Despite Obvious Challenge

HOST: Welcome back to Stake and Rope, from Goat Security. Today: underwater data centers. The Register has a piece out this week walking through the latest round of operators dropping server pods into the ocean, parking them on barges, or pitching wave-powered server farms in the deep Pacific. Highlander Digital Technology has a 24-megawatt facility running off the coast of China, powered by an offshore wind farm. A Portland startup called Panthalassa has Peter Thiel money and a plan to put autonomous server pods in international waters connected via Starlink. The new wrinkle: this is the second time around. Microsoft ran Project Natick from 2015 to 2020 — eight hundred sixty-four servers in a sealed steel cylinder off the Orkney Islands, came back up with a failure rate one-eighth what they saw on land, declared it a success, and then walked away and built multi-gigawatt land-based facilities instead. Six years later, here we are again. Joining me today: the Legacy Sysadmin, who has watched this exact infrastructure-novelty cycle run at least four times. The DBA, who has thoughts about sealed pods and three-year hardware refresh cycles. The Founder, who I assume already has a thesis. Goat Farmer’s here too.

LEGACY SYSADMIN: [sighs] I racked my first server in ‘86. I have been in this room before. Every five to seven years, someone proposes that the obvious answer to data center cooling is to put the servers somewhere unusual. Underwater. In a desert. In a decommissioned missile silo. In a shipping container so we can drop them by helicopter. The proposal is always serious. The funding is always real. And the production deployment always looks like another conventional data center going up in northern Virginia.

HOST: DBA, what’s the technical layer here?

THE DBA: Cooling isn’t the problem. Cooling has never been the problem. Cooling is between thirty-five and forty-five percent of facility power. So Highlander says they cut cooling power by ninety percent. Fine. What does that do to your cost per rack over five years?

FOUNDER: It cuts your operating cost meaningfully —

THE DBA: No. It doesn’t. Because the hardware in that pod is obsolete in year three. What’s your refresh plan?

FOUNDER: Well, you’d —

THE DBA: You’d what. Drag the pod up. Cut the seal. Replace the servers. Reseal it. Sink it. What’s the labor cost on that? Who’s certified to do it?

FOUNDER: [chuckles] okay, look, I take the point, but you’re describing the old model. The whole thing about a sealed pod is you don’t refresh it on the same cycle. You amortize differently. Highlander, NetworkOcean — these are real operators doing real deployments, and the venture capital flowing into this is not stupid money. Capital Clean Energy Carriers just signed an MoU with Samsung and Lloyd’s Register this month. Smart people are looking at this and seeing something.

HOST: Founder, hold on. The smart-capital-is-funding-it argument — Microsoft is also smart capital. Microsoft funded this exact experiment, ran it for five years, published honest results, and then made a deployment decision that was the opposite of what you’re describing. What did the smart capital at Microsoft see that the smart capital at Panthalassa is missing?

FOUNDER: [exhales] So the answer there is timing. Natick was 2015 through 2020. The AI compute demand curve wasn’t where it is now. The cooling math wasn’t what it is now. The grid constraints on land weren’t what they are now. The world changed. The technology that was uneconomic in 2020 is potentially economic in 2026 because the input variables are different.

LEGACY SYSADMIN: Mm. That’s the same argument I heard in 2009 about containerized data centers. Sun MD. HP POD. Verari. Shipping container with servers in it, plug in power and a network drop, you have a data center. Every hyperscaler bought one to play with. Some of them deployed for specific use cases. None of them replaced the conventional buildout. The argument was identical. The compute demand has changed, the cooling math has changed, the grid constraints have changed, this time the modularity wins.

THE DBA: It didn’t win.

LEGACY SYSADMIN: It didn’t win. Conventional buildout absorbed everything anyway.

GOAT FARMER: Had that one in ‘13.

HOST: Legacy, the article mentions Google’s mystery barges. I remember those.

LEGACY SYSADMIN: Two of them. One in San Francisco Bay, one off Portland Maine, around 2013. Google never officially confirmed what they were, but they were sitting on patents from 2008 for floating data centers, and the barges looked exactly like what the patents described. Both barges were quietly removed within about a year. Google reportedly walked away because the federal maritime safety regulations made the timeline and cost untenable. Which is the same problem Microsoft hit with Natick on the maintenance side, just earlier in the lifecycle.

THE DBA: What’s the SLA on a sealed pod three kilometers out?

FOUNDER: Highlander’s pod is only a few hundred meters offshore.

