The plan sounds like something from a speculative fiction novel: floating data centers moored off the coast, fed electricity by separate ships carrying nuclear reactors. Yet Atomarine, a startup tucked somewhere in Boston with backing from Y Combinator, insists it will deploy the first version by 2028. The catch, for now anyway, is that the pilot runs on natural-gas turbines. Nuclear comes later, assuming regulators, investors, and the laws of maritime engineering cooperate.
The company's Summer 2026 launch page lays out the vision in straightforward terms. Start with gas, prove the offshore model works, then swap in marine reactors once the harder approvals clear. It's an acknowledgment of just how difficult the nuclear piece will be—and perhaps a hedge against the possibility that it never happens at all.
Behind the audacity sits a genuine crisis. BloombergNEF publicly reported in July that U.S. data centers face a 63-gigawatt capacity shortfall by 2033, the gap between what surging AI chip deployments will demand and what the power grid can realistically deliver under current constraints. Interconnection queues stretch for years. Communities balk at shouldering infrastructure costs for facilities that consume electricity at industrial scale but employ relatively few people. "States require data centers to act as flexible assets, pay fairly for their own infrastructure, and bring their own generation without socializing those costs," Energy Secretary Jennifer Granholm said at the BNEF Summit in San Francisco in February.
Atomarine is betting that going offshore solves multiple problems at once: cooling, permitting, grid delays, and local opposition. Whether it actually does remains an open question.
The Math Behind the Mania
Data centers are rewriting the electricity demand map faster than utilities can keep up. The U.S. Energy Information Administration reported in January that the sector is driving the strongest four-year growth in electricity consumption since 2000. The International Energy Agency's analysis found that AI-focused facilities now resemble power-intensive factories in their appetite for megawatts, with projections showing data centers could account for roughly half of U.S. electricity demand growth through 2030 in some regional markets. An Uptime Institute scenario analysis published in January ran the numbers on fully built-out planned capacity: it would eat one-third of North America's electricity and a tenth of global demand.
Capital is chasing the bottleneck hard. Dell'Oro now projects that data center capital expenditures could top $3 trillion by 2030. But that spending runs straight into a wall without power. FERC launched targeted action in June, ordering regional transmission organizations and independent system operators to either justify or reform their rules for integrating large loads. Berkeley Lab's "Queued Up: 2026 Edition" documented continued queue growth across more than 50 balancing authorities covering 98 percent of U.S. capacity—a polite way of saying the wait times aren't getting better.
The industry has tried various workarounds. Some hyperscalers are co-locating at existing power plants. Others are financing dedicated generation. A few are reopening shuttered nuclear sites. Atomarine's answer is to skip the grid entirely and float.
The Technical Pitch
Atomarine's design calls for modular compute barges moored at sea, each delivering 75 to 100 megawatts of IT capacity, according to the company's website updated in 2026. Scale that to a six-barge campus and you hit 450 megawatts. Seawater handles cooling through a closed loop, which the startup claims achieves a power usage effectiveness around 1.1, well below the industry average that typically hovers closer to 1.5 or worse. Deployment timelines supposedly run four to five times faster than land-based builds, with a throughput target of 1.5 gigawatts per year once the model is proven.
Power comes from separate vessels anchored nearby. The 2028 pilot relies on natural-gas turbines—proven technology, easier permitting, faster to deploy. Marine nuclear reactors slot in later, under what the company describes as a "gas today, nuclear tomorrow" strategy. That sequencing buys time for reactor licensing, fuel supply chains, and maritime nuclear regulations to mature.
The founders bring credentials that at least look the part. Dimitris Koutentakis, the CEO, holds four MIT degrees and previously worked at Arrowstreet Capital and Amazon. Emile Germonpré, the CTO, has a PhD in nuclear engineering, a Fulbright fellowship, and graduated valedictorian from Ghent University. Crunchbase listed the team size at one to ten employees as of late August and disclosed no funding rounds beyond the Y Combinator batch.
Data Center Dynamics covered the concept in July and August, noting the gas-powered near-term plan and the 2028 target. The startup has not disclosed customers, detailed vessel designs, or specific site selections. That silence could mean deals are in the works or that the pitch is still theoretical.
The Offshore Field Is Getting Crowded
Atomarine is hardly alone in looking seaward. Singapore, facing acute land constraints, advanced two pilots in 2026. Keppel, which received regulatory approval back in 2023, started construction in the first quarter on a 19- to 21-megawatt floating module. The company won a 50-megawatt capacity award in Singapore's DC-CFA2 tender on August 21, according to Data Center Dynamics. Seatrium announced a 30-megawatt offshore pilot on July 29, using six 5-megawatt "Data-In-A-Box" modules with embedded power generation and Bureau Veritas approval.
