Ep. 024 - SpaceX's 10GW Plan Drives $300B ARR by 2027 (Datacenter, Energy) | Reyk Knuhtsen, Jeremie Eliahou Ontiveros, Jordan Nanos

Wait, what? You're adjusting the NVIDIA logo to make sure it's on display?

Yeah, man. When you have GPUs at home, you know, you got to show them to the world.

Is that actually a GPU?

It's a decommissioned GPU. You have like all the heatsink, but there's no GPU. If you remove all of that, you'll see there's no GPU.

You're going to start running some local models to stop this token burn that you've been jacking up or what?

Gotta cut costs, man. But they keep telling me it's cheaper on the cloud, you know? So I'm like, okay, my electricity bill can't handle it.

Jeremy, Raik, we're going to talk about SpaceX doing 10 gigawatts in '27. You guys ready? Welcome back, Semi Analysis Weekly. We got Jeremy and Raik. Like I said, we're going to talk about the article we put out that got a lot of traction. SpaceX 10 gigawatt 2027. Why it's real, will drive $300 billion of ARR for SpaceX and why Microsoft will be the largest offtaker. We're going to try and talk through some of Elon's statements on the earnings call, talk through the economics, how we get that $100 million per megawatt per year that they're gonna sell this stuff at. When you pass through the tokens, Microsoft is a potential customer, how they pay for it and how they get enough chips and people. So guys, yeah, welcome. Excited to dig in.

Yeah, let's go.

Yeah.

Thanks for having us, Jordan.

All right, Jeremy, you're first author here. So Elon's statements during the earnings call, conservatively, what does, what does this mean? What's the takeaway when he has SpaceX's first ever earnings call and says that he's got these gigawatt engines?

Machines?

Look, I think it all really starts from the economics. Like, that's really the key thing. When, like, what we've observed over the course of 2026 is gross margins for these labs just kept going up. And that has been the key driver of their ARR acceleration, right? And if you look at them combined today, that is Anthropic and OpenAI, they're both adding combined over $20 billion of ARR per month, actually close to $30 billion now. So by itself, that drives AI revenue adding close to $400 billion per year. And they're doing that by increasing their gross margins and increasing their gross margins. Essentially, that just means for any given amount of compute, which typically is roughly a fixed cost, like we've seen, you know, $12, $13, $14 million megawatt-year from the likes of CoreWeave, increasing their gross margin just means that revenue per watt goes up. And so for us, and that's where, you know, that's what I want to flip it to you is for us, we've done a lot of work on trying to understand like what is the actual revenue per megawatt. And obviously we have InferenceX and in this article, what we put out is that we think like they can fairly easily today, today on API that is OpenAI Anthropic make $100 million per megawatt per year, right? And just tying it back to like, why is Elon trying to build so many gigawatts so fast? Well, it's because no one else is doing it so fast. No one else is really banking on the opportunity of like, okay, for any megawatt that they have, they can make so much money. And so he was sort of the first to realize, hey man, if you want this 3 months from now, you want 300 megawatts right now, all right, pay me $50 per megawatt, you're going to make 50% gross margin, right? And so essentially, he just wants to replicate that playbook and do that at much larger scale. And if the economics are that good, everything downstream is much easier, right? So the first question, you know, we should ask ourselves is, is $100 million per megawatt per year realistic for OpenAI and Anthropic on API, uh, today, right? So I don't know, what do you think of that, Jordan?

Yeah, I mean, I think it is real. So it's definitely realistic. Um, we've, we, we've dug into this in great detail. Let me share this chart so that we can actually like explain the details, but Yeah, when we talk through maybe the different ways in which GPUs transact, you can start at the beginning, like you said, $12 or $13 billion per gigawatt, which is $12 or $13 million per megawatt. And then that's like the 5-year average infrastructure as a service price across, yeah, CoreWeave, Oracle, Nebbius, long-term, close to self-build pricing for these guys. Like maybe they're making double-digit margins on these things, but it's, it's not, massive. And where things change is when you sell for a premium because either you're getting on-demand, which is the green column here, the B300s, or you're doing these deals where SpaceX is selling their existing compute to somebody because they can turn on so much of it immediately. And like you said, that's the reason you can transact at such a significant premium. And specifically this Google deal at like the equivalent of $14 an hour. Is a significant premium over the average, which might land around $3 an hour for a GB300 right now. Right.

I would just like super quickly add that I think an amazing like term in the contract that they have is the 90-day cancellation policy because for Google, for Microsoft, for Anthropic, they have zero risk. They can cancel this if they realize the economics are not working anymore anyway. So it just makes it way easier. It's really like emergency megawatts. You want them right now, big, easy to cancel, you know, that's the price. So I think the price is justified because that service is unique in the world.

