Subsea and harsh-environment compute

Reading order

Separate three environments before you compare them. A sealed vessel on a seabed is one environment. A barge or shoreline plant that uses adjacent water as a heat sink is another. A land hall in a cold, windy or corrosive climate is a third. Microsoft Natick is the first. Nautilus Data Technologies’ public site describes water-source heat rejection and liquid-cooling equipment that can use natural or industrial water; that is a vendor description of a heat-sink method, not a Swedish permit. Meta’s Luleå page is the third: outdoor air and hydropower on land.

Then classify the evidence state. A retrieved research vessel with a dated company article is stronger than a rendering. A 2024 company statement that the builder is not constructing subsea halls is stronger than an old brochure that still talks in the present tense. A vendor homepage without a dated Swedish deployment is not a Baltic case study. If the source does not name a Swedish municipality, do not invent a pin.

Only after those classifications should you open Swedish coastal questions. The west-coast GleSYS Falkenberg file is a land hall with a 1 October 2020 heat-delivery start to Falkenberg Energi. It is not an underwater vessel. Staffanstorp in SE4, named in Microsoft’s Sweden-region materials, is an inland municipal name in that announcement, not a subsea site. No project in this register is stored as a Swedish subsea data center.

• Sealed seabed vessel, water-side heat sink, cold-climate land hall, or orbit?
• Dated trial, current product, or retired research programme?
• Which country and which water body does the source actually name?
• Is any Swedish municipality, network owner or environmental paper attached?

What a harsh environment actually changes

A sealed subsea vessel changes access, atmosphere and the heat-rejection path. Microsoft’s 2020 article says the Northern Isles used submarine-style heat-exchange plumbing, sat in consistently cool subsurface water, and was filled with dry nitrogen. The company hypothesized that the nitrogen atmosphere and the absence of people bumping equipment explained higher reliability. Those are research hypotheses attached to one retrieved vessel. They are not a Swedish operating manual and they are not a PUE league table.

Water as a heat sink on a barge or shoreline plant is a different mechanism. Heat still leaves the IT load into a fluid loop, then into a larger body of water or an industrial sink. Nautilus’s homepage describes vacuum-sealed, leak-oriented distribution equipment and compatibility with natural or industrial water sources, and it claims zero water consumption when using natural sources. That is vendor language. It does not name a Swedish lake, a Baltic permit or a measured Swedish PUE. A Swedish water-and-cooling file still has to name the water source, the volume assumption and the competent environmental counterpart.

A cold Swedish land climate is again different. Outdoor air in Luleå can reduce mechanical cooling hours. It does not seal the hall in nitrogen, it does not remove neighbours, and it does not replace a network-owner letter. Backup generators, winter construction and icing become part of the noise and access file. Naturvårdsverket’s Environmental Code guidance is the starting page when water, noise or other environmental questions open. Boverket’s planning path still sees a building on a plot. Submerging a rendering in a slide does not move those authorities offshore.

Marine placement, if anyone ever proposed it in Swedish waters, would open a case-specific Environmental Code and planning conversation that this register has not seen. This page does not invent a Swedish marine-permit timeline, a national month count, or a “Baltic Natick” zone. If no primary source names the water body, the depth, the cable landfall and the competent authority, the Swedish coastal product does not exist in this file.

What named primary sources show

Microsoft’s 14 September 2020 article is the retrieval-and-results source. It says marine specialists retrieved a shipping-container-size datacenter from the seafloor off Orkney earlier that summer, that the Northern Isles had been deployed 117 feet down in spring 2018, and that the team monitored it for two years. It records a 105-day Pacific proof-of-concept in 2015 as Phase 1. It says the vessel was manufactured with Naval Group and Naval Energies, with Green Marine supporting deployment and retrieval at the European Marine Energy Centre, where tidal currents and large storm waves are part of the site description.

The same article says researchers collected air samples from a nitrogen-filled vessel, removed racks at Nigg Energy Park, and compared failed servers with a land replica. Microsoft’s Ben Cutler is quoted: the failure rate in the water was one-eighth of that on land. The 2020 article also mentions 864 servers in one photo caption and twelve racks in another. Microsoft Research’s Phase 2 page states that servers in Natick Northern Isles showed a failure rate of one-eighth of a land-based control group, that Phase 2 aimed to manufacture and deploy full-scale modules in under 90 days from decision to power on, and that the vessel was deployed at EMEC on 1 June 2018. Those are Microsoft research sentences. They are not Swedish operating statistics.

