Panthalassa is building ocean-based energy platforms to power distributed AI compute and synthetic fuels

next guest G from Panthalosa building a very cool technology. I'll let him tell you about. How are you doing? What's going on? Hey guys, great to meet you. Thanks so much for being on the show. Love the show. I'm really excited to have you. Uh I have heard about your company. The background. What's going on here?

Is that is that real? Yeah, this is real. We're in Portland right on the river. Wow, that's beautiful. It almost looks tropical with how the lighting looks. Yeah. Very cool. And you have the plant inside the window, so it adds adds to the ambiance. Looking great.

Uh would you mind a quick introduction on you and your and and the company and then we'll go into it. Yeah. Yeah. So company is called Panalasa.

based in Portland about 80 people a lot of people out of places like SpaceX uh Blue Origin Virgin Orbit it's it's a deep tech company love the conversation with Sean just now and you know the the premise of the company is to go to the middle of the ocean and capture energy there very cheaply so this was an idea that got started nine years ago um I was at Bridgewater Associates uh doing investment stuff but before that I had been doing energy stuff at MIT and and working on a startup for energy and my co-founder had been also looking at the ocean and we were saying to ourselves, you know, what if you could find an untapped energy source on the planet uh bigger than many of the ones that people have been going for and uh and maybe more scalable, maybe lower cost of capture.

So, the one that really stuck out to us was the open ocean, far from shore, uh, where there is wind all the time and the wind makes the waves and so you can be out there, uh, capturing hypothetically what we thought back then was at a very low cost and and then you would, you know, use the energy right on site for things.

Mhm. Um and so this was sort of an era before um Cruso, but maybe a similar idea that you go to where the energy is and and we thought that this was really the best place to go, the middle of the ocean.

And so of course, of course, you need to develop a tech stack for that, but there's lots of interesting properties there that you can talk about.

And and what we thought was that maybe you could build a really simple system like you know a lot of the stuff that's been done in ocean energy before was like pretty complicated and pretty expensive.

Um so we thought that if you could develop a really simple system then you'd start to get a lot of those a lot of those properties. I'm hearing simple and nine years it sounds like those are some famous last words. You've been grinding on this for a long time.

Um but it sounds like there's there's actually breakthroughs. I you I saw some photos. You've actually built one of these things.

Can you explain how big it is, how much energy it captures, what you're deploying, what the rollout schedule is because uh it's been a lot of work to get here to to where you are today and I want to hear about like where we are before we go into where you're going. Yeah. Yeah.

I actually think that that length of time is a really important part of any company that's trying to do deep tech. Totally. Um you know, you want to stay small and under the radar while you're sorting everything out, right?

I think that a mistake that a lot of companies make is that they go to market with their first idea or their second idea. And so if you can arrange things so that you can throw things out a lot before you settle on a path because the fact is you are going to get it wrong.

You're going to need to iterate and um particularly when you're trying to do something where the cost is the key consideration or if it's space maybe it's weight. Um you want to give yourself that time.

So, so when I look back on the first four years of that, which was really the heavy R&D time, I'm extremely glad that we took that time to hone the technology and get it exactly right. Um, and so by the time we got to year five and six when we were deploying ocean systems, we were right on the money in terms of power.

You know, we were hitting all of our simulations right away. And uh it comes from the simplicity of the system where what we decided was that you want a system that is no moving parts at all except for one small water turbine. So these systems we call them nodes.

They're about 10 up to 20 meters across at the top and they go down into the water column about 80 meters maybe 100 meters for full scale ones. So so big systems um but pretty small in the scale of things like ship building. So so this is this is sort of the shape of one of them here.

It's a, you know, and it sits in the water like this. Um, and it goes up and down with the waves and all of the dynamics are coming from the inertia of water within that system. And as the system goes up and down, it pumps the water into a reservoir, uh, just like a hydroelectric dam.

And then from there, it drives a water turbine. So, so it's a really simple method of operation and once you have that, you can simulate it effectively and it's really easy to massproduce as well.

