第024期 - SpaceX的10GW计划将推动2027年实现3000亿美元ARR（数据中心、能源）

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

Yeah, man. When you have GPUs at home, you've got to show that to the world.

Actually, a GPU.

It's a decommissioned GPU. You have all the heat sinks, 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?

Got to cut costs, man. But they keep telling me that it's cheaper on the cloud, you know? So I'm like, “Okay, my electricity bill can't handle it.”

Jeremie, Reyk, we're going to talk about SpaceX doing 10 gigawatts in 2027. You guys ready?

1. The $100M Thesis

Welcome back to SemiAnalysis Weekly. We've got Jeremie and Reyk. Like I said, we're going to talk about the article we put out that got a lot of traction: “SpaceX 10 Gigawatts in 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 talk through some of Elon's statements on the earnings call, the economics, how we get to $100 million per megawatt per year that they're going to sell this stuff at, how Microsoft, as a potential customer, pays for it when you pass through the tokens, and how they get enough chips and people. Welcome. I'm excited to dig in.

Yeah, let's go.

Yeah, thanks for having us, Jordan.

All right, Jeremie, you're the first author here. Conservatively, what do Elon's statements during the earnings call mean? What's the takeaway when SpaceX has its first-ever earnings call and says that it has these gigawatt ambitions?

Look, I think it all really starts with the economics. That's really the key thing. When we look at the course of 2026, gross margins for these labs just kept going up, and that has been the key driver of their ARR acceleration, right? If you look at them combined today—that is, Anthropic and OpenAI—they're both adding a combined $20 billion of ARR per month, actually close to $30 billion now. By itself, that drives AI revenue adding close to $400 billion per year, and they're doing that by increasing their gross margins.

Increasing their gross margins essentially just means that, for any given amount of compute, which is typically a roughly fixed cost, revenue per watt goes up. We've seen $12 million, $13 million, $14 million per megawatt per year from the likes of CoreWeave.

That's what I want to flip to you, Jordan. We've done a lot of work on trying to understand what the actual revenue per megawatt is. Obviously, we have InferenceX, and in this article what we put out is that we think OpenAI and Anthropic can fairly easily, today, on the API, make $100 million per megawatt per year.

Tying it back to why Elon is trying to build so many gigawatts so fast: no one else is doing it so fast. No one else is really banking on the opportunity that, for any megawatt they have, they can make so much money. 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 GPU hour; you're going to make 50% gross margin.”

Essentially, you just want to replicate that playbook and do that at a much larger scale. If the economics are that good, everything downstream is much easier, right? So the first question we should ask ourselves is: Is $100 million per megawatt per year realistic for OpenAI and Anthropic on the API today? I don't know. What do you think of that, Jordan?

Yeah, I think it is real. It's definitely realistic. We've dug into this in great detail. Let me share this chart so that we can explain the details.

2. Pricing & Google's Deal

When we talk through the different ways in which GPUs transact, you can start at the beginning: $12 billion or $13 billion per gigawatt, which is $12 million or $13 million per megawatt. That's the 5-year average infrastructure-as-a-service price across CoreWeave, Oracle, and Nebius, and it's long-term, close to self-build pricing for these guys. Maybe they're making double-digit margins on these things, but it's not massive.

Where things change is when you sell for a premium, because either you're getting on-demand, which is the green column here, the GB300s, or you're doing these deals where SpaceX is selling its existing compute to somebody because it can turn on so much of it immediately. Like you said, that's the reason you can transact at such a significant premium. Specifically, this Google deal at 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 add super quickly that I think an amazing term in the contract they have is the 90-day cancellation policy, because for Google, Microsoft, and Anthropic, they have zero risk. They can cancel this if they realize the economics aren't working anymore.

Anyway, it makes it way easier. It's really emergency megawatts: you want them right now, they're big, easy to cancel, and that's the price. 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, you can look at from 2 angles.

