Agent phones poised for explosive growth: the upgrade dilemma, on-device AI, and the battle for the next-generation entry point

Agent phones poised for explosive growth? The upgrade dilemma, on-device AI, and the battle for the next-generation entry point

Titanium Media Download the app | Exclusive early access October 21, 2026, 17:36:00 From Beijing Updated October 21, 2026, 18:18:26

By Silicon Valley 101

Recently, the phone wars have kicked off again: Apple has officially entered the foldable space, Samsung has split foldables into multiple form factors, and Xiaomi has shown off the 18 Fold. At the same time, memory chip prices keep rising, pushing the entire industry into a new cycle and forcing manufacturers to further scale back their mid- and low-end product lines.

But what's more cutting-edge, and also drawing more attention, is the AI agent phone. For the past decade or so, phones have kept getting more powerful—more cameras, faster chips—but the way we operate them has barely changed. What if one day the phone no longer waits for you to operate it, but simply gets things done for you? In 2026, this suddenly started to accelerate.

Qualcomm CEO Cristiano Amon previously said that "2026 is the year of the agent." We've also seen that in July of this year, StepFun released the AI phone STEPX Neo, Honor released the Robot Phone, and more recently, Nubia released the NaviX Ultra, equipped with a new version of the Doubao phone assistant, officially certified as "the world's first AI agent phone."

And in Silicon Valley, Apple, Google, and OpenAI have all jumped into the fray, and suddenly the AI phone market is buzzing.

Why are phone manufacturers starting to bet on "agents"? What is the hands-on experience like with a product like the Nubia NaviX Ultra? And are agent phones really the future of AI hardware?

Why have phone manufacturers recently all started betting on "agent phones"? Behind this, there are actually three forces converging.

▍Phones need a new story

First, the traditional phone industry needs a new "reason to upgrade."

For the past decade or so, the upgrade logic of smartphones has been very clear: bigger screens, more cameras, faster processors, bigger batteries. Every year or two, there would always be an obvious hardware upgrade that made consumers feel "this phone is worth switching to."

But in recent years, this stimulus has been noticeably weakening. Counterpoint data shows that the global smartphone replacement cycle has lengthened significantly—in 2023 and 2024, this figure was around 43 months. By 2026, Counterpoint's latest assessment is that, affected by rising prices of memory chips and other components, the global replacement cycle is moving toward 48 months.

When analyzing the reasons for the lengthening replacement cycle, they mentioned a term: incremental device upgrades, meaning consumers are increasingly unable to perceive the gap between a new phone and the previous generation.

So for manufacturers, they urgently need to find a new stimulus to reignite everyone's desire to buy. That stimulus is, of course, AI—more precisely, agents.

Over the past few years, phone AI features have become more and more numerous, but the imaginative space that agents bring is completely different.

For example, previously, to book a restaurant, a user might need to open the phone's map, search for the restaurant, read reviews, check the hours, and then jump to a booking platform.

In the agent era, we only need to tell the phone, "Find me a restaurant that's close to me, suitable for two people, and still has a table at 7 p.m. tonight." The subsequent searching, comparing, and booking can all be done by the agent.

In this way, the phone's role changes completely—from "a tool that waits for people to operate it" to "an intelligent assistant that can get things done for you." And for phone manufacturers, this may be exactly the "new story" needed for the next replacement cycle.

Previously, AI on phones was mainly about relatively fixed tasks like voice recognition, image recognition, and camera algorithms. The tasks weren't complex, the models weren't large, and targeted NPU optimization was enough.

But once agents enter phones, the complexity of tasks increases greatly—it may involve a whole series of things like voice, text, images, AI inference, and operations. The phone not only needs a lot of computing power—CPU, GPU, NPU, and other dedicated computing units—but also needs to dynamically switch and coordinate around tasks, making on-device AI capability a key issue.

Over the past few years, phone chips have been specifically addressing this problem. We looked back at the Snapdragon iteration record and found that AI computing has gradually evolved from an early auxiliary capability into a core capability of the SoC.

As early as 2015, Qualcomm's Snapdragon 820 carried the first-generation Qualcomm AI Engine, and the chip began helping handle AI tasks in imaging, voice, and sensors. The 2018 Snapdragon 855 added a tensor accelerator specifically for computing neural networks—but at this point, AI was still doing fixed jobs like photography and voice.

