Failure off-ramps
Events that would cut part of the map off. A box is cut off when a hit box sits upstream of it through hard links and it is not yet built. It is slowed when it is reached only through soft links, or is already built. Counts are computed from the links.
| Off-ramp | Trigger | Cut off | Stages cut | Slowed |
|---|---|---|---|---|
| Capex pullback | Revenue lags the buildout and investors cut AI capex. | 79 | 41 of 53 | 41 |
| Taiwan chip disruption | Conflict, blockade or disaster interrupts chip production in Taiwan. | 80 | 41 of 53 | 47 |
| Power shortfall | Grids and generation cannot deliver the electricity datacentres need. | 79 | 41 of 53 | 40 |
| Capability plateau | Returns to both scaling and reinforcement learning flatten before agents become reliable. | 55 | 26 of 53 | 25 |
| Training data ruled out | Courts or lawmakers restrict training on unlicensed work, or order models retrained. | 8 | 6 of 53 | 109 |
| Incident-driven halt | A serious failure of an autonomous system leads regulators to suspend a whole category of deployment. | 1 | 1 of 53 | 13 |
| Misuse catastrophe | A major biological or cyber attack is traced to AI, and governments lock down access to capable models. | 11 | 6 of 53 | 30 |
| Alignment failure in deployment | A deployed system is caught pursuing hidden goals, and labs or governments pause frontier development. | 6 | 5 of 53 | 6 |
| Correlated AI trading shock | Many firms running similar models sell at once, a crash is amplified, and regulators curb AI trading. | 1 | 1 of 53 | 0 |
| Public backlash | Public opinion turns hard enough that politicians restrict deployment, especially where jobs are visibly lost. | 5 | 4 of 53 | 14 |