Intel Details GPU D-Series and Open Racks for Agentic AI
Intel's OCP update pairs a new inference-focused GPU design with high-density Xeon racks, but sampling status and vendor-reported specifications leave deployment evidence incomplete.
Edited by Tyronne Panaino
Intel disclosed new technical details for its GPU D-series and showed Xeon-based rack designs at the Open Compute Project Global Summit on October 12. The update matters to infrastructure teams planning inference capacity because Intel is presenting the GPU, CPUs, software and open rack standards as one heterogeneous system rather than as isolated components.
The affected audience is narrower than every enterprise using an AI assistant. It is aimed at cloud operators, system builders and large organizations deciding how to place retrieval, orchestration, model inference and tool execution across CPUs and accelerators. The hardware is not yet evidence of broad production availability: Intel says customer sampling is underway.
What Intel changed in the D-series picture
Intel previously used the code name Crescent Island for GPU D-series. Its new description identifies the product as an inference-focused design built on the Xe3P architecture, with a standard PCIe form factor, air cooling and a common Xe software foundation. Intel says the design can support memory-intensive inference without requiring a specialized liquid-cooled rack.
The specification needs an important qualification. Intel lists design ceilings of up to 480GB of LPDDR5X memory, 1.5TB per second of memory bandwidth, 1.3 petaflops of FP8 compute and a 350-watt power envelope. Its small print says the Intel-branded PCIe card has 160GB, while the design allows partners to build variants with as much as 480GB. Treating 480GB as the configuration of every card would overstate what Intel announced.
Intel also says a single card is designed for models with as many as 300 billion parameters, while an eight-GPU system is intended to support models above one trillion parameters. Those are vendor-stated capacity targets, not independent workload results. The announcement does not provide a shipping date, public price, third-party benchmark or deployment-scale reliability record.
The rack design connects CPUs and accelerators
At the rack level, Intel showed systems built around Xeon processors and said many use the Open Compute Project's Open Rack v3 standard. Select Xeon 6+ configurations go beyond 46,000 CPU cores per rack at 133.45 kilowatts of total rack power, according to Intel. The company's argument is that agent workflows need general-purpose compute for orchestration, data movement, input and output, and tool execution even when a GPU performs the model inference.
That distinction is useful for buyers comparing accelerator counts alone. A multi-step agent service can stall outside the model call, so the surrounding CPU, memory, network and software layers shape how many concurrent tasks a rack can sustain. Intel has not supplied an independently reproduced concurrency result in this source, however, so the configuration figures should be read as design information rather than proof of delivered application throughput.
Open software and hardware are part of the pitch
Intel says D-series uses the same Xe software base as its other GPU platforms and that it is working through upstream AI frameworks, inference runtimes and deployment environments. The stated goal is application portability across Intel CPUs and GPUs, including support for new open models as they appear.
The company also described work with the Open Compute Project's Open Adapter Card group on a standardized sub-rack accelerator form factor. That effort concerns mechanical, electrical and management interoperability. If the specification and products mature together, system builders could have more freedom to combine accelerators with existing infrastructure. The current announcement does not show how many partners will ship compatible cards or when those systems will be generally available.
What to watch next
The next verifiable checkpoints are final product configurations, partner systems, pricing, general-availability dates and results from workloads that test long-context inference and concurrent agent execution. It will also matter whether a 160GB Intel-branded card and partner-built higher-memory variants share the same performance, power and software behavior.
For now, the material change is architectural clarity: Intel has connected its new inference GPU, dense Xeon racks, common software base and open-hardware work into one agentic-AI infrastructure plan. The evidence is still first-party and pre-general-availability, so it supports the product description but not broad claims about cost, efficiency or production adoption.
Status
Confirmed. Intel published the product and infrastructure details, and the article preserves the company's sampling and evidence limitations. Internal confidence is medium because the only fetched evidence is Intel's own announcement.
Sources
Update note: Last reviewed 2026-10-13. We will revise this post if Intel publishes final configurations, availability dates or independently reproducible performance evidence.
Sources
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.