How much AI capacity
does Sweden need?
Pull the levers and see how assumptions affect Sweden's total AI compute need — from public sector to the full economy.
A91–A93 via the healthcare adoption slider. Fine-tuning (Tier 3) does not affect this figure. 14-jobb.md →
~6 MW = a mid-sized data center. Facebook's Luleå started at ~40 MW.
All assumptions and calculations are open. Help us improve the analysis — especially for sectors with weak data quality. Contribute via PR on GitHub
The public-sector AI question is no longer about isolated pilots. It is about building durable capability where compute, talent, and governance scale together.
The goal is to make trade-offs explicit: what drives compute demand, which risks follow from delayed decisions, and which practical priorities matter most by 2029.
Sweden's compute need compared to Nordic and European AI investments
Why this matters
More than copilots
Agentic workflows, longer contexts, and background agents significantly increase the need. Tier 1 lands around ~2,200 H100-eq in 2029.
→ 03-berakningsmodell.mdSovereignty is a policy choice
Sovereign training accounts for ~4,500 H100-eq — half the base scenario. An active policy decision, not a consequence of user growth.
→ 08-suveranitet.mdCurrent budgets are insufficient
Existing IT budget logic supports ~2,000–4,000 H100-eq. The base scenario requires targeted state investments, EU funding, and public-private partnerships.
→ 03-berakningsmodell.mdThe full economy needs 4–5× more
Public sector is ~20% of Sweden's total AI compute need. Private sector, research, and defense add up. Energy infrastructure and grid capacity must be planned for the whole picture.
→ 11-kompletterande-perspektiv.mdWhy now?
GPU deliveries have 12–18 month lead times. Data centers require grid connections and permits. Every quarter without a decision means a quarter without capacity in 2028–2029.
→ 10-kan-vi-vanta.mdThree recommendations
Start procurement now
Framework agreements, vendor dialogue, site selection, and grid connections must start before demand peaks in 2028–2029.
Build a hybrid model
Start with ~1,000–1,500 H100-eq: on-prem for sensitive data, cloud for burst.
Couple compute with competence
Compute without governance yields low impact. Bundle with accountability, training, and data policy.
Jobs created
AI investments create new roles — not just compute costs
Clinical informatics, MLOps, AI safety & compliance, product owners, change management
System integrators, domain consultants, trainers, independent auditing, vendor support
Gross jobs — net impact depends on how fast administrative tasks are automated and how workforce transition works in practice.
→ 14-jobb.mdAll assumptions and calculations are open. Help us improve the analysis — especially for sectors with weak data quality.
Contribute via PR on GitHubWhat could go wrong?
What if the money doesn't come?
Without targeted investment, capacity stalls at ~2,000–4,000 GPUs — a fraction of what's needed. Sweden falls behind countries investing now.
Hospitals wanting AI diagnostics become dependent on expensive cloud services — or wait.
Processing times stay long. AI tools that could shorten waits in healthcare and government are delayed.
What if AI spreads slower than expected?
The compute investment risks standing unused. But it's a more manageable risk — capacity can be leased out.
AI pilots that work in labs take longer to reach routine care.
The change isn't felt yet. But the world around us keeps moving.
What if we can't buy GPUs in time?
GPU deliveries have 12–18 month lead times. Every quarter without an order means a quarter without capacity in 2028–2029.
AI tools exist but can't run locally — sensitive patient data must be sent abroad.
Sweden has the technology but not the infrastructure. Like having electric cars but no charging stations.
What if the power grid isn't enough?
Data centers need grid connections. Lead times for new connections in Sweden: 2–5 years. That's longer than GPU lead times.
Regional data centers can't expand without sufficient power capacity.
AI infrastructure competes for the same grid as homes and industry. Planning must happen now.
Sources & methodology
The analysis builds on 13 open documents with numbered assumptions (A1–A90), three triangulation tracks, and transparent derivations.
Triangulation
Each key figure is illuminated from at least two tracks: bottom-up, top-down, and big-company triangulation.
→ 01-ramverk.md