A research paper examines how open-weight model demand extends the economic viability of older NVIDIA GPU families like the A100, showing they remain cost-competitive for inference workloads compared to newer generations when self-hosted on rental hardware, challenging assumptions that new GPUs render previous ones obsolete.