Delian Asparouhov and Philip Johnston on making the case for orbital data centers

This is Space Mines from Space News.

This episode of Space Mines is from Space News' recent event on orbital data centers. There, Jeff Foust talks with Founders Fund's Delian Asparouhov about why he's warming to orbital data centers. And in the second half, Jeff chats with Philip Johnston, CEO of Starcloud.

Good afternoon. I'm going to invite Delian Asparouhov. All right. You just got into town, so...

Yes, back-to-back, you know, sort of space summits. I think there were a couple of us that came from New York, from NYC.

Great. Well, thanks for taking the time to join us. You're co-founder and president of Varda Space Industries. You're also a partner in Founders Fund. I thought before we start, just sort of level set the audience on your perspective coming in on this topic of on-orbit compute, orbital data centers.

Um, yeah, let's see. I had an initial, you know, sort of broad set of skepticisms, you know, around this, you know, sort of use case. Partially, you know, at Founders Fund, where I'm a partner, we're an investor in, you know, OpenAI, Anthropic, been investors in SpaceX for a long time, but then also some of the next-gen data center build-out players like Crusoe. And so, we've seen deeply, you know, sort of how the, you know, sort of economics and operations of these types of data centers go, and the sort of, you know, space operator in me had a lot of skepticism around the idea of like, man, these things require just a lot of setup, a lot of build-out. They also require a lot of maintenance. These things do not, like, you know, sort of operate particularly autonomously. And so, the idea of going up into this environment where you now have, you know, sort of more costs around, you know, sort of radiators, you have much more difficulty of access, obviously, you know, sort of much, you know, sort of larger costs on a, you know, sort of per kilo basis. It just did not at all, you know, sort of pencil out to me, you know, from an intuitive basis.

The two things that started to, you know, sort of shift my view were, um, you know, one, our chief engineer at Varda, this guy Andrew McCalip, I think he's probably put together the best, you know, sort of website. Highly recommend, um, you guys check this out. I think it's andrewmccalip.com/datacenters. Uh, but put together this phenomenally, uh, built website that allows you to basically fine-tune all the various parameters on what you need to believe on terrestrial data centers, uh, across various variables like, you know, sort of rent, you know, um, cost for nuclear, solar, battery, whatever you'd like as your, you know, sort of input energy costs, um, and, you know, overall chip costs and maintenance and compares that to space-based data centers and, you know, allows you to put in your inputs on launch costs, solar panel costs, etc. And I think the thing that started to make me realize that there was some potential for this idea was, um, that if you started to, you know, sort of push out the time frame towards end of decade in the 2030s around where launch prices could be, where space solar could be, there started to be some use cases that I think panned out.

The other thing that started to, you know, sort of shift my opinion on this was some of the, you know, both left-wing and right-wing populism that was clearly leading towards data center, you know, sort of moratoriums. Um, you know, I think, um, this may be counterintuitive to the folks in the audience, but it's actually mostly Republican states so far that have, you know, banned, um, you know, sort of data centers, um, from the state. Actually, the only Democratic state that has passed a data center moratorium so far is Maine. And a part of it is people seeing it as, hey, this is these, you know, sort of coastal elite liberal tech companies that are coming into my beautiful pristine state, taking advantage of our, you know, sort of natural resources, our water, our energy. Um, you know, why should we, you know, sort of build these here? And so, at some point if enough of the United States has a moratorium on building data centers, well, maybe space is the only, you know, sort of option. And so, um, I think those two, you know, sort of factors started to, you know, shift my thinking. And then obviously, yeah, yeah, never bet against somebody like Elon in terms of, you know, his, uh, ability to go tackle things that seem impossible and he makes them possible.

We titled the discussion, um, what data centers will be and won't be in 2035. So, let's tackle the first part of that. Looking ahead to 2035, what is sort of the upside for data centers in space?

Um, Elon's clearly going off and tackling the core, um, you know, sort of economic challenges of data centers. He's, uh, establishing Terafab, you know, alongside Intel to start to, you know, sort of build out his own, you know, sort of chips that operate at a higher average operating temperature, uh, which is going to make the, you know, sort of thermal problem in space, you know, sort of much easier. And then the other major input cost in space solar, he, you know, it just got announced, I think, literally yesterday or the day before, where, um, he's going to be establishing the largest North American solar panel, you know, basically, you know, sort of producing, um, line. When I think about, you know, what that looks like between now and 2035, I think more likely than not, a lot of these, um, you know, sort of innovations and pushes on performance largely end up affecting the core Starlink, you know, sort of business. Um, you know, right now, Starlink is still largely a, you know, sort of enterprise, maritime, and aviation, you know, sort of base business. They've struggled a bit with, um, you know, sort of Starlink direct to cellular. There was a, you know, article that came out, um, two days ago, talking about their partnership with T-Mobile, where they expected sort of, you know, hundreds of millions of revenue, you know, sort of flowing through that line of business by this year. Um, it appears to, you know, not necessarily, you know, sort of be the case.

