A Swedish national AI facility: cost and realistic scope
FOUR IMPORTANT CLARIFICATIONS BEFORE PUBLICATION
FOUR IMPORTANT CLARIFICATIONS BEFORE PUBLICATION
1. The electricity use should be stated somewhat more precisely.
- 40–50 MW running through all 8,760 hours of the year corresponds to 0.35–0.44 TWh.
- 200–230 MW corresponds to 1.75–2.01 TWh.
- The earlier ranges of 0.3–0.4 and 1.6–2.0 TWh assume that the facility does not always run at full power.
2. The EU support applies to at most 17 percent of the relevant IT investment — not 17 percent of the whole project cost.
For a full-scale facility the terms are:
- Phase 1: at most 200 million euros from the EU, roughly 2.2 billion kronor, which must be matched nationally.
- Phase 2: a further maximum of 800 million euros, roughly 8.9 billion kronor, conditional on the EU's coming long-term budget.
- Total EU ceiling: one billion euros, roughly 11.1 billion kronor, but never more than 17 percent of the relevant IT capital expenditure.
3. The cost of hardware renewal is a calculation, not a fact.
The EU requires a five-year total cost calculation covering future hardware replacement or decommissioning. But the EU does not state that all hardware must be replaced after exactly three to five years. The wording should therefore be "planning assumption", not a settled fact.
4. Google does not mention 2028 in its official press release.
Google confirms construction start on 2 June 2026, but publishes neither an investment amount nor an opening year. If you wish to stick to primary sources, "planned operation 2028" should be struck.
BASIS FOR THE CALCULATIONS
100,000 processors
EuroHPC defines a full-scale AI gigafactory as at least 100,000 advanced AI accelerators, counted as H100 equivalents. Phase 1 must cover at least 40,000.
The power requirement of 200–230 MW
Nvidia's documentation states:
- 72 GPUs per GB200 NVL72 rack.
- Roughly 120 kW per rack.
- Liquid cooling is part of the design.
The calculation runs:
- 100,000 divided by 72 = roughly 1,389 racks.
- 1,389 times 120 kW = roughly 167 MW for the compute racks alone.
- Networking, storage, cooling and other infrastructure push the total to an estimated 200–230 MW.
For 20,000 processors the corresponding pure rack power is roughly 33 MW, and the estimated total power requirement is 40–50 MW.
Stargate Norway confirms the order of magnitude: 100,000 Nvidia GPUs, 230 MW initially, and the possibility of a further 290 MW — in total 520 MW.
The investment range of 75–125 billion kronor
This is a calculated range, supported above all by the following comparisons:
- EuroHPC estimates that an AI gigafactory requires investment of at least 4–5 billion euros, roughly 44–56 billion kronor. This is a minimum indicative level, not a fixed price.
- Stargate Norway's partners have set aside roughly one billion dollars for the first stage of only 20 MW.
- Microsoft is investing 33.7 billion kronor in Swedish cloud and AI infrastructure, installing 20,000 advanced GPUs.
- Brookfield plans investments of up to 95 billion kronor in Strängnäs, expanding from 300 to 750 MW over 10–15 years. (Brookfield's full press release.)
The range of 75–125 billion is therefore a defensible planning interval, but not a public price quotation.
The cost breakdown
The amounts for servers, networking, data halls, cooling and security are our own calculation items. EuroHPC does, however, confirm which cost categories must be included:
- GPUs, CPUs and memory.
- High-performance storage.
- Internal networks and switches.
- Racks and rack-level power supply.
- Cooling systems.
- System software, cloud platform and licences.
The electricity
The electricity cost of 0.8–1.35 billion kronor is plain multiplication:
- 1.8 billion kWh times 0.45 kronor = 810 million kronor.
- 1.8 billion kWh times 0.75 kronor = 1.35 billion kronor.
The price of 45–75 öre per kWh is a planning assumption. Statistics Sweden (SCB) reports trading price, grid price and total price for Swedish business customers, but a data centre of this size would most likely negotiate its own electricity contracts and grid terms.
SOURCES FOR THE SWEDISH PROJECTS
- Mimer: total budget 29.76 million euros, half from EuroHPC and half from the Swedish Research Council. This corresponds to roughly 330 million kronor at the exchange rate used.
- Evroc, Arlandastad: planned investment of 600 million euros, 18 MW, and room for up to 16,000 GPUs.
- Meta, Luleå: more than 8.7 billion kronor invested and over 300 full-time employees.
- Mistral AI and EcoDataCenter: roughly 12.8 billion kronor in Swedish digital infrastructure, with large-scale AI capacity in Borlänge.
