Qualcomm CEO Says Leading AI Companies Demand Phones to Run 100 Billion Parameter Models Continuously by 2028
Qualcomm CEO Cristiano Amon stated in an interview on October 8 that leading AI companies have requested Qualcomm to ensure that smartphones can continuously run large models with at least 100 billion parameters by 2028.
Amon mentioned that these companies want the models to run continuously rather than shutting down after completing a single task, and they should possess personal context understanding, reasoning, and autonomous action capabilities. He did not name specific companies but noted that some AI firms have begun to manufacture their own phones.
Currently, Qualcomm's latest Snapdragon 8 Elite Gen 6 and Extreme Gen 6 platforms (2nm process) can run hybrid expert (MoE) models with up to approximately 30 billion parameters on devices. The leap from 30 billion to 100 billion represents more than a threefold increase, which raises higher demands on smartphone memory, computing power, and battery life.
Amon described this as a shift towards always-on "agent-based" AI, where software acts on behalf of users rather than waiting for application clicks. He also mentioned that some consumers prefer AI and data to remain on the device and under their control.
Qualcomm previously emphasized at the Snapdragon Summit that agent-based AI will become the center of the experience; smartphones will not disappear but will revolve around the role of agents. The new platform also supports local capabilities to process approximately 1 million tokens daily.
This demand will significantly raise the requirements for terminal computing power and memory, putting greater pressure on chip and storage supply chains under event-driven circumstances. Capital and attention are shifting towards higher-performance mobile platforms and memory, benefiting chip manufacturers and high-end models that can meet the needs of always-on large models, while putting pressure on mid-range and low-end models and overall shipments affected by rising memory prices.
Amon stated that these new capabilities are expected to release the demand for upgrading smartphones that has been delayed due to rising memory prices, driving a super cycle in the smartphone market.
Source: Public Information
ABAB AI Insight
Qualcomm has long dominated flagship Android chips, recently enhancing edge-side MoE model capabilities to approximately 30 billion parameters with the Snapdragon 8 Elite Gen 6, while also expanding into glasses, PCs, and data centers; Amon positions smartphones as the core terminal for agent-based AI rather than being replaced.
In terms of capital and demand pathways, leading AI companies are directly requesting chip manufacturers to provide always-on 100 billion parameter capabilities, with some beginning to develop their own phones; the motivation is to keep personal context and agent capabilities on the device, reducing reliance on the cloud and controlling data.
Similar cases include the early transition of smartphones from application-centric to platform-centric, and the current trend of cloud-based large models moving to edge devices; we are currently at a stage where agent-based AI demands an order of magnitude higher requirements for continuous local inference.
Essentially, this represents a restructuring of the industry chain and a shift in pricing power: moving large models from centralized cloud inference to always-on smartphones makes computing power, memory, and battery new bottlenecks, shifting value from pure cloud services to edge-side chips and high-end hardware, while compressing the competitiveness of mid-range and low-end models.
ABAB News · Cognitive Laws
- Once agents require always-on capabilities, edge-side computing power becomes a mandatory threshold.
- The leap from 30 billion to 100 billion is not an upgrade, but a redefinition of the energy consumption and memory boundaries of smartphones.
- Cloud capabilities sinking into pockets also means that supply chain bottlenecks sink into every device.