NVIDIA Blackwell software cuts DeepSeek V4 token costs 5x

NVIDIA's full-stack inference software on Blackwell reduced DeepSeek V4 token costs by up to 5x in one month, as companies move from peak specs to cost-per-token efficiency. Baseten used TensorRT-LLM to serve DeepSeek V4 Pro on Blackwell, delivering up to 50% more tokens per second for reasoning and coding workloads.