The Future of War, One Year at a Time
A scrolling timeline from the theory of why war happens, through the asymmetries that matter today, to 2036, 2046 and the world after AGI. Every stop is sourced or labeled a projection. The key stops are simulations you can run.
What asymmetries in warfare are currently overlooked?
What will war and deterrence look like in 10 and 20 years? How will robotics and large models shape that future?
Two answers and five signposts
Q1. The overlooked asymmetries today are not exotic. They are cost-exchange and stockpile depth, production capacity, supply-chain chokepoints, space, and the pairing of cyber with machine-speed intelligence. Each one rewards the side that can build, replace and adapt faster, not the side with the best single system. In 2026 that is a different list of countries than the one most defense planning assumes.
Q2. A unified world is unlikely. The likely future is two or three rival powers with separate technology stacks, each able to deny the other. Robotics lowers the human cost of fighting, which shrinks the room for peaceful deals. Large models compress decision time and hide capability in software, which widens the information gap and sharpens the commitment problem. Deterrence survives, but it is measured in factory output, orbital restraint and the speed of machines rather than warhead counts alone.
Signposts that show which way the world is heading:
- AI verification deals. Any agreement that lets rivals inspect each other's compute or models would show the information gap can be narrowed. None exists yet.
- Export-control escalation. Each new round of chip or rare-earth controls, in either direction, hardens the two-stack world.
- Interceptor rebuild rates. If Patriot and THAAD deliveries stay near 20 and zero a month, stockpile depth stays the binding constraint on US deterrence.
- Anti-satellite tests. A new debris-producing test would signal that at least one power no longer accepts mutual vulnerability in orbit.
- Major AI incidents. A documented case of an AI-recommended strike going wrong would force the first rules on human review time.
How this page was built
Stations dated 2026 or earlier are facts with at least one source, or the team's own analysis marked as such. Stations dated 2036 or later are projections: arguments about where the sourced trends lead, not predictions. The dashboard's 1995 and 2026 values are our characterization of the record; everything after 2026 is a projection too.
The simulations are simplified teaching models. Constants marked illustrative are chosen so the mechanism is visible in a browser, not measured from data. Each simulation lists its assumptions under the controls.