China's Agentic AI Phone Sold Out in a Day. Then the Platforms Switched It Off.

China's Agentic AI Phone Sold Out in a Day. Then the Platforms Switched It Off.

The ByteDance and ZTE collaboration on the Nubia M153 generated the kind of launch-day buzz that most consumer technology products spend millions of marketing budget trying to manufacture. Released on 1 December 2025 and priced at 3,499 yuan (about US$494), the phone sold out on its first day of online ordering. The product that drove that demand was not the hardware, which is competitive but unremarkable by Chinese flagship standards. It was the Doubao AI assistant, baked into the operating system at a level deep enough to give it genuine autonomous capability over everything the phone does.

The excitement was real, and so was the subsequent complication. Within weeks of launch, a familiar pattern emerged: the demo videos looked extraordinary, the actual product experience turned out to be more constrained, and the explanation involved platform politics that the original marketing had been careful not to mention.

What Doubao Actually Did

The Doubao Phone Assistant was not a conventional voice assistant with a smarter backend. It combined the Doubao large language model with a second on-device model called Nebula-GUI, giving it the ability to read the phone's screen, navigate app interfaces autonomously, and complete multi-step tasks without requiring users to open individual apps or input data manually. The system watched what was on screen the same way a human would, then acted on it.

In its promotional form, this meant ordering food delivery from a voice command while the user was in a different app. It meant comparing prices across multiple shopping platforms simultaneously, replying to WeChat messages without the user switching to WeChat, changing travel reservations end-to-end^✦, and navigating Mini Programs autonomously. These are genuinely useful functions that existing smartphone voice assistants handle poorly, if at all.

The Platform Pushback

The complication arrived quickly. Major Chinese digital platforms concluded, apparently independently and in rapid succession, that an AI agent with operating-system-level access to their apps was not something they were prepared to support.

Alipay, Taobao, Pinduoduo, and Ele.me all restricted Doubao's ability to autonomously operate within their apps. The stated reasons varied: fraud risk, fairness to other users, security concerns about screen-reading capabilities. Tencent suspended Doubao's ability to control WeChat specifically, entering direct negotiations with ByteDance about what acceptable use cases might look like. Banking and financial applications were disabled entirely. AI features in competitive online gaming were removed to prevent automated advantages.

The practical effect was to substantially reduce what the phone could actually do compared with what the promotional materials had shown. ByteDance clarified that no screen data is stored or used for model training, a statement that addressed one concern while leaving others intact. If an AI can read your screen without storing the data, users may feel better about privacy, but the fundamental question of whether an autonomous agent should be able to initiate financial transactions and send messages on your behalf remains unresolved.

Why the Restrictions Matter for Asia

The platform response to the Doubao Phone reveals a structural tension that will define AI's role in everyday life in Asia for the next several years. Digital platforms in China, and by extension across Southeast Asia where many of the same apps operate, have built their business models around controlling the user experience inside their apps. They own the interface, the data, and the transaction. An OS-level AI agent that can cross app boundaries and act autonomously is not just a privacy question. It is a question of who controls the commercial relationship with the user.

This is not unique to China. Any agentic AI operating across multiple apps and services will eventually collide with the same dynamic: the platform that built the app has strong incentives to prevent a third party from mediating the user's interaction with it. The difference in China is that this collision happened at exceptional speed, in public, and with immediate regulatory visibility.

China's capacity to rapidly integrate AI capabilities with consumer hardware represents a significant technological achievement that emerged outside Silicon Valley.

The recent expansion of humanoid robot training infrastructure across China follows a similar pattern: ambitious demonstration of autonomous AI capability in physical settings, followed by the practical and regulatory work required to make that capability actually deployable. The Doubao Phone is, in that sense, a useful preview of the challenges that the robotics industry will face at the point when humanoid machines are expected to interact with the same commercial and social platforms that smartphones touch today.

What Comes After the Rollbacks

ByteDance is reportedly in discussions with multiple additional device manufacturers, including Lenovo and Vivo, about embedding^✦ the same Doubao-plus-Nebula-GUI system in their hardware. That suggests the company sees the Nubia M153 as a proof of concept rather than a product: a demonstration that the technology works and that there is consumer demand, rather than a fully resolved commercial proposition.

The next generation of agentic smartphones in China will need to solve a negotiation problem as much as a technical one. ByteDance will need agreements with the major platforms before it can deliver the experience the promotional materials promised. Tencent's ClawBot integration into WeChat offers one model for how that might work: the agent running inside the platform, authorised by the platform, rather than an external agent trying to control the platform from outside.

By September 2024, Doubao was processing over 30 trillion tokens daily, more than doubling its earlier capacity.

The deeper question is what happens to user expectations now that the possibility of fully autonomous AI phone assistance has been demonstrated and then partially withdrawn. In consumer technology, the genie rarely goes back in the bottle. Users who saw those demo videos know what is theoretically possible, and they will expect the gaps between the demo and the product to close over time.