Three lines outside the venue.

The first line

The first line has formed near the loading dock. A man in a faded brown jacket has opened his laptop on his knee and is editing code. He squints at the screen. Types. Deletes. Types again. He does not look up.

The second line snakes toward the main entrance. A woman in a black polo tests a microphone. "One, two," she says. Pauses. "One, two."

The third line is just two people, standing apart from the others, looking at the screen above the door: Snapdragon Summit.

The main hall is dark.

By 8, every seat is taken. The lights come up and Cristiano Amon walks out. He wears a navy quarter-zip and dark jeans. He does not wave. He stands at the center of the stage, looks at the audience for a moment, and speaks.

"The most exciting thing about this Snapdragon Summit," he says, "is that I think we have finally entered a stage where the shape of an AI phone is clearly visible."

He pauses. The crowd does not clap. They lean forward.

They did not come for another chip. They came for an answer.

When AI moves from a phone feature to an agent that understands intent across devices, keeps context, and acts for the user, how do CPU, GPU, NPU, memory, and connectivity get rearranged? How do phones, PCs, earbuds, and glasses relate?

Qualcomm's answer filled two days of slides.

I. The phone now serves two users

For twenty years, phones were built around apps. You open one, do one thing, switch to another. The device waits for a command, then responds.

The agent era starts with intent. You do not need to know which app does what. You say what you want. The agent understands, breaks it down, calls tools, and executes.

Amon puts intent at the center of the next phone experience. The interaction changes. The execution changes. The chip's job changes.

"The earliest Snapdragon 8 series platforms had to answer whether a phone could run an advanced OS at 1GHz or higher while lasting a full day," Amon says. "Today the challenge is similar, except it has been replaced by an always-on agent."

The device now serves two users: the person and the person's agent. The agent has its own workflow. Qualcomm has to design a computing architecture that serves both.

The 6th Gen Snapdragon 8 Ultra Premium Edition splits work around that agent workflow. The CPU acts as planner, with a main frequency above 5GHz, a first for the industry. The NPU expands local memory tightly coupled to compute units, cutting frequent access to external DDR. It runs MoE models with up to 30 billion parameters on device.

Qualcomm also launched two flagship platforms at once: the 6th Gen Snapdragon 8 Ultra Premium Edition and the 6th Gen Snapdragon 8 Premium Edition. Both use a 2nm process. The Ultra Premium Edition pushes the limits of agentic AI, gaming, and imaging. The Premium Edition brings the same core architecture and AI features to more flagship phones.

The chip competition is no longer about peak compute alone. It is about organizing perception, reasoning, planning, and execution into one system that runs on low power.

II. One system for perception, reasoning, planning, execution

Both chips use the new Qualcomm Oryon CPU. Two ultra-large cores run up to 5GHz. Six performance cores run up to 4GHz. The Ultra Premium Edition gains 13% CPU performance and 37% energy efficiency over the previous generation. Both chips use the same CPU architecture, implementation, and frequencies. A new cache architecture, FlexCache, lets the CPU share cache resources so the agent's response speed does not drop during heavy tasks.

The Ultra Premium Edition's Adreno GPU gains 44% performance and 40% energy efficiency. It adds Adreno Matrix Cores for AI and Adreno Neural Fusion. AI compute folds into the graphics pipeline, supporting super-resolution, frame generation, and image reconstruction. Gaming, imaging, and AI no longer run on separate pipelines.

The Hexagon NPU supports MoE models with more than 30 billion parameters. Performance is up 35% over the previous generation. Performance per watt is up 33%. Qualcomm adds an Element Accelerator for Transformer workloads. Hexagon's shared memory expands by 50%. More context and KV-cache stay inside the NPU. That cuts DDR access, lowers latency, frees memory bandwidth, and speeds up inference.

Long context, complex reasoning, more tool calls, and concurrent tasks push memory back into the center. The Ultra Premium Edition's Adreno slice architecture pairs with 18MB of dedicated high-speed graphics memory. That cuts power by 12%.

