As demand for data center power skyrockets, available options to provide that power have dwindled. And cohesive frameworks for finding sustainable generation remain few and far between.

In this episode, Shayle speaks with Jake Elder, senior vice president of research and innovation at Energy Impact Partners. The two colleagues dig into the four main generation solutions — on grid, off grid, on edge, and off planet — and consider the viability of each in the years to come.

Shayle and Jake explore topics like:

• A ten-year forecast: Jake’s prediction for how the global “compute pie” will get split up between these four pathways
• Jake’s skepticism around whether a shift towards on-device compute can scale effectively
• The worsening bottleneck facing on-grid connection
• Building “shock absorbers” into the infrastructure of off-grid data centers that enable them to maintain “five nines” of reliability
• The feasibility of making orbital data centers affordable
• The logistics behind creating radiators “the size of a small town” to dissipate heat from orbital data centers

Resources

• Catalyst: PJM and ERCOT are navigating a capacity rollercoaster
• Catalyst: Will inference move to the edge?
• Catalyst: Who benefits from the AI power bottleneck?
• Open Circuit: Are investors losing faith in the AI infrastructure frenzy?
• Open Circuit: The White House AI power pledge: Political theater or policy?
• Latitude Media: The data center boom is a diesel generator boom
• Latitude Media: How Hitachi became a speed-to-power company

Credits: Hosted by Shayle Kann. Produced and edited by Max Savage Levenson, Anne Bailey, and Sean Marquand. Original music and engineering by Sean Marquand. Stephen Lacey is our executive editor.

Catalyst is brought to you by Uplight. Uplight activates energy customers and their connected devices to generate, shift, and save energy—improving grid resilience and energy affordability while accelerating decarbonization. Learn how Uplight is helping utilities unlock flexible load at scale at uplight.com. 

Catalyst is brought to you by Antenna Group, the public relations and strategic marketing agency of choice for climate, energy, and infrastructure leaders. If you’re a startup, investor, or global corporation that’s looking to tell your climate story, demonstrate your impact, or accelerate your growth, Antenna Group’s team of industry insiders is ready to help. Learn more at antennagroup.com.

Catalyst is brought to you by EnergyHub. EnergyHub helps utilities build next-generation virtual power plants that unlock reliable flexibility at every level of the grid. See how EnergyHub helps unlock the power of flexibility at scale, and deliver more value through cross-DER dispatch with their leading Edge DERMS platform, by visiting energyhub.com.

Transcript

Shayle Kann: I’m Shayle Kann. I lead the early stage venture strategy at Energy Impact Partners.

Welcome. So massive data centers on the grid. Massive data centers off grid, small data centers on the edge, huge data center clusters in space. Each of these might get built. Actually, each of them probably will get built. But how much and when? When you boil it down? I think there are actually two basic questions at play here.

The first is the amount of demand for compute in the future, and the second is how to deliver the energy required to meet that demand. I don’t really personally have anything insightful to say about the first one, but boy do I spend a lot of time thinking about the second one. It has occurred to me that I’ve never really seen anyone attempt a cohesive framework to think through all of these different pathways.

You see proponents of one or another adopt kind of a maximalist approach to any given one, but I haven’t seen anybody try to think through how they weigh against each other. So I’ve been trying to organize this in my head and in doing so, I’ve realized that each of these different configurations, sightings.

Types of data centers has a core constraint or two, but each also has its strengths. So to talk through all of them with me, I brought on my colleague Jake Elder. Jake works with me at EIP and he leads our research practice that is focused on the built environment, which these days increasingly means a lot of data centers.

One more thing. Uh, coming up on April 13th in San Francisco, we’re gonna do a live episode of this podcast at the Transition AI Conference. It should be actually a really interesting conversation. My guest is gonna be Amin Vahdat, who’s Google’s chief technologist for AI infrastructure. So obviously relevant to this conversation as well,

I rarely do these in person, so if you’re in San Francisco or you wanna be on April 13th, go sign up for transition ai. register@latitudemedia.com slash events.

Here’s Jake.

Jake. Welcome.

Jake Elder: Thanks. Excited to be here.

Shayle Kann: All right, so the premise here is, let’s just for the, in the interest of not adjudicating this question, let’s just assume compute demand continues to scale. Let’s assume super intelligence, a GI, maybe neither of those things, but that like. The demand for compute and actually the demand for Watts to deliver that compute.

Right. Let, let’s also assume that there is no massive energy efficiency gain that comes and like totally changes the paradigm. So if it is true, and let’s assume that’s true for the next, I dunno, five years, 10 years, whatever we want to talk about, let’s assume that’s true. I think the thing that we wanna talk about here is what are the various options to deliver as much of that demand as possible?

What are the options on the supply side? And so we’re gonna talk about the incumbent solution, which is large hyperscale grid, connected data centers. And then we’re gonna talk about each of the alternatives that I think are currently being proposed. Some of them already being developed, some of them being talked about on X a lot.

