Table of Contents
Snapdragon Summit showed how persistent context, sensory wearables and heterogeneous software could connect the endpoints brought together by Qualcomm’s broader diversification strategy
The first day of Snapdragon Summit provided a framework for understanding the Qualcomm diversification strategy in the context of agentic AI. If agents rather than smartphones increasingly sit at the center of a user’s computing experience, Qualcomm’s presence across phones, PCs, cars, wearables and infrastructure starts to look less like the components of a common computing architecture.
Following the day one vision keynote, speakers on day two dealt with the harder question of how that architecture actually works. An agent that moves between devices needs to maintain context about the user, understand the capabilities available around it and determine where a particular workload should run. That makes personal AI fundamentally a distributed systems problem spanning models, memory, sensors, processors and cloud infrastructure.
Context becomes the connective tissue
Liquid AI CEO Ramin Hasani opened the day with Qualcomm CEO Cristiano Amon and identified context as one of the key requirements for bringing capable agents onto devices. Users, Hasani said, will expect experiences and preferences developed on one device to carry over to another.
Liquid AI used the Summit to announce Liquid Context on Snapdragon, a software layer designed to read, compress and write information generated across a user’s interactions with devices. The concept combines memory with locally running agents, with Hasani demonstrating an agent executing on the Snapdragon Hexagon NPU and accessing user-permissioned information including notes, messages and photos.
Today, applications largely maintain separate stores of user information. Agentic computing requires some mechanism for maintaining persistent state across applications, models and devices. Qualcomm is working on the same problem from another direction with Snapdragon START (scalable, turnkey AI-ready toolkit), its program for accelerating personal AI devices. START combines Snapdragon-based hardware modules, software spanning the device, companion smartphone and cloud, and manufacturing support. During the keynote, Qualcomm said the software can select between different AI models while maintaining a single shared memory so context persists as workloads move between them.
In other words, the agent is portable as is its understanding of the user.
Every device becomes an edge node
Qualcomm’s compute chief Kedar Kondap described the inference-placement problem in economic terms, noting that “sending every single token to the cloud simply doesn’t work.” But that doesn’t mean Qualcomm sees cloud and edge AI as mutually exclusive. The architecture described during the keynote is explicitly hybrid. Large models and deep reasoning can remain in the cloud, while latency-sensitive, privacy-sensitive and continuously available workloads increasingly execute on phones, PCs, cars, robots and other devices.
Qualcomm ultimately sees devices orchestrating among themselves and dynamically determining whether a task should run locally, on another nearby device, at the edge or in the cloud.
This is a more useful way to think about edge AI than simply asking whether a model can fit on a device. The relevant unit becomes the workload; which model, what context it needs, what latency is acceptable, what information should remain local and what compute resources are available at that moment are the determinants.
The expansion of Snapdragon X reinforces that strategy. Qualcomm now has Snapdragon client platforms spanning Windows and the newly introduced Googlebook, and on Day 2 announced official Linux support beginning with Snapdragon X2. Debian support is targeted for the end of 2026 and Canonical is targeting Ubuntu certification in the first half of 2027.
The strategic point is not another operating system. It is increasing the number and variety of capable compute nodes available to an agent.
Wearables become the sensory layer
That helps explain Qualcomm’s heavy emphasis on wearables. Ziad Asghar, who leads Qualcomm’s XR, wearables and personal AI business, said, “Wearables are the foundation for context-aware intelligence.”
Glasses can provide visual context. Earbuds can provide an always-available voice interface. Watches, rings and other body-worn devices can contribute additional sensor data. None necessarily replaces the smartphone; instead, they add new streams of information and new interfaces into the same personal AI system.
The new Snapdragon Sound Elite Gen 2 platform is a good example. Qualcomm designed the platform around on-device AI, audio, sensing and integrated micro-power Wi-Fi 6E, enabling audio wearables to connect directly with agents and cloud services. It supports traditional earbuds and headphones as well as audio glasses and camera-enabled devices. Qualcomm says the platform provides up to twice the AI capability of its predecessor, with up to 40% lower power consumption and a footprint up to 30% smaller, based on its own testing and simulations.
This potentially creates a different device dynamic. Agentic AI does not necessarily consolidate computing into fewer endpoints. It could produce more of them.
Software determines whether the architecture coheres
Hardware heterogeneity creates its own problem. CPUs, GPUs, NPUs and specialized accelerators have different programming environments and performance characteristics; distribute AI across devices and data centers and that complexity compounds.
Chris Lattner, now Qualcomm’s EVP of Advanced AI Software and Platforms following the completed acquisition of Modular, described modern AI as a “full stack distributed systems problem.”
This is where Modular becomes strategically important. Qualcomm describes its platform as a silicon-agnostic software layer capable of deploying workloads across CPUs, GPUs, NPUs and custom silicon. Modular’s Mojo language, MAX platform and Modular Cloud remain distinct products following the acquisition, but Qualcomm plans to extend the software across its own portfolio from Snapdragon devices to data center infrastructure.
That may ultimately be as important to Qualcomm’s convergence strategy as Oryon or Hexagon. A heterogeneous hardware portfolio becomes considerably more valuable if developers can treat it as a coherent compute environment rather than a collection of separate targets.
Day one of Snapdragon Summit made the case that Qualcomm’s diversified businesses are converging around agentic AI. Day two began exposing the plumbing necessary to make that idea real. More and more inference will obviously move from the cloud to the edge. Personal AI extends that distribution in a way that lets context follow the user, inference runs wherever economics and technical requirements dictate, and software abstracts the heterogeneous hardware underneath it.
As this computing model materializes, Qualcomm’s opportunity is to provide many of the nodes — and increasingly some of the connective software — that allow those devices to function as one system.
Author’s note: There was very little 6G discussion during Snapdragon Summit keynotes. However, Qualcomm hosted a 6G Leadership Day event at its San Diego headquarters ahead of the Maui confab. For insights from that event, read here and here.
