With the global market for generative artificial intelligence projected to exceed 1.3 trillion by 2032, companies are eager for enterprise-grade storage solutions that can handle AI workloads.

A majority of gen AI and agentic proof of concepts are struggling to reach production. Storage companies want to help customers take that final leap.

Industry experts from Nutanix, Supermicro, MinIO and Intel talk with theCUBE about their latest collaboration efforts.

“The top reasons that those prototypes didn’t really get into production were that the cost was too high, the performance was not high enough or they didn’t have good IT alignment,” said Wendell Wenjen (pictured, front row right), director of storage market development at Super Micro Computer Inc. “The more recent research that we commissioned with IDC estimates that about a half of these companies are going into production, so I think the numbers are looking very good now.”

Wenjen, Kanchan Mirani (front row left), senior director of strategic marketing at Nutanix Inc.; Keith Pijanowski (pictured, back left), AI solutions engineer at MinIO Inc.; and Gary Brown (pictured, back right), technical product marketing at Intel Data Center Group at Intel Corp., spoke with theCUBE’s Rob Strechay at the Supermicro Open Storage Summit interview series, during an exclusive broadcast on theCUBE, SiliconANGLE Media’s livestreaming studio. They discussed their partnerships and creating storage solutions for enterprise gen AI applications. (* Disclosure below.)

Putting models into production

Supermicro’s storage infrastructure is built on servers using Intel’s Xeon 6 processors, with object storage from MinIO. On top of that base, Supermicro has developed several storage solutions with Nutanix, equipping users to run gen AI applications in hybrid and edge environments.

Companies are increasingly interested in building models on proprietary data, for sectors such as healthcare or finance, according to Wenjen. Supermicro facilitates that development process.

“You’re going to have a trained model that has your enterprise’s unique data and you’re going to be ready to move that into production, which is the last phase,” Wenjen said. “Here, you really want to analyze what you think the maximum concurrent query rate is going to be so that you have sized your equipment and infrastructure appropriately. ”

Supermicro’s partner, Nutanix, specializes in hyperconverged software, which is especially useful for deploying AI on the edge. The collaboration between the two companies includes harnessing the Nutanix Kubernetes Platform, for simplifying platform engineering, and Nutanix Unified Storage, a platform that consolidates data services in a secure manner.

“The first thing people tend to think about when they think about infrastructure for AI is GPUs,” Mirani said. “So, they say, ‘Well, I’ve got my GPU. I’ve got my infrastructure in place.’ But between the AI accelerators, be it GPUs, TPUs or NPUs, CPUs and getting an AI application to production at enterprise-grade, there’s a big gap.”

Because GPUs are so expensive, ensuring that storage is not the “bottleneck” in an AI pipeline is essential, Mirani pointed out, which is why Nutanix ran a benchmark to test the efficiency of its storage solutions.

“The goal was to see how many GPUs can you saturate, get 90% utilization from them, how much can you scale to, and with Nutanix Unified Storage, we were able to scale to over 1,000 GPUs and saturate those GPUs,” Mirani added. “The next vendor was less than half of that. And this was a files workload for image classification.”

Striving for secure storage solutions

Another Supermicro partner is Intel, which has touted its Xeon 6 processors as the most efficient option for modern data centers. The processors can manage overflow from GPU workloads and have greater memory bandwidth than previous iterations, according to Brown.

“You need a massive volume of structured and unstructured data involved when you’re implementing generative AI,” he said. “You also require very high-throughput and low-latency data connection. So, your storage systems have to be built for a huge amount of data to be passed through and stored.”

Rounding out the cohort is MinIO, which offers an object storage system. The company’s AIStor is designed for enterprise-grade AI; its S3 compatibility means the system can adapt to different datasets while ensuring protection for proprietary data.

“What AIHub is all about is we’ve implemented the Hugging Face APIs, we make sure that upon the initial pull, you save your models … to AIStor,” Pijanowski said. “But subsequent versions will automatically be saved back to AIStor and will prevent any mistakes where you push those models back to the hub.”

Some customers are still experiencing choice paralysis over which model to go with and how to start implementing it, but Wenjen is confident that those concerns will be allayed.

“We’ve come a long way in progress in terms of the ease of use,” he said. “It’s not all figured out, we’re still working on this, and there’s going to be a lot more developments when we do this next year. But learning by doing … that is a really good piece of advice.”

Here’s a short clip from our interview, part of SiliconANGLE’s and theCUBE’s coverage of the Supermicro Open Storage Summit:

(* Disclosure: TheCUBE is a paid media partner for the Supermicro Open Storage Summit. Neither Super Micro Computer Inc., the sponsor of theCUBE’s event coverage, nor other sponsors have editorial control over content on theCUBE or SiliconANGLE.)

Photo: SiliconANGLE