What Exactly Are AI-Powered Smartphones Competing For?

'We genuinely sought advice, but ultimately, we decided to take the leap!'

When questioned about 'why a large model company would venture into smartphone manufacturing,' Yin Qi, the Chairman of Stepfun, responded in this manner. Over time, he consulted with numerous experts in the terminal industry, and the consensus was clear: steer clear of hardware. However, Stepfun ultimately chose the most challenging and weighty path.

On July 13, Stepfun unveiled STEPX, its native AI terminal brand powered by large models, and simultaneously launched Step AOS, a native operating system tailored for AI agents, Jieyue Amoo, a personal AI agent, and the first smartphone natively integrated with a large model AI agent, STEPX Neo.

Image Source: Stepfun

Shortly thereafter, Glory announced its collaboration with Alibaba to advance the next-generation terminal operating system, Agentic OS, with its global debut of the robot smartphone, Robot Phone, set for WAIC. Nubia also officially declared its intention to unveil the nation's first mass-produced flagship AI-powered smartphone at WAIC. Additionally, cybersecurity authorities released a list of AI service filings for mobile devices, with products from Apple Intelligence, Huawei, Xiaoyi AI, Xiaomi, OPPO, vivo, and Nubia's Doubao large model for smartphones completing their filings in succession.

In just a few days, a clear trend has emerged: the competition among AI-powered smartphones has shifted from 'who accesses large models first' to 'who can define AI agents and the next-generation operating system.'

This signifies that AI-powered smartphones are entering a crucial phase in their development. However, this also poses a new challenge for all participants.

For smartphone manufacturers, AI agents represent an upgrade to the next-generation operating system; for internet companies, they signify a new service entry point; for foundational large model companies like Stepfun, they even imply a reconstruction of future business models.

The question remains: does a large model company without a smartphone heritage truly need to venture into terminal manufacturing? Can recreating an AI agent operating system truly break through the barriers between App ecosystems, operating systems, and business models?

Looking back at the development of AI-powered smartphones over the past year, we can observe a noticeable shift in the focus of industry competition.

Initially, smartphone manufacturers competed on who could integrate large models first. Whether it was AI image removal, AI summarization, or AI search, these features essentially represented intelligent upgrades to traditional smartphone functions. Subsequently, with Doubao smartphone pioneering the GUI Agent (Graphical User Interface Agent), AI gained the ability to recognize screens, simulate clicks, and complete complex tasks such as booking tickets, shopping, and price comparisons across multiple Apps. For the first time, AI agents truly possessed the capability to 'operate smartphones on behalf of users.'

However, as we entered the latter half of this year, industry competition underwent another transformation.

On one hand, Glory announced its collaboration with Alibaba to advance the next-generation terminal operating system, Agentic OS, aiming to transform smartphones from 'application containers' to 'AI agent carriers.' Stepfun introduced Step AOS and directly proposed the concept of a 'native operating system for AI agents.'

On the other hand, cybersecurity authorities recently released a new batch of filings for generative AI services, with products from Apple Intelligence, Huawei, Xiaomi, OPPO, vivo, and Nubia's Doubao large model for smartphones completing their filings. This indicates that AI-powered smartphones are gradually moving from conceptual verification to large-scale implementation.

Image Source: Weibo

A clear trend is emerging: the competition among AI-powered smartphones is shifting from 'who possesses a larger model' to 'who can define the next-generation AI agent operating system.' However, as competition intensifies, a more practical issue is beginning to surface.

Over the past year, Doubao smartphone has proven one thing: AI can indeed operate smartphones on behalf of users. But it has also proven another thing: being able to operate Apps does not necessarily mean that AI-powered smartphones are truly established.

For internet platforms, users opening Apps, browsing content, clicking on ads, and completing transactions are essential components of their business models. When AI agents bypass the homepage, recommendation pages, and search entries to complete orders and payments on behalf of users, platforms lose their original traffic entry points and commercialization opportunities.

Therefore, the subsequent restrictions imposed on Doubao smartphone by some internet applications were not merely technical compatibility issues but also reflected the fact that AI agents were beginning to touch upon the existing profit distribution mechanisms of the mobile internet. The more efficient AI agents become, the more likely they are to alter the original traffic distribution logic of Apps, which happens to be the most sensitive aspect for various platforms.

Because of this, the focus of today's industry discussions is no longer just whether AI can complete operations but rather how AI agents should complete these operations.

Should we continue along the GUI Agent route, simulating screen clicks like humans? Should we enable AI agents to directly invoke services through open interfaces? Or should we, like Stepfun, redesign a complete operating system for AI agents from the ground up?