China deployed a 24-megawatt offshore data center powered by offshore wind off Shanghai that entered full operation in May. The facility houses roughly 2,000 servers and uses ocean water for passive cooling, according to Tom's Hardware and South China Morning Post coverage. Earlier Hainan subsea modules date to 2022. Microsoft's Project Natick, which ran from 2015 to 2020 with a two-year demonstration on the seafloor near Orkney, remains the most cited proof of concept: the company reported a power usage effectiveness of 1.07, zero on-site water consumption, and up to eight times the server reliability compared to land facilities.
Bluecore Energy, a U.S. startup, raised $10 million in a pre-seed round announced July 21 to build portable small modular reactors on barges. CEO Kofi Asante told TechCrunch that AI data center executives "would not need to pull water or energy from communities" if they could receive their own electricity source and tap water available at sea. The pitch mirrors Atomarine's, though Bluecore seems focused on the nuclear piece from the start.
Alternative ocean-powered approaches are also emerging. Panthalassa, a wave-energy startup, reportedly raised around $140 million and is targeting a late-2026 Ocean-3 pilot with commercial deployment in 2027, according to Ars Technica coverage in May. Mocean Energy in the U.K. unveiled Blue Core on July 28 and 29, a wave, offshore solar, and storage system designed to power AI racks at sea, with a 2027 demonstration planned and a £5 million fundraise underway, Energy Voice reported.
The common thread across all these projects: escaping the grid, the permitting morass on land, and the political friction that comes with asking local communities to absorb massive new power loads.
Nuclear Ships Are Coming Back, Maybe
Atomarine's nuclear vessel concept sits at the intersection of advanced reactor development and a reviving maritime nuclear sector. CORE POWER, a U.K. maritime nuclear developer, signed a memorandum of cooperation with the Port of Corpus Christi on August 19 to study maritime nuclear readiness. Five days later, the company announced plans for a dedicated U.S. nuclear-grade integration shipyard, according to press releases. Lloyd's List reported MARAD collaboration on August 25. CORE POWER also published a June 13 study with Maersk, Lloyd's Register, and the Port of Rotterdam examining port calls for nuclear ships—essentially, what infrastructure and protocols would be needed if commercial nuclear-powered vessels start showing up at major ports again.
Prodigy Clean Energy in Canada announced a collaboration with Lloyd's Register on March 11, 2025, to drive transportable nuclear power plant deployment by 2030, with backing from the Government of Canada. Karpowership markets "Intelliships" to power data centers using LNG, with ongoing conversions and a Seatrium contract for floating storage and regasification units.
Fuel supply is a potential bottleneck. Many advanced reactor designs require high-assay low-enriched uranium, which as of August only Russia and China produce at commercial scale, according to the World Nuclear Association. The U.S. banned Russian HALEU imports in 2024 and awarded $2.7 billion on January 5 to expand domestic enrichment capacity. Centrus Energy secured a $900 million award the same day to expand HALEU and LEU capacity at Piketon, Ohio. The company delivered 900 kilograms of demonstration material by June 2025, with contract extensions running through June 30 this year.
An American Bureau of Shipping study published in 2025 modeled a floating nuclear-powered data center with four BWXT BANR reactors delivering roughly 70 megawatts of electricity to a Nautilus EcoCore-cooled facility. The report warned that heat dissipation from a 200-megawatt-thermal nuclear-plus-IT barge must avoid more than 5°C local temperature rise through adequate water flow. It recommended pier-moored installations over fully offshore moorings to simplify fiber backhaul, crew access, and backup power—practical constraints that could complicate Atomarine's open-ocean ambitions.
Regulators Are Clearing a Path, Sort Of
U.S. regulators moved in mid-2026 to streamline pathways for both large loads and maritime nuclear, though calling it "streamlined" might be generous. FERC's June 18 order on large-load integration and the Department of Energy's July Draft National Transmission Needs Study highlighted data center constraints. The Nuclear Regulatory Commission updated its Maritime Nuclear Applications page on July 23, noting new memoranda of understanding with the Coast Guard and the Marine Minerals Administration to clarify jurisdiction and licensing for floating nuclear plants and civil nuclear propulsion.
