XAI's Radical Plan: Data Centers in Space  ## Document Notes ## Highlights - **Founders Leave As Labs Scale Up** - Jaeden recounts XAI reorganizing and parting ways with some founders as it scales rapidly. - He compares this to other labs where early builders often leave once companies shift to execution mode. Transcript: Jaeden Schafer I think a lot of the address was kind of focused on a whole bunch of people that have recently just left XAI. There was a bunch of kind of the original team members, a lot of the founders of XAI that left. This is what he said about it. He said, as the company grows, especially as quickly as XAI, the structure must evolve. And this was actually an ex-post he put out. He said, this unfortunately required parting ways with some people. I think like reorganizing inside of these companies are growing very fast is quite common. I think what was interesting to me with all of this is kind of the new structure. So XAI is now splitting into four primary teams. They have Grok, which is the voice, you know, the LLM and the voice that goes with it. And basically all of the core chatbot experiences. They then have the coding system powered software generation team. So that's really kind of being like a cloud code competitor is what they're trying to build. And they actually do have a model that ranks quite well on the benchmarks for software development. It might even be number one at times. But because they don't have all the integrations and custom stuff that Cloud Code has, Cloud Code is still far, you know, as far as usage goes is way ahead. ([Time 0:01:51](https://share.snipd.com/snip/927fcea5-b466-470c-ac1c-bae28fb05a94)) - **XAI Splits Into Four Strategic Teams** - Jaeden Schafer outlines XAI's four teams: Grok, a coding-software team, Imagine, and MacroHard for full task automation. - This structure signals XAI is building beyond chatbots toward broad, enterprise-scale AI automation. Transcript: Jaeden Schafer He said, as the company grows, especially as quickly as XAI, the structure must evolve. And this was actually an ex-post he put out. He said, this unfortunately required parting ways with some people. I think like reorganizing inside of these companies are growing very fast is quite common. I think what was interesting to me with all of this is kind of the new structure. So XAI is now splitting into four primary teams. They have Grok, which is the voice, you know, the LLM and the voice that goes with it. And basically all of the core chatbot experiences. They then have the coding system powered software generation team. So that's really kind of being like a cloud code competitor is what they're trying to build. And they actually do have a model that ranks quite well on the benchmarks for software development. It might even be number one at times. But because they don't have all the integrations and custom stuff that Cloud Code has, Cloud Code is still far, you know, as far as usage goes is way ahead. Then they have the Imagine. This is the video and image generation components of XAI. And then they have MacroHard. ([Time 0:02:03](https://share.snipd.com/snip/924c704b-affe-41ca-8247-6ec3b9d324cc)) - **Generative Media Has Hit Massive Scale** - X's image and video tools reached massive scale, producing tens of millions of videos and billions of images recently. - Jaeden uses this to argue AI is mainstream and not just experimental anymore. Transcript: Jaeden Schafer There's a bunch of executives over there that said that imagine their image generator and video generator is generating tens of millions of videos per day and billions of images over The last month. And that's according to a bunch of internal metrics. That is a lot of usage. I think it shows and kind of reflects a bit of a broader thing, which is basically AI is no longer experimental, right? It's pretty mainstream. And we're seeing this with ChadGPT and a lot of these other players. But it is being pushed out on every platform everywhere. And a lot of people might call this AI slop. And especially with Grok and their image and video generation tools, I think there's a lot of problematic things there. But it's definitely hit scale. Like there is, you know, billions of images being generated. ([Time 0:05:27](https://share.snipd.com/snip/d6aebfed-bb37-4fda-8eb3-a3a3c13beccd)) - **Orbit Could Become Cheaper Compute** - Elon Musk argues orbital data centers could be cheapest due to constant solar energy and fewer land constraints. - Jaeden notes SpaceX filed for regulatory approval to build a massive solar-powered satellite data center constellation. Transcript: Jaeden Schafer So he kind of just reiterated what he has said in the past, that the cheapest place to run large-scale AI within a few years is going to be in orbit. Basically, the benefits are