BẢN CHÉP ÂM THANH 🔗
Konstantin Saifoulline: my guest today is Ned Bullard. He spent fifteen years at Bloomberg NEF and rose to chief content officer. today we are talking about the physical bottlenecks behind the AI burlout and as a first question I would be interested in what you think what actually stops us right now from scaling data centers to ten X.
Nat: Well the s the simplest answer to that, Constantine, thank you for having me. pleasure to be here. and I would say that the simplest answer is that while you can manufacture the sort of primary units of compute through their own stack that you probably know very well and your customers know pretty well, everything else depends on a series of other supply chains, whether that's power, equipment, or even sort of basic componentry that would, you know, be within the data center itself. And then you have all of the supply chains that go into delivering all of those sorts of things. And then you've got labor, and then you have planning and permission. And then fin finally, you have power. none of these things exist in a vacuum. And power is to my mind in particular the the absolutely essential step. If you have a powered shell with no power, you have nothing. And so the biggest challenge is that. You need to both provide power of a certain quality and in a certain volume. you need to provide it as quickly as possible. and you need to do it in a way that's got really, really, really high reliability. as high a reliability as almost anything else that we have within our system. And then you do still have to do it with some modicum of economics in mind. I know that training models in particular is relatively less sensitive to the cost of power than many other activities in the economy. But it's not zero sensitivity. So you want to do these things as well in a way where the economics actually make sense, where you're doing something that isn't a penalty on cost. So a lot of factors, but I would say power is sort of most important among them. And then secondarily, all of these other little things that need to be manufactured and deployed sometimes at scale, sometimes specifically and with a lot of mass to make to make power happen.
Konstantin Saifoulline: Mm-hmm. How would you rank the limits for that? So would you say powers on the one and then where comes crit connection, chips, cooling, labor?
Nat: This is this is a good this is a good question and and I the part that I know best is power. Like I would say that that is the most essential. Like if you have everything else and you're not energized, you have nothing. Right. So if you have if you have everything built but you're unable to provide power, then you have no functionality whatsoever. So I would put that sort of first and primary amongst things that you need to have. However, you have to sort of build that in parallel. It's also, I suppose, possible to have a scenario wherein you have power but the rest of things haven't happened. You know, the rest of everything you need to have delivered in on site is not there. Or you don't have you you just don't have key components that you want or that you really, really need in order to be there. So I tend to view this through the lens of power. It's the sector that I know best, it's the sector that I cover that I'm building a business around. But there is of course a whole universe of other things you have to be aware of.
Konstantin Saifoulline: Mm-hmm. Then let's go a bit inside there and what what the main issues are in power that we kind of need to solve right now
Nat: at what timeline do you think we will hit a limit? So what counts as
Konstantin Saifoulline: Yeah.
Nat: a limit? First of all, we have a limit right now in that we have we have demand in excess of supply. So I would say that we're already at a limit, right? the question is how much of that gap can we solve? And even better question, I think, for both of us is that like, are we going to induce more demand? Let's say that we magically then did hit ten X power provision. Would we also have ten X demand from computer that's same time? The answer I would say is probably yes. But what what what keeps us from doing what we want to do are a couple of things. First is simply planning. So are you going to get approval for a project that you want to do to build power? And that's approval at multiple levels. One of them is at the state level, some of them are at federal you have approvals for things such as emissions, air quality, water consumption, things like that. then you have, you know, the city and the municipality which has a say in whether or not it wants something built as well as the community. another thing is if you're moving power across a long distance, do you have the capacity to interconnect generation at one point? with transmission, distribution, and then delivery and consumption on another. This is one of the reasons why you have so many people trying to build power infrastructure, what they call, you know, behind the fence or behind the meter. Essentially trying to do it yourself rather than hand it back to the grid and have the grid do it. And this is