Two paths for consumer AI agents in China and the US: Doubao Phone and Meta Muse

By Ting Wasner-Lian, 22 September 2026

Chart: SE Incubator Consulting

In September 2026, China and the US each got a landmark consumer AI agent: ByteDance's Doubao Phone Assistant, built into ZTE's Nubia NaviX Ultra, and Meta's Muse. Both promise to get things done from a single request, but they take different routes.

Doubao operates native apps on the phone itself, while Muse works from a cloud virtual machine using a browser and service connectors. The split reflects how the two internets are built. Chinese services sit inside closed super-apps; American services live more on the open web and APIs.

What they truly share is that incumbent platforms blocked both: WeChat and Taobao restricted Doubao, Amazon shut out Muse. The conflict is the same in both markets. When an agent completes a task, the user never looks at the screen, and the ad impressions platforms monetize disappear.

My view: advertising will not vanish, but pricing will shift from buying attention to paying for completed transactions. Value will concentrate in three places: the agent entry point, checkout and payment rails, and product data.

The two cases at a glance

The two products launched in the same month with the same pitch, yet differ on almost every structural dimension.

Table: SE Incubator Consulting

Diverging paths: why one lives on the phone and the other in the cloud

The fundamental difference is where the agent works: Doubao operates the user's own phone, while Muse operates a computer of its own in the cloud.

Where the agent runs. Doubao does not rely on official app APIs; it recognises on-screen interface elements and simulates taps (Sohu). Muse runs on a secure virtual machine Meta assigns to each user, with its own browser, acting across the services people use daily (Meta).

The ecosystem dictates the route. Chinese commerce and services are concentrated in closed super-apps such as WeChat, Taobao and Meituan. Many tasks cannot bypass the native app, so the agent has to live on the phone. US services still sit largely on the open web and APIs, so an agent can run in the cloud without taking over the handset.

Distribution. Doubao is tied to hardware, so users must buy a new phone. Muse is an app with a far lower barrier. Doubao's adoption is therefore capped by replacement cycles; Muse's is capped by trust.

Privacy design. Criticism of Doubao centres on screen content being sent to the cloud; Tencent CEO Pony Ma publicly called that approach extremely unsafe and irresponsible (Sohu). Meta's answer is isolation: it says conversations and VM data are not shared with its ad systems, and plans a VM encrypted with a key only the user holds (Meta). Given Meta's track record, whether those commitments hold is the main open question (TechCrunch).

The shared problem: blocked by platforms, fighting over the entry point

Both products were stopped by the largest platforms soon after launch, and both then turned to negotiation and alliances. The obstacle is not technology but control of the transaction entry point.

Table: SE Incubator Consulting

The underlying conflict is the same: when an agent runs the whole task flow, the user never sees the screen, and the platform's ad exposure and traffic allocation collapse (Leikeji). Platforms cite security and fraud controls, and those concerns are real. But the practical effect of blocking is to protect their own ad and recommendation systems and buy time for in-house agents such as Amazon's Alexa for Shopping (TradingKey).

The responses are also similar: Doubao uses a protocol plus bilateral permissions; Muse routes around Amazon through Shopify and Stripe. My judgment: no single agent will cover every platform in the near term. The likelier outcome is that each major platform runs its own agent behind its own walls, and outside agents get in only by negotiation.

Early market signals and availability

Neither product has been on sale for a month, and no third-party sales or retention data exists yet. Only official claims and indirect signals are available.

Doubao phone. Nubia claimed first-day sales passed CNY 100 million within one second (IT Home). The counting method was not disclosed; at the entry price, that equals roughly 17,000 units, modest for a flagship. JD.com pre-registrations exceeded 370,000 before launch (Tencent News), but registration is free and conversion is typically low.

Counter-signals: most versions were still in stock on JD.com half an hour into the sale, and scalpers switched from charging a premium to selling at list price. The first-generation M153 sold its 30,000 units on day one and once resold for CNY 13,000; new units now resell below the CNY 3,499 launch price (SMZDM / Yicai). I think launch interest is real, but buyers have moved from early-adopter excitement to judging real-world use, and core features stay restricted until the SAEP notice period ends on 15 October.

Muse. It overtook ChatGPT as the top free app on the US iOS store (Time News), with roughly 730,000 downloads in its first five days (Defense World). Downloads measure curiosity, not willingness to hand over email and payment access. The metrics that matter, paid conversion and completed transactions, have not been disclosed.

Availability. Both are confined to their home markets. Nubia has announced no sales outside China, and the first generation never left the country. Doubao is a mainland service built to drive WeChat, Meituan and Taobao, so its core features have little use elsewhere (tech-ish). A phone built around ByteDance software also faces clear regulatory headwinds in Western markets (gagadget). Muse is limited to US users aged 18 and over (Tech Insider); no European timeline has been announced.

