08:31.800 the lasers don't miss each other. 166 00:08:32.000 --> 00:08:35.120 That's definitely part of it. The pointing accuracy but the 167 00:08:35.159 --> 00:08:37.360 main reason comes back to the data link speed and 168 00:08:37.399 --> 00:08:39.840 the latency. We'll get more into the links themselves in 169 00:08:39.879 --> 00:08:43.360 a bit, but think about how a huge AI model works. 170 00:08:43.840 --> 00:08:48.440 It's splitting up the job across potentially thousands of processors simultaneously. 171 00:08:49.159 --> 00:08:51.879 All those processors need to talk to each other, share results, 172 00:08:51.919 --> 00:08:54.080 coordinate almost instantly like. 173 00:08:54.039 --> 00:08:55.519 Components inside a single computer. 174 00:08:55.559 --> 00:08:58.559 Basically pretty much in a data center on Earth, those 175 00:08:58.600 --> 00:09:01.639 processors might be connected by a few feet of fiber 176 00:09:01.679 --> 00:09:06.480 optic cable, really short distances, giving you microsecond latency incredibly 177 00:09:06.559 --> 00:09:08.639 high bandwidth. To get anywhere close to that kind of 178 00:09:08.720 --> 00:09:11.200 low latency in space, you just can't have the satellites 179 00:09:11.240 --> 00:09:13.879 drifting far apart. The paper specifies they need to be 180 00:09:13.919 --> 00:09:15.840 separated by kilometers or less. 181 00:09:15.879 --> 00:09:18.799 A kilometer or less, okay, even a kilometer in space. 182 00:09:19.159 --> 00:09:22.360 When you're talking data links, that still sounds like a 183 00:09:22.399 --> 00:09:25.279 long way for data compared to inside a server rack. 184 00:09:25.519 --> 00:09:28.200 It is, it absolutely is, And you can't beat the 185 00:09:28.240 --> 00:09:28.799 speed of light. 186 00:09:28.879 --> 00:09:29.039 Right. 187 00:09:29.320 --> 00:09:32.320 Even though light travels faster in vacuum than in fiber, 188 00:09:32.919 --> 00:09:35.919 that kilometer distance still introduces a delay you don't have 189 00:09:36.000 --> 00:09:38.600 on Earth. So they have to design the whole system 190 00:09:38.639 --> 00:09:42.320 the network protocols, the way tasks are distributed. Specifically to 191 00:09:42.360 --> 00:09:46.559 handle these slightly longer latencies, The whole constellation has to 192 00:09:46.679 --> 00:09:51.559 function like one single massive coordinated computer cluster, even though 193 00:09:51.559 --> 00:09:55.399 it's physically spread out over potentially you know, several miles. 194 00:09:55.080 --> 00:09:58.279 And the altitude they're aiming for around six hundred fifty kilometers. 195 00:09:58.320 --> 00:10:01.399 That adds another complication, doesn't it. It's LAO, sure, but 196 00:10:01.399 --> 00:10:03.879 it's not that high. There's still some atmosphere, some drag. 197 00:10:04.039 --> 00:10:08.279 Precisely, this isn't way out in geo geostationary orbit, where 198 00:10:08.360 --> 00:10:11.600 drab is basically zero down at six hundred and fifty kilometers, 199 00:10:11.679 --> 00:10:14.240 Especially when you're trying to keep potentially hundreds of these 200 00:10:14.240 --> 00:10:17.399 things packed so closely together, maybe hundreds of meters apart 201 00:10:17.720 --> 00:10:19.519 or less. Even tiny forces become a. 202 00:10:19.559 --> 00:10:21.559 Huge problem, like what kind of forces. 203 00:10:21.440 --> 00:10:25.360 Well, the residual atmospheric molecules, for one, Even though the 204 00:10:25.360 --> 00:10:28.080 air is incredibly thin up there, it's not a perfect vacuum. 