Star Cloud founder defends space-based GPU compute ahead of first H100 satellite launch

first guest of the show. We can come back. Um let's go to let's go to space. Let's go to space data centers. Our first guest of the show is Philip from Star Burning up the timeline. Philip, how are you doing? Good to see you. The current thing himself. The current thing himself.

Uh did this all kick off with Nvidia posting about you? Was that when the kuruffle be I think that was the I think that was the catalyst. Yeah. Okay. So thanks so much for having me on by the way. I've been a huge fan since day one. I am glad to have you. This is why the show exists.

Uh we we we we address every timeline that is in turmoil. Um so break it down for me. When did you start the company? What's the pitch? And and then we'll go into some of the like howers. Yeah. Yeah. But but just give me like the recent back backstory. Yeah.

So we started about a year and a half ago um at the beginning of last year. Um we yeah basically in that time built a satellite. We've got the first launch coming up in a week and that's why Nvidia just posted their first thing because we're going to be launching the first H100 to space.

So about 100 times more powerful GPU compute than has been in space before. Okay. Um and yeah, that's that's the the sort of brief history. Okay. What were what were you doing before this? Uh so I started my career as an engineer for 5 years.

Um studied applied math and theoretic physics undergrad and masters and then I spent a few years at Mackenzie working with the space agencies of the UA and Saudi government. And then I founded and sold another company not related to this and then moved my way into this. Here you are. Harvard, Wharton, Colia.

He's been to every academic institution collecting them all. Um what what is the uh what is the strategy for this first launch, this first satellite? Like what what are you trying to figure out with this launch? What is what is the value of putting this H100 into space?

And then kind of like how do you think about the timeline? cuz I think like the I would say like overarching criticism of data centers in space has been like people I don't think people genu generally believe that this will never happen. It's more of like a debate around like the timeline on which it happens. Yeah.

Yeah. Which is a fair I mean it's a fair debate to have I think. Um so the first satellite the reason we're launching this lots of people believe that you can't run terrestrial data center grade GPUs in space because of high radiation environment and you need to dissipate the heat in a vacuum.

So this is really a demonstrator. It's a 60 kg small sat. Um it's launching on a Falcon 9 a quarter plate there. Um and so we'll be running high we'll do a whole bunch of firsts. We'll be the first to run high power inference in space on imagery from working with various DoD customers. We'll announce that soon.

Um we'll be the first to train a model in space. We're training mini GPT from Copathy. We'll be the first to do um uh we're running a version of Gem Gemini called Gemma which is like a cut down version working with Google Cloud on that.

Um but so we'll do some fun demos like for example you can SSH into it and then just ask it things like where are you now and how does it feel to be a satellite and it will give you a kind of coherent answer. Um [snorts] but yeah this is really really a demo.

Um the next satellite launching in October next year is 10 times the power. We'll have by far the largest commercial radiator in space. Um a 30 m wingspan solar panel and then we'll have very very sophisticated optical and um connectivity options on there.

So customers including uh DoD and other satellites will be able to send us raw imagery. We process it on the edge and then we just download the insight and that's that's the road map. How how far are you?

Like there's a launch that's happening very soon and then there's the renders which are like it looks like sci-fi Star Wars like there oh yes there's like 25 shipping containers that are descending like uh what's the delta there and what was the decision between like kind of showing more of a vision document because I think that's what triggered a lot of people was like you were showing an amount of mass in orbit that feels like beyond the ISS like it felt feel it looks huge.

Oh yeah. Yeah. Right. Yeah. No, to to be fair it's extremely difficult to imagine what's about to come down the line with Starship. So like the launch cap launch cost might come down by between 10 and 100x. Launch capacity is set to go up by a thousandx or more.

And the reason is if you build a new Starship every day for a year, which is like a very conservative estimate for what they're trying to do. At the end of the year, you have 365 starships. It has five times payload capacity. You can launch 10 times as frequently.

So we're really talking about a thousand times like more tonnage per year that we can get to space. So and that's coming like 3 to 5 years which is really not that far away. So the idea of the the render like you can launch the the whole ISS in two starship launches.

So the render is to show look there is going to be a lot more mass in space. Um, but yeah.

