Stanford CS153 Frontier Systems | Scott Nolan from General Matter on Energy Bottlenecks

1. SPSpeaker0:09

   We are super lucky to have with us today to talk about energy bottlenecks, Scott Nolan. Welcome, Scott.

2. SNScott Nolan0:14

   Thanks. [audience applauding] Yeah, thanks for having me. Excited to be here.

3. SPSpeaker0:20

   So if you remember, we started the class by talking about how we're going through a great transition, right? So we have the old system stack that is transitioning to the new system stack. And to go back to our organizing mental model and metaphor for the class of the AI factory, you guys remember this, right? This is how intelligence is being manufactured, the frontier, pre-training, mid-training, post-training, deploy to agents, rinse and repeat. And for the last few weeks, as you guys have heard from folks like, um, you know, Matty at ElevenLabs, Robin, uh, sorry, Andy at Black Forest Labs, and Amit at Luma, right? Those are all different types of intelligence that are being figured out in the field right now. Today, we're gonna zoom out a little bit because sometimes it can get easy to forget that what's happening at AI labs, sure, it's exciting, right? We're getting new capabilities that have never been possible before, and that is what's driving so much growth in the industry right now, and excitement and revenue and all of that. Um, but that's just one part of what's going on because to deliver new capabilities to the world, it takes a number of things to come together, right? And we talked about how there are, um, some major bottlenecks on that progress of capabilities, and one of them, as we've talked about before, is compute. But the point of this class is to try to give you sort of a macro systems view of what's going on in the world, not just in model labs, but up and down the stack, okay? And I find the stack, the whole idea of a stack even is quite rigid sometimes because it, it kinda pres- presents this view of, you know, how things work when ul- ultimately it's just one type of mental model and, and scaffolding. And as you know, we're, we've been talking about a different kind of mental model and scaffolding, which is a, a frontier AI pipeline. And what I'd like to do is zoom out a little bit now, okay? This is my, uh, handiwork. I'm not a professional artist, um, but this is my, uh, my attempt to try and be a little, uh, a- as close to, uh, Hayao Miyazaki as I can be. I grew up watching, uh, a bunch of Studio Ghibli movies and, um, this is sort of a stylized mock-up of what I think, you know, is, is a systems-level view of how these capabilities factories are working. And so if you look, right, uh, right at the center of the factory, you've got the pipeline, right? Data, compute algorithms, pre-training, foundation models, mid-training, and so on. But to make that work, it takes a whole other bunch of systems to come together. Now, if you look at the right side of the factory, you've got a little box there that I call, uh, that, that-- th- think about that as analogous to the data center that's providing compute. You know, and we talked about in lecture one how compute is super critical. It's important, but remember, that's just one bottleneck. Sometimes what gets lost in the conversation is that powering the data centers is a whole other important thing called energy and electricity. And to keep your compute running on time, well, somebody's gotta power the data center. And we are going through, well, I, I would say we've been now in four years of relentless pressure on that part of the supply chain 'cause after ChatGPT came out in late 2022, um, you know, that turned out to be this, this-- So for a long time before ChatGPT came out and scaling laws had been discovered, big question on everybody's mind was, "What is this stuff? What is AI going to be useful for?" I mean, it's cool technology, but really, what, how is it gonna change the world? And ChatGPT, I would say, was the first sort of consumer killer app, right? It became this, this way to consume the technology of, um, language models that were, that were legible to everyday people. But the supply chain wasn't ready for that. You know, it takes, like, two years to tape out chips and stand up data centers. And so in early 2023, a few months after ChatGPT came out, there was a huge compute crunch and, for a short window of time, also a huge energy crunch. At that moment in time, a bunch of us in the industry who were paying attention to what was going on started realizing, "Wait, if this continues, we're not gonna be able to keep the progress going." Because at some point in the future, cool, we had a consumer killer app now with ChatGPT, but at some point somebody's gonna figure out an enterprise killer app. You know, so- some tool or way to use this technology that's useful to enterprises and businesses, and that's what happened, right? What, what happened in December 2025, a few months ago? Claude 4.6 came out. Anyone remember that? How many of you were coding over the we- over the sum, uh, winter break? Yeah. Did it feel different? Right? And then all the adults, well, you guys were students, so you had some free time, but all the adults who were, you know, on, on parent duty came back from winter break and started using Claude at work and stuff started to change, right? 'Cause now suddenly you've got enterprises and businesses going, "Hey, this is really useful. We want more of this stuff."And that was a Groundhog Day moment. Um, now for me it wasn't that surprising, 'cause as you know, four years ago when, when ChatGPT came out, that's when I realized compute was gonna be a bottleneck. I started working on a version of, uh, trying to unlock, unblock that bottleneck at a16z. But elsewhere in the industry, there's a guy called Scott [chuckles] who was realizing that energy was gonna go through a similar problem, because if you just keep going down the supply chain, you realize that that's gonna be a huge bottleneck. And so the reason I have the electricity part of this map so much bigger than the data center map is actually from a urgency perspective, even if you have a data center ready to go, if you can't get power to it, doesn't matter. It, it's over. You can't train models. And so, um, for this class, we wanted to make sure you got a, a view into that part of the world as well. And so for the next, um, few minutes, we're gonna get a, a sort of an expert deep dive into this part of the factory, right? Uh, i- ideally all of this is just happening in one place, and at some point in the future maybe we can have modular data centers, like on campus next to the lab with a modular reactor or something. We're not there yet, and so instead we have data centers in one part of the world and e- energy, power generation in other parts of the world. Um, so this is a glorified, sort of idealized utopia schematic here. Um, but for this lecture, we're gonna be zooming in to energy. So with that, um, why don't we start with you, Scott? Thank you for coming. Tell us about yourself. How did you get here? [chuckles]

