MIT founders launch Atomarine to build gas-first nuclear data centers at sea

MIT-trained engineers Dimitris Koutentakis and Emile Germonpre have launched Atomarine, a Boston startup proposing floating data-center campuses powered initially by natural gas and later by marine reactors, with a first pilot targeted for 2028. The company's modular design separates compute barges from power vessels to bypass grid constraints and cooling demands, aiming to support up to 1.5 GW of annual deployment.

Dimitris Koutentakis https://www.linkedin.com/in/dkoutentakis/?ref=runtimewire and Emile Germonpre https://www.linkedin.com/in/emile-germonpr%C3%A9-766045245/?ref=runtimewire , two MIT-trained engineers, have launched Atomarine https://atomarine.co/?ref=runtimewire , a Boston startup proposing data-center campuses made from floating compute barges and separate power vessels. The design is pitched as a way around the grid queues, land constraints and cooling demands holding up new AI infrastructure. The nuclear language is ahead of the hardware. Atomarine plans to power its first pilot with natural gas by 2028, according to its Y Combinator launch https://www.ycombinator.com/launches/Rwg-atomarine-floating-nuclear-data-centers?ref=runtimewire . Compact marine reactors would come later, replacing gas-fired power vessels at the same berth without requiring Atomarine to rebuild the compute platforms. Atomarine frames the sequence as gas turbines today and marine reactors tomorrow, according to its homepage https://atomarine.co/?ref=runtimewire . That rollout gives Atomarine a practical first product while preserving the larger nuclear thesis. It also places the harder part of the proposition beyond the initial pilot: Atomarine must first demonstrate that a serviceable data center can operate economically offshore, then secure the reactors, approvals and financing needed to replace the gas generation. Koutentakis brings an unusual mix of software, finance and maritime exposure. Y Combinator's Atomarine profile https://www.ycombinator.com/companies/atomarine?ref=runtimewire says he earned undergraduate, electrical engineering, mechanical engineering and business degrees from MIT. YC says he later researched trade execution at Arrowstreet Capital and designed electric-fleet charging optimization at Amazon, where his work identified $193 million in potential capital expenditure avoidance. The figure comes from Koutentakis' YC biography rather than an Amazon disclosure. YC also cites his connections to the Greek shipping industry. Germonpre supplies the nuclear specialization. His 2024 MIT thesis https://dspace.mit.edu/server/api/core/bitstreams/ab0a3ff8-4442-4e80-b32d-d9868443828c/content?ref=runtimewire studied nuclear batteries for decentralized hydrogen production. He also co-authored MIT research on nuclear batteries for offshore power https://canes.mit.edu/media/a-preliminary-study-on-the-feasibility-of-nuclear-batteries-for-offshore-power-generation?ref=runtimewire , examining factory-fabricated microreactors as power sources for offshore industrial platforms. That work gave Germonpre a direct research path into Atomarine's central question: whether standardized nuclear generation can be manufactured on land, transported to sea and sold as an infrastructure service. A modular campus built around the power bottleneck Atomarine describes each compute barge as a standardized, shipyard-built platform carrying 75 to 100 megawatts of capacity. A separate vessel provides generation. Atomarine's example campus grows from one 75 MW platform to six, reaching 450 MW by adding barges alongside the existing installation. Separating power from compute is the core architectural decision. Data halls can remain in service when Atomarine changes the generation vessel, and Atomarine can theoretically upgrade the power source without refitting every server platform. Shipyard production also shifts construction from a succession of bespoke land sites toward repeated manufacturing in a controlled facility. Atomarine claims a closed seawater cooling loop can deliver a modeled power usage effectiveness of about 1.1 without cooling towers. Atomarine also says the platform could support 1.5 GW of annual deployment and operate for 20 to 40 years. Those figures are projections for a system that has yet to reach the pilot stage. Atomarine's public materials also give different deployment comparisons. The homepage https://atomarine.co/?ref=runtimewire claims offshore campuses can deploy 4 to 5 times faster than land-based projects, while the YC launch post https://www.ycombinator.com/launches/Rwg-atomarine-floating-nuclear-data-centers?ref=runtimewire says 2 times faster. Neither page provides the construction assumptions behind those multipliers. The argument beneath both figures is consistent: Atomarine expects shipyard manufacturing and on-site generation to remove delays associated with grid interconnection, site preparation and sequential construction. The Atomarine name previously appeared in an MIT Orbit concept https://orbit.mit.edu/launchpad/ideas/atomarine?ref=runtimewire centered on mobile nuclear power plants operating in international waters and supplying energy to cargo ships. The data-center plan redirects the same broad industrial tools - shipbuilding, mobile power and offshore operation - toward the AI industry's immediate demand for electricity. Ocean compute has precedent at a smaller scale Atomarine is entering a field with technical demonstrations, though few operating models at the scale Atomarine proposes. Microsoft's Project Natick https://news.microsoft.com/source/features/sustainability/project-natick-underwater-datacenter/?ref=runtimewire placed a sealed data center on the seabed off Scotland in 2018 and retrieved it two years later. Microsoft reported that the underwater servers had one-eighth the failure rate of comparable equipment on land and required no freshwater for cooling. Atomarine's design makes a different operating trade. Project Natick sealed servers underwater and connected the installation to an existing power supply. Atomarine proposes accessible surface platforms with their own adjacent generation, moving both the data halls and the power plant offshore. That raises Atomarine's potential capacity while adding marine operations, fuel delivery and power-plant integration to the execution burden. The broader nuclear data-center market is moving on land as well. Oklo and Meta announced https://oklo.com/newsroom/news-details/2026/Oklo-Meta-Announce-Agreement-in-Support-of-1-2-GW-Nuclear-Energy-Development-in-Southern-Ohio/default.aspx?ref=runtimewire plans in January 2026 to support a 1.2 GW advanced-nuclear development in Ohio. Atomarine is betting that mobility and shipyard manufacturing can eventually offer a different route to firm power, particularly where coastal access is easier to secure than a grid connection and a large land parcel. The 2028 pilot is the first test Atomarine joined Y Combinator's Summer 2026 batch with a two-person team. Under YC's published standard deal https://www.ycombinator.com/blog/ycs-500-000-standard-deal?ref=runtimewire , accepted startups receive $500,000: $125,000 for 7% and another $375,000 through an uncapped most-favored-nation SAFE. Atomarine's homepage also lists MIT https://www.mit.edu/?ref=runtimewire , The Engine https://engine.xyz/?ref=runtimewire and Cleantech Open https://www.cleantechopen.org/?ref=runtimewire under a Supported by label. Atomarine says it has early relationships with offshore barge builders and multiple small modular reactor vendors, according to its YC launch https://www.ycombinator.com/launches/Rwg-atomarine-floating-nuclear-data-centers?ref=runtimewire , without naming the partners. The 2028 gas-powered pilot is therefore the first concrete credibility checkpoint. A working barge would test whether Atomarine can deliver shipyard construction, offshore cooling, reliable fiber connectivity and self-generated power on a timetable that beats a land development. Nuclear generation comes after that demonstration. Atomarine's gas-first sequence gives Koutentakis and Germonpre a way to begin building the maritime data-center layer with equipment available today. Their larger bet is that the same berth, cooling system and compute barges can become distribution infrastructure for marine reactors once those reactors are commercially ready.