different. Right? Historically we have devolved responsibility for delivering power to the grid unless you are really enormous industrial demand. So something Something like a refinery, for instance, or a very, very big mine, or a really, really big manufacturing facility might have done some of its own power provision in the past. But generally speaking, we cast this forth for good reason into this enormous machine that is the grid. You buy power from the utility, the utility provides it to you. at a at a cost and with a a sort of assumed reliability and stated reliability over time that you can look at it and be like, okay, great. So you've got planning on one side, you've got interconnection on another. You have fuel supply, which is relevant if you are doing a fueled source of power. So basically anything non-renewable,
Load full transcript
meaning are you gonna provide natural gas? Are you gonna truck in or pipe in fuel for combustion to any of these sites? And this is another thing where you have to get approval at multiple levels. Communities rightly I think have a say in whether or not they want a 30-inch gas pipeline running through them, or if they're prepared to do truck movements for days, weeks, months, maybe years worth of delivering power on site. And then finally, you have generation equipment itself. You have to go out and buy the equipment that's going to convert molecules via combustion into electrons. And that's everything from a combined cycle gas plant that has been of the sort deployed in power now for decades, or sometimes that you have what's called a simple cycle turbine, so just one thermal combustion process. Sometimes you even have what are called reciprocating engines, which traditionally have been used for things like backup power or really, really remote power applications, or even like a marine application in the ship. And we're now seeing all of that stuff come together at once in ways that they've Not been deployed before. Rarely would you find a case where somebody would want to do hundreds of essentially gigantic diesel engines, the reciprocating engines, to do power for longer than, you know, a few hours of backup. Now people are planning to run these things pretty much continuously. But you get an indication here of all these steps, and you have to do them all at once. Like you can't not do a step along the way. You have to do all of these things if you want to get power. And without that, again, your data center is nothing.
Konstantin Saifoulline: Hm. Then what I right now the public steps that you see open AI and Thic SpaceX AI taking in order to solve solve the power problem for them?
Nat: So, you know, the the SpaceX thing I think is really instructive because it basically broke a mold that we've had it for a very long time, which is that you use the grid and you ask the grid to provide power for you. in exchange you have to sort of accept the timeline that you've been given. you might be able to accelerate it by being really good at what you do. you might actually, you know, even use renewable power 'cause that might move quicker. You back it up with batteries and whatnot. But that sort of broke a mold. And to be honest, the mold that it broke was one that ran through a lot of steps of planning and also a lot of steps of approval. And I would argue that building things as quickly as possible and running them with no with no conditions, with no emissions controls is probably not tenable in the long run. And in fact, most of these places are changing that. The other moves are to to take what traditionally had been kind of like full on power plant development and bring it in-house and behind the fence or at least connected specifically to an asset. So Meta, for instance, has done this at a number of different sites, wherein a couple of years ago,
again, you would have the grid do this work for you. Now the approach is more to actually just do it yourself. And by do it yourself, that doesn't mean that Meta has stood up a construction company that's going to go build power infrastructure. You're contracting that out, but you're doing it directly in coordination with your own efforts. Like it's devoted to you, even if you work for or work with the utility along the way, you're doing it for your own provision and you're trying to do it in a way that means you get it as quickly as possible and that it's effectively exclusive. Like the power that you're putting together is meant for you and only you, as opposed to when a utility builds an asset, it's building that power to then pre be provided to the entirety of its system. It could be millions of customers at once. This time you're just building it by yourself, effectively in that broad sense, and you're building it for yourself. Sometimes it's within the site, you know, the hundreds or thousands of acres of area that is walled off for at data center. Sometimes it's nearby. Either way, you're doing it just really for yourself, by yourself and exclusively for your own use.
Konstantin Saifoulline: Hm. Have have we seen such move like Meta did it by OpenAI and Entrophic? Have they announced similar public things?
