Qualcomm Technologies outlined a data center roadmap built around its Qualcomm Dragonfly portfolio, spanning a new CPU platform, a near-memory compute architecture, an inference accelerator roadmap, and high-speed connectivity aimed at rack-scale AI inference.
Qualcomm Dragonfly C1000 CPU
The new Qualcomm Dragonfly C1000 is a data center CPU designed for agentic, general-purpose, and AI head node workloads. Qualcomm lists custom Qualcomm Oryon CPU cores with frequencies above 5 GHz, a 250+ core-count chiplet design, and a multi-chiplet architecture intended to support modular integration with advanced packaging for performance and IO scaling.
On the IO side, Qualcomm specifies more than 2 TB/s of PCIe Gen 7 connectivity plus CXL support, targeting accelerator attach, high-speed networking, storage, and memory disaggregation. The platform also includes RAS features such as ECC, fault isolation, and error recovery. Qualcomm says the CPU supports both air and liquid cooling, with deployments in OCP ORv3-compliant racks and servers. Commercial availability is expected in 2028.
High Bandwidth Compute (HBC): near-memory compute
Qualcomm also introduced Qualcomm High Bandwidth Compute (HBC), a 3D-stacked, near-memory compute architecture that bonds compute with highly accelerated memory bandwidth. The goal is to address data movement limits that show up hard in inference, where memory bandwidth and access energy can dominate token economics.
For HBC Gen 1, Qualcomm says the Dragonfly AI250 is designed to deliver 133 TB/s per card and an 18x increase in effective memory bandwidth compared to AI200 with LPDDR5X. For HBC Gen 2, Qualcomm says AI300 is designed for a 54x increase over AI200. The company also claims HBC is designed for a 6x increase in bandwidth per watt versus HBM (normalized at card level) and a 200x increase in capacity per watt versus SRAM (normalized at rack level), both based on competing published specifications. Commercial sampling of HBC Gen 1 with AI250 is expected in mid-2027.
Qualcomm Dragonfly AI300 inference accelerator
Qualcomm Dragonfly AI300 is the company’s third-generation rack-level AI inference platform, following AI200 and AI250 introduced last October. Qualcomm lists both air- and direct-liquid-cooled configurations and says AI300 integrates HBC Gen 2 for compute acceleration with integrated memory and higher effective memory bandwidth (AI250 uses HBC Gen 1). Qualcomm expects 4x–8x better performance per watt versus existing GPU-based architectures on memory bandwidth per watt per card.
For scale-up, Qualcomm lists UALink (Ultra Accelerator Link) and ESUN (Ethernet for Scale-Up Networking); for scale-out, it lists copper and optical. Commercial sampling is expected in 2028.
Custom silicon, and 800G to 1.6T connectivity
Beyond standard platforms, Qualcomm described custom silicon work covering end-to-end co-design across silicon, system, and software, plus advanced packaging and modular architectures. On connectivity, Qualcomm listed die-to-die, copper, optical, and campus-reach interconnects, including 800G and 1.6T across optical, AOC, and AEC, with deployments up to 20 km. For data center architects, the practical point is that Qualcomm is describing a full-stack bill of materials (compute, memory architecture, accelerators, and links), not a single card—though most of the timeline here is still “sampling” and “expected” dates.
Qualcomm also said it has a multi-year, multi-generation agreement with Meta under which Qualcomm Technologies is planned to supply data center CPUs, and that the Dragonfly C1000 is planned to power Meta’s next-generation server fleet.
Source: Qualcomm