The Coast Guard held a maritime nuclear industry workshop on June 23 and announced a new Maritime Nuclear Policy Division created the previous December. Internationally, the IMO agreed in January on a workplan toward a revised Code of Safety for Nuclear Merchant Ships, targeting adoption in 2030, according to the organization's January 29 summary.
Domestic regulations finalized in 2023 under 10 CFR 50.160 allow alternative emergency preparedness frameworks for small modular reactors and advanced reactors, moving away from the 10-mile plume and 50-mile ingestion zones required for large light-water reactors. That change accommodates floating installations with smaller exclusion zones, though it remains to be seen how regulators will handle a floating nuclear ship that could, in theory, move.
Connectivity and environmental permits remain thorny. Any offshore data center campus needs fiber to shore, requiring FCC submarine cable landing licenses under rules updated in late June. NOAA is evaluating cable permitting in National Marine Sanctuaries as of June 26, according to Federal Register and NOAA notices. TeleGeography's 2026 map tracks more than 600 active and planned cables and 1,893 landing stations, so metro backhaul options will depend heavily on proximity to existing landing points.
Thermal discharge and intake structures fall under EPA's NPDES permits, with Section 316(a) regulating thermal plumes and Section 316(b) governing intake impacts on aquatic life. Site-specific modeling will be required to set water quality limits that protect marine ecosystems. The Jones Act and Outer Continental Shelf Lands Act extend coastwise laws to non-mineral energy devices attached to the seabed, per Customs and Border Protection guidance updated in December 2024. That means U.S.-built, U.S.-flagged, and U.S.-owned vessels must handle cargo moves between U.S. points, including offshore installations treated as coastwise points—a requirement that could add significant cost and complexity.
The Case for Seawater
The technical logic for seawater cooling is hard to dispute. NVIDIA's GB200 NVL72 and DGX SuperPOD reference architectures, published in 2025 and 2026, assume high-flow direct liquid cooling and rack-scale thermal budgets topping one megawatt. Uptime Institute's 2026 Cooling Systems Survey documented an accelerating shift toward direct liquid cooling as chip densities climb. Seawater offers effectively unlimited heat rejection capacity if thermal delta-T and intake/discharge permits can be managed—a big if.
Water consumption is becoming a land-based liability. A March 3 arXiv paper estimated U.S. data centers could require 697 to 1,451 million gallons per day of new water capacity through 2030 if current intensity trends persist. The Desert Research Institute modeled Nevada data center water demand for cooling at 9,647 acre-feet per year and for power generation at 12,448 acre-feet per year by 2033 in scenario studies published in January. Nautilus Data Technologies, which operates a 230-foot barge in Stockton, California, with commercial operations since 2021, achieved "zero water consumption" cooling by recirculating a water heat sink without withdrawals, according to Black & Veatch commissioning documentation.
Grid independence sidesteps interconnection delays and community opposition entirely. Synergy Research said in July that U.S. data center capacity will double within three years despite headwinds. That doubling depends on solving the power equation. A Y Combinator post in July summarized Atomarine's thesis as bypassing "interconnection queues and community opposition" with "essentially free cooling" offshore.
That's the theory. Execution is another matter.
What Happens Next
Atomarine's 2028 pilot will test whether the logistics, permitting, and economics actually work. The startup has not disclosed financing, customers, or reactor partners—any of which could derail the timeline. Singapore's Keppel and Seatrium are executing smaller-scale proofs with conventional power and seawater cooling in a jurisdiction with acute land constraints and strong regulatory oversight. China's state-backed subsea deployments at 24-megawatt scale demonstrate technical feasibility under different governance models, though drawing direct comparisons to what a private U.S. startup can pull off feels like a stretch.
The nuclear timeline is considerably longer. The IMO's Code revision targets 2030. HALEU supply ramps domestically through that window, in theory. NRC-USCG-MMA coordination matured in 2026, but licensing a first-of-a-kind floating nuclear installation will take years, not months. Bluecore Energy's $10 million pre-seed and CORE POWER's port studies signal investor and industry interest in portable maritime reactors, but no commercial nuclear power vessel has operated in U.S. waters since the NS Savannah decommissioned in 1971. That's more than half a century ago.
The 2028 gas-powered pilot is the real test. If Atomarine can deploy a multi-barge campus with fiber backhaul, environmental permits, and paying customers inside two years, it validates the offshore data center model and sets the stage for the nuclear swap. If permitting, connectivity, or vessel construction timelines slip—and they usually do—the concept joins a long list of ambitious infrastructure ideas that underestimated execution risk.
The startup's website promises a first unit "ready by 2028." The data center industry's power crisis ensures someone will be watching.