I mean, yeah, I almost feel like you should have done the anderal thing where it's like anime if it's like sci-fi where the vision's going and then it's like render for like what you're actually sending because like the render that I saw of like the multiple shipping containers putting together like that's not what's actually launching next week.

So, I think that's where like a lot of the cognitive dissonance is coming from, right? But yeah, I mean you'll iterate all this. It's fine. Yeah. Yeah. I recently saw Palmer talking about that and I was like, man, that's actually pretty cool. Yeah. Yeah.

It it's nice to have because there is a there is a value for like the vision document like what do you where do you want to go in 10 years? Um but it just hits different when it's like the official Nvidia account is sharing this render and it's [laughter] like okay like that feels like it's happening soon.

They It's not like they don't do their research. They have a whole team of like PhDs that go and look at this [ __ ] and like they have they have a board that approves that this can get shown that it's like feasibly possible. Yes. Yeah.

So, so talk about talk about like Yeah, I guess talk about So, so uh this first this first launch is meant to be like a demo basically like correct me if I'm wrong, but effectively a science project to start to show capabilities in space to start to prove that you guys can actually put GPUs in space that can run sure workloads like what is the actual timeline?

So, and then it sounds like maybe defense is like an an early place where there might be some commercial applications, but then what is what's your timeline on on when uh if or when you know I would be you know prompting a model like chat GBT or Gemini or whatever and and uh open running it's just cheaper to do it in space cuz that that that's the long-term vision, right?

Is like it's just cheaper. It's not anything special or like, oh, we need to be there for some geopolitical or defense reason. It's just like the energy is free and the heat dissipation is free, so it's just cheaper, so that's where you go even though the launch costs are expensive. Yes. Yeah.

Um, so maybe I'll start with the second part. So when will it be cheaper? That's coming with Starship. And so we're looking at the first commercial launches of Starship in 18 months, but we're probably going to be seeing, you know, mass volume production of Starship on a 3 to 3 to 5 year time frame.

So yeah, the the road map is next year in October, we're launching our second satellite. That's the one is about a 10 times the compute of the first one, but much more um uh much more capable in terms of continuous power generation and solar panels.

That one is the first commercial soft service that will produce more cash than it costs to design, build, and la and launch. I mean, that satellite pays for itself just on hosted payloads because that satellite has such high um throughput and processing capabilities.

There's a whole bunch of people that just want to stick stick sensors on that. Um it's not the endstate business model, but that's a uh a nice revenue stream in the meantime.

Um, then we're launching a satellite after that which will be um I mean yeah we can't talk too too much about it but you can s you can see this PEZ dispenser form factor that's coming out of Starship. Um Oh yeah. Yeah. Probably shouldn't say too much. And that was successfully tested on the last Starship launch, right?

The where like there was a whole bunch of like debris blowing around. It was a crazy crazy launch, but it did come back. It landed and the PEZ dispenser thing shot out the the fake Starlink or like the the dummy star. Yeah, exactly. So, so that's about a 4 ton at least at least 100 kow satellite.

Second one's 10 kow, first one's 1 kilowatt. So, we're like going up an order of magnitude each time.

That one could feasibly launch end of 2027 and then from there we basically will launch lots of lots of small versions of or lots of continuous versions of that because that will have very low launch cost and it will be a while before we start launching the full Starship payload base size.

I mean for one we need this clam shell door to be ready before we can do that. Um and yeah but to be frank it's quite a while until we start docking things in space.

The only reason you would dock things in space like in the concept video is if you want to train a large model because then you [clears throat] need the whole data center to be internally. Um, and so that we're probably looking at, you know, early early 2030s before we start docking.

I mean, I I I don't have a problem with thinking in decades like that. That's like what we say in tech we want people to do.

And so I don't I I don't have any I don't think there's any problem with saying like, yeah, like there's some crazy stuff that needs to happen with launch costs and like it's going to take a while until this really works out. Like that's actually exactly what people want from from technology founders ideally.

Uh I guess there is the question of like you know how do you frame it is like how you don't want people to Yeah. If you just needed to generate a lot of revenue quickly in space you could do the first space casino where people could beam up tap into it. All ages ages people to hit me up. Oh really? Yeah. Yeah.

I mean space token. You got to find it in space.

what um how do you think about you know some of the uh we we I think we had um we'd asked Dylan Patel at one point about uh GPUs in space and and I don't think he mentioned like the challenges around heat and heat dissipation but he was talking about how sometimes you just need to like unplug the server and plug it back in again or or switch out uh individual GPUs but like what's the solve on that long term?

