Qualcomm HBC takes aim at HBM costs in AI data centers - digitimes

Qualcomm unveiled its latest AI data center platform Dragonfly at its annual investor day, introducing a new technology called High Bandwidth Compute (HBC) as a key strategy to compete against NVIDIA,... Read original →

Qualcomm’s HBC architecture is a calculated bypass of the HBM supply bottleneck. With HBM reliant on TSV and TSMC’s CoWoS packaging—constrained, costly, and geopolitically sensitive—HBC’s promise of comparable bandwidth via standard packaging could erode NVIDIA and AMD’s memory-driven AI dominance. This pressures memory vendors to revive alternatives like HMC or LPDDR6X and intensifies competition for advanced packaging capacity, especially as the U.S., Japan, and South Korea push semiconductor localization. NVIDIA will likely double down on CUDA lock-in rather than engage in a hardware arms race, while AMD may accelerate HBM3E integration in MI300X to defend its cost-performance edge. Within 18 months, if HBC proves viable in inference workloads, AI data centers could bifurcate into HBM for training and HBC for cost-sensitive inference—decoupling memory and compute in a way not seen since GDDR’s rise.