And I think we’ll compare and contrast, right? We’re gonna talk about the, the constraints on each of them. But why don’t we start with the incumbent thing. The thing we are doing right now where all the data centers are, which is like large hyperscale data centers connected to the grid. What do you think of as being like the core constraint to just delivering 10 x the compute in that way?

Jake Elder: Yeah, no. Great. Great question. I think it’s gonna make for a great, great conversation as we look across the, the different options here. I think the constraints on the grid side right, are fairly well known at this point. It’s a speed issue in particular on the transmission side. How, how long can we, how much time will it take to build out the transmission capacity necessary to interconnect these mega sites, gigawatt scale sites to new power supply, ideally, you know, carbon free power supply.

In many markets, right? That’s running five to seven years now, which is, is a pretty massive timeline for data centers given the speed to, uh, power and speed of deployment on the AI build out that we’re, we’re trying to drive. Maybe the other couple issues that we should at least be mindful of here are power quality, right?

These large data centers, especially as they cluster in certain locations, can have. Bigger impacts on the grid writ large, and the extent to which society, regulators and utilities are willing to serve those customers if they have bigger grid impacts, I think is still a bit to be determined and a space I’m watching pretty closely.

And then of course, maybe the third vector from my side would be, let’s call it, you know, social license. To operate. And we’re seeing in many states, right, just blanket bans on new data center developments. We’re seeing some developments get pulled, you know, years after announcements because of community pushback.

And, you know, if you listen to Elon, for example, and his best cases for going off planet with compute infrastructure, that’s really his argument that at the end of the day, society is, is not gonna move at the pace that the AI build out requires. And therefore, at some point we’re gonna have to, you know, abandon the planet and, and go to the stars.

Shayle Kann: Yeah, we’re gonna get to the orbital compute thing a little bit later. I think that is a good point, right? People, so you mentioned three things. There’s the capacity, actual physical capacity on the grid. And deliverability. There’s the power quality thing, which, and that one of the three, I think is probably the most manageable, honestly.

It’s like a engineering problem. And then there’s the social license to operate, which we’re already seeing kind of burst at the seams in some locations despite being kind of in the early days of this trend. So I think that third one is underappreciated in the question of like, are we gonna be able to deliver all of the compute capacity that we need via the current paradigm?

On the first one, I would say I do think people conflate. The transmission problem and the generation capacity problem. And the thing is, they’re both problems. I mean, you said five to seven years. The five to seven years is the timeline to get new gas turbines if you’re ordering them, and it’s close to the timeline to get new transformers and other, and switch gear and stuff like that.

Like we’re all in the like three to five or maybe seven year timeline for that kind of thing at this point. And maybe the timeline also to get like a substation upgraded. Which is part of the deliverability thing, but I wanna say the timeline to get a new transmission line built, especially if it’s like inter-regional or across state lines or whatever, is not five to seven years.

It is essentially infinite years in the United States, at least in recent history. Like we just aren’t doing it. So there’s a limitation there that might be even more intractable than just the generation thing.

Jake Elder: Totally. And, and I think it’s the, it’s the unique constraint to the grid connected pathway, right? If, if you wanted to go towards some of the other options, we’ll explore down the road and you’re gonna go off grid, for example. You’re still stuck with the timeframes for transformers, the timeframes for generation assets, et cetera.

And I know we’ll talk about some, some ways to shortcut that, but. The transmission site is really unique to this, this first scenario and certainly makes the case that, if you wanna run around that you need to think about some amount of onsite power as the only way to avoid having to build more poles and wires to route power from elsewhere to a, a new site.

And so, you know that, that on the capacity side is one thing that I think gets lost a little bit in this grid connected conversation is it tends to be an all or nothing conversation around how we. Power these data centers. And I do think there’s like, you know, a hybrid option here where you’re still grid connected, but the data center brings some of its own power for a few hours in the day specifically to overcome that transmission bottleneck.

Shayle Kann: And to be clear, that’s what’s happening now. A lot of that there, this, this concept, in fact, I’ve seen people get confused about this. ’cause there was some, I don’t remember who put it out, but there’s some report that came out that saw like. There’s like 50 gigawatts of behind the meter generation and development at data centers.

Right? And they, and some people have interpreted that to be oh, 50 gigawatts of off grid data centers getting built. It’s actually close to zero of those that are true off grid data centers. They’re all either grid connected, but have some behind the meter generation or the behind the meter generation is a bridge and they ultimately intend to be grid connected.

So that is true. There’s a hybrid there, but, okay, so this is the least interesting one ’cause this is the way we do things now and it’s gonna be the way that we do things. As much as we possibly can. Like I think you and I agree that like the first thing that’s gonna happen as is already happening is that developers are gonna find as many sites as possible that can handle hundreds of megawatts or gigawatts of load.

