Qualcomm AI200 NEW

Qualcomm AI200 is a rack-scale AI inference accelerator announced on 27 October 2025, Qualcomm's return to the data center and the first product in its Dragonfly line, now branded Qualcomm Dragonfly AI200. Its defining decision is memory. Each card carries 768 GB of LPDDR5X, the highest per-accelerator memory capacity of any announced AI accelerator and more than five times what a 144 GB HBM3e part offers, deliberately trading bandwidth for capacity and cost so that very large models fit in fewer accelerators. Qualcomm has demonstrated a 350 billion parameter model running on a single card and states the card is designed for models up to a trillion parameters. Almost everything else Qualcomm has published is a rack aggregate rather than a card specification, and is recorded here as such: a rack holds 56 cards in a single-wide Open Rack v3 chassis for 43 TB of total memory and 0.414 PB/s of aggregate bandwidth, uses PCIe 6.0 for scale-up and Ethernet with RoCE for scale-out, supports confidential computing, and is offered with direct liquid cooling or air cooling. Qualcomm's launch materials quote rack-level power of 160 kW while its current product page lists 140 kW, and it has never published a per-card wattage, so no TDP is recorded here rather than dividing a rack figure across 56 cards. Most notably, Qualcomm has published no throughput figure of any kind, at any precision, at either card or rack level. There are no TOPS, no TFLOPS and no FP8, FP16 or INT8 numbers in any Qualcomm document, so this entry carries no performance metrics rather than estimated ones. That is a deliberate change in posture: for the predecessor Cloud AI 100 Ultra, Qualcomm published a product brief listing 150 W, 870 INT8 TOPS and 548 GB/s. Software runs through the Qualcomm AI Inference Suite and the Efficient Transformers Library with one-click deployment of Hugging Face models. The announced launch customer is HUMAIN of Saudi Arabia, targeting 200 MW of AI200 and AI250 rack deployments from 2026.