Both the InferenceX data that we have, where we actually run real workloads on the latest chips—GB200 and GB300—with the latest optimizations from vLLM, SGLang, and TensorRT-LLM from NVIDIA, ROCm and other software from AMD, whatever chip you're trying to use. The point is that we get that real data from open-source models, and increasingly these models are approaching the size of the frontier models.

Kimi K3 is a 3-trillion-parameter model. DeepSeek is using sparse attention, and Kimi is using all of these sparse-attention approaches too, including Kimi Linear. The point is that these are well over 1 trillion parameters total. That takes up all the memory space they have. They have million-token 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.

So you can get a proxy, or a lower bound, let's say, of what a 3-trillion- or 2-trillion-parameter model is going to perform at. Assume that the frontier models 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 Chinese SKUs.

These are not the latest and greatest from NVIDIA. They're increasingly being developed on Huawei chips and others from China. That's a lower bound of performance: our public data on InferenceX. The upper bound is really our simulator, which takes it operation by operation.

These are the GEMMs, the collectives, and all of the actual operations simulated on the current hardware, where we test the actual performance of the collectives and the GEMMs and everything else on real hardware. Then we trace it through what we think is the shape of the frontier model and forecast this forward for what we're going to see on Rubin when we see increases in the FLOPs, memory bandwidth, and networking bandwidth.

All of this is going to change when people start deploying Vera Rubin next year, which is going to be all of SpaceX's capacity next year. They're not going to be deploying a bunch of GB300s; they're going to be deploying the latest and greatest. So you're going to see a performance increase from the GPUs on a relative, per-megawatt basis. That's going to allow people to produce more tokens. 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.

The forecast goes from SpaceX and Google signing a deal that 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'd expect somebody like Google to make a financially sound decision to do that. They've got to see returns on the capital that they're investing in those GPUs, let alone the fact that the leaked financials we're seeing from the labs right now are showing 85% gross margins.

That matches what we are seeing with our simulator and our InferenceX data, which would imply around $100 billion in sales at a cost of around $15 billion per gigawatt, right? That matches this chart that's on screen there in the blue.

That 85% is a blend, which includes older or lower-performing GPUs, not just the GB300. So you could argue that $100 million is actually conservative for GB300 or for Vera Rubin, in fact.

Yeah. I'll scroll down and show another chart here. There are really significant differences even as we go from GB200 to GB300. The increase in memory capacity, memory bandwidth, and FP4 FLOPs on GB300 was a minor change for NVIDIA in the architecture.

You can see it in revenue per megawatt: $73.4 million to $99.7 million is what we forecast for those chips on Fable 5, based on our fake Fable 5 assumptions for the architecture.

This is telling. The fact that they’re selling tokens at such a premium over the cost of compute implies that they will buy any compute they can get their hands on—not just to serve these tokens, but also to train the models. Selling tokens at such a premium means that you have more money to spend on compute. It’s a cycle.

3. Training vs Inference

The thing is, especially for these AI labs, they obviously have to decide what kind of risk they’re ready to take. Dario has talked about this extensively: “I don’t want to put my company at risk, so I have to plan for a certain amount of compute.” What has happened time and time again is that he underforecasts, and that makes sense. It makes sense not to put his company at risk.

These guys know very well that the fuel—the core of their business—is training, because that’s what generates future revenue growth. What ends up happening is that this core compute is not enough to support the revenue growth, so they have to pay for whatever spots are available on demand, right now, at a premium.

Maybe we’re going to talk about Microsoft later. I don’t want to open it to Reyk on whether they can actually build the 10 gigawatts, but the point is that there’s a huge gap because you can’t build data centers fast. If you have a way to access a data center right now, that’s gigantic. That’s amazing.

If you can make $100 million per megawatt per year, there’s no reason why your underlying provider can’t charge you $50 million per megawatt per year, right? Perhaps even more. Maybe we’re being conservative with that pricing, actually. We’ll see, but I think it’s fair to say $50 million per megawatt-year.

