Skip to content
AI.info

The Pulse

Qualcomm AI200 opens for select customer tests on Cirrascale

Select Cirrascale customers can now evaluate Qualcomm’s Dragonfly AI200 on their own AI inference workloads. The rack-scale platform has 768 GB of memory per card and is designed for large language and multimodal models.

Qualcomm AI200 opens for select customer tests on Cirrascale

Select Cirrascale customers can now test Qualcomm’s Dragonfly AI200 inference platform on their own models and workloads through the cloud provider’s AI Innovation Cloud. The companies said on November 17 that they will support validation, benchmarking and proof-of-concept work on the rack-scale system.

Source webwire.com

AI200 is designed for running large language models, multimodal models and other AI workloads. The companies’ release describes a path from evaluation to production inference deployments, while limiting current evaluation access to select customers.

Inference and cost

The platform is designed to deliver high performance per dollar per watt and low total cost of ownership, the release says. Cirrascale provides the cloud environment and managed infrastructure, while Qualcomm Technologies supplies the AI acceleration and rack-scale system expertise.

The release describes Cirrascale’s commercial model as flat-rate, with no ingress or egress fees. It frames power consumption, performance and cost as considerations for companies deploying generative AI at scale.

Dave Driggers, Cirrascale’s CEO, said customers need infrastructure that can support more sophisticated models without power and operating costs becoming limiting factors. Tony Pialis, Qualcomm Technologies’ executive vice president and general manager for data center, said the collaboration lets customers evaluate the technology on their own workloads.

Memory and rack design

AI200 has 768 GB of LPDDR memory per card. The companies say it is architected to support inference for next-generation foundation models at trillion-parameter scale.

Cirrascale lists 36.8 TB of memory capacity and 355 TB/s of memory bandwidth per rack. The company says AI200 serves models from 7B up to 10 trillion parameters, including long-context, retrieval-augmented generation and agentic workloads.

The rack-level system combines direct liquid cooling, PCIe scale-up, Ethernet scale-out and confidential computing. It is supported by Qualcomm AI Management Suite and ecosystem support for leading AI frameworks, according to the release.

Cirrascale describes an integrated rack compliant with OCP ORv3, with direct liquid cooling and a cableless backplane. Its product page also lists software tooling for model deployment as bare-metal, cloud virtual machines or inference-as-a-service.

Earlier cloud offering

The companies have worked together on inference services before. In March 2025, Cirrascale announced an Inference Cloud powered by the Qualcomm AI Inference Suite, following an earlier playground for developers to experiment with AI models, agents and libraries.

The earlier playground used Qualcomm’s Cloud AI 100 Ultra accelerator. The earlier release said Inference Cloud customers paid for their AI model usage and accessed models through API endpoints; the new announcement describes Cirrascale’s flat-rate commercial model for the collaboration.

A demonstration of AI200 is scheduled for Cirrascale’s booth at the SC26 conference in Chicago on November 17–18. Qualcomm engineering executive Rashid Attar is also scheduled to join Driggers and representatives from Meta and Imec for a November 18 panel on near-memory compute and AI infrastructure.

Sources

  1. WebWire
  2. Cirrascale Cloud Services
  3. Cirrascale Cloud Services

Related stories

Explore

More articles