The Token Economy

Compute power has become a first-order economic variable. This working paper starts from a striking fact: the cost of processing a token —the elementary unit of information that AI models consume and produce— has fallen roughly 280-fold in two years, even as the compute used to train frontier models grows almost fivefold per year and the largest cloud providers invest more than half a trillion dollars annually in data centers.

From there it derives an 'automate-or-hire' rule that treats compute as a production factor. Its central finding is counterintuitive: the cost of automating a task can rise above the cost of human labor even as token prices fall, because agentic systems consume substantially more tokens per task. The paper connects token economics to macroeconomic automation frameworks and examines adoption evidence from 2025-2026.

Read the full paper on SSRN.