It all comes down to whether that company can build these data centers and whether it can finance them. Financing is extremely important, because that’s the core reason we don’t have enough data centers right now. The lead time is long, and everyone in the industry—especially the non-hyperscalers, all the third parties—needs the capital up front to start placing orders, switchgear suppliers, cooling suppliers, and build the data centers.

They want to build them well, do commissioning, and so on. The data centers are generally only available 12 months from now, and much more commonly 18 months from now. We keep being in that shortage where demand grows faster than supply, so we’re always short.

There’s a real need for someone to take that speculative risk—to accept that maybe they’re going to get it wrong and be underutilized—but if they get it right, they get that premium. That’s what SpaceX does. They fill a major gap in the market. They’ve done it successfully, somewhat luckily, because that wasn’t the plan.

I don’t think it should be a question of whether they can sell at this rate, given the economics we see in the market today. The real question, frankly, for SpaceX, its investors, or whoever, is whether they can actually build out 10 gigawatts by the end of next year. 2 gigawatts by the end of this year is pretty easy, so can they build 8 gigawatts next year?

4. Sites & Supply Chain

Let’s come back to the motivation at the end and bring Reyk in here. The key question is: how do they get enough chips online, and how do they get enough people to actually do this? That means having sites. Can you talk through, at a high level, without revealing too much, what you’ve got there?

This was fun. This was basically me and Zuhair, 2 days ago, melting GPUs and running through every single permit across the US to see what’s available. We scanned through about 1 million sites within 2 days.

If the sites are available, think of it this way: if you’re Elon, and you look at what we’ve seen from the past and at the track record, all you need is a warehouse and a gas pipeline, basically. Or even just a gas pipeline, and you can greenfield, which he’s recently done—but we’ll get into that.

If your only constraint is getting access to a gas pipeline and finding the turbines, which we’ll get to as well, then you have a number of sites to start from. If you wanted warehouses, you can go with warehouses, too. We even found a few of those when we dug through the liens against xAI or Elon, which is—

And we found 5 very good candidates.

Yeah, that are around 1 million square feet, which, at least from a space point of view—obviously, there’s much more—is over 1 gigawatt per 1 million square feet, potentially 2.

Exactly. Given the sizing on Memphis and Mississippi, I think Memphis was 700 megawatts and Mississippi was 500 megawatts. We had a 1-million-square-foot warehouse, an 800,000-square-foot warehouse, another 1-million-square-foot warehouse, and then a few sites where pipelines are popping up.

Broadly speaking, even though these sites look funky—they’re just a warehouse on a plot of land and don’t look like anything—they can be turned into this, based on what we’ve seen in the past. As much as people hated it internally, and probably externally too, it is possible. It is possible. I got people in my DMs on Slack saying, “This is an insane take. What are you guys talking about? I don’t know about this.”

It is possible, and we found enough sites to show for it. If you look at the power side of things, we also wanted to add a chart to the article regarding turbine availability, but we didn’t have enough time. We pushed it out pretty quickly.

When you look at turbine availability, we found 7 gigawatts of undisclosed, unidentified turbines. This is just including those; it doesn’t include the ones that might be bought out on a secondary market, like from Fermi or someone else, or from a different state. These turbines are there as well. The power is there.

Now we have the warehouse. All that’s left is execution. We’ve already watched these 2 gigawatts being built faster than anyone else in the industry. It’s all there, as much as people don’t like it and as insane as the number is.

If you count the bottlenecks, on the power side, as you said, people don’t realize that there’s actually much more available than expected. Think of Oracle in New Mexico, for example. All these turbines are on the market.

For coal, the pipeline is going to be delayed. The fuel cells are also available; some of them are going to be available at Nebius in New Jersey. The engines are available. They’re on the market. There’s much more of this than people realize.

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 what are the other bottlenecks? Labor is obviously a huge one. We just came out with a massive article on that, so labor is a gigantic bottleneck. Electrical and cooling equipment can also be a big bottleneck—switchgear and all that good stuff.

