It's spouting Shakespeare right now. [laughter] We trained it on the complete works of Shakespeare. It's surprisingly coherent. Actually, the first thing Gemini said was, "Greetings, Earthlings." [laughter]
Stop it. That's brilliant. Thanks for joining.
Thanks so much. Yeah, really appreciate you having me on as well. So, yeah, our company's called Starcloud. We're building orbital data centers initially to provide cloud and edge compute to other satellites and then later to compete with terrestrial data centers on energy costs, essentially. So, I mean, and really we're running up against these huge constraints on energy and water from terrestrial data centers. You know, I think lots of towns and cities are now rejecting huge data center proposals in their neighborhoods for the reasons of they don't want strain on the power grid and on the water supply. And so, yeah, what we're doing is solving that constraint, essentially. We started about two years ago. My background is I spent the first five years of my career on the engineering side and software. Before that, I studied applied math and theoretical physics undergrad and masters, and then spent a few years with McKinsey working with the space agencies of different governments. My co-founder Ezra, who's our CTO, has a PhD in engineering from Imperial and then spent a decade designing and building satellites, most recently with Airbus Defense and Space, where he designed NASA's Lunar Pathfinder deployables. And then my third co-founder, Adi, was previously at SpaceX as a principal software engineer, and he was responsible for tracking beams which enable Starlink for other spacecraft. And then before that, he was 20 years on the data center side with Microsoft, installing the sort of first large GPU clusters in collaboration with Nvidia and OpenAI. So, yeah, we started on this two years ago. We started initially looking at space-based solar and then realized if there was a cheap way of getting the energy consumption part to space — and that is data centers for most new energy projects on Earth — then you could solve one of the biggest problems with space-based solar, where you have huge solar panels and you try and beam that power down, but you lose most of the power in transmission. Whereas if you can move that data center to space, then you know you can consume almost all of that power. We've just launched our first spacecraft. It has the first Nvidia H100 GPU or AI chip in it. It's about 100 times more powerful AI compute than has been in space before. And we've done a whole bunch of firsts on that. So we've trained the first LLM. We've done the first running the first version of Gemini in space on that spacecraft.
And we'll be soon doing running high-powered inference on SAR data in collaboration with various DoD entities.
I was going to ask you about the health status of the GPU that's up there in orbit. So, it sounds like you've now just started to kind of run those models and test it because it's the first time. And just to explain to people about radiation — I mean, no one's ever run a GPU with that level of radiation before, right? And it's just like a colossal amount more than it would ever experience down here on Earth.
Yeah. Well, people have run much smaller GPUs like the Jetson from Nvidia, but nobody's — yeah, as I say, the one we're running is about 100 times more powerful than anything that's been in orbit. So, it's the first ever data center grade GPU that has been operated in space before. And yeah, there's just an enormous amount. It's not at all ruggedized. So, we've had to completely ruggedize it ourselves in terms of new heat sink, you know, switched out the power supply casing, DC to AC converter. We switched out — yeah, AC to DC and...
Fantastic. Well, so we'll get on to the technology. I just want to come back to the origin story of Starcloud. I mean, as soon as I saw the headlines that you were sending up the GPU into orbit, I was kind of immediately like, what is this? Because I hadn't heard of you guys and I was like, how can this company have escaped my knowledge? And it's because you're so new. And so what I love, Philip, was — and there's so many founders, and I'm a founder myself, we all just love taking on these massive challenges, but what you guys have achieved is incredible. And so I was watching you on YouTube over there in your kind of office, kind of workshop, isn't it? And I found one quote from you that I absolutely love and it's this. You said, "Anything that's worth doing is going to be hard. If something is too easy, it won't have the same potential outcome. So, we decided to do the biggest, most ambitious thing we could, and that's build orbital data centers in space." And I just — great quote, but just talk a little bit about the origin story, but also the mindset about how you guys have got this done in such a short amount of time.
Yeah. So the origin story, it started actually about three years ago. So, I went down to Starbase, Texas, where they're building the Starship program. It's this new launch vehicle that is designed to lower the launch cost by between 10 and 100x. And to be honest, it wasn't a work trip or anything. I'm just a huge space nerd. So, it sounded like a fun weekend sort of thing.
