How Panthalassa's Wave-Powered Ocean Nodes Run AI Compute at Sea

Photo by david giling on Unsplash

Most coverage of ocean AI data centers describes Panthalassa's nodes as floating power plants that send electricity back to land. They don't. The electricity never leaves the buoy. The whole point of the design is that moving electrons across miles of seawater is the hard problem, so Panthalassa skips it and moves bits instead.

That single design choice changes what the company actually is. It's not a wave-energy utility competing with offshore wind. It's a data center operator whose buildings happen to float, generate their own power, and have an unlimited heat sink wrapped around them. Once you see it that way, the rest of the engineering starts to make sense.

The misconception: this is not offshore wind for AI

Offshore wind farms exist to push gigawatts of power back to a coastal substation through expensive subsea cables. Panthalassa's economics fall apart the moment you try that. Wave energy density is highest hundreds of miles from shore, and the cost of running HVDC cable that far would erase any advantage over a land-based gas turbine.

So the company inverted the equation. According to Panthalassa's funding announcement on BusinessWire, the autonomous floating nodes generate electricity from ocean waves, run AI inference workloads onboard, and transmit tokens back to land by low-Earth-orbit satellite, avoiding grid interconnection delays, cooling requirements, water use, land availability, and local permitting. The output of an inference query, a few kilobytes of generated text, is trivially small compared to the megawatt-hours of energy that produced it. Send the answer, not the electrons.

CEO Garth Sheldon-Coulson framed the energy thesis bluntly. As reported by Converge Digest, Sheldon-Coulson identified solar, nuclear, and the open ocean as the three energy sources on the planet with tens of terawatts of new capacity potential. The open ocean is the only one of the three you can colonize without a permit fight.

What the hardware actually looks like

Picture a steel structure riding on the surface, with a long neck plunging straight down beneath it. The shape uses wave motion to continuously force water into a pressurized internal reservoir, pumping it through the neck as the structure rises and falls. There is one turbine and effectively no other moving parts.

That geometry matters. A conventional wave energy converter has hinges, articulated arms, or oscillating columns, all of which fail in heavy seas. Panthalassa's node is closer to a giant lollipop with a single hydraulic loop inside. According to Data Center Dynamics, the platform generates energy via water forced through an internal turbine as the unit rises and falls with waves, with the generated power running onboard computing hardware that processes AI workloads and transmits results back to shore via satellite.

The design also throws out the most fragile part of every previous wave energy project: the mooring. Data Center Dynamics reports that the platform has no anchors or cables connecting it to land or the seabed, acting as a self-contained unit. In an interview with CBS News, Sheldon-Coulson described the Ocean-3 as a self-propelled system without an anchor, comparing it to "a little Roomba, except it's enormous." A node that can navigate avoids storms instead of fighting them.

Why seawater is the secret weapon

Land-based AI data centers spend roughly a third of their energy budget on cooling. Phoenix and northern Virginia operators are already losing fights with local water boards over evaporative tower withdrawals. The ocean removes that line item entirely.

The BusinessWire announcement describes the nodes as autonomous, floating energy systems mass-produced from plate steel in coastal factories, with the surrounding ocean providing free supercooling for onboard AI chips. The North Pacific sits at roughly 4 to 12 degrees Celsius year-round at the depths where heat exchangers would operate. That is colder than any free-air cooling regime on land, and it does not vary with the weather.

The non-obvious benefit is chip lifetime. Hyperscale GPUs degrade faster when they cycle between hot and cold; a constant 6°C cold-plate intake is gentler than the thermal swings of a Texas summer. Panthalassa argues this extends silicon lifetimes and addresses a core land-based data center challenge, though the company has not yet published field data to back up the claim.

Inference, not training

This is the part most readers get wrong. A frontier model training run synchronizes tens of thousands of GPUs across a single fabric with sub-microsecond latency. You cannot do that across a satellite link, and you cannot do it across a fleet of buoys drifting hundreds of miles apart.

What you can do is inference. Once a model is trained, every query is independent: a user sends a prompt, a GPU runs it, a response comes back. Latency tolerances are measured in hundreds of milliseconds, well within the budget of a Starlink-class link. Per the BusinessWire release, rather than transmitting energy to terrestrial grids, Panthalassa uses the power directly onboard to drive AI chips and sends the resulting inference tokens to land by satellite.