Data Centre Dynamics’ 17 June 2024 article is the later programme-status source. It quotes Walsh saying she was not building subsea data centers anywhere, that the team had learned about operations below sea level, vibration and server impacts, and that those learnings would be applied elsewhere. It prints a Microsoft statement that the company did not currently have data centers in the water and would continue to use Project Natick as a research platform, for example for liquid immersion. DCD’s later server counts (855 underwater versus 135 on land) differ from the 864 figure on Microsoft’s 2020 page. This register keeps both publications visible rather than inventing a reconciled total.

Nautilus Data Technologies’ homepage describes modular liquid-cooling equipment and water-source heat rejection. It does not, on the cited page, name a Swedish municipality or a Baltic deployment. GleSYS’s 1 October 2020 Falkenberg release remains the dated Swedish heat-delivery start in this set — on land. Meta’s Luleå page remains the dated northern land-campus cooling description. No source here publishes a Swedish subsea megawatt or a Swedish marine-permit decision for a data hall.

What a Swedish coastal question still is

A Swedish west-coast or Baltic conversation, if it ever becomes real, starts as the ordinary establishment file plus water. Who owns the land or the water area? Which detaljplan or shoreline instrument applies? Who is the network owner at the landfall voltage? Which environmental questions — noise, fuel, water intake or heat rejection — have a competent authority? Those questions are the same questions as on an inland plot, with a wetter site visit. They are not answered by reprinting Orkney photographs.

GleSYS in Falkenberg shows how a coastal Swedish hall actually appears in this register: a land building, a utility-power sentence and a dated heat start to a named municipal energy company. EcoDataCenter’s Falun recovery wording with Falu Energi & Vatten is inland district heat, not seawater. atNorth SWE01’s Exergi mention is a Stockholm metro hall. None of those rows is a vessel on a seabed.

Harsh Swedish climate is already in the land file. Winter construction, generator testing in cold air and neighbours who hear fans year-round sit on the noise-and-permit-timelines page. They do not become subsea facts because the marketing deck shows waves. If a future primary source names a Swedish water body and a decision verb, add that row. Until then, the gap is the finding.

Common misreads

The first misread is to treat Natick as a current Microsoft product. The 2020 article is a retrieval story. The 2024 DCD piece records a company statement that Microsoft did not then have data centers in the water. The second is to copy Orkney’s renewable-grid sentence onto a Swedish coastal plot. Energimyndigheten’s national mix still does not connect a property. The third is to merge Microsoft’s 864-server caption with DCD’s later 855-versus-135 comparison into one invented official total.

The fourth misread is to treat a vendor water-cooling homepage as a Swedish marine case. The fifth is to call Meta Luleå a harsh-environment subsea analogue. It is a land campus with outdoor air. The sixth is to invent a Swedish month count for a marine Environmental Code case that has not been located. Guidance pages are not case law, and this page does not write a judgment.

What to ask next

Ask whether the speaker means a retrieved trial, a retired programme, a water-cooled land or barge product, or only a coastal land hall. Ask for the water body, the depth or shoreline instrument, the cable landfall and the competent authority. If those cells are empty, the Swedish product is not in the file.

Then open the water-and-cooling page, the noise-and-permit-timelines page and the ordinary grid-queue questions. A harsh environment does not retire those papers. It usually adds papers. Record the gap instead of pasting Orkney reliability multiples onto a Baltic slide.

1. Name the environment: seabed, barge, shoreline, or cold land hall.
2. Attach the latest dated programme-status source.
3. Do not merge conflicting server counts.
4. Ask for the Swedish water body and authority — or record none.
5. Keep Orkney reliability claims in the Orkney row.

What this does and does not prove

This page proves that Microsoft published a completed Orkney subsea trial in 2020 and that a 2024 news interview recorded Microsoft as not then building or operating underwater data centers. It proves that a vendor describes water-source heat rejection without a cited Swedish site. It proves that Swedish examples in this register that sit near coasts or in cold air remain land halls.

It does not prove that subsea compute is cheaper or cleaner in Sweden. It does not invent a Baltic megawatt or a marine-permit month count. It does not treat Natick reliability as a Swedish SLA. It does not place Nautilus, Highlander or any other foreign vessel on a Swedish map.

Unknown Swedish marine cases stay unknown. Point to the missing municipality, water body and decision. A later primary source may add one. Until it does, Orkney stays Orkney, a 2024 “not building” sentence stays a 2024 sentence, and a Swedish coastal plot stays a land, plan, grid and environment file.