- Microsoft: 33.7 billion kronor and 20,000 advanced GPUs in Sandviken, Gävle and Staffanstorp.
- AWS: more than 39 billion kronor since 2017 in the AWS region with data centres in Eskilstuna, Katrineholm and Västerås.
- Brookfield, Strängnäs: up to 95 billion kronor; planned expansion from 300 to 750 MW. (Brookfield.)
- Google, Horndal: construction start 2 June 2026, 100 planned full-time jobs and thousands of jobs during construction. No total investment amount is stated.
THE ELECTRICITY SYSTEM IN NORTHERN SWEDEN
Svenska kraftnät (the Swedish national grid operator) confirms:
- Sharply rising demand in Norrbotten and Västerbotten.
- A need for grid expansion, new electricity generation and flexibility alike.
- The goal of raising the capacity through Cross-section 1 from 3,300 to 5,500 MW by 2035 at the latest.
Svenska kraftnät's long-term analysis furthermore reckons that Swedish data centres' electricity use may grow from roughly 2.5 TWh in 2026 to between 8.8 and 40.9 TWh in 2050, depending on scenario.
SWEDEN'S AND THE EU'S DECISIONS
- On 30 July 2026 the government instructed the Swedish Tax Agency to take part in a joint EuroHPC procurement and to co-finance access to an AI gigafactory in Finland.
- The EU's call opened on 30 July 2026, closes on 12 November 2026, and covers up to seven AI gigafactories. Decisions are expected in early 2027, and operations must be able to begin within 18 months of contract signing.
- The initiative covers up to 10 billion euros in public financing and is expected to unlock at least 20 billion euros in private investment.
- The rule of at most 17 percent of IT capital expenditure, national matching, and the consortium's responsibility for the remaining investment and operations, is found in EU Regulation 2026/150.
- The exchange rate of 11.0885 kronor per euro on 28 August 2026 is published by the European Central Bank:
SOURCE NOTE TO INSERT IN THE DOCUMENT
The cost ranges in this text are scenario calculations, not quoted prices. They rest on EuroHPC's requirements and financing model, Nvidia's technical data for the GB200 NVL72, publicly announced investments by Microsoft, Brookfield, Evroc, Mistral AI, AWS and Meta, and on Stargate Norway. Electricity use has been calculated as power multiplied by the year's 8,760 hours. The annual total cost covers operations, electricity, and a planning reserve for future hardware replacement. All amounts are rounded to 2026 price levels.
WHAT IS ALREADY HAPPENING IN SWEDEN
The comparisons call for some caution: certain announced amounts cover processors, construction and operations, while others apply only to particular parts of the facilities.
Project: Mimer / Sweden AI Factory Announced scale or investment: 29.76 million euros, roughly 330 million kronor. What it shows: Valuable infrastructure for research and for small and medium-sized companies, but far from gigafactory scale.
Project: Evroc, Arlandastad Announced scale or investment: A plan of 600 million euros, 18 MW, and capacity for 16,000 GPUs. What it shows: A Swedish and European alternative for an independent cloud.
Project: Meta, Luleå Announced scale or investment: More than 8.7 billion kronor and over 300 permanent employees. What it shows: Northern Sweden works technically — but the capacity serves Meta.
Project: Mistral AI and EcoDataCenter, Borlänge Announced scale or investment: 12.8 billion kronor. What it shows: A European AI company building next-generation capacity on Swedish soil.
Project: Microsoft Announced scale or investment: 33.7 billion kronor and 20,000 advanced GPUs. What it shows: The closest existing comparison to a medium-sized national facility.
Project: AWS Announced scale or investment: More than 39 billion kronor since 2017 in the Swedish cloud region. What it shows: Broad cloud infrastructure in Eskilstuna, Katrineholm and Västerås.
Project: Brookfield, Strängnäs Announced scale or investment: Up to 95 billion kronor. What it shows: The strongest Swedish comparison for a genuine private AI megaproject.
Project: Google, Horndal Announced scale or investment: The investment has not been made public; construction starts in June 2026. What it shows: Google's first wholly-owned Swedish data centre, planned to enter operation in 2028.
AWS, Microsoft, Google and Meta bring valuable infrastructure to Sweden. But their machines remain the companies' property, and the capacity, the software and the priorities are controlled by those companies.
A Swedish state facility would serve a different purpose: guaranteed Swedish access when capacity is scarce, Swedish control over its allocation, and capacity for projects that are economically or strategically important even when they are not immediately profitable.
IS NORTHERN SWEDEN THE RIGHT PLACE?