Nat: I mean, they're all moving they're all moving in parallel. A lot of what you see is these places moving towards an asset that might have already acquired power and going as quickly as they can. Meta's a meta's movement has been working directly with some pipeline companies that i have in turn their own power provision, company called Williams here, that are sort of jointly developing assets. Anthropic has generally been going for assets that are already under development somewhere else. there's another thing that that companies do and I think is important is that you apply capital. You you apply capital to solve a problem and that might mean going where a development is already underway and taking part in it. Like you go to a neo cloud or to a powered shell. You can you can also simply look at areas and all of the best of the hyperscalers do this where there might be an availability already within the planning in the grid, right? Places where you've been working for years to get with the utility and say, We expect to be here. We're part we're comfortable paying a lot of our own way for the infrastructure related to power, but we'd like to buy that power from the grid. And to be honest, that's the part that's probably in the long run the most durable. And the reason I say that, I keep coming back to this, is because that means that the reliability that you're able to get from the grid is both quite high, generally speaking, and it's also a responsibility that you can pass to the sort of collective entity that is the power system, as opposed to just doing it yourself. I mean, the you got two things happening at once. You move really quickly to build your own power, it's fantastic. You You both have full ownership of the power, but you also have full ownership of uptime. You are now responsible for making sure that things are working together, however many nines reliability you want. If you are to get power from the grid, then that is the grid's job, and you're providing backup for yourself. And you can see this distinction when you read all of the planning that these companies do. So on one side, you'll have something that says We're doing backup power, right? So let's say I'm reading about a data center in Indiana or Ohio and they're listing hundreds of these reciprocating engines that are traditionally used for backup power. Some of them will say that this is what they call is you know purely backup standby or emergency. So they're getting power from the grid, and that's there in the event that something happens. That's very traditional. The next would be what they call bridge power, meaning it's really essential for me to have this data center up and running now as quickly as possible. And I'm just going to provide power to it through my own capability for as long as I need. Bridging effectively between today when there's nothing and a a an indeterminate tomorrow when I'm buying from the grid. And then there's what they call primary power. I'm building all that same stuff, but I'm going to be using it for essentially forever. I'm gonna be responsible for all of this. Building the plant, purchasing the fuel, managing the operations of that asset, my own responsibility to maintain uptime and reliability and everything like that. And it's interesting, right? Because you get the same the same physical hardware can be doing kind of three steps, three different possibilities rather, using the same stuff, using the same fundamental technical provision. To do three very different strategies, I think.
Konstantin Saifoulline: Mm-hmm. what is if we zoom a bit out the bigger economical impact of the scaling of of this lapse that we have? how will this look like kind of in five to ten years if the scale continues like this? If if it will continue like this, I don't know how how this would change, but what will we see in the word?
Nat: Are we talking power or are we talking just sort of model propagation in general? They're two different things for me. I know one side I won I have a good a good supposition on the one. On the other I'm kind of a taker. I'll tell you the power side. Okay. what I think we're gonna find within five years is a much more nuanced approach to building because people will have will will be will have done their own optimizations and realized what they do best and decide where they need to go. Time right now is so much of the essence. you and y I'm sure your listeners know the same thing that the economics of an extra d of a month of delay in getting your training center up and running is really extraordinary. It could be billions of dollars a year of of cost to
yourself if you're delaying that capability. So you really want to do it now. Let's say we have much more inference in the future. It's just much more nuanced. your scales might be different. So you may not be building just gigawatt scale. You could be sort of building within the kind of parameters of the existing grid. That would be one thing. your your interest in buying power at any cost is probably much, much lower in five years. And you're gonna be much more interested in sort of paying power at a cost that you can that you can afford. It becomes a non-marginal part of the economics of operating something. That's not the case right now. At least you know as of a couple of years ago, there was some great research from Epoch AI that the cost of training a model stock based compensation was a much higher component of that model training cost than electricity. But if you're talking about inference in sort of operations, providing service and not just training from that same fundamental compute. then I think it's gonna be very different. People will be much more sensitive to to costs and a bunch of other conditions that are there. Now, that's the power side, as I say, that I know pretty well. As far as the rest, like model propagation forever everywhere, I think that's probably a that's probably a different question.
Konstantin Saifoulline: Hm, hm. And what do you think where this power will come from? Will it be gas, will it be nuclear, or solar with batteries?
