Qualcomm predicts smartphones capable of running powerful local AI models will arrive by 2028

Qualcomm CEO Cristiano Amon has shared a forecast: within a few years, mobile devices could hit the market that are capable of handling large language models with around 100 billion parameters. According to him, a number of companies are currently working on such devices.

However, a series of serious technical hurdles stands in the way of making this a reality. The key problem is the shortage of memory chips. To run a 100-billion-parameter model with 4-bit quantization, roughly 50 GB of RAM would be required. This figure does not even account for the memory needed to keep the operating system, applications, and the neural network’s context running. If 2-bit quantization is applied, the footprint could be reduced to about 25 GB. High component costs and heat dissipation issues add further complications. Sustained operation of a large neural network places a heavy load on the system-on-chip — primarily the CPU and NPU — which can lead to overheating and performance drops due to throttling.

Amon sees several possible ways out of this situation. The first is the use of more aggressive quantization methods, which would lower memory requirements. However, the trade-off would be a decline in the model’s capabilities, especially in tasks requiring complex computation and reasoning. The second path involves a MoE (Mixture of Experts) architecture, in which only the active portion of the neural network resides in RAM, while the remaining blocks are stored in flash memory and loaded on demand. The downside of this approach is that NAND storage operates noticeably slower than DRAM, meaning AI performance could suffer.

The most practical option, according to Qualcomm’s chief, is to bet on less bulky models — in the 20–30 billion parameter range. Following the Pareto principle (20% of effort yields 80% of results), such networks could come very close in quality to their larger counterparts while placing far more modest demands on the system.