205 00:10:28.480 --> 00:10:32.080 Hit a slightly denser patch, and it creates drag. If 206 00:10:32.120 --> 00:10:36.639 that drag affects one satellite's big solar panel slightly differently 207 00:10:36.639 --> 00:10:38.000 than its neighbors. 208 00:10:37.600 --> 00:10:40.799 It could nudge them out of alignment almost immediately and 209 00:10:40.879 --> 00:10:41.960 break those laser. 210 00:10:41.720 --> 00:10:44.519 Lis exactly, break the links, and the whole cluster stops 211 00:10:44.559 --> 00:10:47.440 working as a single unit. Plus you've got subtle things 212 00:10:47.480 --> 00:10:51.279 like the fact that Earth's gravity isn't perfectly uniform. All 213 00:10:51.320 --> 00:10:54.279 these tiny perturbations add up when you need such precise 214 00:10:54.519 --> 00:10:55.600 relative positioning. 215 00:10:56.279 --> 00:10:59.759 So if the forces are tiny but complex and always changing, 216 00:11:00.679 --> 00:11:03.679 how did Google figure out if they could actually solve this, 217 00:11:03.679 --> 00:11:06.879 this problem of keeping a distributed data center flying in 218 00:11:06.919 --> 00:11:08.679 formation seems almost impossible. 219 00:11:08.720 --> 00:11:10.799 Well, they didn't just guess. This is where the serious 220 00:11:10.840 --> 00:11:14.200 engineering comes in again. They built some really sophisticated physics 221 00:11:14.240 --> 00:11:17.799 simulations to model it. They wanted to analyze the stability 222 00:11:18.240 --> 00:11:20.919 of these tight formations over the long haul, over the 223 00:11:20.960 --> 00:11:22.399 whole mission lifetime. 224 00:11:22.519 --> 00:11:23.960 What did the simulations look at? 225 00:11:24.000 --> 00:11:28.720 Specifically, they modeled two main threats. First, those tiny irregular 226 00:11:28.759 --> 00:11:31.559 gravitational tugs from Earth not being a perfect sphere, what 227 00:11:31.600 --> 00:11:35.360 they're called the non gravitational field effects. And second, maybe 228 00:11:35.399 --> 00:11:40.440 more importantly in Lo, that atmospheric drag, which, like we said, 229 00:11:40.559 --> 00:11:42.840 is small at six hundred and fifty kilometers, but it's 230 00:11:42.919 --> 00:11:46.200 there and it changes depending on things like solar activity, 231 00:11:46.200 --> 00:11:50.120 puffing up the atmosphere. Even tiny differences in drag between 232 00:11:50.159 --> 00:11:54.200 adjacent satellites over time could cause them to drift apart 233 00:11:54.240 --> 00:11:55.960 significantly if you don't correct for it. 234 00:11:56.360 --> 00:11:59.480 Okay, that makes sense. You'd think hitting a slightly denser 235 00:11:59.480 --> 00:12:02.240 bit of air would slow one satellite down relative to 236 00:12:02.279 --> 00:12:06.240 its neighbor. Wouldn't that mean you need constant, powerful thruster 237 00:12:06.399 --> 00:12:09.000 burns to keep nudging them back into place. That sounds 238 00:12:09.000 --> 00:12:09.759 as a lot of fuel. 239 00:12:09.960 --> 00:12:12.519 You'd think so, right, But the conclusion from their models 240 00:12:12.559 --> 00:12:15.399 was actually surprisingly positive, which is good news for the 241 00:12:15.399 --> 00:12:19.240 project's lifespan. The simulations indicated that they should only need 242 00:12:19.559 --> 00:12:23.720 modest station keeping maneuvers to keep the formation stable over time. 243 00:12:24.000 --> 00:12:25.919 Modest that's the keyword there. 