4. SNScott Nolan7:41

   Cool. Yeah, thanks for having me. Um, so Scott Nolan, CEO of General Matter. We are a uranium enrichment company for nuclear energy. I started off as an engineer. I was mechanical, uh, undergrad, aerospace masters. Uh, that was Cornell, and ended up coming to Stanford for, uh, a second master's degree in business of all things. Um, but sat in on a lot of engineering classes. They were pretty much all of my electives, uh, including some CS classes. So, uh, did that, wrapped that up in 2011, joined Founders Fund, the VC firm. Was there for over a decade, uh, just fully focused on anything hard tech, anything engineering, technology-driven, um, and that included energy. And so [clears throat] you know, one, one part of energy I had always been interested in was nuclear. It always felt like this branch of energy production that just had gotten completely forgotten and sidelined as, you know, it being a massive mistake for like the past 50 years. And for the, for that decade, I would meet with so many different nuclear companies, and by 2020, there were starting to be some pretty interesting ones. Um, but they all said the same thing, that they had no fuel, and that they had to get their fuel from Russia, which was really shocking to me. Um, I dug into it for the better part of 2023 and realized it was all because of this one missing step, which is what we're working on. So we'll, we'll get into it, but for today's class, I think the interesting thing is just this energy topic in general, how much of a bottleneck is this? How do we solve it? And we'll go through that.

5. SPSpeaker9:11

   Yeah, sounds great.

6. SNScott Nolan9:12

   So we're gonna, you know, you don't have to take my word for it. We're gonna start with three pretty smart people, um, who talk about energy a bunch as bottlenecks to their, to their businesses. So first one is, um, is Sam from, from OpenAI, and this is him testifying to the Senate. So this is, uh, you know, you, you can't, you can't-- You have to tell the truth when you testify to the Senate, so you know this is true. Um, so everything is going to converge to the cost of energy, to the cost of electricity. Like, chips are gonna get cheaper, um, models are gonna get cheaper, but energy is fundamentally what you consume when you're running these models. And, you know, one version of this is Balaji, uh, who was a Stanford professor at, you know, for a time also, has argued that everything, all costs, all, you know, monetary things should be denominated in joules. Um, and so this gets back to the same sort of, same sort of thing. And then, you know, then you, then you think about Jensen, and probably his, his incentive should be to say that chips are the bottleneck or that something about what he's doing is the bottleneck. But even he would argue or admit on, you know, on the Joe Rogan podcast that energy is actually the bottleneck. So that's pretty powerful. And then, you know, you go to, you go to Elon and, um, there's, you know, many bottlenecks that he could talk about, but the one that he wants to highlight is, is energy. And I think you're seeing this now in, in some of the plans with SpaceX. Um, and so I guess I left that out of my background. I was an engineer. I worked at SpaceX right out of school, um, and then did, did a bunch of other stuff before Founders Fund. So, um-

7. SPSpeaker10:51

   That was before you came back for grad school, right?