Currently, the entire industry lacks a unified answer. For Stepfun, Step AOS represents the third approach. However, whether this path can truly solve the long-standing ecological challenges faced by AI agents or is merely another technical experiment remains to be seen, with the market yet to provide a definitive answer.

If the first part of our discussion focuses on the development direction of AI-powered smartphones, then the question raised by Stepfun is even more pointed: must a large model company possess its own terminal?

In Yin Qi's view, the answer is yes. During a media briefing after the launch event, he admitted that the team had sought advice from many individuals in the terminal industry, and the recommendations were almost entirely consistent—'stay away from hardware.'

Because compared to developing models, making smartphones involves entirely different capabilities, including supply chain management, manufacturing, quality control, and after-sales service. It requires heavier investment, longer cycles, and higher risks. However, ultimately, Stepfun decided to 'ignore the advice.' The reason is simple: without its own terminal, Step AOS would struggle to form a closed value loop, and consumers would find it difficult to truly experience the full capabilities of AI agents.

This judgment was not made on a whim. Yin Qi mentioned that while many people today see Anthropic's leading position in AI programming, few know that it determined its Coding strategy as early as 2021.

In his view, large model startups 'cannot catch up later; they must make choices, sacrifices, and beliefs from the very beginning.' From Day 1, Stepfun believed in the integration of software and hardware, believing that large models would ultimately reach consumers through terminal products rather than remaining in a B2B mode indefinitely. The launch of STEPX now represents the first complete external presentation of this strategy rather than a sudden cross-industry attempt.

From this perspective, STEPX Neo serves more as a 'testbed' for validating Step AOS. Public information indicates that the entire device manufacturing is entrusted to ODM manufacturer Huaqin Technology, with Stepfun focusing primarily on models, operating systems, AI agents, and product definition. By entrusting the heaviest manufacturing stage (manufacturing processes) to a mature supply chain, Stepfun can concentrate its core resources on software and AI capabilities.

This approach indeed lowers the barrier to entering the hardware market for the first time and aligns with the common practices of internet companies entering the terminal market in recent years. However, Stepfun's logic is not without controversy.

Image Source: Canva Image Library

Looking across the entire industry, few large model companies truly choose to make their own terminals. More companies still prefer to continue collaborating with smartphone manufacturers, providing models, agents, or operating system capabilities to OEMs. Glory's choice to collaborate with Alibaba to advance Agentic OS represents more of an upgrade within the existing terminal ecosystem. Companies like Apple, Huawei, and Xiaomi possess mature hardware foundations and gradually superimpose AI agent capabilities.

In contrast, Stepfun needs to simultaneously advance on four fronts: models, operating systems, terminals, and ecosystems. This means it must not only prove that Step AOS is sufficiently advanced but also demonstrate that a large model company without a smartphone brand heritage can truly excel in terminal manufacturing.

Therefore, the truly worthy discussion may not be why Stepfun is making smartphones but whether the path it believes in is valid.

If AI agents truly become the next-generation terminal entry point in the future, then possessing its own terminal may indeed mean possessing more initiative. However, if operating systems ultimately remain in the hands of Apple, Android, and mainstream smartphone manufacturers, whether large model companies must personally enter the market or continue as capability providers remains without a standard answer.

At least for now, this is still an ongoing exploration rather than a proven correct business path.

If the first half of the large model competition focused on model capabilities, then with the advent of the AI agent era, a new question arises for all players: what exactly is the next-generation entry point?

Over the past two decades, the internet has experienced two significant shifts in entry points. During the PC era, browsers served as the entry point; in the mobile internet era, entry points became individual Apps. Users needed to actively open different applications to complete searches, shopping, payments, travel, and various other operations. It was during this process that today's mobile internet traffic distribution system and business models were formed.

Now, whether it's Stepfun introducing Step AOS or Glory proposing Agentic OS, both are essentially attempting to answer the same question: if AI agents can understand user intentions and proactively complete tasks in the future, will users still need to open individual Apps themselves?

This is also why Yin Qi repeatedly emphasizes the importance of operating systems. In his view, while many agents today can write code and articles, their ability to execute complex tasks on real devices like smartphones and PCs remains limited. The root cause is not insufficient model capabilities but the lack of a platform truly suitable for AI agent operation. Without an operating system, an ecosystem, and real user data feedback, AI agents will struggle to continuously improve.

Image Source: Stepfun

Following this logic, users may no longer need to first open Ctrip, Meituan, or Didi in the future. Instead, they could simply tell their AI agent, 'Book me a flight to Shanghai tomorrow morning,' 'Find me a restaurant suitable for a group of four,' or 'Order me an iced latte.'

What truly changes may not be the disappearance of Apps but the addition of a layer of AI agents between users and Apps. However, this is also where the industry's most significant controversy lies.