But when I think about what do the early days of, um, their aggressive performance push on space data centers, um, affect, um, I think what it will affect is that the economic case on direct to cellular closes much sooner rather than later for a much broader set of Americans, where they're able to offer it at a price point and at a bandwidth that makes it effectively, you know, sort of a no-brainer. Probably also gives him some leverage on approaching Apple and eventually getting a phased array antenna in one of these things. And so, now all of a sudden, the, you know, sort of power that you need from one of these to, you know, get into the Starlink network goes much lower. So, I think in 2035, you may have the very early innings, basically, of, you know, sort of, uh, data centers that maybe are doing some of what the previous panel talked about, some of this edge compute, edge inference, you know, sort of edge training. Um, is it a place where, you know, a material, call it like north of 10% of like, you know, AI compute is, you know, headed towards? I think that's probably pretty hard to bet against or to bet, you know, for, but are there going to be a heck of a lot more very powerful, very low-cost Starlink satellites that are providing internet to a much broader, you know, sort of set of the, you know, sort of nation in a more retail perspective? Yeah, I think that's the case. So, that's, you know, probably where I think data centers end up.

Do you see certain obvious early applications for orbital data centers, certain markets, certain types of uses that would be best suited to data centers that are in space versus the ground?

Yeah, I mean, there's definitely some amount of compute that happens in space today, right? You know, in environments like orbital factories, we definitely do some amount of, you know, sort of compute there. And as you start to scale those up, there is going to be more and more, you know, basically on-board, you know, sort of instrumentation to analyze what's going on in the process. And we would ideally want to analyze that and then, you know, sort of autonomously, you know, sort of adjust and do that all entirely in orbit. Earth observation, obviously, has a whole set of application areas where the amount of data that is, you know, sort of captured and analyzed is a, you know, sort of tiny amount relative to what's actually communicated down to the ground. If you have, you know, sort of on-orbit compute, you know, sort of for that, that allows you to analyze likely a much broader set of the data that you capture and then, you know, send down, you know, sort of far, far more, you know, sort of data for it. So, those types of edge compute, you know, sort of applications, I think have a decent amount of use cases.

Yeah, as you start to, you know, sort of imagine that there's a much broader, you know, sort of lunar economy, the robots that are on the surface of the moon, the, you know, sort of data that's being generated there, again, more likely than not that the space data centers is the right place to do that. I do think there's going to be a lot of these like edge applications. When I think about the core, you know, OpenAI, Anthropic, etc., training their like, you know, sort of next generation model in 2035, which at that point, you know, we've got probably AGI times 10, more likely than not, I think that is still happening, you know, sort of terrestrially. You know, for like inference in remote applications, right? Again, if you, you know, both of you have like Starlink direct to cellular going more broad, but you're doing inference for, you know, your ChatGPT on the edge and, you know, it's just, you know, rather than going up to Starlink and then down to an industrial data center and then back to Starlink and then back again, it may end up being cheaper to actually just have that same Starlink satellite be directly connected to the, you know, sort of orbital data center. And so I think there'll be these like, you know, retail mobile edge inference, you know, sort of use cases that make sense for space-based data centers.

So let's tackle that second half of the title. Where data centers won't be in 2035? Were maybe people too optimistic or underestimating the challenges for putting up orbital data centers?

Yeah, I mean, you know, obviously, you know, I found a small, large shareholders of, you know, SpaceX. We're huge fans of, you know, sort of Elon. He does have a propensity to put out, you know, sort of schedules and timelines that, you know, are not necessarily fully realistic, um, to say the least.

Elon time.