8. SNScott Nolan10:53

   Before, yeah. I did everything backwards basically, but it's-

9. SPSpeaker10:56

   Better late than never, man

10. SNScott Nolan10:57

    ... it's okay. Um, and then, you know, and then it's like, okay, well, these are people at the very forefront of data centers, of, of the models, thinking about what's coming next. Um, but then you go mainstream and you realize, well, even the "Financial Times" is realizing this. They're realizing that actually what's upstream of data centers and all the compute is power, and you really need power, and then where are we going to get it from? Um, and we'll have some time after these slides to talk about some theories about this. Um, [clears throat] but you, you then might ask, "Well, okay, how big a problem is this really? Is this really like something that we can easily tackle?" You know, you mention, "Okay, let's talk about energy, electricity." It just sounds to most people like so unexciting and boring and, "Oh, it's big metal wires and infrastructure, and why do we care about this? Certainly someone has this solved. How could that possibly be a problem? We've been doing it for 100 years." Um, and you know, but then you look at the demand and you realize, wait, this is like way super linear, and how are we actually gonna keep up with this? And then you say, "Well, okay, you know, maybe it gets to a terawatt, but, you know, in a decade it gets to a terawatt. That's, that's pretty fast, but maybe we can keep up with that. Maybe it's not so hard."And then you look at what we've actually done over the past, you know, in this chart, you've got over 50 years X-axis, and you look at, like, the last 20 years, and you realize, "Wait, we haven't done much of anything." And in fact, like, one terawatt's kind of a problem based on what we've been doing. Um, we need to be much more on a China-like slope, where you look at the, the yellow portion of this, um, and you realize, "Wait, we need to be on a very different slope even than China." And so we have to go from almost a complete standstill on grid expansion to nearly vertical. And so that's gonna require some very different activities than what we've been doing as a country for a long time, for longer than pretty much anyone in this room has been around. So I think with that, you, you quickly realize, you know, okay, it does seem like maybe electricity is the bottleneck to AI. Maybe, maybe Jensen and Sam and Elon are all on the same page because this is so overwhelmingly obvious that you have to solve this. And then you would say, "Well, okay, how are we going to solve this? This is clearly a big problem. Um, you know, we haven't done much. How do we, how do we go really quickly on, on ramping production?" And if you rewind to, like, five years ago, um, stranded energy was enough. Uh, so there's plenty-

11. SPSpeaker13:32

    Can you define stranded energy?

12. SNScott Nolan13:33

    Yeah. So there's plenty of stranded energy, and stranded energy would be things like, you know, a hydroelectric dam in some s- you know, rural region that there's no population nearby really consuming it, or maybe it's geothermal, um, isolated geothermal with existing technologies that no community consume it, or stranded wind in West Texas. Um, the list goes on and on. But an- anything like that, something where there's supply without real demand. And so what you saw late 2010s, early 2020s, that was completely dominated by companies that said, "Okay, I see that stranded power. I'm gonna go build something there." The very first builds that happened were typically Bitcoin, uh, Bitcoin mining centers.

13. SPSpeaker14:14

    Yep.

14. SNScott Nolan14:14

    Um, you know, you didn't have the really huge AI data center demand, but you did know that, okay, what can I do with stranded power? I can mine Bitcoin. I don't need that much connectivity. I don't need fiber. I can get by with iridium or something if it's middle of nowhere. I can get, I can get, you know, enough connectivity to actually perform that. And so you saw companies like, on the left, um, is a company called Crusoe, which now-

