Qualcomm Dragonfly AI200 Rack Trials Open on Cirrascale, 43 TB LPDDR
Qualcomm lists 43 TB of LPDDR in each 140 kW Dragonfly AI200 rack, now open to select customers on Cirrascale's cloud: over twice the 20 TB of GPU memory NVIDIA lists for an NVL72, at lower bandwidth. No measured results yet.
2 min read

Qualcomm's Dragonfly AI200 racks are open to select customers on Cirrascale's cloud; figures are vendor claims.
Graphic: Voltgrid News
Qualcomm made its Dragonfly AI200 rack-scale inference system available for evaluation by select customers through Cirrascale Cloud Services, the two companies said on Nov. 17, with each card carrying 768 GB of LPDDR memory on Qualcomm's figures.
Customers run their own models on Cirrascale's AI Innovation Cloud, and the companies support validation, workload characterization, benchmarking and proof-of-concept work, according to the joint release. Qualcomm describes the trials as a route into production deployments, The Register reported on Nov. 19. Cirrascale also demonstrated the system at its SC26 booth in Chicago on Nov. 17-18. This is evaluation access, not volume shipments.
Capacity over bandwidth
The AI200 bet is memory size. Qualcomm's product page lists 43 TB per rack; The Register reports 56 cards per rack, and 56 cards of 768 GB matches that. Qualcomm gives the rack a 140 kW thermal design power, in a liquid-cooled Open Rack v3 enclosure. NVIDIA lists 20 TB of GPU memory (HBM3E) for a GB300 NVL72 rack, so Qualcomm's 43 TB is about 2.15 times that, by Voltgrid's arithmetic from the two vendors' figures. Counting the 17 TB of LPDDR5X on NVIDIA's 36 Grace CPUs, a slower tier, NVIDIA lists 37 TB of fast memory per rack, and the gap shrinks to about 1.16 times.
The trade is speed. Qualcomm rates the rack at 0.414 PB/s (414 TB/s) of LPDDR5x bandwidth, about 72% of the up to 576 TB/s NVIDIA lists for the GPU memory of an NVL72 rack, by Voltgrid's arithmetic from the vendors' figures. When a model generates text, every new token rereads the active weights and the conversation's cache from memory, so bandwidth often sets the ceiling on tokens per second per user (large batches shift the limit toward compute), while capacity sets how large a model and how long a context fit. Put another way, on the vendors' peak figures the AI200 rack could read its full 43 TB about 10 times a second, and an NVL72 its 20 TB of GPU memory about 29 times. If Qualcomm's pitch holds, then the AI200 suits big models and long contexts served at moderate per-user speed, where fewer racks per stored terabyte matter more than peak speed.
Vendor figures, not measured: the AI200 rack holds more memory than an NVL72's GPU memory, at lower bandwidth.
Graphic: Voltgrid News
No measured results yet
Neither company has published tokens per second, latency, measured power under load or a price, and no independent test results had been published as of Nov. 19. The 7-billion to 10-trillion-parameter range The Register cites, and the 128K context limit on Qualcomm's product page, are vendor claims. The cloud trials are the first place outside Qualcomm where those claims can be checked, which is why the evaluation program matters more than the demo.
The rack scales up over PCIe Gen6 and out over Ethernet with RoCE, and Qualcomm pairs it with its own management software for provisioning, monitoring and fault handling. The AI250 and AI300 follow the AI200 in the Dragonfly line; Qualcomm said this month that a test stack of the AI250's near-memory design beat HBM4 on bandwidth per watt, as Voltgrid reported. Data center chips are central to the company's pivot away from handsets, including its job cuts.
What to watch next is the first named Cirrascale evaluation customer and the first throughput figure measured outside Qualcomm. Saudi Arabia's HUMAIN, which Qualcomm said in October 2025 is targeting 200 MW of AI200 and AI250 racks from 2026, would be the first large deployment.
Sources
- 01Qualcomm offers test drive of its rack-scale Nvidia AI challengerThe Registertheregister.com
- 02Qualcomm and Cirrascale Collaborate to Advance Next-Generation, Energy-Efficient AI InferenceCirrascale Cloud Servicescirrascale.com
- 03Qualcomm Dragonfly AI200 product pageQualcommqualcomm.comas read on Nov 19, 2026
- Published
- Sources
- 3 cited
