IBM faces AI software test as Nvidia keeps infrastructure lead

theCUBE Research analysts John Furrier and Dave Vellante used the latest episode of theCUBE Pod to frame enterprise AI as a fight over infrastructure, proprietary data and software control. No financing round, valuation, revenue figure or customer count was disclosed, but Vellante estimated Nvidia could hold 75% to 80% of the AI accelerated computing market.

The discussion centered on whether IBM can turn Red Hat, watsonx, governance tools and its data portfolio into a credible enterprise AI platform. Furrier and Vellante said IBM’s recent stock pressure should be read less as evidence of weak enterprise AI demand and more as an execution problem inside IBM’s portfolio.

IBM has parts, but not the system

Vellante said much of the current AI infrastructure spend is going to hyperscalers, neocloud providers and chipmakers rather than to IBM’s traditional infrastructure businesses. He pointed to the company’s mainframe exposure and older product lines as a drag while newer AI products remain too small to carry the business.

The analysts argued that IBM’s issue is not a lack of components. Red Hat gives IBM a hybrid cloud position, watsonx gives it an AI product line, and its governance and data tools fit enterprise requirements. The missing piece, in their view, is a tightly packaged system that connects data, business context, reasoning and applications in a way buyers can understand.

That position is already being contested by Databricks, Snowflake and the hyperscalers, according to Vellante. His assessment was that IBM has not assembled its software portfolio into a platform with comparable market attention. Furrier rejected the idea that IBM’s answer is a breakup, describing the problem as focus and execution rather than structure.

Nvidia’s lead still leaves a market for others

On infrastructure, the analysts described Nvidia as the clear leader because of its position across GPUs, networking, software and rack-scale systems. Vellante’s 75% to 80% share estimate leaves Nvidia dominant, but not alone.

AMD and Broadcom were identified as credible beneficiaries of demand that is expanding beyond what a single supplier can serve. Vellante said AMD Chief Executive Lisa Su has used acquisitions, including Xilinx, Pensando Systems and ZT Systems, to compress years of infrastructure ecosystem building into a much shorter period.

Broadcom’s opportunity is different, according to the discussion: custom silicon and networking. Furrier said enterprises will also need smaller and less expensive inference systems outside hyperscale data centers, where budget constraints are more visible than they are at the largest cloud providers.

The software layer may carry more value

Furrier and Vellante said the larger strategic prize may sit above compute. Their argument is that companies will try to turn proprietary data, domain knowledge, models and applications into internal systems that can reason and take action.

That requires more than a frontier model API. Furrier said enterprises will need combinations of relational, vector and graph databases, governed data pipelines, specialized models and general-purpose models. He questioned whether companies should outsource the full intelligence layer to a small set of model providers.

The analysts also said AI spending discipline is likely to move from raw token consumption toward measuring output value. Furrier described a potential token-to-value model in which companies monitor whether AI usage produces results that employees or customers use. That would put AI consumption closer to a real-time cost and performance discipline, rather than a broad experimental budget.

Human roles would shift under that model, according to the discussion. Workers would spend more time setting context, checking outputs and supervising AI systems, while vendors compete to own the infrastructure and software layers underneath that operating model.

This story draws on original reporting from SiliconANGLE.