Qualcomm CEO: Phones May Run 100B AI Models by 2028 [2026] - shattered.io

Qualcomm's 2028 roadmap for 100B-parameter on-device models is less a product announcement than a structural redefinition of mobile SoC compute ceilings. The technical cascade is immediate: NPUs must exceed 200 TOPS, memory bandwidth must leap to LPDDR6 or HBM-lite architectures, and this directly accelerates the capacity race at TSMC (Taiwan, China) on 2nm nodes and advanced packaging. Upstream, Samsung and SK Hynix HBM-lite lines become the new bottleneck; downstream, the developer ecosystem shifts from cloud-API dependency to a local-inference-plus-cloud-finetuning hybrid paradigm. On compliance, if US export controls extend to on-device AI compute thresholds, Snapdragon flagships could face scrutiny comparable to data-center GPUs, inflating supply-chain compliance costs materially. Competitively, Apple's A-series already ships 7B-class on-device models. Qualcomm's framing reads as a defensive counter to Apple's "AI phone" narrative rather than a genuine first-mover claim. MediaTek, leveraging TSMC capacity headroom, is positioned to launch a price war in the 2027 window. The real variable over the next 18 months isn't raw FLOPS—it's thermal design and battery endurance under sustained inference loads. That determines whether 100B is a spec-sheet number or a daily-driver experience.