They’re gonna develop those into data centers. So like we just assume that happens and we should just assume it’s not enough or, or maybe it doesn’t happen because of community pushback, but either way, we’ll assume it’s not enough. Now let’s talk about the other, I think, three categories of ways of configurations to get a lot of new compute online.

The first one is maybe the least distant, which is you still grid connect data centers, but they’re smaller and you put them at the edge. So I’ve talked a little bit about edge compute on the podcast before. You and I have spent a lot of time thinking about it separately. First of all, define what you think of as edge compute. ’cause it is sort of malleable. And then like, what is your latest thinking on what role that plays in the market.

Jake Elder: Yeah, so, so this is a really tricky question, right? Edge computing has been around for a while. Historically it evolved to serve certain use cases like telecommunications, and more recently, video streaming, for example, is something that happens much closer to the edge than other hyperscale data center activities.

Moving forward, I think there’s a school of thought that says that AI inference in particular might move to the edge. The, you know, first principles argument that folks tend to make is that latency’s gonna matter more. And so citing compute infrastructure closer to demand just has a performance benefit that can’t be met via, you know, large central sites in West Texas, for example.

As we’ve dug in a little more, I think that’s a little bit of a red herring. And so let’s come back to that in a second and talk about why you would actually pursue edge data centers and edge computing. But latency has certainly been one of the reasons historically. That edge computing can mean a few different things to your point.

Right. So in the extreme scenario, I think, you know, as you move out, you know, 10, 10 plus years, more and more is gonna happen on device. We already know that. You know, like Waymo cars for example, have a lot of their day-to-day or all of their day-to-day navigational tools and driving decisions get made in the car directly.

And increasingly as we have models that can operate on a phone, for example, you might have a version of Chat GT or Gemini that just operates natively on your phone and doesn’t need to go out in the world at all to get. Access to basic inference, you know, results. On the other end of the spectrum, we’ve seen a few folks announce larger scale projects.

Really think about 20 megawatts style data centers, maybe 15 to to 30. And those folks are basically building many hyperscale sites, but they’re trying to build them in locations where they think they can get power sooner, and perhaps in a regional node where they could serve. You know, some late more latency sensitive applications.

But from a design perspective and a deployment perspective, they kind of look like much of what we’re building today. Just, you know, small scale relative to the gigawatt scale assets. My suspicion is that’s probably the most economic piece here. And so if this becomes a cost play that, that’s the space that becomes most interesting.

But again, let’s come back to that. And then I think there’s this third category, which is really more kind of true. You know, what we might thought, have thought about is edge computing, where you’ve got a. You know, a hundred kilowatts at a given site or a couple of megawatts at a given site, you could think about these being located at the utility substations or in a commercial real estate, you know, office basement.

And the reason to pursue that right is probably cost at the end of the day. There, there, we know across the, the folks that we know well, right? That there are a number of individual parcels of land that were, you know, provisioned for five megawatts of power and are only using two. And so I think the theory to pursue that is probably more around speed, where you could probably suck up a bunch of assets relatively quickly and start to build out a network. But if you end up in a cost game and you’re trying to be the cheapest form of inference strikes me that that probably struggles. ’cause you’re, you’re subscale relative to, to bigger sites.

Shayle Kann: Yeah, you said a couple things that resonate with me based on what I’ve learned. The first is that is, is that latency is a bit of a red herring. The latency benefit of being edge, not for zero applications, but for very few, does it seem that you need such low latency that Edge has a, has a big benefit over, you know, the sort of like regional hyperscale model that we have today.

And people use the example of things like autonomous vehicles. That was like a classic case people would talk about is, well you need edge. Computing for autonomous vehicles, but as you said, most of what a Waymo needs is inside the car. And so as I understand it, they can operate with compute inside the car, and then when they need to go pull something from the cloud, it’s generally not solely latency sensitive that they can’t handle the hyperscale.

So this concept of edge being necessary for latency purposes, I’m, I’m yet to have that proven to me, I’m, I’m waiting for it, but it does seem unlikely. Secondly, it’s hard to imagine it’s cheaper. Now people do make the argument that you might get free land, right? And that could be true, like if you’re taking land that’s already getting paid for because it’s at a commercial property or whatever, it’s in a parking lot.

It could be any of those places at a utility substation that’s already substation the land could be pretty cheap. But if you look at the fully loaded cost of a data center, land is not a big. Portion of it. It’s a very, very small portion of it. The cost is actually in the GPUs, obviously in the building, in the labor, all those kinds of things.

And as you said, being subscale is tough. Maybe you can make some modularization argument. You know, you have the standardized shipping container and the shipping container is like super cheap and easy to deploy. You just plug it in. But, as is true in many other sectors, my guess is. You know, your 300 megawatt data center on a fully loaded levelized cost of flop, uh, is just gonna be cheaper.

