Qualcomm unveils AI200 and AI250 data center accelerators

San Diego based Qualcomm on Monday unveiled two artificial intelligence accelerator chips for data centers, the AI200 and AI250, aiming to expand beyond smartphones into large scale AI infrastructure. The company said the AI200 will be commercially available in 2026 and the AI250 in 2027, and that both will be offered as accelerator cards and in liquid cooled rack appliances designed for inference workloads.

Qualcomm described the chips as optimized for inference, the phase of AI that runs pretrained models, and said they are intended to improve cost efficiency measured as performance per dollar per watt. The AI200 offers 768 gigabytes of memory per card and is built on the companys Hexagon neural processing unit architecture. Qualcomm said the AI250 will deliver more than 10 times the memory bandwidth of the AI200, a change the company said is aimed at supporting larger, denser models, but it did not disclose precise power use or raw performance figures for either chip.

The company presented the processors as rackscale appliances that use direct liquid cooling and PCIe connectivity. Qualcomm stated a fully configured rack would consume roughly 160 kilowatts and that customers...