Nat: i a little bit of a little bit of all. At the moment, my company Halcyon, you know, we we we we track all of this stuff very, very closely. We see five hundred gas fired power plants under development in the US right now, as of this month. Or last month rather, September. A huge percentage of that is going directly towards data centers. it's a like I'll put it in power terms, it's more than two hundred and eighty gigawatts worth of power, which that that's under development from gas. That's about half of the all the capacity that already exists for gas-fired power in the US right now. And of that, 40 gigawatts of it is is expected behind the meter, like it's being done, you know, behind the meter for specifically for data centers. That moves quick. That's Some of the biggest and best established of the power developers that are in play. Like for instance, Intersect, which is now owned by Google, builds a little bit of gas, relatively speaking, alongside a really big solar, sometimes wind, definitely paired with batteries, renewable power provision, because the economics of that work, and that can also be done pretty pretty quickly, and you're not captive to the gas supply You're not worried about just the economics of putting deposits down on that. You also see fuel cells. So using still gas infrastructure and using methane, natural gas, as the fuel input, but without combustion and without the without the emissions. Those are also something that can move reasonably quickly, though we don't see them at the same scale as combustion. And then finally your question on nuclear. You know, my short answer to that is I do think we'll see quite a bit of this nuclear power directly devoted to data centers that's being newly built. However, not probably in the next five years. It's just a longer, slower process than building, certainly than building solar and batteries, and probably than building even a big elaborate combined cycle gas plan. And I think what's important there is you do see a lot of nuclear attention that is actually happening right now that is very, very relevant, but it takes a much more nuanced form. It's a restart of a nuclear plant that's effectively been mothballed, that's been sort of kept off market for a couple of years, with one customer in mind. That's usually a hyperscaler. And then you also see what's called up rating, which is simply going in and making mechanical improvements to the existing assets to get more power out of them that will then be devoted or sold directly to a data center. So those are those are like fixes within the system that exists already, which is different than saying a newly built green field did not previously exist, nuclear power facility is gonna be up and running by the year twenty twenty nine.
Konstantin Saifoulline: Mm. And well well that's that's interesting. w what's what's going to be the the safest power source of them long term? I mean there's also no some things moving into space and a lot of startups that build things for space and well in space get said energy is free. Well what is not completely true, but yeah there's
Nat: No, it's not. It we it's never it's never free. it's never
Konstantin Saifoulline: Yeah.
Nat: free. And it's funny because it was the nuclear industry that said, you know, decades ago that eventually the our power will be, quote, too cheap to meter. So I would I would caution anybody who says it's ever going to be free. remember there's energy requ there, you know, lifting costs and getting things to orbit is essentially energy of a different form. so I I would I'm I'm always I'm always careful to watch the economics of doing that very quickly. When you say safe, tell me what you mean by s what do you mean by safe? Like in terms of the safest way. Tell me what you mean by that and then I can give you a good answer.
Konstantin Saifoulline: Yeah. So I I mean with that w what will have the most long term impact so that it can be used in the most long term way. So if we think like in twenty years that it will be still there.
Nat: This I'm like, okay. That that's a great that's a great way to phrase it actually and helpful. Thank you. So if you want to go for sort of the most future-proofing, it's probably a renewable asset where you can think about it as an an embedded option. Like you will be able to go repower that over time. As renewable technologies improve, you can take a site that is, you know, a thousand acres. 400 hectares worth of renewable
power. And in a number of of years, maybe in 10 years, technologies will have improved that you could replace the hardware that's on that site and have something that gives you more power and more storage of that, more energy storage over time. That to me feels like the one that has the most, the most embedded optionality. So that's one thing. The thing that you know will be able to operate reliably over time, if done properly, is a combined cycle gas plant. They are designed to run, you know, sometimes up to like an 80% utilization or higher. So basically running almost all the time for decades. Like that is what they're built to do. So if if that's your your your prime concern that you just want to have that power provision over a long time, that's one. Nuclear obviously is capable of running for decades, very reliably and and but on a very, very known kind of timeline. Like not particularly flexible, most of the existing nuclear technologies. So you you need to sort of like think about what you want to optimize for. If you want to optimize for options, like taking advantage of what's coming new and new economics, then a renewable asset's probably going to be the thing that gives you the most optionality. And probably the lowest cost right now in terms of the power. If you're looking at something that's going to sort of meet this very particular demand profile, then you would probably want to do or probably be considering rather gas-fired power without anything else. N not doesn't necessarily have an energy storage asset attached, does not have the ability to change fuels over time. One thing that you might want to consider, depending on your view on regulation over time. or the social license to operate, if you will, would be whether or not you're going to capture emissions at some point, whether or not you're going to try to get as close to neutral emissions or zero zero net emissions from your project as you possibly can. And then if you want to, you know, be planning for the very, very long future and you have the appetite for it, then perhaps there's a there's a case to be made for for doing a nuclear play. The challenge is that like at a time when everything's moving really fast, and at a time when the the reward for getting things done really quickly is very high, people are gonna try basically everything they can all at once.
