At the World Artificial Intelligence Conference (WAIC), Western Digital pointed out that AI is not just a computing system, but a data system. Computing cycles can be reused, while data continues to accumulate. How data is stored, managed, and retained long-term has become a critical factor in determining whether AI can achieve sustainable and cost-effective scaling.
The Overlooked "AI Data Amplification Effect"
Tim Rausch, Senior Vice President of Development Engineering at Western Digital, pointed out that the public's intuitive impression of generative AI often stays at inputting prompts and getting instant results. However, at the backend of data centers, every interaction triggers a data loop, where generated content migrates down the storage hierarchy and is ultimately stored in Hard Disk Drives (HDDs).
Figure | Tim Rausch, Senior Vice President of Development Engineering at Western Digital; Source: Western Digital
Tim Rausch shared a real-world test case using a common AI tool: a user inputted a prompt to generate an 8-second short video with a file size of only 3.97MB. However, at the backend, up to 44GB of data flowed through GPU (Graphics Processing Unit) processing; more importantly, for subsequent model optimization, compliance review, and auditing, the system ultimately retained encoded outputs, inference trace logs, compliance records, and telemetry metadata in the storage layer, totaling 509MB of data.
The actual amount of data retained by AI is up to 130 times the final delivered content. Meanwhile, when this video is published, shared, and subsequently recognized and analyzed again by other AI systems, new copies and derivative data will accumulate in an endless loop.
Therefore, industry analysts predict that global data volume will face exponential growth in the coming years. This self-reinforcing loop, where data flowing through computing systems is continuously enriched to generate even more data, puts unprecedented expansion pressure on the original warm and cold data storage tiers.
Economics Dictates the Tiering of Cloud Storage Architecture
Tim Rausch pointed out that to cope with this data tsunami, cloud storage architecture is being forced to evolve from the traditional2-tier (flash memory + traditional HDDs) to a specialized "AI 4-tier storage architecture."
Figure | Cloud storage architecture is being forced to evolve from the traditional 2-tier to a 4-tier storage architecture; Source: EEFocus
Currently, leading AI infrastructure is restructuring sophisticated data tiering systems around GPUs. The top tier consists of HBM (High Bandwidth Memory) and DRAM (Dynamic Random Access Memory) located adjacent to the GPU, responsible for rapidly feeding data to the GPU; immediately followed by the KV Cache, which handles the short-term memory for conversational context; the next tier is the semantic database running on high-performance SSDs (Solid State Drives), used to process Vector, Embeddings, and RAG (Retrieval-Augmented Generation) data. The underlying infrastructure that supports all this, the bottom tier, is the high-capacity HDD storage layer responsible for preserving logs, raw training data, and inference replicas.
This intelligent tiering stems from extreme cost considerations. With the significant surge in flash memory and DRAM costs in recent years, when hyperscale cloud providers manage exabyte (EB)-level data, even a tiny price difference of just 1 cent per GB of storage media, when scaled up to a 100EB data volume, means a rigid cost fluctuation of up to $1 billion. According to data cited by Tim on-site, currently about 80% of AI content is stored on HDDs and about 20% on SSDs, reflecting the tiered synergy between the two media in terms of capacity, performance, and cost. Therefore, migrating performance-insensitive warm and cold data to the lower-cost, high-capacity HDD storage tier is an inevitable choice to maintain the economic viability of the infrastructure.
Microscopic Grains and Hardware Restructuring
Meanwhile, to fit more data into limited data center space, underlying hardware vendors are concurrently advancing two magnetic recording technologies: ePMR (Energy-Assisted Perpendicular Magnetic Recording) and HAMR (Heat-Assisted Magnetic Recording), attempting to push single-drive capacity beyond 100TB in the coming years.
Figure | ePMR (Energy-Assisted Perpendicular Magnetic Recording) technology; Source: Western Digital
Figure | HAMR (Heat-Assisted Magnetic Recording) technology
The key to achieving 100TB capacity lies in the dual breakthroughs of areal density and internal structure. HAMR can currently achieve 4TB capacity per disk platter, and 11 platters can form a 44TB hard drive. At the microscopic level, by shrinking the diameter of the magnetic grains in the recording medium (breaking through the existing 8nm limit), more data can be accommodated on the same platter.