The real turning point was the Snapdragon 8 Gen 2 at the end of 2022. This generation was the first to support INT4 low-precision computing and was also specifically optimized for Transformers.

Data released by Qualcomm shows that on some AI workloads, performance improved by up to 4.35 times. From this generation on, phone chips began preparing for "running large models locally."

The 2023 Snapdragon 8 Gen 3 went even further, announcing support for generative AI models with up to 10 billion parameters running on phones.

By last year's Snapdragon Summit, where the fifth-generation Snapdragon 8 Elite was released, Qualcomm had for the first time written "the agent on the phone" into its design goals. The NPU is 37% faster than the previous generation, and it also added capabilities such as Personal Scribe and Personal Knowledge Graph. The role of these two things is to let the on-device Agent remember you and keep learning.

From helping with photography, to computing neural networks, then being able to run large models, and now beginning to have "memory," chipmakers have actually been moving step by step toward the Agent direction—it's just that in previous years no one was saying that word out loud.

▍Device-cloud collaboration becomes reality

But a fast chip isn't enough; there also has to be a model that can actually fit into a phone.

Cloud-based large models often have hundreds of billions of parameters, and there is no way to fit them into a phone, so for a long time, "large models" and "phones" were almost two separate worlds.

But in the past two years, these two worlds have begun to converge rapidly. On one hand, the models themselves are rapidly getting smaller. Dr. Xu Hao, Qualcomm's global vice president, said in our previous interview: for the same intelligence, fewer and fewer parameters are needed.

Xu Hao

Qualcomm Global Vice President

For example, we did a study. The best LLaMA model in 2024 could reach a certain order of magnitude in performance. Within a few months, less than a year, these 100-plus-billion or several-hundred-billion models—with very small models, with 4 billion or a few dozen billion—could potentially achieve the same effect, or even better.

On the other hand, the "device-cloud collaboration" architecture also gives the Agent more room to exercise its capabilities on the phone.

As demand, computing power, and models—these three forces—come together, the basic conditions required for an AI agent phone are already in place. At this point, a problem also emerges: once AI starts operating the phone on behalf of the user, who decides how it should operate? This is exactly where the industry is currently forking.

▍Two paths for agent phones

At present, there are mainly two different approaches in the industry.

The first approach can be understood as a "whole-device Agent." When you turn on the phone, it still looks like what we are familiar with, and the commonly used apps are all there, while AI features are integrated into the system. The biggest advantage of this path is that the experience is direct enough. The Nubia NaviX Ultra we tried this time, equipped with the Doubao Phone Assistant (consumer version) and the Snapdragon flagship mobile platform, takes this path.

The second approach is to put a SuperApp in the phone, that is, a super AI entry point. Users first enter an AI app, such as ChatGPT, Doubao, Gemini, and so on, and then it calls various services such as maps, shopping, and food delivery. The advantage of this path is that implementation is relatively easy. Phone makers do not need to restructure the entire operating system all at once; they only need to first build the AI entry point well, then continuously add services it can call. The learning curve for users is also relatively low.

The industry is currently exploring both directions and looking for balance and breakthroughs.

Let's first look at some moves on the Silicon Valley side, starting with Apple, which just held a launch event recently.

In September, Apple just released the iPhone 18 series. Starting from this year's WWDC, one of Apple's biggest changes is that it is pushing Siri from an "assistant that answers questions" toward a system-level Agent.

At this year's launch event, Apple even directly proposed the concept of Intelligent Personal Hub at the opening, describing the new-generation iPhone as an "intelligent personal hub."

In terms of computing, Apple Intelligence also continues to use a combination of "on-device processing + private cloud computing": whatever can be done on the device is kept on the device as much as possible, and when stronger capabilities are needed, it is handed over to Apple's cloud. Its advantage is that the experience is relatively easy to unify, but correspondingly, its openness to the third-party ecosystem is relatively limited, and many AI capabilities need to run under Apple's own rules and frameworks.

The potential problem this brings may be that iteration will be slower.

On the other side, Google's path is somewhat different. It looks like it is building a system-level Agent for Android, but judging from the ecosystem behind it, it also looks very much like a SuperApp.

Because Google itself has a very complete service system, when Gemini enters Android, it gains two advantages at the same time: on one hand, it can go deep into the operating system and understand what the user is doing; on the other hand, it itself has a large number of services that can be called directly.

