The Economics of Open-Weight Inference

A study on the economics of open-weight inference examines how open-weight demand might support the useful life of NVIDIA GPU families. The paper presents forward term-price retention data for various GPU families, including A100 SXM4, H100 SXM, H200, B200, and B300, over different periods (6M, 1Y, 3Y, 5Y). However, the study notes that rental observations do not isolate the contribution of open-weight demand and does not definitively link it to A100 occupancy or rental-price behavior, acknowledging alternative explanations like device retirements and scarcity effects.

This paper uniquely presents forward term-price retention data for NVIDIA GPU families, unlike other studies that often focus solely on current market prices or performance metrics.