Supermicro alliance tackles the storage bottlenecks holding back enterprise AI
SiliconANGLE Mark Albertson ● Covered by 3 sources
Supermicro, Hammerspace and Sandisk are pushing storage tech built for AI instead of old data piles. The bet: keep GPUs busy by moving data faster, not just storing more of it.
Based on reporting by SiliconANGLE, Mark Albertson — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Supermicro is making a case that storage is no longer a back-office afterthought for enterprise AI. The old conversation was about capacity. Now it is about whether storage can actually keep up with AI systems that need data delivered fast, flexibly and without a mess of copies and migrations.
That shift matters because a lot of companies are trying to bolt AI onto storage setups that were built before AI was the point. Those systems tend to be fragmented, siloed and aging. Allen Liu, a senior product manager of architecture solutions at Super Micro Computer, said the goal is to preserve performance and scale while also keeping flexibility, composability and efficiency in the same package.
The company’s answer runs through partners. Hammerspace is pitching a software layer that sits on existing infrastructure, whether that infrastructure is in the cloud, a data center or at the edge. Molly Presley said it activates data where it already lives, so enterprises do not need a big copy-and-migrate project just to make datasets usable for AI models and workflows. That is a cleaner pitch than asking everyone to rebuild their storage stack from scratch.
Sandisk is addressing a different pressure point: GPU utilization. Praveen Midha said average GPU use can sit around 30% to 50%, which leaves a lot of expensive hardware underfed. Sandisk’s QLC SSDs are aimed at giving AI storage environments more capacity and efficiency, with flash also moving closer to the GPU in the stack as workloads get more demanding.
There is also a standards story here. Supermicro has adopted NVMe over Fabrics, which extends low-latency NVMe storage across data center networks using technologies such as Remote Direct Memory Access. Presley argued that standardization matters because these environments need to work across different vendors and locations. Midha added that AI reasoning is driving larger key-value cache sizes, and that flash is increasingly part of the memory tier for getting cached data back quickly.
The broader message is simple: enterprise AI is becoming a storage problem as much as a compute problem. Supermicro wants to be the foundation, Hammerspace the orchestration layer and Sandisk the storage media layer. That is a tidy division of labor, and a pretty blunt reminder that the fastest GPU in the room still spends its day waiting for data.
My take — AI-written commentary, not fact-checked reporting
A lot of AI talk still treats storage like plumbing, which is how companies end up with expensive GPU fleets idling like bored interns. The real story here is not flash or protocols by themselves; it is the return of systems thinking, the thing the industry keeps rediscovering every time a shiny model hits a very ordinary bottleneck.
Read more about this at: SiliconANGLE