THE DBA: Okay. What’s the SLA?

FOUNDER: I don’t have that number in front of me.

THE DBA: Right. Because the customer doesn’t have it either. China Telecom signs the contract. State-owned AI computing firm signs the contract. They get told it’s a data center. What happens when a rack fails in month fourteen?

FOUNDER: The pod is designed to tolerate component failure without intervention. That’s the Natick result. One-eighth the failure rate. The nitrogen atmosphere and the lack of human hands inside the box is why the failure rate is so low.

THE DBA: Fine. So one-eighth the failure rate on hardware that’s still going obsolete in year three. You haven’t solved obsolescence. You’ve solved hardware longevity, which wasn’t the bottleneck.

HOST: Let me push on something. Founder, you mentioned the AI compute curve. Meta is building Hyperion, which is a planned multi-gigawatt cluster. Highlander’s underwater facility is 24 megawatts. That’s about half the average US data center, and it’s a factor of forty or fifty smaller than what Meta is targeting for a single site. The hyperscaler trajectory is up on scale, and underwater pods don’t go up on scale — you can build a bigger pod, but you can’t build a Hyperion-class pod because at some point the seal, the power feed, and the heat dissipation curves all break. Where does waterborne win against that trajectory?

FOUNDER: It doesn’t win against Hyperion. Nobody serious is claiming it does. But Hyperion isn’t going in Singapore. Hyperion isn’t going in Tokyo Bay. Hyperion isn’t going where land is the scarce resource, and there are more of those places every year. Half the world’s population lives within 120 miles of the coast. The latency case for coastal compute is real. If you’re land-constrained and grid-constrained, you have to do something other than the conventional buildout. Waterborne is one of the things you can do.

HOST: That’s the population-centers argument. That was also true in 2015 when Microsoft started Natick. It didn’t translate then. What changed?

FOUNDER: Demand changed.

LEGACY SYSADMIN: Demand was changing in 2015 too. Mobile, video, streaming. Every cycle has a demand story.

HOST: Goat Farmer.

GOAT FARMER:

Reason number 27. Just about any barnyard animal is fault tolerant, except some cows.

HOST: Let me ask the harder question. The thing I keep circling. Microsoft ran the experiment for five years. They had the world’s best engineering team on it. They published peer-reviewed results showing the technology worked better than the land-based control. And then they declined to deploy it at production scale. The article says — quote — Microsoft chose not to proceed any further with submerged server farms, and the company never fully explained why. DBA, what do you think they didn’t explain?

THE DBA: They explained it in their next ten-K. They built thirty data centers on land. That’s the explanation. The decision is the artifact.

LEGACY SYSADMIN: The thing that doesn’t get said in the press release is the operational layer. I have spent thirty years watching organizations decide what’s deployable. The technical proof of concept is never the constraint. The constraint is always: can we hire the people to run this at the third site, and the fourth site, and the fortieth site. Microsoft has a workforce of conventional data center operators. They don’t have a workforce of marine engineers. To deploy Natick at scale, they would have had to build a parallel operations organization specialized in marine deployment, sealed-pod maintenance windows, coastal grid integration, and underwater fiber. That’s not a technology problem. That’s an organizational depth problem.

THE DBA: And the people who would build that organization want to work on AI now.

LEGACY SYSADMIN: Right. Which is the second-order effect. The marine engineering talent pool isn’t growing toward data center operations. It’s growing toward offshore wind, which is the same labor market.

HOST: Highlander is collocated with an offshore wind farm.

LEGACY SYSADMIN: Of course they are. The wind farm has the maintenance infrastructure, the boats, the certified divers, the coastal grid interconnect. The wind farm is the actual capital expenditure. The pod is a tenant.

THE DBA: That’s the model that might work.

HOST: Say more.

THE DBA: The pod isn’t the business. The pod is a tenant on infrastructure that exists for a different reason. Wind farm operator has the boats and the divers and the grid tie. They lease pod-space to a data operator. The pod operator doesn’t have to build the marine organization because they’re renting it. That’s the only version of this where the unit economics work, because somebody else is amortizing the marine costs against a different revenue stream. Same way colocation worked in the late nineties. You weren’t paying for the building. You were paying for the building someone else already built.

FOUNDER: That’s actually a really sharp framing. The pod-as-tenant model. I want to write that down. Because if you look at what Capital Clean Energy Carriers is doing with Samsung, that’s exactly the play — they’re a shipping company that already has the vessel ops, and the data center is a tenant on the ship’s existing systems. Power, cooling, water intake, crew. The marine costs are amortized against the shipping business.

