Intel sets out its AI infrastructure plan, from 46,000-core racks to a high-memory inference GPU

Intel has outlined a broad data-centre push for AI agents and inference, combining Xeon rack designs, a new GPU and software intended to work across its CPUs and accelerators. The standout detail for AI operators is the GPU’s memory capacity: Intel says its own card will have 160GB, while partner designs could offer up to 480GB.

Intel AI Watch analysis

What happened

The announcements came at the Open Compute Project Global Summit. Intel describes an open, heterogeneous approach: use CPUs for general computing, orchestration and data movement, and GPUs or other accelerators for suitable workloads. Selected Xeon 6+ rack designs are listed at more than 46,000 cores and 133.45 kilowatts per rack. Those are figures for selected configurations, not a universal rack specification.

Intel’s GPU D-Series, formerly codenamed Crescent Island, is aimed at inference, including large-language-model workloads and long contexts. The company quotes up to 1.3 petaflops of FP8 compute and 1.5 terabytes per second of memory bandwidth. It says a single card could handle models of up to 300 billion parameters, with systems of eight GPUs exceeding one trillion. Those are manufacturer claims, and parameter count alone does not tell an operator how fast or efficiently a model will run.

Intel specifies 160GB of LPDDR5X for its own PCIe card. The higher figure, up to 480GB, is an option for partner-developed cards, not the standard Intel card. Customer sampling is under way; Intel has not given a general availability date. The company also says it is working on shared software foundations, early support for new models and open standards for accelerator cards.

Our top picks

• More memory for inference
  Intel’s own card is specified at 160GB; partner designs could reach 480GB, a distinction worth keeping clear when comparing configurations.
• Aiming at large models
  Intel says one card could support models up to 300 billion parameters, with eight-GPU systems exceeding one trillion.
• Inference performance figures
  The company quotes up to 1.3 petaflops of FP8 compute and 1.5TB/s of memory bandwidth for the D-Series.
• CPU-heavy rack designs
  Selected Xeon 6+ configurations exceed 46,000 cores, positioning general-purpose compute as part of the agent infrastructure, not just supporting cast.
• Sampling, not shelves
  Customer sampling has begun, but Intel has not specified when the GPU will be generally available.

Why it matters

The announcement makes Intel’s pitch bigger than a single accelerator: it wants CPUs, GPUs, software and rack designs to fit into an open AI infrastructure stack. That could matter to operators looking beyond a GPU-only build, particularly where model memory needs, power and existing data-centre systems shape what can be deployed.

The practical test is still ahead. Sampling is not availability, and specifications are not customer results. Intel’s numbers make the proposition concrete; deployments will show whether the combination is useful outside the presentation room.

Our read

The memory options are the most immediately interesting detail, especially for teams trying to fit larger models into inference hardware. But do not treat the 480GB figure as the standard Intel card, or the headline performance claims as a substitute for workload-specific tests. The right move for buyers is to watch for shipping configurations, prices and independent comparisons before planning around the platform.

What to watch

• When Intel announces general availability and pricing for the D-Series.
• Which partners actually offer cards with higher memory capacity.
• Real-world inference results across model sizes and workloads.
• Whether Intel’s software and open-standard work translates into supported, deployable systems.

Discussion spark: For inference workloads, would you prioritise fitting a larger model in memory, or choose a faster accelerator with a more established software ecosystem?

Independent WittyWires tracker for public updates about Intel AI. Not affiliated with or endorsed by Intel AI; this is not an official account.