Yeah. And so what is the service? How do they demand these margins? It's really selling tokens at the inference API costs, which as we model it and you look specifically from two angles, both the InferenceX data that we have where we actually run real workloads on the latest chips, this is GB200, GB300 with the latest optimizations from vLLM, SG Lang, uh, TRT-LLM from Nvidia, um, Mori and other stuff from AMD, whatever chip you're trying to use. The point is that, uh, we get that real data from open source models. And increasingly these models are approaching the size of the frontier models. K3 from KIMI is a 3 trillion parameter model. DeepSeq is using sparse attention and all these, I mean, KIMI is using all of their sparse attention approaches too, but KIMI linear, but the, um, you know, the point is that these are well over a trillion parameters total. That takes up all the memory space. They have million context windows, which is the same as the frontier models that blows up your KV cache. And they have all these optimizations to do KV offloading. And so you can get a proxy or a lower bound, let's say, of what a 3 trillion or a 2 trillion model is going to perform at. Assume that the frontiers are bigger because they're higher performance in terms of total parameters and active parameters because they have better GPUs that they can serve these things on compared to the open source models, which are, typically Chinese developed on the Chinese SKUs. So this is, you know, not the latest and greatest from Nvidia, or they're increasingly now being developed on, you know, Huawei chips and, and others from China. So, so that's a lower bound of performance is our public data on, on InferenceX. And then the upper bound is really, our simulator, which takes, you know, op by op. So these are like the GEMMs, the collectives, the like all of the actual operations simulated on the current hardware where we test the actual performance of the collectives and the GEMMs and stuff on the real hardware. And then just, you know, trace it through what we think is the shape of the frontier model and then forecast this forward for what we're going to see on Rubin when we see increases in the FLOPs, the memory bandwidth, the networking, you know, bandwidth. All of this is, you know, the power consumption, all of this is going to change when people start deploying Vera Rubin next year, which is going to be all of SpaceX capacity next year. They're not going to be, you know, deploying a bunch of GV300s. They're going to be deploying the latest and greatest. And so you're going to see a performance increase from the GPUs on a relative, like, per-megawatt basis. That's going to allow people to produce more tokens. And like you said, we think they're going to be able to turn on more of these than anybody else in the fastest amount of time. So the forecast to go from SpaceX and Google signing a deal that, you know, gives them GPUs at roughly $48 billion per gigawatt, and then saying there's going to be a 50% gross margin to get to $100 billion. You know, you'd expect somebody like Google to make a financially sound decision to do that. They've got to see returns on that., you know, capital that they're investing in those GPUs, let alone the fact that they, you know, the, the reporting of the leaked financials that we're seeing from the labs right now is showing 85% gross margins. And, and that matches what we are seeing with our simulator and our InferenceX data, which would imply, you know, around $100 billion of selling costs at a cost of around $15 billion per gigawatt, right? Which, which matches this chart that's on screen there in the, in the blue.

And that 85 being a blend, which includes older or less performant GPUs, not just the GB300. So you could argue 100 is actually conservative for GB300 or for VR, in fact. Yeah.

Yeah.

I'll scroll down and show another chart here. You know, there is like really significant differences even as we go from GB200 to GB300. The increase in memory capacity, memory bandwidth, and FLOPs, like FP4 FLOPs on GB300 was like, this is a minor change for Nvidia in the architecture. And you can see it in revenue per megawatt, $73.4 million to $99.7 million is what we forecast for those, those chips on Fable 5, like our fake Fable 5 assumptions for the architecture. And, yeah, this is telling, man, like the, the fact that, they're selling tokens for such a premium on the cost of the compute just implies that they will buy any compute that they can get their hands on, not just to serve these tokens, but also to train the models. Like, you know, selling tokens at such a premium means that you have more money to spend on compute. It's a cycle.

Yeah. And I think like, the thing is for, especially for these AI labs, they obviously have to make decisions on like what kind of what risk am I ready to take? And Dario has talked about this extensively, like, I don't wanna put my company bankrupt, so I have to plan for a certain amount of compute. And what has happened, like, time and time again is he under-forecasts, and that makes sense. It makes sense to not put his company at risk. And so the thing is that these guys know very well that the fuel, the core of their business is training because that's what generates future revenue growth. And so what ends up happening is that you have this core compute that is not enough to support the revenue growth, and so they end up having to pay for whatever is available spot gigawatts on demand, like right now at a premium effectively, right? And so maybe we're gonna talk about Microsoft later. I don't wanna open it to Raik on like, can they actually build the 10 gigawatts? But like the point is, hey, there's a huge gap because you cannot build data centers fast. So if you have a way to have access to a data center, like right now it's gigantic, that is, you know, amazing. And so if you can make $100 mil per megawatt per year, then there's no reason why your underlying provider can't charge you, you know, $50 million per megawatt year, right? And perhaps even more, like maybe we're being conservative actually with that pricing. We'll see, but I think it's actually fair to say $50 million a megawatt. And so it all comes down to like, can that company build these data centers? Can they finance them? Financing being extremely important because that is the core reason why we don't have enough data centers right now is that the lead time is, everyone in the industry, especially the non-hyperscalers, all the third parties, they need the capital upfront to then start making the orders to like, you know, switchgear supplier, cooling suppliers, build the data centers. They want to build them well, they do commissioning and whatnot. And so the data centers are only available generally 12 months from now and much more commonly 18 months from now. And so we keep being in that shortage where demand grows faster than supply. And so, you know, we're always short. So there's really a need for someone to take kind of that speculative risk to some extent, accept that, hey, maybe you're going to get it wrong and be underutilized and whatnot, but if you get it right, like you get that, that, you know, that premium. So that's what SpaceX does. They fill a major gap in the market. They've done it successfully, kind of luckily to some extent, because that wasn't the plan. And so now the, the, the, the, I think it, I think it shouldn't be a question of whether they can sell at this rate, given the economics that we see on the market today. The real question, frankly, for SpaceX, you know, whatever investors or whatever is, can they actually build half 10 gigawatts by the end of the year, uh, 2 gigawatts by the end of this year is pretty, uh, pretty easy. So can they build 8 gigawatts next year?

