An agent that poses existential-level risk needs compute to operate. Compute costs money. Someone has to pay. This calculator traces the economic chain that would need to hold for a dangerous autonomous agent to actually run, and shows what breaks at each link. Companion to the Half-Life Tax model.
These describe the current frontier. Adjust to explore different capability levels.
The Weibull model makes compute costs gentler than exponential, but verification cost takes over as the dominant expense. Under realistic review fractions, human review accounts for the majority of total agent cost.
Total cost (compute + verification) as a ratio to human cost. Green (<1×) = viable.
The minimum T50 below which even free compute cannot make an agent viable, because review time alone exceeds the cost of a human doing the work.
| Task | At 10% review | At 20% review | At 30% review |
|---|
For a dangerous autonomous agent to operate, all four conditions must hold simultaneously. Each is evaluated at the current parameter settings.
The T50 doubling trend is funded by capital investment. The telecom precedent (1996–2001) suggests that infrastructure overbuilding based on exponential growth projections ends with a capex collapse, fire-sale pricing on excess capacity, and a deep stall in new buildout. Starting from the T50 set above, with κ fixed at 0.70.
| Milestone | Baseline | Post-bust | Delay |
|---|
Does the pair survive the bust? Under the telecom-precedent default of fire-sale pricing, margins become absurd. The bust kills growth potential without killing the pair.
| Inference cost | Cost/delivered | Market price | Margin | Survives? |
|---|
Can multiple pairs coordinate to exceed the individual T50 horizon? No. When κ < 1 the hazard rate declines over time, so a surviving agent enters a progressively safer state. Decomposition throws this away: each fresh agent restarts from peak hazard.
| Strategy | P(160h plan) | vs single attempt |
|---|