Intel has announced a broad push to optimize its data center infrastructure for agentic AI, pairing new Xeon-based rack designs with its GPU D-series and a software strategy intended to tie CPUs, GPUs, and open models into a single deployment path.
The announcements came at the Open Compute Project (OCP) Global Summit. The company’s main argument is that agentic AI changes inference from a single model call into a multi-step workflow involving retrieval, reasoning, orchestration, tool use, and execution. That, Intel says, makes general-purpose CPUs more important because they coordinate workloads, move data, and keep accelerators busy.
“Open models are expanding the options enterprises have for deploying AI,” said Nicolas Dubé, corporate vice president of the Data Center Systems and Solutions Group at Intel, in a statement. “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.’”
At the OCP show, Intel showcased Xeon-based rack designs developed with ecosystem partners, many based on the OCP’s Open Rack v3 standard. Select Xeon 6+ configurations can scale beyond 46,000 cores per rack at 133.45 kW of total rack power, a configuration Intel frames as evidence of rising CPU density for agentic workloads.
Rather than measuring systems solely by raw throughput, Intel says AI infrastructure should be evaluated by how many concurrent agents a rack can sustain at a target latency. In that view, Xeon’s role is to manage I/O, execute tools, and maintain data flow so GPUs and other accelerators remain productive.
Intel also disclosed the specs of its GPU D-series, codenamed Crescent Island. The accelerator is built on Intel’s Xe3P architecture and is specified with up to 480 GB of LPDDR5X memory, 1.5 TB/s of memory bandwidth, and 1.3 petaflops of FP8 compute within a 350 W power envelope.
Intel is positioning the part around memory-heavy, sustained inference rather than peak training performance. It says a single card can support models with up to 300 billion parameters, while an eight-GPU system can address models with more than one trillion parameters.
In addition to the hardware announcements, Intel highlighted software initiatives designed to improve portability across its CPU and GPU platforms.
The company is integrating its hardware with established AI frameworks, inference runtimes and development tools, while using a common Xe software foundation to simplify application deployment across Intel CPU and GPU products.
On the hardware side, Intel is contributing to the OCP Open Adapter Card workgroup, which is developing a standardized sub-rack accelerator-card form factor intended to improve mechanical, electrical and management interoperability across system designs. Intel said it has made 116 contributions to OCP since 2011.