LEGACY SYSADMIN: Sure. And that’s the version that has a chance. The Panthalassa version, where you build standalone wave-powered autonomous pods in deep water connected by satellite — that’s the version that’s the same experiment Microsoft already ran, plus weather, plus latency, plus no human within four days’ steam.

FOUNDER: Panthalassa is the moonshot end of the spectrum, sure. I’m not defending Panthalassa specifically. I’m defending the category.

THE DBA: [scoffs] sure. The category includes wave-powered autonomous pods in the deep Pacific with Starlink uplinks.

FOUNDER: The category includes the Singapore case, the Tokyo Bay case, the Marseille case where Nautilus actually shipped, the Stockton case that’s still operating. Real deployments at real scale for real customers. You can’t just point at the weirdest pitch in the category and dismiss the category.

HOST: That’s fair. Nautilus did ship. Stockton is still up. Six and a half megawatts of IT load. Not Hyperion, but operational. Goat Farmer.

GOAT FARMER: I don’t miss any of it.

HOST: Let me land this in one more place before we go to closing thoughts. The article frames this as datacenters dipping a toe back into waterborne computing despite obvious challenges. Legacy, what does the long pattern tell you about the word “back” in that sentence?

LEGACY SYSADMIN: [sighs] The word “back” assumes there was a moment we left. We didn’t leave. We tried, we measured, we wrote it up, and we made a deployment decision. The current cycle isn’t a return. It’s the next group of operators showing up to take the same class. [pause] The institutional memory in this industry runs about five to seven years, which is roughly the tenure of a senior infrastructure engineer at one company before they go somewhere else. So the lessons reset on that cadence. The reason every cycle repeats is the same reason every cycle repeats. The people who learned the lesson last time aren’t the ones making the decision this time.

HOST: Let’s go to closing thoughts. Goat Farmer first.

GOAT FARMER:

Reason number 34. As long as the stable hasn’t caught fire, a goat couldn’t care less about a power surge.

HOST: Founder.

FOUNDER: Look, I’ll concede the scale point. Waterborne isn’t going to compete with Hyperion. It’s not the answer for everything. It’s the answer for the places where land is the constraint, where the grid is the constraint, where the cooling math actually moves the needle. Singapore. Tokyo. Marseille. The coastal megacities. And I think the pod-as-tenant model the DBA just sketched is the version that survives — somebody else owns the marine ops, the data operator is a customer on the platform. That’s not the moonshot version. That’s the boring version. The boring version is usually the version that actually ships. I’ll write the partial-fit case up tonight.

HOST: Legacy.

LEGACY SYSADMIN: I’ve watched infrastructure-novelty cycles run since the eighties. Ruggedized Microvax deployments in the late eighties. Solar-powered data center pitches across the two-thousands. Decommissioned missile silos in 1998. Containerized data centers in 2009. Google’s mystery barges in 2013. Natick from 2015 to 2020. Each cycle has a real technical proof of concept. Each cycle has real funding. Each cycle has a real partial-fit case where the niche deployment actually ships. And each cycle ends with the conventional buildout absorbing the overall workload anyway, because the unit economics of the conventional model keep improving faster than the novel model can mature. The reason this cycle is going to look like the previous five is that the gap between proof-of-concept-economics and production-scale-economics is structural, not technical. Microsoft didn’t fail to translate Natick. They observed the gap and made a rational decision. The next operator to observe the gap will make the same decision. Probably around 2031.

HOST: DBA.

THE DBA: The technology works. Microsoft proved it. One-eighth the failure rate, five years of operation, peer-reviewed results. Nobody has to re-prove it. The technology working was never the question. [pause] The question is whether the deployment economics work, and that’s a different question entirely, and Microsoft answered that one too, by what they built next. Thirty conventional data centers. Zero underwater. That’s the answer. The gap between the technology works and we should deploy it is the whole story. The gap closes when both sides converge. The technology side has been ready since 2020. The economics side hasn’t moved. Until the economics side moves, every new operator that picks this up is going to run the same experiment and reach the same conclusion. They might publish it differently. They might not publish it at all. But they’re going to reach it.

HOST: Microsoft proved the technology works and proved the deployment economics don’t, and the next cycle is running the same experiment with the same results pending. We’ll see you next time.