Are you thinking about, you know, these being like fully robotic systems that allow you to uh make changes on the fly? Like what's the solution there? Yeah, I think probably two points on that.

I mean, the first is for these massive data centers where you've got like, you know, a kilometer long or a mile long row of racks when the people I speak to now are saying, look, if if an H100 breaks in that, you just leave it in there. Nobody's going down there and unplugging and putting that back in.

They're just writing software around it and they'll just leave it basically. Um so that's number one.

But secondly, um for the first few iterations of the satellite, it'll be exactly like Starlink in that we won't be able to touch it and so we'll have to have redundancy on the critical systems and then we overprovision things like solar panels and radiators which degrade over time.

Um over over time the whole industry is or space industry in general is moving towards robotic maintenance but also terrestrial data center industry is moving towards robotic maintenance and that's where on a sort of fiveyear time frame you know some of the people we speak to in data centers are saying we we are fully expecting data centers to be maintained by humanoid robots on a 5year time frame and you know this is where it starts to sound a little bit wacky and sci-fi but um it I mean humanoid robots do not require too much modification to work in space in the sense that a a human space suit would take care of radiation and thermals, you would need to retrain it slightly, but um you know that's maybe sounds a little bit wacky, but robotic.

Yeah, it doesn't have to be a humanoid, right? Like there's a bunch of other form factor. It could be any form factor, but humanoids will be mass-produced, so probably the probably the cheapest. It's interesting. I want you to react to Andrew Mallalip's viral post.

It gets 7,000 likes breaking down a bunch of different things. Uh the core claims I see in here are um questions about what if demand for payload to to orbit doesn't decline. Uh he does talk about uh the difficulties of uh heat dissipation.

He says uh you just finished watching a Scott Manley video on uh radiative heat transfer and now you think you're going to disrupt AWS with a few solar panels in a ride share slot. you're going to believe that right up until the next month when you crack open DWIT. And he goes into a lot of the technical level.

But I was I was wondering if if if there were any key points in here that you think were uh you know easily debunked from actually being in the trenches.

You know, obviously he's he's you know, he does work in in space and does something similar and does seem like an expert who'd be equipped to know this stuff, but uh you've obviously focused on this case specifically. So, what was your reaction when you saw his sort of uh comedic breakdown?

Yeah, I mean I like the post and I mean I'm I'm a fan of Andrew. I've been following his stuff and he's a great engineer. Same with by the way. Very impressive what they're doing. Yeah, totally. Yeah. Yeah.

Um so, I mean on the on the thermals, the criticism usually is in order to dissipate that heat, you need a large surface area. And they think for some reason that that's super impractical. Mhm.

Um I mean my co-founder Ezra has a PhD in engineering spent 10 years designing and building large deployable structures, solar panels and radiators and I mean you just have to build a large surface area and that's what we're doing.

So half our engineering team is building a very large lowcost and low mass deployable radiator. So that is the core IP of our company is that. Um so that's number one on the launch cost side of things.

Um, so, so sometimes people say if there's no competitors to Starship, which it doesn't look like there'll be for at least five years, um, then will there be any incentive for Starship to for SpaceX to drop the pricing? Yep. Now, I have a bit of a different opinion than most people.

Most people say, well, no, they'll just keep it high. I I don't think so. And it's the same reason that they keep that that that teller Tesla is sold at under the price. They could they could potentially sell Tesla at a higher price, but in order to reach a much larger market, you need to sell it at a lower price.

So for example, they're going to have all of these starships sitting there. They can only launch them two years once every two years to Mars. Their marginal cost of launch on each starship is $2 or $3 million. Now either they can leave it there because there's no demand at the, you know, $200 million launch price.

Or they can say to people like us, hey, you can launch it for 20 or $30 million. They still make shades of money. That's still very profitable for us. And so it's like a win-win. So my expectation is that actually the launch price will come down even if they don't have huge a huge number of competitors. Yeah. Yeah.

Like the the elasticity of demand there like they might just be able to sell more volume at lower prices. So they might wind up lowering the cost. Uh or yeah kind of giving you like a bulk discount if you want to put it in that terms. Uh interesting.

