Honor shipped MagicOS 9.0 across its 2024 flagship lineup in October 2024 with an on-device AI agent, YOYO, that the company says comprehends 600 user intents and executes 900 tasks automatically. The demonstrated example was a cross-app purchase rather than a chat window. Ask the phone for a latte, and YOYO opens the user’s delivery app, finds the preferred merchant, and places the order without a server round trip. The handset, not the marketplace, is now the first surface to see the query. The chokepoint that built Taobao, Shopee, and Amazon search economics has begun to migrate upstream of the platform, and the device maker, not the retailer, decides what counts as discovery.

The architecture change is the story, not the feature list. A three-billion-parameter model runs locally on Honor’s mid-range and premium hardware — with cloud compute available for more demanding tasks — and YOYO can remember, per Honor, 950 personal habits. The agent reads a spoken intent, checks its memory of prior behaviour, and completes the action before any marketplace ranker is consulted. The older shopping path ran from user to marketplace search to a ranked list of listings. The new architecture runs user-to-agent-to-result, with the marketplace absent from the routing table. The platform never receives the query, which means it never runs the auction that ordinarily decides whose listing appears first.

That auction is the marketplace business model, not an adjacent feature. Alibaba’s marketing-services arm Alimama, which sells the CPC inventory attached to Taobao and Tmall search and recommendation surfaces, generated more than $55 billion in ad revenue in fiscal 2024. Most of that sits inside customer-management revenue, the single largest line in Alibaba’s core marketplaces group. Amazon’s sponsored-listing business depends on the same query-reaches-platform flow. An impression that never occurs cannot be sold. Remove the query and you have not discounted the inventory; you have deleted it.

Honor has accelerated this architecture rather than retreated from it. At MWC 2026 the company advanced its existing Alpha Plan and promised international rollout of agentic capabilities. The phrase is marketing. The direction it marks is not. If Samsung’s One UI 8.5 and the next tiers of Apple Intelligence follow the same pattern of locating action at the OS layer, rather than inside any one merchant’s app, the default shopping surface on a 2026 flagship will have become the phone itself.

A query the marketplace never sees is an impression it cannot sell.

The cloud-AI version of this migration is already visible in the retail numbers. eMarketer reported a 670% year-over-year jump in generative-AI referral traffic to US retail sites in the first month of the 2025 holiday season. AI-driven shopping flow is arriving at scale, which is why platforms have stopped treating it as noise. Each of those clicks still travelled through the retailer’s own ranker; Perplexity or ChatGPT sent the shopper to an Amazon page, and Amazon closed the sale. On-device agents collapse that last step.

The strongest counter-argument is that the catalogue remains the moat. Taobao, Amazon, and Shopee hold the deepest inventories, the live prices, the review corpora, and the fulfilment hooks; any on-device agent is brittle without access to that data. For the main thesis to fail, the platforms have to win the API layer, forcing every handset agent to query their ranking engine before surfacing a product and attaching paid impressions to that query. Walmart is running the web-agent version of this defence already, as eMarketer reports, adding site rules that block external agents from completing checkout and preserving its own funnel. Those rules govern third-party browser agents. They have nothing to say to YOYO, which is the operating system and reads screen, habit, and intent on the user’s side of the TLS handshake.

The price is that retail-media budgets priced against the search impression do not survive the impression never occurring. Global retail media spend reached $184 billion in 2025 on Forrester’s estimate, most of it placed against on-platform surfaces the marketplace controls. When the on-device agent decides, the merchant’s route to the sale runs through whoever sets the agent’s rules, not through the marketplace’s ranker. That whoever is currently a Chinese handset maker and its silicon partners, with the rules being written inside Shenzhen product teams rather than inside any retail group. If this architecture keeps shipping at Honor’s pace, and if Samsung and Apple complete the same move, the next retail-media auction will not be Alibaba’s to run. It will belong to whoever owns the handset.