Qualcomm enters AI data-center chip race

Qualcomm has introduced the AI200 and AI250 data center accelerators with a focus on memory capacity and data movement efficiency. The AI200 card supports 768 GB of LPDDR memory , which allows large language models and multimodal models to run with fewer parts and without constant off-card transfers.

The AI250 uses a near-memory compute layout designed to increase effective memory bandwidth during inference, which is where workloads often stall due to repeated key-value cache access. Qualcomm states that this architecture provides greater than 10x higher effective memory bandwidth while lowering power use.

“Qualcomm launches AI200 and AI250 chip-based accelerator cards and racks—delivering industry-leading rack-scale inference performance and memory efficiency for data center AI workloads.”

— Cristiano Amon

Both products are offered as accelerator cards and as full liquid-cooled racks. The racks use PCIe for scale-up configurations and Ethernet for scale-out clusters. A full rack operates at 160 kW, which places it in the same class as existing GPU inference racks from other vendors. Availability is planned in stages, with AI200 scheduled for 2026 and AI250 for 2027. Qualcomm has also disclosed an annual update cadence for its data center roadmap and identified HUMAIN as one of the early customers with deployments planned up to 200 MW of capacity.

Source: Qualcomm

In terms of memory comparisons, AMD’s Instinct MI350X offers 288 GB of HBM3e and 8 TB/s bandwidth. The successor (MI400) will feature up to 432GB. NVIDIA’s B200 class configurations provide around 180 GB of HBM3e per GPU. Intel’s Gaudi 3 provides 128 GB of HBM2e. These figures place Qualcomm’s approach into a category defined not by HBM size, but by total on-card memory capacity.

Source: Qualcomm