AI推理背后的隐性存储成本问题

AI推理过程中的数据存储成本远被低估。虽然企业通常将关注点集中在计算成本（GPU）上，但AI推理产生的大量后端数据存储需求正迅速成为重要的成本因素。1一个8秒的AI生成视频在用户端呈现为3.97MB的720p文件，但在后端却会产生超过500MB的关联数据，相当于输出文件大小的130倍。1

根据IDC对763名IT和业务决策者的全球调查，95%的受访者在过去12个月内报告了因生成式AI导致的数据量增长，其中61%的增幅达到25%或更高。1McKinsey的预测显示，AI推理需求将以35%的年复合增长率增长至2030年，这一速度高于训练需求的22%，推理最终将占数据中心需求的40%以上。1随着这一趋势的发展，存储成本可能在12至18个月内从项目预算中被忽视的部分转变为重要的支出项目。1

调查中96%的受访者预计需要更快的存档检索能力以支持推理或检索增强生成（RAG）应用。1一家超大规模云服务提供商的高级产品经理指出，如果GPU处于空闲等待数据的状态，相关成本与缺乏GPU一样昂贵，这进一步突显了存储基础设施与计算资源之间的紧密关联。1

The computational costs of artificial intelligence inference have long dominated industry discussions, but a significant hidden expense is being overlooked: the data storage required to support these operations. An analysis examining AI infrastructure reveals that the backend storage footprint generated during inference far exceeds what most organizations anticipate, with storage expenses poised to become a major cost component within months rather than years 1.

A striking example illustrates the scale of this problem: an 8-second AI-generated video that results in a 3.97MB file for end users generates over 500MB of associated data on backend systems—roughly 130 times the size of the output file itself 1. According to McKinsey projections, AI inference demand will grow at a 35% compound annual growth rate through 2030, a faster pace than the 22% growth rate for training workloads, and will eventually account for more than 40% of data center requirements by the end of the decade 1. An IDC survey of 763 IT and business decision-makers found that 95% reported increased data volumes over the past 12 months due to generative AI, with 61% experiencing growth exceeding 25% 1. Senior infrastructure planners acknowledge the urgency: 96% of survey respondents anticipate needing faster archive retrieval capabilities to support inference and retrieval-augmented generation (RAG) applications 1. The competitive pressure is intensifying as well—an executive at a major cloud provider noted that idle GPUs waiting for data access incur costs equivalent to having no GPU capacity at all 1.