开源权重推理的经济学分析

Ornn Data发布研究论文，分析了开源权重模型推理的经济学特征。1研究表明，在自托管场景下，较旧的NVIDIA A100 GPU在运行稀疏模型时成本优势明显。1以稀疏模型gpt-oss-120b（5.1亿活跃参数）为例，A100在满使用率下的成本为每百万输出token 0.29美元，低于H100的0.64美元；而在完全利用的情况下，A100的成本进一步降低至0.12美元，相比之下H100为0.27美元。1

这项研究挑战了新一代硬件自动淘汰前代GPU的传统假设。1数据显示，A100的五年期租赁价格保持其一个月价格的80.2%，而Hopper和Blackwell的保留率分别为43.7%-59.8%和53.8%，表明A100具有更强的价值保持能力。1在租赁市场上，A100的占有率从2026年3月的74%上升至9月的90%，期间列表容量增长了13%，现货指数上升20%。1按照现有租赁周期，五年A100合约将持续至2031年9月，届时Ampere系列芯片已服役超过11年。1

研究未能确立开源权重模型需求与A100占有率上升、租赁价格上涨之间的因果关系。1论文认为，只要存在适配的工作负载且运营者可自由部署，较新硬件代就不会自动使前代GPU经济上过时。1

Ornn Data has published research examining the economic implications of inference workloads using open-weight models, challenging conventional assumptions about hardware obsolescence.1 The study demonstrates that in self-hosted environments, older NVIDIA A100 GPUs can deliver lower costs than H100s when running sparse models, with A100 five-year lease prices retaining 80.2% of their one-month pricing compared to 43.7%-59.8% for Hopper and 53.8% for Blackwell.1 This suggests that newer hardware generations do not automatically render previous-generation GPUs economically obsolete, provided suitable workloads exist and operators retain deployment flexibility.1

The research quantifies substantial cost advantages for open-weight models at inference time.1 Open-weight models cost approximately one-fifth that of comparable closed-source alternatives at sampling thresholds.1 For the sparse model gpt-oss-120b, which contains 510 million active parameters, A100 inference costs $0.29 per million output tokens at full utilization or $0.12 per million tokens at complete capacity, both undercutting H100 rates of $0.64 and $0.27 respectively.1 Market indicators show A100 share rising from 74% in March 2026 to 90% by September, with overall list capacity growing 13% and spot pricing increasing 20%.1 Five-year A100 contracts extend through September 2031, at which point Ampere architecture will have operated for over eleven years.1

Despite these findings, the researchers could not establish a causal relationship between demand for open-weight models and observed changes in A100 market share and lease pricing.1