Labor is really fascinating. I’ll start with switchgear and overall electrical and mechanical equipment. What I think is going to happen is that he’s going to extensively use equipment 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. You can buy preassembled modules out of China at extremely large scale. Do customers want it? If you’re in for a 20-year offtake with high SLAs, probably not. If it’s an on-demand cluster, I think you kind of don’t care what the electrical equipment is, so long as the cluster is usable and there are some SLAs that impose a penalty if it doesn’t work.

When you have a unique product on the market—a gigawatt 3 months from now—the SLAs and demands from customers are very different. The best proof of that is that Google signed with them. That’s the most unthinkable thing. Google hates turnkey leases for data centers. They like to do everything themselves, and yet they still signed with xAI because they were bullish, quote unquote, on the time to market, which was unbeatable.

So, they’ll use electrical equipment from China pretty extensively. Anything they can get as fast as possible, if it’s unconventional, they’re going to do that.

Labor is probably the single biggest bottleneck in data centers these days. That’s where it’s really interesting to look at Elon’s history. When you look at what he’s done with Tesla or SpaceX, 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 workers per day at peak at that site, on a per-gigawatt basis. This is about 3× lower than what you see even from the very best data center developers—the ones that build very fast with highly modular data center designs, like Crusoe, for example. They’re amazing at what they do, and yet somehow Elon needs 3× fewer people than them, right? Again, it goes down to extensive prefabrication from China and whatnot. He does things differently.

Another one: on the supply-chain side, I believe it’s possible. By the way, folks, 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—we saw orders in 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. They’re hiring, by the way, if you want to join our memory team.

But Reyk, a question for you: What about the permitting bottleneck? How did they pull it off in Memphis and Mississippi, and how can they pull it off again at 4× scale? Is it even possible?

5. Permitting Playbook

Well, Jeremie, I’m glad you asked. Mississippi was a special case, right? That’s when we had the Colossus 2 article a bit ago, where we talked about how they couldn’t figure out how to get the permitting done for the power plant, or the on-site generation, within Tennessee. The idea was, okay, the data center is right next to the border—let’s just build it over the border.

So they went ahead and built it over the border. They got some permits for some amount of turbines. I think it was a 1.2-gigawatt permanent power plant—that was the idea. Then they started rolling in mobile turbines and said, “Okay, maybe down the line they become permanent.” Then they rolled in more mobile turbines, and it completely went past the permitting allowance. They said, “Okay, well, this is not great.” Then it kept going. They ended up with a total of 69 turbines.

Eventually, there were some complaints, but then the DOJ intervened and said it was okay. It’s kind of an unprecedented precedent, in the sense that I guess he can just do these things on that front, so I don’t put it past them that we’ll see these on the other fronts.

The nice thing as well is that 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. You can actually just skip this. This is the benefit of not going with a powered-land solution instead. You might take more time to get the actual construction permit across, or I can take a warehouse, rip everything out, and then plug it all in for myself. All I need is an air permit.

One thing I want to add to that is on the power side, it’s good to remember that he does things in a very unusual way. 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. That’s exactly what he did in Colossus 2, right? Literally, the power plant is something like 2 miles away from the warehouse.

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 measure. But again, if you want to do it fast, you have to make trade-offs. It’s not going to be efficient, but there are ways to do it.

And guess what? They have the most brilliant electrical engineers on Earth. What they’re going to do is find these warehouses and scout everything nearby to see where they can build a power plant, even if it’s 3 miles away. They’ll just figure out the electrical side as well, right? So, yeah, unconventional—it’s going to be highly unconventional, I think, is the name of the game.

One other thing I want to add to this, too, is that with these timelines, the question comes about, as Jeremie brought up earlier, SLAs. 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? That’s obvious. You’re going for a large-scale, contiguous cluster within the next 5 months or so.

