Amidst the increasing urgency of powering data centers, a new solution has entered the mix: send them out to sea.
In this episode, Shayle speaks to Garth Sheldon-Coulson, co-founder and CEO of Panthalassa. The company is building 85-meter steel “nodes” – taller than Big Ben – that it deploys into the deep ocean. These untethered, self-propelled nodes harness wave energy to power AI clusters, then beam their data back to land via satellite. The technology isn’t without its fair share of logistic complications, but it nonetheless offers a pathway to powering the AI boom that’s largely independent from grid or fuel constraints.
Shayle and Garth cover topics including:
- The physics and mechanics that power Panthalassa’s nodes
- The significance of building an autonomous fleet
- The energy generation waiting to be tapped in the open ocean
- The logistics and unit economics behind scaling Panthalassa’s technology
- Why deep-sea compute is well-suited for long-running workloads like inference and reinforcement learning
Resources
- Catalyst: AI scaling pathways: On grid, on edge, off grid, off planet
- Catalyst: How to build more hydropower
- Latitude Media: Are Thiel-funded floating data centers enough to make wave energy pencil?
- Open Circuit: Grid utilization vs expansion: The 100 GW debate
- Latitude Media: What geothermal can learn from offshore wind’s demise
Credits: Hosted by Shayle Kann. Produced and edited by Max Savage Levenson. Original music and engineering by Sean Marquand. Stephen Lacey is our executive editor.
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.
Catalyst is brought to you by FischTank PR, an award-winning climate and energy tech, renewables, and sustainability-focused PR firm dedicated to elevating the work of both early-stage and established companies. Learn more about their PR approach and how they can support your company’s messaging by visiting fischtankpr.com.
Tune into Critical Capital, a brand new podcast from Crux and Latitude Studios. Hosted by Crux CEO Alfred Johnson, Critical Capital explores the interlocking forces powering clean and critical infrastructure. Join us every other Tuesday for in-depth conversations at the intersection of energy, government, finance, and global markets. Listen here, or wherever you get podcasts.
Transcript
Shayle Kann: I’m Shayle Kann. I lead the early stage venture strategy at Energy Impact Partners. Welcome to Catalyst.
So a little while ago, my colleague Jake Elder and I walked through what I have been thinking of as the four archetypes of data centers that people are building today: The first is hyperscale grid connected data centers. That’s basically the entire market today, but then slightly more revolutionary are edge data centers, off-grid data centers, maybe even orbital data centers. But actually maybe there’s a fifth category that we didn’t talk about, which is coming from a company called Panthalassa. Panthalassa makes 85 meter steel nodes. That’s roughly the height of Big Ben that get towed or self-propelled into deep ocean water, flip vertical, and then bob with the swell. And as they bob, water flows through an internal turbine and generates electricity. No anchor to the seafloor, no cable to land.
That power runs AI compute that’s sealed inside the same structure cooled by seawater and connected to the world over Starlink. It sits in an interesting place relative to the other four. In theory, it’s scalable in the same way that orbital data centers are and in theory without the risk of the same political and local blowback that terrestrial data centers are starting to face. So if that works, you can imagine it being an enormous unlock in the same way that has generated a lot of excitement about orbital. Of course, it also raises a lot of hard questions. How the physics hold up over decades potentially in the open ocean, what the unit economics actually look like once maintenance and decommissioning are baked in, which workloads could tolerate that kind of latency, who the buyer is, even what jurisdictions apply to a compute facility hundreds of miles away from any coastline.
Well, let’s see if we can figure it out. My guest today is the co-founder and CEO of Panthalassa, Garth Sheldon-Coulson. We’re going to work through all of it after the break.
Shayle Kann: Garth, welcome.
Garth Sheldon-Coulson: Thanks, Shayle. Good to see you. Nice to be here.
Shayle Kann: Good to see you. Excited to have you here. Okay. We’re going to dive right in. Tell me how a Panthalassa generator works. What is it and how does it work?
Garth Sheldon-Coulson: Cool. Yeah. So a Panthalassa generator, we call it a node. It’s a new energy technology. We created it from scratch. We created it to do a very particular thing, which is go far from shore and capture energy where the resource is really good. The resource being the waves. And we wanted it to be able to do it hundreds of miles from shore, thousands of miles from shore. So I can explain how the power generation piece works and I will, but it’s also two other things at the same time. So a node is also a vehicle. It drives itself. It can be towed, but we designed it so that once you deploy it, it can walk out to the resource on its own. It can walk back under command, it can stay in a region. And that’s essential because it’s untethered. It doesn’t have electrical cables coming home.
