Meituan Bets on "Physical AI" Infrastructure Play, Sidestepping the Model Arms Race
Meituan is restructuring its artificial intelligence operations around a calculated wager: rather than outspend Alibaba and ByteDance on foundation models, it will embed AI into the physical logistics network that processes tens of millions of daily transactions — and position that network as indispensable infrastructure for every AI agent in China.
The strategic pivot crystallized this week with the formation of an AI Transformation division within Meituan's Core Local Commerce (CLC) unit. The new department sits alongside the food delivery and flash-retail verticals and reports directly to CLC Chief Executive Wang Puzhong — a deliberate elevation from its previous line into Senior Vice President Li Shubin, the platform chief. The structural change signals that AI has graduated from a technology experiment to a board-level commercial priority at China's dominant on-demand services platform.
Heading the new unit is Mu Yao, former general manager of Dianping, Meituan's reviews-and-discovery subsidiary. The appointment of an operator rather than a technologist is consistent with Meituan's stated philosophy: AI must serve the business, not the other way around.
Reorganizing Elevates AI From Lab Project to Revenue Mandate
The reporting-line change carries operational weight. Previously, Meituan's in-house large language model team, its consumer-facing AI apps and its business-to-merchant tools all funneled decisions through Li Shubin's platform organization — a structure that critics inside the company said slowed cross-departmental resource allocation. With AI Transformation now under Wang Puzhong, the unit can mobilize CLC's merchant network, rider fleet and transaction data without negotiating across organizational silos.
Meituan CEO Wang Xing set the directional tone as early as Q1 2025, telling analysts the company would pursue AI "offensively, not defensively." Since then, the LongCat model series — first released in September 2025 — has expanded to 12 models spanning text, image, audio and video modalities. The pace of iteration, roughly one new model per month, mirrors the cadence of dedicated AI labs rather than a diversified internet conglomerate.
Talent acquisition has reinforced the push. Pan Xin, formerly a partner at Shanjike AI and head of ByteDance's visual foundation model platform, joined Meituan in late 2025 and now leads multimodal AI innovation, including the LongCat App. Pei Peng, LongCat's base-model lead, has appeared at multiple academic forums since 2025, a rare instance of public visibility for a team that has otherwise operated with minimal external profile.
Tabbit's 100-Day Sprint Tests Meituan's Consumer AI Ambitions
The most visible consumer product to emerge from the restructuring is Tabbit, an AI-native browser developed by the GN06 team — the unit formed from the integration of Guangnian Zhiwai, the AI startup Meituan acquired for RMB 2.065 billion (approximately US$287 million) in 2023. Guangnian Zhiwai was founded by Wang Huiwen, Meituan's co-founder and former senior vice president, making GN06 one of the most closely watched internal ventures in Chinese tech.
Tabbit entered public beta after roughly 100 days of testing, during which the team shipped more than ten version updates. The product entered a crowded field: domestically, it competes with Quark, ByteDance's browser offering and 360 AI Browser; globally, it faces Google Chrome, Microsoft Edge and OpenAI's Atlas. Tabbit's differentiation strategy relies on model aggregation rather than proprietary model dominance — the browser integrates DeepSeek, Douba, Qianwen and Kimi alongside Meituan's own LongCat-Flash-Chat.
The international dimension is now live. Tabbit has quietly launched an overseas version and, after pausing international hiring in mid-2025, has reopened recruitment for offshore roles — a signal that GN06 is testing distribution beyond China's domestic market. The team's product history includes AI image and video generation tool Miaoshua and emotional companion app WOW, launched in 2023 and late 2024 respectively, suggesting an appetite for consumer AI experiments that extends well beyond Meituan's core delivery use case.
"To A" Strategy Redefines AI Agents as Customers, Not Tools
The most strategically distinctive element of Meituan's AI roadmap is what Wang Xing, during the Q1 2026 earnings call, labeled the "To A" model — treating AI agents as a third customer category alongside consumers (To C) and merchants (To B).
