Qualcomm CEO Cristiano Amon said AI companies are requesting the capacity to run 100-billion-parameter large language models continuously on devices by 2028, a demand that signals a fundamental shift in mobile chip requirements and device economics.
The request reflects AI companies' strategic pivot from pure software and cloud services toward hardware integration. By embedding their models directly on devices, they reduce reliance on expensive cloud GPU infrastructure while improving data privacy—a calculus that makes economics work for both the model developers and the chipmaker.
For Qualcomm, the 2028 target represents a substantial engineering and capital challenge. Current flagship mobile chips run smaller models, typically 7 billion to 13 billion parameters. Scaling to 100 billion parameters while maintaining mobile power envelopes requires major advances in neural processing unit architecture, memory bandwidth and thermal management.
Qualcomm's Snapdragon platforms already dominate mobile silicon, with integrated NPUs designed for on-device AI. The company's licensing model—earning royalties on intellectual property while selling chips to a broad range of manufacturers—positions it to capture demand from both traditional phone makers and new entrants. Competitors like Apple design their own silicon; MediaTek competes on cost. Qualcomm's advantage lies in its depth on power efficiency and an established software ecosystem for edge AI.
The economic case is clear: local LLM execution cuts operational expenses and latency compared to cloud routing. For AI companies weighing vertical integration into hardware, the margin math on high-volume devices makes the bet rational. Whether Qualcomm can deliver the required performance in a mobile power budget by 2028 will determine which companies win the next cycle of device-level AI.
