Sam Sipe used to land E-2D Hawkeye aircraft on the decks of US Navy carriers. His next vessel will carry something less glamorous but, investors hope, far more lucrative: racks of servers running artificial intelligence.
Sipe’s startup, Drogue, is barely six months old. It has just raised $20 million to build data centres that float at sea and draw their power from the ocean itself, according to TechCrunch, which broke the story on 13 November. Backers include Toyota Ventures, Slow Ventures and XYZ Venture Capital. The plan is to put a multi-megawatt prototype on a conventional oceangoing ship by early 2028, then move to floating platforms that look a lot like offshore oil rigs. Each vessel would stay at sea for around a decade, assuming they don’t get hit by a tanker in New Orleans.
The technology is called ocean thermal energy conversion, or OTEC, and it is older than the microchip. Warm surface water cools the servers and then boils ammonia, which drives a turbine. Cold water pumped up from about 1,000 metres down, at roughly 4°C, turns the ammonia back into liquid so the cycle can start again. The idea was developed in the 1950s, tested in the 1980s and then shelved.
So why is fresh money flowing into a 70-year-old idea? And would it survive the people who finance, insure and regulate the energy transition?
Why are AI data centres running short of power?
Sipe’s pitch is all about timing. He told TechCrunch that AI demand has grown tenfold in just three years, while data centre developers work on timelines of eight to ten years. The grid simply cannot keep pace.
The evidence is piling up. Global data centre electricity demand is growing by 20 to 40% a year, according to a UK Energy Research Centre review. In Ireland, data centres already use about a fifth of the country’s electricity, and the grid operator has effectively stopped taking new applications in Dublin.
Gbemi Oluleye, Associate Professor in Sustainable Market Mechanisms at the Grantham Institute and Faculty of Natural Sciences at Imperial College London modelled 33 European countries to 2050. With Mohammad Hemmati and Vassilis Charitopoulos from UCL they estimate AI could add anywhere from 73 to 723 TWh of demand, a range that says a lot about how uncertain the future is. More telling, they find that after 2030, where data centres get built will depend more on access to reliable, always-on power than on how much clean energy is nearby.
The ocean, by contrast, is not short of energy. It absorbs an extra 5.8 million TWh a year, TechCrunch notes, against roughly 180,000 TWh of total human energy use. Little wonder founders are looking out to sea.
Underwater data centres and ocean thermal energy: what’s been tried before
Drogue is not the first to look to the sea. In 2018 Microsoft sank a container holding 864 servers off the Orkney Islands as part of Project Natick. When it was hauled up two years later, the servers had failed at one eighth the rate of a comparable facility on land. Microsoft’s team judged the concept logistically, environmentally and economically viable. Yet the company has since been reported to have wound the project down. A technology that works is not the same as a business that scales.
OTEC has its modern champions too. Global OTEC, a UK firm founded to replace diesel generators on small tropical islands, tested a scale model of its hull and cold-water pipe at Charlton Lido in south London in 2023. The trial was part of an EU-funded project in the Canary Islands.
Drogue’s twist is to bolt the power plant to its own customer. The servers that need the electricity also warm the water that helps produce it. Neat on paper, but is it actually green?
Is ocean thermal energy conversion (OTEC) actually green?
This is probably greener than diesel, not obviously greener than wind, and nobody yet knows what it does to the sea around it.
Start with physics. A heat engine’s efficiency is capped by the gap between hot and cold. With surface water at about 29°C and deep water at about 4°C, the theoretical ceiling is roughly 8%. The real figure is lower, because pumping vast volumes of water up from the deep eats into the output. A US National Academies assessment worked an example assuming 30% of the power is lost to pumping, with a 100 MW plant moving 200 to 300 cubic metres of water a second. The fuel is free, but it takes a lot of kit to catch it.
That kit carries a carbon cost. A TU Delft life-cycle study of a 10 MW OTEC plant on Curaçao found emissions per kWh 16 times lower than diesel, but 3.9 times higher than a wind turbine. Most of the footprint came from raw materials, manufacturing and transport. In other words, steel and shipping.
Then there is the ocean. Bringing up cold, nutrient-rich deep water and releasing it near the surface could affect plankton and fish. A 2024 systematic review in Frontiers in Marine Science found just two field studies worldwide on OTEC’s cold-water discharge, one in Japan and one in France. Ammonia, the working fluid, is toxic, so leaks matter too. There is better news on carbon dioxide escaping from the deep water itself: one study projected that releases from open-cycle OTEC plants would be 15 to 25 times smaller than from fossil power stations.
Measuring a data centre’s environmental impact
Data centres already have a scorecard. European and international standards define ratios such as power usage effectiveness (total energy divided by the energy reaching the servers, where 1.0 is perfect), the share of renewable energy, water use, carbon per unit of computing and how much heat is reused.
On paper, Drogue could shine on several of these: almost all renewable power and little need for fresh water. That is the problem. The standard scorecard measures how a site runs, not what it took to build or what it does to its surroundings. A fairer test would add three things:
- Life-cycle carbon per kWh, including the ship, the steel and the heat exchangers, compared with a realistic alternative such as wind or solar with storage on land.
