Parallel futures
Six futures for AI, power and the economy, and the odds of each as of 2026-10-22. The odds come from a scenario tree moved by dated signposts (programs/future-odds); they are a described model, scored over time by the calls below.
The odds
TechnofeudalismAI as Landlord · A few platforms own the models, the compute and the data; everyone else pays rent to use them.
39%
hyperscaler-compute +0.14 · labour-share +0.14
The Cooperative MarketAI as Navigator · Firms still compete, but shared models and priced externalities pull them toward the global optimum.
17%
hyperscaler-compute −0.07 · labour-share −0.07
FractureAI as Terminator · Blocs race for AI and the energy and minerals under it; capability is built for advantage before safety.
14%
us-china-dialogue −0.05 · ai-autonomy +0.03
Gradual DisempowermentAI as Governor · No one seizes power; it drains away to systems that no longer need people's work, votes or attention.
13%
labour-share +0.11 · ai-autonomy +0.08
The Managed OrderAI as Clerk · Governments license and pace frontier AI, and agree limits with each other.
10%
us-china-dialogue +0.06 · open-model-gap −0.03
The CommonsAI as Gardener · AI becomes public infrastructure, and its gains are widely held.
7%
labour-share −0.15 · hyperscaler-compute −0.12
Paths
What each future looks like
| The Commons | The Cooperative Market | The Managed Order | Technofeudalism | Fracture | Gradual Disempowerment | |
|---|---|---|---|---|---|---|
| The economy | Gains spread through cheap open models, public compute and data dividends; labour's share stops falling. | Markets stay competitive, but data and models interoperate and costs like emissions and grid upgrades are priced in. | Slower growth at the frontier; licensed national champions; public investment in compute. | Value flows to whoever owns cloud capital; wages and labour's share keep falling; a rent on every transaction. | War economies; supply chains split into blocs; capital flows to defence and chips. | Firms and states run on AI that no longer needs most people's work; income comes loose from contribution. |
| Politics | Citizens can see and question the models that shape policy; decisions carry their evidence. | Regulators use shared models to price harms; antitrust keeps markets contestable. | Strong agencies license large training runs; expertise concentrates in government. | Platforms out-lobby and out-know governments; rules are written by those they regulate. | Security states; dissent framed as risk; labs drafted into national programmes. | Decisions made by systems optimising metrics; democracy in form, little influence in fact. |
| Geopolitics | Science stays open across borders; safety research is shared; the most dangerous capabilities fall under arms control. | Competition with guardrails: trade continues, incident hotlines work, standards converge. | A verified US–China framework limits the race; others join it or are bound by it. | Two digital empires, each with client states that rent their AI. | An arms race in AI, chips and minerals; chokepoints such as Taiwan, Hormuz and rare-earth refining become flashpoints. | States compete through automated systems no one fully understands. |
| Energy and resources | Grids and compute are planned as public goods; towns that host data centres own a share of them. | Flexible data centres use the grid's idle hours; capacity is built where it is cheapest to connect. | Planned build-out tied to licences; strategic compute gets priority on the grid. | Private grids behind the meter; households carry the transmission and capacity bill. | Energy and minerals secured by force or alliance; prices volatile. | Energy goes to compute first, because compute earns the most. |
| Who holds the frontier | Many labs and open projects within months of each other; no single owner. | A handful of leading labs plus a strong open tier a few months behind. | A few licensed labs under oversight; open models capped below a threshold. | Three to five firms hold most of the world's AI compute and the channels that reach users. | National champions behind classification; safety traded for speed. | Who owns the frontier matters less each year; the systems set the pace. |
| Daily life | Expert help becomes a utility: a doctor, a lawyer and an analyst in every pocket, free at the point of use. | Better services from rival firms that must interoperate; switching is easy. | Safer, slower, more paperwork; AI arrives through approved channels. | Convenient and personal, paid for in data and attention, with little bargaining power. | Rationing, surveillance and the risk of war; AI met mostly as a weapon or a censor. | Comfortable for many, with less and less say in what happens. |