On one hand, Yin Qi explicitly states that Apps will not disappear but will continue to exist in new forms, transitioning from direct user entry points to services invoked by AI agents in the background. On the other hand, both internet platforms and smartphone manufacturers find it difficult to easily abandon the traffic entry points and user relationships they have built over many years.

More importantly, when AI agents begin making decisions on behalf of users, new questions arise.

If a user wants to book a hotel, should the AI agent prioritize Ctrip or Meituan? If they want to order coffee, does it recommend Luckin, Cotti, or Starbucks? Are these results based on the user's historical preferences, real-time price comparisons, or potential future commercial collaborations between platforms?

In the past, these choices were mostly made by search rankings, recommendation algorithms, and users themselves. In the future, if more decisions are entrusted to AI agents, how will new traffic distribution rules be established?

Therefore, rather than saying that the industry has already entered the AI agent era, it would be more accurate to say that the entire industry is vying for the right to define the AI agent era.

Whether it's Stepfun, Glory, Apple, Huawei, or Xiaomi, what everyone is competing for is not just a new operating system but who can become the most important entry point between future users and the digital world.

However, whether this entry point will ultimately belong to AI agents and how it will take shape remain without standard answers today.

From a capital market perspective, the significance of STEPX Neo may extend beyond just being an AI-powered smartphone.

Just two months prior to the launch event, Stepfun successfully secured a new round of financing, amassing nearly $2.5 billion. Concurrently, it dismantled its red chip structure and expedited its preparations for a Hong Kong IPO. At the same time, consumer electronics supply chain companies, including Huaqin, Longcheer, OmniVision, and ZTE, along with institutions such as the Hong Kong Investment Corporation (HKIC), successively became shareholders. This move encompassed multiple facets of the industry, ranging from device manufacturing and core components to terminal ecosystems.

It is noteworthy that Tencent, an existing shareholder, also took part in Stepfun's latest financing round. Tencent first joined Stepfun's shareholder roster during its Series B financing and has continued to invest in subsequent rounds. Not long ago, the two companies further strengthened their partnership. Tencent Cloud inked a strategic cooperation agreement with Stepfun, integrating Stepfun's foundational large models with Tencent's content and application ecosystem to jointly develop AI cockpit agents. This collaboration links Tencent's music, video, mapping, payment, travel services, and other sectors, achieving a closed-loop of in-car services.

Image Source: Weibo Screenshot

Compared to previous financing rounds that mainly involved financial investors, the most notable change in this round is the growing participation of industrial capital in betting on Stepfun. This context also imbues the launch of STEPX Neo with additional significance.

Over the past few years, when assessing large model companies, the capital market has primarily focused on model capabilities, parameter scales, and leaderboard performances. However, as foundational models gradually reach the top echelon, more investors are posing the same question: beyond models, where lies the path to commercialization?

For OpenAI, there is ChatGPT; for Anthropic, there is Claude and the enterprise market. For domestic large model companies, establishing their own user entry points, data flywheels, and commercial closed loops remains a challenge that the entire industry is grappling with.

To a certain extent, STEPX Neo is precisely addressing this question. At the launch event, Yin Qi repeatedly stressed that STEPX would not be a traditional mobile phone company, nor would it rely on selling hardware or pre-installed advertisements for profit. Instead, he envisions connecting models, operating systems, intelligent agents, and real users through AI-native terminals, enabling the terminals to serve as a continuous data source for model iteration, rather than merely focusing on one-time hardware sales.

This undoubtedly presents a compelling and comprehensive narrative for the capital markets. However, telling the story is just the first step; the real challenges are only beginning.

On one hand, STEPX Neo is still in its nascent stages, with the product not yet officially launched. The extent to which ecosystem partnerships can deepen, whether users are willing to engage long-term, and whether intelligent agents can truly establish a continuous data flywheel remain to be validated by the market. On the other hand, terminal manufacturers such as Apple, Huawei, Xiaomi, and Honor are also continuously advancing their AI and intelligent agent strategies. Whether large model companies can definitely establish new competitive advantages through self-built terminals remains an open question.

In other words, StepOn is betting on a future where intelligent agents will become new user entry points, and owning terminals means having more initiative.

This may also be what STEPX Neo truly aims to prove.

If this logic ultimately holds, then what StepOn presents to the capital markets will not just be a foundational large model company, but an AI platform company with models, an operating system, terminals, and a user ecosystem. However, if intelligent agents ultimately remain at the application layer or if mainstream mobile phone manufacturers take the lead in evolving intelligent agent operating systems, then today's business logic may also need to be re-evaluated.

Therefore, whether STEPX Neo represents a paradigm shift in the era of intelligent agents or merely one of many technological approaches remains uncertain, and no one can provide a definitive answer yet.