theCUBE Pod: AI skeptics get louder as IBM bets on commodity models and Marvell chases inference

Furrier, a research executive analyst at theCUBE Research, called the public backlash "an AI freakout moment," saying the view from inside supplier companies looks different from the sentiment outside. "This new AI layer is a system; the data center as a system is now the computer," he said, pointing to rack-scale changes and to chip work at Marvell, Broadcom and elsewhere aimed at efficiency.

Much of the discussion centered on takeaways from IBM's recent Analyst Council meeting. Vellante, chief analyst at theCUBE Research, said customers are spending first on the infrastructure and models needed to produce intelligence, and only then turning to IBM to put that intelligence to work. "The alpha is going to be in your data, your processes and your operating model," he said, adding that the competitive advantage will not come from Palantir or from models themselves.

IBM has promoted an architecture built on agents, data, automation and hybrid infrastructure, with an orchestration layer that can integrate agents from different vendors. Rather than relying only on proprietary models, the company is positioning itself around Red Hat's open-source foundation. "Their bet is the commoditization of the models," Furrier said. "Having a nice model management capability will be good for IBM; but when you have a commoditization bet, you better be right." He added that IBM does not need its own proprietary operating system because Red Hat gives it an abstraction control layer around open infrastructure, and that running all the agents could be an opportunity.

Vellante said customers trust IBM but do not see it as an innovator in the current AI landscape. If the company can establish itself as a provider of the operational infrastructure around models and agents, he said, it could own "the scarcity around the abundance." He added that IBM has "got to play in both camps" and needs an IBM version of Oracle's Red Stack. Asked last year what IBM's platform is, chief commercial officer Rob Thomas answered Red Hat, which Vellante agreed with, though he said it remains unclear what the surrounding solution looks like even if the pieces are in place.

On the chip side, Marvell has set a revenue target of $70 billion to $90 billion for fiscal 2031. The maker of custom AI processors and optical networking technology announced a partnership this year with Nvidia's NVLink Fusion. "What impressed me about Marvell is that they had a 10-year vision," Furrier said. "They made the argument that AI research is becoming a connectivity problem as much as a compute problem. And they're not wrong."

Furrier said he does not believe the AI market is approaching a bubble, and Vellante said the breaking point is farther off than many expect. Vellante said that a year ago about 90% of workloads at CoreWeave were training, while today the split is roughly 60-40 between training and inference. Marvell and the neoclouds stand to gain as spending on AI infrastructure grows.