And then because it doesn’t have electrical cables coming home, it also has the payload on board. So each one has a computing cluster or each one has an electrolyzer and it’s using the power on board to do things. So that’s part one is it’s three things all at the same time. As far as the power generation piece goes, this is the piece that we developed first back in, let’s see, it would have been 2016 to 2019. And the idea is to convert wave energy into hydroelectric power for the first time. Nobody had really figured out how to do this. And we said, if you can find a way to just spin a water turbine pretty constantly, then we know what to do with that. That is low cost, reliable, simple. And then downstream of there, it’s pretty easy. So we had to develop a shape, a system hull that because of the up and down motion of the waves causes water to be forced up into a pressurized reservoir.
And from there you drive the turbine. So at the top level, that’s really what it is. It’s a system that does all those other things, but it’s moving up and down, pumping water because of the shape of the hull into the reservoir, driving a water turbine. The water goes on a circuit and it just does that over and over hour in, hour out, day in, day out.
Shayle Kann: So it’s almost like this, I’m sure this is not true, but like it’s offshore hydroelectric sort of.
Garth Sheldon-Coulson: It is, yeah. And Dan Place, on our team, he was our first hire apart from me and my co-founder, Brian. He came up with that term, ocean hydro. And that’s really what it is maybe even more than wave energy. Although of course we’re getting the energy from the waves too, but we’re trying to get the cost structure of hydro. We’re trying to get all the advantages of hydro without the limitation of relying on rivers of which there is a finite supply, obviously.
Shayle Kann: And the water that’s going through the turbine, that’s seawater. It’s not self-contained fresh water that you’ve shipped out there, right?
Garth Sheldon-Coulson: That’s correct. Yeah.
Shayle Kann: And is it self-contained? Is it filtering in new seawater running it through the cycle or is it just like ever cycling the same seawater?
Garth Sheldon-Coulson: It’s mostly cycling the same seawater. The seawater as it goes up into the reservoir down through the turbine then returns back to the main tube that sends it forth into the reservoir, but it is open at the bottom and the water in that tube that is sort of, it’s sort of like a liquid piston that’s doing the compression of the water up into the reservoir under the inertia of that water and the system. That water does have an opportunity to mix with the seawater. So there is some mixing. We’re not totally sealed, but by and large we’re not pulling water through. We’re not sucking in new nutrients for things to grow. It’s mostly a closed cycle.
Shayle Kann: Okay. And you mentioned it’s self-propelled that can propel itself out into wherever it wants to be and then it keeps itself in position. How does it do that? Like do you
have a small separate auxiliary generator on board that’s doing that or what does that look like?
Garth Sheldon-Coulson: So early on we said we could try to use some of our electrical power to drive propellers and stuff like that.
Shayle Kann: But it sort of seems like a chicken or egg, especially the initial propelling, right? Like you don’t have enough resource on shore, so you need to have an auxiliary generator, I assume.
Garth Sheldon-Coulson: Exactly. And it’s not even a generator. So we said, what if we could … We’re all about shapes at Panthalassa. We like shapes that do things. An airplane wing is a shape that does something, a boat is a shape that does something. We wanted to find a shape as well as … I was describing the shape earlier that pumps the water. We wanted to find a shape of the hull that because of the up and down motion pushes water backwards and that causes the system to move forward. So in the same hull shape, we have both of those behaviors happening. We have the pumping action into the reservoir. We also have a shape towards the bottom of the system that pushes water backwards so the system moves forward. The system is always moving forward. It’s always moving forward as it moves up and down. And so all we have to do is steer it.
And so it’s like a Roomba … You can’t stop it. It’s always moving forward, but then if you can steer it, you can drive it around. There are videos I can show you of us doing figure eights out there at sea with these systems. You actually have quite a lot of authority to steer them and drive them in circles and drive them any way you want to.
Shayle Kann: But then presumably your propulsion is driven by the amount of wave energy available. So you plan to operate out in the middle of the ocean where the resource is really good and we’ll talk more about what that actually means, but I’m still interested in the question of how do you get the thing out there in the first place and you’re certainly starting where the resource is not very good. And so I would presume to the extent that it is propelling itself forward, if you drop it offshore, just offshore, it’s certainly going very slowly initially, I would guess, right if at all.