However, Google's advantage is also its biggest challenge. What users really want to use on their phones is far more than just Google's services. When the Agent begins to execute tasks across apps, different third-party applications need to open permissions, interfaces, and execution capabilities, and there are many Android devices, with differences in hardware, systems, and permission mechanisms among different manufacturers.

So the problem Google may need to solve is how to let Gemini reliably complete tasks for users in an open, huge, and highly fragmented Android ecosystem.

Of course, Silicon Valley also has a class of new players, such as OpenAI.

In May 2025, OpenAI announced a partnership with former Apple chief designer Jony Ive's design company LoveFrom, and acquired its hardware team io Products.

At first, what Ive and Sam Altman envisioned was to break away from the "screen + app" interaction model and create a more AI-native device, and at one point they even envisioned a device without a screen. But later, due to development setbacks, reports said OpenAI's product design might still take the form of a traditional smartphone, with the goal of mass production in the first half of 2027.

▍Domestic manufacturers: ecosystem, models, and terminals begin to converge

Silicon Valley's current exploration is still mostly concentrated at the system layer and terminal layer, while in the Chinese market, manufacturers' approaches are more diverse and livelier.

In July of this year, the Cyberspace Administration of China released for the first time centralized filing information for on-device large models. Apple, Xiaomi, OPPO, vivo, Huawei, Samsung, and Nubia—all seven mainstream manufacturers completed filings for on-device AI generative services.

This also means that agent phones in China have gradually moved from manufacturers' own product exploration into a deployment stage with filings and regulation.

Among them, Nubia's approach is somewhat special: it takes the whole-device Agent route. So how is the experience?

Compared with traditional smartphones, the Nubia NaviX Ultra feels very different in several ways.

The first change is the most intuitive: for some app operations, after authorization, you no longer have to do them yourself. In the past, when we wanted to do something, we had to open the corresponding app and then search, page through, and query step by step. On an agent phone, you only need to long-press the independent AI button and state your request, and after authorization it will automatically operate for me.

The second is improved execution capability: it can operate across apps, specifically mainly in three steps.

In terms of reasoning, for some tasks that require the assistance of other apps, it can automatically invoke the corresponding app, so I no longer need to open it manually.

More importantly, in terms of execution, it can run on a virtual screen in the background, so the display does not occupy the main screen, and I can do other things at the same time.

The third is memory capability: as long as I authorize it, or proactively tell it some information, it can remember it for me. For example, my itinerary and ticket prices—when I need them, I can directly ask Doubao and no longer need to open an app to look for them.

Xu Hao

Qualcomm Global Vice President

It is a brand-new user interface. From choosing an app to not needing to choose an app, these are all changes in user experience and interaction methods that we will gradually adapt to.

Of course, as a brand-new category, the experience of agent phones still has a lot of room for improvement. During our trial, we found that the current smooth experience mainly revolves around system apps, ZTE, and ByteDance-family apps. To invoke more third-party apps, we will have to wait for gradual integration in the future.

But from a technical perspective, this complete chain of "understanding—planning—execution" has already been proven. This is like the early state of the autonomous driving industry: first build the technical foundation, and next the more important thing is to repeatedly verify and iterate in real commercial environments.

For these complex capabilities to run well and run long on phones, what kind of support does the underlying hardware need?

▍From being able to run to running well

First, the Agent must be on standby at all times, but if the chip runs at full power all the time, the battery will not last long. This requires low-power sensing.

Second, the hardware must better support parallel scheduling of multiple tasks. In the past, CPU, GPU, and NPU each undertook relatively clear tasks, but on an Agent, there is task scheduling, inference computing, and the background also has to render the operation interface. These tasks need to switch efficiently among different computing units, and no part can be a weak link.

Moreover, the computing power on a phone is ultimately limited, which brings us to the third requirement: collaboration between the cloud and the local device.

Xu Hao

Qualcomm Global Vice President

We do preprocessing on the phone side, simplifying and compressing personal information as local memory.

If I want to make a big plan, if it can be done on the phone, then we do it on the phone; if it cannot be done on the phone, then we can put the planning in the cloud. I provide certain information, the cloud makes the plan, and then sends it back.

During device-cloud collaboration, the biggest impact on experience is latency, and the Agent's computing is not centralized, which also places greater demands on distributed dynamic scheduling capabilities.

If we compare these points with reality, where have we already reached? It can be said that the basic framework has been built, but the future is even more worth looking forward to.