So it’s probably not a cost thing either, which means it’s a speed thing, right? And speed is the, the name of the game right now. But I think what remains to be proven in edge world is that it can actually be faster at the same scale. This is what we need to find out.

Jake Elder: I, I think that’s right. And, and I do think at some point the speed game is gonna slow down and cost is gonna matter, especially in the inference world. I don’t know exactly when that happens. And you know, in our future scenario where we’re in some kind of relatively quick takeoff around AI capabilities, maybe speed matters for longer because models continue to improve kind of indefinitely.

But at some point when we have age agentic, you know, employees in most Fortune 500 companies in, in this kind of future, right. The cost of those workers matters. And so I do think at some point if there’s an Edge build out and you’re looking at two or three different edge deployment models, well speed matters.

The cheapest of those models might be the one that that ends up winning at scale.

Shayle Kann: Yeah, but I think speed remains a question mark. Like in principle, if you have an existing interconnect, as you said, there’s some commercial site that has like a five megawatt interconnect and is using two megawatts, you put three megawatts on there. That should be much faster than waiting for an upgrade in the system.

But of course, to match the speed with which you need to go, you’re gonna go deliver your 300 megawatt data center. You then need to go find a hundred of those sites and develop a hundred of them. And like in principle, I can understand how that could be faster, but I’m waiting for somebody to show me that that is true.

Jake Elder: Yeah. And certainly requires a lot more conversations and turning over of rocks and, you know, dead leads as you try to build it out, right? You, you’ve gotta have a hundred success outcomes in terms of site evaluation, not just just one.

Shayle Kann: Right. Okay. So that’s edge. So both of those are grid connected. Uh, let’s assume the grid becomes the constraint. It just is the constraint. Right. Okay. So now we’re either gonna, we’re gonna get into like. Increasingly distant in a literal and a metaphorical sense. But let’s start with the one that I think is, is maybe least talked about relative to how interesting.

I find it as an answer to this question, which is just off grid like, and again, we’re not talking about a hybrid version where you have behind the meter generation, you’re still grid connected. Let’s just say put a data center anywhere it, it has an amazing relaxation of a constraint if you remove the grid as a constraint.

Right. We have plenty of land available. Right. That is not the constraint here and, and you can go where there’s the cheapest labor you can go where there’s the easiest permitting and siting, like it does change the game in that manner, but it does have its own set of challenges and constraints, which is why it hasn’t happened a lot historically.

So what’s your perspective on just straight off grid?

Jake Elder: Yeah, I mean. You make a pretty good case, it should be pretty attractive, right? There was this foundational study that came out about two years ago, that was co-authored by Stripe and Paces and Scale microgrids, and they found over a terawatt of opportunity in the American Southwest alone with high levels of renewable development, being able to support those assets like 50% solar.

Plus batteries at cost priority to using all gas and the ability to get up to, I think, 80 or 90% solar without a meaningful cost increase. So like from a land perspective and a resource perspective, it makes a lot of sense. And to your point, it can also move really quickly. You can avoid the places where the public really doesn’t want data centers, right?

You’ve got such geographic flexibility. It should be the opportunity if you, if you just take a first principles approach. And we certainly don’t need to be thinking about going to space until we think about going to remote parts of, of the Earth. Right? But to your point, it’s not happening at scale yet.

And I think there’s a couple reasons for it. There are some projects that are happening that we can learn from, right? And we’ve got some, some anti data, to support that. I think at the end of the day, the grid’s a marvel of humanity and it does a lot of really good things. In particular, being a giant shock absorber for any one individual asset.

And so if you go off grid and you have to operate on an island, you have to build the whole shock absorber yourself, all of the inertia, the fault response, the ability to blackstar the asset. And that’s just not, not just expensive. It’s really complicated and there’s not a lot of folks out there that know how to run a gigawatt scale grid.

Right. When you think about. The risks that these new data center developers are needing to take in the values of these assets. Betting on a model where you can’t be, you know, comfortable or can’t guarantee that you’re gonna have 99% uptime, is, is possibly a non-starter in some cases. And, you know, when we, we’ve heard some of the early data, from some of the off-grid projects that have been built so far, the an data suggests they’re not able to stay above even 90% up time yet.

Will they get there over time, probably. Right. Like, this is a, a learning curve and we, we know that there are power quality solutions that can manage a lot of these issues, but it’s a big risk if you’re gonna be a first mover for a $10 billion asset, to, to design it in a way that you don’t know how to manage and operate it and keep it running.

Shayle Kann: Right. It strikes me as one of these things that like, clearly that is, it should be solvable. It is a real engineering. Challenge It appears, and I’ve, I’ve, you know, you and I have looked at some of that same data, like it does appear that there are, there are actually projects that are mostly these ones that are bridge power projects.

