NVIDIA Reveals Blackwell Cost Details: GPU Price Doubles, Token Cost Drops 35x

AIMPACT Update, April 30 (UTC+8): According to monitoring by Beating, NVIDIA published a blog post analyzing hardware selection for inference, with one core argument: when evaluating inference infrastructure, focus on “cost per token,” not “cost per GPU per hour.” In terms of GPU unit price, Blackwell is more expensive; but in terms of cost per token, Blackwell crushes its predecessor. The blog used DeepSeek-R1 (an MoE inference model) as the test case, comparing Blackwell (GB300 NVL72) with the previous-generation Hopper (HGX H200). Based on cloud market rental rates, Blackwell costs $2.65 per GPU per hour—nearly double Hopper’s $1.41—but single-GPU token output per second jumps from 90 to 6,000, a 65x throughput increase that reduces cost per million tokens from $4.20 to $0.12. Token output per megawatt improves by 50x. Important caveat: the $0.12 figure assumes all software optimizations are enabled, including FP4 low-precision inference and MTP (Multi-Token Prediction, which allows the model to generate multiple tokens in a single pass to accelerate throughput). Raw data from SemiAnalysis InferenceX v2 shows that on the same GB300 NVL72 running DeepSeek-R1, without MTP, the cost per million tokens is approximately $2.35; with MTP enabled, it drops to around $0.11—just this single optimization creates a 21x difference. All figures above are based on testing with the DeepSeek-R1 model alone; results will vary across different model architectures and scales. (Source: BlockBeats)