As artificial intelligence transitions from training-centric models to large-scale inference, agentic AI, and physical AI applications, enterprise infrastructure is undergoing a fundamental transformation. Organisations are increasingly seeking scalable, energy-efficient, and cost-effective architectures that support complex AI workloads across data centres, cloud environments, edge deployments, and intelligent devices.

In response to this shift, Intel is advancing a chip-to-system strategy that integrates processors, accelerators, networking, software, and rack-scale infrastructure into a unified AI platform. The company believes that emerging workloads such as agentic AI, where autonomous systems reason, plan, and act and physical AI applications in robotics and industrial automation, will redefine infrastructure requirements and elevate the role of CPUs as critical orchestration engines.

In this exclusive interaction with AI Spectrum, Akshay Kamath, Director – Enterprise Client and Data Center Category, Intel India, shares his perspectives on the evolution of enterprise AI infrastructure, the growing importance of inference-driven computing, the role of ecosystem partnerships in accelerating AI adoption, and Intel's vision for enabling AI across the continuum from cloud and data centers to edge devices and real-world physical AI deployments.

What are the key market shifts driving Intel’s broader chip-to-system AI strategy?

The AI market is pivoting now from training to inference, and the infrastructure implications are significant. Industry projections estimate AI inference to account for nearly 40% of all data centre power demand by 2030. The traditional 1:8 CPU-to-GPU ratio built for training workloads is expected to move towards a 1:1 ratio for agentic AI deployments. That is structural reorientation, not an incremental upgrade.

Agentic AI is the primary driver of this shift to inference. Unlike prompt-response workloads, agentic systems reason, plan, act, and reflect, continuously orchestrating data movement, memory management, and coordination across multiple models and services. This makes the CPU central again as the orchestration layer, and raises the bar for processors, networking, accelerators, and the software stack to function as a unified system.

Physical AI adds a third dimension — robotics, autonomous machines, and industrial deployments where real-time action, edge processing, and cost efficiency all matter simultaneously. Processors built on Intel's latest 18A process node are already featured across 130+ edge designs and 100,000+ active deployments across manufacturing, robotics, and retail, with TCO improvements of up to 67 per cent in humanoid robotics applications.

Intel's chip-to-system strategy addresses all three tiers. Integrated solutions spanning processors, networking, accelerators, software, and rack-scale infrastructure — deployed across cloud, data centre, edge, and device — and built with an open ecosystem to ensure system-level AI isn't constrained by proprietary stacks.

How does Intel see enterprise AI infrastructure evolving over the next few years, particularly regarding scalability and efficiency for inference and agentic AI?

Enterprise AI infrastructure is shifting around inference, not training — and the architecture requirements are materially different. Training demanded GPU density and parallelism. Inference and agentic AI demand high concurrency, low latency, and cost-efficient performance at scale. These are system-design problems as much as silicon problems.

The CPU is being revalued in this shift. Already today, a large share of Enterprise AI inference workloads run efficiently on AI-optimised CPUs, especially Intel Xeon with AMX - delivering predictable cost and lower energy use. Agentic workloads — where systems reason, plan, and iterate across multiple models — require continuous orchestration that GPUs alone aren't designed to manage. Intel projects the CPU-to-GPU ratio moving from 1:8 towards 1:1 for agentic inference deployments.

At the infrastructure level, Intel sees enterprises adopting disaggregated, rack-scale architectures where compute, networking, and acceleration are independently optimised. Intel's Xeon 6+ processors — the ficentreta centre CPUs built on the Intel 18A process node — are designed to improve efficiency for agentic AI with ultra-high-density cores optimised for orchestration and allowing data centres to consolidate multiple legacy servers and increase compute capacity without using more power or physical space with up to 9:1 server consolidation compared to 2nd Gen Intel Xeon processors.

The broader direction is open and composable. As enterprise AI moves from pilot to production, proprietary stacks become operational liabilities — and Intel's platform approach is built around avoiding that lock-in.

