Qualcomm targets $15B data center revenue by 2029 via Dragonfly; High-Bandwidth Compute claims 6x HBM efficiency

Qualcomm unveiled its Dragonfly platform at its June 25 Investor Day, a full-stack data center infrastructure push including the Dragonfly C1000 CPU (250+ cores running at 5GHz), AI200/AI250/AI300 accelerators, High-Bandwidth Compute (HBC) memory technology, and custom silicon, targeting $15 billion in data center revenue by fiscal 2029. The strategy represents Qualcomm's formal pivot from a handset-centric semiconductor company to a systems-level player competing directly with NVIDIA, AMD, and cloud-native ASIC producers across the full AI inference stack.

HBC's core innovation places compute cores directly beneath LPDDR DRAM stacks via through-silicon vias, eliminating the expensive silicon interposer required by HBM systems. Qualcomm claims HBC delivers 6x the bandwidth per watt versus HBM, 200x the capacity per watt versus SRAM, and up to 8x more tokens-per-watt than traditional GPU configurations. AI250 equipped with HBC Gen 1 achieves 133 TB/s of memory bandwidth per card, an 18x jump from the AI200. Commercial sampling begins mid-2027 with AI250, followed by HBC Gen 2 in AI300 on an annual cadence.

Meta and Microsoft have endorsed the architecture as early adopters. Humain (Saudi Arabia) committed to deploying 200 MW of Qualcomm racks, signaling deployment at power-plant scale. Qualcomm also acquired Modular for $3.9 billion to challenge NVIDIA's CUDA dominance with an open, architecture-agnostic software stack. For architects evaluating memory-bandwidth-constrained AI workloads, Qualcomm's HBC eliminates a core bottleneck in inference scaling. The combination of commodity LPDDR instead of scarce HBM plus 6x power efficiency improvement could reshape TCO calculus for long-context and agentic AI deployments.