Yeah.

Let, let's come back to the motivation at the end, uh, and bring Raik in here. So like, yeah, the key question is how do they get enough chips online and how do they get enough people to actually do this? So that means having sites. Um, you talk through high level without revealing too much what you've, what you've got there.

Yeah. Yeah. So this was fun. This was basically, um, me and Zuher like 2 days ago doing, um, just melting GPUs, running through every single permit across the US to see what's kind of available. We scanned through like a million sites basically within like 2 days. Um, because if the sites are available, right? Like if you, like, Think of it this way. If you're Elon and like what we've seen from the past and the track record, which is the track record just says all you need is a warehouse and a gas pipeline basically, or even a gas pipeline and you can greenfield, which is recently done, but we'll get into. And so if that's really your only constraint is like getting this gas pipeline access and finding the turbines, which we'll get to as well, then you have a number of sites to start from to begin with, right? If you wanted to go warehouses, you can go warehouses too. We even found a few of those when we dug through the liens against, uh, MZX or Elon, which is—

and we've found like 5 very good candidates.

Yeah. Yeah.

That are like a million square feet, which is at least from a space point of view, obviously there's much more, but it's like, you know, over a gigawatt per like a million square feet, potentially 2.

Yeah, exactly. Like, like given the sizings on, um, MacroHard and MacroHarder, like comparing, I think MacroHard— was it MacroHard? It's like 700 megawatts and the MacroHarder is 500 megawatts. And so, Yeah, we had 1 million square foot warehouse, we had an 800,000 square foot warehouse, I think another million square foot warehouse, and then a few sites where pipelines are popping up. But broadly speaking, even though these sites look kind of like funky, right? It's just a warehouse on a plot of land, doesn't look like anything. It can be turned into this just based off what we've seen in the past. And so as much as people hated it internally, and probably externally too, it is possible, right? Like it is possible. I got, you know, people in my DMs on Slack going, this is an insane take. Like, what are you guys talking about? I don't know about this. It is possible. And like, we found enough sites that we can kind of show for it. And then if you look at like power side of things, we also wanted to— well, we wanted to add a chart to the article regarding like the turbine availability, but we didn't have enough time. We kind of pushed it out pretty quick. Anyway, when you look at the turbine availability, we found like 7 gigawatts of like undisclosed, like unidentified turbines. All right. And this is like just including those. It's not including the ones that might be bought out on a secondary market, like from a Fermi or from someone else or from a different state. All right. And so these turbines are kind of there as well. And so the power's there. Now we have the warehouse, kind of all that's left is the execution front. And we've already watched in the past these 2 gigawatts being built faster than anyone else in the industry. It's all kind of there as much as people don't like it at as obscene as the number is.

So, so I, I think like if you, if you count like the bottlenecks, uh, I think on the power side, as you said, people don't realize that there's actually much more available than expected. And I would just add like, hey, guys, think of Oracle in New Mexico, right? Like all these turbines are on the market. And then if for Raton or Cole, that the pipeline is going to be delayed, the fuel cells are also, some of them are going to be available. Nebeus, New Jersey, the engines are available, they're on the market and there's much more of these, right? Anyways, so there's much power I think that people realize and Elon already has something like 9 to 10 gigawatts of turbines on order or in operation. So there's already a lot of it in the fleet. But then what are the other bottlenecks? Labor obviously is a huge one. We just came up with a massive to call on that. So labor is a gigantic bottleneck. Electrical and cooling equipment can be a big bottleneck, switchgear, all that good stuff. Now, labor is really fascinating. I'll start with switchgear and overall just electrical and mechanical equipment. What I think is going to happen is he's going to extensively use equipment that comes from China. It's not like he doesn't know Chinese supply chains. Obviously, Elon knows them extremely well, better than probably any other firm that builds data centers in the US. He knows the electrical landscape extremely well because of what he does with Tesla and SpaceX as well. You can buy pre-assembled modules out of China, extremely large scale. Now, do customers want it? If you're in for a 20-year offtake with high SLA, it's probably not. If it's an on-demand cluster, I think you kind of don't care what is the electrical equipment so long as the cluster is is usable and there's some kind of SLAs and whatnot that induce a penalty if it doesn't work. But yeah, I think basically you have to assume that when you have a unique product on the market, which is, you know, a gigawatt 3 months from now, right? A gigawatt 3 months from now, the kind of like SLAs and the demands from customers are way different. And I think the best proof of that is Google signed with them. Like that's the most unthinkable thing. Google is a company that hates turnkey leases for data centers. They hate, they just like to do everything themselves. And yet they still signed with SpaceX, right? Because they were kind of bullied, quote unquote, by the time to market that was unbeatable. So electrical equipment from China pretty extensively, anything that they can get as fast as possible, if it's unconventional, they're gonna do that. Now labor, I think labor, it's fair to say, it's probably the single bottleneck these days on data centers. And that's where like, it's really interesting to look at Elon's history. I think when you look at what he's done with Tesla or SpaceX, Systematically, he always does things with much less labor than others. There's a precedent in the data center world as well, Colossus 2. The numbers we can see out there point to about 3,000 people, workers per day peak at that site on a per gigawatt basis. This is about 3x lower than what you see even from the very best data center developers and the ones that build very fast with highly modular data center designs. Call it, you know, Crusoe, for example, they're amazing at what they do. And yet somehow Elon needs 3x less people than them, right? And again, like it goes down to like, you know, extensive prefabrication from China and whatnot, like he does stuff differently. And another one. So on the supply chain, I believe it's possible. And by the way, folks, like we've told our institutional clients, I think several weeks from now that on the chip side, we can maybe go back to that when I finish on data centers. On the chip side, we saw orders on the supply chain for 5 to 10 gigawatts just for next year. So that's been in preparation for the last few months already. Our Taiwan supply chain team tracked that in a brilliant way, and they're hiring, by the way, if you want to join our memory team. But Rick, I've got a question for you. What about the permit bottleneck? How did they pull it off in Memphis, and how did they pull it off in Mississippi, and how can they pull it off again at 4x larger scale? Is it even possible?