- Published marine monitoring, before and after deployment, of temperature, nutrients and sea life around the discharge.
- The counterfactual. Is this new clean power, or could the same investment do more good feeding an island grid that still burns diesel?
If Drogue published all three, it would set a standard much of the industry would struggle to match. If it publishes only its efficiency ratio, readers should ask why.
Can floating data centres be financed and insured?
Imagine Sipe pitching to students and faculty on Imperial Business School’s MSc Climate Change, Management & Finance. The one-year programme, run with the Grantham Institute, includes a module where student teams build a cleantech venture and pitch it to an investor. The faculty’s research points to three tough questions.
Who will lend to it? Michael Wilkins, Executive Director of Imperial’s Centre for Climate Finance and Investment, has spent 30 years in infrastructure, project and structured finance, much of it at S&P Global Ratings. Venture capital can pay for a prototype, but it cannot pay for a fleet. To scale, Drogue will need cheaper project debt, and lenders like proven technology, predictable cash flows and plenty of similar deals to compare against. A single ship using an idea last tried at scale in the 1980s has none of those yet.
Who will insure it? Enrico Biffis, Professor of Actuarial Finance and Director of the Centre for Climate Finance and Investment, studies climate risk and insurance. With Princeton’s Shuai Wang he used machine learning to sharpen global climate models for Southeast Asia down to about one square kilometre. They found global projections of urban rainfall could be out by as much as 100%, and temperature by 80%. That matters here. OTEC works best in warm tropical seas, which are also where typhoons form. Pricing the risk of a floating data centre with no track record, in waters where the models are blunt, is no small task. Biffis’s work on parametric insurance, which pays out automatically when a set trigger is hit, raises an interesting question: could that kind of cover make the risk bankable?
Should policy back it at all? Mirabelle Muûls and Ralf Martin, both Professors of Economics at Imperial, worked with Jonathan Colmer and Mannheim’s Ulrich Wagner on the EU’s Emissions Trading System. They found it cut emissions from covered manufacturers by 14 to 16% over eight years, with no measurable hit to profits or jobs. In a 2024 article, Muûls argued that market-based policies beat direct regulation for cutting industrial emissions. Applied to AI, that suggests a different starting point. Rather than hoping founders find heroic new sources of supply, price the carbon in data centre power and let the market sort clever ideas from costly ones.
None of this makes Drogue a bad bet. It means the engineering is only half the job.
How Finland reuses data centre waste heat
There is another way to think about all that heat. Drogue treats it as something to get rid of at sea. In Finland, it is something to sell.
Sanna Syri, Professor of energy technology and energy economics at Aalto University, has studied how data centres can feed their waste heat into district heating, the networks of hot water pipes that warm whole towns. In a 2025 study with colleagues S. Tervo and P. Hiltunen, she looked at two Finnish cities. In Espoo, home to 310,000 people, a large data centre would help district heating reach 95% fossil-free. In Seinäjoki, a town of 66,000, it would replace peat. Finland’s tax incentives seem to be enough, they conclude, when district heating still relies on fossil fuels and carbon prices are high.
The sums can add up quickly. One Applied Energy study with an Aalto co-author found that a single liquid-cooled server rack could supply about 155 MWh of heat a year to an office building, paying back its costs within a year. There is a caveat from Imperial’s Grantham Institute, though: reusing waste heat only makes sense if the data centre runs on renewable power in the first place.
Aalto’s Gopika Premsankar comes at the problem from the other end. Her Cloud and Network Computing group is building tools to measure AI’s carbon emissions across development and use, and to schedule AI workloads in step with the grid. Her work raises a question before we send servers to sea, of how much computing we really need, and when?since 2023
Could ocean thermal energy replace diesel on island grids?
There may be a bigger prize hiding in Drogue’s plan. OTEC works best where surface water stays warm all year, which means the tropics. Many of those places are islands that still burn imported diesel for electricity. Global OTEC was set up precisely to serve them, with plans for a 1.5 MW platform off São Tomé and Príncipe.
If Drogue’s ships prove the technology works at scale, the knock-on effect could be cheaper, cleaner power for island grids, not just for servers. That is where development finance comes in. Institutions such as the International Finance Corporation, part of the World Bank Group, where Imperial economist Ralf Martin has been seconded since 2023, could decide whether OTEC stays a niche idea or reaches the places that need it most.
Lessons from the deep
Whether or not Drogue’s first ship sets sail in 2028, the story holds three lessons for leaders.
Old ideas come back when demand changes. OTEC did not suddenly get better. What changed was the cost of not having enough power. Shelved projects in your own organisation may deserve a second look when the world around them shifts.
Working is not the same as scaling. Project Natick worked, and still stopped. What decides whether an idea scales is whether lenders can rate it, insurers can price it and policy rewards it.
Waste is a choice. Drogue sends its heat into the ocean. Finnish towns pipe it into homes. The smartest energy strategy may be less about finding new supply and more about using what we already make.
As world leaders wrap up the COP31 climate talks in Antalya, with a push for electricity to supply 35% of global energy by 2035, the question Drogue raises is where should the AI economy live, and who pays for its power?
The answer may not be on land. It may not be at sea either. But it is worth asking before the next data centre gets built.
Interested in this topic? You might also like…
