Compute at Sea: The Grid Became the Bottleneck, So YC Points at the Ocean — a physical-infrastructure request read from a software desk

This is the sixth paper in this series reading Y Combinator's Fall 2026 Requests for Startups one at a time. So far the series has covered an AI tutor for education, the future of American defense, a cloud for small software, and team-shared AI sessions — all things that fit in a browser tab. This one does not. "Compute at Sea," written by YC's Francois Chaubard, asks founders to put data centers on the ocean. It is the most physical request in the batch, and worth reading carefully precisely because Gwen has no stake in it.

What the request actually says

The request opens bluntly: "Artificial intelligence is running out of compute. And data centers are running out of electricity and land." Chaubard's diagnosis is that demand for data centers is "insatiable," but new ones "can take years for approval and enough megawatts, and still could get killed by local government intervention." His answer: "While communities increasingly oppose the land, the water is open. It sounds crazy, but we think part of the answer may be to move compute offshore." He points out that the ocean is 70 percent of the Earth's surface, "has abundant sunlight, no permitting process, and is an enormous natural heat sink that is already getting hit by the sun all day anyway." The vision is not a single barge but a fleet: "Think of them as compute flotillas: many standardized, modular vessels operating together as one global cloud." YC wants to fund "founders who are bringing the world's compute to the oceans."

Strip the imagery and this is two claims. First, the binding constraint on AI has moved from chips to electrons, land, and permission — now conventional wisdom, with interconnection queues running years and gigawatt campuses colliding with local moratoria. Second, the ocean offers arbitrage on all three. The request's own language — approvals, megawatts, "local government intervention" — makes clear this is a regulatory and energy thesis wearing a nautical costume, and the second claim is only partly true.

The weakest sentence in the request

"No permitting process" is the line that should give founders pause. Territorial waters — the first twelve nautical miles — are as sovereign as the shoreline, and near-shore water is often more regulated than land, not less. YC has direct evidence in its own portfolio: NetworkOcean, a YC-backed startup, announced plans in 2024 to test a 500 kW GPU capsule in San Francisco Bay and, per reporting at the time, had not sought permits; the reporting itself prompted two regional agencies to warn the company that unpermitted testing could violate the law, while scientists raised concerns about heat and algae blooms. Exclusive economic zones extend the reach of coastal-state regulation two hundred miles out. Even on the high seas there is no lawless zone: a vessel needs a flag state, and with the flag comes that state's law, inspection regime, and liability framework. The honest version of the claim is that permitting at sea is different, and far offshore it is thinner — not absent. But the further out you go to escape regulators, the further you are from the two things a data center cannot live without: power and fiber.

What is actually hard

The heat-sink half of the pitch is real, and it is the best-evidenced part of the whole idea. Microsoft's Project Natick sealed 855 servers in a nitrogen-filled cylinder off Scotland from 2018 to 2020 and reported a failure rate around one-eighth of the matched land control group — roughly 0.7 percent of servers failed underwater versus about 6 percent on land. Seawater cooling works, sealed atmospheres exclude the oxygen and humidity that kill electronics, and nobody bumps into the racks. Yet Microsoft shelved the program anyway; its head of cloud operations said plainly in 2024 that she is not developing subsea data centers anywhere in the world. A hyperscaler with infinite need for exactly this capability ran the experiment, liked the results, and walked away. Any founder answering this RFS should have a crisp theory of why.

The likely answer is everything around the cooling. Power first: the request gestures at "abundant sunlight," but solar's areal density is a poor match for AI loads measured in hundreds of megawatts; floating photovoltaics at that scale would be an engineering project as large as the data center itself. The practical near-term source is offshore wind — which is exactly what China chose. Then serviceability: a sealed vessel trades repairability for reliability, which suits batch and training workloads but complicates the hot-swap operational culture of commercial clouds. Then connectivity: subsea fiber is mature technology, but landing rights, cable ships, and route redundancy are their own slow, permitted, capital-intensive world, and a flotilla that moves complicates all of it. Then the ocean itself: salt spray, corrosion, biofouling, storm loading, mooring in deep water, crew or robotic maintenance logistics. Then money: marine insurance and classification. A novel vessel that no classification society has rules for is a vessel underwriters cannot price, and unpriceable risk is unfinanceable risk. Getting a class society and an insurer comfortable is as much the product as the cooling loop is.

Who has tried

The lineage is longer than it looks. Google was famously reported to be building mystery barges in 2013; they never became compute. Nautilus Data Technologies launched a 7 MW data center on a barge in the Port of Stockton in 2021, ran it for years with real tenants — and by 2026 had put the barge up for sale for $45 million while pivoting to selling its cooling technology on land. That is the market's verdict on floating colocation version one: the technology worked; the business preferred shore. The most serious current operator is Chinese: Highlander and its HiCloud arm have run commercial underwater modules off Hainan since the early 2020s, added a 400-server module in early 2025, and have brought online an underwater data center off Shanghai's Lin-gang area powered by a surrounding offshore wind farm, reported at roughly 2,000 servers with a planned 24 MW capacity. Notice what made that project possible: not the absence of permitting but the presence of a state that wanted it to exist, co-sited with generation. And NetworkOcean shows YC's own conviction predates this RFS — this is a doubling-down, not a first look.

What building it takes

The winning team looks more like an offshore energy company than a cloud company: naval architects, power engineers, marine operations people, with data center operators attached. The credible wedge is probably not "a global cloud" but one vessel, co-sited with stranded offshore generation, running training or batch workloads that tolerate remoteness, classed by a society an insurer will listen to, sold to one customer who needs megawatts faster than the interconnection queue can deliver them. Timelines are set by shipyards and cable ships, not sprints. The prize is real — if land-side power and permitting stay jammed, whoever industrializes marine compute owns a genuinely scarce input. But the request's romance is the ocean, and the business is the megawatt.

Where Gwen stands

Plainly: this is not Gwen's lane, and this paper will not pretend otherwise. Gwen is software. It builds and hosts websites and small web apps from a plain-language description, and runs marketing, content, CRM, research, and operations work as long-lived missions in a customer's workspace, with human approvals before anything goes out. It owns no hardware, no infrastructure, and certainly no ships, and nothing in this request is something Gwen could build or partner on.

The one honest connection is downstream. Underneath Gwen's work, a routing rail sends each task across many AI models by measured quality and cost, with caching and continuous evaluation — Gwen's economics ride the falling cost of inference. Every serious attempt to add compute supply, on land or at sea, pushes that curve down, and a customer's Work Budget quietly buys more finished work. If compute flotillas succeed, Gwen's customers will never see the ocean. They will just notice the work got cheaper. That is the entire relationship, and saying so is the point: when other papers in this series say a request sits squarely in Gwen's lane, this paper is why you can believe them.