that space has, you know, constant solar energy, fewer land permitting bottlenecks, although, I mean, you still got to get your satellites approved and Stuff. But apparently, that's easier than land permitting on Earth. And then there's also a path to scale compute beyond just terrestrial constraints. So SpaceX has already requested regulatory approval to build solar powered orbital data centers. I think, you know, we've seen that the administration looks like is kind of looking at this favorably and is going to grant this. I think they put out a tweet recently where they were saying, you know, they're requesting comments. If anybody has any objections, speak now or forever hold your peace, basically. And I think this is going to have their application that they put out for was like for a million data centers. So there's going to be a massive satellite constellation. ([Time 0:06:26](https://share.snipd.com/snip/284c1b26-2ea9-4a79-914a-802a0f518306)) - **Moon Manufacturing Could Solve Lifecycle Limits** - Lunar manufacturing could enable on-moon production and retrofitting of orbital data center hardware. - Deorbit rules force lifecycle plans like reentry burn-up or capture and refurbishment on the Moon. Transcript: Jaeden Schafer So literally building like factors on the moon can launch and develop and create these data centers. And I think one of the parts of this is as far as like regulation goes, you have to, I think, deorbit your satellites every five years or have a plan for deorbiting so that we don't just have Unlimited space junk. These things have to be able to basically come down. A lot of people have it where they just reenter the atmosphere and burn up. And that's kind of their, you know, de-orbiting plan. And it feels like we might have perhaps ways to capture them and bring them to the moon to get retrofitted, or perhaps we're still going to let them just burn up on re-entry and just keep, You know, our factory on the moon pumping out more and more solar paneled, solar panel covered space compute data centers. ([Time 0:07:42](https://share.snipd.com/snip/d9677c5e-16b7-4e34-8148-8cd996bf2f94)) - **Starship Is The Economic Linchpin** - The business case hinges on Starship cutting launch costs drastically via high reusability and low per-kilogram prices. - If Starship achieves its goals, orbital compute economics shift dramatically versus current expensive launches. Transcript: Jaeden Schafer If Starship is able to, you know, achieve its intended cost reductions per kilogram to orbit, then I think the business case is going to shift dramatically. But right now, because it is so expensive, even for Starship to get things into orbit, this doesn't feel like a profitable business. You know, he's like, you know, it's way cheaper because there's like no land permitting. And, you know, we don't have to get water to cool these things. And we have unlimited solar. And so those things are true. But just getting these satellites into space is expensive. But Starship has a plan to decrease their payload costs. So being able to get things into space for cheaper, they're planning on doing a lot of optimizations. And if they are able to do those optimizations, perhaps this becomes a profitable business. Today, the Falcon 9 already delivers basically the best industry pricing for getting stuff into space. Starship is trying to push that down by an order of magnitude. There are no other launch vehicles in development. And right now, they're targeting comparable reusability at that scale. So part of basically saving a lot of their costs is how much of the rockets can be reused if they shoot something up and they bring it back down. You know, and that's kind of the thing that they saved a killer of money over NASA was just that the rockets now were not exploding or crashing on reentry. They're actually able to reuse them. So how many times can they reuse them? How many parts can be reused? How much they need to be repaired? If they can bring all those costs down and just have a vehicle that goes up and down to space over and over again, that's where they start saving a lot of money. And they have a goal and a plan for that. ([Time 0:08:55](https://share.snipd.com/snip/2b749aac-40f9-45d6-91c5-0e246339e3b1)) - **Operational Experience Is A Moat** - SpaceX's Starlink experience shows they can mass-produce and operate thousands of satellites at scale. - Jaeden argues that operational expertise gives SpaceX a unique moat for turning satellites into data centers. Transcript: Jaeden Schafer SpaceX has already shown that they can obviously mass-produce and operate thousands of satellites, right? They have the Starlink constellation, which by the way, if you didn't know, Starlink generates 80% of SpaceX's revenue. I like they charge money to shoot things