Floating wind and GPUs: the 800kW test of gridless AI

Exascale Labs and EnergyBank have signed a non-binding memorandum of understanding to develop AI data centers alongside floating offshore wind and long-duration battery storage in Norway, with the pilot sited at Hywind Demo, the Equinor-developed floating turbine that Zephyros acquired in 2019 and that has been producing power since 2009 at roughly 2.3MW. The bet is that power delivery, not chip supply, is now the constraint on AI infrastructure.

The pair wants to build an 800kW modular data center that is completely independent of the grid—an Exascale unit packing GPU server racks, power distribution, UPS battery backup, switchboards, and management systems into a single deployable enclosure—while EnergyBank expects to fund the pilot through private investment and potential public money.

A non-binding MoU is a statement of intent, which makes the deal a feasibility study with a near-term engineering schedule rather than a financed certainty—and the case behind it is that AI infrastructure growth is no longer constrained by chip availability but by the ability to secure power at scale on commercially viable timelines. Exascale CEO Hoonsao Lee puts power delivery, rather than advanced semiconductors, as the constraint, and EnergyBank CEO Tim Hawkey argues that a single off-grid system combining floating wind, long-duration storage, and modular compute can turn variable offshore wind into firm, high-value compute without waiting years for a grid connection.

What Hawkey describes is more interesting than the hardware: wind developers gain a direct route to market, compute operators gain access to new power capacity, and the project captures substantially more value from the same energy asset—the language of an energy project rather than a data center build-out.

The companies plan to start with technical feasibility work, produce a reference architecture for the pilot, and build a roadmap for scaling beyond it, targeting up to 10MW of compute per turbine—a figure that assumes next-generation units more than four times the size of Hywind Demo, which produces only 2.3MW.

Aikido launched a US floating wind platform aimed at the same pairing in March, but the Norway project is rooted in a turbine with an actual operating history, which gives it more credibility than a render.

Grid access has become the scarce asset in the AI buildout: interconnection queues and consent timelines, not chip supply, now decide where capacity gets built, and the Norwegian pilot tests an alternative—making the grid optional by integrating generation, storage, and compute into one package that floats off the coast.

The announcement names no hyperscaler offtaker, no fixed price for compute, and no utility subscription; funding is expected from private investment and potential public money rather than a signed power purchase agreement, placing the project in the merchant category where public-market underwriters have been most cautious.

Completion without an offtake is merchant risk wearing an infrastructure costume; for now, the pilot is small enough that the risk is manageable, but the 10MW-per-turbine target, if reached, would mark a shift from proof-of-concept to production.

The 800kW pilot is small enough that the test is the all-in cost per megawatt of firm compute—generation, storage, and data center together. If the pilot demonstrates that off-grid wind can deliver compute at a cost close to grid-fed power, the concept will carry; if not, the pilot will stay exactly that—a pilot. The reference architecture will show whether a floating turbine can carry the electrical load of the GPU servers and whether long-duration batteries can smooth the wind's intermittence, and the first results are worth waiting for, but the next turbine class—the 10MW units the companies are targeting—will be the real test.