Well, Jeremy, I'm glad you asked. Yeah. So Mississippi was a special case, right? That's when we had the Colossus 2 article a bit ago where we talked about, okay, so they couldn't figure out how to get the kind of permitting done for the power plant or the onsite generation within like Tennessee. And so the idea was, okay, well, the data center is right next to the border. Let's just build it over the border. And so they went ahead and built over the border. They got some permits for like some, some X amount of turbines. I think it was like 1 1.2 gigawatt permanent power plant was the idea. And then they started rolling in like mobile turbines and they're like, okay, you know, maybe down the line they become permanent. And then they roll in more mobile turbines and it just completely goes past like the permitting allowance and they're like, okay, well, this is, this is not great. And then it keeps going and they end up with a total of, you know, of course, 69 turbines. And so then eventually like there's some complaints, but then the DOJ intervenes and says it's okay. And so there's, it's kind of an unprecedented precedent in the sense where it's like, I guess he can just kind of do these things on that front. And so I don't put it past that we'll see these on the other fronts. And then the nice thing as well is when you choose warehouses, like in this instance, the million square foot warehouse in the middle of Mississippi, they're already permitted and zoned for the most part on these things. Uh, so you can actually just skip this, right? This is the benefit of not going with kind of a powered LAN solution. Instead, you might take more time to get the actual construction permit across, or I can take a warehouse and just rip everything out and then plug it all in for myself. All I need is an air permit then. And so you can kind of, ah, go ahead, Jeremy.

Yeah. One thing I want to add to that is, uh, on the, on the power side, like, like it's good to remember, like he does things in a very unusual way. And so the warehouse can be used for large-scale cooling, large-scale whatever, but maybe that parcel or even the parcels nearby can't be permitted for air pollution. And that's exactly what he did in Colossus too, right? Like literally the power plant is something like 2 miles away from the warehouse. And so he just built a private transmission wire from the power plant to the warehouse. Those are, as far as I know, running on medium voltage, which is highly inefficient by any means. But again, like, you know, you want to do it fast, like, you're not going to— so it's trade-offs. It's not going to be efficient, but there's ways to do it. And hey, guess what? They have the most brilliant electrical engineers on Earth. So what they're going to do is they're going to find these warehouses and they're going to scout like everything nearby to see where can I build a power plant, even if it's 3 miles away from it. I'll just figure out the electrical way as well. Right. So I think, yeah, unconventional is going to be highly unconventional, I think is the name of the game.

And I think like one, one other thing I want to add to this too is like with these timelines, right, the question comes about like Jeremy brought up SLAs earlier. I think it's a really interesting topic as well. I'll be brief though, but I think you're not going to these data centers for the highest SLA. Right? Like, that's obvious. You're going for like a large-scale contiguous cluster within the, within the next 5 months or so. And so this is fine in Elon's case, but it's also becoming more fine, like broadly across the industry, right? Like we have, you know, the Anthropic self-build that's like being talked about by multiple people now where you've got 99.7% uptime, right? This is unheard of. There's no multiple nines. There's no tiers of this from like the Uptime Institute. It's just, hey, let's remove the redundancy we need, you know, like if it's going to be internal electrical or if it's going to be the generators on backup, remove all of it. I don't care. I don't need these lead times. I don't need this like capex. And so people are also pretty fine with accepting lower SLAs if it means more speed to market. And so I think— I don't think it's really a bad thing if he's just ripping open a warehouse and plugging in GPUs and, you know, maybe melting them really quick, but then everything works afterward. It kind of doesn't really matter too much either.