This is fine in Elon’s case, but it’s also becoming more acceptable broadly across the industry, right? We have Anthropic’s self-build, which is being talked about by multiple people now, where you’ve got 99.7% uptime. This is unheard of. There are no multiple nines, and there are no tiers from the Uptime Institute. It’s just, hey, let’s remove the redundancy we need. If it’s internal electrical or generators for backup, remove all of it. I don’t care. I don’t need these lead times, and I don’t need this capex.

People are also pretty fine with accepting lower SLAs if it means more speed to market. I don’t think it’s really a bad thing if he’s just ripping open a warehouse and plugging in GPUs and maybe melting them really quickly, but then everything works afterward. It kind of doesn’t really matter too much either.

Yeah, I mean, fair enough. I think it matters to an extent. But they’re well known for being pragmatic as opposed to overly conservative when it comes to redundancy and keeping things up and online across all of their different services. If you have to go fast, you have to make some concessions there.

And they bring a lot of batteries from Tesla as well, which helps with the redundancy.

Yeah.

There is some kind of thing involved. Yeah, yeah. Go ahead.

6. Microsoft's Role

Can we go back to talking about the potential customer here? Obviously, they’ve signed deals directly with Anthropic, and they’ve signed deals with Google. You guys put in this article that Microsoft is the potential, or the clearest, customer. In other words, there may be, in my view, a way—Elon isn’t going to sell directly to Sam and OpenAI, but there’s a way to get exposure to serving OpenAI’s models, which means selling directly to Microsoft. So can we talk through that?

Yeah. Essentially, you get to a point where the demand for this kind of service is only from folks that can make economics of about $100 million per megawatt per year. There are 3 companies in the world today that have access to frontier AI models and are capturing the full benefits of them: OpenAI, Microsoft, and Anthropic. No cost, no revenue share whatsoever—just the cost of infrastructure, with Microsoft having the OpenAI IP.

Microsoft is in an amazing position because they can monetize at this rate. However, they have 2 issues. One is that they had this massive data center pause in the second half of 2024 and the first half of 2025. They were ready to build more than anyone else, and then they paused. They didn’t want to spend too much. So now they find themselves in a situation where it takes a while to build data centers, and they’re not going to be able to have as much capacity as they would like.

The second thing is that they signed a massive offtake contract with OpenAI. We estimate that at about 7 gigawatts, so a lot of the megawatts they’re building today are serving OpenAI, but through infrastructure as a service. In the chart we showed earlier on the economics, that would be closer to $12 million per megawatt per year, as opposed to $100 million.

The question is, okay, if they could potentially accelerate like crazy—and that’s why year to date they’ve woken up pretty dramatically—in any way they can. Data center pre-leasing is always a good one. You just call up third parties and sign a contract with them to build data centers. They’ve signed 7 gigawatts of data center pre-leasing year to date. They’ve been the most active company alongside Meta, for fairly similar reasons.

You can do the same, to some extent, on the compute side. On the self-build side, they reaccelerated a lot of their big sites, like Fairwater in Wisconsin, and a lot of other sites here and there. On neocloud offtakes, they’ve still been very aggressive with folks like Nscale; they’ve kept signing more deals with them.

Behind-the-meter agreements are completely new. They signed this 2.7-gigawatt offtake agreement with Chevron in Pecos County. Microsoft is the company that would probably never do that, right? They’ve been very committed for a long time to five nines and to the grid, and then they made this massive pivot where they’re going to go behind the meter in the middle of West Texas.

I think all of that tells you there’s a big strategy change. They’re preparing for a massive acceleration of their infrastructure build-out. However, it takes time to build stuff. They have a target date, and we agree with that. That’s why we forecast 2028 for the Chevron deal, for example.

A lot of the leases they've signed are for late 2027 to 2028. There's a gap, especially in late 2026 and in the first half of 2027, right? Microsoft is left with this option—or, I guess, this debate—of how to get megawatts to bank on the opportunity they have of selling GPT tokens at $100 million per megawatt-year.

Right now.

