Need smarter insights in your inbox? Join our weekly newsletters to get solely what issues to enterprise AI, information, and safety leaders. Subscribe Now
As AI purposes more and more permeate enterprise operations, from enhancing affected person care by way of superior medical imaging to powering complicated fraud detection fashions and even aiding wildlife conservation, a essential bottleneck usually emerges: information storage.
Throughout VentureBeat’s Rework 2025, Greg Matson, head of merchandise and advertising, Solidigm and Roger Cummings, CEO of PEAK:AIO spoke with Michael Stewart, managing accomplice at M12 about how improvements in storage know-how allows enterprise AI use circumstances in healthcare.
The MONAI framework is a breakthrough in medical imaging, constructing it quicker, extra safely, and extra securely. Advances in storage know-how is what allows researchers to construct on prime of this framework, iterate and innovate shortly. PEAK:AIO partnered with Solidgm to combine power-efficient, performant, and high-capacity storage which enabled MONAI to retailer greater than two million full-body CT scans on a single node inside their IT atmosphere.
“As enterprise AI infrastructure evolves quickly, storage {hardware} more and more must be tailor-made to particular use circumstances, relying on the place they’re within the AI information pipeline,” Matson stated. “The kind of use case we talked about with MONAI, an edge-use case, in addition to the feeding of a coaching cluster, are nicely served by very high-capacity solid-state storage options, however the precise inference and mannequin coaching want one thing totally different. That’s a really high-performance, very excessive I/O-per-second requirement from the SSD. For us, RAG is bifurcating the varieties of merchandise that we make and the varieties of integrations we’ve to make with the software program.”
Bettering AI inference on the edge
For peak efficiency on the edge, it’s essential to scale storage right down to a single node, as a way to deliver inference nearer to the information. And what’s secret is eradicating reminiscence bottlenecks. That may be finished by making reminiscence part of the AI infrastructure, as a way to scale it together with information and metadata. The proximity of knowledge to compute dramatically will increase the time to perception.
“You see all the massive deployments, the massive inexperienced discipline information facilities for AI, utilizing very particular {hardware} designs to have the ability to deliver the information as shut as potential to the GPUs,” Matson stated. “They’ve been constructing out their information facilities with very high-capacity solid-state storage, to deliver petabyte-level storage, very accessible at very excessive speeds, to the GPUs. Now, that very same know-how is going on in a microcosm on the edge and within the enterprise.”
It’s turning into essential to purchasers of AI methods to make sure you’re getting probably the most efficiency out of your system by operating it on all stable state. That lets you deliver enormous quantities of knowledge, and allows unimaginable processing energy in a small system on the edge.
The way forward for AI {hardware}
“It’s crucial that we offer options which can be open, scalable, and at reminiscence velocity, utilizing among the newest and biggest know-how on the market to try this,” Cummings stated. “That’s our objective as an organization, to supply that openness, that velocity, and the size that organizations want. I feel you’re going to see the economies match that as nicely.”
For the general coaching and inference information pipeline, and inside inference itself, {hardware} wants will hold growing, whether or not it’s a really high-speed SSD or a really high-capacity resolution that’s energy environment friendly.
“I’d say it’s going to maneuver even additional towards very high-capacity, whether or not it’s a one-petabyte SSD out a few years from now that runs at very low energy and that may mainly exchange 4 instances as many onerous drives, or a really high-performance product that’s virtually close to reminiscence speeds,” Matson stated. “You’ll see that the massive GPU distributors are taking a look at the best way to outline the subsequent storage structure, in order that it will probably assist increase, very intently, the HBM within the system. What was a general-purpose SSD in cloud computing is now bifurcating into capability and efficiency. We’ll hold doing that additional out in each instructions over the subsequent 5 or 10 years.”