How does the new rack-scale AI infrastructure, combining Xeon processors with SambaNova RDUs, help enterprises address the demands of large-scale workloads?

The combination addresses a core tension in enterprise AI infrastructure: inference at scale requires both high-throughput acceleration and intelligent orchestration — and no single processor type handles both well.

In Intel's rack-scale architecture, Xeon processors manage orchestration, workflow coordination, and data movement, while SambaNova SN-50 Reconfigurable Dataflow Units accelerate AI inference. The result is a production-ready system optimised for large-scale inference and agentic AI — delivering higher performance per watt and better cost efficiency than GPU-heavy alternatives built primarily for training workloads.

Foxconn's system integration further compresses the gap between silicon availability and enterprise deployment, bringing validated, production-ready rack configurations to market faster than customers could assemble independently.

Future CPU-dense variants, without additional acceleration, are also planned for orchestration-heavy workloads that don't require full inference throughput, giving enterprises the flexibility to right-size infrastructure rather than over-provision.

How important are ecosystem partnerships, such as those with Foxconn and Siemens, in accelerating real-world AI adoption?

Silicon is the starting point, not the finish line. Real-world AI adoption stalls at the gap between validated hardware and production deployment, and that's precisely where ecosystem partnerships matter most.

Intel's partnerships with companies such as Foxconn and Siemens demonstrate this approach. Foxconn brings deep system integration capabilities and expertise in infrastructure deployment, helping accelerate the rollout of rack-scale AI solutions. Siemens contributes expertise across industrial automation, manufacturing, digitalisation, robotics, and lifecycle management, enabling AI solutions tailored to real-world industrial requirements.

By combining Intel's silicon and platform technologies with partners' domain expertise, these collaborations help customers move from AI experimentation to production-scale deployments faster. They also enable industry-specific solutions that address unique requirements in manufacturing, healthcare, infrastructure, robotics, and other vertical markets.

What specific advantages will the Intel Xeon 6+ processors bring to data centres managing complex AI applications?

Intel Xeon 6+ processors are purpose-built for next-generation AI infrastructure and bring several key advantages for data centres running complex AI applications.

Built on the Intel 18A manufacturing process, Xeon 6+ is engineered for high-density, scale-out deployments that support cloud-native, agentic AI, and network-intensive workloads. The platform delivers greater performance density, improved power efficiency, and enhanced operational scale for AI environments.

Key benefits include support for large-scale agent hosting, high memory bandwidth through 12-channel DDR5, extensive PCIe Gen 5 and CXL connectivity for accelerated data movement, and getting real-time insights and application-level visibility into CPU energy use and activity with Intel Application Energy Telemetry (Intel AET). Xeon 6+ is also designed to handle the orchestration, concurrency, and data movement demands of agentic AI while enabling enterprises to scale infrastructure without requiring disruptive data centre redesigns. Together, these capabilities make Xeon 6+ a strong foundation for orchestration and emerging agent-based workloads.

How is Intel positioning itself to support the expansion of AI workloads from cloud to edge devices and physical AI applications?

Intel is taking a full-stack approach to AI, spanning cloud, data centre, edge, client devices, robotics, and physical AI systems. The company believes AI will increasingly become hybrid, operating across distributed environments rather than being confined to centralised cloud infrastructure.

At the data centre level, Intel offers Xeon processors, networking solutions, AI accelerators, and rack-scale infrastructure optimised for inference and agentic AI. At the edge, Intel's Core Ultra Series 3 platform is enabling AI deployments across manufacturing, robotics, retail, transportation, and smart city applications.

Intel is also expanding collaborations with ecosystem partners to develop industry-specific AI solutions, including robotics, industrial automation, healthcare, and physical AI use cases. By leveraging common architecture and software frameworks across cloud, edge, and endpoint environments, Intel aims to help customers deploy AI wherever data is generated and decisions need to be made.

This chip-to-system strategy allows Intel to support AI workloads across the entire continuum—from cloud-based inference and agent orchestration to real-time edge intelligence and physical AI applications.