AI and R&D

Figures converted from renminbi at historical period-end FX rates — see data/company.json.fx_rates. Ratios, margins, and multiples are unitless and unchanged. Financials are reported in renminbi and converted here to US dollars; the shares trade in Hong Kong dollars.

AI and R&D

Meituan's research and development spend rose 23.5% to $3.6 billion in FY2025 — a record, and the first increase in four years after three years of falling R&D intensity [1]. Management attributes the step-up to central AI investment: the proprietary LongCat models and the Xiaomei and Xiaotuan assistants [2]. At $0.7 billion the increase is small next to the $5.6 billion subsidy-war marketing surge, but unlike that surge it looks structural — and so belongs in normalised earnings.

Source: FY2025 Annual Report, MD&A — Research and Development Expenses [3].

R&D broke a three-year discipline

For three years Meituan held its research budget flat while revenue grew, and let the ratio fall. R&D was $3.0 billion in 2022, up from $2.6 billion in 2021, holding at 9.4% of revenue [4]. It stayed at $3.0 billion in 2023 as the ratio dropped to 7.7%, and $2.9 billion in 2024 as it fell again to 6.2% [5]. The absolute number barely moved for three years; the operating leverage came from spreading it over a larger business.

FY2025 reversed that. Spend rose to $3.6 billion and the ratio ticked back up to 7.1%, the first increase in either measure since 2022 [6]. The pace accelerated into the fourth quarter, when R&D rose 29.7% year over year to $1.0 billion and reached 7.6% of revenue [7]. The consolidated income statement carries the same figures — R&D of $3,640 million against $2,884 million a year earlier [8].

Source: derived from FY2022, FY2023 and FY2025 Annual Reports, MD&A — R&D Expenses [9] [10] [11].

The intensity number carries a useful calibration. At 7.1% of revenue, FY2025 R&D is still below both 2022 (9.4%) and 2023 (7.7%) [12] [13]. What changed is not the size of the budget relative to the company — it is the direction. Three years of declining intensity ended, and a research line that had been a source of operating leverage became a source of drag.

Where the money is going

Management is explicit about the destination. Meituan launched its in-house large language model, LongCat, in early 2023, and has since built out a model family — foundational models through business applications — and open-sourced core versions covering dialogue, reasoning, omni-modal understanding, and image and video generation [14]. On top of LongCat it has released consumer-facing assistants: Xiaomei as a standalone app and Xiaotuan embedded in the Meituan app for all users, meant to turn search into natural-language requests [15].

The chairman frames this as strategy, not experiment. The FY2025 letter commits Meituan to "driving the AI transformation of the physical world," and the 2026 outlook elevates it to a "Retail + Technology" corporate strategy built on combining the LongCat models with the company's merchant data, consumption behaviour and user reviews [16] [17]. The spend reaches beyond consumer apps: an internal AI coding effort put a dedicated team in front of over 6,000 frontline engineers to raise development productivity [18], and the company continues to fund L4 autonomous delivery vehicles and a Shenzhen robotics institute as longer-dated fulfilment bets [19].

The competitive context argues that at least part of this is table stakes. Meituan's largest instant-commerce rivals are making the same commitment on a larger scale — Alibaba, whose Taobao Instant Commerce is the escalating second front in the delivery war (The Delivery War), told shareholders it is "scaling up investments in our full-stack AI capabilities" including AI infrastructure and proprietary chips [20]. A local-services platform whose largest competitors are all building AI-native assistants cannot credibly not spend here. The open question is whether Meituan's version buys more than parity — whether proprietary models fed by the platform's dispatch, merchant and review data become an efficiency and discovery advantage, or simply a cost of staying in the game.

Small next to the subsidy war — but it does not reverse

The AI budget is not what broke the 2025 income statement. Of the operating-cost increases that turned a $5.0 billion operating profit into a $3.5 billion operating loss, selling and marketing dominated, rising $5.6 billion year over year as Meituan defended share against JD and Alibaba [21]. The $0.7 billion R&D increase is the second-largest cost step-up of the year, but it is roughly an eighth the size of the marketing surge.

Source: derived from FY2025 Annual Report, Consolidated Statement of Comprehensive Income [22].

The distinction that matters for valuation is between the two. The marketing surge is cyclical: it is the cost of a subsidy war that regulators have moved to cool and that management expects to ease, and the recovery case (What the Price Implies) rests on most of it reversing. The R&D step-up carries no such expectation. Management describes AI as a permanent strategic pillar, not a defensive burst, and has told the market to expect more of it, not less [23]. When the marketing spend rolls off, the research spend most likely stays.

That has a specific consequence for the earnings anchor the valuation leans on. Meituan discloses R&D only at the group level, and the segment accounts allocate operating costs — R&D among them — into Core Local Commerce and New Initiatives rather than parking them in the corporate line: unallocated items were essentially flat in 2025, at $1.1 billion against $1.1 billion in 2024 [24]. So the $0.7 billion increase sits inside the segment cost bases. The pre-war Core Local Commerce operating profit of $7.2 billion in 2024 [25] was earned while group R&D was $2.9 billion. If research resets to $3.6 billion or higher and that cost stays allocated to the segment, a full restore of Core Local Commerce volumes would land at a normalised profit modestly below the $7.2 billion headline — by the segment's share of the step-up — unless AI efficiency offsets it. The trough reverses; part of the cost structure does not.

The read

The evidence points to an AI program that is defensive in origin and optional in payoff: a necessity Meituan could not skip, and a bet it might win. The necessity is clear: every major instant-commerce rival is funding frontier models, and a platform that mediates local-services discovery cannot cede the natural-language interface [26]. The optionality is real but unproven: LongCat fed by Meituan's dispatch, merchant and review data could lower fulfilment cost and deepen the density advantage that underwrites the moat (Demand Engine), or it could settle into a permanent ~7%-of-revenue cost that buys parity rather than advantage.

The strongest fact against treating this as a problem is that the spend is modest and self-funded. R&D intensity at 7.1% remains below where it sat in 2022 and 2023 [27], the increase is a fraction of the marketing swing, and it is being financed from a business that still converts profit to cash (Cash Conversion) rather than from new leverage. This is not a company betting the balance sheet on AI; it is a company redirecting a research line it had been shrinking. What would change the read is direction and payoff: R&D intensity climbing toward the 9%–10% range with no visible efficiency return in Core Local Commerce margins would mark a genuine reset of normalised earnings power; evidence that AI is lowering cost per order or lifting conversion would move it from cost to moat. Both are quarterly-checkable, and neither is settled yet.