Abstract

The latest AI summer brought us large language models (LLMs)—generative powerhouses with surprising reasoning, coding, and text generation abilities. Unsurprisingly, the database community started applying them to every aspect of data processing, including data compression. However, these methods come with painfully low inference speed—and recent papers have conveniently left out runtime considerations. In this paper, we address this gap and ask: When will LLM-based compression methods become economically viable? To answer this, we first benchmark existing approaches and optimize them for speed. We then introduce a cost model showing that, given current compute and storage costs, LLM-based compression would take 10 years to amortize its computational expense. Compared to traditional compression methods, LLM-based compression would take 120 years to break even. Finally, we project future trends and speculate that LLM-based compression could become economically viable within the next decade with anticipated improvements in hardware and model efficiency.