DeepSeek V4 Pro lands on NVIDIA build.nvidia.com — 1M context, >150 tok/s on GB200

DeepSeek V4 Pro — 1.6T parameters with 49B activated and a 1M-token context — is now hosted on NVIDIA's build.nvidia.com. Early benchmarks on the GB200 NVL72 (Blackwell Ultra) report throughput above 150 tokens per second per user, alongside a 75% V4 Pro API price cut and 1/10 cache pricing that's already forcing competitor recalibration.