Qualcomm CEO says AI firms want 100B-parameter models running on phones by 2028 - Crypto Briefing

Qualcomm's 2028 target for 100B-parameter on-device inference is less a roadmap than a supply-chain ultimatum. The binding constraint is not NPU FLOPS—2nm-class silicon already clears the throughput bar—but memory bandwidth and thermal envelope. Even at 4-bit quantization, a 100B model demands roughly 50 GB of resident weights, shattering the 8–16 GB ceiling of current LPDDR5X. Expect HBM-class stacking to migrate into flagships by 2026–2027, making advanced-packaging capacity in Taiwan, China the scarcest input in the mobile AI stack. On compliance, tightening US export controls amplify Qualcomm's single-point dependency on that same packaging ecosystem, while the EU AI Act's transparency mandates for on-device inference add certification overhead that mid-tier rivals absorb more easily. Strategically, Apple's unified-memory architecture is structurally better positioned for large local models; MediaTek will undercut on good-enough inference. Qualcomm's moat is migrating from baseband-plus-GPU toward an AI inference orchestration layer. If it fails to secure next-gen memory IP by 2027, flagship ASP premiums erode. Twelve-to-twenty-four-month outlook: hybrid cloud-edge inference becomes the default paradigm. Pure on-device 100B before 2028 is marketing narrative, not engineering reality. The real inflection: a commercial phone with >32 GB unified memory by Q3 2026.