The AI CEO’s Dilemma

Whether we like it or not, we’re all AI economists now: will progress hold, will the buildout tank the economy, and what will AI do to work?

deslop.media’s AI Economy Model v1 is a fictional world with rules taken from our real one. Every year, AI CEOs decide how much Gold goes to data centers and how much compute to training, hoping to earn it back over a long horizon. Smarter models open new markets only they can serve, faster frontiers let laggards catch up cheaply, and nobody knows how much work the frontier will find.

What the model lets you steer, what it holds fixed, and what it leaves out is set out in the Model Boundary. Two omissions matter most: nothing is discounted, which flatters payback, and there is no funding market, which removes both rescue funding and debt costs.

What can be learned from this game?

The analysis covers 136,400 policy-seasons, split evenly between eta 0.3 and eta 0.8; a policy-season is one strategy played through one 12-year world, not an independent economy design. Eta is the share of serving capacity left idle by refusals that counts toward training, so the two settings count 30% and 80%. The runs vary trajectory, elasticity, interaction effects, and work volume, with a baseline for comparison. Each of the three labs starts with 1,000 Gold; industry payback is measured against their combined 3,000 Gold.

Early revenue separates payback from failure

In the default eta 0.3 on-course simulations, the frontier-following strategy paid back 85.0% of the time. By year 3, cumulative industry revenue strongly separated eventual payback from failure, with an AUC of 0.91. An AUC of 0.5 is chance; 1 is perfect separation.

In the same sample, a year-3 revenue cutoff of about 2.4% of starting industry capital, or 73 Gold, separated the groups. Above it, 97.8% of simulations paid back; below it, 55.0% did. The cutoff was chosen on these same seasons and has not been tested on new ones.

Observed IQ barely separated payback from failure, with AUCs of 0.52 at year 3 and 0.56 at year 6. Average market prices through year 6 were far more informative, with an AUC of 0.92. Under the balanced strategy, 99.0% of simulations above a separately selected year-3 cutoff paid back, versus 50.8% below it.

Missing the curve narrows the path

The on-course benchmark is the median cumulative revenue among eventual payback simulations in the same year. At eta 0.3, every frontier-following simulation at or above that benchmark eventually paid back at years 3, 6, and 10.

Falling far behind cut the odds without making failure certain. More than 80% below the benchmark, 49.1% of simulations eventually paid back at year 3, 34.4% at year 6, and 32.1% at year 10.

At eta 0.8, the same deepest-shortfall group paid back 57.3% of the time at year 3, 38.0% at year 6, and 39.2% at year 10.

Less accessible work reduces payback

Price elasticity controls how quickly affordable work expands as prices fall across the model’s task tiers. Actual payments still vary by task, lab, and scarcity.

Intelligence elasticity controls how quickly addressable work expands as model capability rises. The observed IQ measure above tracks realized performance rather than either configured relationship.

Formal definitions, the finite grids, and exact denominators appear in the companion manuscript.

At eta 0.3, every tested combination through intelligence elasticity 0.89 paid back in all simulations. On the baseline price curve, payback fell to 67.8% at 0.69 and 35.4% at 0.57. No adjacent decline reached 40 percentage points, so the tested grid showed no single elasticity boundary.

At the lowest tested intelligence elasticity, price shape determined the outcome: payback ranged from 100% on the flattest tested curve to 0% on the steepest. At eta 0.8, the baseline price curve crossed the 40-point threshold between intelligence elasticity 0.69 and 0.57.

Where the Gold and unused capacity sit

All 4,000 baseline simulations paid back at the industry level at eta 0.3. The all-serving strategy also finished above its 1,000-Gold starting balance in every run. Build costs averaged 605 Gold, ranging from 192 to 995. Mean buyer headroom was 255 Gold.

At eta 0.3, every modeled task had at least one capable lab. Of work that still went unmet, restrictions accounted for 39.1% in the baseline and 14.8% in the on-course simulations; exhausted capacity accounted for the rest. Separately, refusals left 24.8% of baseline serving capacity and 23.3% of on-course capacity idle.

In these simulations, early revenue separated investments that paid back from those that failed; observed IQ barely did. Falling far behind narrowed the path, and price shape mattered most when capability growth was scarce.