Author｜Xie Jiabaoshu

After Doubao unveiled its 'Second-Generation Doubao Smartphone,' Alibaba is stepping up its game.

On September 22, 2026, Alibaba hosted the 2026 Yunqi Conference, where it showcased the 'Qwen Tablet' QwenBook, a device still in the development phase.

According to LatePost, the QwenBook is being developed by Alibaba Cloud's Wuying team and is envisioned as a native agent computer. The team is taking full charge of product definition and hardware development. Leveraging Wuying's extensive experience in cloud computing, including system and end-to-end cloud capabilities, will be integral to the product's development.

Indeed, as the AI competition intensifies, not only Doubao and Alibaba but also a growing number of tech firms are shifting their focus toward hardware. In September 2023, Meta, in collaboration with Ray-Ban, introduced the Ray-Ban Meta AI glasses, which quickly gained popularity due to their lightweight design. In November 2025, OpenAI CEO Sam Altman announced that the company had developed its first AI hardware prototype, with plans to start production within the next two years.

With agents emerging as an industry trend, the focus of large model competitions is transitioning from Q&A services to task execution. In this context, tech companies are investing in hardware not only to explore innovative product forms but also to target control layers such as default access points, system permissions, and app orchestration.

The success of AI hardware developed by tech companies in overcoming challenges like third-party app barriers, security concerns, and commercialization will determine whether this round of hardware competition leads to the emergence of new platforms or simply results in a batch of new devices equipped with chatbots.

01 The Agent Era: Large Models Seek a 'Physical Form'

Reflecting on the evolution of the AI industry over the past few years, while large models have remained the foundational technology, the implementation products and usage scenarios have undergone significant transformations.

Initially, spurred by ChatGPT, large models were packaged as Q&A robots, providing responses based on user queries. Since this involved only conversational interactions without the need to orchestrate external tools, chatbot services could be hosted on ordinary web pages or apps.

Image Source: Doubao

While Q&A robots reduce the cost of information screening for users compared to traditional search engines, their limited revenue potential and high operational costs have made it difficult to sustain a profitable business model. According to LatePost estimates, as of the first half of 2026, Doubao's average daily revenue was less than RMB 1 million, while costs reached tens of millions.

Fortunately, AI has not been limited to the chatbot format. As technologies such as reasoning, memory, and orchestration mature, large models are evolving into agents capable of executing complex tasks with a single command, helping users solve real-world problems and offering greater commercial potential.

Compared to traditional Q&A robots, agents can directly invoke tools to execute tasks but require higher permissions. Only by accessing the system layer and obtaining permissions such as local file modification, screen access, and cross-app operations can the full potential of agents be unlocked.

Image Source: Apple

This is precisely the driving force behind tech companies' efforts to develop innovative hardware around AI technologies. As computer scientist and graphical user interface pioneer Alan Kay famously said, 'People who are serious about software should make their own hardware.' AI technologies demand a higher degree of software-hardware synergy. To deliver differentiated user experiences, tech companies must develop both software and hardware, creating innovative products with underlying permissions.

While the tech industry has recently witnessed a surge in AI hardware innovation, this does not imply that model companies are transforming into consumer electronics firms. Rather, these companies recognize that without device-layer permissions, agents will struggle to transition from 'talking' to 'doing.'

02 Smartphones, Tablets, and Glasses: Targeting Different Contexts

Although AI has sparked a new wave of hardware innovation, unlike the desktop internet and mobile internet eras, which were dominated by PCs and smartphones, respectively, AI hardware has yet to adopt a unified form.

Broadly speaking, AI hardware developed by tech companies can be categorized into smartphones, tablets, and wearables.

This is partly because the AI hardware race is still in its infancy, with manufacturers in the exploration phase. On the other hand, it relates to the technical requirement that large models need sufficient context to better serve users.

Take smartphones, for example. As a mass computing platform, smartphones handle tasks such as communication, socializing, and payments, containing rich lifestyle-oriented contextual data. Given the importance of mobile terminals, many tech companies are developing AI-powered smartphone products.

Image Source: Doubao

On September 14, Doubao launched the consumer version of its Doubao Smartphone Assistant, which can automatically execute tasks via MCP, A2A interfaces, or GUI after receiving user commands. On September 22, Alibaba introduced Qwen Intelligence, a full-stack AI smartphone solution, providing an agent technology platform based on the Qwen large model to enable smartphones to execute complex tasks across apps.

