餐饮产业 · 一餐的生命周期 × AI · 2026-07 版The Restaurant Industry · A Meal’s Lifecycle × AI · Jul 2026

餐饮被 AI 劈成两半:能标准化的后端被算到极致,那口锅气与那桌人情一寸未动——真正的战争,发生在中间那道「用了要不要告诉你」的缝里 AI splits dining in two: the standardizable back end is computed to the limit, while the wok-breath and the shared table stay untouched — the real war is fought in the seam between, over “must they tell you they used it”

「预制菜渗透率」没有唯一真值——同一个词,惊悚全靠选口径。引爆信任危机的从不是「能不能用」,而是「用了要不要告诉你」:知情权,而非食品安全。“Pre-made penetration” has no single true value — one phrase, and the shock is entirely in the chosen basis. What detonates the crisis is never “can they use it,” but “must they tell you they did”: the right to know, not food safety.

看清一件事:万店连锁的品牌方,是披着 2C 外衣的 2B 供应链公司——赚的是卖货给加盟商的钱,不是你那顿堂食的利润。单店九死一生,总部旱涝保收。See one thing clearly: a ten-thousand-store brand is a 2B supply-chain company in 2C clothing — it earns by selling goods to franchisees, not on the profit of your dine-in meal. The single shop mostly dies; headquarters is rain-or-shine.

最小单元:一餐的完整满足。主脊是一餐生命周期八节点双轨(想吃什么→找店→点单→制作→交付→吃→评价复购→回流供应链),每节点并置「传统 vs AI」并挂成熟度四档(存量已普及/规模化/试点/概念营销)——防吹牛的关键装置。三条诚实层张力:锅气 vs 预制菜口径战、平台税两套算法、加盟与资本线;供给侧暗线是一家店的死法图鉴与杠杆排序;外加龙头矩阵与产品指南(2C/2B)。交叉互挂:种草与投流 → traffic,饭桌上的关系仪式 → sell,食材上游订单农业 → farm。 Atomic unit: the complete satisfaction of one meal. The spine: an eight-node meal lifecycle on twin tracks (craving → finding → ordering → production → delivery → eating → review & repeat → flowing back to the supply chain), each node pairing «traditional vs AI» and tagged with a four-step maturity scale (ubiquitous / scaled / pilot / marketing concept) — the anti-hype device. Three honesty-layer tensions: wok-breath vs the pre-made basis war, the platform tax’s two arithmetics, and the franchise-capital line; the supply side’s shadow spine is a single shop’s death catalog and leverage ranking; plus a leader matrix and a product guide (2C/2B). Cross-links: seeding & ad-bidding → traffic, the ritual of the shared table → sell, upstream contract farming → farm.

传统节点Traditional

AI / 数字化节点AI / digital node

供应链 · 去厨师化Supply chain · de-chefing

平台税 · 调度Platform tax · dispatch

死法 · 割韭菜Death · the franchise trap

+20.4→+3.2%

餐饮增速的急刹车曲线:+20.4%(2023)→ +5.3%(2024)→ +3.2%(2025,全年 57,982 亿,统计局 A 级)。疫后报复性反弹两年内熄火——增长的蛋糕没了,AI 成了存量厮杀的武器The growth curve slams the brakes: +20.4% (2023) → +5.3% (2024) → +3.2% (2025; ¥5.80T full-year, NBS, grade A). The post-Covid rebound died within two years — with no growing pie left, AI became the weapon of a zero-sum fight

口径战

预制菜「渗透率」没有唯一真值:广义规模 2023 约 5165 亿、2026 预测破万亿(C);B 端使用率常说 60–80%,C 端仅个位数;2026 卫健委国标征求意见稿把央厨菜肴/净菜/主食排除在外——监管定义在收窄,公众直觉在扩张Pre-made “penetration” has no single true value: broad market ~¥516.5B (2023), forecast past ¥1T by 2026 (C); B-side use cited 60–80%, C-side single digits; the 2026 draft national standard excludes central-kitchen dishes, washed produce and staples — the regulator narrows the definition while public intuition widens it

27.83元

一笔 46.9 元外卖订单,商家实际到手(C 级案例)。名义技术服务费仅 6–8%,但佣金+配送+投流+满减+包装的综合负担:客单 <20 元可达 30–40%——客单越低,平台税越狠What the shop actually keeps of a ¥46.9 delivery order (grade C case). The headline tech-service fee is just 6–8%, but the all-in burden of commission + delivery + ad-bidding + discounts + packaging hits 30–40% under a ¥20 ticket — the lower the ticket, the heavier the tax