And but when I was down there, what really struck me was not just the fact that this new launch vehicle is going to lower the cost. What was really interesting is the capacity is going to go up by an enormous amount, maybe a thousand times or more. And the reason is they're building these two Starship gigafactories. They're the most enormous buildings. I think they're the largest buildings in the world. And it's designed to produce three Starships a day. And I mean, it's just mind-blowing capacity because you can — yeah, it's crazy. It's absolutely crazy because they're reusable. If with the current launch vehicle, Falcon 9, if you build a new one every day for a year, at the end of the year, you still only have one Falcon 9 because it's expendable. Whereas, if you build a new Starship every day for a year, at the end of the year, you have 365. And they have five times the payload capacity of the Falcon 9. They can fly 10 times as frequently. So, it's just completely mind-boggling what's to happen with the capacity. So, that got me thinking about this concept from, you know, in the 1940s, Isaac Asimov was writing about space-based solar, which is where you have these huge solar panels and you beam the power down. The main conclusion we came to was, well, most new energy projects on Earth are being used for data centers. And so either directly or indirectly it looks like that power is going to be contributing to data centers somehow. So then, yeah, that's how we came to the idea of space-based data centers. But I would credit the speed we're moving mainly to my co-founders, to be frank, because they have basically built by hand that first spacecraft and we're still only a team of 12 engineers. People are often shocked that we did this all with — I mean, the first spacecraft we did with three engineers, maybe three or four, and that's just because they have the experience already. You know, Ezra's been building satellites for 10 years.
But it's also sort of a mindset thing. We are very deliberately staying aggressively lean and engineering focused as long as possible. At the moment, I'm the only sort of product person, product and commercial type person in the company. The rest are all engineering from SpaceX, Kuiper, AWS, Azure — like proper hardcore engineers.
Yeah. So you just kept it lean. You got focused on the problem statement and then it seemed like you just solved stuff in house. There just didn't seem to be this like, oh, we've run into a problem. You just test it, fix it, come up with the solution right there and then in house. Maybe not right there and then, but certainly in house.
So you've got the GPU up into orbit right now. You mentioned that you've already been running some models on there. I think you mentioned Gemini. What's the goal for you, Philip, with this first launch?
Yeah, I mean, just the very first goal is proving that we can actually run this H100 in space. So people were worried about two things. Lots of people actually thought for a long time you just couldn't do it on a chip that powerful in space. So number one was the heat dissipation and number two was the radiation. And so in order to solve those, we've come up with a pretty innovative cooling solution for the chips I mentioned that involved replacing the heat sink on the H100. And then we actually submerged the entire motherboard, chip, power subsystems, and everything else in this phase change material that melts. So it's pretty — I don't think anybody's ever tried to do anything like that before. Then for the second one, yeah, and the other thing on radiation is that just requires a lot of testing. So we put a whole bunch of H100s in a cyclotron for proton beam particle accelerator. And then for heavy ions, we went down to Brookhaven National Lab because they have a particle accelerator there for heavy ions. So I think we're the only people in the world that know where an H100 will fail if you blast it with these high velocity protons and heavy ions now, and that informs then our choice of shielding so we can have a lower mass on the shielding side of things. So, yeah.
So you've ruggedized it, essentially. I want to come on to heat dissipation because you've also said actually we are a radiator in space company, was I think the term you used, and that sounds like an incredible challenge that you're focused on. But going back to the current deployment, so what have you managed to achieve? What are the current results of that H100?
Yeah. Well, we've run, as I said, the first training run of an LLM. We actually trained Andrej Karpathy's nanoGPT model, if you're familiar with that.
We did — so it's actually spouting Shakespeare right now. [laughter] So we trained on the complete works of Shakespeare. [laughter] It's a model that predicts not just the next token, but it's predicting the next letter. It's surprisingly coherent, actually. But I mean...
And then, yeah, we've run this version of Gemini. The first thing Gemini said was, "Greetings, Earthlings." [laughter]
Stop it. That's brilliant.
We've got — yeah, we've got the best out of it. [laughter] So at least we know it's got a sense of humor, so that's good. [laughter] I don't know what the macro prompt behind that was, but somebody did well there.
No. Well, and I guess — well, let's talk use cases because you've mentioned a couple of really interesting use cases and what you're looking to do once you move this from test into commercial. I think you might have your first commercial launch next year. So let me get your perspective on that. But for, you know, what's the purpose here initially on the commercial side of this, Philip, before you get to the data centers in space piece? What, yeah, you know, how to pay the mortgage along the way, you know.