So they’re currently off grid, intending to be on grid eventually, but as they’re operating off grid, they are not operating at the normal five nines of reliability or whatever. Now, interestingly, you may or may not need that. That’s in some ways it’s sort of a legacy of the cloud business where. AWS and Azure.

And, and Google basically promised in their SLAs to their customers that they would be able to offer really high uptime. And so they have this, you know, huge redundancy requirement and and so on. And the new world of ai, sometimes you do need that. Sometimes you don’t need that. And so there may be a class of data centers that can accept sub 99.999% reliability.

There’s an economic impact, of course, to lower uptime. But again, in a world where we’re so constrained on the grid side, it seems inevitable to be that that is gonna happen to some degree, and that the engineering challenge is gonna get at least partially solved.

Jake Elder: Yeah, I, I think that’s right and I think over time, we’ll, we’ll figure out better ways to, you know, have more and more checkpoints as you’re doing model training runs and whatnot. Such that you could tolerate a major outage. I think the key is you could make it work at 90% uptime if you know when that 90% is.

I don’t know whether you could handle. Total randomness with that 10% downtime. And if they all, all the downtime happens to come in the middle of big, long, expensive model runs that it takes down, right? I don’t know what that does to the economics of those, those projects. And, you know, I do think we’ll learn a lot here.

I think it’s critical also to acknowledge that, you know, the, the, those that operate our larger grid don’t yet know how to manage these sorts of voltage swings and harmonic distortions that are coming from these data centers. And so if we can’t solve the problem when the data centers are a small part of the overall load on the.

System. Then it tells me that’s probably gonna take us some time to figure out how to solve it when they’re the only load and it’s, you’ve got a much more constrained set of tools to, to manage the impact.

Shayle Kann: Yeah. And, and though cost is not the determinant factor in this stuff right now, it’s not nothing. And you know, the way to. Engineer yourself into five nine’s. Reliability off grid right now is, is to overinvest in both capacity and sort of, and and storage, and you can do that, but it does come at a significant cost.

It starts to actually matter for your economics. And back to your point, like who finances your $10 billion asset if it is, you know, at the top end of the cost curve

Jake Elder: Yeah. And then you start getting into, you know, do you need two different fuels, right? If you’re gonna use, you know, some, some kind of base load resource, or if it’s just gas, you need two separate gas pipelines and that constrain sites add costs. And oh, by the way, we’ve just jumped in, assuming that location doesn’t matter in this world, right?

And that you can do everything in remote parts of the, the country, for example. I’m curious for your take there. I, my suspicion is that at least to date, folks are still generally sensitive to where they’re being cited for not all projects, but for most projects. And, if there were lots of off-grid opportunities in, in Virginia, for example, I think we’d see them being pursued more quickly than we’re seeing some of the, the stuff move forward in West Texas, New Mexico, et cetera.

Shayle Kann: I think that’s changing in real time. Like historically, you know, there were these tier one markets like Northern Virginia or Chicago or Phoenix or whatever, Atlanta and uh, and they were where 90% of the. Demand for new data centers was gonna be, and there’s still that. But it is broadening out quickly, right?

And you see all this development in West Texas, for example. So many data centers going into Texas. And I think that’s just because of speed to power and availability and scale, right? And so I, I think that the constraint of like, you need to be in certain locations, it still matters from a, is there a workforce?

Is there, you know, can you get enough labor electricians and construction workers, and. Water and all that kind of stuff. But I think apart from that, my sense is that it is not the most important thing. The one thing I do wanna say though, about the grid thing, and you mentioned this before, but let’s reiterate it.

Your fundamental, assuming you can solve for sufficiently uh, advanced engineering to get to whatever reliability you need, your constraints then on scaling predominantly become. Power generation and delivery. So you’re still, you still need to, ’cause you’re probably gonna need some gas, you still need turbines.

Or if you’re doing a lot of solar and storage, you need solar and you need batteries, you need transformers, you need switch gear, you need whatever, all that kind of stuff. And that, that you’re now still in that supply chain problem. And I wanna, I wanna mention that because if that is the constraint on really massive scale off grid.

In a minute we’re gonna talk about Orbital, and so we can compare and contrast like which is the more challenging constraint between those two.

Jake Elder: Yeah, no, that’s a great, great reminder. You’re still stuck with all the generation and supply chain issues. Maybe with one possible exception, which is that your gas infrastructure is gonna likely be smaller and more mod modular, right? Like, you’re not gonna have a 500 megawatt, you know, combined cycle turbine.

That’s your sole generation asset for a, a massive data center just because of the redundancy issues. And so. You can get a lot of one mega reciprocating engines today. I know you can find some smaller, a derivative turbines or, you know, all repurpose jet engines if you wanna get a little bit crazy. But I do think the off-grid option in some ways maybe shortcuts the actual supply chain bottlenecks on the generation equipment side, at least to some extent relative to the other options.

But agree with you. There’s still a bunch of other pieces of equipment, transformers, et cetera, that, that you’re stuck waiting for.

