Qualcomm's AI200 rack inference ships 2026 with 768GB LPDDR

Qualcomm's push into data-center AI racks builds on its Hexagon NPUs, the same processor line it uses in edge devices, but applies them to a 160 kW power envelope per rack. The company is betting that memory bandwidth, not raw compute, is the bottleneck for serving large generative models: the AI250's near-memory computing design is claimed to deliver greater than 10x higher effective memory bandwidth at much lower power.

The launch customer, Humain, is a Saudi AI venture that 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. That commitment gives Qualcomm a named anchor customer for a product line that has yet to ship.

The AI200 and AI250 enter a rack-scale inference segment currently owned by Nvidia and AMD, and Qualcomm did not release per-card throughput or latency figures, so direct performance comparisons are not yet possible. The launch is Qualcomm's 11th item logged by this outlet in 90 days, following a defense laptop with Anduril and a round of performance-focused layoffs.