For two years “AI phone” was a slogan. Now there is a permit system — and the tickets are in hand.
On July 15, China’s Cyberspace Administration published the filing records for seven on-device generative AI services, covering seven phone makers: Apple, Huawei, OPPO, vivo, Xiaomi, Samsung, and Nubia. It is the first time regulators have collectively published on-device model filings — and after two years of shouting about AI phones, the vendors finally hold official entry tickets.
Until now, generative-AI filings centered on standalone apps and cloud services. Filing at the phone-OS level means on-device models can touch personal information — photo libraries, calendars, device state — and may gain cross-app execution capability. Regulation has formally reached the system layer. One industry insider’s read was blunt: phone makers will push system updates through the second half of the year, and only filed on-device models can ship publicly and commercially inside them. The timing coincides with WAIC 2026, where Nubia and ByteDance’s second-generation Doubao phone NaviX Ultra, StepFun’s STEPX Neo, and Honor’s ROBOT Phone all appeared — AI phones are no longer a slide concept but the most crowded exhibit on the floor.
Three Signals in One Day
Signal one: regulation moves from after-the-fact to before-the-fact
For two years vendors operated in a gray zone of “launch first, file later” — trumpeting AI on stage while users found features throttled, partly because filings were incomplete. A collective filing regime means a reusable approval framework now exists: every on-device AI function — photo editing, voice assistants, agent operations — must clear filing before commercial release. That draws a compliance starting line for the industry: whoever files first, ships first.
Signal two: the Doubao phone accelerated everyone
The catalyst was ByteDance and ZTE’s nubia M153, an engineering device released late last year running the Doubao phone-assistant preview: speak a request and it reads the screen, opens apps, and completes operations — price-comparison shopping, ordering food, booking flights, replying in WeChat. That is a qualitative leap beyond the photo-retouching and summarization of traditional AI phones. The M153 sold out instantly and hit 7,999 yuan on the second-hand market; vendors across the industry accelerated their agent roadmaps in response. The second-generation NaviX Ultra at WAIC is stocked at 200,000 units with 100,000 in the first batch — ByteDance is pushing the experiment into commercial scale. But the experiment also exposed the core conflict: M153’s GUI-agent route (simulating human screen operation) crossed platform risk boundaries — WeChat, Taobao, and Alipay each issued risk warnings, and Doubao had to suspend WeChat operation. It was the first head-on collision between phone agents and internet platforms’ risk-control systems.
Signal three: model companies are building phones themselves
On the eve of WAIC, StepFun simultaneously launched a hardware brand (STEPX), an agent operating system (Step AOS), a personal agent (Amoo), and the STEPX Neo phone — the first time a foundation-model company has stood onstage as a hardware brand. OpenAI’s play is more aggressive still: after acquiring Jony Ive’s io in 2025, it is reportedly developing smartphone processors with MediaTek and Qualcomm targeting 2028 mass production. Model companies are done renting the entrances Apple and Android provide — they want ownership of the entrance itself. StepFun chairman Yin Qi framed the logic: the phone is only the agent’s first carrier into the physical world, and personal memory becomes the AI era’s new lock-in — once your preferences, habits, and history live in some agent’s cloud, switching phones stops being “buy a new handset” and becomes “abandon your digital personality.”
The Real Battlefield Is Distribution, Not Hardware
The essential change in the AI phone is a migration in how services get distributed — from “users open apps” to “agents decide on their behalf.” That reorders the commercial stakes completely. Old logic: user wants to shop → opens Taobao → browses recommendations → orders; Taobao captures traffic and ad revenue. Agent logic: user says “buy a robot vacuum under 3,000 yuan that arrives tomorrow” → the agent decides which platforms to search, compares prices, allocates the order. The app becomes a back-end service stripped of its advertising and recommendation surface. Hence the three-way power contest:
-
Internet platforms are loosening, unevenly: “This year there’s 10-20% openness; last year it was zero. But vertical aggregators like Meituan will stay conservative.”
-
Phone makers are diverging: Huawei bets on Ascend silicon for domestic compute; Xiaomi leans on its IoT ecosystem; OPPO/vivo focus on small on-device models; Honor runs a platform-AI route where the OS recognizes intent and dispatches to different internet companies, adjusting partners on price and experience.
-
Model companies want global scheduling rights. ByteDance’s negotiations with phone makers push for system-level entry and dispatch capability, while the makers want to keep the on-device model and core scheduling for themselves.
OPPO ColorOS smart-product R&D director Jiang Yuchen’s judgment offers the likely technical settlement: A2A (Agent-to-Agent) is the only route that holds long-term — each party keeps its own agent and data boundary, exchanging only the information a task requires, with no need to open underlying permissions to each other. That may be the rational convergence point of a multi-party game.
Four Hurdles Ahead for the AI Phone
-
Privacy trust. More agent capability means more system permissions. NaviX Ultra’s answer is “every operation is user-authorized and fully visible” — whether consumers buy it awaits market proof.
-
Ecosystem openness. How far platforms open up decides which implementation path agents can take. The industry currently runs four routes — GUI, API, MCP, A2A — and none pleases every stakeholder.
-
Inference cost. The more complex the agent’s task, the higher the cloud inference bill. Vendors cannot subsidize forever; users may end up buying token packages directly. The three carriers have launched consumer and family token plans, but willingness to pay remains unproven.
-
Return on investment. One leading vendor’s AI spend has reached the tens-of-billions-of-yuan scale against output well below expectations; internal dissatisfaction with AI division ROI is showing, and self-developed foundation-model spending is being cut.
Three Things Worth Watching
-
The second-half upgrade wave. Seven vendors will push on-device AI system updates over the coming months. If you plan to change phones, H2 2026 is when AI-phone capability actually ships.
-
The token economy sprouts. Vendors charging B-side commissions and C-side token subscriptions is taking shape; tomorrow’s phone plan may read “X yuan a month, 2 million tokens included.” Watch this space.
-
The regulatory dividend. Filing is both barrier and moat: the first cohort gains first-mover advantage not only in shipping speed but in the user-habit and data window before the next filing round, perhaps six months away.
One interviewee put it vividly: if the iPhone was the ape first standing upright — freeing eyes and hands — then the AI agent phone is the first ape looking up at the stars. And after looking up, the long evolution has only begun. The AI phone stands exactly at that point.