into space. And that's kind of what their company was built on. But Starlink is actually the biggest revenue generator for the company right now. So obviously, like they're able to do it, they can mass produce, they can, you know, mass operate 1000s of satellites, and they're doing a good job of that. And so turning those satellites into data center, so is kind of the next step of can they do it. And if any company can, it would appear, you know, they have the most experience they would be able to, but it's tricky. And I think the one thing they have going for them, there's not a ton of competition has obviously some insane operational experience. And that's a huge moat uh not a lot of people are really competing directly but i will say google's uh project suncatcher is focusing on there's a startup called star cloud and then even Amazon and blue origin um are kind of showing that this orbital ai infrastructure is something that they're interested in so there are other players that are competing google amazon And star cloud And so and then of course, SpaceX, we have like four players in this. ([Time 0:10:40](https://share.snipd.com/snip/50227d83-dba6-4db6-a2f9-3809b0638c05)) - **Start With Inference, Not Training** - Focus initial orbital deployments on inference workloads, not full-scale training, to match current practical limits. - Use many small distributed GPU clusters in satellites to serve customer-facing and generative tasks efficiently. Transcript: Jaeden Schafer So basically early orbital deployments are more likely to focus on inference rather than full scale training, right? So these data centers up there, they're not going to be doing full scale training. They're probably going to be doing inference, which is just when I'm chatting with talking to chat GPT, it has to run inference to basically give me my answer. And that's what they're probably going to do. So inference workloads, they can run on smaller GPU clusters, and they can tolerate distributed architectures a lot more easily, right? And so, you know, there's not like one giant data center in space, it's going to be a ton of these little tiny satellite data centers all linked together. And because they're distributed like that, you have to have unique architectures that can actually make that function. So basically, that opens up some near-term paths. There's like, you know, you can imagine like customer service agents, generative media workloads, AI-assisted software tasks, which could be computed in orbit and then powered By continuous solar panel. That kind of stuff all makes sense. ([Time 0:12:05](https://share.snipd.com/snip/fd84a349-3a99-4f0b-b1c2-3d5d78262737)) - **Shift Workloads Dynamically To Cheapest Layer** - Dynamically allocate workloads between Earth and orbit to use the cheapest available compute and avoid vendor margins. - Prioritize internal orbital capacity when terrestrial servers are overloaded to control costs and latency. Transcript: Jaeden Schafer So they can, they can work their system that way. And alternatively, they would just have to be going with a competitor, which would be upselling up charging them a lot AWS or Google. And so, you know, which have to add their own margins and markup. ([Time 0:13:38](https://share.snipd.com/snip/ee081fe5-a7f0-47ea-8980-1db952ac8e57)) - **Vertical Integration Enables Bold Experiments** - Elon frames AI as a civilizational force multiplier and is combining energy, compute, manufacturing, and distribution vertically. - Jaeden highlights this rare full-stack integration as a strategic advantage few competitors match. Transcript: Jaeden Schafer I think Elon for many years has talked about AI, not just as a product category, but as a he's called it like a civilizational force multiplier. I think right now we're seeing a lot of those pieces align. There's reusable heavy lift rockets, there's massive satellite manufacturing capacity, there's vertically integrated AI research, and then of course he's got a social platform For distribution, but also for data. And like he's got data, he's got AI, he's got manufacturing, like he's got so many pieces, it's hard for me to think of anyone else that could pull this off to the scale that they are and Have all of the pieces to this. So it's obviously a massive vision, but it's also kind of consistent with his experience and what his other companies are able to produce if he kind of merges them together. AI on Earth is already very, you know, this is massive for productivity. So perhaps putting it in space is going to remove the energy and land constraint that's limiting scaling. I mean, those are two of the biggest scale limiters. I've talked to many investors focusing on this kind of stuff. ([Time 0:15:48](https://share.snipd.com/snip/f3360376-b4c8-4bec-9284-991ecd13f2c8))