Possible Futures for Cloud Cost Models

TL;DR

The paper analyzes how cloud cost models oriented toward AI/ML create barriers for scientific workloads. It surveys historical cost models and current challenges (chip shortages, resource pooling, elasticity, measured service). It proposes future directions including micro-commitments, resale of unused capacity, predictive scheduling, time transparency, and other hybrid models to better align cloud economics with scientific discovery. The work highlights need for coordinated policies and tool support to maintain scientific access to cloud resources while preserving vendor profitability.

Abstract

Cloud is now the leading software and computing hardware innovator, and is changing the landscape of compute to one that is optimized for artificial intelligence and machine learning (AI/ML). Computing innovation was initially driven to meet the needs of scientific computing. As industry and consumer usage of computing proliferated, there was a shift to satisfy a multipolar customer base. Demand for AI/ML now dominates modern computing and innovation has centralized on cloud. As a result, cost and resource models designed to serve AI/ML use cases are not currently well suited for science. If resource contention resulting from a unipole consumer makes access to contended resources harder for scientific users, a likely future is running scientific workloads where they were not intended. In this article, we discuss the past, current, and possible futures of cloud cost models for the continued support of discovery and science.