Garth Sheldon-Coulson: So this is a question of where we put the factories and how close are they to places where you can deploy it from its towable horizontal configuration into its operational configuration and then you’re right, how good are the waves there to start producing power and start doing the propulsion? And so for example, we wouldn’t tend to put factories on the coast of North America or something like that. We want to put them in the regions, near the regions where the energy is the best and in the locations where we want to put them, you can absolutely just tow them 50 miles offshore, flip them, and then they can work their way out into the resource under their own propulsive power.
Shayle Kann: I see. Okay. Let’s talk about the resource.
Garth Sheldon-Coulson: And by the way, you don’t have to do it only 50 miles. You could tow them much further if you wanted to, but the economic optimum is to tow them as short of distance as possible.
Shayle Kann: Right, because you pay for that towing and because they’re big. I mean, I guess we should maybe for folks who haven’t seen it as I have, how big is a node?
Garth Sheldon-Coulson: Yeah. So a node is anywhere from 10 meters across at the top, like our ocean two that we did two years ago and ocean three is about that as well, but up to 30 meters across at the top and you sort of get diminishing returns after about 25 or 30 meters and then it goes down in the water column anywhere from 70 meters to a hundred meters. And so, a big system in the scale of human objects, but quite small actually in the scale obviously of the ocean. When you get out there and you’re at sea and you see one, it actually feels very small and it’s also very small compared to ships. So it’s the right size for what we’re trying to do.
Shayle Kann: I guess maybe a better comparison would be like, how big is it in comparison to equipment you send out to an offshore oil rig, right? It’s probably like a big piece of equipment that you might send out to an offshore oil rig.
Garth Sheldon-Coulson: Yeah. Yeah. It’s certainly quite a bit smaller than an oil platform, but it’s same order of magnitude probably as like a particularly big workboat or something like that.
Shayle Kann: You mentioned that this is new and different. I think we should put a finer point on it. What is it that is distinct about this approach versus historical wave energy approaches? What makes this unique?
Garth Sheldon-Coulson: Yeah, great question. So the most distinctive thing first I think is that we are not doing it in coastal areas and this was marine energy for the most part for all of history before we decided to pursue this approach. There were patents, scattered patents in the distant past of people thinking about doing something like this, but most marine energy has been coastal because of course the idea that most people have had is you want to get the energy back on a cable and if you’re going to run a cable, you better be close to shore. And also many of the at least wave energy technologies, but also wind energy technologies have all relied on a sea floor connection of some kind either just to more the system and keep it in place, keep it from drifting off, or in many cases to actually push or pull against to create the reaction forces that you need to drive your generator or power takeoff or something like that. And so there’s been this historical center of gravity to do it close to shore.
We decided to cut the cable and go to the middle of the ocean and that’s for many reasons. It’s number one because that’s where all the energy is, frankly. If you look at the entire globe and you say, “I’m going to look at it for how much energy is there in the wind, in the waves,” the coastal regions are very small in terms of both the energy flux through them because you’ve lost a lot of the energy coming up the continental shelf. There’s just not a lot of energy in those regions on an area basis and then those areas are very small because it’s just these slivers next to the coastlines and you’re also competing with lots of other human activities, fishing, you’re touching the sea floor typically, which has all kinds of regulatory and environmental consequences. And by contrast, if you can get out into the middle of the resource, you’re far away from all of those competing uses, you have the energy right where it’s being generated by the wind and it’s much more intense, much more regular and you can deploy much more without having any consequence for conflict with other activities or ecological consequences related to the seafloor.
So I think that’s number one, that’s the first part of the difference. And then the second part of the difference that’s necessitated by that is the technology stack and there’s many pieces of that and we’ve discussed some of it, but I would say that the general thesis of all of our technology development is like solid state, just rigid steel holes. If you can get the behavior with the shape, then that’s much better than getting it in any other way. And so our company is largely a company about making cool shapes that do things when you put them in the water and we devote a lot of energy to that.
Shayle Kann: You mentioned that if you go out into these areas in the middle of the ocean, you get a much more reliable resource. Let’s talk about the resource. What is it like? How consistent is it in the areas that you’re targeting? How much variability by season or weather conditions or time of day? I don’t know. How should I think about the resource profile?
Garth Sheldon-Coulson: Yeah. So let’s start with literally what is the resource. As you know, the wind first of all is created by a combination of thermal gradients created by sunlight and also a little bit of coriolis earth rotation. So you get wind and the wind is sort of a concentrated form of sunlight is one way you can think of that. And then as the wind blows over long distances of water, it first creates ripples and then those ripples present more of a normal area to the wind and then that can push more energy in. And so you get this compounding injection of energy into the water from the wind that creates the waves and the waves propagate over long distances without significant loss of energy.
