NVIDIA Has Deployed Massive Computing Power to Support DeepSeek V4
Not a simple software push — NVIDIA tied hardware to a fresh model design and rolled out support in tandem. The company says Blackwell GPUs can already push roughly 3,500 tokens/sec per card, though that's an early snapshot and will shift with tuning and real-world workloads.
DeepSeek V4 itself was reworked to cut costs: computation per token is down to 27% of what the previous generation required (i.e., ~73% lower compute per token), and KV-cache usage is reported at 10% even with a 1,000,000-token context. Two model flavors landed at launch: the Pro build (up to 1.6T parameters) and a Flash variant at 284B params — different trade-offs for different deployment needs.
A few implementation notes matter more than headline numbers. FP4 quantization (e.g., lower-precision formats) is being used to trim latency and memory pressure; combined with optimized CUDA kernels and parallelization schemes, that’s the performance delta NVIDIA highlights. Still, throughput will depend on batching, sequence length, and the particulars of the inference stack — so your mileage may vary.
Finally, the story isn’t limited to Western silicon. Huawei’s Ascend 950 chips are slated to support similar instructions in 2026, which means DeepSeek V4 could run efficiently on non-Blackwell hardware as well. That widens the field of viable accelerators and raises practical questions about software portability, ops, and cost for those deploying large-context models.