Intel's reference card ships with 160GB of LPDDR5X, and manufacturing partners can build versions with up to 480GB. Intel lists 1.5TB/s of memory bandwidth and 1.3 PFLOPS of FP8 compute. One card is rated for models of roughly 300 billion parameters, and an eight-GPU system for more than one trillion. The card is a 350W, air-cooled PCIe 5.0 x16 design that fits conventional servers. Sampling is underway with server makers, cloud providers and software vendors, but pricing and general availability are unannounced. One report puts broader availability in 2027. The D270 is the product name for the chip previously called Crescent Island. AMD's Instinct MI350P reaches 2.3 petaflops of FP8 but draws up to 600 watts. Intel promises day-zero support for leading open models through vLLM and PyTorch. For inference-heavy startups, the pitch is capacity per dollar rather than peak speed. LPDDR5X gives less bandwidth than HBM but more capacity at lower cost and power. Large models could fit on fewer cards, which cuts multi-GPU sharding. The gains depend on batch size and workload, and the FP8 figure is theoretical. Software maturity and real pricing are the open questions.