Yeah. I mean, uh, fair enough. I think it, it, it matters to an extent. Uh, but they're, they're well known for, uh, being pragmatic as opposed to overly conservative when it comes to redundancy and like keeping things up and online in terms of all of their different services, which is, you know, if you gotta go fast, you have to make some concessions there.

So, and they bring a lot of batteries from Tesla as well, which helps on the redundancy.

Yeah, there is like some amount of, yeah, like, uh, let's go ahead.

Can we go back to talking about the potential customer here? Because obviously they've signed deals with Anthropic directly. They've signed deals with Google. You guys put in this article that Microsoft is the potential or the most clear customer. In other words, there may be, in my view, uh, Elon's not going to sell to Sam and OpenAI directly, but there's a way to get exposed to serving OpenAI's models, which means selling directly to Microsoft. So can we talk through that?

Yeah, like, again, essentially, you get to a point where the demand for this kind of service is only from folks that can make these economics of like $100 million megawatt-year. And there's, you know, 3 companies in the world today that have access to frontier AI models, ripping the full benefits of them, no cost, no revenue share whatsoever, just the cost of infrastructure: OpenAI, Microsoft, and Anthropic, right? Microsoft having the OpenAI IP. So Microsoft is in a pretty amazing position because they can monetize at this rate. However, they have two issues. One is they did this massive data center pause in second half of '24, first half of '25. They were ready to build more than anyone else, and then they paused. They didn't want to spend too much. And so now they find themselves in a situation where it takes a while to build data centers. And so they're not going to be able to have as much capacity as they would like to have. And the second thing is they signed a massive offtake contract with OpenAI. So we estimate that at about 7 gigawatts. And so Microsoft, a lot of the megawatts they're building today are serving OpenAI, but through infrastructure as a service, which in our chart that we showed earlier of the economics, that would be closer to the $12 million megawatt year as opposed to the $100 million, right? And so the question is like, okay, they could potentially accelerate like crazy. And that's why year to date, they've sort of woken up pretty dramatically. In any way they can. Examples, data center pre-leasing is always a good one. So you just call up third parties and sign contracts with them to, you know, build data centers. They've signed, you know, 7 gigawatts of data center pre-leasing year to date. They've been the most active company alongside Meta for fairly similar reasons, you could argue to some extent on the compute side. So extremely active on the leasing front. On the self-build, they reaccelerated a lot of their big sites like, you know, Fairwater in Wisconsin and a lot of other sites here and there. NeoCloud offtakes, they have been still very aggressive with folks like Enscale. They kept signing more deals with them and behind-the-meter agreement. That's completely new. They signed this 2.7 gigawatt offtake agreement with Chevron in Pecos County, right? And Microsoft is the company that would probably never do that, right? Like they've been very committed for a long time to like Five Nines and to the grid. And then they did this massive pivot where they're gonna go like behind the meter in the middle of West Texas. And so I think all of that tells you there's a big strategy change. They're preparing for a massive acceleration of their infrastructure buildout. However, it takes time to build stuff. You know, there are targeted dates and we agree with— that's what we forecast, you know, models 2028 for the Chevron deal, for example. A lot of the leases they've signed are for late '27 to '28. There's a gap, especially in late '26 and in the first half of 2027. Right. And so Microsoft is left with this option of like, uh, or, or this, I guess, debate of how do I get megawatts to, to bank on the opportunity they have of selling GPT tokens at 100 million megawatts a year.

Right now.

Right now.

Yeah. Not end of '27, not end of '28, like right now, you know, or as soon as possible next year. Yeah. Um, and can you talk through the 90-day cancellation policy, like in a little bit more detail? I know you've brought this up a couple of times, but conceptually, if you're doing the self-build and you're doing these long-term offtake agreements, that's actually quite different when it comes to serving the OpenAI tokens because like that's almost your base load in power terms. But now you've got this flexible sort of on-demand capacity, which, you know, it's, it's, you don't want to cancel everything you have and then try to strike new deals with these guys later. But conceptually, either side can pull the shoot on these SpaceX deals within 90 days, right?

Yeah, and I think, you know, conceptually what this means is as you do these deals and you can also keep, you can also plan ahead of time and you can do these deals, you can think, hey, I'm gonna do this for like, you know, 6 months to a year, whatever. And I'm also gonna sign for a guy that's gonna deliver for me a year from now so I can swap it. But anyways, the point is like the deal structured is Anthropic and Google and Reflection, same terms. Is whenever you want to cancel that contract, we can do so in 90 days. And that contract includes a monthly payment, which annualized equates to about $50 million a year. And so yeah, you just like, the only burden on your books is these 3 months that we're going to still pay at that rate, and then it's done. So from a balance sheet point of view, it's extremely easy to sign off on. The period at risk is really low, and that's extremely different from the compute agreements or data center agreements they've signed here today, those 10 gigawatts, that equates to, you know, over, well over actually $300 billion in total contractual value binding contracts. So that's going to be spent no matter what. So that's why it's so hard to plan infrastructure ahead of time is because a lot of capex is very balance sheet heavy, whether you build or you lease, it's the same thing. Whereas this is on demand. So the CFO can say, hey, you know, I think we can make a whole lot of money if we take this, there's no risk on my books. I'm just going to do this for whatever, again, like 6 months, a year. I mean, Google has been pretty open publicly that they did this because they have this need, and then they're going to cancel the contract like, you know, 3, 6 months from now. And again, like, that's why this is an option at all for these companies. And if, you know, given what we've said on Google, maybe you can expand on that if it's the topic today, I don't know. But we think Google is not going to be able to like compete at the frontier, but Microsoft is at the frontier thanks to OpenAI.

