Sweden's AI Strategy Under Uncertainty

• The government should commission an assessment with low, central and high estimates. It should cover the compute needed to train, adapt and run models; demand from universities, public authorities, defence and companies; security and data-control requirements; facilities and processors already available; the cost of renting compared with owning; and how capacity should be divided among Swedish, Nordic and European users.
• The assessment should set the numbers for targets, procurement and financing, and be updated as processors and models change. Its purpose is to speed up implementation, not to reopen the decision to prepare large-scale computing infrastructure.

Report section 3.5

• Sweden uses an estimated 120–150 MW of AI compute in 2026. With demand growing by roughly a quarter each year, this reaches 290–390 MW by 2030. Guaranteeing half of that means approximately 145–195 MW; covering 80% would mean 230–310 MW.
• Guaranteed access does not require the state to own data centres. It can come from long-term contracts with private providers, Swedish shares of European facilities, or Nordic capacity with priority rights. A facility in Sweden adds no resilience if Swedish users have no right to use it. Contracts should set the amount and performance of compute, priority during shortages, security conditions and hardware renewal.

Report sections 3.3, 3.4 and 3.7.2

• Sweden’s electricity, climate and land make it one of the best places in Europe for large data centres, and much AI computing can run far from its users. Publicly known Swedish projects add up to approximately 1.06 GW by 2030, so approximately 620 MW more is needed to reach 10% of current European plans.
• Commercial capacity should be financed privately. The state’s task is to cut the time from investment decision to a running facility. It should prepare a small number of AI Infrastructure Acceleration Zones, where grid capacity, planning, environmental assessments and fibre are in place before an investor arrives, with one public body coordinating each site. Because 1.7 GW at full use would need approximately 17 TWh of electricity a year, compute and electricity expansion must be planned together.

Report sections 3.6 to 3.9

• The EU is building AI Factories, planning up to five AI Gigafactories financed through InvestAI, and negotiating the Cloud and AI Development Act. Sweden hosts one AI Factory, Mimer in Linköping, and has agreed to co-finance a gigafactory in Finland in return for public-sector access to its computing resources. It is not public whether a Swedish site is in the running for a gigafactory; this should be established before the formal call.
• Joining every initiative would spread Swedish resources as thinly as trying to build everything at home. Public co-investment should pass a simple test: would the project be built at a similar scale and on similar terms without Swedish money? Sweden should also introduce the Act’s data-centre acceleration zones early and treat the EU framework as a minimum.

Report sections 1.3, 3.3, 3.7.2 and 3.8.2

• Frontier AI developers and hyperscalers need electricity, land, grid connections and predictable permitting. Sweden can offer these, and growing local resistance to data-centre construction in the United States may make them more valuable. In return, Sweden should seek commitments that Swedish and European firms, researchers and public institutions get the same models as American customers, without delays or weaker versions.
• Commitments should be made through contracts, procurement or co-investment, not attached to permits, and should name the users covered, the models, how quickly new versions arrive and when access may be withdrawn. No contract can stop the provider’s home government from imposing restrictions, so Sweden should also keep enough secured compute to run European or open-weight models.

Report chapter 4

• Sweden should take the initiative for a Nordic Compute Union. A joint Nordic offer of several gigawatts of future capacity and prepared sites is much harder for a provider to replace than any single country’s offer, and pooled public and research demand gives larger contracts and better terms.
• The core group should stay small, while customers can come from anywhere in Europe. Cooperation already works: the LUMI supercomputer in Finland is financed and used jointly by ten countries, including Sweden and Norway. The infrastructure should also be designed so that capacity in different countries can later be used together for AI training.

Report chapter 5

• The institute should monitor developments at the AI frontier, research emerging security risks, run targeted evaluations where Sweden needs independent evidence, advise senior decision-makers and support the government during serious AI-related incidents. It should develop a few technical specialisations, chosen after assessing Swedish needs and existing expertise; verifying claims about AI systems and the compute behind them is one candidate.
• It should not copy the UK institute’s more than 100 technical staff, but work through the International Network for Advanced AI Measurement, Evaluation and Science. It should be advisory rather than a regulator, report directly to the minister responsible for AI, and start small with a strong founding director.

Report chapter 6

• AI could raise productivity and wages, or reduce demand for cognitive work and move income from wages to profits. Sweden should stress-test transition organisations, omställningsstudiestöd, the Public Employment Service, unemployment insurance and adult education against several scenarios, and the Ministry of Finance should test how tax revenue and public spending would change.
• Sweden should also protect the junior roles through which people become experienced professionals, track underemployment, wages and career moves rather than unemployment alone, agree in advance what evidence would trigger a larger response, and use collective bargaining to combine fast AI adoption with retraining and a share of productivity gains for workers.

Report chapter 7