Qualcomm Dragonfly AI200 Now Available for Evaluation

Qualcomm Technologies and Cirrascale Cloud Services have opened evaluation access to the Dragonfly AI200, a purpose-built rack-scale inference accelerator designed for trillion-parameter foundation models. The collaboration brings the hardware platform to select customers through Cirrascale's AI Innovation Cloud, targeting enterprises and cloud providers navigating the shift from AI model development to production-scale deployment.
Release Date and Availability
Officially launched on November 17, 2026, the Qualcomm Dragonfly AI200 is now available for evaluation by select customers through Cirrascale's managed cloud environment. Live demonstrations run at the SC26 conference in Chicago from November 17-18, 2026, in Cirrascale's booth 1046. The companies position this release as a foundation for Qualcomm's multi-generation data center AI inference roadmap, with the current evaluation program supporting platform validation, workload characterization, and proof-of-concept activity.
Memory Capacity and Architecture Design
Each Dragonfly AI200 card delivers 768GB of LPDDR memory, providing the capacity required for large-scale AI inference workloads. The architecture combines PCIe scale-up and Ethernet scale-out within a 160kW rack configuration, supported by direct liquid cooling to manage thermal loads at scale. Qualcomm explicitly designed the platform for trillion-parameter model inference, distinguishing it from general-purpose GPU infrastructure.
Tony Pialis, EVP and GM at Qualcomm Datacenter, stated the platform brings Qualcomm's high-performance, low-power AI computing expertise to rack-scale infrastructure, with an architecture designed around the performance, memory, and cost requirements of inference. The system includes confidential computing capabilities, support for Qualcomm AI Management Suite, and ecosystem integration with leading AI frameworks.
Performance Positioning Against Nvidia
Qualcomm positions the Dragonfly AI200 against Nvidia GPUs on a tokens-per-watt basis, targeting the economics of inference as a primary competitive vector. Dave Driggers, CEO of Cirrascale Cloud Services, noted that customers need efficient, scalable infrastructure that can support increasingly sophisticated models without allowing power and operating costs to become limiting factors. The evaluation program enables customers to assess the platform using their own models and workloads, moving beyond vendor-supplied benchmarks to real-world performance validation.
Cirrascale brings a transparent, flat-rate commercial model without ingress or egress fees to the collaboration, addressing one of the cost pain points in cloud-based AI infrastructure. The partnership focuses on inference-optimized software stacks rather than training workflows, reflecting the industry's growing emphasis on deployment economics as foundation models mature beyond the research phase.
SC26 Conference Demonstration and Panel
Beyond booth demonstrations, the companies organized a TechArena breakfast panel on near-memory compute and the future of AI infrastructure for November 18, 2026, from 8:30 am to 10 am. The panel includes Rashid Attar, Qualcomm Technologies SVP of Engineering, David Driggers from Cirrascale, Pavan Balaji from Meta as Research Scientist and Technical Lead, and Greg Yeric, VP of Portfolio, Compute Systems Scaling at Imec.
What This Means
The Dragonfly AI200 evaluation launch represents Qualcomm's entry into rack-scale AI infrastructure, challenging Nvidia's GPU dominance with an inference-first architecture. For enterprises, the 768GB memory per card and trillion-parameter model support address the deployment bottleneck as models grow beyond what single-accelerator systems can handle efficiently. The Cirrascale partnership provides immediate evaluation access without requiring customers to build or manage dedicated hardware, lowering the barrier to assessment. As inference becomes the dominant data center AI workload, purpose-built accelerators optimized for tokens-per-watt rather than training throughput may reshape procurement decisions across cloud and enterprise segments.
on Emergent today





