Relocate compute into submerged or floating offshore modules cooled by seawater

The proposal

Place compute in sealed, nitrogen-filled pressure vessels on the seabed in shallow water, or in modules on floating platforms, and use seawater as the heat sink. Power and network arrive by cable from shore, or from co-located offshore generation. The community argument is straightforward: the noise, the generator yard and the municipal water draw are removed from any neighbourhood, because there is no neighbourhood.

Why it would work, and the boundary of that claim

The cooling case is real. Seawater is an effectively unlimited heat sink at stable temperature, eliminating evaporative water consumption and the chiller plant that produces the tonal noise.

The reliability case is real but is commonly misattributed. Microsoft's Project Natick Phase 2 put 864 servers 117 feet down off Orkney from 2018 to 2020 and measured a failure rate one eighth of that on land, 6 failures in 855 units against 8 in 135. Microsoft attributed this to the dry nitrogen atmosphere and to the absence of humans touching the hardware (Microsoft). Neither cause requires water. A sealed, inert-atmosphere, technician-free rack on land should capture most of the same benefit at a fraction of the cost, which is the honest reading of the strongest result the field has produced.

Evidence and current state

• Demonstrated and abandoned. Natick worked and Microsoft chose not to commercialise it. In June 2024 the head of Microsoft Cloud Operations and Innovation said: "I'm not building subsea data centers anywhere in the world" (DCD).
• Demonstrated commercially, at small scale. The Shanghai Lingang facility reached full commercial operation in May 2026: 2.3 MW in phase one of a 24 MW build, about 2,000 servers, claimed PUE of 1.15, zero freshwater, sited between phases of an offshore wind farm, at an investment of about $226 million (Interesting Engineering). Highlander completed a shallow-water unit off Hainan in December 2022, claiming shallow-water design cut costs 80% against deep-water (DCD).
• Claimed only. Atomarine describes floating platforms of 75 to 100 MW with a PUE of about 1.1 that its own site labels "modeled", powered by gas turbines initially and marine reactors later. There is no deployed unit, no named customer and no disclosed funding; every figure is company-sourced (atomarine.co).

Implementation path

Shallow water near an existing landing point and, critically, near offshore generation. Note what actually drives the Shanghai result: siting inside a wind farm. The cooling advantage is worth perhaps 0.25 on PUE; the power supply is the rest of the problem and is not solved by immersion.

Trade-offs and limitations

• It does not address the load. Cooling is roughly 10 to 30% of a facility's total draw. The compute electricity still arrives by cable, and the grid interconnection question is unchanged.
• Serviceability versus refresh cycle. Natick's design was deploy-sealed and retrieve-whole. That is incompatible with a three to five year accelerator refresh, and no operator has published a credible in-place upgrade path.
• Unmeasured marine impact. Scientists consulted on a proposed San Francisco Bay test warned that even small temperature increases could trigger harmful algae blooms and attract invasive species (Gizmodo). Electromagnetic fields from power cables, underwater noise, seabed disturbance, biofouling and decommissioning are all raised, and no source provides measured thermal plume extent, biofouling rates or decommissioning data from any operating facility (CleanTechnica). The absence of evidence runs in both directions and is itself the finding.
• Regulatory exposure. A US startup planned a Bay test without consulting the Bay Conservation and Development Commission or the Regional Water Quality Control Board, both of which warned of heavy fines (DCD).
• The constituency swap. The affected party changes from a residential community with legal standing to a marine ecosystem with no baseline monitoring and no representative. That is an improvement in the politics of siting, not necessarily in the physics of harm.

Conditions under which this is the right answer

Adjacent to offshore generation, at modest scale, in a jurisdiction with clear marine consenting, for workloads that tolerate the latency of a coastal landing point and a hardware refresh measured in whole modules. Outside those conditions the evidence points to sealing the rack on land instead.