Disaggregated Storage – the Latest AI Trend
An Enterprise AI solution isn't complete without high-performance storage. The huge quantities of data that must be served up quickly to the GPUs make high throughput and low latency storage critical to maximizing any AI solution investment. In a general sense, AI is a subset of High-Performance Computing, so storage solutions that have dominated the HPC space for decades were the obvious solution. These solutions combine controllers and flash storage into a single chassis that scales by adding additional nodes to the environment.
More recently, some storage OEMs have begun to challenge the status quo for how storage for AI solutions should be designed. The traditional method of combining storage, compute and capacity into a single chassis limits the flexibility of scaling options. If a system needs more capacity but is performing at or above requirements, a customer would have no choice but to scale up both. The same would go for situations where a solution is performance-bound; ie, the storage tier would have to add multiple nodes to meet performance requirements, and the customer ends up with 2- 3x the capacity the solution needs.
Shared-nothing architecture is a distributed computing design where each node in a system has its own dedicated resources (CPU, memory, storage) and doesn't share them with other nodes. This design was ideal for horizontal scaling and designing solutions with no resource contention, but it has disadvantages when used in a modern AI deployment. Data consistency becomes an issue, as synchronization of data between nodes takes time and resources. To balance workload across nodes, data is often broken into shards and distributed across all available nodes; this can add complexity and make querying data more difficult. This approach also tends to add a lot of east/west network traffic, which can impact overall performance.
Disaggregated storage solutions address these limitations by scaling the storage compute nodes independently from the capacity nodes. VAST Data is synonymous with storage for AI and was the first AI-focused OEM to offer a disaggregated storage solution. Pure Storage recently announced FlashBlade//EXA, which uses the FlashBlade storage for high-performance metadata and scales out the compute and storage nodes as needed. Hammerspace is another emerging storage for AI solutions that features disaggregated storage. HPE also adapted the disaggregated architecture for their Alletra MP platforms: B10000, X10000, and GreenLake for File. NetApp will soon be announcing its own disaggregated solution, DASE, that will pair with its A1000 storage solution. By coupling disaggregated scaling capabilities to traditional high-performance file storage, OEMs can now deliver top-tier performance and provide increased flexibility on deployment that AI solutions require.
There are some challenges that can be introduced with a disaggregated solution when compared to a more traditional combined storage node. Networking requirements tend to increase, since separating nodes typically increases the overall port count, which can lead to additional interconnect switch requirements. It can also take more detailed planning for both initial deployment and future scaling, since storage, compute and capacity would require separate monitoring. And while larger deployments would benefit from the independent components, it will also increase the size for smaller deployments due to minimum node counts and RAID requirements.
Where do customers go from here?
Coexistence, not immediate replacement, is the prevailing trend for enterprises adopting disaggregated AI storage. Rather than ripping out existing general-purpose storage platforms, most companies are augmenting their environments with new high-performance file or object storage clusters specifically for AI workloads. This allows them to retain traditional SAN/NAS systems for established business applications, databases, and user files, which continue to run effectively on these proven, stable platforms. This phased adoption minimizes risk, leverages existing investments, and addresses skill gaps associated with new storage technologies, enabling IT staff to master the new systems in a controlled environment.
This strategy is driven by several practical considerations. Existing workloads perform optimally on legacy systems designed for low-latency IOPS and high availability, making migration unnecessary and potentially disruptive. Furthermore, there's a natural workload segmentation: AI training and big data analytics thrive on the high-bandwidth, scalable performance of disaggregated storage, while transactional databases and virtual machines remain on traditional SAN/NAS optimized for fast random I/O. This ensures each workload receives the appropriate service level and prevents performance conflicts. Companies also maintain legacy storage as a risk management strategy and to preserve vendor relationships, especially in industries with strict compliance or certification requirements.
Looking ahead, a complete displacement of traditional arrays by AI-focused disaggregated storage is unlikely in the near to mid-term. While capabilities are converging, and some specific use cases (like large-scale file storage for research or backup targets) are seeing a shift, general-purpose arrays will remain relevant for their optimized performance with certain traditional workloads. Enterprise storage refresh cycles are long, meaning companies will continue to leverage existing infrastructure for years to come. However, as disaggregated platforms mature, become more cost-effective, and offer broader workload support, they may gradually become the new standard for enterprise storage, integrating more deeply into IT roadmaps as part of a broader shift toward fluid, platform-based resource pools.
WWT can help!
WWT has been helping customers navigate the HPC and AI space for over a decade. We will work with customers to set up solution overviews, workshops, or lab POCs to help you and your team make the best decision for your environment.