Empirically estimate data’s contribution in AI production functions

Establish the functional form by which data enters AI production (e.g., Cobb–Douglas or CES), derive the marginal product of data and its elasticities with compute and labor, and empirically estimate these parameters using firm-level training datasets, model performance, and revenue outcomes, complemented by controlled and natural experiments when firm-level data is unavailable.

Background

The paper argues data should be modeled explicitly as a factor in production functions, distinct from capital, labor, and technology due to nonrivalry and emergent rivalry via contamination. However, it does not specify the functional form and emphasizes the need for empirical estimation of marginal products and elasticities.

Section 6 proposes combining controlled experiments, natural experiments, and evidence on firms’ data investments to isolate causal effects and quantify returns to scale and complementarities.