Signposts
| Signpost | Reading | Direction | Moves the odds (likelihood ratios) | Source |
|---|---|---|---|---|
| Hyperscalers' share of world AI compute Up from 62.5% in Q1 2024; Google, Microsoft, Amazon, Meta and Oracle. | 70.7 % (Q4 2025) | rising | technofeudal:1.4 · commons:0.75 · cooperative:0.85 | epoch.ai · 2026-04-14 |
| How far open models trail the frontier Epoch's capability index; earlier estimates were about 3 months. | 4 months | narrow | commons:1.3 · cooperative:1.2 · technofeudal:0.85 · managed:0.9 | epoch.ai · 2026-05-29 |
| US labour share of income The lowest on record; 65.5% in 1947, 54.1% at the end of 2025. | 52.8 % (Q2 2026) | falling | technofeudal:1.4 · drift:1.3 · commons:0.7 · cooperative:0.85 | www.bls.gov · 2026-09-15 |
| Top 1% share of US household wealth 30.9% at the end of 2024. | 32.5 % (Q2 2026) | rising | technofeudal:1.3 · commons:0.8 | fred.stlouisfed.org · 2026-09-19 |
| AI capital spending by the four largest hyperscalers About 77% more than in 2025. | 725 $bn (2026 guidance) | rising | technofeudal:1.2 · drift:1.1 · fracture:1.05 | www.lpl.com · 2026-07-31 |
| US–China AI dialogue The USA-China AI Dialogues formalised on 21 Sep 2026; a US incident hotline proposed; chip controls largely unchanged at the September summit. | formalised (Sep 2026) | opening | managed:1.3 · fracture:0.8 | www.aa.com.tr · 2026-09-21 |
| US chip export controls on China Case-by-case licences for H200 since January 2026; some easing after the May summit. | largely in place (Sep 2026) | steady | fracture:1.1 · managed:1.05 · cooperative:0.95 | www.koreatimes.co.kr · 2026-09-27 |
| Antitrust against platform and AI giants No Chrome divestiture (Sep 2025), no ad-exchange sale (Sep 2026), Meta won the FTC case; the EU ordered WhatsApp opened to rival AI assistants (Jun 2026). | behavioural remedies (2025–26) | weak | technofeudal:1.15 · cooperative:0.9 | ppc.land · 2026-09-02 |
| Public and sovereign compute US NAIRR operations centre funded ($35M over five years); the EU's call for up to seven AI gigafactories closes 12 Nov 2026. | small (2026) | growing slowly | commons:1.15 · managed:1.1 | eurohpc-ju.europa.eu · 2026-09-01 |
| Local opposition to data centres At least 75 projects in Q1 and 45 more ($68bn) in Q2; moratorium proposals in 14 states. | $130bn blocked or delayed (Q1 2026) | rising | cooperative:1.1 · commons:1.1 · technofeudal:0.95 | datacenterwatch.org · 2026-07-15 |
| How long AI agents can work on their own METR: the horizon has doubled about every 3–4 months since 2024; measurements above 16 hours are unreliable. | ≥16 hours (50% task horizon) | rising fast | drift:1.3 · fracture:1.1 · managed:1.05 | metr.org · 2026-09-08 |
| Binding AI rules EU general-purpose model duties since Aug 2025, high-risk duties delayed to Dec 2027; a US push to pre-empt state laws; California SB 53 in force. | partial (2026) | mixed | managed:1.05 · technofeudal:1.05 | www.hunton.com · 2026-07-27 |
| PJM capacity price, the largest US grid 6,831 MW short of the reliability requirement; data centres drive most of the new demand. | 325 $/MW-day (2028/29, at the cap) | high | technofeudal:1.1 · fracture:1.05 · cooperative:0.95 | www.pjm.com · 2026-07-14 |
Calls on the record
| Call | Our odds | Resolves by | Result | Source |
|---|---|---|---|---|
| Open models trail the frontier by 3 months or less on Epoch's index at any point in 2027 | 40% | 2027-12-31 | open | epoch.ai · |
| Epoch puts the hyperscalers' share of world AI compute above 75% for Q4 2026 | 60% | 2027-06-30 | open | epoch.ai · |
| The US labour share (BLS, nonfarm business) is below 52.5% in any quarter of 2027 | 45% | 2028-03-31 | open | www.bls.gov · |
| A US–China AI incident hotline is announced as operational | 30% | 2027-12-31 | open | www.aa.com.tr · |
| A court orders a major AI or cloud company to divest a business | 12% | 2028-12-31 | open | www.mlex.com · |
| A US state enacts a statewide moratorium on new large data centres | 25% | 2027-12-31 | open | datacenterwatch.org · |
| A published evaluation puts a frontier model's 50% task horizon above 40 hours | 50% | 2027-12-31 | open | metr.org · |
parallel-futures · 6 futures · 13 signposts · 7 calls · every figure is read from data/*.csv