Qualcomm presents Dragonfly, its new brand for data centers, and lets us see the rack at Computex 2026

This week we have been in Computex 2026 And, among everything we have been able to see both in the Qualcomm keynote and later in demos, we have been able to witness one of the most cutting-edge, but at the same time most mysterious, projects of the brand. We were able to take photos of a Qualcomm Dragonfly rackwhich is the company’s bet to fully enter the data center market for AI.

Dragonfly name introduced during CEO’s Computex keynote Cristiano Amonwho did not give many details about this new platform, since they reserved this information for the investor event that will take place at the end of this month. The brand comes to join the Qualcomm ecosystem along with the already known snapdragon and Dragonwingthe latter launched last year for the IoT and robotics market, which has also had news at Computex.

The Dragonfly rack that we were able to see in the demonstration venue for a select group of media, among which we have been lucky enough to meet. In this rack we could see solutions such as accelerators AI200not the future AI250water-cooled in a very compact rack system where they ran models with up to 350 billion parameters in a single accelerator.

The AI200 is the first to hit the market and stands out for incorporating up to 768 GB of LPDDR memory per cardan amount designed to cover the inference needs of today’s large AI models. The AI250, scheduled for 2027, goes a step further with an innovative memory architecture based on near-memory computing, which according to Qualcomm promises more than 10 times the effective memory bandwidth with lower energy consumption, which translates into a significant generational leap, combining new levels of processing capacity, with greater scalability within data center environments dedicated to AI.

Both solutions will arrive in integrated rack format. The systems use PCIe for interconnection within the rack itself and Ethernet for external interoperability scale, with a consumption target of 160 kW per rack and direct liquid cooling, figures that should allow Qualcomm to compete with similar solutions from NVIDIA and AMD.

Qualcomm is also working in parallel on server-grade CPUs, and that’s where Dragonfly comes in. The company wants to take advantage of its experience in this regard to also provide distributed agentic AI interoperability within its business solutions.

The Computex announcement is, for now, a concept demo. Qualcomm has promised a more detailed presentation of its plans for data centers on June 24, during its Investor Day 2026. It seems that one of its main collaborators for the proof of concept will be HUMAIN, a company that we were also able to meet a few months ago at another Qualcomm event.