Konstantin Saifoulline: Hm. Well w what does it also mean for for data centers when they get built and they try to build them as fast as possible in let's say like twelve twelve months instead of thirty six months? w w what does it take for that?
Nat: I mean, all kinds of stuff, right? You probably know better than me. I would I would wager that where you sit, you probably see this better than I do in terms of modularity for one thing. Like how much modularity can you do? How much can you do manufacturing versus assembling? Big distinction. Or rather, manufacturing versus sort of constructing, if that makes sense.
Right. You want to manufacture something somewhere, assemble it elsewhere as opposed to be constructing stuff that is sort of custom or semi-custom. on site. So that's one thing that I think is neat and I've I feel pretty pretty good about watching those kind of innovations because I think they're durable. Like when people make these improvements because there's a reason to do it, then ideally they stick with us. And I think I think to me that kind of stuff is pretty cool. Like I get a kick out of that and I'm happy to see that that that sort of thing is happening.
Konstantin Saifoulline: Hm.
Nat: The other thing to think about, and and I I'm I'm not exactly certain, mostly because of this time premium right now, it's not part of the primary condition, is going to be what people's view is of that data center over time. And this comes down to like nuanced things about the depreciation schedule for chips. Like they're on a depreciate a GPU has a depreciation schedule financially, in terms of like how you treat it from an accounting perspective. But I still think we're kind of learning what's that residual value. You know, we still have people running, we have people running GPUs that are would be considered obsolete in some cases, still running quite a bit and at really, really good economics. So the question is going to be like, are you building that data center and saying in five years I expect to be removing basically everything from here, you know, everything but the shell and completely replacing all of the compute? Or is it gonna be the case? That are like, this thing is fully depreciated, right? I've done all the financialization I can with it, but there's still demand for definitely less than cutting edge, definitely not on the frontier compute, because there's just that much demand for services at like all these all these levels. What's complicated about this, but what also I think is neat about this, is that it's sort of outside the control of any one particular player along the way. that you have to have models that are capable of using you know using older GPUs. You have to have an application that wants those models, and you have to have an economics that makes sense to do all of it. So there's every scenario from I build this 500 megawatt or one gigawatt data center for training, and I'm just gonna be running it for training forever, all the way down to this is a site to be re recomputed. I'm gonna put totally new compute into this thing in five years time because I'm gonna train again, or I'm going to do inference with the existing compute, or I'm gonna do inference with different compute, or I'm going to sell the asset to somebody else that has another plan for it. So there's like like my answer is kind of multifaceted, probably not like hugely satisfying, but I think interesting in the sense of like there's just a lot of possibility what could happen.
Konstantin Saifoulline: Hm. w one of the outputs from data centers is also low heat. I I wonder are there any applications where where this low heat can be actually used in a useful way in order to reuse some of the energy that goes out?
Nat: It's great. It's a great question. And I'm glad you mentioned because Northern Europe does this. You use that waste heat as part of like, you know, doing things like heating water for municipal purposes or doing district heat in an area. The challenge in the US is that we build these enormous things in the middle of nowhere. So you don't you don't have the same ability to say heat a public bath the same way because your data center is nowhere close to anything. But I would I would love to see that sort of innovation take up. low temperature heat is is interesting because it it's also not necessarily of like high economic highest economic value. So the higher the heat,
Konstantin Saifoulline: Hm. Yeah.
