Training versus inference in the data-centre energy story

Reading order

Start with the IEA Energy and AI landing page already in the source registry. It is a 2025 report, not a later key-questions landing page, and it is the IEA energy-and-AI document this page treats as the report row. Then open the demand chapter of the same report. That chapter is the place that says training and deploying AI models take place mainly in data centres, and that assigns approximate electricity shares to servers and to cooling. Do not treat the landing page and the chapter as two independent global totals. They are one report with two URLs.

After the IEA rows, read Swedish operator pages that actually use training, deployment, GPU, AI-focused or customer-workload words. EcoDataCenter’s DeepL notice talks about NVIDIA DGX GB200 systems and occupancy from 11 June 2025. EcoDataCenter’s 11 February 2026 Mistral notice talks about an AI-focused facility in Sweden. The 24 September 2025 Borlänge construction-start notice talks about 250 megawatts of access at launch and potential to scale to 600 megawatts — those are access and potential figures on a different date, not an IT-load split between training and inference. evroc’s 3 February 2025 Arlandastad notice talks about 4,890 square metres of building rights, a design of up to 16,000 GPUs and a EUR 600 million intention. Conapto’s 24 June 2026 CoreWeave notice is a colocation agreement, not a new campus and not a training-versus-inference label.

The cited Meta Luleå careers page describes outdoor-air cooling, local hydropower and an operator PUE of 1.09. It does not, in the wording this register stores, name the campus as a training hall or an inference hall. Only after those operator rows should you open the parent AI and HPC guide and the high-density racks guide. Those pages already own GPU-count-is-not-megawatts and rack-density-is-not-site-IT-load. This page does not repeat their matrices. It only asks whether a named Swedish sentence is about training, about deployment, or about neither.

• IEA Energy and AI landing page and demand chapter first — one report, two URLs.
• Then named Swedish operator sentences that actually say AI-focused, GPU, GB200 or colocation.
• Keep the 24 September 2025 Borlänge access figures off the February 2026 Mistral notice.
• Parent AI/HPC and high-density guides last — they own density and GPU-count rules, not this split.

What the sources actually separate

The IEA demand chapter is written as a global and regional energy story. It says training and deploying AI models take place mainly in data centres. It does not publish a Swedish training megawatt total and it does not publish a Swedish inference megawatt total. A reader who needs a national split will not find it there. The same chapter says that in a modern data centre, servers can account for around 60 percent of electricity use, while cooling can range from about 7 percent in efficient hyperscale halls to more than 30 percent in less-efficient enterprise halls. Those shares are about equipment classes inside a hall, not about whether the workload on the servers is training or inference.

The same chapter says electricity consumption in accelerated servers is projected to grow by 30 percent a year in the Base Case, and that those servers account for almost half of the net increase in global data-centre electricity. Conventional server electricity grows more slowly, at 9 percent a year, and accounts for around 20 percent of the increase. Those are global growth rates for equipment classes. They are not a Swedish training table and not a named hall in Falun, Borlänge, Arlandastad or Luleå.

The Commission data-centre energy-performance page, already in this register, quotes IEA Energy and AI figures of about 415 terawatt-hours of global data-centre electricity, about 1.5 percent of global electricity, and a path toward about 945 terawatt-hours by 2030. Those are Union-level sentences that rest on the IEA work. They are not a Swedish training-versus-inference table, and they are not a named Swedish hall inside the 415 terawatt-hour figure.

Swedish operator pages that use AI words do not fill the IEA gap. EcoDataCenter’s DeepL page and DeepL’s own 11 June 2025 notice say DeepL is deploying NVIDIA DGX SuperPOD with DGX GB200 systems in Europe, with occupancy from that date. That is an occupancy and product sentence. It is not a training megawatt and not an inference megawatt. EcoDataCenter’s 11 February 2026 Mistral notice says the parties will build an AI-focused data centre in Sweden. That notice, as this register already stores it, does not publish a Borlänge megawatt. The 24 September 2025 construction-start notice is the row that says access to 250 megawatts at launch and potential to scale to 600 megawatts. Those two dates stay on two rows.

evroc’s 3 February 2025 notice gives 4,890 square metres of building rights, a design of up to 16,000 GPUs and a EUR 600 million intention. It does not publish a megawatt figure. A GPU count cannot be turned into a training load or an inference load on this site. Conapto’s 24 June 2026 notice says CoreWeave is taking colocation at Conapto. That is a contract sentence. It does not say whether the racks will train models, serve inference, or do both, and it does not add a new campus megawatt. atNorth’s SWE01 page gives 12 megawatts of site power and a design PUE of 1.2 with liquid cooling. It does not label SWE01 as training or inference. SWE02 is written as 30 megawatts announced. SWE04 is written as 300 megawatts announced on 23 February 2026, with a separate 28 May 2026 municipal report of a 200 megawatt E.ON connection agreement. Announced site power and a reported connection agreement are still not a training-versus-inference split.