15. SPSpeaker14:40

    Right

16. SNScott Nolan14:40

    ... has, you know, is doing the Stargate project in West Texas, and that project is linked up with wind and natural gas and all sorts of things, some of it which was stranded. Um, and so that was, that was a great strategy for a long time. At this point, most of those great resources that were stranded without nearby demand have been claimed. People have gobbled those up. And, you know, the capacity that we need is increasing quite a bit. And so even those small chunks of electricity that were available would not even be enough today to satisfy things. And so, um, things are really moving to ask the questions of, how can we create massive net new power production? And so this was something I was starting to think about, both the stranded topic and the bigger topic, um, at Founders Fund late 2010s, early 2020s, and coincidentally, um, invested in... If the top left is Crusoe, uh, then invested in all of these companies. And so top left is a data center just like Crusoe. Um, I don't actually think it's a Crusoe one. Uh, you've got SpaceX, which is now talking about in-orbit power production, and then you've got a company called Pantalaassa doing distributed energy in, in the ocean. And so lots of different angles on this. People have different theories. We can talk about in-orbit, we can talk about other options. But today we'll talk about on land because that really dominates things, and that's what, that's the reality that we're living in. And so you say, "Okay, well, we need to produce a lot of power on land. What are, what are the constraints? What are we designing for? Um, what are the things that the data centers actually care about?" And one of the big things they care about is uptime. So, you know, data center, can you run it on solar? Can you run it on wind? You could, um, but you're gonna need a lot of batteries. And by the time you had enough batteries to get this uptime, at least as batteries exist today, grid scale, um, your cost is gonna be pretty high. And so people have, have gone away from that. Uh, what you're seeing today, the last couple years, is a lot of natural gas-powered data centers running on turbines. Turbines are getting pretty scarce. The lead time for turbines is a few years now, um, which has increased drastically, and the producers of turbines generally are not ramping production quickly enough to even remotely keep up with this. Um, and so then you say, "Okay, well, we need, we need something that's not natural gas, can be base load. Where do we look?" And you might say, "Well, okay, what are the other factors? Maybe we don't wanna put out a lot of carbon. Maybe we want it to be pretty safe." And so then you look at the historical statistics factoring in, you know, every plant that's ever been built, and you realize that, um, here's, here's the base load chart. Then you realize looking at safety and cleanliness of power source, that actually nuclear is pretty good. Um, it's actually lowest carbon emission of any of them, [chuckles] and it's essentially tied for safest with wind. Um, and so those two things together, if you care about safety or, or emissions, it's gonna push you pretty hard towards nuclear, and that's why all the hyperscalers are, are looking to that. I think they all realize nuclear is not gonna be something where you build a plant overnight. It's not a one-year project. It's something that we're gonna see ramping in the next five to 10 years, truly ramping and moving the needle. Until then, it's kind of a, a race. Who can find stranded power? Who can find enough turbines? Who can maybe stand up solar with enough battery storage if they're less cost-sensitive? Um, but long-term, everyone's looking to nuclear. And so then, then you say, "Well, okay, well, if nuclear's the long-term, you know, scaling limiter to electricity, five to 10-year timeframe-"And electricity is the bottleneck to AI, then you probably realize, well, that's kind of unexpected, but maybe nuclear is actually the bottleneck to AI scaling, um, if you're talking about here on land at least. And so then you might ask the third question, "Well, okay, is there a bottleneck to nuclear?" Uh, which brings us to what we're working on, and every nuclear reactor runs on fuel. I think a lot of people hear nuclear and you would think it's a magical technology. It's like a perpetual motion machine. But no, you actually need to refuel it every, every year or two, depending on the reactor. For more advanced reactors, there's some that design for five to 10 year refueling cycles. But, um, it does require constant fuel, and it constantly burns up fuel, just like any other type of engine. And that fuel comes from five different steps. You s- you start by mining, you turn it into a gas, you enrich it, uh, you turn it back into a solid, and then you make your fuel pellet. And you might then think, just like electricity, "Well, may- this is a solved problem. What could be the issue?" Um, but you actually look at these five steps, and it turns out that the U.S. has less than 0.1% market share today of enrichment, which is the middle step. And so the U.S. is actually unable to produce its own nuclear fuel at any scale whatsoever, and we rely completely on European firms and even to this day, Russia. Um, even though there's sanctions, we still, we still import, uh, because we really need to. And, um, you know, so there's this missing piece right in the middle. And so we can't really scale nuclear fuel as a country, which means we can't really scale nuclear, and which will mean that we can't scale, uh, data centers and AI. Um, and you know, if scaling is one thing, cost is another. At some point, the cost will matter a lot. People will start being more price sensitive. It won't just be an arms race for who can stand up a data center the fastest. There'll be margin compression. Cost will matter. But in fact, cost is the biggest, you know, of the cost of, uh, advanced nuclear fuel is the biggest cost in many cases, and the biggest portion of that is actually enrichment, which is why we're working on it. And so you, you know, you do the build one more time, and you realize enrichment is kind of the bottleneck all the way through to, to AI on a, on a five-year timeframe. And so that's why we're almost in a race against time at General Matter, going as quickly as we possibly can to bring enrichment back online in the U.S. at scale, like with a highly scalable method that we think can completely win on cost. Um, and we're getting a lot of support from that. So when we started the company, there was, uh, no ban on Russian uranium. There was no AI data center boom. Um, it, it was under Biden administration. We started by working on advanced fuel for advanced reactors, and that was a big push. And then now this administration is very focused on energy production, and there's kind of follow through on that push, um, across administrations. And bottom right, you can see our August groundbreaking of our facility in Kentucky. Um, in that image, there's people from, like the full range of political spectrum all getting together around this. And so going from the very beginning, you can see the tech leaders are realizing energy's the bottleneck all the way to DC. Everyone realizes that energy is k- upstream of everything, not just AI, but also manufacturing and pretty much any in, any industry that you want does rely on it. And, uh, the current state of it in the U.S. is, is far worse and far less ready to scale than a lot of people realize. So that's what we're working on.

