The AI industry has an energy problem it can no longer hide. Data centers are projected to consume roughly 1,050 TWh globally by 2026 — enough to rank them as the fifth-largest energy consumer on Earth, wedged between Japan and Russia. Every major AI lab is scrambling for power, and the grid can’t keep up.

Panthalassa thinks the answer is floating in the ocean.

A Giant Lollipop That Thinks

The Vancouver, Washington-based company has spent a decade in semi-stealth building something that sounds like rejected sci-fi: autonomous, self-propelled data centers that ride ocean waves, generate their own electricity, and beam results back to shore via Starlink.

Their latest design, the Ocean-3, is roughly 20 meters across at the top and plunges 80 meters into the water. As it rises and falls with waves, water funnels through internal channels, pressurizes, and spins a turbine. Servers packed with GPUs sit onboard. The ocean provides both the power and the cooling.

No anchor. No cables. No connection to the seafloor whatsoever.

CEO Garth Sheldon-Coulson describes it as “a little Roomba, except it’s enormous.” The Ocean-3 navigates up to 30 miles per day, positioning itself in optimal wave conditions. AI workloads get processed at sea. Results travel via satellite.

Why This Isn’t as Crazy as It Sounds

The knee-jerk reaction is skepticism — and honestly, wave energy deserves it. Despite decades of investment, it’s never achieved the cost curves that solar and wind have. Saltwater corrodes everything. Storms destroy equipment. Maintenance at sea is a logistics nightmare.

But Panthalassa has a few things going for it that previous wave energy ventures didn’t.

First, the demand is real and desperate. U.S. power demand is projected to hit a record 4,244 billion kWh this year. Data center capacity is expected to nearly double between 2025 and 2028. AI companies aren’t just looking for power — they’re begging for it.

Second, Microsoft already proved the ocean works for computing. Project Natick sank 864 servers to the seafloor off Scotland from 2018 to 2020. The result: a failure rate one-eighth that of land-based equivalents. The sealed environment eliminated corrosion and human interference while seawater provided consistent cooling. Microsoft didn’t commercialize it, but the data was compelling.

Third, the integration is elegant. Traditional data centers require land acquisition, grid interconnection, water rights for cooling, and community approval. Panthalassa’s model collapses all of that into a single manufactured unit. No NIMBY battles. No permitting fights. No water usage. Scale by deploying more units, not by negotiating more real estate.

The Workload Question

Not everything can run on a floating data center. Real-time inference for autonomous vehicles or high-frequency trading? The satellite latency kills it. But batch processing, model training, and large-scale inference jobs that can tolerate milliseconds of delay? That’s a massive market where energy cost is the primary concern.

And energy cost is exactly where Panthalassa claims to win. The ocean contains an estimated 2.64 trillion watts of wave energy potential. “The ocean is really unlimited in terms of how much energy is available,” Sheldon-Coulson told CBS News. “It will really be the cheapest energy on the planet.”

Bold claim. But when your competition is paying grid rates that are spiking thanks to the very demand you’re trying to serve, “cheapest” doesn’t have to mean “cheap” — it just has to mean “cheaper.”

The Infrastructure Arms Race Gets Weirder

Panthalassa isn’t the only company thinking creatively about AI infrastructure. Google, Microsoft, and Amazon have all signed nuclear power deals. Small modular reactors are seeing a renaissance. Geothermal is gaining traction. Some companies are exploring space-based solar.

The pattern is clear: AI’s energy appetite has outgrown conventional solutions. The companies that figure out unconventional power sources first will have a structural advantage in the compute race.

Panthalassa says construction of the first Ocean-3 units is already underway, with deployment targeted for August 2026. If that timeline holds, we’ll have real-world performance data within months.

What to Watch

Whether Panthalassa specifically succeeds is almost beside the point. The company represents an important signal: the AI industry has accepted that incremental grid improvements won’t cut it. The demand curve is too steep, the timeline too compressed.

The ocean covers 71% of Earth’s surface. If even a fraction of its wave energy could power AI workloads, it would fundamentally reshape the economics and geography of computing infrastructure.

We’re watching software’s appetite collide with physics. The solutions emerging — floating data centers, nuclear reactors, orbital solar farms — will define not just AI’s future but the global energy landscape for decades.

Panthalassa’s bet is that the answer has been rolling beneath our hulls all along. By August, we’ll start to find out if they’re right.