NVIDIA's software stack cuts DeepSeek V4 token costs by 5x on Blackwell

NVIDIA's inference software stack on Blackwell GPUs reduces token costs by up to 5x on DeepSeek V4 within a month. Customers like Baseten, Cognition, and Together AI are leveraging TensorRT-LLM and Dynamo for production inference. DigitalOcean helped Hippocratic AI achieve 30% higher throughput on Blackwell.