Um on the on the heat dissipation uh question, how much do you think that that is uh in the domain of science and we need to have a breakthrough discovery and you need to discover the solution to that problem versus an engineering uh problem where maybe it just comes down to can you make manufacture something that is already follows the laws of physics and all the physicists agree it works if you can just build it profitably.

Yeah, I mean it it's from what we're seeing it's very much an engineering challenge. It's a it's really like a manufacturing. So we know that it works because the International Space Station has a radiator that dissipates 70 kow. The problem is it's ridiculously expensive.

So our radiator on our second satellite is 10 times lower mass per watt of dissipation and 100 times lower price per watt of dissipation than the radiator on the International Space Station. That's the whole the whole game is making this radiator cheap and light.

Um and yeah, we've got very some very innovative or in very unusual manufacturing techniques that we are working on basically. Fun. Uh you posted a few days ago, people don't understand Starlink is going to just be the internet direct to sell is going to hit hard. Basically, all telco is doomed.

Let's check in this time next year. Uh expand expand on that. I um we [laughter] that was at 1:00 a. m.

and I had a very heated argument with somebody and like I was in the cabin and I was like I'm just going to post this and then like 300,000 views later I was like okay this was probably like a little bit too extreme too aggressive.

[laughter] Well no yeah obviously if you want to get attention on Axe just say the most inflammatory thing possible.

uh say that uh you know all these multi do no but but uh what's what's what's your view on how quickly you know we we saw news on Apple exploring something with with uh Starlink on direct to cell seems like every major manufacturer is going to be interested in this and then um there's also that other company that that's like a competitor right that uh what is it that is Verizon propping them up they don't have any satellite iss Yeah.

Yeah. Yeah. No, I don't think they have any hope in hell. And the reason they don't [laughter] Sorry, I know there's a there's a whole bunch of people that got They have a I mean, they have a retail army retail traders. The retail traders are going to come for you, dude. Good luck. What do you know about space?

Well, you got you [laughter] you got the private fine. You got the private markets Varta guys coming for you. Now you got the retail traders at AS coming for you. You got you got no friends in this foxhole, dude. You got to get out of here. ASS is up 7% just today. So that's Oh, wow. Amazing. Well, no. I mean, okay.

I I want them to succeed and I'm I'm like, okay. Well, in that case, I reverse my entire position. [laughter] 7% move. Let's go. That's very funny. Yeah. Uh I I don't know. It's a launch cost thing. It's a launch cost thing. They can't compete with Starlink on the launch cost. So, yeah. Yeah.

Um that argument can be leveled against us, by the way, as well. Well, so so on on what what are the what are the telco companies going to do because they're multiundred billion dollar companies. They're not just going to roll over. Is it is like how do you see them kind of reacting?

I mean it it feels like there's lots of multiundred billion dollar companies that just get backed into a corner and die. I mean Blockbuster hundred billion dollar company. You are such a space maxi.

This is well no I guess you might be right over like you know it just depends on the time frame but like I mean it just doesn't seem like it just doesn't seem like like internet back hall is not going to move to Starlink in the next like 5 years you know this is like not happens for sure but the thing people are underappreciating is the amount of capacity they can launch per starship is absolutely mind-blowing and that is coming in 18 months.

So like and that's like proper hardcore, you know, very high bandwidth director cell in buildings and that's appreciate buildings. It's pretty crazy. Yeah. Yeah. Yeah. No, I mean I guess a lot of Starlink generally they the telecom companies will could end up being like Starlink resellers. Yeah. Right.

Is that is that one scenario? I mean they could be. Yeah. doesn't sound like the best business to be in, but yeah. Anyway, [clears throat] uh thank you so much for hopping on and defending your honor. I think you did a great job defending your honor. I think you're thinking in decades.

I'm excited for you to continue working on this. How how you're you're in the Washington area, right? Yes. Washington State. How how's the what's the scene like in Redmond before you go? Good. How's how's the energy? I mean, there's a lot of it.

It's where Kyper and Stalin castle all of the people like I think 95% of satellites launched in the last two years were designed in Milton and Redmond.

So there's a lot of space talent out there but uh yeah it's not the liveliest of cities in the world but [laughter] well you guys you guys have work to do so thank you thank you so much for joining fun conversation and we're we're rooting for you. I really appreciate it. We'll talk. Let me tell you about linear.

