Qualcomm AI250 NEW

Qualcomm AI250, branded Qualcomm Dragonfly AI250, is the second generation of Qualcomm's rack-scale AI inference line, announced on 27 October 2025 and expected to reach commercial sampling in mid-2027. It keeps the AI200's defining trade, 768 GB of memory per card, and changes how that memory is reached. Where the AI200 attaches conventional LPDDR5X, the AI250 introduces Qualcomm High Bandwidth Compute, or HBC, a near-memory design that bonds a compute die directly beneath the DRAM so that data-movement-bound operations are handled next to the memory instead of being shipped across a narrow external interface. Qualcomm quotes the result as 133 TB/s of effective memory bandwidth per card, an eighteenfold increase over the AI200, and 7.455 PB/s across a full rack. That figure needs reading carefully, and this page deliberately does not list it as a memory bandwidth specification: Qualcomm defines effective bandwidth as the internal bandwidth of the memory rather than the external interface, and it is available to the HBC compute die for a specific class of low arithmetic intensity operations, while the main Dragonfly accelerator continues to talk to memory conventionally. It is therefore not the same measurement as the HBM bandwidth quoted for competing accelerators, and Qualcomm's own like-for-like claim against HBM is a ratio, six times the bandwidth per watt, rather than a raw comparison. As with the AI200, nearly everything else Qualcomm publishes is a rack aggregate: a rack holds 56 cards in a single-wide Open Rack v3 chassis for 43 TB of total memory, uses PCIe 6.0 for scale-up and Ethernet with RoCE for scale-out, and is offered with direct liquid cooling or air cooling. Qualcomm also quotes over 6 TB of HBC memory per server, enough for models beyond ten trillion parameters, without saying how many cards a server holds. Rack power is stated inconsistently across Qualcomm's own documents, 160 kW in the launch announcement and 140 kW in the current product page, and no per-card wattage has ever been published, so no TDP is recorded here. Qualcomm has published no throughput figure of any kind for the AI250, at any precision, at card or rack level, and this is a stated position rather than an omission: its engineering blog argues that headline compute figures matter less than a system's ability to keep compute fed, and that buyers should evaluate on delivered performance on memory-bound workloads and on performance per watt and per dollar. This entry therefore carries no performance metrics rather than estimated ones. A third generation, the AI300 with HBC Gen 2, is already announced for 2028 sampling with a claimed fifty-fourfold bandwidth increase over the AI200. The announced launch customer is HUMAIN of Saudi Arabia, whose 200 MW programme begins with AI200 racks and is stated to include AI250 hardware.