What named primary sources show

IEA Energy and AI (landing page, 10 April 2025) is the report row. The Commission energy-performance page quotes the 415 terawatt-hour and 945 terawatt-hour figures from that work. The demand chapter shows that training and deployment take place mainly in data centres, that servers can be around 60 percent of a modern hall’s electricity, that cooling can run from about 7 percent to more than 30 percent, and that accelerated servers grow faster than conventional servers in the Base Case. Those are global and regional sentences.

EcoDataCenter DeepL and DeepL’s 11 June 2025 notice show GB200 systems and an occupancy date. EcoDataCenter Mistral (11 February 2026) shows an AI-focused Sweden sentence without a Borlänge megawatt. EcoDataCenter’s 24 September 2025 Borlänge notice shows 250 megawatts of access and 600 megawatts of potential. evroc Arlandastad shows GPU count, building rights and a euro intention, not megawatts. Conapto–CoreWeave (24 June 2026) shows a colocation agreement. atNorth SWE01, SWE02 and SWE04 show site-power and announcement figures already owned by other guides. The 28 May 2026 Sollefteå notice shows a reported 200 megawatt E.ON connection agreement. Meta Luleå shows outdoor air, hydropower and PUE 1.09, not a training or inference label. None of those operator rows is a national training-versus-inference total.

A matrix instead of a workload winner

The table below is a permission table, not a ranking. A filled cell is a claim the named URL can carry. An empty cell means this page will not invent the missing figure. Training, deployment and inference stay off any row that does not use those words.

Do not invent a Swedish training megawatt from the IEA 60 percent server share, from a GPU count, or from an access figure. Those remain different objects from different years and methods.

Common misreads

The first misread is to take the IEA sentence that training and deployment happen in data centres and turn it into ”Swedish data centres are training data centres.” The IEA sentence is about where those jobs sit in the global energy system. It is not a label on Luleå, Borlänge, Arlandastad, SWE01 or a Conapto hall. The second is to treat the IEA server share of around 60 percent as a training share. That 60 percent is servers versus other hall loads such as cooling. A training job and an inference job can both sit on servers. The chapter does not split the server share by workload.

The third misread is to convert evroc’s 16,000 GPUs, DeepL’s GB200 sentence, or Mistral’s AI-focused sentence into megawatts. The parent AI and HPC guide already forbids that conversion. The fourth is to treat Meta’s outdoor-air sentence as proof that Sweden is an inference country, or Mistral’s AI-focused sentence as proof that Borlänge is a training country. One operator sentence is one operator sentence. The fifth is to add 250 megawatts, 600 megawatts, 12 megawatts, 30 megawatts, 300 megawatts and 200 megawatts into one Swedish AI-training total. Those figures live on different rows: access, potential, site power, announced power, and a reported connection agreement. This page will not add them. The sixth is to paste the February 2026 Mistral notice onto the September 2025 access figures as if they were one instrument.

What to ask next

If a brief says a Swedish hall is a training hall, ask which named URL uses the word training for that hall. If the URL only says AI-focused, GPU, GB200, colocation or outdoor air, the brief has already left the source. If a brief uses the 415 terawatt-hour or 945 terawatt-hour figures as a Swedish training total, ask for the Swedish table. The pages cited here do not have one.

Then open the parent AI and HPC page, the high-density racks page and the colocation-versus-build-to-suit page. If the operator will not name the workload, record the gap. Do not fill it from an IEA global share or from a GPU count.

1. Does the named URL say training, deployment, inference, or only AI / GPU / colocation?
2. Is the megawatt figure site power, access, potential, announced, a connection agreement, or IT load?
3. Is the IEA or Commission number global, European, or (it is not) Swedish?
4. Has anyone converted a GPU count into a training or inference megawatt?
5. Are the February 2026 Mistral notice and the September 2025 access figures still on two rows?

What this does and does not prove

This page proves that IEA Energy and AI (2025) treats training and deployment as data-centre jobs in a global and regional energy story, and that named Swedish operator pages use AI-focused, GPU, GB200 and colocation words without publishing a national training-versus-inference megawatt table. It proves that the Commission page quotes global 415 and 945 terawatt-hour figures from that IEA work.

It does not prove which Swedish hall trains models. It does not prove which Swedish hall only serves inference. It does not prove a Swedish share of the 415 terawatt-hour or 945 terawatt-hour figures. It does not treat the cited Meta Luleå page as an inference label.

Empty cells stay empty. Until a later primary source names a Swedish hall as training or inference with a dated megawatt and a boundary, the split stays a method, not a census.