Yeah, Elon time. Now, I do think it like helps, you know, sort of guide the companies towards like a very ambitious, you know, sort of target that like eventually they get to, right? Uh, but I think if you were to look at, you know, where we think data centers will get to in 2035, I think ultimately looking at the, you know, sort of path of the Starship program is probably the thing to pay, you know, sort of the most attention to. That was a program that really started getting going in like 2017-18 time frame. The original time frame was to have that fully operational by like, you know, sort of mid-24. We're obviously, you know, sort of way past that time frame and there's still a lot of challenges in that program. You know, they've landed the booster a handful of times. They've reused, you know, sort of one of those boosters. The actual ship itself has not yet, you know, sort of landed on land. Obviously has not yet, you know, sort of been reused. Getting that up to an operational cadence. My sort of one-liner that I've used sort of internally as like a mental model for how to think about Starship is it's roughly on the same sort of timeline that, you know, sort of Falcon 9 was, and, you know, it's a much more complex vehicle, but the order of magnitude of increase of complexity is roughly the equivalent order of magnitude of increase of resources that SpaceX has. And so I kind of imagine that it's going to roughly follow all same time frame, right? 2015 is when the Falcon 9, you know, sort of first really, you know, sort of lands for the first time, uh, and 2019 is when I would argue when it fully becomes operational. Um, in, you know, November of 2019, the first ever, you know, sort of Falcon 9 booster launched and landed four times in a row. That to me was clear, you know, sort of operational cadence. That same booster went on to land almost 20 times. Um, you know, Starship has not yet accomplished what I would call that end-to-end equivalent of 2015. We're sort of close. Obviously, you know, you've gotten the stage one reusability, so you could argue that it has, but you kind of need stage two, and so I expect that that'll happen, you know, sort of this year, but then that means that it's probably still another four to five years away from true operational cadence. That means operational cadence in 2030, 2031. Um, and so, as you think about the scale of, you know, sort of what you need to be launching in order to get these data centers off the ground, it's not just the like, you know, hundreds of flights a year, but it's a thousand plus, more launch pads. And so, you know, where will it not be? I think it's not going to be, you know, call it north of 10% of market share of all, you know, sort of orbital, you know, either training or sorry, AI, you know, sort of training or inference. Um, but, yeah, will it at least be at some, you know, reasonable early demos? Will there be some of these like edge, you know, sort of use cases? You know, absolutely. And then in the interim, Starlink is going to become the most monster telecom business that we've ever seen, and SpaceX will probably be a five-plus trillion dollar company at that point.

Where do you see sort of the investment opportunities in orbital data centers? If someone pitched you on a startup now, would you say, 'No, sorry, SpaceX and Blue Origin and Starcloud, they're already here, it's too late,' or you say, 'No, you're too early,' or what?

Um, you know, I think we are in a similar time frame to, you know, call it the mid-1990s with Microsoft, where the like, you know, common phrase that VCs would use at the time is if you're going to, you know, sort of invest in a mouse, make sure you're paying a lot of attention to the elephant and where it's stomping. Um, SpaceX is definitely that, you know, sort of elephant. You know, Blue Origin is maybe like a, you know, antelope, and then, you know, everybody else in the entire market is basically a mouse in comparison. Um, I don't want to be where the, you know, sort of elephant's going to be, you know, sort of stomping, and so going directly up against SpaceX on anything in, you know, sort of comms, in, you know, sort of data centers and anything that they think of as like their core business, you know, lunar or Mars, etc. Not something that I want to be investing into. Things like, as an example, you know, one of our portfolio companies, Northwood, that develops like, you know, sort of ground stations that, you know, sort of an alternative to something that SpaceX has built entirely in-house, but a lot of other companies need but can't afford to build in-house. That I like because I see it as like, you know, sort of fully parallel and, you know, sort of off that path. Within data centers being the users of those, you know, sort of edge data centers. So I definitely think that like we have not yet made like a, you know, a real lunar investment, you know, sort of at Founders Fund, but if there's a lunar investment that is like autonomous lunar ice mines that are heavily dependent on data centers to do the like onboard processing compute so that that ice mine is running 24/7. Yeah, I see that like out of SpaceX's, you know, sort of core, you know, sort of window, um, but clearly something that, you know, sort of has huge benefits from the orbital data centers. Trying to compete with them on space solar and a launching alternative data center, probably not the thing that we want to, you know, sort of plan.

All right. You're probably going to get a lunar ice mine pitch here sometime soon from someone in the audience. Delian, thank you very much.

Thanks so much for having me, Jeff.

So we've got one more discussion here. We have Philip Johnston of Starcloud.

Thanks for joining us. Um, I'll give you an opportunity to start off here with a little bit of the elevator pitch of what Starcloud is doing here in the orbital data center space.

Yeah, I'll talk a minute about that constellation. 88,000 satellites. I was about to say only 88,000 because SpaceX has filed for a million. Then I realized how crazy only 88,000 satellites sounds. But why 88,000? Why not 8,000 or 800,000? How did you get to that particular architecture?

So what's the roadmap to get to that 88,000? You've got one satellite now, you're doing another in eight months. You just raised a significant funding round. Um, what's the path to that 88,000 steps that we talk about in the title of the session?