Infrastructure optimization
Offshore wind-powered subsea data centers
Evaluating the potential of offshore wind to support gigawatt-scale AI computing through spatial optimization, hourly electricity dispatch, and infrastructure assessment.
Offshore wind
Subsea compute
Project Natick photo by Microsoft; offshore wind photo by Nicholas Doherty on Unsplash.
Optimization framework
Integrating offshore wind, computing siting, and hourly electricity supply
Research manuscript in preparation. The framework integrates spatial optimization, hourly electricity dispatch, and environmental resource accounting across four regions.
The problem
The rapid growth of AI data centers is increasing electricity demand faster than power infrastructure can expand. Offshore wind offers a potential source of low-carbon electricity near coastal demand centers, while subsea deployment could provide additional cooling benefits. However, abundant wind resources do not necessarily translate into reliable or cost-competitive computing capacity. The challenge is to determine where offshore wind can support large data center loads, how much storage and grid integration are needed, and what environmental and infrastructure trade-offs follow.
What I built
I developed a large-scale spatial optimization framework that integrates project-level offshore wind resources, candidate subsea data center locations, and hourly electricity supply across Europe, Asia-Pacific, Brazil, and North America.

The study compares a conventional onshore data center (a) with an offshore wind-powered subsea data center that can draw on storage and, optionally, limited grid imports (b), across six evaluation dimensions (c).
The optimization model decides IT capacity allocation, offshore wind procurement, battery and hydrogen storage, grid connection, and hourly operation, across three offshore wind development portfolios (construction-based, scheduled delivery, and pipeline potential) and three IT deployment targets (40%, 60%, and 80% of each region’s reference). Its hourly capacity expansion and dispatch formulation includes approximately 43 million variables, using regional decomposition to make the optimization computationally tractable.
A post-optimization assessment then evaluates cooling suitability, electricity-related carbon emissions, water consumption, and infrastructure material requirements; these are calculated from the optimized systems rather than optimized directly. The framework compares islanded and grid-supported subsea configurations with conventional onshore grid-powered data centers serving equivalent computing demand. A supplementary spatial screening evaluates potential connectivity and latency constraints using geographic distance proxies; it is not used to reduce the reported IT capacity.
Key findings
Scale
Under the scheduled-delivery wind portfolio, offshore wind could support an annual-energy-equivalent IT capacity of 98.6 GW-IT across the four regions. The representative medium deployment case, set at 60% of that reference, allocates 59.1 GW-IT. At this level, Europe’s 23.7 GW-IT exceeds its projected 2025–2030 data center growth and Asia-Pacific’s 21.9 GW-IT represents approximately 48% of projected regional growth, while North America’s 2.1 GW-IT is a small fraction of its projected increase.

Annual-energy IT capacity references under three offshore wind development portfolios, compared with projected 2025–2030 data center growth (a); locations and sizes of the 59.1 GW-IT medium deployment allocation (b); and how concentrated that capacity is by site size (c).
Economics
Islanded systems, which rely only on offshore wind and storage, cost 1.96 to 3.64 times regional industrial electricity prices. Allowing grid imports to supply up to 20% of annual facility electricity reduces supply costs to USD 116.9/MWh in Europe and USD 86.8/MWh in Asia-Pacific, 78% and 92% of their industrial price benchmarks, mainly by cutting storage requirements. Brazil and North America remain above their benchmarks. Limited annual imports do not mean limited connections, however: optimized grid connection capacity reaches 63–70% of installed IT capacity, while average connection utilization is only about 25%.

Electricity-supply costs relative to regional industrial electricity prices at medium deployment (a), across low, medium, and high deployment targets in each region (b–e), and broken down by component (f). Costs cover electricity supply only, not total data center costs.
Environmental trade-offs
Electricity-supply emissions are 94.5% lower for islanded configurations and 80.1% lower with grid support than matched onshore grid-powered facilities, a reduction of roughly 135 to 158 Mt CO₂e a year at medium deployment. These figures exclude embodied emissions from storage and marine infrastructure. Other outcomes depend more on location and configuration. Seawater cooling suits 90.9% of allocated capacity in Europe and all of it in North America across ten years of monthly temperatures, but only 35.9% in Asia-Pacific. Displacing conventional generation saves the most water in Asia-Pacific and Europe. Islanded systems can raise copper and lithium requirements, and grid support cuts the additional lithium requirement by about 84%.
What it is built to answer
Where could offshore wind support large-scale AI data centers, and how do electricity costs, reliability-related requirements, cooling conditions, and infrastructure needs vary across locations?
The framework helps identify regions where renewable electricity supply and computing demand could be better coordinated. It also reveals why abundant offshore wind resources alone do not guarantee economically or environmentally favorable deployment.
Research manuscript in preparation.
Limitations
The study evaluates regional deployment potential and electricity-supply configurations rather than detailed engineering designs for individual subsea facilities. The annual-energy capacity references are not guarantees of firm hourly supply, and cooling suitability is screened using historical monthly temperatures.
Cost comparisons exclude several subsea construction and maintenance components, while emissions and material accounts do not represent complete facility life-cycle assessments. Network latency is evaluated using geographic proxies rather than detailed routing simulations. Site-specific engineering and infrastructure comparisons are needed before drawing conclusions about actual project feasibility.