Qualcomm outlined its artificial intelligence server strategy at the Deutsche Bank 2026 Technology Conference, highlighting progress in High Bandwidth Compute (HPC) and a new server CPU program designed to address memory bottlenecks in AI inference workloads.
Durga Malladi, Qualcomm’s Executive Vice President of Tech Planning and Edge Data Center, said the company’s HPC initiative places compute directly adjacent to memory, cutting transport latency and power consumption. The approach targets the decode stage of AI inference and can be deployed as a standalone solution, integrated with Qualcomm’s CPU and AI accelerator stack, or combined with third-party components. Commercial availability for HPC Gen 1 is slated for 2027, with HPC Gen 2 expected in 2028.
The company also provided details on its Dragonfly C1000 server CPU, which is scheduled for launch in the second half of 2028. Meta has been secured as the first customer for the platform, which will be evaluated using the SPECint 2K methodology. Malladi noted that the CPU-to-XPU ratio in data center designs has shifted from 8:1 to approximately 2:1 as agentic AI workloads expand, reflecting a broader industry transition toward more balanced compute-memory architectures.
Performance metrics shared by Qualcomm indicate HPC delivers roughly a sixfold improvement in tokens per second per watt compared with High Bandwidth Memory (HBM), which typically achieves 21 to 22 terabytes per second of bandwidth. Qualcomm’s HPC solution targets hundreds of terabytes per second of bandwidth, with some Dragonfly C1000 configurations planning more than 250 cores per rack. The Modular software stack, acquired by Qualcomm, can deliver up to 50% better performance on non-NVIDIA hardware versus native stacks, supporting a shift away from proprietary ecosystems.
Qualcomm’s market valuation stands at $171 billion with a price-to-earnings ratio of 19, a gross profit margin of 54%, and a return on equity of 34%. The company is advancing parallel architectures, including RISC-V alongside ARM, and has bolstered its data center portfolio through acquisitions such as Antenna Microsystems for RISC-V capabilities and Alphawave for custom silicon and SerDes networking technology.
Malladi, who joined Qualcomm in 1998, emphasized the company’s long-term focus on addressing the “memory wall” problem, where compute growth outpaces memory bandwidth, limiting AI inference efficiency. He noted that Qualcomm recognized this challenge around 2021 and has since restructured its R&D and product planning to commercialize solutions by 2027-2028.
