Nvidia, IBM and the race for enterprise intelligence

Artificial intelligence companies are racing to control the infrastructure, data and software layers that will power enterprise intelligence.

Nvidia Corp. remains far ahead in accelerated computing, but AMD Inc., Broadcom Inc. and other challengers are positioning themselves for a market in which demand may support multiple winners. At the same time, IBM Corp. is under pressure to prove that its software, data and hybrid cloud assets can secure a place in the emerging AI stack.

The next competitive advantage will come from turning proprietary data and domain expertise into an internal system capable of reasoning and executing work, according to John Furrier (pictured, left), executive analyst at theCUBE Research.

“The moat of a business in the future is the brain of the company, the data that they have,” Furrier said. “Can it be organized in a way to reason — system of intelligence, systems of execution, systems of agency? Can they reason?”

On the latest episode of theCUBE Pod, Furrier and Dave Vellante (right), chief analyst at theCUBE Research, discussed IBM’s execution challenges, Nvidia’s infrastructure lead and the growing opportunities for AMD and Broadcom. They also examined open versus closed AI models, token economics and the race to build the enterprise system of intelligence.

IBM faces an AI execution test

IBM’s sharp stock decline prompted speculation that enterprise AI adoption is progressing more slowly than expected. IBM’s performance says more about the company’s portfolio and execution than the health of the broader AI market, according to Furrier and Vellante.

Much of today’s infrastructure spending is flowing toward hyperscalers, neocloud providers and chipmakers. IBM’s mainframe and infrastructure businesses are not as closely aligned with that buildout, while its newer AI offerings have yet to grow enough to offset pressure on older products.

“Their infrastructure business is not aligned with the wave in the industry. They’ve got mainframes,” Vellante said. “What always happens in these waves is the new is not big enough to offset the decline in the old, and I think that’s what’s happening with IBM.”

IBM has many of the assets needed to participate in enterprise AI, including Red Hat, watsonx, governance software and a broad data portfolio. The missing piece is a unified system of intelligence that connects data, context, reasoning and applications. Databricks Inc., Snowflake Inc. and the hyperscalers are already competing for that position. IBM has the components, but it has not assembled them into a platform that commands the same attention, according to Vellante.

“They have all the ingredients, but they’re not putting them together that way,” he said. “Their software portfolio could be. They just need a little bit better focus, in my opinion.”

Furrier dismissed calls to break up IBM, describing its problem primarily as one of execution. Enterprises may also be delaying software decisions while they rebuild data pipelines and prepare infrastructure for agentic workloads.

“I think there is nothing wrong with IBM other than execution,” he said. “I think they could have been more data specific in that system of intelligence. They had all the piece parts. That’s just focus, right?”

Nvidia’s lead leaves room for challengers

Nvidia remains the dominant force in accelerated computing, supported by its graphics processing units, networking, software and rack-scale architecture. Vellante expects the company to retain between 75% and 80% of the AI accelerated computing market. That dominance does not eliminate opportunities for AMD and Broadcom. Demand for AI infrastructure is growing quickly enough that competitors can build large businesses without displacing Nvidia.

AMD has expanded beyond central processing units through graphics accelerators, networking and systems capabilities. Its acquisitions of Xilinx Inc., Pensando Systems Inc. and ZT Systems have helped transform the company into a broader infrastructure provider.

“[AMD CEO Lisa Su] is basically compressing 20 years of ecosystem development by Nvidia, and she’s compressing that into five years of capital allocation,” Vellante said. “It’s actually remarkable what she’s done when you think about that, because she recognizes she’s got to move at the speed of Nvidia.”

Broadcom is pursuing a different opportunity through custom silicon and networking. Furrier expects more alternatives to emerge as enterprises seek smaller, less expensive inference systems that can operate outside hyperscale data centers.

“Everything that AMD and Nvidia makes will sell; Nvidia specifically because they are the leader,” Furrier said. “But the enterprise, they have to start thinking about the budget for the AI infrastructure because they don’t have the big bucks.”

Companies race to build enterprise intelligence

The greatest value may ultimately sit above the infrastructure layer. Companies will need to connect proprietary information, domain knowledge, models and applications into systems capable of reasoning and acting.

That enterprise brain will require relational, vector and graph databases, governed data pipelines and specialized models working alongside general-purpose models. Furrier questioned why companies would outsource all of that intelligence to a small number of model providers.

“If you’re a company and you have to build the next 20-year competitive advantage, you’ve got to build the company brain, the intellect, intelligence, cognition for the company,” he said.

OpenAI, Anthropic PBC and other frontier model companies will need to expand deeper into software and build ecosystems around specialized models, according to Furrier and Vellante. Open-weight models will also increase competition, though they may not match frontier systems in both capability and efficiency.

Enterprises are meanwhile becoming more disciplined about AI costs. The focus is shifting from consuming the largest number of tokens to measuring how efficiently AI produces a useful result. Companies may begin using token-to-value ratios to connect model consumption with measurable outcomes, according to Furrier. That shift could create an AI version of financial operations, with organizations monitoring spending, performance and return in real time.

“Right now, in real time, you can actually peg the value to the outcome and be like, OK, you built an app, everyone’s using it. That’s good. Or you built an app and no one’s using it,” Furrier said.

Human workers will remain involved, but their roles will change. Rather than completing every task themselves, people will manage context, validate outputs and supervise AI systems. The winners will be the companies that combine infrastructure, intelligence and human judgment into a reliable operating model, according to Furrier and Vellante.

The conversation also previewed theCUBE’s upcoming coverage of AMD’s Advancing AI and Neo4j’s GraphTalk events. Stay tuned for more reporting and interviews as theCUBE continues tracking the infrastructure, data and software layers shaping enterprise AI.

Here’s the full episode of this week’s theCUBE Pod:

Image: SiliconANGLE