The waves that you might have on the beach in Hawaii are often being generated by storms in Alaska or storms in the Southern Hemisphere, very long distances with very little loss, which means that when the wind stops, the waves keep going. Even if the wind stops momentarily, you’ve got this accumulation in this big battery really. And so a thing that we often say is that this energy resource, particularly in the southern hemisphere, is the world’s biggest solar battery by far and will always be. It’s just an enormous storehouse for solar energy. And if you can create the system that just goes and sits in it capturing that energy, you can achieve very high power density, very high availability. And so we can talk about what those numbers are, but that’s the general concept is you just want to sit in the best battery and you’ve got very high power density.
So to give you a sense first of the power density, if you put an object that’s say like 15 meters across in this region, you can calculate how much energy is fluxing through you in the form of the waves and it amounts to like well over a megawatt. It’s two and a half megawatts on average in those areas and that’s because you have the wave height in those regions averaging around four or four and a half meters across the entire year. And you also get periods where it’s larger and then you also get periods in the summer where it can die down a little bit to like three meters, three and a half meters, but those are the typical and it never really stops. You don’t get really below three meters in these regions. So you’ve got this thing that’s on all the time. What that means in terms of your power capture is that you can have very high capacity factors.
You can have very high availability. That’s all a function of how you size your system, how you size your payload, your generator, et cetera. But in all of our optimizations we can be achieving for payloads, we can be achieving with very little battery like 99.5, 99.8% power availability with far less battery than you would need for an equivalent solar installation, for example.
Shayle Kann: So you have a battery integrated into a node as well?
Garth Sheldon-Coulson: We do. In most node configurations we do have a battery. It’s usually between like two and four hours of payload capacity. You don’t strictly need to. Depending on what your payload is, you might not want to, but for very high value payloads that are high in terms of capital cost, you obviously want to be amortizing those all the time so it pays to put in some battery into the system.
Shayle Kann: Right. Which presumably is compute if you’re talking about high value payloads. And so those seasonal differentials, they happen, but the way that you design the system and the capacity of the system, you’re sort of like designed, I guess, to a layperson’s version of it. If the waves drop to three or three and a half meters, that is still enough to be operating near full capacity for your system or you run off the battery for short periods of time.
Garth Sheldon-Coulson: Correct. Yeah, exactly. And so we often run our optimizations over like 10 years of data. The ones we’re currently running are over 11 years of data. So it’s the actual historical meteorological data in these regions captured from a mix of buoys and satellites and big government agencies devoted to creating these models. And so we can actually drop our simulated systems right into that resource and run them over those entire periods of time and say like, how often do you see a drop below nameplate of the payload, for example? And what turns out to be true is that for optimized … You don’t want to put in so much battery that you’re paying for what you don’t need. There is an optimum depending on the value of your payload and so forth. And what you see is that usually optimum is somewhere between like 99 and 99.8% availability on the payload.
So you do have a couple days, usually in the summertime, the Southern Hemisphere summer, where you’ll see for half a day or something that you’re dropping below nominal, but you’re never dropping to zero. You have some draw down and you’re running at like 50% on your payload and then you go back up again. So very different from solar in this sense. It’s like we’re not dropping to zero at all. We’re just riding it and sometimes there’s some drawdowns.
Shayle Kann: Okay. So that’s a good segue into then the unit economics. So I’m curious where you think this ends up in terms of a total delivered cost of energy, but maybe before we get to that, walk me through the big drivers. I mean, it’s like any other renewable in the sense that it’s almost all CapEx. What are the big CapEx drivers for you? Is it steel basically?
Garth Sheldon-Coulson: Yeah, it’s mostly steel. About half the system’s cost is steel. Well, if we exclude battery for the minute because that can-
Shayle Kann: Yeah, that can be a big … I mean, I would presume if you’re doing a two or four hour battery and given what you’ve got in the rest of the system, like the battery could be as expensive as everything else potentially or more.