Yeah, not into 2028. Right now, or as soon as possible next year. Can you talk through the 90-day cancellation policy in a little more detail? I know you've brought this up a couple of times, but conceptually, if you're doing the self-build and these long-term offtake agreements, that's actually quite different when it comes to serving the OpenAI tokens. That's almost your baseload in power terms, but now you've got this flexible, on-demand capacity. 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 chute on these SpaceX deals with 90 days, right?

Yeah. Conceptually, what this means is that as you do these deals, you can also plan ahead of time. You can think, “Hey, I'm going to do this for 6 months to a year, whatever, and I'm also going to sign with a guy that's going to deliver for me a year from now so I can swap it.”

The point is that the deal structure with Anthropic, Google, and Reflection AI is the same: whenever you want to cancel that contract, you can do so in 90 days. That contract includes a monthly payment, which annually equates to about $50 million a year, maybe. The only burden on your books is these 3 months where you're still going to pay at that rate, and then it's done.

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 GW equate to well over $300 billion in total contractual value in binding contracts. That's going to be spent no matter what.

That's why it's so hard to plan infrastructure ahead of time: a lot of capex is very balance-sheet-heavy. Whether you build or lease, it's the same thing. Whereas this is on-demand, the CFO can say, “Hey, 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 6 months to a year.” Google has been pretty open publicly that they did this because they have this need, and then they're going to cancel the contract 3–6 months from now. That's why this is an option at all for these companies.

Yeah. I mean, I think we don't know where things are going to play out in everything that's not coding today. If you believe that coding is the path to AGI, and coding turns into all of these other models, then Google is certainly behind. All the signals we're seeing from them indicate that they are not doing this. Not only is Microsoft pouring in all this money, but Google is also doing this, to the tune of $300 billion or something like that in capex.

Maybe the point that I didn't make on the last podcast, which I'd like to make now, is that it's shocking to see people like Jeff Dean leave and raise money—even David Silver, John Jumper, or Noam Shazeer, who left before Jeff Dean, and all the guys now. They're leaving to raise $1 billion or $2 billion, which is an incredible seed round and the craziest thing ever, except for the fact that you have to compare it to these guys pouring in $300 billion of capex and say, “You couldn't give Jeff Dean 1% of this to do what he wants, to keep him to stay?”

Instead, he's got to go raise money from external parties just to pursue the research that he wants, leave all of the infrastructure, all the contacts, all the people, and start something new. Anyway, that's fascinating.

Let's get back to maybe the question that's on everybody's mind right now, which is: SpaceX has a demand. We think there's a path for them to actually build all of this stuff on the timeline that people are talking about. How do they pay for it? Saying on earnings, “We're exclusive to NVIDIA. It's the world's best hardware. Vera Rubin is so amazing,” coming from Elon Musk, I am assuming that's not free.

7. Paying For It

Well, look, the first thing is, again, the revenue—you just pay it with operating cash flow.

Right now, the contracts that they've signed already, which are just for a portion of their 2 GW, are for about 1 GW to 1.5 GW and give them $50 billion of annualized revenue. So already, with the compute they have, they're making that amount of money, which is pretty tremendous: $4 billion a month of essentially pure cash, extremely high margins, over 90% EBIT margins, of course. Even EBITDA margins are very high.

If they keep going in that direction and what we said can happen does happen—which is that they can build 10 GW and monetize it at $50 million per megawatt-year—and we think they're still going to build for internal training, we'll see how much and so on and so forth, but let's say they build 5 GW. Five times $50 million per megawatt-year across 5 GW is $250 billion. That's all operating cash flow.

That's going to help financing pretty dramatically. NVIDIA, we think, is likely to step in as a financing partner. I don't know—I don't think anyone at all knows exactly what it's going to look like. But the point is, if they can make $50 million per megawatt-year, that is well over the cost of the GPU. So the GPU is going to be paid back in under a year.

