Chips & Compute

Qualcomm's Dragonfly AI200 Opens for Customer Testing via Cirrascale Cloud

Qualcomm and Cirrascale have opened the Dragonfly AI200, a rack-scale inference platform with 768GB of LPDDR per card, for evaluation by select customers through the AI Innovation Cloud.

Qualcomm & Cirrascale open Dragonfly AI200 for tests
Qualcomm & Cirrascale open Dragonfly AI200 for testsValentin Lacoste / unsplash
By Sophie Lindqvist5 min read

Why it matters

  • Qualcomm's Dragonfly AI200, with 768GB of LPDDR memory per card, is available now for evaluation by select customers via Cirrascale's AI Innovation Cloud.
  • The platform is architected for inference at trillion-parameter scale and includes direct liquid cooling, PCIe scale-up, Ethernet scale-out and confidential computing.
  • Cirrascale's evaluation service uses a flat-rate commercial model with no ingress or egress fees.
  • Qualcomm says the Cirrascale work could anchor its multi-generation data-centre AI inference roadmap.

Qualcomm and Cirrascale have opened the Qualcomm Dragonfly AI200 for evaluation by select customers through Cirrascale's AI Innovation Cloud, marking the first hands-on access to the companies' rack-scale AI inference platform.

Each card in the system carries 768GB of LPDDR memory, and Qualcomm says the platform is architected for inference at trillion-parameter scale. That hardware profile — big memory, low power, rack-level integration — is the whole pitch: run enormous models cheaply, rather than train them.

The evaluation program lets users test the system with their own models and workloads inside Cirrascale's cloud environment. Qualcomm and Cirrascale have designed the offering around platform validation, workload characterisation, benchmarking and proof-of-concept work, with a defined route from testing to production inference deployments.

The announcement matters because AI spending is shifting. After years of investment concentrated on model development and training, suppliers are now competing on the cost and energy demands of actually running those models in live data-centre services. Inference — the act of serving a trained model to users — is where those bills land.

What does the Dragonfly AI200 actually offer?

The hardware combines 768GB of LPDDR memory per card with a set of rack-level design features. Qualcomm lists four:

  • Direct liquid cooling
  • PCIe scale-up
  • Ethernet scale-out
  • Confidential computing

That combination signals a system built for dense, efficient serving rather than general-purpose compute. Qualcomm describes the platform as designed around the memory, performance and cost profile needed to serve AI models — large language models, multimodal systems and other AI workloads — rather than train them.

The context here is Qualcomm's own trajectory. The company is better known for mobile and edge computing, and it has been expanding into data-centre AI inference. The Dragonfly AI200 is part of that push, and Qualcomm positions it as a purpose-built alternative to repurposed general-purpose infrastructure in a market where power use and operating costs are under growing scrutiny.

The takeaway: this is a bet that inference deserves its own silicon and rack design, not leftover training hardware.

What did the companies say?

The announcement comes with two quotes that frame the strategic logic on each side.

Tony Pialis, Executive Vice President and General Manager, Data Centre, Qualcomm Technologies, argues that inference has outgrown generic infrastructure. "AI inference is becoming one of the largest and most important computing workloads in the data center, and the industry needs a more efficient approach. Qualcomm AI200 brings our leadership in high-performance, low-power AI computing to rack-scale infrastructure, with an architecture designed around the performance, memory and cost requirements of inference. Our work with Cirrascale gives customers a direct way to evaluate this technology on their own workloads and experience the resulting data center economics," Pialis said.

Dave Driggers, Chief Executive Officer of Cirrascale Cloud Services, frames the shift in economic terms. "The economics of AI are increasingly becoming the economics of inference. Our customers need efficient, scalable infrastructure that can support increasingly sophisticated models without allowing power and operating costs to become the limiting factor. By working with Qualcomm Technologies, we can give customers evaluation access to a compelling new architecture purpose-built for the next generation of AI inference," Driggers said.

Both statements point to the same pressure point: power and operating costs as the constraint on AI deployment. The takeaway is that both companies see inference economics, not raw model capability, as the next competitive battleground.

Why offer evaluation through a cloud?

Cirrascale provides the cloud environment for the evaluations. Its service includes managed AI infrastructure and a flat-rate commercial model with no ingress or egress fees. That pricing detail matters — data-transfer fees are a common complaint about mainstream cloud providers, and their absence gives customers a cleaner basis for comparing inference systems on real workloads.

For Cirrascale, the arrangement adds a new system to its AI Innovation Cloud at a moment when cloud providers and infrastructure specialists are trying to attract customers seeking alternatives to mainstream compute platforms. Offering evaluation access in a managed environment lowers the barrier for enterprises and cloud users who want to measure performance, energy use and operational cost before committing to deployment.

Select customers can move directly from evaluation to production inference deployments through the same route. In a market where many AI buyers want faster ways to test before scaling workloads, that continuity from trial to production is part of the pitch.

The takeaway: try before you buy, with a paved road from benchmark to deployment.

What about the software layer?

Hardware alone doesn't make an inference platform usable. Qualcomm says the Dragonfly AI200 is supported by its AI Management Suite and by ecosystem support for leading AI frameworks.

That software layer is intended to help customers assess how the platform handles production-style inference tasks alongside the underlying hardware design. In practice, it means evaluators can test not just raw throughput but the operational tooling they would need to run the system in a live service.

The takeaway: the evaluation covers the full stack, not just the accelerator.

What does this signal for the road ahead?

The collaboration signals longer-term intent from both companies. Qualcomm says the work with Cirrascale could serve as a base for further work across its multi-generation data-centre AI inference roadmap.

That phrasing is worth noting. It positions this evaluation program as a first step in a sustained product line rather than a one-off launch, with Qualcomm planning successive generations of data-centre inference hardware.

For the broader market, the move adds another serious entrant to the inference infrastructure race, at a time when enterprises are actively comparing options on energy use and total cost rather than peak performance alone. A chipmaker with deep low-power computing experience — Qualcomm's core strength in mobile — applying that heritage to rack-scale systems is a direct challenge to incumbents whose hardware was largely designed around training.

The Dragonfly AI200 is available now for evaluation by select customers through Cirrascale. How many of those evaluations convert into production deployments will be the first real test of whether inference-specific rack design, backed by flat-rate cloud economics, can shift buying decisions in the data centre.

Source: itbrief.com.au

Share this article:

More from Sophie Lindqvist

Sophie Lindqvist

Show full bio

Staff writer covering regulation, policy and their impact on artificial intelligence and machine learning at AI In Context.

749 articles

Related articles

  1. OpenAI and Broadcom Unveil Jalapeño, a Custom LLM Inference Chip
  2. Qualcomm's new Snapdragon chips push AI agents on phones
  3. AI Compute in 2027: Serving Economics Will Reshape Data Centers
  4. KORA Unveils Linux Desktop With On-Device AI for Snapdragon PCs
  5. MiTAC Splits AI Inference Across Two Servers With 2 TB Decode Memory

« Previous articleNext article »