Jevons Paradox In AI

Jevons paradox in AI is the E155 argument that falling per-token cost can increase total AI consumption rather than reduce aggregate compute demand. The episode compares token efficiency to fuel efficiency: if each use becomes cheaper, people and agents may use the system more often, for more rounds, across more tasks, and inside more products.

Key Claims

• Better chips, inference architecture, routing, and engineering can lower the cost of one token.
• Lower cost can unlock more calls, longer contexts, deeper reasoning loops, more users, and more always-on agents.
• Agentic workflows create non-human token demand because software agents can call models repeatedly while planning, acting, checking, and repairing.
• Cost decline therefore does not automatically reduce demand for GPUs, power, cooling, storage, or network capacity.
• The paradox turns efficiency gains into an infrastructure-scaling problem: the system needs cheaper tokens and more total capacity at the same time.

Connections

• AI Inference Cost Structure — per-token and workflow-level cost pressure.
• MaaS Infrastructure — serving layer that must convert compute into stable token supply.
• AI Investment Metrics — token growth is a useful metric only when interpreted with cost and revenue.
• Agentic Workflow, Token-Driven Software, and AI Skills — usage patterns that can increase token demand.
• Human Resource Deflation Compute Infrastructure Inflation — aggregate shift from human labor costs to compute infrastructure demand.