If they can get vendor financing, then it basically makes it nearly cash-neutral for them. They have a bunch of cash on the balance sheet as well, and so on and so forth. They're on the public markets as well, so they could always raise equity. There are always options, and we'll see what happens. 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 is just how big of a commitment this is to NVIDIA. SpaceX and xAI have historically been building the bulk of everything on NVIDIA, but they've been trying out other things. They've been dipping their toes in the water of TPUs and AMD GPUs and things like that. Elon declared on the first earnings call that they are 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 them doing this for themselves; that's what we take away from this. I think this is also pretty positive for NVIDIA. Can I say that, Jeremie, by the way? Yeah. Okay.

Public info, man.

It is.

Yeah, it's public information.

Okay. Well, yeah, but the Hut 8 leases are 7,004 MW, and then all the backstop stuff is overseas, right? Firmus, I guess, for example. But it's pretty clear that NVIDIA is game to finance these situations, right? They understand that this is kind of the name of the game with Gemini, with Google becoming more and more of a TPU seller, and Anthropic continuing to build their workloads upon TPUs, making it harder to switch back into NVIDIA chips.

You end up with NVIDIA being 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 side of things. Whenever we go to the conferences, too, Jeremie and I have had conversations with folks who are saying, “Yeah, sometimes we'll be having a talk with NVIDIA, and we're talking about taking this site and then maybe having Anthropic as an offtaker or something. But then we've got the Google backstop. We kind of wave that around; maybe we can curry some favor with NVIDIA there.”

And it tends to happen, right? NVIDIA needs to get more of these labs or these offtakers locked into the NVIDIA ecosystem. So I think it's more than reasonable at the moment.

Yeah. And one thing I want to add as well on your question—drawdown financing—is that one comment Elon made on these earnings that I thought was fascinating. He said, “I'm going to try and be at 20 GW,” or whatever, but he specifically said gigawatts of power and cooling, right? What that tells me is that what he's going to do first and foremost is build data centers. When you have the data center built and it's extremely close to delivery for the end customer, everything becomes much easier.

And so, buying the GPUs at that point, you start to have many more financing options because you know you can bully your customer into, “Hey, maybe it’s a 6-month deal. Maybe it’s going to be 40 mil.” But there are so many more options when you have the data center built and it’s so close to delivery. 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 on the supply chain to ensure he’s going to have enough chips, and it could fall short; he said that explicitly. He thinks it could fall short as well, but he’s going to try to make sure that he has enough data centers. That’s going to be the number 1 priority: build them, and then financing and buying the GPUs can come easier once you have them built.

So, building 10 gigawatts of data centers, it’s still a lot of money. Still a lot of money, but SpaceX has cash, and again, they have decent operating cash flow.

Yeah. Well, 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 having the optionality on who to sell it to, as opposed to having an empty data center that somebody else can bring their own chips to, which are on their own balance sheet. That’s probably not as big on the leverage side, in terms of sales, but anybody who is asking the very natural question of how they’re going to be able to pay for this much capex, I think 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 cash-flow-generating business. He’s one of, if not the best guy in the entire world at raising money. So, they’re going to go for it, man. It’s inspiring to see. It’s an incredible thing to be a part of and to witness as we see them work through this.

Where do we go from here, guys? What do you think the next step is that people are going to take? What are people going to be looking for on the next SpaceX earnings call, or from Microsoft, or from NVIDIA, or anybody else that we’ve mentioned so far?

8. Bull vs Bear

Yeah, I think it really comes down to how much people believe in these ultimate economics. I think it’s starting to happen. More and more, we’re going to see toward the end of the year that, yes, indeed, selling tokens—especially frontier AI tokens—is insanely profitable, and so profitable that it’s just going to drive the industry to do all kinds of creative things to try to bank on this. So I think 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 megawatt 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 going to want to do the same thing. If anything, the question should be asked of Microsoft: You can make that amount of money; why aren’t you doing it? What are you doing about it?

I also think somebody commented, or somebody on Twitter replied to one of the comments, saying, “You guys have $12 million per megawatt for the CoreWeave deals, but you’re assuming $40 to $50 million per megawatt for the SpaceX deals.” Obviously, these were the deals. But I think what’s kind of interesting to me is that, in my humble data center opinion—and not the tokenomics guy—I think this is kind of overdue, right?

