Nvidia Blackwell cuts DeepSeek V4 cost by 80%

Nvidia says its Blackwell platform can now run DeepSeek V4 at one-fifth the cost per token compared to a month ago. The company calls cost per token a key metric for AI total cost, claiming it's now the lowest in the industry.

The optimization happens at three layers: operations for service orchestration, application acceleration for compute and communication overlap, and infrastructure for GPU and network access. Blackwell's single GPU token throughput can increase up to 20 times.