Position: The Machine Learning Community Must Treat Compute Inequality as a First-Class Research Problem

This is a position paper. We argue that compute inequality—systematic disparities in who can ac- cess modern machine-learning compute and at what cost—should be treated as a first-class re- search problem by the ML community. Training compute for state-of-the-art models has grown dramatically, while the practical ability to run large experiments remains concentrated in a small set of well-resourced labs and regions. This con- centration shapes what questions get asked, what results can be reproduced, and who gets to partici- pate in setting research agendas. We propose that conferences, funders, and model developers adopt concrete norms: low-compute benchmark tracks, mandatory lightweight baselines, and standard- ized reporting of compute and energy. We also address the common view that cheaper hardware or ad hoc cloud credits will resolve the problem on its own, and explain why that expectation is incomplete.