17. SPSpeaker21:49

    Th-thank you, Scott. Uh, why don't we take a beat there because you said a few things I wanna kind of double-click on. Scott mentioned Bitcoin mining, [chuckles] and he's me- you know, sort of mentioned that in passing. But, uh, the reason I wanna zoom in on that is because, um, you know, sometimes the cultural commentary around a piece of technology can often make the underlying progress that's quite real and, you know, clearly sort of fundamental, um, confusing. Um, you know, nominally from a memes perspective, you know, the way this manifests itself on the timeline and so on is people saying SBF funded, you know, Anthropic. SBF was running FTX at the time. Um, like, you know, the fact that people in the crypto community were investing in the AI stuff, you know, you can-- we, we, we can disagree on whether crypto ended up delivering on its promises or not, but what we have to acknowledge is that, um, you know, c- Bitcoin mining was, was a bit of a dress rehearsal for AI.

18. SNScott Nolan22:53

    Mm-hmm.

19. SPSpeaker22:54

    Um, and sometimes, you know, I, I, I get all these questions where there'll be a, a, a team working on a pretty important fundamental bottleneck at the infrastructure layer, but just because they've raised money or something, or have done some political donation or something with somebody from the crypto community, their underlying progress just gets thrown out. You know, it's like a baby, throw out the baby with the bathwater sort of moment because people go, "Oh, if, if crypto people are involved or Bitcoin mining is involved because that didn't work out." And yeah, right, who knows? At, at some point we may have this decentralized, you know, uh, sort of, um, censorship resistance and so on.

20. SNScott Nolan23:29

    Mm-hmm.

21. SPSpeaker23:29

    Uh, and then we, we will have truly decentralized computing. Uh, these things take a long time, but the, f-f-from a first principles perspective, I, I find it quite sad and, and disappointing when people aren't able to decouple those two things. Um, you know, uh, you, you m- you mentioned Crusoe as an example of, of a company that's been, that you've worked with before.

22. SNScott Nolan23:50

    Mm-hmm.

23. SPSpeaker23:51

    Um, I think Crusoe was originally a Bitcoin mining-

24. SNScott Nolan23:54

    Yep

25. SPSpeaker23:54

    ... company.

26. SNScott Nolan23:54

    Yep.

27. SPSpeaker23:54

    Right? And then they, they sort of, some of the, many of the innovations that they ended up realizing, uh, during the Bitcoin era have ended up translating into, to, uh, building infrastructure for the AI era. Um, and I, and I think those learnings have ended up becoming valuable. Um, you know, venture capitalists sometimes like to call these evolutions pivots, and I think there's a-... unnecessary stigma around that, when in fact pivots are just one step of the continuous feedback loop we've talked about before, in my view, right? Um, and it, it's an update, and the best leaders update their priors. Similarly, as, as you've been approaching the sort of en- energy, um, discussion, you know, nuclear is, is another one of these areas that ha- has been unfortunately, uh, plagued by a bunch of confusion, politics, you know, s- social divisiveness. Um, w- what advice would you have for people here who are, who, who believe that, uh, you know, they, they'd like to work on energy, they, they'd like to work on nuclear, but for whatever reason feels like because of the past, um, uh, political climates or social objections to that technology, uh, the fundamental progress is not legible. Am I making sense? Is my question making sense?

28. SNScott Nolan25:12

    Yeah. Yeah, yeah. I mean, you can even go back to that chart, the emissions and the safety track record of nuclear, and then what we'll talk about nuclear. But going back to what you were saying before of these pivots, I think, you know, pivots is one thing, but I would say if a company is building something that you can think of more like a primitive-

29. SPSpeaker25:29

    Right

30. SNScott Nolan25:29

    ... like a fundamental building block, which you might say, well, utilization of stranded electricity feels like a primitive. And yes, what we might do with it today is mine Bitcoin.

Episode duration: 1:00:27