In the run‑up to the Chinese New Year, when firms compete with red‑envelope giveaways, Alibaba chose a different gambit: it set aside RMB3 billion to fund a promotion that let users order free milk tea through its new Qianwen app. The campaign reportedly generated 10 million orders in nine hours and pushed Qianwen to the top of Apple’s App Store free download chart, a blunt demonstration of scale that reads as both marketing and systems test.
The user experience was deliberately frictionless. Every Qianwen user received a RMB25 voucher usable across Taobao flash sales, takeaways and, crucially, milk‑tea purchases; the app’s AI populated addresses, recommended vendors and completed checkout via an Alipay AI payment flow. For consumers it was convenience; for Alibaba it was a live experiment in binding language models, commerce flows and payments into a single closed loop.
What makes the stunt noteworthy for a global audience is not the novelty of free drinks but the engineering story beneath it. Qianwen runs on Alibaba’s in‑house model lab (Tongyi), sits on Alibaba Cloud, and relies on proprietary silicon developed by Pingtouge, known as the Zhenwu PPU. That stack reduces marginal token and compute costs and, more importantly, gives engineers end‑to‑end control when millions of simultaneous inference requests threaten to produce a ‘compute tsunami’ that would overwhelm ordinary setups.
The move echoes broader industry talk about the next phases of AI. At CES, Nvidia’s Jensen Huang argued for AI that reaches into the physical world; industry practitioners increasingly talk in stages — sensing, generation, agency and physical effect. Alibaba’s campaign is a practical validation of the ‘agent’ stage: models not merely producing text but triggering multi‑step actions across real services, and beginning to close the loop with real‑world feedback.
Alibaba’s advantage is not only cheaper tokens or cheaper chips. Its sprawling consumer ecosystem — Alipay, Taobao, AutoNavi (Gaode), Fliggy and others — shares single sign‑on and internal APIs, allowing Qianwen to operate inside the flows that deliver real goods and payments. That kind of native integration is something international model providers and cloud vendors cannot easily replicate, and it gives Alibaba a privileged position to convert model outputs into executed transactions.
That privileged integration also creates a powerful training and product feedback loop. When a recommended tea shop is closed or an address fails, those operational errors become supervisory signals that can be used to refine models and orchestration code. The result is rapid, repeated real‑world validation — exactly the kind of data‑rich loop that accelerates an AI system’s move from clever text generator to reliable digital agent.
There are broader strategic stakes. Running model, cloud and silicon in‑house reduces unit costs and raises the barrier to competition, while the ability to route execution through consumer apps concentrates control over user interactions and valuable behavioural data. Regulators and rivals will watch whether such integrations entrench dominant positions, give rise to new privacy risks, or permit preferential treatment of in‑ecosystem merchants.
For observers of the AI arms race, Alibaba’s milk‑tea giveaway is a test case: a marketing stunt that doubled as a stress test for full‑stack AI engineering and a milestone on the road from generative language models toward agent‑based and physical AI. The coming months will reveal whether this kind of closed‑loop deployment scales beyond promotions and becomes a structural advantage in retail, logistics and local services.