CUDA Poor

Adjective · Slang & Abbreviations

Definitions

  1. A joking description of an AI team or researcher without enough GPU capacity to train or run the models they want, a variant of the more common GPU-poor. The phrase took off in 2023, when a widely shared SemiAnalysis piece split the field into the GPU-rich, a handful of labs holding tens of thousands of accelerators, and the GPU-poor, everyone else. The CUDA framing points at the real bottleneck: most training and inference software targets Nvidia's CUDA platform, so capacity means Nvidia hardware specifically, which has been scarce and expensive. Being CUDA poor shapes technical choices: smaller or quantised models, parameter-efficient fine-tuning such as LoRA, rented spot instances, careful batch sizing, and research on efficiency rather than scale. The term is self-deprecating rather than insulting, and the constraint it describes has produced much of the open-source work on running capable models on modest hardware.

    In plain English: Lacking enough GPU compute resources, especially CUDA-capable ones.

    Example: "They were CUDA poor enough that every training experiment had to justify itself before consuming the tiny shared GPU window."

Related Terms