AI campus density versus traditional colo

Reading order

Name the commercial wrapper before you name the rack. Who owns the IT load? Who operates the building? Do retail customers share the hall? Is high-density compute a current contracted requirement or only a future design target? A CoreWeave-on-Conapto sentence and an evroc GPU envelope can use the same marketing language and still be different objects: one is a customer on a colo platform, the other is an owner-led hall design without a published megawatt.

Then classify every power figure. IT load, facility load, site power, utility import and an unclear announced number are different cells. atNorth SWE01’s 12 MW is operator site power. EcoDataCenter’s 41 MW and 65 MW annual-report sentences are campus capacity rows with different dates and bases. The 80 MW Falun marketing sentence is available power. None of those cells is a watts-per-rack series, and none is created by writing “AI” on the slide.

Then ask what sits at the rack and what rejects heat outside. Liquid loops, coolant distribution units and a defined outdoor path appear when the design target is high density. Air may still serve a traditional colo hall. Each path opens a different water, noise and heat-reuse question. Do not start with a national ranking of AI counties or with a claim that training must sit in SE1.

• Owner-occupied campus, retail colo, or a customer on someone else’s floor?
• Design GPU envelope versus installed kit.
• IT, facility, site, utility — or unclear MW.
• Rack-level cooling versus outdoor rejection.

What density changes — and what colo still is

Higher density concentrates heat in a smaller footprint. The cooling plant must move that heat to a fluid or to a larger air volume, then reject it outside. Liquid-to-chip or rear-door loops can raise the temperature of recoverable heat, which is why density and district-heat conversations often arrive together in Stockholm operator copy. They also add leak-detection, coolant chemistry and a maintenance aisle that a traditional air hall may not have. Those are establishment facts when a named source states them. They are not a published Swedish watts-per-rack table. This register does not invent one.

Power quality and import shape change. A training cluster can ramp faster and sit at a higher load factor than a mixed retail colo floor. The network owner still answers a requested import at a named voltage and point of connection. Svenska kraftnät’s indicative map does not replace that answer. A cheap hour in SE1 does not prove that an AI campus can connect in Luleå, and a dense rack in Kista does not prove residual capacity on the next Stockholm plot.

Neighbours hear a different plant. High-density rejection can mean more outdoor equipment, more heat-pump skids and different generator-test assumptions if the critical load is larger. Naturvårdsverket’s environmental guidance is the starting page; it is not a project acoustic decision. Boverket’s detaljplan page does not contain an “AI-ready” zone. A building permit does not certify that the hall can run a training cluster.

Traditional colo remains a commercial pattern, not a low-density law. A colo hall can install liquid cooling, as SWE01’s operator page describes. A colo hall can host an AI customer, as Conapto’s CoreWeave notice describes. The wrapper still matters: the customer contract does not create a second campus, and the operator’s site-power sentence does not become the customer’s IT load. Meta’s Luleå campus shows the opposite wrapper — owner-occupied hyperscale with outdoor air and no GPU envelope on the cited page. One label cannot rank those rooms.

What named primary sources show

evroc’s 3 February 2025 notice records 4,890 square metres of building rights in Arlandastad, a design to accommodate up to 16,000 GPUs, a EUR 600 million investment intention, and company aims to break ground in the first half of 2025 and operate in the latter half of 2026. Megawatts are not stated. Ground-breaking after the H1 2025 aim is not confirmed in a later primary source used here. The GPU number is a design envelope, not installed kit and not a rack-kilowatt.

EcoDataCenter’s Falun site page markets the campus for HPC and AI and states 80 MW available power and a PUE of 1.2 according to ISO standards. The 2025 annual report says buildings C and D were completed and commissioned and that site capacity increased from 29 MW to 41 MW, with E and F under construction and 65 MW once operational. Those megawatt sentences stay in separate rows. They are not a GPU count and they are not added together.

atNorth’s SWE01 page is the clearest colo-density contrast in the set: an operating Kista hall, 6,400 square metres, 12 MW site power, design PUE 1.2, liquid cooling and Exergi heat recovery. Occupied IT load is unpublished. The page is one metro hall, not a national AI census. Conapto’s 24 June 2026 CoreWeave notice is a colocation agreement to expand AI cloud capacity on Conapto’s Swedish platform. It does not publish a new site megawatt and does not create an extra campus in this register.

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 do not publish a GPU envelope. The 2011 Associated Press energy-need figure remains a dated announced claim with an undefined metric. Microsoft’s 16 November 2021 Sweden-region opening names Gävle, Sandviken and Staffanstorp without facility megawatts in the cited launch materials. No source here publishes a comparable Swedish kW/rack series for AI versus colo.

A comparison you can defend

A defensible comparison has the same wrapper, the same megawatt definition and the same date type on both sides. You can compare two operator site-power sentences if you keep them labelled as operator site power. You cannot compare evroc’s GPU envelope with SWE01’s 12 MW and call the result a density ranking. You cannot compare Conapto’s customer agreement with EcoDataCenter’s annual-report 41 MW and call the result a Swedish AI-campus census.

You can compare mechanisms. An owner-occupied training hall will usually own the cooling choice and the heat-buyer conversation. A retail colo hall will usually publish a hall product and a shared rejection path. An AI customer in colo inherits the hall’s plant unless the contract says otherwise. Write that inheritance. Do not invent the rack kilowatts the contract does not publish.

Common misreads

The first misread is to convert 16,000 GPUs into megawatts. The notice did not. The second is to add Falun’s 41, 65 and 80. They are different rows. The third is to treat SWE01’s 12 MW and design PUE as a league score against Meta or EcoDataCenter. Different years, methods and wrappers. The fourth is to treat the CoreWeave agreement as a new Swedish site. The fifth is to publish a kW/rack figure that no cited source printed. The sixth is to send every training workload to SE1 because the hour looks cheap.

What to ask next

Ask for the wrapper, the megawatt definition, the rack-level cooling device and the outdoor rejection path, each with a date. Ask whether any kW/rack figure is actually published. If it is not, leave the cell empty. Ask the network owner which paper exists for the requested import. Ask the heat company only after the loop temperature exists.

Then open the parent AI/HPC page, the cooling page, the district-heating page and the grid-queue page. Density is a reason to ask better questions. It is not a reason to invent a Swedish rack standard.

1. Wrapper: owner campus, colo hall, or customer deal.
2. MW cell typed and dated — or empty.
3. Cooling device at the rack and outside.
4. No invented kW/rack.
5. Customer contract not stored as a new site.

What this does and does not prove

This page proves that AI-labelled Swedish projects in the reviewed set change wording discipline and cooling questions, and that colo can host dense customers without becoming a second campus. It proves that a GPU envelope, a site-power sentence and an annual-report campus megawatt are different objects.

It does not prove a Swedish watts-per-rack standard. It does not rank counties for training. It does not convert CoreWeave into commissioned IT load. It does not improve anyone’s PUE by comparison.

Unknown rack density stays unknown. Point to the named source gap. A later primary source may add a measured rack figure or a commissioning date. Until it does, the GPU cell and the MW cell remain separate, and traditional colo remains a wrapper rather than a density law.