by QualcommUS
Qualcomm's data-center inference accelerator cards and racks, designed to generate tokens efficiently for large language and multimodal models. Targeted at enterprises seeking alternatives to NVIDIA.
AI200 available 2026, AI250 targeting 2027; 160 kW per rack with 768 GB LPDDR5X memory.
Checked 10 Oct0 updates this week
Large memory per card with LPDDR5X bandwidth keeps data in motion during token generation, reducing trips to slower off-chip storage.
Cards within a rack scale via PCIe; racks link across data centers via Ethernet, allowing unlimited horizontal scaling.
Designed for total-cost-of-ownership advantage: efficient memory use and power delivery reduce cooling and operational costs versus GPU-based racks.
| Spec | AI200 / AI250 accelerators |
|---|---|
| Memory per card | 768 GBR (reported) |
| Memory type | LPDDR5XR (reported) |
| Rack-level power | 160 kWR (reported) |
| Cooling | Direct liquid coolingR (reported) |
R reported by the company
Qualcomm
Snapdragon, AI accelerators, automotive chips, wireless
Qualcomm designs and licenses semiconductor technology and wireless standards. The company expanded into AI infrastructure in 2025 with the AI200 and AI250 rack-scale inference accelerators, positioned to challenge NVIDIA in data-center inference.
