Future scenarios

AI Scenarios 2026–2040

🎧
Listen to this essay
English voice via Kokoro (bf_emma) · approx 2 min

Three scenarios — Acceleration, Friction and Divergence — year by year from today to 2040. Not forecasts and not value judgements, but consistent stories of how today's uncertainties may combine. Based on the report Sweden's AI Transition (May 2026).

AI-skiftet · Rolf Skogling · Last updated June 2026

Three paths diverge from the same present: Acceleration, Friction and Divergence.
Three scenarios from the same present — Acceleration, Friction and Divergence — tested against the same variables.
0%
Swedish firms (10+ empl.) using AI in 2025 · SCB
0×/yr
Growth in training compute for frontier models since 2020 · Epoch
0%
US share of global AI venture capital 2025 — EU27: 6% · OECD
0 bn USD
Build cost of a 1 GW AI data centre · Epoch
How to read this. Each year shows how the three scenarios might look — Acceleration, Friction and Divergence. They are not forecasts and not value judgements: Acceleration is not “success”, Friction not “failure”, Divergence not merely a bleak sidetrack. Each scenario holds both opportunities and costs, and is tested against the same underlying variables — capability, adoption, capital, energy, geopolitics, Swedish response and EU coordination. Reality will likely blend elements of all three. Click a card to expand it.
2026
Baseline: documented signals
Acceleration
Short lead times — speed is the question
  • Frontier models improve fast on digital tasks (HLE, OSWorld, SWE-bench); training compute has grown ~5×/year since 2020.
  • 35.0% of Swedish firms (10+ employees) use AI — 71.9% among large, 30.8% among small (SCB).
  • AI firms account for 61% of global venture capital in 2025 (OECD); large players secure compute and models.
  • Read in Acceleration: this is the start of a longer concentration phase, not the peak of a hype cycle.
Click for details ↓
Friction
More time — but a passivity risk
  • The same model gains, but benchmarks saturate fast and invalid test items occur (AI Index).
  • McKinsey: nearly two-thirds have not yet scaled AI across the whole company.
  • Goldman/Walker: no clear productivity effect is yet visible at the economy level.
  • Read in Friction: 2026 is a starting point where the gap between demo and production is wide.
Click for details ↓
Divergence
Uneven access and control
  • Large firms use AI far more than small ones; lack of expertise is the most common reason to abstain (SCB).
  • Capital and GPU clusters are concentrated: US ~75% of AI venture capital, EU27 6% (OECD); US ~¾ of GPU performance (Epoch).
  • Read in Divergence: the unevenness of 2026 is not transitional but an emerging structural pattern.
Click for details ↓
2027
Early divergence
Acceleration
Short lead times — speed is the question
  • The time horizon of agent tasks grows (Hinton: doubling ~every seven months) — tasks taking a working day start to be handled.
  • Firms rebuild workflows around agents in code, support, analysis and case handling.
  • Early labour-market signals in exposed occupations — entry jobs and recruitment — before total unemployment moves.
  • A political asymmetry: firms adapt fast, the state's lead times are longer.
Click for details ↓
Friction
More time — but a passivity risk
  • Robust operation requires control, documentation and human review — the demo→production gap persists.
  • Broad but shallow use; local gains don't show up across the whole P&L.
  • The EU begins to mature standards and supervision; Sweden gains time to sequence.
Click for details ↓
Divergence
Uneven access and control
  • Large and digitally mature players pull ahead; smaller firms lack specialists and procurement capability.
  • The public sector splits: some agencies build capability, others get stuck in pilots.
  • Dependence on external clouds and models starts to show in price and terms.
Click for details ↓
2028
The threshold question
Acceleration
Short lead times — speed is the question
  • A possible qualitative threshold: Legg's 50/50 for “minimal AGI” falls here — general agent capability in digital environments.
  • Large parts of administrative and knowledge work get new cost and production conditions.
  • On early threshold passage (2028–2031), normal-tempo lead times don't suffice — rapid sequencing comes into play.
Click for details ↓
Friction
More time — but a passivity risk
  • AGI claims may prove premature; episodic memory, continual learning and robust world models remain unsolved.
  • Adoption broadens but stays uneven; a skills shortage is a real brake.
  • More time to sequence reforms — but only if the time is actually used.
Click for details ↓
Divergence
Uneven access and control
  • If the first general systems are controlled by a few players in the US/China, European dependence becomes structural.
  • The EU can regulate the systems but not produce them on relevant timescales.
  • The gap between those with and without access to the best capacity becomes a new line of division.
Click for details ↓
2029
The middle phase begins
Acceleration
Short lead times — speed is the question
  • The agent horizon approaches quarter-length lead times; autonomous and semi-autonomous flows in more functions.
  • Tax base and social insurance are pressured by faster service transformation than the support systems can follow.
  • Risk that decisions arrive in the wrong order if indicators don't sit early in the chain.
Click for details ↓
Friction
More time — but a passivity risk
  • The productivity gap between firms with and without capability can become long-lasting.
  • Physical occupations (care, construction, eldercare) are affected more slowly — robotics lags machine cognition.
  • A reasonable stance: reversible steps where possible, early investment where lead times are long (education, the grid).
Click for details ↓
Divergence
Uneven access and control
  • Data centres are built where capital, power, land and permits can be assembled; other regions fall behind.
  • Use and sovereignty drift apart: high AI benefit, low influence over the technology.
  • EU coordination can reduce dependence — but only if regulation is matched by actual capacity.
Click for details ↓
2030
The sequencing test
Acceleration
Short lead times — speed is the question
  • AI may add significant GDP (Goldman) — but institutional stress remains if reforms weren't built ahead of the pressure.
  • Data centres' electricity demand is large (IEA); big loads compete with industry and households for power.
  • Proactive adaptation (reforms 2026–2030) yields a steerable position; otherwise the transition turns reactive.
Click for details ↓
Friction
More time — but a passivity risk
  • McKinsey projection: ~30% of work hours automatable — but the outcome depends on adoption depth.
  • A slower course conceals that lead times in education, infrastructure and government are still long.
  • The main risk is passivity and misread security, not overload.
Click for details ↓
Divergence
Uneven access and control
  • Gains concentrate in capital, data and competence; costs land on smaller players and exposed occupations.
  • Different service levels in the public sector become an equality issue, not just competitiveness.
  • The price of dependence: pricing, security and continuity are decided in markets Sweden does not control.
Click for details ↓
2035
Institutionalisation — or catching up
Acceleration
Short lead times — speed is the question
  • On late threshold passage (2032–2037) the middle phase becomes decisive; reforms must be accelerated.
  • If readiness was built early, the conflict is over adjustments — not late crisis management.
  • Rapid sequencing's reversible instruments can be activated within 30–90 days when indicators tip.
Click for details ↓
Friction
More time — but a passivity risk
  • Reforms can be well sequenced if the time was used; otherwise study places and support are built in catch-up mode.
  • Reactive adaptation (2030–2035) is manageable but costlier — and easily perceived as punishment or panic.
  • EU standards and Nordic energy coordination can become Swedish assets if made operational.
Click for details ↓
Divergence
Uneven access and control
  • A two-speed pattern between players and regions may have set in.
  • Strategic dependence becomes institutional and democratic, not merely economic.
  • The realistic answer is qualified receiving capacity — test environments, a data resource, compute, procurement skill — not full self-sufficiency.
Click for details ↓
2040
Three paths, four outcomes
Acceleration
Short lead times — speed is the question
  • Proactive adaptation: indicators, transition support, financing readiness and transparency in place before the pressure — steerable, not friction-free.
  • On early threshold passage without rapid sequencing: discontinuity — the right reforms for normal tempo meet a tempo that isn't normal; proactive becomes reactive in two years instead of five.
  • The question is not whether Sweden can guarantee a good 2040, but whether 2026–2030 built institutions that make 2040 less dependent on luck.
Click for details ↓
Friction
More time — but a passivity risk
  • Reactive adaptation: you get there, but with less trust and a higher price.
  • Reforms work differently well depending on whether they took effect before the pressure set in.
  • Outcomes aren't binary — Sweden can be proactive in one track and delayed in another at the same time.
Click for details ↓
Divergence
Uneven access and control
  • Delayed adaptation: several long-lead systems are repaired simultaneously under pressure; some groups feel they bore the cost.
  • The word crisis should be used precisely — it means lost time and uneven distribution, not necessarily societal collapse.
  • The annual review should ask which reform cluster is moving which way — not just which scenario the country is in.
Click for details ↓

