Asymptotically-Free Inference
The cost of running a model trends toward zero — so the edge moves to whatever stays scarce when inference is free.
Cost per unit of intelligence is falling fast — a top-tier answer costs a small fraction of what it did a couple of years ago, and the curve keeps bending down the way transistors and bandwidth did before it. Yet total inference spend is exploding, not shrinking. That’s Jevons paradox: when the unit gets cheap, you use vastly more of it. Agentic systems make the point vivid — they burn far more tokens per task than a single prompt ever did.
Why it matters: don’t build a business that assumes inference stays expensive. Assume it approaches free, and ask what becomes scarce in that world — proof-of-work, verification, trust, attention.