Shayle Kann: Okay, so let’s shift to the most fun one. We talked about off grid. Let’s go off world, and talk about orbital data centers. There’s. Such a long conversation to be had about orbital data centers here, but I wanna frame it in the context of these other things. Again, again, I think the premise here, and certainly the way that Elon talks about it, is the most prominent, uh, proponent of orbital data centers is this is gonna be the only way.

It’s a scalability thing. I mean, he says too, let’s, okay, let’s, let’s dispense with the premise. He says he thinks orbital data centers are gonna be the cheapest way to get compute in three to four years.

Jake Elder: Correct.

Shayle Kann: I do not believe that. Do you believe that?

Jake Elder: I do not believe that. I think like we need to start this conversation with a bit of the, the acknowledgement, right? The moving off planet for lots of reasons is a crazy proposition, right? And if you listen to Elon talk through it, it starts to sound like a logical end game in a world where we’re building hundreds of gigawatts of compute infrastructure a year, and Elon asserts that, that’s gonna start happening in, in three or four years, right?

I don’t think it is likely, I don’t think it is gonna be the cheapest source of new compute capacity in three or four years, nor do I think that we’re gonna be building hundreds of gigawatts of compute infrastructure per year in the US alone in three or four years. But in a world where we’re assuming that we’re somewhere between, you know, a GI and some, you know, more super intelligent, you know, computing infrastructure. It’s kind of the end game, right? It’s kind of the only place you could go to build, you know, infinite amounts of compute capacity. Whether that’s in five years or 500 years. You know, I, I’m not quite sure, but I agree it’s not, not before 2030.

Shayle Kann: And okay, so people have, as this has become a bigger conversation, people have talked about lots of things that they think are gonna be the like. The killer of the idea of orbital data centers. I think we should dispense with them. ’cause despite what you and I just said, which is both like fairly skeptical on the cost side, I think we both think it’s not like totally insane and it doesn’t seem like the technical challenges are insurmountable.

So people talk about like heat transfer as one of the big problems. I think It doesn’t seem like actually that is likely to be, it’s not nothing but it, it doesn’t seem likely to be the thing that kills orbital data center.

Jake Elder: Agreed. I think the, the, the heat transfer conundrum right, is that space is a vacuum and it’s very, very hard to dissipate heat in a vacuum. I think the, the whole international Space Station, for example, rejects less than a hundred kilowatts of heat in total. And they have a radiator the size of a soccer field.

Right. And when you think about the compute infrastructure where we’re building out, like a single Nvidia high density rack could soon be more than a hundred kilowatts. It may already be, in some cases more than a hundred kilowatts. The flip side, of course, heat dissipates to the fourth power of temperature.

And so it turns out that the hotter and hotter you run chips and the denser and denser you run chips, the better your heat rejection gets on its own. And so it does seem like as we move to a world of denser and denser computing infrastructure, it gets easier and easier to reject chips. But if you just do the simple math on a single gigawatt scale.

Space-based data center, you end up with a radiator the size of a small town, right? Like between the radiator and the solar panels required. I think you end up with a four square kilometer, orbiting asset. And that’s obviously complex to manage, but it’s also a target. I saw this really great, piece of analysis this morning from a analyst called Thunder Set Energy.

And I think the stats on like the odds that a starlink system. It gets hit today, by a piece of space debris is like a couple percent maybe per, per year. If you scale that up to a single floating thing that’s four kilomet, four square kilometers large, you can basically expect to have a piece of space debris, debris hitting that, data center every hour. And I don’t know how you operate something that’s gonna, you know, get knocked off, over and or destroyed, just every hour like by a piece of space debris every single hour. That sounds really complicated.

Shayle Kann: Yeah. I mean, to me, the thing that seems, this is sort of related to it, the thing that seems like the hardest to solve. It’s all hard. But the thing that is the hardest to solve with orbital data centers is o and m. ’cause actually data centers on land require a lot of maintenance and you can’t really do a lot of complicated maintenance, uh, to a satellite, right?

And so either we solve that with some robotics that’s gonna be very clever. That seems difficult for me to imagine. Or it’s an economic thing. You lose a bunch of, you just have some loss rate, and you have to account for that.

Jake Elder: Yeah, I mean, you know, in, in a hyperscale data center today, right? Like there’s a meta engineer or a Google engineer that is going to replace every CPU or GPU as it breaks more or less in real time and in space. If it breaks, at least today, you’re, you’re kind of stuck with it broken. And to your point, maybe in 20 or 30 years, if we’re really in some super intelligent future, there’s, you know, robotic replacement and ways to update chips in real time and whatnot.

But until then, it just adds economic drag on the, the overall project. And, you know, we kind of skipped over cost, but it, it’s not clear that there’s a real economic advantage here. I mean, the economic reason to do this right is free, free power. You could effectively get 95% capacity factor on the solar panels at a.