So, you keep saying you keep saying simple and I'm looking at your website and it's like the most complicated thing I've ever seen, but I mean it speaks to it speaks to your background and and and how ambitious this project is, but I keep thinking you're you really simple stuff, John. Yeah. Yeah. It's simple.

Just just just weld and build this massive thing. And there's all this wiring and we're pulling up your website right now and it's fantastic. I mean, congratulations. It's amazing. Um I I I want to know more about the applications.

how much energy are we are we uh generating with one of these and then um and then I want to go into some of the applications because it feels like you could not have possibly predicted the the the reinforcement learning revolution that's happening right now the ability to do uh distributed AI training and and where like it is the perfect time for this thing and uh maybe you just made your own luck but I want to know about the evolution of the application uh the stranded energy thesis and kind of uh the like like the the scale of what we're thinking about doing here.

Yeah. Yeah. Great. So, each one of these systems makes anywhere from 200 kilowatts to a megawatt. So, you can build anywhere in that scale. Got it.

And um you know, when we got going, it it was interesting because, you know, obviously we were jett jettisoning the cables back to shore and the moorings and all of that stuff. And we said to ourselves, you know, this was back in 2017 or so.

uh what are the applications for which this can be very effective and we we decided on two platforms. Uh one is computing and that's the primary and first platform and then the other is making synthetic fuels and getting the fuels back to shore.

Uh but if you look at our patents and everything back from 2017 2018 you know we have we have a whole bunch of them on computing. Uh and you know every back then we were thinking about AI you know often in the patents we would call things brain simulation so we had AI brain simulation crypto. Sure.

Probably Bitcoin at a certain point. Crusoe was doing the same thing. Yeah. Exactly. And you know I had been at Bridgewwater. I'd been doing AI stuff with Dave Fuchi there. You could see even at that time where this was going. You could see it in the exponential growth of energy.

You could see it in the growing power of the models even then. You know this is when AlphaGo was happening. So, you know, you could sort of put two and two together. And, um, what we believe, of course, is that, you know, this is going to 100 gawatt scale, terowatt scale.

Uh, and there's not that many options on the planet for making that go really fast, at least if you're not doing something like coal in China. Uh, so, so we are positioning ourselves to be an option that can scale really rapidly, much much more rapidly than terrestrial infrastructure.

Um, and and that's that's really the whole play. We want it to be, you know, more sustainable, more affordable. It's much cheaper, uh, and be able to scale to much larger levels even than, you know, than most terrestrial infrastructure could. Wow. Got it. So, uh, one, uh, you said up to one megawatt. Is that right?

Yeah. Yes. The the nominal system might be 400 kilowatts, but you can go larger than that. So, like how how big of a are you thinking like it's basically a data center inside? Like you could put a whole bunch of H200s, maybe get up to like a few thousand maybe. Is that roughly correct?

Yeah, ex exactly like a few hundred uh up to a thousand GPU equivalents and there's enough room in that in that device to actually run. Yeah, I mean that's amazing. So um what what are you uh what are you tracking in the in the reinforcement learning training market?

Is that important because you're at this point where you're actually going to market?

Um and and will it even like are there specific decisions that you make need to make on the equipment that you load onto these machines or uh or are you trying to serve the GPU capacity or or or data center capacity at like a higher level of abstraction? Exactly.

So so those are those are um those are sort the sort of questions about the go to market. But as you identified, there's, you know, the the massive push towards test time compute, just tons of longunning inference workflows all the time. Um, combined with Yeah.

I mean, we think it's going to be mostly RL on the training side, you know, and I mean, I'm sure you had people on last week and you saw the semi analysis thing drop. Yep. Was great yesterday.

But yeah, I mean you could sort of put it together even when you had alpha go back in the day of like you put together LLMs with reinforcement learning you get very powerful models.

