Running optimized 800-billion-parameter open-weight models on private GPUs can reduce token processing costs to $1 to $2 per million tokens