Qualcomm Debuts AI200, AI250 Data-Center Inference Racks

Qualcomm will sell data-center AI racks that carry 768GB of LPDDR memory per card. Tom's Hardware reports the AI200 lands in 2026 and a follow-on AI250 in 2027. Both are built around Qualcomm's Hexagon NPUs, cooled by direct liquid, and sit inside a 160 kW power envelope per rack. PCIe handles scale-up inside the rack; Ethernet handles scale-out between them.

The AI250 carries the architectural bet. Qualcomm's claim is that its near-memory computing design will deliver "greater than 10x higher effective memory bandwidth" at much lower power, aimed squarely at the cost of serving large generative models in production.

The first named customer is Humain, the Saudi AI venture, which has committed to a 200-megawatt deployment of combined AI200 and AI250 systems starting in 2026 to power its ALLaM models and enterprise and government workloads. "With Qualcomm AI200 and AI250, we're redefining what's possible for rack-scale AI inference," said Durga Malladi, SVP and GM of Technology Planning, Edge Solutions & Data Center at Qualcomm Technologies. He added that the systems let customers "deploy generative AI at unprecedented TCO."

This is the 11th Qualcomm item we've logged in 90 days, after a defense laptop with Anduril and a round of performance-focused layoffs. Nvidia and AMD still own the rack-scale inference segment the AI200 is aimed at, and no per-card throughput or latency figures were released alongside the launch.