So the future is largely about who can deploy the most RL and then once people really start to use this you know we think it's going to move to um all kinds of applications beyond text you know this has to move to engineering it has to move to medicine um and you're going to want to have agents working together around the clock so you have longunning inference and and basically what this all comes down to is you want the cheapest throughput for tokens and so what we're targeting is the that long tail of very massive horizontal scaling of distributed energy units.

Um, and that's what we that's what we think our platform can be. How over time would somebody live on one of these systems? Would they visit it frequently? We had Dylan Patel on last week and he said one of the challenges with putting data centers in space is that chips are unreliable.

Sometimes you need to like plug, you know, basically unplug a system, plug it back in. It's the meme. It's the IT meme, you know, unplug and plug it back in. Basically, you know, even with cheap energy, the IT work will still be necessary.

So, I'm I'm I'm wondering is one person going to be responsible for a number of these systems and they sort of visit them and make sure they're operating efficiently or um there's probably other solutions to Yeah, I can imagine you're building like some type of like gyroscopic tech to kind of keep the system stable when they're, you know, shifting up and down.

Maybe that doesn't even matter. I don't know. Yeah, but I'm I'm curious. It it's it's really important. So, this is part of the tech. So, you not only need a way to produce the power on board really simply and cheaply, you also need to be able to drive them around. So, these are these are self-propelled systems.

Um, and they're self-propelled without things like propellers. So, they create their own thrust just because of their shape. So, we've done trials where we're driving these things in figure8s out there and it's all happening just because of the shape of the system. Wow.

And what that allows you to do is now this is sort of like SpaceX boosters coming back. You can deploy the systems relatively close to shore. They work their way out into the resource. They hang out for a period of time capturing energy and you know maybe it's three years, maybe it's five years.

When it's time for a payload swap, they bring themselves back in and then you swap the payload. That's remarkable. So you can configure that. Yeah, you can configure the payload swap time for whatever you need to get the numbers, you know, for the meantime between failure on your payloads.

Y um and at the beginning, I'm sure we'll start with the most reliable stuff. Uh and then as we get into worlds of inference AS6 and stuff like that that are you know more reliable than the GPUs then we start deploying those as well. Yeah that makes sense.

Uh are there are there lessons that you're learning from like the the Bitcoin ASIC era uh like like other kind of out ofthebox stranded energy projects. Crusoe was kind of widely known for the I think it was gas beaker plants initially, but obviously those are are very much terrestrial.

You can just drive the ATV or the the truck out to them to check on them. Um, what what what have you seen or learned from the uh the more uh the more creative uh energy solutions to the Bitcoin ASIC problem? Um, well, I mean, I think the big theme of course is everyone shifting towards AI energy.

um just because the the amount of demand that's coming is enormous. Um Bitcoin is you know Bitcoin is always going to be there as a fallback for anyone who's got very cheap energy on board our systems will have you know between two and three cents per kilowatt hour probably less in the future.

So that's always an option for us. But um but I I think that you know as as people look at this it becomes clear that the the model layer will become more commoditized.

I think there's a lot of comp comparative advantage there for a long time chip layer too but the fundamental bottleneck is going to be who can scale tens of gigawatts per year or even 50 gigawatts per year. You know that's that's a really hard thing to do in the west.

Um obviously we've been doing net zero gigawatts per year for a long time. So if we don't relax that constraint then you get prices going up and that becomes a major impediment to the to the buildout. I mean so the device already is massive. If we're adding a gigawatt you're talking about making a thousand of these.

walk me through the actual manufacturing, how you scale that up to to to go from, you know, kind of one exquisite system to making dozens of these every month to every day. Like the the roll out needs to be uh intense.

Um I'm sure there's a lot that you're learning from from what's going on at SpaceX and how quickly they've made those massive uh Starship rockets. It always surprises me when there's a new one on the pad because I watched it blow up just a few days ago and they have a new one ready.

Uh, and it really speaks to their ability to manufacture something at scale. Uh, both in terms of speed and just the size, you know, it's it's it it's one thing when you see the Statue of Liberty.

