Quantization and distillation let capable models run on a single GPU. For focused tasks, a tuned open model can beat a general giant while keeping data on-prem.