Intel Pushes Open AI Infrastructure for Agentic Era

Overview

Intel is accelerating the rollout of open infrastructure solutions tailored for the emerging agentic AI era, emphasizing the need for heterogeneous computing that balances cost, power, and performance. At the recent Open Compute Project (OCP) Global Summit, the tech giant unveiled new technical specifications for its Intel GPU D-series and showcased next-generation Intel Xeon-based rack designs aimed at supporting complex, multi-step AI workflows.

The company argues that as open models reshape the economics of artificial intelligence, enterprises require more flexible infrastructure to handle the escalating demands of inference workloads. By combining general-purpose compute with purpose-built acceleration and open software standards, Intel aims to provide turnkey solutions that simplify deployment while optimizing efficiency across the entire AI lifecycle.

Redefining Agentic AI Infrastructure

Agentic AI represents a significant shift from static model generation to dynamic, multi-step workflows involving retrieval, reasoning, orchestration, tool use, and execution. This evolution places substantial new demands on general-purpose compute resources, particularly x86 CPUs, which must manage data movement, execute tasks, and ensure accelerators remain productive.

Nicolas Dubé, corporate vice president of Intel’s Data Center Systems and Solutions Group, highlighted the strategic importance of this transition. He noted that open models are expanding enterprise options for deploying AI, but realizing their full potential requires infrastructure capable of matching specific workloads with the appropriate compute resources and economics.

“Open models are expanding the options enterprises have for deploying AI. To realize their full potential, organizations need infrastructure that can match different workloads with the right intelligence, right compute resources and right economics. That's what we mean by 'put every prompt in its place.'”

This philosophy underpins Intel’s approach to open, heterogeneous AI infrastructure. Rather than relying solely on specialized accelerators, the strategy integrates CPUs, GPUs, and other accelerators through interoperable software. This allows organizations to assign each task to the most efficient resource, improving overall utilization and reducing the bottlenecks associated with memory constraints and power limitations.

Xeon Processors as the Orchestration Foundation

At the heart of Intel’s agentic AI strategy is the Intel Xeon processor, which serves as the orchestration layer for complex workflows. As these workflows grow in complexity, the CPU must handle heavier loads related to input/output management and task execution to keep downstream accelerators fed with data.

Intel showcased new Xeon-based rack designs developed in collaboration with ecosystem partners, many of which adhere to the Open Rack v3 (ORv3) standard from the Open Compute Project. These designs provide customers with deployment-ready infrastructure that simplifies integration within existing open data center environments.

Key technical highlights for the updated Xeon configurations include:

• Massive Core Density: Select configurations scale beyond 46,000 cores per rack.
• High Power Capacity: The systems support up to 133.45 kW of total rack power.
• Enhanced Throughput: The increased CPU compute density supports the concurrent execution of numerous AI agents.

Performance in this context is increasingly measured by the number of concurrent agents a rack can sustain at target latency. By leveraging two decades of rack-scale innovation and a broad x86 software ecosystem, Xeon processors ensure that data flows efficiently and tools execute without delay, maximizing the delivered work across the entire system.

Intel GPU D-Series: Optimizing Memory-Intensive Inference

While CPUs handle orchestration, the Intel GPU D-series, formerly code-named Crescent Island, is engineered specifically for high-efficiency, memory-intensive AI inference. Built on Intel’s Xe3P architecture, the GPU addresses the growing need for economical deployment of large language models and long-context workloads.

The hardware specifications represent a significant focus on memory capacity and bandwidth, which are critical for sustaining large models. Key features include:

1. Up to 480 GB of LPDDR5X Memory: Providing the vast capacity needed for large model parameters.
2. 1.5 TB/s of Memory Bandwidth: Ensuring rapid data access to prevent compute bottlenecks.
3. 1.3 Petaflops of FP8 Compute: Delivering high throughput for inference tasks.
4. 350-Watt Power Envelope: Maintaining energy efficiency even at high performance levels.

Designed for applications ranging from multi-tenant AI services to latency-sensitive robotics, the GPU supports models with up to 300 billion parameters on a single card and over one trillion parameters in an eight-GPU system. Its standard PCIe form factor and air-cooled design allow it to integrate seamlessly into existing data centers without requiring specialized cooling or rack infrastructure.

Anil Nanduri, vice president of AI Products and Go-to-Market at Intel’s Data Center Group, emphasized the practical benefits of this design. He stated that inference is becoming the defining AI workload, and enterprises need infrastructure that delivers the right performance at the right cost. The goal with the D-series is to make high-capacity inference both practical to deploy and economical to scale.

Ecosystem Collaboration and Deployment

Intel is not pursuing these hardware advancements in isolation. The company is actively collaborating with a broad ecosystem comprising OEMs, ODMs, neocloud providers, independent software vendors, appliance providers, and AI-native companies.

These partnerships focus on creating validated platforms and deployment-ready solutions that reduce complexity for end-users. By ensuring software integration and compatibility across the stack, Intel aims to help customers move more quickly from model evaluation to production deployment.

This collaborative approach aligns with the broader industry trend toward open standards, allowing organizations to build flexible AI infrastructure that can adapt to rapidly evolving model architectures and workload requirements.