95%+

蜜雪冰城商品/设备销售占收入比(加盟费仅个位数,C);锅圈 90%+ 收入来自向 11,566 家加盟店卖货(财报 A)——万店连锁真相=披着 2C 外衣的 2B 供应链公司;门店平均寿命 ≈500 天(旧口径 C)Mixue’s goods/equipment sales as a share of revenue (franchise fees a single digit, C); Guoquan earns 90%+ by selling goods to its 11,566 franchisees (filings, A) — the ten-thousand-store truth: a 2B supply-chain company in 2C clothing; average shop lifespan ~500 days (dated basis, C)

口径警告:本页综合四份深度研究交叉整理,机制与模式高可信;数字按 A(财报/统计局)、B(权威转述)、C(第三方估算)、D(厂商自述)分级标注,D 级一律降权。预制菜四口径(广义规模/B端使用率/连锁后厨使用率/C端渗透率)不可互换,「万亿」大头是速冻面米与净菜;外卖费率有平台口径与商家口径两套算法(6–8% vs 综合 20–40%),本页两边并列;闭店三口径(标记停业 339 万/关闭 296 万/注销)不能互换,企查查注销数含大量空壳;损耗率两口径(门店餐损 8%+ vs 供应链生鲜损耗 2.5–3%)是「选基线讲故事」的典型;炒菜机器人的性能数字几乎全部来自厂商(D)。门店端「炒菜机器人」多为营销,真「去厨师化」靠中央厨房——千玺机器人餐厅(2020–2023)是「机器人餐厅经济性不成立」的实证墓碑。海外侧(欧美 drive-thru 语音、零工立法)四份文档覆盖单薄,本页以中国为主、海外为注脚。 Basis warning: cross-compiled from four deep-research reports; mechanisms and models are high-confidence, figures graded A (filings/NBS), B (authoritative citation), C (third-party estimate), D (vendor claim) — D always down-weighted. The four pre-made bases (broad market / B-side use / chain-kitchen use / C-side penetration) are not interchangeable, and the bulk of the “trillion” is frozen staples and washed produce; delivery rates carry two arithmetics — the platform’s and the shop’s (6–8% vs an all-in 20–40%), shown side by side; the three closure bases (339M flagged dormant / 296M closed / deregistrations) don’t mix, and deregistration counts include shells; the two waste-rate bases (8%+ in-store vs 2.5–3% supply-chain) are a textbook case of “pick your baseline, tell your story”; cooking-robot performance figures are almost entirely vendor claims (D). Store-side robots are mostly marketing; real “de-chefing” runs on central kitchens — Qianxi’s robot restaurant (2020–2023) is the empirical tombstone that robot-restaurant economics don’t hold. Overseas coverage (US drive-thru voice AI, gig-work law) is thin in all four sources; this page is China-first with overseas as footnotes.

◆ 诚实层 · 地基The Honesty Layer · foundation

餐厅卖你的故事 vs 账本里的事The story the restaurant sells vs what the ledger records

读这张图前,先把四句话摆在桌面上。AI 在餐饮真正跑出规模的只有四类——外卖调度算法、送餐机器人、扫码点餐/排号、团餐视觉结算台;其余大多卡在「试点→规模化」或干脆是「营销概念」。性感的机器人餐厅,反而是反面教材。Before reading the map, put four sentences on the table. Only four AI categories truly scaled in dining — delivery dispatch, serving robots, scan-to-order/queue, and canteen visual checkout; the rest are stuck between “pilot-to-scale” and outright “marketing concept.” The sexy robot restaurant is the cautionary tale.

立场声明:本页是批判性、祛魅的行业结构分析,用 A–D 角标区分财报/统计局一手与厂商自述,口径打架处标 ⚠️ 两边并存。既不做「AI 万能」的吹鼓手,也不做「AI 无用」的哀悼者;产品指南是市场地图,不构成经营或投资建议、不构成推荐背书。核心判断一句话:AI 不一定能做出最好吃的一餐,但它决定这一餐被谁看见、由谁生产、多久送达、能否复购,以及这家店能不能活下去。 Stance: a critical, demystifying structural analysis; A–D badges separate filings/NBS primaries from vendor claims, with ⚠️ where bases clash — both sides kept. Neither an AI hype-man nor a doom-monger; the product guide is a market map, not business or investment advice, nor an endorsement. The core judgment in one line: AI may not cook the best meal, but it decides who sees it, who produces it, how fast it arrives, whether you reorder — and whether the shop survives.

从这张图看到的五条规律Five patterns this map makes visible