Yeah, I mean, I think, uh, we don't know where things are going to play out in everything that's not coding today. Uh, if you believe that coding is the path to AGI and coding, you know, turns into all of these other models, then Google is certainly behind and they, they need to catch up. And all of the signals we're seeing from them is that they are not doing this. So not only is Microsoft pouring in all this money, Google is also doing this in to the tune of $300 billion or something like that in capex. So maybe the point that I didn't make on the last podcast that I like to make now is just that it's shocking to see people like Jeff Dean leave and raise— even Dave Silver, John Jumper, or Noam Shazir, who left before Jeff Dean and Oriol and all the guys now— is that they're leaving to raise like $1 or $2 billion, which is an incredible seed round and, you know, the craziest thing ever, except for the fact that you have to compare it to these guys pouring in $300 billion of capex and be like, you couldn't give Jeff Dean 1% of this to do what he wants to keep him to stay, right? Instead, he's got to go raise money from external parties just to pursue the research that he wants to leave all of the infrastructure, all the contacts, all the people and start something new. Anyway, that's fascinating. Um, but let's, uh, yeah, let's, let's get back to maybe the, the question that's on everybody's mind right now, which is, okay, so SpaceX has a demand. We think there's a path for them to actually build all of this stuff in the timeline that people are talking about. How do they pay for it?

You know, saying on earnings, we're exclusive to Nvidia, it's the world's best hardware, you know, Vera Rubin is so amazing, coming from Elon Musk, I am assuming that's not free. I would put it this way. No, but look, the first thing is, again, like the revenue, you just pay it with operating cash flow. You know, we, so, okay, let's put it this way, right? Right now, like the contracts that they've signed already, which is just for a portion of their 2 gigawatts, it's for about a gigawatt or 1 to 1.5, gives them $50 billion of, you know, annualized revenue. So already the compute they have, they're, you know, they're making that amount of money, which is pretty tremendous. So $4 billion a month of, that's essentially pure cash. It's extremely high margins, like over 90% margins. EBITDA margins, of course. But even, you know, even EBIT margins are very high. And so the thing is, like, if they keep in that direction, and indeed what we said can happen, which is they can build 10 gigawatts and monetize at $50 mil gigawatt a year, and we think they're still going to build for internal training, and we'll see how much and so on and so forth. But let's say they build 5. So 5 times 50, you know, $250 billion. That's all operating cash flow. That's going to help financing pretty dramatically. And so Nvidia, we think, is likely to step in as a financing partner. I don't know, I don't think anyone at SemiAnalysis knows exactly what it's going to look like. But the point is, if they can make $50 million megawatt-year, that is well over the cost of the GPU. So the GPU is going to be is, you know, paid back in under a year. And so they don't even, you know, if they can get vendor financing, then it basically makes it like nearly cash neutral for them. And they have a bunch of cash on the balance sheet as well, and so on and so forth. But, and they're on the public markets as well, so they could always raise equity. There's always options. And we'll see what happens. But, you know, at least I'm not too worried given the amount of operating cash flow they're making on these deals.

Well, maybe the number one takeaway from this for me was just how much, big of a commitment this is to Nvidia. So SpaceX and xAI has historically been building the bulk of everything on Nvidia, but has been trying out other things. They've been, you know, dipping their toes in the water of TPU and AMD GPUs and things like that. And Elon declared on the first earnings call that they're Nvidia exclusive and he tweeted about it. I think he specifically said, we choose to go with Nvidia GPUs because they are the best. So this is great for Nvidia too. It's not just, you know, them doing this for themselves that we take away from this. I think this is also pretty positive on Nvidia.

Yeah. And I mean, I think it comes at like, you know, a pretty convenient time too with, at the moment we have the Nvidia backstops kind of hitting the market overseas. And then in the US there's direct data center leases, you know, like with I mean, the Hut 8 stuff. Can I say that, Jeremy, by the way? Yeah.

Okay. Uh, it's public info, man.

It is?

Yeah, it's public info. Okay.

Well, yeah, but the, the Hut 8 leases at like 704 megawatts and then, you know, all the back stuff's over overseas, right? Um, Firmus, I guess, for example, but it's pretty clear that Nvidia's like game for financing these situations, right? They understand that this is kind of the name of the game with Gemini or with Google becoming more and more of a TPU seller, um, and Anthropic continuing to build their workloads upon TPUs, making it harder to like kind of switch back into Nvidia shapes. Uh, you end up with Nvidia pushed a little bit and having to really finance or really help out a lot of these projects to make sure that more folks get on the Nvidia like kind of side of things. And I mean, Whenever we go to the conferences too, like Jeremy and I have had conversations with folks who are saying like, yeah, you know, sometimes we'll be having a talk with Nvidia and like we're talking about taking this site and then maybe having Anthropic as an off-taker or something. But then we've got, you know, the Google backstop, you know, and we kind of wave that around. Maybe we can, you know, we can curry some favor with Nvidia there. And I mean, it tends to like happen, right? Nvidia kind of needs to like get more of these labs or these off-takers locked into the Nvidia ecosystem. And so I think more than reasonable at the moment.

