Intel Urges Enterprises to Put AI ROI Ahead of Hardware Specs

Key Points

• Intel says enterprises should prioritize AI ROI over hardware specifications by defining the business problem, expected productivity gains and cost economics before selecting infrastructure.
• AI workloads vary significantly: interactive applications require low latency, while batch tasks can use less expensive resources. Agentic AI also depends on orchestration, security, memory, storage, networking and data access—not just token-generation speed.
• Enterprises can improve returns through observability, prompt optimization, caching, efficient agent loops and matching model costs to specific tasks. Intel is positioning its CPUs, GPUs and planned Crescent Island accelerator for heterogeneous, cost- and energy-efficient AI deployments across on-premises, cloud and hybrid environments.

Enterprise AI infrastructure decisions should be centered on return on investment rather than hardware specifications alone, according to Anil Nanduri, vice president of AI Products and Go-To-Market for Intel NASDAQ: INTC Data Center.

Speaking during The Six Five Summit: AI Unleashed 2026, Nanduri said organizations moving from AI pilots into production need to identify the business problem they intend to solve, the expected productivity gains and the cost economics before selecting infrastructure.

“A lot of our customers, they start thinking about hardware first,” Nanduri said. “I think we really have to change the conversation to ROI.”

Token efficiency and workload requirements

Nanduri said enterprises should not treat all AI tokens as equal. Some use cases demand low-latency, interactive responses, such as credit-card fraud protection, while other applications, including audit-report generation, can be handled in batch mode with less urgency.

Rather than pursuing the highest possible token output, companies should focus on using tokens efficiently to achieve a business result, he said. That distinction becomes more significant as companies deploy agentic AI systems designed to plan, coordinate and execute tasks rather than simply generate responses.

Agentic workloads require more than token generation, Nanduri said. They also require systems for code execution, testing, verification, sandboxing, orchestration, security and access to enterprise data. CPUs can play a role in execution, verification and control-plane functions, while GPUs and other accelerators can support different token-generation requirements.

“The wall clock time has shifted from just the AI running to actually waiting on answers,” Nanduri said, referring to the time required for an end-to-end workflow to access data and interact with applications across an enterprise.

He said organizations should assess complete system throughput rather than evaluate a single component in isolation. Agentic applications may put pressure on memory, storage, databases, networking and application programming interfaces as they retrieve information from separate financial, supply-chain and customer systems.

Measuring AI business value

Nanduri said observability is important for managing AI costs and determining ROI. He cited comments from Uber executives regarding efforts to reduce AI implementation costs while expanding AI use, describing an approach that included improving operational efficiency, providing users visibility into token consumption and determining where expensive portions of an AI workflow reside.

According to Nanduri, enterprises can reduce costs by improving prompts, context caching and agent loops, allowing them to reach the same result with fewer tokens. They can also match model choices to specific portions of a workload.

He said frontier models may be appropriate for some tasks but can be expensive, while open-weight models may be sufficient for other work. A company could use lower-cost models for a large portion of a workflow and reserve more expensive AI capabilities for tasks that require them, he said.

Nanduri expects AI deployments to develop into hybrid environments, similar to the mix of on-premises and cloud infrastructure that emerged over time. Those environments could include local systems, lower-cost cloud services and more expensive cloud resources depending on workload needs.

Infrastructure layers and data strategy

For enterprise planning, Nanduri described several infrastructure layers that need to be balanced:

• Control plane: Orchestration, model routing, agent routing and security controls, typically supported by CPUs.
• Agent or data plane: Token generation, task execution, sandboxing and verification, using CPUs, GPUs or other compute architectures based on performance and cost needs.
• Network plane: Connectivity across chips, servers, racks and data centers as models and workloads expand.
• Storage and data systems: Databases, KV cache, context-memory pooling and access to enterprise information.

Investments need to be aligned across those layers, Nanduri said, because a weakness in orchestration, storage or networking can limit the benefits of token-generation improvements.

He also characterized enterprise data protection as a central strategic issue. As AI models seek more data for training and refinement, companies must decide whether to send their data to external AI services or bring AI tools to where their data resides, he said.

Nanduri said enterprises can use open-weight models, fine-tuning and model distillation to create smaller, domain-specific models suited to their own data. In his view, companies that protect their data while deploying AI capabilities around it will be better positioned for business success.

Intel’s approach

Nanduri said Intel’s strategy is based on heterogeneous computing, combining CPUs, GPUs and other compute elements for different AI tasks. He pointed to Intel’s Xeon processors, its Arc GPU lineup and its planned Crescent Island product as components of that approach.

He said Crescent Island was designed around cost and energy efficiency, using low-power DDR memory rather than GDDR or HBM memory. The product is intended to serve enterprise workloads that do not require the most latency-sensitive AI tokens and can be deployed through PCIe cards, he said.

Nanduri also highlighted Intel’s partnership with SambaNova Systems for high-throughput, low-latency token generation for interactive workloads. Intel’s broader objective, he said, is to offer building blocks that help customers optimize total cost of ownership and business outcomes across on-premises, cloud and hybrid deployments.

About Intel (NASDAQ:INTC)

Intel Corporation, founded in 1968 by Robert Noyce and Gordon E. Moore and headquartered in Santa Clara, California, is a leading global designer and manufacturer of semiconductor products. The company is historically notable for introducing the first commercial microprocessor and for driving the x86 architecture that underpins many personal computers and servers. Intel's core business spans the design, fabrication and marketing of processors, chipsets and related components for a wide range of computing applications.

Intel's product portfolio includes client and mobile processors marketed under brands such as Intel Core and Pentium, as well as high-performance Xeon processors for data centers and cloud infrastructure.

This instant news alert was generated by narrative science technology and financial data from MarketBeat in order to provide readers with the fastest reporting and unbiased coverage. Please send any questions or comments about this story to contact@marketbeat.com.

Continue following MarketBeat

Add MarketBeat as your preferred source on Google to see our latest stories in your feed.

Add As Preferred Source