The D-Series is built on Intel's Xe3P architecture and pairs LPDDR5X memory with 1.5 TB/s of bandwidth and 1.3 petaflops of FP8 compute. Intel says one card can serve models up to 300 billion parameters, and an eight-GPU system can go past one trillion. The cards are air cooled and use a standard PCIe form factor, so they fit existing servers without special racks or liquid cooling. ASUS, HPE, Dell and Supermicro are among the system partners. Using LPDDR5X instead of HBM gives up bandwidth in exchange for more capacity at lower cost, aiming at long-context and many concurrent requests rather than training. The launch came the same day Microsoft detailed Maia 300 and Nvidia opened its Vera Rubin rack designs. For startups serving large open-weight models, a cheaper high-memory card that fits ordinary air-cooled servers could cut inference costs. That depends on Intel's software stack supporting the runtimes teams already use.