You're significantly warming up the planet just purely by pumping heat from these data centers out into the atmosphere. And at some point then the only way, if you want more compute, that you can have it is by figuring out ways to dissipate that heat in a vacuum into the infinity of deep space. I actually like to think we're digging for gold, where the gold is the sunlight, and we are harvesting that energy resource in space.
I think Sam Altman spoke a few days back that we inevitably have to move to space, or we do go build them in space. I don't know, it sounds cool to try to build them in space because the current land data centers, the pollution and the amount of heat generated becomes massive.
That is actually a very small portion of the end state problem with building enormous amounts of data centers. Even if just using solar panels, for example, you are significantly warming up the planet just purely by pumping heat from these data centers out into the atmosphere.
So imagine if you're a satellite that is taking enormous amounts of raw imagery. There's many use cases where you would want to get that data back very quickly. Most of those satellites, they'll pass over the target site where they take a photograph and then they have to wait 60 or 90 minutes to get to a ground station before they can actually downlink any of that data. With what we're doing, so in space-to-space, you can use this optical mesh network that there is. We then process it on orbit and then we just downlink the insight.
You see Elon Musk talking about the pursuit of AGI going forward, and you're playing a role in this whole process. What's your personal vision when it comes to AGI?
All right, Philip, welcome to our podcast. I've been doing a bit of research on you and I found that you initially were a pure mathematician and you really loved that.
Before you sort of switched to algorithmic trading, which you in one of your interviews called the job for boring losers, and sort of moved to McKinsey and then scaled the Digital Brands marketplace in Dubai, as I understand. But during your days in algo trading, you figured that it helps to sort of learn the pattern recognition.
So what kind of patterns you recognized at the moment when you decided that the future is in space, in orbit computation?
I'll just clarify because I don't want to lose friends in algo trading. So algo trading as a job I find boring. I have lots of good friends that do algo trading who are not boring people.
So in terms of pattern recognition, or at least what we're doing, is the convergence of three main trends. The first main trend is there is very rapid increase in demand for tokens and for compute. The second is that is driving very rapid increase in demand for energy. And the third is that the launch cost is coming down very rapidly. And what we're doing is at the intersection of these three trends essentially. And I'd say if there was to be any pattern recognition in what we're doing, it's noticing that these three trends are converging and are heading in a certain direction.
And basically Starcloud now raised one of the record numbers in terms of the seed funding among the YC companies, and without even sending any satellite in space. So I'm curious, what was the sort of part of engineering a vision of Starcloud versus engineering an actual satellite in a way that helped you to raise that? What's the vision you sort of convince your investors in?
I think to invest in us you need to believe a few things. Essentially the three trends. If you buy into those three trends, we are a pretty obvious bet. If you don't, then we are a very risky sounding bet. So if you believe, for example, that Starship is going to lower the cost of launch by 10 or 100x, and if you also believe that that is going to drive the launch capacity up by maybe 10,000x, then you're probably looking for things that are going to happen in space that don't currently happen in space. Likewise, if you think that the growing demand for AI is not going to stop growing, it's not going to slow down anytime soon, which I do, then I think we are a pretty obvious bet.
But the balance between believing that narrative and the actual engineering, I think, is a fine one. And people wouldn't invest in us if we hadn't proved to some degree that we are capable engineers. So all of the team is from SpaceX and NASA and Microsoft Azure on the data center side or on the hardware side. And we've also actually designed, built and tested the first satellite. We're just waiting for the launch. So we're scheduled to launch between May and August. It's been pushed back to August. So now we're launching in August. Yeah.
Like you mentioned, the investors who believe that the cost of sending things to space basically will be decreased by 100x. That's a sort of a bet on Elon Musk's vision on Starship.
Do you have like a plan B if it doesn't happen, or happens 5 years later, or it will not be like a target $50 per kilogram, but end up like $500?
Yeah, $500 is about the break-even of where the longer-term vision of competing with terrestrial data centers breaks even. There's two mitigants to that. The first is that there are other launch vehicles coming online: the New Glenn vehicle from Blue Origin, the Neutron rocket from Rocket Lab, Stoke Space is trying to do some stuff. My view is they are quite far behind Starship. So if Starship doesn't work, it will be pretty delayed. But that's where the second mitigation comes in, which is if we are delayed, let's say 3 years or 5 years, the first business we're building is actually only reliant on the Falcon 9 launch prices. So the first service we're providing is an edge compute workload, or like a cloud-on-orbit compute workload for other satellites, particularly for Earth observation and government satellites.
And that's where you want to also sort of propose the new way of solving this bottleneck with downlink. Could you share a little bit more about that?
Yeah. So imagine if you're a satellite that is taking enormous amounts of raw imagery. So for example, you might be in the case of the government, you might be monitoring a battlefield. In the case of commercial, you might be monitoring shipping lanes. And you use that data to then forecast a whole bunch of things. There's many use cases where you would want to get that data back very quickly, particularly if it's a battlefield situation. What happens right now is satellites are flying, they're in low Earth orbit, which means they're orbiting once every 90 minutes. Most of those satellites are not connected to anything for most of their orbit. So they'll pass over the target site where they take a photograph and then they have to wait half an hour or 60 or 90 minutes to get to a ground station before they can actually downlink any of that data. And when they are above a ground station, they're getting very low data rates, usually a few hundred megabit a second data rates.
