Panthalassa, a startup developing autonomous floating platforms that perform AI inference using power generated from ocean waves, has raised $140 million in a Series B led by Peter Thiel. New investors include John Doerr, Marc Benioff’s TIME Ventures, Max Levchin’s SciFi Ventures, Susquehanna Sustainable Investments, Korea’s Hanwha Group, and Australia’s Fortescue Ventures. The Financial Times framed the company at the $1 billion mark. Panthalassa will deploy its first Ocean-3 pilot nodes in the northern Pacific in 2026 and commercial nodes in 2027.

Three Constraints, Pick Two

An AI data center built on land in 2026 has three problems. Power: interconnection queues at every major US grid operator run into years. Water: cooling a hyperscale site uses tens of millions of gallons a day, and the host community usually has opinions about that. Permits: the host community usually has opinions about everything else too.

An ocean platform makes the three problems disappear. There is no grid because the platform generates its own electricity from wave motion. There is no water bill because the surrounding ocean is the heat sink. There is no community board because there is no community.

In their place are a different set of problems. Marine engineering is hard. Saltwater is corrosive. Service crews need boats. And the only way to move data on and off the platform is by satellite, which is bandwidth-constrained and latency-bound.

The argument for going to sea is that the constraints accepted out there are easier to engineer through than the constraints faced on land. The numbers Panthalassa is putting in front of investors are a 90 percent capacity factor (against roughly 40 percent for offshore wind and 25 percent for onshore solar) and a potential generation cost of $0.02 per kilowatt-hour. Whether either number survives contact with commercial deployment is the question.

Inference-Only Is Doing All the Work

The most important word in the pitch is “inference.” Training a frontier model requires keeping tens of thousands of GPUs in lockstep at millisecond latency. That cannot be done over a satellite link. So training stays on land. Inference is a different workload. A request arrives, the model runs, tokens go out. Inference is usually tolerant of batching, queuing, and a few hundred milliseconds of latency that would be fatal to training.

So Panthalassa is making the same bet that the Amazon-Anthropic $100 billion structure is making, that the eight-vendor Pentagon deal is making, and that the specialist-tier neocloud build-out is making: the marginal compute demand of the late 2020s is inference, not training. Training stays on land at a handful of frontier campuses. Inference goes wherever the power is cheapest. If the cheapest power is sitting in the middle of the Pacific Ocean, the inference goes there.

Floating Data Centers, A Field

Panthalassa is not alone in the category, just the best-funded US entrant. Nautilus Data Technologies runs barge-based data centers cooled by surrounding water; its Stockton site has been operating for years. Subsea Cloud sells pressure-equalized pods that sit on the seabed. Microsoft’s Project Natick demonstrated subsea operation in 2018 and was retired. Google’s floating data center concept was retired earlier.

The company actually operating commercial subsea AI compute at scale is Chinese. Highlander, in partnership with state offshore-engineering company COOEC, has been running the Hainan Underwater Intelligent Computing Center off Sanya since late 2023. The cluster is currently two 1,300-ton sealed pressure vessels at 35 meters’ depth, with roughly a thousand servers between them. The roadmap calls for 100 modules within five years. The total project cost is approximately $880 million.

Panthalassa just raised $140 million for a 2026 pilot. Highlander has been processing AI workloads commercially under the South China Sea for more than two years. Whether wave-powered, untethered, satellite-uplinked autonomous nodes turn out to be technically superior to anchored capsules wired to shore is unknowable today. What is knowable today is that one country has subsea AI infrastructure running, and the other has just written its first nine-figure check toward building one.

That is consistent with how the rest of the buildout is going. Land compute is constraint-bound, and the response in 2026 is to find ways around the constraints. Decommissioned bitcoin mines. Fiber-carrier campuses. Hyperscaler joint ventures. Sovereign neoclouds. And now floating platforms in the open Pacific. The buildout is happening in every place that has a constraint someone is willing to engineer around. The Pacific Ocean is just the latest one.

Sources: BusinessWire, Data Center Dynamics, Dataconomy, TechRadar, TechCrunch, DCD on Highlander, Light Reading

–
By the Control Plane Editorial Team