据Digitimes报道，西部数据CEO在近期财报电话会上直言，AI本质上是数据系统，任何工作负载都会生成海量数据，这为HDD带来了比算力更具持久性的市场机遇。其前七大核心客户订单可见度已排至2030至2031年。
这一判断直接反映在产能上。西部数据2026年HDD产能已全部售罄，长期供应协议甚至签到了2028年。公司2027财年第一季度营收同比大增55%，下季度指引远超预期，但管理层坦言，收入与出货量仍受限于供给能力，而非需求不足。
AI对存储的拉动，远不止GPU需要HBM和高速SSD这么简单。GPU负责训练和推理，企业级SSD承接热数据，而HDD则保存训练数据集、历史版本、日志、视频及备份。
随着AI从训练转向推理，KV缓存管理等任务正从昂贵的GPU显存转移到共享高性能存储，进一步推高了容量需求。
具体到单机配置，差距极为悬殊。单台AI服务器的存储配置高达1.7TB，而传统服务器仅有0.5TB，前者是后者的三倍以上。
当云厂商和超大规模企业扩建AI集群时，采购清单绝不会只有GPU服务器，海量数据的落地、归档与长期保存，才是存储需求真正的放大器。
换言之，GPU决定了AI算得多快，但存储决定了AI能走多远。没有足够的容量托底，再强的算力集群也无法高效运转。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
According to Digitimes, Western Data CEOs, speaking out at recent financial teleconferences, AI is essentially a data system and any task load generates big data, which presents a more persistent market opportunity for the HDD than for the economy. The first seven core client orders have been scheduled to be visible from 2030 to 2031.
This judgement is directly reflected in productivity. Data for the west show that HDD capacity was fully sold in 2026 and LTAs were even signed in 2028. While the company increased its revenue by 55 per cent in the first quarter of the year, and the guidelines for the next quarter were well ahead of expectations, management frankly stated that revenue and deliveries were still limited to supply capacity rather than demand.
AI's pull on storage is much more than GPU needs HBM and high speed SD. GPU is responsible for training and reasoning, enterprise-level SSD carries heat data, while HDD keeps training data sets, historical versions, logs, videos and backups.
As AI moved from training to reasoning, tasks such as KVcache management were shifting from expensive GPUVRAM to sharing high-performance storage, further pushing capacity requirements.
In terms of individual configurations, the gap is extremely wide. The storage configuration of single AI servers is as high as 1.7 TB, compared with only 0.5 TB for traditional servers, which are more than three times higher than the latter.
When cloud manufacturers and mega-enterprises expand the AI cluster, the purchase list will never be the real amplifier of storage needs only for GPU servers, the landing, archiving and long-term preservation of big data.
In other words, GPU decides how fast AI counts, but storage determines how far AI can go. There is not enough capacity to operate efficiently with any stronger calculator clusters.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.