Hyperscale versus enterprise colocation — a decision file, not a league

Reading order

Start with the holder of the papers, not with the IEA cooling share. Write who holds the network-owner letter, who operates the building, who owns the IT load, who would file the Union energy-performance report, and who would sit as heat counterparty. A shared hall can put those papers on the operator. An owner-occupied campus can put them on the occupier. A wholesale suite can split them. The IEA 7 percent and more-than-30 percent figures do not assign those inboxes.

Then classify the Swedish row you actually have. Equinix SK1 and SK2 are published Stockholm colocation halls. Digital Realty’s Stockholm campus is written as six halls. STACK’s Stockholm page states 18 megawatts on a multi-tenant metro site. Bahnhof publishes Pionen without a megawatt on the cited page. Conapto’s 5 September 2024 south-campus opening states 24 megawatts; the 24 June 2026 CoreWeave notice is a colocation agreement. GleSYS publishes Stockholm and Falkenberg. Those are shared-hall rows. Meta Luleå is operational from 2013 with outdoor air and PUE 1.09 and no current megawatt on the cited careers page. Microsoft named Gävle, Sandviken and Staffanstorp in 2021 without facility megawatts. AWS Europe (Stockholm) opened 12 December 2018 without a municipality on the cited launch pages. Google Horndal has a 2 June 2026 groundbreaking and no campus megawatt on the cited pages. Those are occupier or cloud-region rows.

Only then open the parent pages. Types, colo-versus-build-to-suit and edge-versus-hyperscale already own pattern, wrapper and scale. The training-versus-inference page already forbids converting a GPU count into a workload megawatt. This page does not repeat those matrices. It only asks whether the brief is choosing a holder of papers or pasting a global cooling share onto a Swedish hall.

• Name the holder of the connection, heat and reporting papers.
• Keep IEA 7 percent and 30 percent off any named Swedish hall.
• A colo agreement is not a new campus.
• Parent type, wrapper and scale pages last.

What the decision actually splits

The decision splits holders, not climates. A shared hall can sell a cabinet or a suite while the operator holds the connection conversation, the building permit file and, often, the heat counterparty. An occupier campus can put the connection letter, the environmental matrix and the Union reporting duty on the occupier. A cloud region can name municipalities, as Microsoft did, or withhold a municipality, as the cited AWS launch pages do. Those are evidence states. They are not efficiency medals.

The IEA cooling shares are about equipment classes inside halls labelled efficient hyperscale or less-efficient enterprise. They are not a measurement of SK1, Pionen, SWE01 or Luleå. Servers around 60 percent are likewise a modern-hall equipment share, not a Swedish type census. All three IEA types, the chapter says, contribute to growth. That sentence forbids a slide that says Sweden must choose one type.

A colo customer can still be a hyperscale-like workload. Conapto’s 24 June 2026 CoreWeave notice is a colocation agreement for an AI-cloud customer. It does not relabel Conapto as a hyperscale campus and it does not publish a training megawatt. evroc’s 3 February 2025 notice gives building rights, a GPU design and a euro intention without a megawatt. A GPU count is not a type decision.

Megawatts stay on their own dated rows. STACK 18 megawatts, Conapto 24 megawatts, atNorth SWE01 12 megawatts, SWE02 30 megawatts announced, SWE04 300 megawatts announced, EcoDataCenter Falun 41 / 65 / 80 on separate dates — the market hub already keeps those unsummed. This page will not add them into a hyperscale total or an enterprise-colo total.

What named primary sources show

The IEA demand chapter shows three growing types and the 7 percent / more-than-30 percent cooling shares plus the 60 percent server share. Equinix, Digital Realty, STACK, Bahnhof, Conapto and GleSYS show shared-hall or multi-tenant publication. Conapto–CoreWeave shows a dated colo agreement. Meta, Microsoft, AWS and Google show occupier or cloud-region publication with the megawatt and municipality gaps already stored in this register.

atNorth SWE01, SWE02 and SWE04 show operator campus pages and announced site-power figures. evroc shows land and GPUs, not megawatts. None of those rows is a national type census.

A matrix instead of a type winner

The table is a permission table. A filled cell is a claim the named URL can carry. An empty cell means this page will not invent the missing Swedish type census or convert a cooling share into a hall ranking.

Do not treat the IEA 7 percent as Equinix inefficiency, and do not treat Meta’s outdoor-air sentence as a hyperscale mandate for every Swedish buyer.

Common misreads

The first misread is to treat the IEA 7 percent cooling share as a Swedish hyperscale medal. The share is a global efficient-hyperscale equipment class. The second is to treat a Stockholm colo hall as the IEA’s less-efficient enterprise class. The chapter does not name SK1 or Pionen. The third is to treat a CoreWeave colo agreement as a new hyperscale campus. The notice is a contract sentence.

The fourth is to convert evroc’s 16,000 GPUs into a type decision. The fifth is to add 18, 24, 12, 30 and 300 megawatts into one Swedish hyperscale total. Those figures live on different rows. The sixth is to treat a missing AWS municipality as proof that cloud regions are not Swedish files. The gap is a publication gap, not a type.

What to ask next

If a brief picks a winner between hyperscale and enterprise colo, ask which named URL assigns the connection letter, the heat paper and the reporting duty for that buyer. If the URL only shows an IEA cooling share or a GPU count, the brief has left the source.

Then open the types page, the colo-versus-build-to-suit page and the edge-versus-hyperscale page. Those pages already own pattern, wrapper and scale. This page only asks who holds the papers.

1. Who holds the network-owner letter?
2. Who would file the Union energy-performance report?
3. Is the IEA 7 percent or 30 percent being pinned to a named Swedish hall?
4. Is a colo agreement being treated as a new campus?
5. Are mixed megawatt rows still unsummed?

What this does and does not prove

This page proves that the IEA demand chapter uses three growing types and global cooling shares, and that named Swedish pages already show both shared halls and occupier or cloud-region files. It proves that a colo agreement and a campus announcement are different objects.

It does not prove which Swedish hall is efficient hyperscale. It does not prove which Swedish hall is less-efficient enterprise. It does not invent a national type census.

Empty cells stay empty. Until a later primary source names a Swedish hall as one IEA type with a dated megawatt and a boundary, the decision stays a method, not a ranking.