Probably — but the grid connection matters more than the average price of electricity.
Northern Sweden offers:
- Historically lower wholesale electricity prices.
- Hydropower and other fossil-free generation.
- A cold climate and efficient cooling.
- Land for future expansion.
- Existing experience of industry and data centres.
- Possible connections to Arctic fibre routes.
But the standard political claim that "there is plenty of cheap electricity in the north" is becoming dangerously oversimplified. Steel, hydrogen, mining, battery manufacturing and other industrial projects are competing for the same electricity.
Svenska kraftnät states that electricity demand in Norrbotten and Västerbotten is expected to rise sharply, and that this will require grid reinforcement, new generation and greater flexibility. Current plans aim to increase transmission capacity in the north, among other things by raising the capacity through Cross-section 1 from 3,300 MW to 5,500 MW by 2035. (Svenska kraftnät.)
Sweden should therefore look first for:
- An existing industrial site with a secured grid connection of 100–250 MW.
- Two independent high-capacity fibre routes.
- Proximity to district heating networks, greenhouses or industry that can use the waste heat.
- Room for liquid cooling and later construction stages.
- Access to technical labour and transport.
- A second, smaller facility elsewhere in Sweden, for resilience.
A completely new site chosen solely because it lies in price area SE1 may wait years for its grid connection — longer than the useful life of the processors originally planned for the facility.
WHAT SWEDEN DOES TODAY
Sweden already has Mimer at Linköping University, co-financed by Sweden and EuroHPC. The system, costing around 330 million kronor, is aimed chiefly at small and medium-sized companies, start-ups and researchers. It matters — but it is an AI factory at research scale, not a national industrial public resource.
More important still: on 30 July 2026 the Swedish government instructed the Tax Agency to take part in a EuroHPC procurement and co-finance Swedish access to an AI gigafactory in Finland. The announced policy, in other words, is that Sweden buys access to Finnish capacity instead of building the main facility in Sweden. (The Swedish government.)
Your proposal is more ambitious: Sweden would keep the physical access, the jobs, the technical competence and the strategic control inside the country.
THE OPPORTUNITY WITHIN THE EU
The timing is unusually favourable. On 30 July 2026 the EU opened its call for up to seven AI gigafactories. The deadline is 12 November 2026, with selection expected in early 2027. The EU reckons that public financing will unlock at least 20 billion euros in private investment. (EuroHPC's call.)
Within the programme:
- EU support can take the form of guaranteed purchases of computing capacity.
- A participating country must contribute at least as much as the EU.
- The EU's contribution can amount to at most 17 percent of the relevant investment expenditure, with the limits set out in the tender.
- The largest projects can receive up to one billion euros in EU support in two stages, provided the country contributes a matching amount. (EuroHPC's governing decision.)
For a Swedish project of 75–125 billion kronor, this could mean:
- 5.5–11 billion kronor in Swedish guaranteed capacity purchases.
- A similar amount from the EU.
- The remaining 53–114 billion kronor raised by the operating consortium and its investors.
This is considerably easier to defend than placing the entire risk of future processor replacement directly on the Swedish state.
THE BEST OWNERSHIP MODEL
I do not recommend that a government ministry should design and operate a competitor to AWS. I recommend instead a free-standing, professionally governed infrastructure company — provisionally named Sveriges AI-kraftverk AB.
The model could look like this:
- The state owns the site, the grid connection and the strategic control.
- Sweden, the EU and private industry together finance the computing equipment.
- An experienced Swedish or European operator runs the physical facility.
- The state has guaranteed capacity for public services, universities, healthcare, defence and Swedish foundation models.
- Companies buy capacity at open and transparent prices.
- Small and medium-sized companies receive compute vouchers rather than unlimited free use.
- Citizens get access to Swedish AI services and a national basic allocation of AI capacity — not unlimited access to the raw processors.
- Hardware is procured in successive generations, so that the whole facility does not become obsolete at once.
- The software interfaces are open enough that workloads can be moved between Nvidia, AMD and future European hardware.
A hall full of processors is not, in itself, technological sovereignty. Sovereignty requires control over allocation, data, software, operators, security and the next generation of hardware.
SWEDEN AND NUCLEAR POWER
The government's road map is at least 2,500 MW by 2035, and an expansion that could correspond to roughly ten large reactors, about 10,000 MW by 2045. But the special support model so far covers only about 5,000 MW. The larger 2045 goal is thus an ambition, not a fully financed construction programme.
