Qualcomm hands select customers a 43 TB rack to test against Nvidia
Cirrascale's cloud now hosts Qualcomm's Dragonfly AI200 for trials, betting LPDDR memory beats Nvidia's HBM on power bills rather than raw speed.

Qualcomm has opened cloud-based trials of its Dragonfly AI200 rack-scale inference system to selected customers through Cirrascale Cloud Services, with datacenter chief Tony Pialis pitching it as a more efficient alternative to Nvidia-sized power and operating costs. For buyers weighing AI infrastructure spend, the trial is a rare chance to test a 43 TB LPDDR-based rack against Nvidia's HBM3e-based Blackwell systems before committing capital.
Qualcomm has cracked open the door on its answer to Nvidia's rack-scale dominance, and it is doing it through someone else's cloud. The chipmaker has opened cloud-based trials of its Dragonfly AI200 inference system to selected customers via Cirrascale Cloud Services, giving those buyers a route into production deployments if the hardware holds up. Tony Pialis, Qualcomm's executive vice president and general manager for datacenter, framed the pitch bluntly: "AI inference is becoming one of the largest and most important computing workloads in the datacenter, and the industry needs a more efficient approach." Translation for anyone who signs the power bill: Qualcomm is selling the idea that you can run large language, multimodal, and other AI models without Nvidia-sized electricity and operating costs.
The headline spec is memory. Each AI200 accelerator card carries 768 GB of LPDDR memory, and 56 cards in a liquid-cooled rack add up to roughly 43 TB. Qualcomm says the system supports models ranging from 7 billion to 10 trillion parameters, including long-context, retrieval-augmented generation (RAG), and agentic workloads. That is a deliberately wide net. A 7 billion parameter model is the kind of thing a startup runs on a single node; a 10 trillion parameter model is frontier territory. Qualcomm is arguing one rack architecture can serve both, which is exactly the claim enterprises want to hear before they standardize on a vendor for the next several years.
Here is where the argument gets genuinely interesting. Nvidia's Blackwell-based NVL72 systems lean on HBM3e stacks, which are fast and costly. Qualcomm went the other way, choosing LPDDR, the memory more commonly found in mobile devices. As The Register has previously explored, packing in memory does not necessarily mean matching Nvidia's performance. The trade is capacity and cost against raw bandwidth, and for inference workloads, which are less about training a model and more about serving it cheaply at scale, that trade can look attractive. Whether it looks attractive in practice is precisely what the Cirrascale trial is designed to settle.
Pialis claimed the AI200 brings Qualcomm's experience with high-performance, low-power computing to rack-scale infrastructure, in an architecture designed around the performance, memory, and cost requirements of inference workloads. Cirrascale CEO Dave Driggers made the buyer's case: "Our customers need efficient, scalable infrastructure that can support increasingly sophisticated models without allowing power and operating costs to become the limiting factor." He added that working with Qualcomm gives customers evaluation access to "a compelling new architecture purpose-built for the next generation of AI inference." Note the word evaluation. Nobody is being asked to rip out an Nvidia deployment today.
The AI200 is one piece of a broader Dragonfly datacenter lineup that also includes the AI250 and AI300 accelerators and the C1000 server CPU. Qualcomm set out those ambitions at its Investor Day in June, which is the moment this stopped being a side project and became a stated strategic bet. The company also has its own software stack, the Dragonfly AI Infrastructure Management Suite, covering provisioning, monitoring, workload orchestration, and fault handling across rack-scale deployments. That software layer is easy to skim past and hard to overstate. Rack-scale systems live or die on orchestration and fault handling, and a challenger that ships silicon without a credible management layer tends to stall in pilot purgatory.
Cirrascale demonstrated the system at SC26 in Chicago on November 17-18, which puts the hardware in front of the supercomputing crowd rather than a closed-door sales room. That is a signal of confidence, and it is also a test. Supercomputing audiences are unsentimental about benchmarks, and a memory-heavy, power-light pitch invites exactly the scrutiny Qualcomm needs to survive.
The strategic stakes stretch well beyond one chip. Inference is where the money is heading, because training runs are episodic while serving runs forever. Every enterprise that has moved a model into production has discovered that the recurring cost is not the GPU purchase, it is the electricity, the cooling, and the utilization rate. If Qualcomm can demonstrate that a liquid-cooled LPDDR rack delivers acceptable latency at a materially lower operating cost, it does not need to beat Nvidia on peak performance to win business. It only needs to be good enough on the workloads that dominate the bill.
For executives and boards, the practical takeaway is that the AI infrastructure market is finally developing a second option worth pricing. That changes negotiating leverage even for buyers who never deploy a Dragonfly rack. Watch three things from here: whether the selected customers convert from evaluation to production, whether Qualcomm publishes performance numbers that survive independent scrutiny, and whether the AI250 and AI300 arrive on schedule. The trial is the opening move, not the verdict.
This story's Key Insights and Take-aways are locked.
Create a free account to unlock Executive Actions for one credit.
Register to UnlockAlways free for Executives Club members. Join the Club
More in Technology
Anthropic mathematician claims Claude helped crack the Riemann hypothesis
If the proof holds, AI has toppled math's most famous open problem; if it does not, it is a case study in announcing breakthroughs before anyone verifies them.
AI's $4 Trillion Wipeout Erases Nasdaq's Gains for the Year
A brutal week took roughly $4 trillion off the AI sector and left the Nasdaq flat for the year, and the damage ran the full length of the supply chain.
Anthropic mathematician says AI proved the Riemann hypothesis. Verification will take time.
A single unverified claim from inside Anthropic now sits against more than 160 years of open mathematics, and the gap between asserting a proof and surviving peer review is the entire story.