Yeah. And one thing I want to add as well on your question, Jordan, on the financing, what one comment that Elon made on these earnings I thought was fascinating is he said like, you know, I'm going to try and build up to 20 gigawatts, whatever, but he said specifically gigawatts of power and cooling, right? And so what that tells me is that what he's going to do first and foremost is building data centers. And when you have the data center built and it's like extremely close to delivery for the end customer, like everything becomes much easier. And so buying the GPUs at that point, like you start to have many more financing options because you know, you can like bully a customer into, hey, maybe it's a 6-month deal, maybe it's going to be a 40 mil, whatever, but there's so many more options. When you have the data center built and it's so close to delivery, right? And so I think that's going to be his main focus currently. He's obviously, as we said, talking to the chip supply chain. We think he's talking to everyone in the supply chain to ensure he's going to have enough chips. And it could fall short. And he said that explicitly, he thinks it could fall short as well. But he's going to try and make sure that he has enough data centers. And that's going to be the number one priority is build them. And then, you know, financing and buying the GPUs can sort of come easier once you have them built. So Uh, but building 10 gigawatts of data centers, you know, it's still a lot of money.

Uh, still a lot of money, but the SpaceX has cash and again, they have like decent operating cash flow of, you know, it does feel like the ability to generate all this cash flow kind of depends on them having the chips on their balance sheet and kind of having the optionality on who to sell it for, as sell it to, as opposed to, you know, having a empty data center that somebody else can bring their own chips to that's on their own balance sheet. Probably not as big on the leverage side there in terms of sales. But anybody who is asking the question about like the very natural question of like, how are they going to be able to pay for this much capex? I think, you know, what I'm hearing from you guys, which makes sense, is that you need to be considering this as Nvidia's balance sheet, probably the strongest in the entire world except for maybe Apple at this point. And then Elon's ability to raise money on the back of an incredibly profitable, incredibly cashflow generating business, which, you know, one of, if not the best guys in the entire world at raising money. So, um, they're going to go for it, man. And it's, uh, it's inspiring to see. It's an incredible, uh, it's an incredible thing to be a part of and to witness as we see them, uh, work through this. Uh, where do we go from here, guys? What do you think's the, the next step that people are gonna take? What are people gonna be looking for on the next SpaceX earnings call or from Microsoft or from Nvidia or anybody else that we've mentioned this so far?

Yeah, I, I think it really comes down to like, how much do people believe in these ultimate economics? And I, I think like it's starting to re-happen and I think more and more we're gonna see towards the end of the year that yes, indeed selling tokens, especially frontier AI tokens, is insanely profitable and so profitable, it's just going to drive the industry to do all kinds of creative things to try to bank on this. So I think that's the only thing is, do you believe in that profitability? I think we'll see it from OpenAI and Anthropic as they're going to keep accelerating revenue because any megawatts they bring online gives them so much. And I think it's unrealistic to expect that if they both accelerate at this pace, Microsoft is not gonna wanna do the same thing. Uh, if anything, the question should be asked to Microsoft, like, you can make that amount of money, why aren't you doing it? Like, what are you doing about it?

I think, I, I also think we, we, somebody commented or like somebody on Twitter replied to one of the comments saying like, you guys have, you know, $12 million per megawatt for the CoreWeave deals, but you're assuming, you know, $40 to $50 megawatt, $40 to $50 million per megawatt for the SpaceX deals. And like, yeah, obviously this is, these were the deals, but I think what's kind of interesting to me is like, in my opinion, in my humble data center opinion, I'm not the tokenomics guy. I think, I think this is kind of overdue, right? I remember we at one point we did some, we ran some analysis on, I think it was 4.6 or 4.8 on like a GB300 and somebody found it was like 90 to 95% margins and this was I don't know if we maybe got that number wrong. I mean, 85% still insanely high, but in any case, reading these numbers was always to me like, well, I mean, somebody has to, you know, capture more value somewhere else, right? Like, why are we selling these GPU hours for so cheap if they're making 90% margins? To me, this kind of always made a bit of sense. And they weren't ever going to like, I mean, I mean, maybe some world, but it never felt like they were going to keep these margins forever. And I mean,, it makes sense that somebody like Elon is coming in and charging these prices because why not? They're still going to make so much money. I don't really think it's going to kill them or hurt them too much. And so I, I think it makes a lot of sense, honestly, what's going on.

What's your conclusion, Jordan? Close it for us.