Scope: Videberg/Ringhals Production: 1,410 MW, roughly 12 TWh per year. Official construction estimate: Roughly 117 billion kronor by the government's standard figure. Status: Three Rolls-Royce SMRs, development phase.
Scope: The current support programme Production: 5,000 MW, roughly 39 TWh per year. Official construction estimate: 413 billion kronor. Status: Financing framework decided.
Scope: The political 2045 ambition Production: Roughly 10,000 MW, roughly 78 TWh per year. Official construction estimate: Roughly 826 billion kronor, calculated linearly. Status: Not fully decided or financed.
The government's standard figure is just under 83 million kronor per MW, or 413 billion for 5,000 MW. It is explicitly a construction cost excluding financing costs. The assumption is roughly four reactors of 1,250 MW each and seven years of construction.
The first concrete project is Videberg at Ringhals: three reactors of 470 MW each. The first reactor still aims for the mid-2030s, but the project price and the final investment decision are not yet public.
MY TOTAL ESTIMATE FOR 10,000 MW
Scenario: Official technical floor Calculation: 826 billion kronor of construction plus 122 billion for the fixed waste system. Total: roughly 0.95 trillion kronor.
Scenario: Planning scenario Calculation: 826 times 1.25 for cost growth, times 1.15 for construction financing, plus 122. Total: roughly 1.31 trillion kronor.
Scenario: Stress scenario Calculation: 826 times 1.50, times 1.25, plus the maximum waste framework of 183. Total: roughly 1.73 trillion kronor.
The government estimates the fixed costs of a new nuclear waste programme at 122 billion kronor through 2159, with a further 61 billion as a safety margin. The idea, however, is that the reactor owners shall bear these costs once the expansion is large enough; the amount is therefore not automatically a bill to the taxpayer.
My 25 and 50 percent cost overruns are scenario assumptions, not the government's forecasts. The government's own reserve in practice allows up to 100 percent cost overrun for the first financing round — which shows how large the uncertainty is judged to be.
HOW MUCH IS THE STATE RISKING?
For roughly the first half of the support programme, the state has set up:
- 220 billion kronor in state loans.
- A separate risk reserve of 220 billion for cost overruns.
- A two-way price guarantee with an expected total payout of around 110 billion over up to 40 years, but with a maximum authorisation frame of 400 billion.
- For Videberg: 60 percent direct state ownership and up to 34.3 billion in future capital injections. Since Vattenfall also owns 20 percent, public ownership is indirectly about 80 percent.
One should therefore not call 220 + 220 + 400 billion a "cost of 840 billion". It is a mixture of loans, risk reserve and price insurance — rather like adding together the price of a house, the mortgage, and the maximum payout of the fire insurance.
The loans are to be repaid, the state's shares may hold or lose value, and the price guarantee can bring the state income when the market price sits above the guaranteed price. The state's true net cost will be known only when construction prices, interest rates, the guarantee price and future electricity prices are known.
Over 60 years of operation, 10,000 MW would produce around 4,680 TWh. At the government's calculation level of 80 öre per kWh, that corresponds to roughly 3.7 trillion kronor in combined capital, operating, fuel and waste costs. But that sum is financed mainly through the sale of electricity, and should not be stacked on top of the construction cost.
OTHER GREAT SWEDISH INFRASTRUCTURE LEAPS
Area: Transport infrastructure Order of magnitude: 1,171 billion kronor. Period and comment: 2026–2037; roads, railways, maintenance and new investment.
Area: Swedish electricity grids in total Order of magnitude: 890–945 billion kronor. Period and comment: Estimated need 2021–2045; regional, local and transmission grids.
Area: Svenska kraftnät's share Order of magnitude: 215 billion kronor. Period and comment: 2026–2035; part of the larger grid picture and not to be added separately.
Area: Water and sewerage Order of magnitude: At least 560 billion kronor. Period and comment: Through 2040.
Area: LKAB's transformation Order of magnitude: Up to 400 billion kronor. Period and comment: A corporate strategy over 20–25 years, not a final investment decision for the whole amount.
Area: SSAB's new Luleå works Order of magnitude: 4.5 billion euros. Period and comment: Roughly 50 billion kronor at a rounded exchange rate.
Area: Stegra's steelworks in Boden Order of magnitude: Roughly 6.5 billion euros. Period and comment: Roughly 70 billion kronor; mainly private financing.
Taken together, nuclear power, transport, electricity grids, water and sewerage, and the largest industrial transformations point toward investments of roughly 4–5 trillion kronor through about 2045. This is not one joint state bill: the time periods differ, and the financing comes from the state, municipalities, grid fees, electricity prices and private companies.