00:00100 gigawatts. That's the annual AI compute capacity SpaceX says it ultimately wants to
00:06deploy in orbit, a lofty goal made no easier by the challenges of building and maintaining
00:11networks of thousands, even millions of these specialized spacecraft in a harsh, harsh orbital
00:19environment. What does a space-based data center look like? Bloomberg did a mock-up of how these
00:24craft would work. So this is the body inside radiation-tolerant AI chips right in here. Around
00:31them, networking layer, power, thermal management, flight control systems. Then for communications,
00:38the satellites could rely on laser links around here rather than your traditional radio frequencies.
00:44That's kind of the squiggly line beaming back down to Earth. Handling that enormous flow of data to
00:49power these systems. Absolutely massive solar array panels as well as batteries to store that energy
00:57when the sun's not there, right, when it's blocked. This up here is the radiator. On Earth, data centers
01:03use air or they use water for cooling. In space, radiators disperse the heat into deep space and
01:10they're going to need to handle far more heat than today's current designs. Out of this world? Yes.
01:16Our next guest has already started down this complicated road. Philip Johnson is the CEO
01:21and co-founder of StarCloud, which launched the first NVIDIA H100 chip aboard its satellite this past
01:29November. I put myself on the spot here. You tell me, what was my explanation of the basics of orbital
01:36data center like? Did we get it right? Excellent. Excellent explanation.
01:41So for you, this is real, right? An H100 inside the form factor of a satellite. Explain how you got
01:51to that point, the engineering challenge, the deployment.
01:56Yeah, for sure. So we, as you rightly mentioned, we launched the first NVIDIA H100 in November last year.
02:03Actually, we launched a satellite which has five GPUs, three from NVIDIA, three from ARM, but the H100 was the
02:09most
02:10important and interesting one. And the way we got there is we started the company about two and a half
02:16years ago.
02:17And we had a very quick development cycle for that first spacecraft. And I think on the screen here,
02:22actually, you can see one of our early renders of a very large four kilometer by four kilometer solar panel
02:27with a five gigawatt cluster in the middle. And that would be for something like training.
02:31We're more going to be doing things like, right now, we're very focused on these small inference nodes.
02:37So we'll be launching, we just filed with the FCC for a constellation of 88,000 of them,
02:42which enables us to deploy about 20 gigawatts of compute.
02:45And, Philip, we're zooming out now. This is a case study of a five gigawatt data center.
02:51Can you just reiterate what you said about the scale of the solar arrays? What were the numbers on that?
02:57Yeah. So this solar array on the screen is about four kilometers by four kilometers with about a one
03:01kilometer by four kilometer radiator down the back there. I would say that that renders from a couple
03:08years ago when most AI workloads were for training. And it seemed a more, you know, it seemed we wanted
03:14to show that if you wanted to do any kind of workload, you could. And so we showed the hardest
03:18thing.
03:18You know, actually, now it looks like 99% of all AI workloads will very soon be inference.
03:23And so we don't need to dock together a large structure like that. And that's
03:26the reason we're now doing this distributed constellation of much smaller satellites.
03:31So, you know, Philip, I want to be honest with the Bloomberg Tech audience. I invited you onto the
03:35program because this is SpaceX's pitch, right? A vast network of orbital data centers. And I guess
03:45that we, in SpaceX's case, we're still waiting for the designs. And Elon Musk has said that,
03:50you know, in the coming weeks, he will show his hand on that. You're a company that is
03:54already doing this. Yes. At right now, a limited and small scale. I just wondered if you'd account
04:00for that, you know, the idea that SpaceX plans to do this and, you know, how StarCloud will fit in
04:07with that. Yeah, for sure. I mean, I think we really are talking about potentially the largest market
04:15opportunity ever. We're talking about trillions of dollars per year of CapEx will be deployed in
04:19space. And so everybody's going to have to have a solution for space compute. Some of that will be
04:24using SpaceX, but, you know, certainly lots of other people are going to want more independent
04:28clouds. And so we, that's the customer base we're going after. Initially also we'll be serving workloads
04:35to, you know, uh, other spacecraft providing edge and cloud services for other spacecraft.
04:42What about, um, the deployment of your satellites? Are you a customer of SpaceX on the ride share
04:48program or how does that work? Yes, we are a very happy customer of SpaceX on both the ride share
04:54program and, uh, potentially for dedicated Falcon 9 launches. Um, so yeah, don't want to speak too soon,
05:01but, uh, we should have some interesting things to say about that soon. You come back on the show
05:05when you're ready. Let's end here. Um, what's the ripple effect of this SpaceX IPO going to be for
05:11you, right? Is there going to be a trickle down? Yeah, we're, we're already seeing it. Um, I mean,
05:17there's just enormous interest in the space industry that, that wasn't there before. Um, so we raised
05:22actually the fastest, uh, unicorn round coming out of YC ever. So in 17 months, we went from basically
05:28zero to $1.1 billion valuation with $170 million raise. I think a lot of that is fueled by people
05:34now understanding space, realizing there's a huge market opportunity there. And, and,
05:39you know, that's in large part because of the SpaceX, uh, IPO.
Comments