This explains the choice of investors. The BusinessWire announcement lists the $140 million Series B as led by Peter Thiel, with participation from John Doerr, Marc Benioff's TIME Ventures, Max Levchin's SciFi Ventures, Super Micro Computer, and returning backers Founders Fund and Lowercarbon Capital. Super Micro is on that list because it builds the inference racks. The economics only close if Panthalassa is selling tokens by the million, not megawatts by the hour.

The numbers, and the ones to be skeptical of

Panthalassa's most striking claim is on cost. Data Center Dynamics reports the company asserts power generation costs potentially as low as $0.02 per kWh if scaled. That would undercut every utility-scale solar PPA signed in the United States in 2025. It would also require a manufacturing volume the company has not yet demonstrated.

The verified facts about scale and timing are more grounded:

• Funding: $140 million Series B at a valuation reportedly approaching $1 billion, per TechStartups citing the Financial Times.
• Headcount: Panthalassa was founded in 2016 and has grown to 120 employees, with new funding earmarked to complete a pilot manufacturing facility near Portland, Oregon, GeekWire reported.
• Track record: The BusinessWire release cites Ocean-1, Ocean-2, and Wavehopper prototypes tested at sea in 2021 and 2024.
• Schedule: Ocean-3 pilot nodes are planned for the northern Pacific Ocean in 2026 ahead of commercial deployment in 2027.

The $0.02/kWh figure deserves the same scrutiny SpaceX's early launch-cost projections did. It might be directionally correct. It will not be the price of the first hundred units.

What the open ocean actually does to hardware

The physical environment is the part of this story that gets glossed over in funding coverage. Data Center Dynamics flags saltwater corrosion, biofouling, and operational reliability during storm conditions as primary engineering risks, with repairs in open water adding further cost.

Biofouling is the unglamorous one. Within weeks of submersion, a steel hull in the North Pacific picks up barnacles, algae, and tube worms that change its hydrodynamic profile and clog seawater intakes. Every offshore oil platform spends millions a year on antifouling coatings and divers. Panthalassa has not publicly described how its nodes will handle this over a multi-year deployment.

Then there are storms. The North Pacific routinely produces wave heights above 15 meters during winter cyclones. The company's answer is autonomy: a node that can move under its own power can theoretically steer away from the worst weather. CBS News reported that Ocean-3 units are expected to be operating offshore by around August 2026, with the company hoping to eventually deploy thousands of nodes far out at sea.

The team has the resumes for the work. TechStartups reports Panthalassa draws on veterans from SpaceX, Tesla, NASA, Google, Apple, and Blue Origin. Co-founder Brian Moffat developed wave energy technology for Spindrift Energy before launching the company, GeekWire reported.

The regulatory story almost nobody is telling

Operating in international waters is being marketed as a permitting advantage. It is also a policy frontier nobody has mapped. TechStartups notes that regulators are likely to examine impact on marine ecosystems, shipping routes, and communications before approvals.

A fleet of floating compute nodes raises questions UNCLOS was not written to answer. Who has jurisdiction over data processed in international waters? How do export controls on AI chips apply to a node flagged in one country, manufactured in another, and operating outside any EEZ? What happens when a node carrying GPUs subject to U.S. export rules drifts into the territorial waters of a country those rules restrict?

Panthalassa is structured as a public benefit corporation headquartered in Portland, Oregon, per the BusinessWire release, which may help with stakeholder reviews but does not preempt environmental or maritime regulators. The next two years are when the regulatory questions stop being theoretical.

How this fits into the broader AI infrastructure picture

The Series B lands in the same six-month window as comparable unconventional compute infrastructure bets. GeekWire notes that space-based data center startup Starcloud raised $170 million in March 2026 at a $1.1 billion valuation. The connecting thread is that AI demand has outrun the grid's ability to provision new capacity, and capital is now flowing into every plausible workaround, including exotic compute substrates that would have been laughed out of a 2020 pitch meeting.

The ocean bet is the most pragmatic of the three. Space is harder to service. Nuclear is slower to permit. The Pacific is already there, already in motion, and already cold. As Converge Digest frames the company's positioning, traditional nearshore wave energy generates less power, while Panthalassa is targeting more energy-dense offshore zones.

Whether Panthalassa specifically wins is a separate question from whether the architecture works. The 2026 Ocean-3 pilot will produce the first real public data on capacity factor, biofouling rates, and inference economics in salt water. Until then, the company has built the most coherent argument anyone has made for moving AI compute off the grid entirely, and a $140 million check from Peter Thiel saying it deserves to be tested at scale.