Intel Unveils Data Center AI Inference GPU "D-series" Specs and New Xeon Rack Designs, Targeting Enterprise AI Deployment

What Intel announced

Intel says that at the Open Compute Project (OCP) Global Summit on October 12, 2026, it did three things: shared new technical details of the Intel GPU D-series (formerly code-named Crescent Island), introduced new Xeon rack designs, and highlighted open software and ecosystem innovation. OCP is an industry initiative focused on open data center hardware standards. A rack is the standard cabinet in a data center that holds multiple servers.

"Inference" means using a trained AI model to actually answer questions and generate content. Intel points out that agentic AI (AI that completes tasks in steps on its own) turns inference into a multi-step process spanning retrieval, reasoning, orchestration, tool use and execution. Intel also says power limits, memory bottlenecks and cost pressure are making AI scaling an increasingly system-level challenge.

Intel GPU D-series specifications

Intel says the GPU is built on the Xe3P architecture and combines up to 480 GB of LPDDR5X memory, 1.5 TB/s of memory bandwidth and 1.3 petaflops of FP8 compute, with a 350-watt power rating. FP8 is a low-precision number format commonly used in AI computing, and petaflops is a unit of quadrillions of floating-point operations per second. Intel also explains that its own-brand PCIe card carries 160GB of LPDDR5x; partners can build their own-brand cards on an ODM (original design manufacturer) basis, with memory options up to 480GB. So 480 GB is not the configuration of every card.

Intel says the product is designed for large language models, long-context workloads and concurrent inference serving, with a single card supporting models of up to 300 billion parameters and an eight-GPU system supporting more than one trillion parameters. Intel also says it uses a standard PCIe form factor (the expansion card slot format common in ordinary servers) and air cooling (fan airflow, no liquid cooling needed), along with a common Xe software foundation, aiming to simplify deployment in existing data centers.

Xeon racks and open software

Intel says many of the new Xeon rack designs it developed with ecosystem partners use OCP's Open Rack v3 (ORv3) standard. Some Xeon 6+ configurations scale to more than 46,000 cores per rack, with a total rack power of 133.45 kW. Intel says Xeon processors orchestrate workloads, execute tasks and manage data movement to keep accelerators such as GPUs busy.

On software, Intel says it provides Day 0 support for Meta's Muse Glimmer, meaning a new model can run on its hardware the day it launches. The Intel AI for Enterprise Agent Toolkit is now optimized for Xeon and deployable, with an architecture designed to extend to Intel GPUs. Intel also says it has made 116 contributions to OCP since 2011, and this year showcased its contribution to the Open Adapter Card workgroup, which aims to define a standardized sub-rack accelerator card form factor (a card format that fits inside a rack's systems).

What this means for enterprises

This news mostly concerns data center hardware planning; ordinary consumers won't use it directly. According to Intel, the goal is to make it easier for enterprises to scale inference in existing data centers, reuse existing investments and keep infrastructure choice.

Intel says customer sampling is under way; the press release gives no official launch timeline and no independent benchmark data.

Browse the latest news in this topic

• Tech news