Qualcomm Unveils AI200, AI250 Inference Chips

Qualcomm unveiled two AI inference chips, the AI200 (commercial 2026) and AI250 (2027), built around its Hexagon neural processing units and positioned for inference rather than model training. The AI200 supports large memory footprints (cards referenced up to 768 GB of LPDDR per card) and the AI250 uses a near‑memory computing design promised to deliver substantially higher effective memory bandwidth and improved power efficiency. Qualcomm will offer the chips as standalone silicon, PCIe accelerator cards, and liquid‑cooled rack servers, and it emphasized lower total cost of ownership and energy efficiency. Early customer commitments include Humain, which is linked to Saudi Arabia’s Public Investment Fund. The announcement sent Qualcomm shares up roughly 20–22% intraday to about $205 and also lifted related names such as Arm. Qualcomm left several per‑chip performance and scaling details unspecified (for example NVLink/Fusion support and exact metrics), leaving open questions about how quickly it can challenge entrenched GPU providers.