They built that once, but it's a different thing when you see a new Statue of Liberty come off the manufacturing line every single day. So, uh, what needs to change at your company in terms of process? What are what are the requirements to make big things fast and regularly? Yeah.

So, so this comes back to the design of the system. So, it's the system is just this really simple steel shell. It's like entirely hollow inside and it's made with mild steel. It's got this axiymmetric shape so that there's a central axis and all of the whole system is built around that central axis.

So, that allows you to use equipment that lets you make these concentric shapes, put them together, weld them really simply in a lot of ways. I mean I keep saying simple but in a lot of ways it's actually simpler than a car. Oh interesting.

And uh you know this just comes back to the design of it but what we can go and do because it's that way is we can cost all of the capex that you need in a factory for these things. We know how many people will be involved. We know how much land it'll take up.

And so even now we can say that look if we were to go and build a factory of a certain capex investment you know let's say you built a factory at the scale of a gigafactory in terms of cost how many gigawatts would that spit out and for example we know that if we built factories with the amount of capex that Tesla has put into g gigafactories we'd be spitting out 20 gawatts a year of capacity so the the manufacturing story is really nice these are still small enough to like mass produce.

It's not ship building. Um, and it's not people crawling all over them. It's things that you can to a large extent automate. So, we're actually building our first uh pilot line here in the Pacific Northwest. Parts of it will be operational in January. And the whole thing should be operational later in the spring.

And then once that's going, then we can really start to to carbon copy that elsewhere. It looks and feels so sci science fiction book. It really It really does. It's extremely sci-fi. Um, how do you actually make these shells?

Like I, you know, Tesla has like the Giga Press or something like that that can form aluminum or stainless steel into a particular shape. Once you get so big, I don't even know. Is there a bigger machine that just kind of will stamp these things together?

Do they have to be bent or rolled out of some sort of is it is it a mold? I don't even know how it gets made. Yeah. Yeah. it. So, so if you're using mild steel, the key thing is to keep the surfaces single curvature as much as possible.

Um, so you get to use just like plate rolls and basically they're like big pasta rolls that make, you know, cylindrical shapes and then conicle shapes. Um, and and that's just, you know, it's it's just some metal rolls and you push the metal through and it comes out in the right shape.

then you can put that into your jig or put that into your what's called a growing line configuration that that lets you align them and then use an automated welding head to weld them. Um so so luckily a lot of this stuff has been developed. It's it's pretty well understood. They use it for like wind turbine monopiles.

They use it for other axismmetric shapes like that. So we get to piggyback, you know, sort of piggyback on a lot of the Oh, that makes sense. Yeah, it does kind of have the same shape as like one of those massive wind turbines. Uh the the huge windmills that you see. I mean, it it's a fantastic project.

It's so fascinating to be working in this industry. Um, I'm sure tons of lessons learned and a lot more good stuff in the future. But thank you so much for stopping by. This is fantastic. Yeah, we're going to our our power demands are immense. Pretty pretty immense doing a lot of reinforcement.

Expect us to fill out the form on the site to to get one of these uh ASAP. Yeah. Awesome. Awesome. Well, we'll come back on when we're when we're kicking out those first nodes from this from this factory here. We love that. Thank you so much. Awesome. Thanks for joining, G. Have a great one. You too.

Uh we have our next guest in Yeah, it's wild. Just the imagery of seeing people like, you know, just climbing on top of it. You don't really realize the the scale. It's simple. It's simple. It's just it's a small little thing really downplaying the size of massive. I love it.

Uh anyway, uh if you're heading to Antarctica or, you know, anywhere else to travel the world, uh go take advantage of some lowcost energy across the globe. Book a wander. Find your happy place. Find your find your happy place. Even gold retrievers joined in. It's fantastic.

Hotel grade amenities, dreamy beds, top tier cleaning, 247 concier service. It's a vacation home but better, folks. Use code TBPN when you sign up. Um do it. Our next guest is coming in to the studio uh from