Qualcomm Launches AI250 and AI200 with Memory Footprint for AI Workloads (hothardware.com)

Qualcomm announced two rack-scale, inference-optimized accelerator cards — AI200 (shipping 2026) and AI250 (2027) — that prioritize large memory capacity alongside NPU compute. Each card supports up to 768 GB of LPDDR, and the AI200 rack pairs that memory with a hexagonal NPU and direct liquid cooling. The AI250 keeps the 768 GB memory footprint but introduces a near‑memory computing architecture that Qualcomm says yields a >10× increase in effective memory bandwidth and much lower power for inference workloads. For the AI/ML community this signals a shift from raw NPU flop counts toward memory-centric system design tailored for LLMs and large multimodal models, where data movement and memory bandwidth are often the bottlenecks. Qualcomm emphasizes rack-scale TCO, compatibility with major AI frameworks, an open software stack and one‑click model deployment to simplify integrating pre-trained models. Early customer Humain plans to deploy both solutions in a 200 MW Saudi Arabia facility, highlighting these cards’ positioning for hyperscale inference. The technical implication is a push toward architectures that reduce data movement (near‑memory compute), increase effective bandwidth, and lower power — important levers for scalable, cost‑efficient generative AI deployments.