AI factories package Nvidia reference architecture, but who's buying?
Nvidia partners bundle the OpenShell Reference Architecture's infrastructure-level AI safeguards in AI factories, but enterprise adoption numbers remain a mystery.
This week at GTC Berlin, Nvidia launched a set of hardware and software products for securing agent workloads, which were quickly packaged for distribution by enterprise partners. But while those vendors say enterprises are interested, so far it's unclear precisely who is buying and ready to deploy them.
Nvidia's OpenShell Reference Architecture combines its Vera Rubin computing platform with infrastructure-level guardrails for agent workloads. Nvidia partners such as Red Hat, HPE and Dell are incorporating the architecture into their respective AI factories. AI factories are pre-integrated hardware and software systems meant to help enterprises run AI workloads in private data centers.
Enterprises have been attuned to agent security in recent months, in part due to several high-profile incidents involving agents, the growing need for sovereign AI, especially in Europe, and shifting liability requirements for software dependencies. This makes a private AI factory potentially more interesting, but such systems also introduce new operational challenges, wrote Mike Leone, vice president and principal analyst at Moor Insights & Strategy, in an email to TechTarget.
Mike Leone
"The [OpenShell Reference Architecture] pulls enforcement out of the agent itself, and in some cases off the server's main processors entirely. Monitoring runs on a separate processor on the network card (Nvidia's BlueField), so a compromised agent can't switch off whatever is watching it," Leone wrote. That's also where the new problems come in. "Companies that invest in private AI will end up with many agents to manage, each carrying a machine identity, plus another infrastructure layer to run," Leone wrote. Most existing data centers weren't built for the power demands and liquid-cooling systems that Vera Rubin rack-scale systems require.
Moreover, the rack-scale Vera Rubin platform, which experts estimate could reach hundreds of millions of dollars in total cost of ownership, requires sophisticated management so that the hardware is utilized enough to pay for itself, Leone wrote.
Vera Rubin buyers, whoever they are, have choices
So far, while Nvidia's AI factory partners say these systems are gaining enterprise traction, it's difficult to be specific about who is buying them and who will be in the market for the still-pending Vera Rubin stacks when they ship.
"Right now, it's mostly the big cloud providers, the neoclouds, the frontier model labs and national AI programs. Most of the enterprises using Vera Rubin are renting it through those clouds," Leone wrote.
For example, in September, TechTarget reported that neocloud provider CoreWeave had put seven Vera Rubin racks into production. Scaling beyond a single rack is an architectural milestone that comes with significant operational complexity.
Enterprise customers have placed orders for Vera Rubin systems, said Tushar Katarki, senior director of product for Red Hat AI.
"I know that enterprise customers, banks specifically, have [ordered Vera Rubin]," he said.
Those customers include telecommunications companies and cloud service providers using Red Hat AI Factory with Nvidia, Katarki said. They plan to use Red Hat Enterprise Linux (RHEL) for Nvidia 26.03 and Nvidia's OpenShell Reference Architecture when deploying Vera Rubin systems, both of which were launched this week.
Buyers interested in taking on a private AI factory will likely choose a vendor whose infrastructure already runs in their data center, Leone wrote.
"Enterprises don't want to build a second, separate stack just for AI. They want it to land inside what they already run, with the same Linux, Kubernetes, identity system, ops team and support contract," Leone wrote. "That's also where these AI factories differ from each other. The Nvidia design underneath is largely the same. What changes [are] the software, management tools and services wrapped around it."
Thus, Red Hat's AI Factory will appeal most to existing Red Hat shops, said Torsten Volk, principal analyst at Omdia, a division of Informa TechTarget.
Torsten Volk
"RHEL for Nvidia lets the existing platform team run basically the same OpenShift it runs everywhere else," Volk said. "The AI nodes just receive some special treatment from the kernel and a couple of Linux packages. This seems pretty attractive for existing Red Hat shops."
Unlike Dell, which lets customers choose their AI software stack, and HPE, which sells Private Cloud AI as an integrated hardware/software system, Red Hat AI Factory runs on hardware from multiple vendors, including Dell, HPE, Lenovo and Cisco, Volk said.
However, RHEL for Nvidia and the OpenShell Reference Architecture both depend on Nvidia hardware. The specialized OS also follows a quarterly release schedule, requiring customers to stay on the latest version to maintain support, a departure from standard RHEL's 10-year lifecycle that could increase operational overhead, Volk said.
HPE declined to comment on whether it has received orders for Vera Rubin or how many orders it has recieved, citing a pre-earnings quiet period. The company says it has more than 200 Private Cloud AI installations overall.
A Dell spokesperson also declined to provide a Vera Rubin order count but said Dell customers are deploying AI systems using multiple generations of Nvidia GPUs, from the H100 to the latest Vera Rubin NVL72. Dell touts more than 6,500 customers using Dell AI Factory with Nvidia, up 1,500 from the previous quarter.
Building AI factories chip by chip
While actual purchase orders for pre-packaged AI factories are difficult to track, market research indicates some interest among enterprises in running AI applications on-premises. Seventy-five percent of 400 IT professionals surveyed by Omdia in April said their organizations prefer to run AI applications on-premises, according to the firm's June 2026 IT Modernization Report. AI factories offer one way to do that.
"The question is whether their organization is mature enough in their process and use case definition, [and whether] their data is the right shape for their organization to justify the cost of these larger full-stack [systems]," said Scott Sinclair, principal analyst at Omdia, who authored the IT Modernization Report.
The question is whether their organization is mature enough in their process and use case definition, [and whether] their data is the right shape for their organization to justify the cost of these larger full-stack [systems].
Scott SinclairPrincipal Analyst, Omdia
"The expectation is that the larger and more sophisticated organizations lead the way," Sinclair said. "It's about which vendor can provide the right mix of technology, services and expertise to get the organization where they want to go."
Scott Sinclair
Leone wrote he expects smaller enterprises to gradually adopt the AI factory model over time, and with it, pieces of Nvidia's platform.
"The same pressures that pulled [current buyers] in, like rising token bills and agents that need access to sensitive data, will eventually hit everyone else too," Leone wrote. "Most will rent first, then bring smaller pieces in-house, like the Vera CPU servers for running agents, or use co-location facilities that already have the power and liquid cooling."
Dell and HPE both announced dedicated Vera CPU racks at GTC Berlin. Vendors claim CPUs are best suited for the work that agents generate beyond model inference and that they are generally more efficient than using GPUs everywhere, but it's too early to treat these claims as fact, said Stephen Sopko, practice lead for semiconductor and deep tech at HyperFrame Research.
To determine whether AI factories are working as intended, Leone recommended monitoring the number of agents moving from pilots to production, whether cluster utilization remains high, and whether customers continue to expand their deployments. Adoption will also need to spread beyond regulated industries and early adopters, and AI factories will need to become part of routine operations.
"When a company's regular IT ops team runs this day-to-day, instead of a dedicated AI team, it becomes enterprise infrastructure," Leone wrote.
Ben Lutkevich is an award-winning writer for TechTarget.com
Mike Leone
Torsten Volk
Scott Sinclair
