Qualcomm announces AI chips to compete with AMD and Nvidia — stock soars 15% (www.cnbc.com)

Qualcomm announced plans to enter data-center AI with two accelerator families — the AI200 (shipping 2026) and AI250 (planned 2027) — that can be deployed in full, liquid‑cooled server racks. The chips scale up Qualcomm’s Hexagon neural processing units (NPUs) from smartphones to rack-level systems and are positioned for inference workloads rather than large‑scale model training. Qualcomm says its cards support 768 GB of memory (higher than current Nvidia/AMD offerings), target lower power consumption and total cost of ownership, and have a rack power footprint of about 160 kW. The company will sell rack systems and discrete parts (including CPUs) to hyperscalers and partners, and has an announced deployment partnership with Saudi firm Humain; pricing and per‑rack NPU counts were not disclosed. This move matters because Nvidia currently dominates AI accelerators (>90% market share) and most cloud/data‑center investments through 2030 are expected to target AI hardware. Qualcomm’s entry increases supply competition and gives cloud providers another inference‑optimized option with different memory and power tradeoffs, potentially lowering costs and diversity risk. For the AI/ML community, the main implications are more heterogeneity in accelerator architectures, new choices for inference deployments, and pressure on incumbents to differentiate on performance, memory architecture, and TCO.