Tim Rausch believes that through this technology iteration, single-platter capacity can be increased from the current 4TB to about 7TB, and even reach 10TB in the future.
Furthermore, in HAMR technology, to heat the medium for high-density writing, extremely tiny micro-laser components will be integrated into the read/write head, thereby freeing up space inside the hard drive. In the future, 14 7TB platters can form an approximately 100TB hard drive; 14 10TB platters can reach 140TB.
At the same time, structural restructuring can also enhance the throughput performance of hard drives. According to Tim Rausch, High Bandwidth Drive Technology (HBDT) uses a triple-stage actuator, allowing simultaneous reading and writing of multiple paired tracks, directly doubling the sequential throughput.
Figure | High Bandwidth Drive Technology (HBDT); Source: Western Digital
Dual Pivot Technology (DPT) distributes the read/write heads on actuators with two independent pivots, increasing sequential I/O performance by 2 times, and creating conditions for accommodating more platters by reducing the spacing between them.
Figure | Dual Pivot Technology (DPT); Source: Western Digital
When these two technologies are combined, the sequential I/O performance of the hard drive will be improved to 4 times the original level overall. While achieving 100TB HDDs, it can still maintain the customers' current relative I/O rate per TB.
Tim Rausch revealed that current High Bandwidth Drive samples have entered the customer validation phase and will gradually move towards mass production in the coming years. Western Digital plans to start providing Dual Pivot Drive samples in 2027 or early 2028, followed by advancing mass production.
Additionally, Tim Rausch also mentioned power-optimized HDDs. Its principle is to reduce the platter rotation speed (reducing power consumption by about 20%, with only a 5%-10% drop in performance), providing an economically sustainable solution for cold data storage.
How to Unlock the Potential of Existing Hardware?
The vision of 100TB is certainly promising, but facing the immediate data pressure, data centers urgently need cost-reduction solutions that can be deployed right now. The commercialization of Shingled Magnetic Recording (SMR) and Variable Capacity technologies is exactly a product of this background.
Figure | Ultra-High Density Shingled Magnetic Recording (SMR) technology; Source: Western Digital
SMR shingled technology breaks the limitation of traditional Conventional Magnetic Recording (CMR) tracks being independent of each other, allowing tracks to overlap like roof tiles, thereby increasing usable capacity by 20%. Although overlapping tracks require garbage collection when rewriting data, in the SMR architecture shown in the presentation, the control is completely handed over to the host side, allowing enterprises to flexibly schedule according to the peak and valley periods of AI workloads.
Tim Rausch revealed that in the simplified five-year TCO (Total Cost of Ownership) example shown in the presentation, this technology can reduce the TCO per TB by about 8.4% when calculated.
Variable Capacity technology breaks the delivery limitations of traditional fixed tiers. In traditional manufacturing, usable capacity that exceeds the standard but does not reach the next tier is usually masked and wasted. Variable Capacity allows hard drives to be delivered based on the highest stable capacity actually tested physically.
This technology not only immediately brings an initial capacity increase of about 3%, but more flexibly, as the production line process matures, the average usable capacity of the same hard drive design can increase by up to 10% within a year. Enterprises only need to pass the software architecture certification once, and in the future, they can directly obtain the TCO savings brought by process improvements.
Conclusion
According to Grand View Research, China's AI data center market is expected to exceed $66 billion by 2033, with a compound annual growth rate (CAGR) of 26.2% from 2026 to 2033. A sample survey conducted by Western Digital among core customers in the Chinese market shows that 81.8% of respondents prioritize meeting AI training and inference workloads and TCO optimization, and HDDs account for 50% to over 75% of the data storage infrastructure for the majority of Chinese customers.
Tim Rausch emphasized that in the battle of scaled AI, computing power and storage cannot be separated. HDDs and SSDs are not a zero-sum game. Playing their respective roles and intelligently tiering within a professional 4-tier storage stack is the pragmatic solution to support the sustainable and scaled development of AI.