Qualcomm CEO Cristiano Amon expects smartphones capable of running local 100 billion parameter AI models to hit the market by 2028. Speaking in an interview with Sources, the executive said leading technology firms are already developing hardware designed for continuous on device machine learning. But current memory limits and processor thermals raise serious questions about how manufacturers will pull it off.
Amon argues that mobile phones will remain the primary gateway for artificial intelligence agents. The advantage comes down to physical access. Handsets continuously collect sensor data, user location, and private context that external hardware cannot easily replicate. Amon claims unnamed technology leaders are already prototyping systems capable of running 100 billion parameter models locally around the clock. He declined to name specific hardware partners or reveal internal architectural designs.
The technical math behind Amon claims paints a tough picture for current mobile hardware. Running a 100 billion parameter model with 4 bit quantization demands roughly 50 GB of dedicated RAM just for the model weights. Adding the operating system and background software pushes that requirement higher. Dropping down to 2 bit compression cuts the memory footprint to roughly 25 GB, but heavy compression usually damages response accuracy. Sustained execution also creates intense heat inside a fanless chassis.
Engineers could bypass memory limits using a Mixture of Experts design. Under this setup, a phone only pulls active parameters into RAM while leaving inactive weights on flash storage. The trade off is latency, as reading from flash memory slows down processing speeds. Another path involves compact models sized between 20 and 30 billion parameters, fine tuned to match larger platforms. Amon did not explain which approach manufacturers will pick. He also highlighted smart glasses as the most promising wearable category for mobile AI.
