According to Ornn Data, on September 8th, the spot rental rates for A100 and H100 were approximately $1.02 and $3.08 per hour, respectively. In their forward models, the five-year rental rate for A100 is equivalent to 80.2% of the one-month rate, while for H100, H200, and B200 it's only 43.7%–59.8%.
In specific single-card tests, the cost per million output tokens for A100 running gpt-oss-120b is $0.29, lower than H100's $0.64. Although the model has 117 billion total parameters, each token only needs to activate 5.1 billion parameters, thus allowing for adaptation of older, lower-priced cards. In tests using NVIDIA's optimized software stack, the cost for H100 can be reduced to approximately $0.09.
This data truly weakens the depreciation assumption that "new chips immediately replace old chips"—GPU value is more likely to migrate in layers based on workload, rather than being reduced to zero across generations.