244 00:12:26.120 --> 00:12:29.080 It really is, because if they needed constant high thrust 245 00:12:29.159 --> 00:12:34.120 chemical rockets firing, they'd burn through propellant incredibly fast. That 246 00:12:34.159 --> 00:12:37.159 would limit the mission life drastically, and the whole economic 247 00:12:37.279 --> 00:12:38.399 argument kind of falls apart. 248 00:12:38.720 --> 00:12:42.799 So modest maneuvers probably means they're thinking about using things 249 00:12:42.840 --> 00:12:47.799 like electric propulsion, ion drives, hall thrusters, stuff that SIPs 250 00:12:47.840 --> 00:12:50.840 fuel but provides tiny pushes over long periods. 251 00:12:51.000 --> 00:12:54.679 That's almost certainly the implication Yeah, low thrust, high efficiency 252 00:12:54.720 --> 00:12:58.519 propulsion is perfect for this kind of longeration fine adjustment job. 253 00:12:59.000 --> 00:13:01.879 They can provide continue uk US really small corrections for 254 00:13:02.039 --> 00:13:05.080 years without needing massive fuel tanks, which means the mass 255 00:13:05.120 --> 00:13:08.600 needed for station keeping stays manageable within the overall satellite budget, 256 00:13:08.639 --> 00:13:11.440 and that helps keep the launch cost down. It all connects. 257 00:13:11.679 --> 00:13:15.919 The source material actually mentions NASA's Stereo Observatory spacecraft as 258 00:13:15.919 --> 00:13:20.039 a kind of parallel. Those probes needed really precise orbital 259 00:13:20.080 --> 00:13:23.399 control at relatively low altitudes to maintain their view of 260 00:13:23.440 --> 00:13:26.559 the Sun. The fact that Google's modeling suggests only modest 261 00:13:26.559 --> 00:13:29.240 maneuvers are needed implies that while the formation is tight, 262 00:13:29.759 --> 00:13:33.600 the natural orbital dynamics aren't totally fighting them. They can 263 00:13:33.720 --> 00:13:37.080 use these efficient, low thrust systems to nudge things back 264 00:13:37.120 --> 00:13:40.360 into place. It means that huge energy advantage they get 265 00:13:40.360 --> 00:13:43.360 from the sun synchronous orbit isn't immediately canceled out by 266 00:13:43.399 --> 00:13:46.679 needing tons of fuel just to hold the constellation together. 267 00:13:47.159 --> 00:13:49.279 Okay, all right, So we've got the massive power source, 268 00:13:49.320 --> 00:13:51.840 sorted things of the orbit, and we've got a plausible, 269 00:13:51.879 --> 00:13:55.360 simulation backed path to keeping the satellites flying to gather 270 00:13:55.399 --> 00:13:59.759 in formation using efficient thrusters. Now the third huge piece 271 00:14:00.480 --> 00:14:03.639 communication because if you can't shift data between all those 272 00:14:03.679 --> 00:14:06.679 processes at blinding speed was really low latency, then you 273 00:14:06.679 --> 00:14:08.240 don't have a data center. You just have a bunch 274 00:14:08.279 --> 00:14:11.679 of well solar powered computers sloating near each other. Not 275 00:14:11.879 --> 00:14:12.879 useful for big AI. 276 00:14:13.279 --> 00:14:16.559 Yeah, this communication challenge, it's truly the make or break 277 00:14:16.600 --> 00:14:20.759 piece for Suncatcher. We talked about AI workloads needing high 278 00:14:20.759 --> 00:14:23.679 bandwidth and low latency because the tasks get split up 279 00:14:23.679 --> 00:14:27.200 across maybe thousands of processors. If these connections between the 280 00:14:27.240 --> 00:14:30.360 processors are slow, the whole calculation grinds to a halt. 281 00:14:30.840 --> 00:14:33.440 It doesn't matter how fast the individual chips are. The 282 00:14:33.480 --> 00:14:35.639 system runs at the speed of its slowest link. 283 00:14:35.919 --> 00:14:36.039 Right. 