The sensor hub is redesigned. Two Micro NPUs deliver 85% more performance and 20% better energy efficiency. The hub supports models with up to 200 million parameters. Even in a low-power state, the phone can process ambient sound and other sensor tasks, giving the agent always-on input.

Amon previews one more item. HBC, which already appeared in data centers, will enter Snapdragon platforms as a coprocessor. It pulls memory and AI compute closer. More details come at MWC Barcelona 2027.

III. Agents start scheduling devices

Qualcomm used to think about AI from devices. A phone had imaging features, so AI made image processing better. A PC had more compute, so AI made users more productive.

Now Qualcomm starts from the personal agent and asks what each device can do for it. What services can a phone, PC, earbud, or pair of glasses provide? What capabilities do those services need?

The phone is still the core node. But the competitive question for this year's flagships has changed. Xiaomi 18 Pro debuts the 6th Gen Snapdragon 8 Premium Edition. NPU prefill performance on some models improves by up to 80%. It runs HyperOS 4 and Super Xiao Ai 2.0. Xiaomi has pushed AI from Q&A to task orchestration, cross-app execution, and cross-device collaboration. Phone makers now ask: what can this phone let an agent complete for the user?

The PC is becoming the agent's productivity node. PCs powered by the Snapdragon X2 series are what Qualcomm calls "agentic PCs." Snapdragon X2 Elite reaches 85 TOPS of AI performance. PCs can now run agents locally for long periods. Qualcomm is expanding Snapdragon X2 support to Linux, starting with a Debian developer preview by the end of 2026 and Ubuntu certification targeted for early 2027.

PC makers are also redefining their role. Google launched Googlebook, combining Gemini Intelligence, Android, and the PC, with cross-device experiences between PCs and Android phones. Microsoft continues to strengthen Surface's local AI. The Surface Pro 12-inch and Surface Laptop 13-inch shown at the summit both run Snapdragon X2 Plus. Local AI inference is up 80% over the previous generation. Battery life is over 15 hours.

Some players outside the traditional PC camp are redefining the computer. Qwen Book calls itself an "agent computer." It uses a tablet-plus-keyboard form and redesigns interaction around an AI entry point, Skills, and other mechanisms.

Earbuds and glasses are changing too. In the past, audio chips made earbuds sound better, connect more stably, cut latency, and improve noise cancellation. Qualcomm calls the 2nd Gen Snapdragon Audio Platform Premium Edition the first audio platform built for agent experiences. It combines high-quality audio and voice, on-device intelligence, and connectivity. It introduces Snapdragon Sound apps and agents, letting earbuds, earphones, and glasses connect directly to AI assistants, services, and content.

The numbers: the 2nd Gen Snapdragon Audio Platform Premium Edition doubles AI capability. Peak AI compute is 128 GOPS. It executes 256 MAC operations per cycle. It integrates ultra-low-power Wi-Fi 6E, letting earbuds and glasses connect directly to cloud AI services. Power consumption drops by up to 40%. Platform size shrinks by up to 30%.

Earbuds go from a playback device to an always-on sensing and interaction entry point for the personal agent. You do not need to pull out a phone, open an app, and type a prompt. Voice interaction is direct. Real-time translation, environmental understanding, contextual assistance, and voice interaction all run.

Qualcomm also showed a concept device with Cleer: earbuds with a camera. Qualcomm stressed that these camera-equipped audio devices are not for photos. They continuously capture the context of the user's environment. The agent knows what the user is looking at and what scenario they are in. It can then provide more real-time assistance. Compared with smart glasses that have displays and fuller visual processing, camera-equipped audio devices emphasize lightweight context capture.

Qualcomm also introduced wearable forms such as the AI Pin. A person expresses intent, and the device summons the agent.

The core of next-generation personal computing may no longer be a single device. It may be contextual information and personal intelligence that persist across devices as the user moves. Wherever intelligence flows, it needs an underlying layer of compute, connectivity, and sensing.

IV. Connectivity, 6G, and software

If future computing is no longer centered on devices but on a person's agent, the boundaries of computing must be reset.

