At its annual Investor Day, Qualcomm Technologies, Inc. announced a major expansion into full-stack data center infrastructure, debuting a roadmap of AI-focused processors, custom silicon solutions, and high-speed connectivity products.
The hardware portfolio—spanning the Qualcomm Dragonfly C1000 CPU, Dragonfly AI300 inference accelerator, and an architecture dubbed Qualcomm High Bandwidth Compute (HBC)—marks the company’s formal push to challenge incumbent silicon providers in hyperscale data center infrastructure. Alongside the product reveals, Qualcomm announced a multi-generation agreement to supply its upcoming data center CPUs to Meta.
The Silicon Portfolio: Powering Agentic AI
Qualcomm is positioning its new lineup as an “inference-first” platform optimized for agentic AI—autonomous AI workloads requiring high token throughput, continuous reasoning, and minimal latency.
Qualcomm Dragonfly C1000 CPU
The bedrock of the compute announcement is the Dragonfly C1000, a purpose-built data center CPU.
- Core Architecture: Built using custom-designed Qualcomm Oryon CPU cores, the chip utilizes a multi-chiplet architecture featuring a core count of over 250 cores.
- Clock Speeds: Core frequencies are rated at greater than 5 GHz to handle high-throughput agentic orchestration and general-purpose server workloads.
- I/O and Interconnects: The architecture delivers greater than 2 TB/s of PCIe Gen 7 and CXL connectivity to facilitate next-generation memory disaggregation.
- Efficiency: Qualcomm estimates the platform will achieve greater than 2x better performance-per-watt compared to existing benchmarked competitive server CPUs.
- Timeline: Commercial availability for the Dragonfly C1000 is targeted for 2028.
Qualcomm High Bandwidth Compute (HBC)
To bypass traditional memory data movement bottlenecks, Qualcomm introduced High Bandwidth Compute (HBC), a near-memory 3D-stacked silicon architecture intended as an alternative to High Bandwidth Memory (HBM).
According to Qualcomm, HBC is designed to offer a 6x increase in bandwidth-per-watt compared to standard HBM specifications normalized at the card level, alongside a 200x increase in capacity-per-watt versus SRAM at the rack level.
- HBC Gen 1: Integrated into the previously announced AI250 accelerator, it is designed to achieve 133 TB/s per card—an 18x effective memory bandwidth increase over LPDDR5X-based AI200 solutions. Commercial sampling begins in mid-2027.
- HBC Gen 2: Integrated into the newly announced AI300, it is designed to offer a 54x increase in effective bandwidth over the AI200 line.
Qualcomm Dragonfly AI300 Accelerator
The Dragonfly AI300 represents Qualcomm’s third-generation rack-level AI inference platform. Compatible with Ultra Accelerator Link (UALink) and Ethernet for Scale-Up Networking (ESUN), the AI300 is designed for large language model (LLM) and multimodal model inference. The platform is projected to yield a 4x to 8x performance-per-watt advantage over existing GPU-based architectures on memory bandwidth metric scales. Commercial sampling is expected in 2028.
Custom Silicon and High-Speed Connectivity
Expanding beyond standard platforms, Qualcomm disclosed an end-to-end custom silicon business to design and manufacture bespoke chips for specific cloud infrastructure workloads. This initiative leverages the company’s existing IP portfolio and packaging pipelines to manage execution risks from initial design to high-volume manufacturing.
Supporting this infrastructure is a new optical and copper connectivity portfolio. Capable of handling high-bandwidth 800G and 1.6T networking, these interconnects scale from short-reach intra-data-center links up to campus-reach deployments spanning 20 kilometers. The networking stack utilizes Qualcomm’s internal SerDes, PAM4, and coherent-lite DSP technologies to maintain signal integrity across distributed, disaggregated environments.
Meta Partnership and Ecosystem Adoption
A cornerstone of Qualcomm’s data center strategy is a newly solidified, multi-year collaboration with Meta. Under the agreement, Qualcomm will act as a silicon supplier for Meta’s data center CPUs, with the Dragonfly C1000 slated to power portions of Meta’s next-generation scale-out server fleet.
Beyond Meta, Qualcomm highlighted ecosystem backing from over 35 technology hardware, infrastructure, and semiconductor partners. The list includes server manufacturers, memory vendors, and testing providers such as:
- Infrastructure & Hardware: Supermicro, Lenovo, GIGABYTE Technology, Foxconn, Arista, and Quanta.
- Memory & Storage: Samsung SDS, Micron Technology, SK hynix America, and Nanya Technology.
- Semiconductor & Design Ecosystem: Microchip Technology, Advantest, Teradyne, and UMC.
Qualcomm stated that it remains committed to an annual cadence for its data center roadmap, focusing future updates on compounding metrics for AI inference execution, energy efficiency, and total cost of ownership reduction.
Cristiano Amon, President and CEO, Qualcomm Incorporated, said:
Agentic AI is driving a significant increase in demand for AI inference in the data center. As these become the dominant workloads, infrastructure has to deliver much higher performance at lower power and cost. That plays directly to Qualcomm’s strengths, and we’re well positioned for this shift. With Qualcomm Dragonfly, we’re bringing our high-performance, low-power computing into the data center, with multi-year, multi-generation agreements with leading customers.