284 00:14:36.240 --> 00:14:38.399 So, to really perform like a data center on Earth 285 00:14:38.440 --> 00:14:41.679 where data zips around on fiber optic cables at tens, 286 00:14:41.759 --> 00:14:45.200 maybe hundreds of gigabits per second, these links in space 287 00:14:45.320 --> 00:14:47.919 have to hit similar numbers across that kilometer or so 288 00:14:48.000 --> 00:14:51.080 gap we talked about, and Google set the bar incredibly 289 00:14:51.120 --> 00:14:54.440 high here. The goal they state is links supporting tens 290 00:14:54.440 --> 00:14:57.039 of terabits per second between satellite. 291 00:14:56.639 --> 00:15:00.879 Tens of terribus per seconds that's astronomical way beyond current 292 00:15:00.960 --> 00:15:02.240 inner satellite links, isn't it. 293 00:15:02.279 --> 00:15:05.000 Oh yeah, orders of magnitude faster. So how do you 294 00:15:05.039 --> 00:15:08.320 possibly achieve that kind of speed across kilometers of empty space, 295 00:15:08.840 --> 00:15:13.519 and crucially without using huge power hungry radio antennas. The 296 00:15:13.600 --> 00:15:18.320 answer is lasers. They're relying entirely on laser based optical communication. 297 00:15:18.440 --> 00:15:19.960 That's the key technology here. 298 00:15:20.279 --> 00:15:23.320 Lasers. Okay, why lasers instead of radio? 299 00:15:23.919 --> 00:15:26.720 Well, unlike radio waves which tend to spread out over 300 00:15:26.759 --> 00:15:28.679 distance and need a lot of power to keep the 301 00:15:28.679 --> 00:15:33.120 signal strong, lasers can focus light into an incredibly tight, 302 00:15:33.440 --> 00:15:38.080 narrow beam think laser pointer versus a floodlight. That tight 303 00:15:38.159 --> 00:15:41.480 focus means much less power gets wasted and you can 304 00:15:41.519 --> 00:15:45.120 pack way more information, much higher throughput into that beam. 305 00:15:45.360 --> 00:15:47.440 We've seen demos of this kind of thing before, right, 306 00:15:47.559 --> 00:15:48.799 leaser cons in space. 307 00:15:49.039 --> 00:15:49.240 Yeah. 308 00:15:49.279 --> 00:15:53.000 Absolutely. The sources mention the opal US experiment on the 309 00:15:53.039 --> 00:15:57.159 International Space Station, for example, optical payload for lasercom science 310 00:15:57.559 --> 00:16:00.320 that proved you could beam high bandwidth data using less 311 00:16:00.320 --> 00:16:03.120 pretty effectively. But the challenge here is taking that proof 312 00:16:03.120 --> 00:16:06.080 of concept and scaling it up massively, like by a 313 00:16:06.120 --> 00:16:08.559 factor of ten or twenty maybe more to get to 314 00:16:08.639 --> 00:16:09.879 those tens of terrabits. 315 00:16:09.919 --> 00:16:12.720 Okay, so the basic tech exists, but going from say 316 00:16:12.799 --> 00:16:16.120 gigabits or maybe single terabt up to tens of terrabits, 317 00:16:16.399 --> 00:16:18.840 that's a giant leap. How do they plan to squeeze 318 00:16:18.840 --> 00:16:20.919 so much more data down the same laser plate. 319 00:16:21.120 --> 00:16:25.200 They're planning to use some really advanced multiplexing techniques, basically 320 00:16:25.320 --> 00:16:28.480 finding clever ways to pack more separate data channels into 321 00:16:28.519 --> 00:16:32.960 the same physical path. The paper details two main methods working. 