Abstract:
According to Digitimes,
Western Digital CEO bluntly stated in a recent earnings call that AI is essentially a data system, and any workload will generate massive amounts of data, which brings market opportunities for HDD that are more durable than computing power. The order visibility of its top seven core customers has been scheduled to 2030 to 2031.
This judgment is directly reflected in production capacity. Western Digital's HDD production capacity in 2026 has been sold out, and the long-term supply agreement has even been signed until 2028. The company's revenue in the first quarter of fiscal year 2027 increased by 55% year-on-year, and its guidance for the next quarter far exceeded expectations. However, management admitted that revenue and shipments are still limited by supply capacity, not insufficient demand.
The impact of AI on storage is far more than just the need for HBM and high-speed SSD for GPUs.
The GPU is responsible for training and inference, the enterprise-level SSD handles hot data, and the HDD saves training data sets, historical versions, logs, videos, and backups.As AI shifts from training to inference, tasks such as KV cache management are being moved from expensive GPU memory to shared high-performance storage, further pushing up capacity requirements.
When it comes to stand-alone configurations, the gap is extremely wide. The storage configuration of a single AI server is up to 1.7TB, while a traditional server has only 0.5TB. The former is more than three times the latter.
When cloud vendors and ultra-large-scale enterprises expand AI clusters, the purchase list will never only include GPU servers. The implementation, archiving and long-term storage of massive data are the real amplifiers of storage needs.
In other words, the GPU determines how fast the AI can calculate, but the storage determines how far the AI can go. Without sufficient capacity, no matter how powerful the computing cluster is, it cannot operate efficiently.
Comments