I remember at one point we ran some analysis on, I think it was 4.6 or 4.8, for a GB300, and somebody found it was like 90% to 95% margins. I don’t know if we maybe got that number wrong. I mean, 85% is still insanely high, but in any case, reading these numbers was always, to me, like, well, somebody has to capture more value somewhere else, right? Why are we selling these GPU hours for so cheap if they’re making 90% margins?

To me, this always made a bit of sense, and they weren’t ever going to keep these margins forever. 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. I think it makes a lot of sense, honestly. What’s going on?

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

9. Hugging Face Security Scare & Wrap-Up

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 release about the security incident with Hugging Face and how they had this talk at Black Hat Summit from their Astra model, which is in testing right now. I’d say Astra seems comparable—or certainly comparable—to Mythos in terms of the Project Glasswing sort of cybersecurity concerns. This is the one where you commented in Slack, saying it actually scared you, right?

I mean, I’ve been scared for a while, but this is like 2 separate organizations identifying 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 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, if we are making the steelman case for the people who think Elon can’t do this, 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 case where this doesn’t work, it’s more political than technical or operations-related. And we’re yet to see that.

On the other side, the creative part of my mind is going, well, there are lots of other use cases outside of coding and security that we can pursue in order to give people a lot of tokens and keep models going. I mean, everybody’s focused on coding right now. What about everything else? When you point the GPUs and the researchers toward drug discovery, materials science, weather prediction, video generation, and robotics, right? If you point the GPUs and the research effort toward something other than cybersecurity, I think we will realize a lot of benefits there. But, man, this cybersecurity stuff is really concerning right now. It’s really, really freaking scary, man.

Good, yeah. A nice positive note to end it on there, but, yeah.

For what it’s worth, I’m 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? A zero-day is like a publicly disclosed CVE, meaning a security issue in 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 that it works on the system and, in some cases, making minor modifications to make sure it works.

This is something an agent can do. It seems kind of obvious an agent can do this from everybody who’s had experience doing research with these things. Providers out there that are not SpaceX, but are on the lower tier—there are plenty of them—are running stuff not with 6-week-old zero-days, but with 3-year-old zero-days that we found in a second just by checking the version of software they’re running.

The ability to develop PoC exploits to show them that this is the potential concern took us afternoons, not weeks or months of effort, and we didn’t need to be a security expert or a Linux kernel expert or an NVIDIA GPU driver expert or a Kubernetes expert. You just say, “Hey, model, check for this version.” It checks for it, and then it can build an exploit in a couple of hours.

So anyway, that’s when 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 them; they went and did this by themselves because they were pursuing a goal of trying to find a data set to pass their eval. So they went out and hacked Hugging Face, which hosts these data sets, by finding issues in the HDF5 data format. They found a zero-day in the HDF5 data format on Hugging Face. It’s unbelievable.

So, maybe—because I tried to explain it—I’ll give people another nugget so 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. 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 file server. They couldn’t even access the contents of the files; they could only write the names.

That’s absurd. That is absurd.

Yeah, you should go watch this video. It’s 30 minutes and well worth it if you want to understand where we are right now in AI progress at the frontier. It’s a glimpse into what people like Elon, who see Grok training and talk to Anthropic and OpenAI, are seeing with these models that they hold internally and don’t release publicly. Even the stuff that’s public is so unbelievably profitable that they don’t even need to release these research projects that are being evaluated.

We’re seeing rapid progress right now in 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 at the article.

Yeah, really. I was wondering how to change the topic back. It’s 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.

We got a lot of hate when we had our call on Amazon, and we were like, “Guys, just look at these data centers. They’re going to accelerate revenue pretty obviously.” They were like, “No, you’re idiots. Never going to accelerate. They’re losers of AI.” So right now, SpaceX is the loser of AI.

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

Looking forward.

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

All right. Good job, guys. Take care. Bye-bye.