Qualcomm announces AI200 and AI250 rack-scale data-center accelerators targeting inference

Qualcomm

Qualcomm unveiled the AI200 and AI250 accelerator cards and racks for generative AI inference in data centers. The AI200 offers 768 GB of LPDDR5X memory per card in a 160 kW rack; the AI250 adds near-memory computing for an order-of-magnitude performance jump. Both target enterprise customers seeking alternatives to NVIDIA, with HUMAIN named as the first customer planning 200 megawatts of deployment beginning in 2026.

• AI200 targeting 2026 availability; AI250 targeting 2027
• HUMAIN (Saudi Arabia's public AI company) is first major customer with 200 MW planned
• Focuses on inference efficiency and low total cost of ownership vs. NVIDIA's high-performance training GPUs
• Supports confidential computing for secure workloads; scales via PCIe within racks and Ethernet across data centers