As mentioned earlier, in addition to smartphones, Alibaba has also developed the 'Qwen Tablet' QwenBook. Featuring a tablet-plus-magnetic-keyboard design, QwenBook is marketed as 'your first native agent computer' and is positioned as a productivity tool.

Clearly, Alibaba's primary objective with QwenBook is to capture contextual data from users' work and learning scenarios. By delivering results based on users' productivity contexts, QwenBook could redefine traditional tablets' entertainment-focused positioning and unlock new consumer demand.

Image Source: Li Auto

The same rationale applies to tech companies' focus on wearables. While smartphones and computers have broad utility, users do not use them continuously. In contrast, wearables like glasses and earphones can be worn for extended periods, capturing environmental, locational, and auditory information to provide large models with real-world user contexts.

Because smartphones, tablets, and wearables offer different dimensions of context, many tech companies are not betting solely on one type of device but are diversifying their portfolios across multiple categories. Some are even expanding into products like voice recorders, smart cars, and smart home appliances, attempting to collect contextual data from more specialized scenarios.

For AI companies, the value of hardware lies not just in adding a business line but in providing a stable gateway to obtain user contextual data. This suggests that the future focus of AI hardware competition will shift from individual devices to cross-device personal agents.

03 Proprietary Hardware Is Just the First Hurdle; the App Ecosystem Is the Bigger Challenge

For well-funded tech companies, developing AI hardware using mature solutions and relying on contract manufacturers is relatively straightforward. The real challenge lies in gaining developer support.

In December 2025, Doubao launched a technical preview version of its Doubao Smartphone Assistant, which enabled cross-app automation through GUI simulation click technology.

For example, when a user instructs the Doubao Smartphone to select a product on Taobao, the assistant can automatically open the Taobao app, choose a product that meets the user's needs, and proceed to payment. The user only needs to verify the password to complete the purchase.

While Doubao Smartphone significantly enhanced the agent experience on mobile devices, the GUI simulation click technology bypassed the mobile internet ecosystem's security and risk control systems, leading to restrictions from popular apps like WeChat, Meituan, and Alipay.

In response, Tencent CEO Pony Ma stated in early 2026 that Tencent has always been firmly opposed to methods that involve 'black market plugins' to record and upload user screens from phones and computers to the cloud for processing, calling such practices 'extremely unsafe' and 'irresponsible.'

Doubao later clarified that the smartphone assistant only invokes necessary capabilities with explicit user authorization and adheres to a 'no storage, no training' principle for cloud processing.

Image Source: Doubao

Today, while the second-generation Doubao Smartphone still retains GUI technology, it has introduced a 30-day notice period. During this period, third-party apps will not be operated by GUI agents. After the notice period expires, apps that do not explicitly refuse being operated will be gradually opened up based on risk levels, while those that explicitly refuse will not be operated.

While it remains uncertain which apps can be operated by GUI agents, given the experience with the technical preview version, the consumer version may still struggle to control super apps like WeChat, Meituan, and Alipay.

Given the difficulty of breaking through the mobile internet ecosystem's barriers, QwenBook emphasizes system-level AI capabilities. Alibaba's product team has independently developed a suite of AI tools, including system-level note-taking, Wiki, smart folders, a proprietary email client, and knowledge base MyNote and memory management app MyZone. Additionally, QwenBook will collaborate deeply with Kingsoft to integrate WPS.

The issue is that after years of mobile internet development, most user contextual data and consumption scenarios are concentrated in third-party apps. Relying solely on proprietary AI tools makes it difficult to persuade users to incur additional costs for a product whose basic experience is indistinguishable from traditional devices.

In essence, developing proprietary hardware is just the first step toward agent maturity. The real challenge is whether AI companies can get app developers, service platforms, and users to collectively accept a new set of authorization and benefit-sharing rules.

Rather than simply selling more AI devices, companies that strike the right balance between system permissions, ecosystem barriers, and user acceptance are more likely to emerge as winners in the AI hardware race.

Interactive Topic

What AI hardware have you used? What advantages does it offer over traditional hardware?

This article is original content from Farsight and is not authorized for reproduction.