Garth Sheldon-Coulson: Yeah. The battery would usually be about the cost of the steel. So about a third of the cost structure in a system that did have battery, but if we exclude, and that’s with like nominal battery sizing, but of course you can go higher or lower, but in a system where if we just exclude the battery for the minute, it’s basically half steel and then you’ve got about a quarter is your powertrain and then about a little less than a quarter is your marine coatings and then you’ve got the onboard systems, you’ve got initial deployment and some other things that are small parts of the cost structure, but mostly it’s steel, marine coatings and powertrain. The powertrain is composed of the generator, the turbine, the power is the generator, and your power electronics, which are converting the generator power into what the batteries and payload want to consume.
We have all of that stack in house. It’s a pretty short stack. We make the turbines ourselves, we make the generators ourselves and we make the power electronics ourselves and then you have something that you can integrate a payload directly into.
Shayle Kann: And if the payload is compute, do you end up … I mean, I guess on a relative basis, like how much of the value of the full thing is the payload versus the device?
Garth Sheldon-Coulson: Hugely much more on the payload side. And so you can almost think of … When you’re doing computing and I don’t think we’ve talked about our other platforms as well, but if you’re doing computing, it can easily be the case that the node itself is like a fifth to a 10th of the cost structure, especially if you count multiple replacements of the payload over time. So it’s almost negligible and your goal is just to make that thing not really as cheap as possible anymore, but rather as reliable as possible, as scalable as possible. And this is a big difference from when we started the company. We had at the beginning lots of goals related to like one cent LCOE and things like that. What turns out to be true with the nature of chip costs and demand is that we’re very happy now to optimize in the direction of higher node costs in favor of higher uptime, more battery, better capacity factor on the geometry, manufacturability.
So making the shapes in the system all such that they can be made with very simple steel making, steel forming equipment so that you can really rapidly manufacture systems.
Shayle Kann: Right. Yeah. I think that’s probably right. You could imagine just describing it as like, well, it’s not the cheapest power, but in theory, if you could do it, it’s maybe the most scalable source of power short of orbital, which we could talk about separately.
Garth Sheldon-Coulson: Right, right. Well, I mean, that’s where the economics pull us, which is different, but actually our power is very cheap. It’s like we have designs that are two cent per kilowatt hour on the power and the optimum tends to be, given everything I was just describing, the optimum tends to be in the four cent, three and a half to four cent per kilowatt hour range. But keep in mind, that’s a very high capacity factor.
It’s over 90% capacity factor, at least considered in terms of like standard capacity factor metrics and we can achieve like over 99% availability once you get the battery in there. So very cheap power, but yes, it’s like probably the most interesting parts of the platform are very high availability because that’s what helps you amortize your chips, very high manufacturability. And then that the really interesting aspect of like we get free cooling too. We are actually in a resource that gives us free extremely good convective cooling and that’s huge because it essentially lets you eliminate the entire cost structure of the data center and our object is actually replacing both power plant and data center and that’s the cost structure to be comparing us to not just the power.
Shayle Kann: What’s the capacity of one node?
Garth Sheldon-Coulson: So an individual node is on the order of 200 kilowatts up to a megawatt depending on design, size, optimizations and so forth. We think that the economic optimum for most applications will end up being in the 400 kilowatt range, but there may still be breakthroughs that we make that shift that optimum higher. We’ll see.
Shayle Kann: I mean, it’s interesting, this is probably the wrong way to think about it, but that’s one server rack in a hyperscale data center. So one of your nodes, which is cut to 30 meters wide at the top and 70 meters deep in the ocean is one server rack in a hyperscale data center. It’s such a different picture of density, but I guess the more apt comparison is how much available space and resource do we have in the middle of the ocean versus what we have in Phoenix or whatever.
Garth Sheldon-Coulson: Exactly. Yeah. When you’re talking about super dense racks, that is true and that’s like a crazy concept to think about. As an aside, I’m not sure that we will always be running the most dense racks because it has a lot to do with what is the economic optimum chip for us to run. And so that’s a whole separate set of questions. So what you’re saying is true and that is crazy to think about at the same time too, if you think about how much land is required for powering a server rack using solar, that’s also a very expansive piece of real estate and on an area basis we are far more compact than that for example. Even when you consider our factories too The footprint of the node and the footprint of the factory, when you take that together on a power basis and you compare that to the sort of weighted average land occupation by other energy technologies, including solar, hydro, wind, you throw in fossil fuels and nuclear as well, we’re about one 100th the footprint on the planet compared to the weighted average of those others.
Shayle Kann: That makes sense to me for solar or wind. I mean, that can’t be true for gas, natural gas.
Garth Sheldon-Coulson: Oh, no. Yeah. I’m doing a weighted average and it’s dominated by solar.