The practical implementation is already underway. In May 2026, Meituan launched "Errands Skill", packaging its on-demand courier capability as an encapsulated skill that third-party AI assistants can call directly. Multiple AI assistant platforms have already integrated the service. More consequentially, Wang Xing disclosed that Meituan's AI agent Xiaomei is preparing to go live inside Tencent Yuanbao, enabling users to place food delivery orders through the Tencent AI interface without switching to the Meituan app.
The architecture differs meaningfully from Alibaba's closed-loop model, in which Tongyi Qianwen handles both the query and the transaction within Alibaba's own ecosystem. Meituan's approach is deliberately open: the company is willing to cede the conversational interface to whoever wins that battle — whether Yuanbao, Doubao or a future entrant — so long as Meituan's merchant network and rider infrastructure execute the physical fulfillment. Wang Xing has framed this as an information-asymmetry advantage, noting that even a hypothetically omniscient AI cannot know in real time whether a specific restaurant has available seating — only Meituan's live operational data can.
That data advantage spans over 2,800 cities and counties across China, processing tens of millions of transactions daily. It is the asset Meituan is monetizing through its "Physical AI" strategy: a bet that the value in the AI economy will accrue not only to model builders but to whoever controls the ground-level data loops and last-mile execution networks.
Drone Network Extends the Physical Perimeter
Underpinning the Physical AI thesis is Meituan's drone delivery program, a business the company has developed since 2017 when Mao Yining joined with his founding team. By September 2022, the drone unit had grown to more than 300 engineers, many of them established leaders in AI and hardware. In February 2024, Wang Xing restructured reporting lines so that the drone division reports directly to him — the same elevated status he assigned to Meituan's overseas operations — signaling that autonomous aerial delivery is a strategic priority, not a research curiosity.
Drone delivery extends the Physical AI perimeter from road-level logistics into low-altitude airspace, a domain that China's regulators are actively developing under the "low-altitude economy" policy framework. For Meituan, capturing this channel early would widen the fulfillment moat that anchors the entire To A strategy.
Heavy Spending, Muted Visibility Creates a Valuation Perception Gap
Meituan's AI investment is substantial by any measure. Wang Puzhong has publicly stated annual AI spending exceeds RMB 10 billion (approximately US$1.39 billion). Wang Xing has noted that since early 2023, capital expenditure and AI talent hiring have accelerated, with Meituan ranking among China's top spenders on proprietary model development — trailing only major cloud computing platforms.
Yet the company's AI narrative has struggled for mindshare. Xiaomei's Android download count stands at approximately 1.69 million — a modest figure relative to the hundreds of millions of monthly active users on the core Meituan app. LongCat's team rarely surfaces publicly. Wentuan, the AI assistant promoted to the bottom navigation bar of the Meituan app ahead of the May Day holiday in 2026, has generated limited consumer buzz compared with Alibaba's red-packet campaigns around Qianwen-powered food ordering.
Internally, however, AI adoption metrics tell a different story. A Meituan technical manager quoted by Huxiu in early 2026 stated that more than 95% of coding work in at least one core business unit now runs through CatPaw, Meituan's proprietary AI development tool. The company has formally designated "AI-driven productivity improvement" as a key 2026 operational target, encompassing both software development acceleration and unmanned delivery expansion in complex environments.
The divergence between heavy internal deployment and weak external perception is the central tension in Meituan's AI story heading into the second half of 2026. The formation of AI Transformation under Wang Puzhong is in part an organizational answer to that gap — consolidating scattered product efforts under a single mandate with direct access to the company's largest revenue engine.
Whether Meituan's infrastructure-first positioning generates the kind of AI premium that investors have rewarded at pure-play model companies remains the open question. The company's answer, implicit in every structural decision it has made this year, is that the physical world cannot be tokenized — and that whoever owns the fulfillment layer owns the terminal node of the AI economy.
Related Coverage:
Meituan Loosens Its Grip on the Front End to Survive China’s AI Shift