Cross-cutting themes

Labour market & entry paths
Exposure is uneven; younger workers in exposed occupations may be affected before total unemployment moves. Measure earlier than after-the-fact metrics.
Tax base & financing
If productivity and profit concentrate in capital and models while costs land on the wage base, an imbalance arises that must be addressed.
Energy & data centres
AI is software on the surface but physical infrastructure beneath it. Power, grid, land and build rate become strategic scenario variables.
Democratic institutions
New attack surfaces: deepfakes near elections, opaque AI decision support in administration, gradual function creep in surveillance.
EU coordination
EU law provides a frame, time and standards — but does not replace Swedish capacity. Regulation can coexist with strategic dependence.
Capacity threshold & readiness
The threshold (Hassabis, Hinton, Legg) sets the tempo. Rapid sequencing is reversible, indicator-driven readiness for early passage — not panic decisions.

About the scenarios

The framework rests on three symmetrically constructed scenarios — Acceleration, Friction and Divergence — each tested against the same seven variables (capability curve, adoption rate, investment durability, energy and data centres, geopolitical concentration, Swedish response and EU coordination). They should not be read as value judgements or as a guess about which future “wins”, but as three tests of the same reform package: a reform that works in only one scenario is too fragile.

The timeline projects each scenario's trajectory 2026–2040 from the report's logic and documented signals. The analysis distinguishes the documented (established measurements and expert assessments) from the forecast (a guess about outcomes) — and does not try to predict which outcome occurs, but to build readiness that holds across several possible courses. The May addendum adds an eighth variable, the capacity threshold (Hassabis, Hinton, Legg), and rapid sequencing as a readiness track for early threshold passage.

All data points and positions derive from published sources — including SCB, Epoch AI, OECD, IEA, McKinsey and the IMF. The basis is the report Sweden’s AI Transition (May 2026) and is revised as the source base ages. See Sources.