Impact on platforms: intent-capture advertising takes the hardest hit

A common claim is that once agents spread, recommendation algorithms stop working and the ad model collapses. That is imprecise. Algorithms keep working; their target shifts from persuading a human to click to getting chosen by an agent. The real impact depends on the type of advertising:

• Intent capture: search ads and sponsored listings inside marketplaces. The user already knows what they want, and the ad intercepts that intent. An agent that orders directly skips this step first.
• Demand creation: social feed ads. The user had no need until content sparked one. Agents execute existing intent; they do not create desire, so this layer is hard to replace in the near term.

Not all shopping will be delegated, either. Task-like purchases such as replenishment, price comparison and ticketing go to agents first. Purchases where browsing is the pleasure, such as fashion, home goods and gifts, will stay largely human.

Table: SE Incubator Consulting

An easily missed implication: if everyone used Muse, Meta would not lose; it would win, moving from selling ad slots to controlling where purchases happen. The real question is not whether platforms can still make money, but who owns the agent and therefore the new toll booth. Markets are already pricing it this way: Shopify's stock rose on the Muse partnership while Amazon sat out the rally (24/7 Wall St.).

How business models will shift: from selling attention to charging for outcomes

Advertising, at its core, buys attention in order to produce a purchase. Once an agent completes the purchase directly, platforms can skip the attention step and charge for the result. Five models will coexist, ranked by likelihood:

Table: SE Incubator Consulting

The paid-placement dilemma. An agent is valuable because it works for the user. If recommendations are quietly bought, users will stop trusting it with their wallet. I think if platforms do sell placement, it will be far more restrained than search ads, for example paying only to rank among several options that are all good enough, with clear labelling.

How merchants compete. Agents do not look at banner ads; they read price, stock, specifications, return policies and reviews. Competition shifts from the most eye-catching ad to the cleanest, most machine-readable data. That opens a new risk as well: merchants may try to embed instructions in product data to manipulate agents.

Attribution is unresolved. Agents draw on review sites and creators, but users no longer click their links, blurring who earns the affiliate commission (Affiverse). How content producers get paid in an agent economy remains an open question.

The largest share of value will go to whoever controls three positions: the agent entry point, checkout and payment rails, and product data.

What this means for Europe and Austria

Neither product is available in Europe yet, which makes this a window to prepare.

Europe may take a third path. In China, super-apps decide through permissions what agents may do; in the US, commercial deals between platforms decide. In Europe, regulation is likely to draw the lines first. Relevant questions include:

• Disclosure of paid placement: EU consumer rules already require paid ranking in online search results to be labelled, and extending that principle to agent recommendations seems likely.
• Data access: agents like Muse read email, calendars, payments and even health data, which sits in natural tension with GDPR's data-minimisation principle. This may be one reason Muse is reaching Europe later.
• Platform blocking: Amazon, Meta and Google are all gatekeepers under the EU Digital Markets Act. Whether a gatekeeper may unilaterally refuse third-party agents is an open question.

For European SMEs. The first real demand will not be advertising inside agents but making product data agent-readable: accurate prices, stock, specifications, delivery and return terms. Merchants already on Shopify can start by auditing catalog completeness; for now, Muse is open only to merchants that ship to the US (Naughton & Bird).

Signals to watch:

• Whether WeChat, Meituan and Taobao formally opt out once Doubao's SAEP notice period ends on 15 October;
• Muse's paid conversion and actual transaction volume;
• Whether Meta introduces paid placement in Muse;
• A European launch timeline for Muse or similar agents, and the regulatory response.

Sources

Information as of 22 September 2026. Both products are less than a month old; most figures come from official claims or early reporting and may change. Chinese-language sources are marked (ZH).

• Meta: Introducing Muse
• Axios: Meta debuts Muse
• TechCrunch: Meta debuts its Muse AI agent
• Tech Insider: Muse pricing and rollout
• TradingKey: Amazon blocks Muse
• Seeking Alpha: Shopify partners with Meta
• 24/7 Wall St.: Shopify and Amazon stock reaction
• Time News: Muse app-store ranking and subscriptions
• Defense World: Muse early downloads
• Affiverse: Muse and affiliate attribution
• Naughton & Bird: Muse for Shopify merchants
• VerityScore: Muse and the Shopify catalog
• Caixin Global: Nubia releases second-generation AI phone
• Sohu: Doubao phone goes mainstream (ZH)
• Tencent News: second-generation Doubao assistant and SAEP (ZH)
• Tencent News / Kuai Keji: NaviX Ultra launch and pricing (ZH)
• Tencent News: NaviX Ultra pre-registrations (ZH)
• IT Home: NaviX Ultra first-day sales (ZH)
• SMZDM: launch-day details and first-generation resale prices (ZH)
• Sina: Doubao assistant app support and pricing (ZH)
• Leikeji: GUI agents and app traffic models (ZH)
• tech-ish: NaviX Ultra is China-only
• gagadget: NaviX Ultra outside China

Ting Wasner-Lian is the founder of SE Incubator Consulting in Vienna and helps European decision-makers understand Chinese AI.

All China–US AI Insights · About the author