AI and HPC infrastructure

Reading order

Read an AI or HPC sentence as an establishment claim, not as a product review. First ask whether the GPU or node number is a design envelope or installed equipment. Then ask what cooling device sits at the rack and what rejects heat outside the building. Then classify every power figure as IT, facility, site, utility, or unclear. Only after those three cells exist should you ask which workload — training or inference — is actually contracted.

Do not start with latency slogans or with a national AI ranking. Training versus inference is a later specialist distinction and does not, by itself, locate a hall in SE1 or SE4. A cheap hour in SE1 Luleå does not prove that an AI campus can connect. A dense rack does not prove that the municipal detaljplan allows the use. Keep the marketing line visible, then walk the power, cooling and land files in that order.

• Is the GPU number a design envelope or installed kit?
• What cooling device sits at the rack, and what rejects heat outside?
• Is the power figure IT, facility, site, utility, or unclear?
• Which workload — training or inference — is actually contracted?

How AI and HPC change a Swedish establishment file

Higher rack density changes the cooling plant before it changes the press release. Air may still serve lower-density halls. Liquid loops, coolant distribution units and a defined rejection path appear when the design target is high-density compute. That path may reject heat to outdoor air, to a tower, or toward a district-heat network. Each path opens a different water, noise and neighbour question. Naturvårdsverket’s environmental guidance is the starting page when those questions become Environmental Code matters. A liquid design is not a permit.

Power wording becomes stricter, not larger by default. An AI label does not create megawatts. The network owner still answers a requested import capacity at a named voltage and point of connection. Svenska kraftnät’s indicative transmission map does not replace that answer and does not show underlying-network capacity. Bidding areas SE1 Luleå, SE2 Sundsvall, SE3 Stockholm and SE4 Malmö remain wholesale-price areas. They do not reserve GPU power on a plot in Sigtuna, Falun or Luleå. A national fossil-free mix is still not a local connection offer.

Networks and workload sit in different cells from power. Training and inference impose different latency, storage and cluster-size questions, but this page refuses universal latency or flexibility claims. Fiber diversity, carrier entry and a contracted interconnect are establishment facts only when a named source states them. A cloud-region product, a retail colo contract and an owner-occupied training hall can all use GPU language. The commercial wrapper does not tell you the rack density or the cooling fluid.

Municipal planning remains a land-use file. Boverket’s detaljplan guidance does not contain an AI-ready zone. A building permit does not certify that the hall can run a training cluster. Inspektionen för strategiska produkter may screen certain investments on activity and investor facts; an “AI factory” headline is not the screening test. Skatteverket’s energy-tax pages do not convert a GPU count into a deductible. Keep each institution in its own letterhead.

What named primary sources show

On 3 February 2025 evroc said it had purchased 4,890 square metres of building rights in Arlandastad, aimed to break ground in the first half of 2025, and expected operations in the latter half of 2026. The same notice restated an intention to invest EUR 600 million and a design to accommodate up to 16,000 GPUs. Megawatts are not stated. Ground-breaking after the H1 2025 aim is not confirmed in a later primary source used here. The euro figure is an intention, not a recorded spend. Commissioning is not treated as documented.

EcoDataCenter’s Falun site page markets the campus for HPC and AI and states 80 MW available power. The 2025 annual report says buildings C and D were completed and commissioned and that site capacity increased from 29 MW to 41 MW. Buildings E and F were under construction, with a stated 65 MW once they are operational. Those three megawatt sentences are separate rows from different measurement bases. They are not a GPU count, and they are not added together. Heat delivered versus heat-recovery design is not quantified on those pages.

atNorth’s SWE01 Stockholm metro page describes an operating Kista hall with 6,400 square metres, 12 MW site power capacity, a design PUE of 1.2, liquid cooling and heat recovery with Stockholm Exergi. Occupied IT load is not published. The page is an operator description of one metro hall, not a national AI inventory. The design PUE is an operator figure and is not compared here with other halls as a league table.

Meta’s Luleå materials describe an owner-occupied campus operational from 2013, outdoor-air cooling and local hydropower. They do not state current megawatts on the cited page, and they do not publish a GPU envelope. That silence matters: a long-running hyperscale campus can exist without an AI slogan, and an AI slogan can exist without a megawatt. The 2011 Associated Press energy-need figure is a dated announced claim with an undefined metric and is not used as current operating capacity.

Common misreads

The first misread is to convert GPUs into megawatts with an unpublished watts-per-GPU factor. evroc’s notice did not state that factor and did not state MW. Inventing one would be a new figure, not a reading. The second misread is to treat “AI factory” as a Swedish planning class. Boverket does not publish such a class. The third is to add EcoDataCenter’s Falun megawatt sentences because the marketing page also says HPC and AI.

A fourth misread is to treat liquid cooling as proof that heat is sold. atNorth SWE01 describes heat recovery with Stockholm Exergi; delivered annual megawatt-hours are not on the cited page. A fifth is to treat a northern climate or a SE1 price as an AI site offer. Outdoor-air cooling in Luleå is an operator-described design, not a connection agreement and not a GPU permit.

• Minting megawatts from a GPU design envelope.
• Adding Falun’s 41 MW, 65 MW and 80 MW rows.
• Reading design PUE as a national AI ranking.
• Treating liquid cooling as delivered district heat.
• Using SE1–SE4 as a spare-MW allocation for training halls.

What to ask next

Ask for the dated document behind each AI word. If the claim is a GPU envelope, ask whether any kit is installed and which generation the envelope assumed. If the claim is power, ask the network owner which paper exists — inquiry, reservation, agreement or energized load — and which megawatt definition it uses. If the claim is cooling, ask for the rack-level device, the outdoor rejection path, and whether water volumes have a source.

Then ask which workload is contracted, for whom, and for how long. Refuse a universal statement that training must sit in the north or that inference must sit in Stockholm. Open the cooling page for fluids and rejection, the heat-reuse page for buyer and contract state, and the capacity-definitions page before you compare two MW cells. If the operator will not put a number on paper, record the gap.

• Design envelope versus installed GPU or node count, with a date.
• IT versus facility versus unclear megawatts.
• Rack-level cooling versus outdoor heat rejection.
• Contracted training or inference, not a headline workload.
• Which Swedish authority or network-owner paper is still missing?

What this does and does not prove

This page proves that AI and HPC language in the reviewed Swedish set changes cooling, density and wording discipline, and that it does not by itself create a megawatt. It proves that evroc published a GPU design without MW, that EcoDataCenter published HPC and AI marketing beside separate campus-MW rows, and that atNorth published a liquid-cooled metro hall with a design PUE and an Exergi mention.

It does not prove how many AI halls Sweden has, what national GPU capacity exists, or which county is best for training. It does not publish watts per GPU. It does not rank operator PUE figures. It does not treat company aims for ground-break or operation as completed events. It does not turn a hydropower description in Luleå into a connection offer for an AI plot elsewhere.

Unknown capacity stays unknown. Point to the named source gap and to the capacity-definitions methodology. A later primary source may add a megawatt boundary or a commissioning date. Until it does, the GPU cell and the MW cell remain separate, the company aim for ground-break or operation stays an aim, and the slogan stays a slogan rather than a commissioned load.