Space-based data center because you put it in kind of permanent, sun, right? From an orbital perspective. And then there’s much better solar radiance. So you get somewhere between five or 10 x the energy output per panel over the life of the panel than you would on Earthbound panel. And so, you know, power is really cheap.

But as you mentioned earlier, you know, total cost wise, energy’s only, you know, five to 15% of a. AI focused data center and chips and maintenance are the rest, and you’re stuck with the same chip cost, whether you put the thing in, in space or or on earth. And, the maintenance piece gets, gets much more expensive.

And so. I kind of have a hard time seeing it being a cost play, even in a world where launch costs go way down and if you buy Elon’s Elon’s view of the world, that the star ship’s gonna get super reusable and be able to launch at a hundred bucks a, a kilogram. And so I kind of come back to like, it just has to be the sort of thing that we pursue from a physics perspective because we can’t build at the pace needed for for a GI on Earth.

Shayle Kann: I think that’s right. Okay. So but that gets us then maybe to, to close it out into what I think is the interesting comparison that I don’t hear people making very much, which is. Orbital data centers versus off-grid data centers. Let’s just compare those two. As we said, the rate limiter. We have plenty of land.

I mean, you know, in the long arc of history to build many terawatts. Sure, we’re gonna run outta land, but like to a first order for the next decade. I don’t think we’re running outta land. So we, we’ve got land, Then the rate limiter is all of the other stuff we talked about, you know, turbines or whatever, power grid infrastructure and so on.

And we certainly don’t have enough of that today to go build hundreds of gigawatts a year of off grid data centers. The rate limiter on orbital data centers is sure there’s gonna be some like solar. Solar for space, right? Uh, Elon is saying that X ai, or I guess now, now SpaceX is gonna. Develop a hundred gigawatt solar manufacturing, uh, presumably for space.

They’re also, Tesla’s gonna do it for land. But let’s let, let’s say that that’s the lesser constraint, the bigger constraint is Starship. Starship has to launch a lot, like a lot, a lot to get that kind of capacity into space. And they’ve got a ways to go. So as I think about it, I’m like, okay, if you’re, if your, if your binding constraint is like capacity of Starship launch on one side versus.

Ability to scale up the supply chain for power generation and delivery on land. It’s not clear to me that like space is eminently more scalable on that, on that measure of the problem. Like, can we not as a planet go develop, you know, 200 gigawatts a year of new turbine manufacturing capacity? Seems possible.

Jake Elder: I, I think, I think that piece we could maybe, maybe the question back to you is, do you think that society over time is supportive of us building, you know, 200 plus gigawatts of incremental gas infrastructure year over year for the next next 20 years? And I know that’s, that’s. One of the other concerns that Elon raises right, is at some point, the, the conversation around carbon free energy will, will shift back in a different direction.

And

Shayle Kann: Totally. So, but then, but then, right. But so, but then be maximalist on solar and storage. Be a maximalist on geothermal. Be a maximalist on new nuclear. Like, are those things all so much crazier than like five Starship launches a day? You know?

Jake Elder: When you hear him talk, you through it and it’s basically the, the ship lands and then takes off again within, you know, a few minutes, uh, that, that certainly does sound pretty crazy. And, you know, solving fusion might even be easier than, than cracking that code.

Shayle Kann: Yeah. Again, I think for me it’s not that like it’s totally insane to do orbital data centers. That’s not my takeaway here. It’s just, I think if we’re going straight to space, I’m surprised that we’re. Making, stopping at a way point along the way of doing a lot of off grid. I’m surprised that hasn’t happened.

Jake Elder: Agreed. And I, I think the other, the other constraint that, that obviously exists across both scenarios and we’ve kind of washed over in the, you know, a decision to talk about a world where we’ve, we’ve continued to see massive AI progress is just the chip supply chain, right? And in a world where we’re building a couple hundred gigawatts a year, I don’t know how many chips that actually turns into, but I know that we don’t have the semiconductor fabrication space today to build at that level.

And so we probably end up bottlenecked by chips before we’re. Really in a world where we can’t build everything in on, you know, on the ground, for example. And probably before we’re in a world where Starship launch costs are so cheap that space becomes the cheapest option. So, you know, if you take that as a fundamental constraint, then I think you probably do bet on the off-grid stuff, moving materially faster.

Yet same, same as you. I think, you know, we shouldn’t dismiss the, the orbital option. And I think in a world where, you know, compute build out does rapidly accelerate in 20 years or 30 years, there’s gonna be a lot of AI models being trained in particular in space. And that’s maybe just the one last topic we, we didn’t quite hit on is latency in space.

Right? If you’ve got latency concerns. Building in West Texas, then you’re certainly gonna have latency concerns building, you know, a few miles north of the or above the, the South Pole. And so I do still think in that world, right, we’re still gonna have to build a lot of our infrastructure here, even if we’re training the, you know, brain that is a thousand times as smart as a human, the atmosphere.

