The Thesis
AI inference is rapidly becoming a commodity. As models proliferate and providers compete on price, the cost of a unit of AI-generated output is falling on a trajectory that resembles earlier technology deflation curves—semiconductors, bandwidth, storage. But unlike those precedents, there is no standard unit of measurement for what AI actually delivers per dollar spent. The Capability-Seconds™ Index exists to fill that gap.
The core insight comes from an unexpected place: metrology. The SI second—defined by the fixed frequency of a cesium-133 atom—is the most precisely measurable quantity in physics. In an economy increasingly dominated by AI-generated output, the second becomes a natural invariant unit of account. How much capability can a model produce per second of latency, per dollar of cost? That ratio is the CSI. It gives us a single number that tracks whether AI is getting more efficient in terms that matter to the people and businesses paying for it.
The Capability-Seconds™ Index exists to test this thesis with real data. Every day, we measure frontier models on identical tasks, record their speed, accuracy, and cost, and publish the result. If the thesis is right—if AI inference really is deflating along a measurable curve—the CSI will show it. If not, the data will show that too. Either way, the measurement comes first.
The Copernicus Problem: Time as the Invariant Denominator of Value in the Age of Artificial Intelligence
This paper argues that the SI second provides the natural invariant unit of account for pricing artificial intelligence. It introduces the Capability-Second as a composite measure and presents the CSI as an empirical test of the AI deflation thesis.
Key Concepts
The amount of useful AI output (score) delivered per second of latency. The speed dimension of the index.
The amount of useful AI output delivered per dollar of cost. The economic dimension of the index.
The composite: score divided by the product of latency and cost. A single number capturing efficiency across all three dimensions.
The observed trajectory of declining AI inference costs over time. CSI tracks whether this curve is accelerating, decelerating, or holding steady.
Experimental: Forward Curve
We also maintain an experimental forward curve projecting CSI scenarios over 3–24 months. This model is in beta and uses structural assumptions, not statistical fitting. It reflects different cost deflation rate assumptions applied to current model pricing and will transition to empirical fitting once 12+ weeks of measurement data are available.