322 00:16:32.679 --> 00:16:36.639 Together multiplexing like splitting the beam sort of. 323 00:16:36.799 --> 00:16:41.960 The first is called dense wavelength division multiplexing or DWDM. 324 00:16:42.360 --> 00:16:45.799 This is already standard practice in long haul fiber optics 325 00:16:45.840 --> 00:16:49.120 on Earth. Imagine the laser beam is like a single highway. 326 00:16:49.519 --> 00:16:52.919 DWDM is like painting say thirty or forty different color 327 00:16:53.000 --> 00:16:56.519 lanes onto that same highway. Each color, each specific wavelength 328 00:16:56.519 --> 00:16:59.320 of light carries its own independent stream of data. 329 00:16:59.480 --> 00:17:02.960 Ah. Okay, so multiple data streams riding on different colors 330 00:17:03.000 --> 00:17:05.960 of light within the same beam. That multiplies the capacity 331 00:17:06.079 --> 00:17:07.039 right there exactly. 332 00:17:07.079 --> 00:17:09.799 It leverages the spectrum really efficiently. But then they add 333 00:17:09.799 --> 00:17:11.079 a second layer on top of that. 334 00:17:11.160 --> 00:17:12.440 Okay, what's the second layer. 335 00:17:12.480 --> 00:17:15.880 That's spatial multiplexing technology. So if d do UDM gives 336 00:17:15.880 --> 00:17:19.519 you multiple lanes on one highway, spatial multiplexing is like 337 00:17:19.559 --> 00:17:22.640 building several parallel highways right next to each other. It 338 00:17:22.680 --> 00:17:26.759 means sending multiple separate laser beams between the two satellites simultaneously, 339 00:17:27.200 --> 00:17:30.720 all aimed very precisely at an array of sensitive receivers 340 00:17:30.720 --> 00:17:31.440 on the other end. 341 00:17:31.759 --> 00:17:35.200 Wow, okay, so multiple beams and each beam is carrying 342 00:17:35.279 --> 00:17:37.960 multiple colors or wavelengths of data precisely. 343 00:17:38.079 --> 00:17:43.799 You combine DDM and spatial multiplexing multiple colors times multiple beams. 344 00:17:44.359 --> 00:17:46.720 That's how they aim to multiply the data rate so 345 00:17:46.839 --> 00:17:50.079 dramatically and hit that collective target of tens of terabits 346 00:17:50.119 --> 00:17:50.599 per second. 347 00:17:50.839 --> 00:17:53.680 That sounds incredibly complex keeping all those beams and colors 348 00:17:53.680 --> 00:17:56.720 aligned across a kilometer of space. Is this still just 349 00:17:56.920 --> 00:17:59.160 theory or have they actually tested any of. 350 00:17:59.119 --> 00:18:02.240 This is important. They have done some concrete validation, at 351 00:18:02.319 --> 00:18:04.640 least on a lab bench, which is crucial right to 352 00:18:04.640 --> 00:18:07.039 show the physics works. The research teams set up a 353 00:18:07.079 --> 00:18:11.119 demonstration combining these approaches, and they successfully transmitted data at 354 00:18:11.119 --> 00:18:13.920 a total rate of one point six terabits per second. 355 00:18:14.160 --> 00:18:17.039 Okay, one point six tvps. It's still a long way 356 00:18:17.039 --> 00:18:19.839 from tens, but it's well, it's not zero. It shows 357 00:18:19.839 --> 00:18:22.359 the combined technique actually functions exactly. 