Qualcomm’s Dragonfly AI200, a rack-scale system built to run AI models, can now be tested by select customers through Cirrascale Cloud Services. The two companies announced the evaluation program on November 17, 2026, and say each AI200 card carries 768 GB of LPDDR memory.
Qualcomm is best known for phone and edge chips, so this is a notable step into data center hardware. Cirrascale is a specialist cloud provider, and it is the route through which customers get access for now. Nothing is on general sale yet, and neither source lists a price.
What is the Qualcomm Dragonfly AI200?
The AI200 is a rack-scale inference platform. Inference is the stage where a trained model answers requests, as opposed to the training stage that builds it. According to Qualcomm, the system is designed for large language models, multimodal models and other AI workloads, including foundation models at trillion-parameter scale.
The figures and features the sources list:
- 768 GB of LPDDR memory per card
- 43TB of memory across a rack, according to Qualcomm’s product page
- Direct liquid cooling at rack level
- PCIe for scale-up, meaning links between cards inside a system
- Ethernet for scale-out, meaning links between racks
- Confidential computing, which keeps workloads protected while they run
IT Brief reports that Qualcomm also points to its AI Management Suite and support for the major AI frameworks, so teams do not have to rebuild their software from scratch. Qualcomm has not published speed or throughput numbers in the pages we could read, so claims about how fast it runs a given model cannot be checked yet.
How does the Cirrascale evaluation work?
Select customers can try the AI200 through Cirrascale’s AI Innovation Cloud. They bring their own models and workloads and use the hardware for validation, benchmarking and proof-of-concept work. The companies describe a path from those tests to production inference, although they have not said how long that takes or what it costs.
Cirrascale runs the infrastructure as a managed service. It offers a flat-rate model with no ingress or egress fees, which matters for teams that move large datasets in and out of a cloud and are used to paying for each transfer.
The partners are also showing the system in person. A demonstration is running at SC26 in Chicago on November 17 and 18 at Cirrascale’s booth #1046, and a TechArena breakfast panel is scheduled for Wednesday, November 18, from 8:30 to 10:00 a.m. Both fall before this story was published, so readers who missed them will have to look for recordings or follow-ups from the companies.
Why does inference matter here?
Running models at scale is where many organizations now spend their money, and IT Brief notes that power use and operating costs are under scrutiny. The release carries the word energy-efficient in its title, but it gives no wattage or efficiency figure, so that remains a claim rather than a measured result.
The large memory per card is the most concrete selling point on the page. Models with very many parameters need a lot of memory to hold their weights, and Qualcomm is betting that LPDDR capacity, rather than the more expensive memory used on many rival accelerators, can serve that need. That is our reading of the specs, not a comparison either company has made.
IT Brief adds that the partnership could support Qualcomm’s multi-generation inference roadmap. Qualcomm’s chief of data center business, Tony Pialis, and Cirrascale’s chief executive, Dave Driggers, are both quoted in the release.
Prices
Neither Qualcomm nor Cirrascale lists a price for the Dragonfly AI200 hardware or for evaluation access. The only commercial detail given is Cirrascale’s flat-rate cloud model with no ingress or egress fees. Any quote will have to come from Cirrascale directly, and the sources name no country, currency or region limits for the program.
Who it suits
AI platform teams. If you run large language or multimodal models in production and your bill is dominated by serving them, the evaluation lets you test your own workloads before committing to anything.
Teams that move a lot of data. The no ingress or egress fee model removes one common surprise on cloud invoices, though you should confirm the exact terms with Cirrascale.
Security-minded buyers. Confidential computing is listed at rack level, which may matter for regulated workloads. The sources do not explain how it is implemented, so ask for details.
Who should wait. Hobbyists, small startups and anyone who needs a published price, a benchmark or a general release date. Access is limited to select customers, and no wider timing has been given.
What is still unknown
The pages we read do not give speed figures, power draw, pricing, the number of cards in a rack, how customers are selected, or when the AI200 will be widely available. If any of those decide your purchase, ask Cirrascale for numbers on your own model before drawing conclusions.
Next step
Read the Qualcomm and Cirrascale announcements, and if you want to test your workloads, contact Cirrascale about the evaluation program.