In the future, people may perceive the real world through glasses, hear the environment through earbuds, learn information through phones, handle complex tasks on PCs, access larger maps in cars, and hand more complex reasoning to the edge or cloud. Agents need to flow across devices and allocate compute resources according to task complexity.

Qualcomm wants to recombine capabilities accumulated over decades. The problem it is beginning to solve is: how can an agent find the right place in the entire computing network to complete a task and keep running?

Doing high-performance computing at extremely low power is a relatively easier question for Qualcomm. Starting from phone SoCs, one of Qualcomm's long-term core problems has been how to pack more compute into mobile devices under limited power, size, and thermal conditions.

Today, the new sensor hub carries two Micro NPUs. It can continuously process ambient sound and other sensing tasks in a low-power state. The 2nd Gen Snapdragon Audio Platform Premium Edition doubles AI performance and cuts power by 40%, letting earbuds stay online all day. On phones, the Ultra Premium Edition's CPU and GPU improve energy efficiency by 37% and 40% respectively. Offset PoP packaging and the 2nm process push sustained peak performance to twice the previous level.

In the agent era, connectivity is no longer just a channel for transmitting compute results. It begins to participate in deciding where compute happens.

The new Snapdragon Mobile Platform Premium Edition carries Qualcomm FastConnect 8800. It brings 4×4 Wi-Fi into the mobile connectivity system for the first time. It supports Wi-Fi 8, reaches a peak rate of 11.6Gbps, and adds capabilities such as proximity-aware AI. Qualcomm's thinking about Wi-Fi 8 is no longer just "faster." It is about bringing AI-native intelligence to connectivity to meet real-world scenarios created by current AI demands.

6G is defined in Qualcomm's technology roadmap as AI-native infrastructure. Devices can use faster uplinks to send visual, audio, and environmental information from the real world into the network in real time. That gives agents more complete context. Compute no longer concentrates at a central node. It extends from terminals to the edge, base stations, regional data centers, and large data centers. It is dynamically allocated according to tasks and environments.

Amon has said before that AI is becoming a core driver of wireless data traffic growth, making 6G R&D and deployment inevitable. As a generation of wireless technology designed for AI, 6G will be built around three pillars: connectivity, computing, and sensing. It will provide the industry with critical data processing capabilities needed for AI model evolution. It will create new services based on sensing infrastructure. It will push networks from simple connectivity carriers to core computing systems supporting many industries.

The software stack decides how far the agent can go. Qualcomm filled that gap with its July acquisition of Modular. The company wants developers to deploy models to Snapdragon, Dragonwing, and Dragonfly in a more unified way.

V. Qualcomm's boundary moves from processors to agent infrastructure

From Snapdragon to Dragonwing to Dragonfly, Qualcomm is building an AI computing continuum across personal terminals, edge devices, and data centers.

Qualcomm's past strength was getting connectivity and compute into more devices. In the agent era, it must prove how to keep intelligence flowing across those devices.

In chip industry history, the next market is often decided not by a performance upgrade but by who redefines their capability boundary first when the computing model changes. ARM wants to move from an IP company into a broader computing world. Intel is looking for a new computing position beyond the PC.

Qualcomm has given its answer first on the device side. From phones and PCs to earbuds and glasses, from low-power sensing and heterogeneous computing to connectivity and the AI software stack, Qualcomm is moving AI from a feature on a single chip to the core logic running through its entire roadmap. Its goal is no longer just to provide faster processors. It wants to become the compute and connectivity foundation for agents running across terminals, edge, and cloud.

In the smartphone era, Qualcomm sold the chip that connected phones to the internet. In the agent era, it wants to sell the layer that keeps intelligence running, no matter which device the user picks up.

On the last day

On the show floor, a pair of camera-equipped earbuds was connected to Wi-Fi 6E. A small light next to the camera was on. On the phone screen, real-time translated subtitles scrolled line by line.

Amon did not say "the future is already here." He walked off the stage. The audience lined up to touch the real devices.