Shayle Kann: And hydro.
Garth Sheldon-Coulson: Yeah.
Shayle Kann: Okay. So onto then I think maybe the obvious really big question, which is O&M. Both energy generation equipment and data centers require a fair bit of maintenance. How does one do maintenance in the middle of the ocean?
Garth Sheldon-Coulson: Yeah. Let’s talk about the nodes and then the servers as well. So the nodes, this has been our philosophy from the beginning. We wanted to design something that really doesn’t require maintenance during operation. And how do we get that? We get that by having the hull be just solid state, completely just steel marine coatings and that’s it. And then you have your one water turbine, which is fed by the reservoir. And that is also just a very simple rotary moving part, spins on bearings. You can design that for whatever lifetime you choose, five years, 10 years, 15 years. And those are the only moving parts on the whole system. The whole is solid state. It moves up and down in the waves and the turbine spins inside. None of that should require maintenance, at least not on the timeframes that are relevant. So it remains to be seen whether we achieve that design goal, but we don’t know what about that breaks because we’ve run our systems at sea.
They of course survive. We’ve run our turbines and endurance testing and they survive and those are the elements of the system that you need to survive in order for the systems to last a long time.
Shayle Kann: I would think the generator and the power electronics too. I mean, that stuff fails on land. It’s not necessarily because it’s at sea that it would fail, but we’ve seen that. Inverter failures are not uncommon and solar power electronics are … They’re pretty reliable, but they’re not perfect generators, same thing.
Garth Sheldon-Coulson: Yeah, that’s true. And so this goes to the design philosophy that we have on those things. For our power supplies, for example, the team that we have working on them is a team that came out of Raytheon and Collins Aerospace, places where they have a need for extremely high reliability, power supplies for avionics and basically what is the power supply that powers your triple seven? And there’s a whole bunch of design principles in that related to not using software. It’s all analog logic that runs our power supplies. There’s no firmware, no capacitors with liquids inside that can evaporate. There’s a whole bunch of other design principles that if you follow those, your power electronics really ought to last for the design life without failure. In the event that one does, then that node, which it would be one in a thousand, would potentially be dead in the water, or at least you’d have a fraction of your powertrain go down.
It can hopefully be a graceful degradation. And in the worst case, we have to go recover it, we bring it back, we fix it, and we don’t make the same design mistake again on the next one. But on the whole, on the average, the fleet should have extremely high reliability for these reasons that we’ve been talking about.
Shayle Kann: But my guess is, tell me if I’m wrong, just the economics of going to grab it and tow it back and then redeploy it would be challenging, right? That’s why you built the self-propulsion system into it so that they don’t have to do that all the time because that ends up being pretty expensive. So you really want to design for pretty robust operations for whatever designed lifetime, which guess the other thing I think you were going to describe, which is in the case of compute, what does O&M look like on the compute side? Because again, there’s a lot of maintenance on servers on land too.
Garth Sheldon-Coulson: Yep, exactly. And so you’re right, you would not want to be going and recovering or deploying each individual node over distances of hundreds or thousands of miles. That would start to break your cost structure, but if you’re doing it on the occasional node that fails and we build all of this into our models, then that’s fine as long as you have sufficiently high reliability on all these components, which I think we will. And by the way, which we show in our labs, we run a lot of long endurance testing on all of these components and including in sea water and so forth. So that’s all just an applied engineering problem, honestly. And in the scheme of applied engineering problems, it’s not the worst by any stretch. It’s much easier than landing a rocket, for example. So we think we’re going to be able to achieve that on compute.
Yeah, this is a whole interesting conversation. And so the problem that you’re identifying is that if your servers have a sufficiently high failure rate, then you’re deploying them, you might have really low cost of energy, all of these things, really high scalability, but if half of them are dead within three months, then that’s not a good way to deploy compute. And so we spend a lot of time on this. We spend a lot of time on this in two different ways. One is to model the actual failure rates of the most failure prone GPUs, for example, in our environment and build all of our economic models in a way that includes that. So we actually have models where you have the servers degrading according to data that we have from real deployments in the world, the switches, the other components of the cluster, they all degrade at certain rates.
Some of those have different blast radiuses and so you can model what is the probability that your cluster is degraded in X way by X time and build that into this whole simulation model of your entire fleet of your whole system. And it’s an empirical problem, but it turns out that given the empirical failure rates, it’s all fine. We have somewhat like maybe 1% lower availability on average than systems on land if we’re deploying and recovering at the right cadence and that’s okay in the scheme of all of the economic benefits that we’re getting.