Shayle Kann: All right, so I’m gonna put you on the spot to wrap up here. 10 years from now, you’ve got a fixed pie of all the global compute. That exists. We have four categories here, hyperscale, grid, connected edge. Let’s define it as like sub 50 megawatts or something like that. So a broad definition of edge, off grid, off world.

Jake Elder: 10 years, all compute infrastructure that’s operating.

Shayle Kann: That’s right.

Jake Elder: You know, if I were to look forward about about 10 years, and assume we’re talking about all compute infrastructure that’s operating. I still think the majority of it’s gonna be in hyperscale data centers, and that’s probably, you know, 50 to 60% of the total. Let’s assume that on top of that, there’s another 10 to 15% that gets built off grid in a similar hyper hyperscale like format, but never connects.

And so that puts us at, you know, 65 or 70% that’s built in more of a traditional way, whether grid tied or not. I suspect that the bulk of the rest comes in the edge markets for certain use cases or applications, call it 15% or so. There. And I do think we’ll see a couple, you know, efforts to really build out some infrastructure in space.

And we know SpaceX and Google in particular are gonna take their shot there. And so I wouldn’t be surprised if we’re training some models. We’ve got five to 10% of our overall compute capacity out there over time. I’m curious, which of those are you, uh, are you buying or selling?

Shayle Kann: Hmm. That’s interesting. It’s so hard. Okay, so again, it comes down to this like, how bullish are you on compute demand? Like if you told me that the total number, the size of the pie in 10 years is 10 terawatts, I have a very different answer from if the size of the pie is 300 gigawatts, right. And that, that like dictates the, the shares to me.

So it’s really hard to know. I would say, I generally agree with you. And to be clear, that’s actually like a fairly bullish statement on. It’s, I would say what you’re saying is bullish on off grid and bullish on orbital just because you’re starting from zero in both of those. And so getting to 5%, even 5% of hundreds of gigawatts is gonna be a big number to do in 10 years for orbital.

So it’s actually like a fairly bullish statement on all of them. Again, depending on how big the size of the pie is. I’m, I’m filibustering ’cause I’m trying to figure out which one of these I disagree with the most. I guess I may be. Where I currently sit, I’m a little bit even more bullish on off grid.

Jake Elder: Mm-hmm.

Shayle Kann: It, it has the scalability. I think it can have the cost. There are challenges, engineering challenges, but if we’re really gonna be in this world where we’re that heavily constrained, like it’s just seems inevitable to me

Jake Elder: Do you think that comes from the grid, tide, large sites, or where do you think that that comes from?

Shayle Kann: where the,

Jake Elder: Like what, what? From, from my view of the world, which of those categories do you see losing market share? Let’s call it if, if more’s gonna go off grid.

Shayle Kann: Oh, I see. I’m still having, I mean, you didn’t put a lot into the edge category in the first place, but where I currently sit, I don’t, I don’t know why we’re gonna have a lot of edge in the grand scheme. We’ll have some, but like in the, as a portion of overall compute, I dunno why that’s gonna be a lot.

Jake Elder: Which is frustrating ’cause it’s the, it’s the least cost in many ways. It’s the, the most obvious and theoretically fastest way. To, you know, deploy, compute. Right? Like, this is why you and I have spent a lot of time thinking about this over the last three or four months,

Shayle Kann: Totally. And look, I’m,

Jake Elder: should be the right answer.

Shayle Kann: And I reserve the right to change my mind, right? Like I think you and I have spent a few months trying to like convince ourselves of Edge, and I think we haven’t done so yet, but that, that’s a matter of time. In fact, if a listener wants to convince us of Edge, I would welcome it.

Jake and I both, but, uh, but yeah, we’re, we’re. We’re struggling to find the, like it ha it’s gonna happen and here’s why. For all these reasons. Anyway, I, I would maybe take a little bit away from edge and I guess I’d take a little bit away from grid connected hyperscale, but I agree with you that that’s like most of what we’re gonna do is just build more grid connected hyperscale.

Alright, Jake, all the time. We’ve got, thank you so much. This was fun as always.

Jake Elder: This was a pleasure. Thanks for having me.

Shayle Kann: Jake Elder is a Senior Vice President of Research and Innovation at Energy Impact Partners. This show is a production of Latitude Media. You can head over to latitude media.com for links to today’s topics. Latitude is supported by Prelude Ventures. This episode is produced by Max Savage Levenson, Anne Bailey and Sean Marquand. Mixing and theme song by Sean Marquand.

Stephen Lacey is our executive editor. I’m Shayle Kann, and this is Catalyst.

The post AI scaling pathways: On grid, on edge, off grid, off planet appeared first on Latitude Media.

via Latitude Media https://ift.tt/hu5TPZo