358 00:18:22.440 --> 00:18:25.680 It proves the principle. And what's really interesting here is 359 00:18:25.720 --> 00:18:28.759 how this lab success ties straight back to that need 360 00:18:28.799 --> 00:18:32.359 for tight formation flying we discussed earlier keeping the satellites 361 00:18:32.400 --> 00:18:36.200 separated by kilometers or less. That tight spacing it isn't 362 00:18:36.240 --> 00:18:39.640 just for orbital stability. It's absolutely essential for making sure 363 00:18:39.680 --> 00:18:42.839 these complex laser links can actually be scaled up and 364 00:18:42.880 --> 00:18:46.720 work reliably. The closer the satellites are, the easier it 365 00:18:46.759 --> 00:18:50.160 is to maintain that pinpoint accuracy needed for multiple laser 366 00:18:50.160 --> 00:18:53.880 beams carrying multiple wavelengths to hit tiny receivers without the 367 00:18:53.920 --> 00:18:55.440 signal degrading or getting lost. 368 00:18:55.599 --> 00:18:58.960 Ah I see. So maintaining that close distance is basically 369 00:18:58.960 --> 00:19:01.559 what allows the fancy multiplexing tricks to work at their 370 00:19:01.559 --> 00:19:04.319 full potential. It protects the signal quality for those super 371 00:19:04.400 --> 00:19:07.559 high data rates. The whole system is interconnected. 372 00:19:07.240 --> 00:19:11.279 Totally interconnected. The formation flying enables the high speed links, 373 00:19:11.319 --> 00:19:13.920 and the need for high speed links dictates the type formation. 374 00:19:14.359 --> 00:19:17.200 One can't work without the other. If the formation drifts 375 00:19:17.240 --> 00:19:20.640 too far apart, the beams misalign, the data rate crashes, 376 00:19:20.880 --> 00:19:23.720 and your space data center just isn't a data center anymore. 377 00:19:23.759 --> 00:19:25.079 It's just separate computers again. 378 00:19:25.200 --> 00:19:29.200 Okay, let's switch gears then to the computers themselves, the processors. 379 00:19:29.319 --> 00:19:31.799 You can have all the power in the universe, perfect formation, 380 00:19:32.000 --> 00:19:36.200 amazing laser links. But if the actual chips doing the work, 381 00:19:36.240 --> 00:19:39.319 the TPUs, get fried by space radiation on day one, 382 00:19:39.960 --> 00:19:43.519 then the whole project is obviously useless. And the usual thinking, 383 00:19:43.599 --> 00:19:46.680 right is that regular commercial computer chifts just can't handle 384 00:19:46.680 --> 00:19:49.880 the radiation environment in space. You normally need special rad 385 00:19:49.880 --> 00:19:52.359 hardened components, which are expensive and usually slower. 386 00:19:52.519 --> 00:19:55.359 That has definitely been the big hurdle for using standard 387 00:19:55.400 --> 00:19:57.599 off the shelf hardware in space for a long time. 388 00:19:58.160 --> 00:20:00.799 But what's really striking in the Sunketcher page is what 389 00:20:00.839 --> 00:20:03.880 they found when they tested Googles on chips. It suggests 390 00:20:03.920 --> 00:20:06.759 maybe that old constraint isn't quite as rigid as we thought, 391 00:20:06.920 --> 00:20:10.119 at least for their specific architecture. They took their Trillium 392 00:20:10.359 --> 00:20:14.279 V six E Cloud TPU, that's their custom chip designed 393 00:20:14.279 --> 00:20:17.920 specifically for AI and machine learning, and they blasted it 394 00:20:17.960 --> 00:20:20.480 with radiation to see how it held up, and the 395 00:20:20.519 --> 00:20:23.480 results they're actually much better than even the researchers themselves 396 00:20:23.480 --> 00:20:24.880 seemed to expect better. 397 00:20:24.920 --> 00:20:28.079 In what way, how do you measure holding up against radiation? 398 00:20:28.200 --> 00:20:30.480 We're talking about accumulated damage over time, right, Yeah, and 399 00:20:30.519 --> 00:20:33.599 also maybe sudden failures from like a direct hit bioparticle. 400 00:20:33.680 --> 00:20:36.799 Yeah, both are concerns. The study focused heavily on the 401 00:20:36.799 --> 00:20:40.079