00:32:11 · Philip Johnston
[00:32:11] S1: Yeah. So the whole thing is actually contingent on the launch cost. So if the launch cost gets below around $500 a kilo, then it will be much cheaper to do this in space than terrestrially. And I think a sort of good comparison for this is if you build a new solar project on earth to power a data center, because solar is kind of the cheapest form of energy we have terrestrially, and so the way you can think about it is solar terrestrially has these three big costs. The biggest cost is the cost of permit to land, depending where you are. Second biggest cost is the cost of battery storage and backup power, and then the last cost is the cost of the solar cells themselves. And so in space number one, don't need permitted land, so your biggest cost is gone. We don't need batteries and backup power because they're 20 fourseven in the sun, and then we need eight times less solar cells because one square meter of solar panel in space produces about eight times the energy of one in space. So we have all these savings in space. The main additional cost we have is the launch cost, and so that's where this trade off comes, and that's where we see when you run those numbers, you come out to about $500 a kilo, where it makes more sense to do it space than terrestrially. And, of course, right now with Starship, we're at about $5,000 a kilo. Sorry, with Falcon nine, The design of Starship is designed to produce launch costs of, internal launch costs for SpaceX of sort of 20 to 20 to $40 a kilo. So so long as they pass that launch cost through to us at anything less than, say, 10 times their marginal launch cost or even more. That makes sense for us. Right now, by the way, on Falcon nine, they charge two or three times their marginal launch cost, so it seems reasonable.
00:33:48 · Molly O'Shea, Philip Johnston
[00:33:48] S0: Yeah. Thing that I listened to... So I listened to a handful of your interviews. One of the other very concerning points is chips and chip failures. So how does that work into the model? [00:34:00] S1: Yeah. For sure. It's actually a super, like, relevant point. I think it's actually more relevant and pertinent than even cooling, which is people think is the biggest challenge. We've basically solved the cooling challenge. The other part of it that we're working on is the radiation shielding and testing all the chips in different particle accelerators to make sure they can survive the radiation environment in space. And the reason we put so much time into that is if the chips have, you know, like a 10 or 20% higher failure rate in space than they do on the ground, then the whole economics of the model fall apart, and the reason is the chips are by far the most expensive part of what we're doing actually, And so there's a few ways that you can go around that. I mean, we're not going to have access to the satellites once they launch, so they operate a bit like the way the Starlink satellites work, which is you need to have redundancy on critical systems, you know, things like pumps, and then you over provision things which fail over time like the solar panels, you're going to lose a few percent per year, so you slightly over provision that. And then on the chips, we have a longer burn in period on the ground. Most chips actually fail in the first period of their life and so if you can, put them through a more rigorous burn [00:35:09] S0: Why is that?
00:35:09 · Molly O'Shea, Philip Johnston, Ezra Feilden
[00:35:09] S0: Why is that? [00:35:10] S1: I don't know. It's basically the inform... They call it infant mortality of the chips. Oh, no. Yeah. Oh my god. I know. It sounds kinda terrible. [00:35:18] S0: That's horrible. [00:35:20] S1: So what you want is to get through the infancy period and find out all those failures. So for Starcloud two, we had a three month burn in period. It basically means you run the chips hot for three months so you can find all those failures early. It's basically like, yeah, if you can survive... If the chips can survive three months at full power, the chance they have a failure from that point onwards diminishes very dramatically versus the chance they have a failure in the first three months. [00:35:46] S0: So you have five chips up in space right now, including one, H100. [00:35:51] S2: Yes. [00:35:51] S0: What are they doing?
00:35:53 · Philip Johnston
[00:35:53] S1: So, actually, we have... Yes. We have three chips from NVIDIA, two from ARM, and we also have some CPUs from AMD on Starcloud one. And what they're doing is, at the moment, we used them to do some interesting workloads, like train the first model in space, nanoGPT from Andrej Karpathy. We did the first... Ran the first version of Gemini, but now they're actually doing useful workloads for government customers, so they are, mainly in sort of as a test phase, but we're doing inference on satellite imagery, so on synthetic aperture radar data, on hyperspectral data. So we actually just, will do an announcement about some of the results of this, but we've got some very interesting results. And the use case for that is, let's say the government has a bunch of military satellites. They collect enormous amounts of, mostly imagery, but also other types of data. They don't want to have to wait for a ground station to downlink all of that data, so if they can run inference on that at the edge, we can then send down the insight much quicker. And so, for example, if they're monitoring a patch of water, they might want to know is there a vessel crossing some, some boundary. At the moment, what they do is they take an enormous amount of imagery, and then they have to wait at least an hour and a half to get to a ground station, and then they try and downlink as much as they can, but that's usually not very much. Once we're in space, we have a bunch of optical terminals on our satellite. In real time, they can send data to us. We can then run the inference workloads, on our satellite, and then we can downlink the coordinates of a vessel, for example, in minutes rather than hours and days, which is the current way that this works.
00:37:27 · Molly O'Shea, Philip Johnston
[00:37:27] S0: That's wild. Oh my gosh. So our initial use cases for these satellites and what they're doing right now, all the satellite operators and that kind of data, when do you think it'll reach out into other parts of the economy? [00:37:40] S1: Yeah. So it's entirely dependent on the launch cost hitting $500 a kilo. So the business as it is today runs really economically with the current launch price. In order to be competitive with terrestrial data centers on energy and infrastructure cost, that's where we need the launch cost to be $500 a kilo or less. My expectation is that sort of we've already seen Starship deploying its first payloads. We'll see reusability of Starship probably within a year, or at least we'll see the capture of the upper stage. Reflight might take a little bit longer. And then 2028 onwards, we're hoping to see SpaceX selling capacity to commercial customers. You know, If they don't, we also are considering other options and so we've booked on a bunch of different other launch vehicles, not just Falcon nine, but also some others. Think we announced Stoke already, we announced Firefly, and we'll announce, some others soon. [00:38:38] S0: Congrats. Well, if the goal on Earth is to get to a one gigawatt facility, a five gigawatt facility, what is realistic for space and what does that look like in mass?
00:38:38 · Molly O'Shea, Philip Johnston
[00:38:38] S0: Congrats. Well, if the goal on Earth is to get to a one gigawatt facility, a five gigawatt facility, what is realistic for space and what does that look like in mass? [00:38:50] S1: Yeah. So very large data centers on Earth are usually used for training large models, so you need a very highly internally networked cluster to train a frontier model. In space, for a long time, we're probably not going be docking together things, and so most of the workloads we're going to be doing in space is running inference, and that's usually for things like, coding agents and back office business processing agents and chat HBT queries. So the the the road map is we launch a bunch of things like this. So the 200 kilowatt one, we can launch 50 per Starship, so you're talking about close to 10 kilowatts... Sorry, 10 megawatts of new compute capacity per Starship launch, and we've filed with the FCC for a constellation of 88,000 of those. [00:39:36] S0: That's crazy. [00:39:37] S1: Yeah. So... And that's just the first 88,000. So that would... Just the first 88,000. Yeah. Mean, we're not stopping there. So that that enables us to deploy the first 20 gigawatts. Mhmm. We think we can fit about 10 terawatts in this dawn dusk sun synchronous orbit, so that's about 20 times the entire US power grid. And then once we run out of that, we can go to cislunar, MEO, beyond GEO, and to the Lagrange points. Essentially, if you want to scale much more compute, then you really need to do it in space. [00:40:08] S0: So will we have access to sunlight on Earth if we have so many solar panels