▍The mobile computing platform in the Agent era

Take the Nubia NaviX Ultra in our hands: it uses the Snapdragon flagship mobile platform.

For the low-power sensing mentioned earlier, it has two Micro NPUs that usually keep watch over voice wake-up and motion status with very low power consumption, and only wake the main NPU when AI needs to do heavy work.

As for multitasking scheduling, its approach is "heterogeneous computing," assigning different jobs to the most suitable units.

Xu Hao

Qualcomm Global Vice President

Our architecture includes CPU, GPU, and NPU, each with somewhat different functions. The CPU mainly handles control, while most of the actual compute power is on the NPU, so the NPU can perform AI operations to the utmost while reducing power consumption. The NPU's biggest advantage is that it is fixed-point, so it is tailor-made for AI and uses the most energy-efficient approach. The GPU is mostly more flexible, but consumes more energy. So our heterogeneous architecture allows each hardware component to play to its strengths.

For those things that really cannot be handled on-device, we use the low-latency 5G link to coordinate with the cloud.

Xu Hao

Qualcomm Global Vice President

We have very strong 5G connectivity to transmit this data to the cloud, let the cloud finish training (inference), and then send it back. So all of this is part of a unified solution.

From this perspective, the hardware requirements for agent phones have already been "basically" met. Why "basically"? After all, the Agent concept only began to land this year, and as AI capabilities grow, the demand for compute will continue to rise. Let's take a look at how the mobile chip industry will respond.

For example, Qualcomm just held the 2026 Snapdragon Summit and released a new generation of dual flagship mobile platforms (the sixth-generation Snapdragon 8 Elite Edition and the Super Elite Edition), officially claimed to be flagship mobile platforms designed specifically for the agent era.

The first block is the CPU. Looking at the specs alone, the main frequency reaches 5GHz. But more noteworthy than the main frequency is a cache architecture it brings called FlexCache. Simply put, it allows different compute cores to share the same cache pool and dynamically adjust the allocation according to task load. Because the way an Agent works is: a request comes in, is split into a series of subtasks, and these subtasks need to run in turn across different cores.

Under traditional architecture, every time a task switches cores, the data may need to be reloaded from memory. FlexCache lets everyone share one cache pool, so when tasks are handed off, the data is still in place and does not need to start over. The benefit is that it can support agents in retaining longer context, with faster response and lower latency.

The second block is the GPU, and this time the idea is to give the GPU itself AI capabilities. The newly released Adreno Neural Fusion puts AI acceleration directly into the graphics pipeline, and also adds a set of GPU cores dedicated to running AI, called Adreno Matrix Cores, used for things like AI super resolution, frame generation, and Transformer computation.

The third block is the NPU, with mainly two updates here. First is the Element Accelerator optimized specifically for Transformers, which contains two groups of units: vector compute units, responsible for high-throughput AI math calculations, making operations faster; and scalar compute units, responsible for decision logic, routing, and orchestration, making inference more accurate while lowering power consumption.

Another is large-capacity shared memory, allowing model state, context, and KV-cache to be stored closer to the accelerator, so that when an agent switches between multiple tools and tasks, it does not need to reload from scratch every time.

It also supports MoE mixture-of-experts models, as well as multiple precision strategies from INT2 to FP16. When executing tasks, whoever is best suited takes the job, making simple tasks power-efficient and complex tasks more effective.

So the industry's next move is to have AI acceleration across all three lines: NPU, CPU, and GPU, which is exactly what Agent-style multi-step tasks value most.

Whether it is elderly people reading novels, Gen Z "scrolling social media," or office workers who use several phones at once and treat them as mobile computers, this form of hardware has become an indispensable part of our lives and work. And this close relationship formed by the mobile internet is now facing a huge upgrade with the support of AI.

This upgrade is: from "I operate it" to "it completes it for me." And when this change truly happens, competition in the next generation of phones may no longer be just about specs, cameras, and chips. Whoever can become the Agent closest to users, that understands users best, and is most worthy of users' trust may be the real battle for the entry point of the next round of mobile computing.

"Special statement: The content above (including videos, pictures, or audio if any) is uploaded and posted by users of "Dafeng Hao," a self-media platform under Phoenix News. This platform only provides information storage space services.

Notice: The content above (including the videos, pictures and audios if any) is uploaded and posted by the user of Dafeng Hao, which is a social media platform and merely provides information storage space services."