With what we're doing, they will be able to, you know, there's a constraint from space to Earth. There's less of a constraint from space to space. So in space-to-space, you can use this optical mesh network that Starlink has essentially produced, and so they'll be able to ship enormous amounts of raw imagery to us either through the Starlink network or just directly. We then process it on orbit and then we just downlink the insight. So the insight might be, for example, there is a ship in this location moving in this direction and there's a 90% probability that it's this type of vessel, that you can downlink with less than a kilobyte rather than having to downlink hundreds and hundreds of gigabytes of raw imagery.
Correct me if I'm wrong. So essentially, instead of sending the raw data back to Earth, you do the compute in space and then you send only the output that is required.
Yes. So we run the inference workload on that data, and the inference then produces some insight, and the insight, as I say, could be information about the...
So you mentioned Blue Origin among the companies. I was just wondering if this concept of space compute and clouds in space would be that attractive, what stops, say, Jeff Bezos from building AWS in space?
Not very much, to be honest. No, I mean, yeah, I don't want to speak too quickly for them, but we know they are very keenly looking at space. So I'm not trying to encourage competitors.
Well, maybe it's not something bad because, you know, you can grow the market together, push the market in a way.
Yeah. I mean, you're right though. It's a very good point you raised. If what we're saying is correct, which is my personal opinion is that within 10 years almost all new data centers will be being built in space purely because of the constraints we have on energy terrestrially. So if that statement is correct, there's two data center companies that have a space arm. One is xAI with SpaceX and the other is AWS with Project Kuiper. And so it would probably make sense that those will enter the space. The other data center companies right now do not have a space arm. Google doesn't, Meta doesn't, Microsoft doesn't, Oracle, OpenAI, etc. And so they would then need to partner with somebody who has that capability. And at that point we might become a viable partner for one of those guys.
The sort of current model, I mean, traditionally in space, the compute power is compromised by other characteristics like longevity of this hardware in space because of immense radiation. So you need to invest a lot in sort of hardening all the electronics to be able to survive for long missions of 15 years.
And in this context of Starcloud, as I understand, you're sort of betting on the latest chips from Nvidia, which essentially would become out of date within 5 years of working in space. So you're kind of betting on replacing them sort of like a cartridge in a printer every five years. Is that correct?
Could you tell us, I mean, how viable would this model be in the long term?
It's viable if the launch cost is low. That's the... so everything has a break-even point dependent on the launch cost. So the key unlock in all of this is really the launch cost falling very, very rapidly. I would say though that there's nothing different about what we're doing than how terrestrial data centers work. So some people misunderstand, might be the right word, but some people don't see why if a chip is out of date in 18 months, that means we're going to have a problem with replacing it. And actually, we don't have a problem because it's the same in a terrestrial data center. If you have a chip that lasts 5 years, you put it in a data center, in 18 months a new chip comes out, you don't turn that one off. You keep that one running because you've spent the capex on it already. And then you just start using new chips, and it's exactly the same for us. So yes, there'll be new chips coming out in 18 months, but we're still going to be drawing value from those old chips.
Maybe let's touch a little bit on sort of the radiator in the room. In your white paper, it says that it will require a massive, not massive but rather huge scale of radiators to be able to cool down the chips. So basically we're talking about 4 and 4 km of...
That's the solar panel, and then the radiator is 1x4 km. Yeah.
Yeah. Which in the current industry benchmarks would require you to basically invest a lot in building a new type of materials, more lightweight, and essentially become a sort of space radiator company.
Oh, do you mean radiator or... you need to construct this somehow in space, using robotics. So essentially, does it mean that you will become like a space construction robotics company rather than a data center company?
It really depends how complex the deployment methodology for the radiator is. We're working on some very simple deployment mechanisms. I mean, for example, with the solar panels, one of the simplest mechanisms for deploying very large surface solar panels is just to unroll them. And you don't need a huge amount of... so you can roll these very, very 30 micron thick solar panels in quite a tight package and then unroll them. We're working on a very simple radiator design as well that will not require too much robotic assembly.
But it's still rather huge... still huge, definitely huge... sort of objects. What's your vision? How do you see that making this a reality?
I mean, it's a great question. So half of our engineering team is working on radiator design and deployment. The other half of our engineering team is working on radiation hardening of the chips.
Are you still investing in hardening? Okay.
We're not making new chips, but we are working on ways to make the chips survive the space environment. So yeah, that's both on the hardware side, mainly shielding, and also on the software side. So algorithms that produce redundancy. But yeah, so half the team is working on this radiator. We've got, I mean, the best space radiator engineers in the world. So we've got the guy from NASA, from Jet Propulsion Laboratories. That's the deep space arm of NASA that designed the radiators on the Europa Clipper mission, NASA's most expensive deep space probe. He also designed the radiators on a recent lunar lander and three of the NASA payloads on that. So I think he's maybe the best in the world in terms of designing space radiators. You can almost think of us as a space radiator company. That is the core technology that we're trying to solve for. Yeah.
Yeah, that's interesting. And you also proposed this very smart idea of placing this in a terminator area. Could you explain a little bit more on that? How does it work?