Nat: the higher the pressure, the higher the value of what you're using. Like when you're talking about the kind of heat that you could use for really big petrochemical processes or for you know doing things in say metals, that's very specific. That's really hard to do with anything besides dedicated heat heat generation and storage capabilities. But you might find a case where you can do other things with it. And this would be an interesting an interesting thing to think about. One of the great things you can do with low temperature heat, for instance, is food preparation. So, you know, are you gonna find yourself with like a canning and packing plant next door to your data center? Because it's now able to get very, very low cost or possibly even free low temperature heat. The waste coming out of that thermal system in in the data center is an input to doing other stuff. This all comes down to temperature and quality. Like in the by quality, I mean things like the steam, the steam concentration within it. How reliable is the provision? What else do you need to add to it? Can it be stored? All these sorts of things. So if it's just hot air, that can be tricky. Like there's not there's not a there's not that much of a market for it besides, you know, residential space heating. If it's something that's more approaching the way that we think about industrial heat, then there should be an industrial application. But I love this question because it's a fun experiment to think ahead. You know, take this to your clients and be like, if you were to be given this and you knew that in five years' time you'd have this kind of heat and people will rate heat the way that they rate power, they're gonna rate it in megawatt hours, they're gonna rate it with quality and in particular temperatures. Maybe there's somebody out there who'll be like, okay, I can plan alongside that. But then you've also got to think about, is that d is that durable enough? Is that data center gonna be there? Let's put it another way, Constantine. What happens if somebody comes up with an ultra efficient chipset that actually
does not reject, as they say, that much heat that doesn't kick that much heat out. Then what have you got? Then you've then you built something that went from having a free input to now need to pay for one. So I love questions like this because I think there there's not an immediate answer. There's more a kind of portfolio of probabilities that we have to consider over time. but what I've always enjoyed in in in my role now founding a company in my in my role as an analyst for a long time is asking the right kind of questions about what might come next. So I I appreciate that one. That's a good one.
Konstantin Saifoulline: Yeah. It's interesting to think about a bit of how this could play out because well we could say we we invest more in this data center and and there has been really funny concepts there that I saw on on X flying around, making days data centers really, really beautiful and like adding a bath to it. But I I yeah, but but
Nat: Yeah, yeah, you're right. Making them lovely.
Konstantin Saifoulline: but it's not I I feel like it's not that stupid. If if you could make if you create something around the data center where people would love to live and you can make it somehow that it's like not looks like a huge thing that is just like annoying or like not nice
Nat: Okay.
Konstantin Saifoulline: to be in the surrounding and and you can
Nat: Right.
Konstantin Saifoulline: implement it well then w wh wh why not? Why why should people not decide to live there and like there why should there not be positive economic outcomes of it?
Nat: I I agree. And this is this is where I think the the the the inference landscape will become interesting over time. Like do we find smaller assets closer to load where they're not making that same impact? They're not sitting on five hundred or a thousand acres of land. they're not you know, they're not as impactful in that way. The question of course is still going to be about reliability. They still need backups. They still need emergency power provision and things like that. Those can be lovely, but they also have to run. It's it my my family had a farm and we had an emergency generator set there for the same reason, effectively, that a data center has one, which is power will go out and you need to be able to keep, you know, farm operations going. Now, fortunately, we didn't have to use it that much, but you know what we did have to do is every week on a clock it fires up just to make sure that it works. So you're still going to have that. No matter how much beautification you can do, you've still got some of that. But again, questions, questions that could have good answers. Okay, if you're connected to the grid, so you're not just sort of generating your own power from reciprocating engines or a gas turbine. You have backup, but maybe you have uninterruptibility, designed not just as like thermal power generation, but as a UPS, a power source or by batteries on site. What are those gonna look like? So We've got a lot of we've got a lot of things that have potential answers, I'll put it that way, right? They have potential solves and solutions that we haven't we haven't fully explored yet because it's still early. But my hope is that people are always gonna be thinking about this opportunistically. Like to your point, I think it's I think it's more constructive to for us to think about good possibilities here. than the sort of just maintaining the status quo when that's it.
Konstantin Saifoulline: Mm. Of all the data center capacity that has been announced so far, what do you think how much percent of it will will get actually built?