Linear is a purpose-built tool for planning and building products. Meet the system for modern software development. streamline issues, projects, and product road maps you can start building. Uh, do you want to go into bub talk? Are you still interested in talking bubbles or do you want to move on to something else?

We can move on to something else. There's there's a bunch of bub talk in here. Uh, [laughter] maybe more importantly, the makeup. We need to hammer bub talk more. I mean, there's been a lot of bubble talk lately, but I feel like we've hit it a bunch. There's a bunch of charts. Yeah.

If you're trying to figure out if we're in a bubble or not, I recommend going outside, picking a daisy. Mhm. And just picking the leaf, the picking the pedals off, it's a bubble. You like where we are right now. It's a bubble. It's not bubble. It's not.

And then whatever you end up landing on, that's what's going to happen. So, good luck. Um, in Oreo World, Mandles, the owner of Oreo, has trained their own video model for television advertising. They invested $40 million, and they say it cuts production costs by 30 to 50%.

I don't know how that's possible that that like it should cut costs by 99%.

Like do you know I think I still but I still think there's a huge amount of budget goes to coming up with good ideas and then the the like editing and then making infinite variations of it which I'm sure AI will will uh can already do well in some cases.

So, um, but I still think you pay you pay people to come up with great ideas and then execute, but the the idea is is like the actually the highest margin part, right? And so the ad some of the ad agencies I think will do well.

Um, apparently the Oreo's parent company only spent 100 million on digital print and national TV advertising in the last year. Wait, so they spent a hundred million on advertising and 40 million on this? [laughter] Okay, Tyler, what's your take on Oreo Net?

So, this was my idea when we were talking about Sora where I said that Coke should or Coca-Cola should release a video model open source um where you can do anything, but every single video has a Coke. Oh, yeah. Yeah. Yeah.

So, so they should release this model, make it open source, but there's just Oreos everywhere in the video. You just like can't do anything else. You ask it to like make a picture of your dog and your dog just is an Oreo for a head or something like that. It just gets crazy. Yeah.

I yeah that that would be that's a brilliant idea that would be abused so badly that it would be shut down with does this 24 hours I mean does this math make any sense like where did the 40 million go like what were they were they fine uh what is it like that 40 million is ours now that is ours now was like I I really want to know I really want to know the breakdown of that was it actually like running the fine-tuning was it generating you know training data like what was the actual project that got them there?

I mean, I I I could imagine, you know, uh it was it was uh somebody took the $40 million and they made an app layer on top of Sora, too. Yeah.

Also, if they have a whole bunch of footage of Oreos throughout the 30 years that they've been advertising and doing video ads, maybe 50 years, uh and they load those in, can they count that as like depreciation or something?

It's hard not to see this being like a huge waste of money that somebody somebody high up said like I'm going to be the AI guy at Mandelas, right? Because you have to imagine whatever they were able to do today by investing $40 million 40% of the of their ad.

So again, like production cost, which is where they're saving money, is not advertising budget. I would I would look at the way they're talking about their ad budget as like what they spent on ads plus the production cost. Like that's the ad budget. It's very possible that the production cost was like 10% of that. Yeah.

So you spent 40% of your overall advertising budget to slash 10 potentially 10% of your budget by 30 to 50%. Meaning that the payback Yeah. I mean it's a it's a it's not the craziest thing. I don't know. Yeah. But but the models are gonna the models will just be oneshotting.

I promise you they will be oneshotting people eating Oreos better than your model your custom model by now. Like by the time they finish by the time they finish it's it's like they they get on sore too and they do it and they're like oh no why did we finetune like some you know oh they did it with Accenture.

I called it that way they did it [laughter] with Accenture. I knew I knew I knew it. I I knew this was just like Yeah. like big consulting, you know, big strategy firm says, "Yeah, we're we're happy to give you if you want it. What's your but okay? $40 million. We're happy to give you an AI play here.

You're going to be you might just be the open AI of food. [laughter] You might become we can help you be the anthropic of cookies. " Of cookies and then and then it's just immediately a write-off. It is so crazy that they created their own model. How did this leak? It's [laughter] so bad.

