To Store or To Regenerate? A Cost Model for AI-Generated Content at Scale
原文链接: https://arxiv.org/abs/2609.30448
作者: Yunjia Zheng et al.
发布时间: 2026-09-24
源: arxiv
摘要
arXiv 2609.30448 建立一个成本模型比较 AI 生成内容的持久存储与按需重生成,涵盖语料增长、HDD/磁带价格趋势、驱动器替换、电价、请求偏斜、缓存、generator FLOPs 与未来 GPU 价格性能改进。对图像生成:prompt-based regeneration 要到 2040 年前后才比存储便宜,因每次条件生成都远比一次存储贵。数字反直觉但推演透明,适合做茶余谈资。
English Summary
arXiv 2609.30448 develops a cost model comparing persistent storage against on-demand regeneration for AI-generated artifacts. The model accounts for corpus growth, HDD and tape price trends, drive replacement, electricity, request skew, caching, generator FLOPs, and projected GPU price-performance improvements. For image generation, prompt-based regeneration does not become cheaper than storage until around 2040, because each conditional generation pass costs far more than a stored copy. The headline number is counterintuitive but the derivation is transparent, making the paper a useful talking point rather than an action item.
为什么值得关注
主题线扩展 AAIF storage-cost/regeneration-cost/cost-model/gpu-price 等主题。