Nat: So I I don't know. The challenge is also like what counts as an announcement. So one of the things I do in my capacity with Halcyon is like we're looking we're looking as early as possible for a good signal. And what I look at are things like an air permit to allow you to emit particulate emissions from from basically from running a power plant. That's a signal that I think of as this asset is going to happen. Because by that point, a lot of things have to be true. You have to know where it's going to be. Like you have to have a site. You're not just doing it in like in the middle of nowhere. Two, you have to know what that is. Like I'm going to operate X number of generating units, Y number of hours per year equals Z emissions, both on a concentration and on an absolute basis. That to me feels real. The challenge is that like it's very easy to say you're going to do something and then not follow through with it. And this is a neutral statement on my part. That's the nature of a development business. You have a lot of things planned that may not happen at all. And it doesn't even necessarily mean that you have, quote, canceled an asset that you had early on. You just haven't advanced it through the pipeline. So what I tend to think of is by the time the power the power provision is available, by the time you've gone to that point, you are likely to have it happen. the question is how quickly does it happen compared to the timeline that you told yourself you've been planning on the timeline that you maybe had promised to investors. Not everything will get built. there are a lot of other factors that do go into this, such as capital. You could reach the point where you've done an early application and say you need to build this much power for this asset, but if you cannot finance it, then that won't happen either. So I don't I don't I don't like to go and do tight bands. Some people do they're you know, there are people who are more knowledgeable and there are people that have that do that f specifically for a living. What I like to watch is the power component. That's where my company sits. It's what I see most closely and That's the part that's the part that I think is the determining step. And the one where I can get the most resolution, I can get the most granularity. What I don't do is actually give much
credence to things that are just a press release. Until they've advanced further, I tend to think of them as potential only. and I would say very, very low potential, and then they move sort of medium, high. to different levels of conviction in my mind until finally we reach the point where I'm like, Yep, that sounds like a real thing, that's gonna happen.
Konstantin Saifoulline: Hm. how how do you see this with data centers in space? Do you think this is something that's actually coming?
Nat: I I am so fascinated to watch this because it's an area about which I confess to not know enough. I don't know enough about the science, although I do know enough to know that you know, dealing with heat in in a vacuum is actually pretty challenging. What will be interesting to me, I guess, is if I think about a the blonde instrument version of this, is just lifting cost. Right? Lifting cost has a way of solving for lots of problems. If it turns out that getting things into orbit is ex is really extremely cheap, then a lot of other problems can be solved by building and adding on to them. Right. And in fact, in this case, it's the equivalent to me of of thinking about solar energy, wherein ultra, ultra, ultra low unit cost and then cost of delivered power on the margin. allows you to spend money on doing other things, if that makes sense. So that that to me is the most determining thing. Like at a certain lifting cost, a whole bunch of stuff is doable. And that that changes the nature of does it make sense or not. I again, what purpose is that compute going to be serving up there? what's the assumed lifespan? What's the the end state of it? Does it have a repurposing or optionality to it? I think these are all very much to be determined still. And what I'm grateful for is that minds much smarter than mine are like hard at work at doing this and earnestly. So I never like to dismiss out of hand a technology where you could have some kind of constructive discontinuities in the economics of doing something.
Konstantin Saifoulline: Hm. And when it comes to Earth, who do you see there right now as the the short term and the long term winner in terms of built out for for energy? Is it the US ha has the China has China a monopoly about it? where are things they're standing?
Nat: the US and China are t you know, the two biggest power systems in the world. China is much, much bigger. What China has done is said effectively, we we collectively, the state and the state owned enterprises that generate power, will build the power infrastructure for you so that you don't have to go and source this yourself. The US has a much more competitive landscape where there's very little federal building of anything and it's not directed very much at a federal level. So you have fifty states, you have then regional regional transmission operators and regional grids. You have you have a lot of different ways to then get power, whether you're going to have the utility build it, you're gonna build it yourself, many different options. The Chinese model is this you know, is that sort of the country as a whole provides the power and then you can shop around for where it works best for you, including a high percentage of renewables, depending on where you are. And the US says, you know, let a thousand flowers bloom. Let's think about every possible way we could be competing here. And they're just very different systems. It's hard to say which one, you know, which one is better. Depends on what you're looking for. The Chinese system is also just, my last thought is just significantly larger. So even like a very big data center build out has proportionally much less impact on total power demand than it does in the US. It's just it like like the system is like two and a half times the total amount of power generation is the US. And that means that that means that there's n there's not as much pressure from each individual new unit of demand coming online as there would be in the US.