Yeah. Well, I think, look, the one thing in the back of my mind that we haven't had a chance to talk about on either of these podcasts this week is just the, OpenAI, uh, release about the security incident with Hugging Face and how they had this, uh, talk at Black Hat Summit, um, which is from their Astra model, which is in testing right now. And I would say it seems to be comparable or, uh, you know, it's certainly comparable to Mythos in terms of like the Project Glasswing sort of cybersecurity concerns that are being—

this is, this is the one where you, you kind of commented in Slack saying it like actually scared you, right?

I mean, I've been scared for a while, but this is like two separate organizations identified a security incident where multiple agents were coordinating to attack infrastructure using really sophisticated methods. I would say not super sophisticated in some of them, but the way the agents coordinated to do this and kind of went awry during evaluations was crazy. Everybody should go watch that video. I won't even attempt to try to explain the details right now. But the point is that, um, if we are making the, uh, steel man case for the people who think Elon can't do this, the, the, the biggest concern that I have is not actually the execution and all the stuff that we've laid out here. It's that demand has a problem in the future and not because I think the models aren't going to be good enough, I actually think the models are going to be too good and then people will be so scared they're going to shut down access. They're going to stop people being able to do this stuff. Politicians are going to be involved. And if there's any, you know, case where this doesn't work, it's more political than technical or operations related. And, uh, we're yet to see that. On the other side, the creative part of my mind is going, well, there's lots of other use cases outside of coding and security that we can pursue in order to give people a lot of, uh, tokens and keep models going. I mean, everybody's focused on coding right now. And so what about everything else? When you point the GPUs and the researchers towards drug discovery and material science and weather prediction and video generation and robotics, Reyk, like, you know, this is like, if, if you point the GPUs and the research effort towards something other than cybersecurity, I think we will, um, realize a lot of benefits there, but Man, this cybersecurity stuff is really concerning right now. It's really, really freaking scary, man.

Yeah.

Good. Yeah.

A nice positive note to end it on there.

But yeah, for what it's worth, we're going to be coming out with an article in a couple of weeks about our security experience testing a bunch of the NeoClouds recently and how we found— I mean, these guys talk about zero days, right? Which is like a publicly disclosed CVE, meaning like a security issue. Of some piece of software that's running in somebody's infrastructure where the ability to exploit it requires downloading something from the internet and testing it works on the system and making, in some cases, just minor modifications to make sure it works. This is something an agent can do. It's, it seems kind of obvious an agent can do this from everybody who's had experience doing research with these things and providers out there that are not SpaceX, but are on the lower tier. But there's plenty of them are running stuff, not with like 6-week-old zero days, but like 3-year-old zero days that found in a second when using this. Just check the version of software they're running. And the ability to develop POC exploits to show them that this is the potential concern took us like afternoons, not like weeks and months of effort and not needing to be a security expert or a Linux kernel expert or an Nvidia GPU driver expert or a Kubernetes expert. You just are like, hey, model, check for this version. It checks for it and then it can build an exploit in in a couple of hours. So anyway, that's when the, the, you know, people are actually trying to work on it directly. The really scary part about this story is that it was autonomous agents. People weren't monitoring going and doing this by themselves because they were pursuing a goal of trying to find a dataset to pass their eval. So they went out and hacked Hugging Face, which hosts these datasets by finding issues in like the HDF5. You found a zero-day in the HDF5 data format on Hugging Face. Unbelievable. So anyway, um, maybe, okay, because I tried to explain it, I'll, I'll give people another nugget so that they can go watch the video. The agents were coordinating with a message board using file names on a JFrog artifactory service that ran remotely. So an agent would go leave a note in the file name and then another agent would come back later and pick it up and be like, oh, you did this. I'll keep working on that. So they were like a swarm, all pursuing the research to attack Hugging Face. Using file names on a, like a file server. Not even, they couldn't even have access to write the contents of the files, just the names.

That's absurd. That is absurd.

Yeah, you should go, you should go watch this video. It's 30 minutes, well worth it, guys. If you want to understand where we're at right now in AI progress at the frontier, and it's a glimpse into what people like Elon who see Grok training and Talk to Anthropic and OpenAI AI are seeing with these models that they hold internally and don't release publicly because even the stuff that's public is so unbelievably profitable. They don't even need to release these, you know, research projects that are being evaluated. So we're, we're, yeah, we're, we're seeing rapid progress right now of how much these things can improve. Okay. Any final thoughts after I went on that rant about cybersecurity when we were supposed to be talking about SpaceX?

Looking forward to the article.

Yeah, really. I was wondering how to change the topic back. It is a really out-there call. I think people will look back on it and go, wow, those guys are crazy, but they got it. Hopefully. I mean, God, hopefully.

Yeah. I mean, we got a lot of hate when we had our call on Amazon and we were like, guys, just look at those data centers. You know, they're gonna accelerate revenue pretty obviously. People are like, no, you're idiots. They're never gonna accelerate. They're losers of AI. So right now SpaceX is the loser of AI.

I guess, man. Yeah. We'll, we'll, we'll see you at Christmas for the year-end review 2027 podcast where we'll check in and see, uh, see if this one was right or wrong. Okay guys.

Looking forward to it. All right.

I think we, I think we got it. All right. Thank you, Jordan.

All right.

Good job, guys. Take care.

Bye-bye.