Qualcomm CEO says phones may run 100 billion parameter AI models by 2028 - Global Sources

Qualcomm's 2028 roadmap for on-device 100B-parameter inference is less a product forecast than a supply-chain stress test. The technical implication is severe: sustaining that model locally demands NPU throughput beyond 500 TOPS and memory bandwidth pushing past LPDDR6 into HBM-lite territory, making advanced packaging capacity in Taiwan, China a single-point-of-failure bottleneck. Geopolitically, tightening US export controls on sub-3nm tooling amplify Qualcomm's exposure to TSMC's node schedule; by 2026 the company will likely accelerate qualification of Samsung's 2nm GAA as a hedge, compressing its margin structure. Competitively, Apple's in-house NPU trajectory and MediaTek's Dimensity 9500 cycle are closing the gap on raw AI FLOPS. If Qualcomm delivers, on-device LLM shifts from marketing copy to a spec-sheet line item, forcing the entire Android SoC ecosystem into a forced refresh cycle. The real 18-month variable, however, is not the silicon itself but whether memory suppliers can hold 12 GB LPDDR6 under 3 W sustained draw. That power envelope, not peak TOPS, is the kill criterion for 100B-parameter inference on a 5,000 mAh battery. Whoever cracks the memory wall first owns the 2028 mobile AI narrative.