What this is
Qualcomm CEO Cristiano Amon publicly revealed a target this week: by 2028, AI companies want phones to continuously run 100-billion-parameter (100B) large models.
What's the magnitude here? Today's mainstream phones can run 1B–7B parameter models locally — these are quantized open-source models, where parameter precision is compressed so the model can execute on-device. 100B is dozens to hundreds of times larger. Qualcomm is the largest phone chip supplier in the Android ecosystem, so its targets effectively set the direction for the industry.
This statement matters because it pushes "on-device large models" from "possible" to "a deadline-bound commitment."
Industry view
The bull case is straightforward: today, AI runs in the cloud and the phone is just a mouthpiece. Once 100B-parameter models can run continuously on-device, AI assistants can work fully offline, response latency drops from hundreds of milliseconds to tens, and private data never leaves the device. Qualcomm's NPU (neural processing unit — a chip module purpose-built for AI compute) roadmap is genuinely improving, roughly doubling performance each generation.
But the sober voices come from three places:
First, physics. Three constraints hit simultaneously: phone battery capacity, thermals, and memory bandwidth. A 100B-parameter FP16 (half-precision floating-point) model needs 200GB just for weights — far beyond any phone's memory ceiling. Even quantized to 4-bit (representing each parameter with 4 bits, compressing size to one-quarter), it still needs 50GB. "Continuous operation" without throttling requires simultaneous breakthroughs in battery and thermal technology.
Second, economics. Qualcomm's flagship chips cost tens to over a hundred dollars each. To make every phone vendor ship this tier of compute, device prices could push past 10,000 RMB (~$1,400).
Third, ecosystem. Even if the hardware lands, the training data, training compute, and user scenarios for 100B-parameter models aren't ready. Who trains an ultra-large model for the phone? Who maintains it?
Impact on regular people
For consumer markets: if this actually happens, 2028 flagship phones will hit new price highs — but voice assistants, AI photo editing, and on-device translation will be transformed: fully offline, zero latency.
For individual professionals: business travel, cross-border work, and poor-network scenarios will genuinely benefit. Privacy-sensitive jobs — lawyers, doctors, journalists — may no longer need to send conversations to the cloud.
For enterprise IT: localized AI will pressure the pricing power of many cloud AI providers, but enterprise-grade security and customization needs will still depend on the cloud in the short term.