AI Labs Want 100B Models on Phones by 2028, Says Qualcomm

Major artificial intelligence labs are asking chipmakers to run always-on models with at least 100 billion parameters directly on smartphones by 2028. That target was disclosed by Cristiano Amon, chief executive of Qualcomm, speaking on the Sources podcast hosted by journalist Alex Heath.

Amon did not name the specific labs making the request, though he noted that some of them are starting to build their own hardware. He pointed to existing precedents like Nubia handsets shipping with ByteDance's Doubao assistant. However, scaling to what Amon calls a looming "battle of agents" running locally requires an aggressive technical leap.

The benchmark today is far smaller. The Snapdragon 8 Elite Extreme Gen 6, introduced in late September, runs an on-device Mixture-of-Experts (MoE) architecture with 30 billion parameters, while Apple rolled out a 20-billion-parameter MoE model in June. Reaching 100 billion parameters would mean more than tripling model capacity within two chip generations—all while shifting from brief, on-demand queries to persistent, background execution.

The physical constraints start with memory capacity. Even under 4-bit quantization, a 100-billion-parameter model consumes roughly 50 gigabytes just for weights, before accounting for the key-value cache or the operating system. Typical flagship phones currently carry only a fraction of that in total system RAM.

Memory bandwidth is the second bottleneck. Generating tokens on a phone is limited by how quickly weights can be loaded from RAM into compute cores. If developers employ a sparse MoE design with around 10 billion active parameters per forward pass, a device still needs to read roughly 5 GB per token. Mobile memory channels hovering around 100 GB/s would yield roughly 20 tokens per second at best, assuming next-generation LPDDR architectures materialize.

Power consumption complicates the roadmap further. An always-on agent maintaining user context and sensor feeds burns battery far faster than answering isolated prompts. Meanwhile, hardware economics are already strained: citing a recap by TMTB, Amon noted memory prices have jumped five- to sixfold, trimming roughly 20% off smartphone sales this year, mostly in budget tiers. Shoving 64 GB or more into pocket devices would put phonemakers in direct competition with cloud data centers for scarce DRAM.

Source: habr