I think No, I think they're bragging, John. They're bragging. They're bragging. Mandelas is doing It's like, why are they telling on themselves? They're not, John. They're bragging. Wait.

Yeah, they said the Yeah, the head of the glo the global senior vice president of consumer experience said they're going to use it in the 2027 Super Bowl. I don't know. Is there any reason why you'd want a three-year-old cartoon tune? You could have just run eight Super Bowl ads. You [laughter] could have run eight.

That is actually crazy. That's actually an under so inspiring. If you want to make money in the AI boom, go and get $40 million for a project like this. It's genius. Like, you just made epic consulting moment. Uh yeah, Coupe says, "Sounds bubbly to me. " To totally agree. Nobody has applied. So So the friend.

com billboard strategy is potentially a new strategy to just like buy so much advertising that people can't stop talking about it. Um why has nobody done that for the Super Bowl?

Like has anybody bought like if you really wanted to break through with the Super Bowl you would buy 10 ads or you would buy try to I mean friend tried to buy Can you imagine doing a whole Super Bowl buyout? [laughter] It's just it's just four. How many minutes of Super Bowl ads are there?

Could you do 40 minutes of ads? The um we drove we we we can't get away from the from obvious billboards. We drove by today this one that just says uh when you're on the street it just says end. com. Um is it the end?

Yeah, I would I would have I I definitely think that Oreos just being like we want to be drilled into the mind of every American. We're going to buy 20 Super Bowl ads this year. That's our budget for the year. Yeah.

I mean, there is like some sort of bullcase where it's like you need to generate a whole bunch of Oreo creative across Facebook and and you know, Instagram ads and whatever. But at the same time, it's like I don't know why you need a custom model. the the $40 million custom model.

You have to assume that every model is training on every Oreo ad ever in history. Do Am I crazy here? Is there any reason why you'd want to fine-tune for $40 million? I would I would go farther. I would train a language model. Oh, yes. From the ground up. Yeah, cuz then you have the Do a full pre-train. Do a pre-train.

Do a pre-train. Do a full base model. Y you know, what's it going to score on committee's last exam? What's it going to score on? and and let's let's assume I'm a Fortune 500 buyer and you're an Accenture consultant. How much is that going to cost me? Tyler, uh couple billion.

I mean, and you're telling me that that's worth it, right? Of course. Yeah. We got to get this guy a job at Accenture. [laughter] He's going to he's going to 100x revenues. Yeah.

I mean this just uh to close it out, this is Publicist Group, French multinational advertising company, one of the largest advertising agency conglomerates in the world.

Uh and Accenture serving up a farm-to-table AI strategy for an executive at Oreo who wants to be able to say we don't we're not just using generative AI, we are generative AI. We are the future of food. I I hope they uh they need to vertically integrate. They need to vertically integrate.

They need their own data center. They need their own chips. They need to bake this Oreo algorithm into an ASIC. They need to go to Broadcom. They need to go to TSMC. They need line time at TSMC because I I mean it's possible that they need to go to ASML and get an entirely new lithography machine.

They need to get to rare earth. They [laughter] need to get into rare earth. They need to get into rare earth. They need to go all the way down to maybe acquire a desert just to get enough sand. They need to teach sand to think. They need to teach sand to make Oreo hats. That's what they need to do.

They need to acquire sand. Best of luck out there, Oreo. We're luck. We're rooting for you. Well, you know what Oreo should do? They want to use AI. They got to use numeralhq. com. Put your sales tax on autopilot. Spend less than five minutes per month on sales tax compliance.

If you're if you're selling Oreos online, you got to get numeralhq. com. Uh Lulu says it's the year 2025. You wake up and a progressive white billionaire with a blackface NFT profile picture is saying GM to you, saying wag me, friend. This is a crazy choice. Why did he pick this one?

This is like the how do you do fellow kids meme with the [laughter] It's It's such a crazy choice on a million different levels. Even the smoking, even the NFT. Uh Jay, maybe he's maybe he maybe Reed secretly likes heaters. Jay, that would be sick, I guess.

Uh Jay said this this GM is proof that LinkedIn is four years behind on every news cycle. Uh pretty pretty remarkable. How did the How is this I think the crypto community loved it to be honest. Really? I think I think they're pretty fired up.

If you're one of the 10,000 people that own Crypto Punk, a punk, this is a How are the punks doing? Are they still