So that’s point one is that even if you’re using the most failure prone stuff that works in the economic model, then part two is there’s also lots of compute that is really valuable to run that is not failure prone at all. And so for example, there’s lots of compute for reinforcement learning, for tool use that is CPU heavy. There’s lots of new accelerators that don’t use as much high bandwidth memory and other components that are particularly failure prone. And that’s where a lot of the industry will be going for reasons related to reliability, but for other reasons too, related to cost, related to availability, related to economics. And so there’s many different kinds of payload that we will be qualifying. We will be able to run the most high performance and sometimes unreliable stuff. We’ll also be able to run the least failure prone and most economical stuff and we’ll form a mix that depends on what is best for our customers and best for the low cost of intelligence that we’re going to provide.
Shayle Kann: You provided a good segue to the next thing I wanted to ask about, which is the market and like what types of use cases make the most sense. But before we get there, just to make sure I understand it, what you’re saying about the economic model still working is basically that if the servers do fail out of the middle of the ocean and you have a logistical system such that you can go pick them up and return them, replace them, fix them, send them back out there, you probably are over provisioning a little bit in order to have sufficient overall reliability or you’re not over provisioning and you’re just accepting some measure of downtime as a function of basically the transit time to and from shore, something like that.
Garth Sheldon-Coulson: Yeah, exactly. And so just to highlight, we do have this really nice advantage, which is that we can command the systems to come home and you can do that either on a schedule according to just the statistical degradation if you have that and you can do that in the case of catastrophic failure and you can just command them to come home. It takes a week or two depending on how far they are out and then yeah, you do a formula one swap of that payload and you send the system back out, you refurbish, retrofit the old payload and then you have that ready to go and put into new systems. And so this is, it’s certainly a big applied engineering effort to make, design, qualify those very bespoke payloads that are suitable for this environment. It’s not like we have exactly a rack in a data center.
We’re going to be designing these custom enclosures with custom racks inside very specific servers that we’ve qualified, but we have the partners to do that. We’re working with several big server makers, we’re working with several partners, chip companies to qualify the specific payloads and then we’ll have basically a menu both for ourselves internally and for external customers of payloads that we know will work on our platform be sufficiently reliable and in many cases we believe that the reliability of the chips, the failure rates, the failure rates will actually be lower on our platform because we can provide much colder cooling temperatures than is typical on land and we have no oxygen. We eliminate the oxygen from the payload by replacing it with nitrogen because it’s hermetically sealed. We don’t have vibration on the platform. We don’t have dust and so there’s all of these advantages that we have that in the empirical data from Iceland and from underwater deployments in the past have actually been shown to produce significantly lower failure rates.
And to the extent that that’s true, then our platform is not just better for cost of power, it’s also better for chip economics overall. And that’s a huge effect if you can achieve that. So we have strong reason to believe that that will be true as well.
Shayle Kann: All right, let’s talk about the market. So you deploy some number of thousands, tens of thousands of these nodes. Let’s stick with the version where it’s compute that is on board. You obviously are not offering a low latency product, but there’s a whole swath of the market that doesn’t really care that much about latency, at least not to the degree where that would matter as evidenced by the excitement around orbital, which also is not the lowest latency kind of thing. You’re offering maybe good uptime, probably not best in class uptime, but you also probably don’t need that in every use case. So how do you think about like who’s your customer for the compute?
Garth Sheldon-Coulson: Yeah. So it’s anyone you can think of who wants either a lot of intelligence applied to problems or to make the models better so that when the intelligence is used, it’s more powerful. And so there are these two buckets. The first one is just long running inference and long running inference means whether it’s for coding, you’ve got a code base, you want to send it somewhere, have the agents churning on it. Our platform is the perfect place for that. It’s very low cost. You can send it all of the inference chips are running around the clock and you can swarm agents onto problems. They can be working together on problems. They can be communicating with each other and it actually can be very interactive. The additional latency that we have is only like a hundred milliseconds that vanishes into even the latency of time to first token on most prefill, certainly on the interactive latency of a human waiting for an output, which can often take minutes or if you send something away for a long time, it can take even hours.
