Physix Frontier · News Briefing Card (IT Home · Nov 20, 2026)
Qualcomm AI200 inference accelerator enters customer cloud evaluation
KEY FACTS
- Qualcomm has partnered with AI private-cloud company Cirrascale to open evaluation of the Dragonfly AI200 inference accelerator to select customers.
- Customers can test the chip's performance on their own AI workloads through a cloud environment.
- The AI200 is Qualcomm's first rack-scale AI solution, with 768GB of LPDDR5X memory per card.
- The accelerator supports PCIe Gen6 scale-up and RoCE scale-out.
- Each 140kW liquid-cooled ORv3 rack holds 56 cards, for 43TB of total memory and 0.414 PB/s of memory bandwidth.
KEY DATA
768GB LPDDR5XPer-card memory
56 cardsCards per rack
43TBTotal memory per rack
0.414 PB/sMemory bandwidth per rack
PHYSIX OBSERVATION
Qualcomm is entering the inference accelerator market through cloud-based evaluation, bypassing the asset-heavy path of building its own data centers and letting customers first test the waters on real workloads. The 43TB of memory and 0.414 PB/s of bandwidth per rack target ultra-large-model inference, but whether it can pry open Nvidia's ecosystem depends on software stack maturity and customer migration costs; performance during the evaluation period is the key signal.
Source: IT Home report
Physix Frontier