In 1858, a group of investors spent today’s equivalent of $500 million to lay a telegraph cable across the Atlantic Ocean. The cable worked for three weeks before it died. The press called it a hoax. Engineers called it impossible. Queen Victoria had already sent a congratulatory message through it.
Eight years and several spectacular failures later, they got it right. The 1866 transatlantic cable changed global communication forever. And when people asked why anyone would attempt something so absurd, the answer was always the same: because the demand was so enormous that normal infrastructure couldn’t keep up.
That’s exactly where AI is right now. And one startup just decided the answer is, once again, the ocean.
The Power Wall Nobody Talks About
Here’s a number that should reframe how you think about AI infrastructure. The International Energy Agency projects that global data centers will consume 1,000 terawatt-hours of electricity in 2026. That’s equal to Japan’s entire national electricity consumption. The whole country. Every factory, every bullet train, every vending machine on every corner in Tokyo.
And it’s not enough.
Nearly 50% of all data center projects scheduled for completion this year are delayed because there isn’t enough power to turn them on. Not because the buildings aren’t built. Not because the chips aren’t available. Because the electrical grid literally cannot deliver the watts.
Morgan Stanley forecasts a 49-gigawatt power shortfall in the US alone by 2028. To put that in perspective, 49 gigawatts is roughly the output of 49 nuclear power plants. Lead times for high-voltage transformers, the boring metal boxes that connect data centers to the grid, have stretched from 18 months to nearly four years. A $2 billion data center campus can sit completed and dark, waiting on a $40 million transformer order.
And then there’s the most emblematic example. OpenAI’s Stargate Project, the single largest AI infrastructure announcement in history, is as of May 2026 an empty field. Bloomberg confirmed no physical progress on the campus. The reason? Power infrastructure shortages, compounded by Chinese-supplied equipment delays and tariffs. The biggest AI company on Earth announced a megaproject and the grid said “no.”
Big Tech is already spending five times its payroll on AI infrastructure, and the bottleneck isn’t money or chips anymore. It’s kilowatts.
Enter the Ocean
While the rest of the industry fights over land permits, grid connections, and transformer delivery dates, a 10-year-old startup from Portland, Oregon did something that sounds like a rejected sci-fi pitch.
Panthalassa just raised $140 million in Series B financing, led by Peter Thiel, to build autonomous floating data centers powered by ocean waves. The company is now valued at $1 billion.
Let me describe what they actually built, because it matters.
The Ocean-3 is an 85-meter steel structure, roughly the height of Big Ben, shaped like a lollipop. It sits mostly beneath the ocean surface. As waves pass, the structure bobs up and down, forcing seawater through an internal turbine. That mechanical energy generates electricity. The electricity powers AI chips housed in a hermetically sealed, seawater-cooled container. The chips receive inference queries via SpaceX’s Starlink satellite connection and beam the answers back the same way.
No anchor. No cables to shore. No connection to any power grid. No hinges, flaps, or gearboxes that could break down in ocean conditions. Just steel, water, and physics.
The company claims power generation costs as low as $0.02 per kilowatt-hour. For comparison, the average US industrial electricity rate is around $0.08. And the cooling? Free. The ocean handles it.
Garth Sheldon-Coulson, the CEO, spent years at Bridgewater Associates before co-founding Panthalassa in 2016. His co-founder Brian Moffat had already built a wave energy system for a previous company. They’ve been quietly developing this technology for a decade while everyone else was obsessing over GPU architectures.
Why This Isn’t Crazy (Even Though It Sounds Crazy)
I know what you’re thinking. Floating data centers in the ocean sounds like something a startup pitches when they’ve run out of real ideas. Salt water corrodes everything. Waves destroy equipment. Satellite latency kills real-time applications.
And you’d be right to be skeptical. But here’s why the smart money disagrees.
First, the power crisis is real and getting worse. When Anthropic needed 220,000 GPUs, it had to rent them from Elon Musk’s SpaceX because even the biggest cloud providers couldn’t guarantee the power capacity. Meta is building a $3.2 billion gas power plant in Louisiana just to run one data center campus. Microsoft is restarting Three Mile Island’s nuclear reactor. When your options are “build a nuclear plant” or “put a computer in the ocean,” the ocean starts looking reasonable.
Second, Panthalassa isn’t trying to replace hyperscaler campuses. They’re targeting AI inference workloads, the kind of processing that powers chatbots, recommendation engines, and language models in production. Inference is less latency-sensitive than training. A 50-millisecond satellite round-trip is invisible to someone asking Claude a question but would be devastating for training a foundation model. Panthalassa picked its use case carefully.
Third, the engineering is simpler than it sounds. No moving parts that contact each other. No gearboxes. Earth-abundant materials (steel, not rare earth metals). The design is built to be manufactured at scale in shipyards, not custom-fabricated in cleanrooms. Multiple Ocean-3 nodes network together to function as a single distributed data center.
And fourth, the economics are genuinely different. Land-based data centers pay for the building, the land, the grid connection, the cooling infrastructure, the backup generators, and the electricity itself. Panthalassa pays for the steel structure and the satellite link. The ocean provides the rest.
What This Means If You’re Building a Startup
If you’re a founder reading this, the actionable insight isn’t “go build a floating data center.” It’s that the AI infrastructure stack is being restructured around energy access, not chip access, and the biggest AI opportunities might not be software at all.
Three things to watch.
The compute cost curve is about to get weird. For three years, the assumption has been that inference gets cheaper as models get more efficient and chips get faster. But if power costs dominate the total cost of ownership, and power is constrained, inference costs could actually go up. Founders who’ve modeled their unit economics on “AI gets cheaper every year” might be holding a wrong assumption.
Location becomes a competitive advantage. If power scarcity drives compute offshore, into the ocean, near geothermal vents, next to nuclear plants, or into countries with excess energy capacity, then where your AI runs becomes part of your competitive strategy. Today, most startups don’t think about this at all. They pick AWS us-east-1 and forget about it.
The “wild infrastructure” category is investable. Panthalassa isn’t alone. Nuclear microreactors, geothermal data centers, natural gas microgrid companies like VoltaGrid (which just raised $1 billion from Blackstone and Halliburton), modular nuclear from companies like Oklo. These were fringe energy plays two years ago. Now they’re where the smart infrastructure money is going, because the grid can’t scale fast enough.
The Pattern
There is a pattern in technology infrastructure that repeats so reliably it should be a law. When demand for a new technology exceeds what existing infrastructure can deliver, someone builds infrastructure in a place that seemed absurd five years earlier.
Telegraph demand put cables on the ocean floor. Oil demand put rigs in the North Sea. Internet demand put fiber optic cables under every ocean on Earth. Space demand is putting data centers in orbit.
AI demand in 2026 is consuming more electricity than most countries. The grid can’t keep up. Half the planned data centers are delayed. The biggest project in the industry is an empty field.
And a startup from Oregon just put a computer in the ocean, powered by the waves, connected by satellite, cooled by seawater.
It sounds insane. But so did laying a cable across the Atlantic in 1858.
The question isn’t whether AI will find its power. It always does. The question is whether you’ll recognize the infrastructure shift before your competitors do.