So the satellite latency isn’t really a problem, but we’re not going to be the thing for like that result at the top of Google.com or something that’s like driving a self-driving car. So there’s a whole class of like super latency sensitive applications where you wouldn’t want to use us, but that’s not where the bulk of energy will be going. The bulk of energy will be going to very long running processes that are churning, churning, churning, trying to develop products, trying to develop new software and so forth. So that’s category one is just long running inference. Category two is reinforcement learning. It’s the part of training that is devoted to making the models better by lots of practice effectively, where you want to give the models an opportunity to pursue many different paths. This all happens not working with a user, but as part of the training process, pursue many different paths, see which of the paths were better and worse according to different scoring functions and then feed all of that back into your training clusters on land.
And this is an area that is becoming huge in terms of energy demand. It will be probably already as bigger than what we have historically considered training, pre-training and it’s really important for making the models better at these long running tasks, working together over long periods of time. That’s also an application that’s perfect for our platform because it’s just things that are sitting there trying out many approaches, getting all of those results sent back to shore. So huge amounts of workload that are quite good for our platform. The things that we will not be good for are like traditional training where you need 100 megawatt clusters tightly interconnected all in one place and the super latency sensitive stuff.
Shayle Kann: Yeah. I was going to ask that question on the cluster size thing. Do you view there as being any benefit to clustering to concentrating a bunch of your nodes near each other and networking them? Or should I think of this as the equivalent of is if you were to put a ton of individual couple hundred kilowatt server racks spread throughout the country on land?
Garth Sheldon-Coulson: For most of the applications that I’ve been talking about, you don’t actually get that much benefit from side to side communication among the individual nodes. So you can largely think of it as lots of independent clusters operating on these problems. We do have the ability to provide mesh networking between the nodes when they’re within a certain distance of each other, which is radio side to side, not bouncing up to the satellite. And I think there may be applications for which we do that, but it’s not a major part of the value proposition at this point for most of these workloads.
Shayle Kann: Alright, final question for you. What have you built and what are you building next?
Garth Sheldon-Coulson: Yeah. So the big things that we’ve built so far, we’ve obviously created the core technologies, created the business model. We’ve put a series of full scale prototypes in the water. So Ocean One, which was back in 2019 … No, sorry, 2021. Ocean two, which was 2024, Wavehopper, which was also 2024. So a series of prototypes that have proven computing, hydrogen production, propulsion, operating at sea for reasonably long periods of weeks. And then the next set of devices that we’re putting in the water are part of what we’re calling the Ocean three series. So Ocean three is our first real commercial pilot series. It’s our first series. It’s all on the similar design. It’s all designed for manufacture. So these are systems that are designed to be made using capital equipment that we can put in a very factory-like setting and really turn the systems off of the line.
This series, we start to put in the water in October of this year and it will be 3.1 starting in October, then 3.2, 3.3 by spring or summer of next year, we’ll have an autonomous fleet out there really demonstrating all of these aspects, including propulsion, power generation, inference compute. We’ve done compute in the past, but this will be the first group that’s doing inference compute at sea and all of this is happening simultaneously with all of our payload qualification on land so that when we go to the really full scale system starting in early 2028, we have the payloads ready to go and we can start scaling on those full scale payloads.
Shayle Kann: When you say full scale systems, the distinguishing factor between those and what you’ll be deploying starting later this year is size or something else?
Garth Sheldon-Coulson: Yeah, I should be more clear. So the systems that we’ve put out have been full scale for the North Pacific Ocean here off of Oregon and Washington. When I talk about those future full scale systems, that’s full scale systems for the Southern hemisphere oceans. And so those can be quite a bit bigger, more powerful, very similar design, but much higher power levels and then you can make them bigger so the power levels really go up.
Shayle Kann: Got it. I see. Okay. Garth, super interesting. I mean, you and I have been chatting about this for years and I’ve been excited to see every new iteration of the payloads get out there into the ocean. So I will be excited to see this next wave as well so to speak.
Garth Sheldon-Coulson: Come visit. When we start putting threes out there, then I mean, we can take you out on a plane, we can take you out on a boat, whatever is comfortable for you.
Shayle Kann: I’ll take the boat. Thank you. Cool. Sounds good. Well, thank you so much for the time. This was really fun.
Garth Sheldon-Coulson: My pleasure. Great to see you.
Shayle Kann: Garth Sheldon-Coulson is the co-founder and CEO of Panthalassa. This show is a production of Latitude Media. You can head over to latitudemedia.com for links to today’s topics. This episode was produced by Max Savage Levenson, mixing and theme song by Sean Marquand. Anne Bailey edits the video version of the show. Stephen Lacey is our executive editor. I’m Shayle Kann and this is Catalyst.
