Amon says unnamed AI firms want 100B-parameter models on phones by 2028

Qualcomm CEO Cristiano Amon said unnamed AI companies wanted phones capable of running models with at least 100 billion parameters by 2028, according to The Next Web’s account of his remarks, published Oct. 16, 2026. Amon described a goal those companies had shared with Qualcomm, not a named phone or model.

Why run AI on a phone?

Amon argued that some AI work should happen on the device because phones hold personal context and sensor data and run apps. Other AI work, he said, would remain in the cloud. The idea is to put some processing where that information and those apps already are, while leaving other tasks to remote computing.

Amon described the desired model as running “kind of all the time.” That phrase gives no operating schedule; it doesn’t specify a charging routine, either.

What 100 billion parameters means for storage

A parameter is a value a model learns during training. For a dense model with 100 billion parameters, idealized storage for the weights alone works out to 200 decimal GB at 16 bits per parameter, or 50 decimal GB at 4 bits. Those figures apply the formula explained by Aaron Mekonnen: parameter count × bits per parameter ÷ 8.

The figures cover the model weights, not a phone’s entire memory budget. Inference also uses memory for quantization metadata, higher-precision tensors, activations, runtime allocations and the key-value cache, Mekonnen explains.