AI 硬件全景图 · 一次人机交互的入口占用 · 2026-07 · 消费电子×AI 原生设备AI Hardware · One Seized Interaction · Jul 2026 · Consumer Electronics × AI-Native Devices
硬件之争,是入口之争
The entrance war
硬件之争是入口之争:最小单元不是「一台设备」,而是「一次人机交互的入口占用」——谁截住你张嘴提问、抬手拍照、开口说「帮我订一下」的那个瞬间,谁就拿到 Agent 时代的入场券。而这一代最贵的教训已经写好:「替代手机」是幻觉,「寄生手机」才是活路;硬件是订阅的容器,赚钱的从来不是那个壳
The hardware war is an entrance war: the atom is not «a device» but «one seized moment of human-machine interaction» — whoever intercepts the instant you open your mouth to ask, raise your hand to shoot, or say «book it for me» holds the ticket to the agent era. And this generation’s costliest lesson is already written: «replacing the phone» is a mirage; «parasitising the phone» is the way to live — hardware is the container of the subscription, and the shell was never where the money was
本图相信:30 天留存是硬件唯一的真话。一 AI 原生设备约 10 万买家、五个月后同一时刻仅约 5000 人日活(≈95% 弃用);另一台退货多于销量、烧掉 2.3 亿美元后以 1.16 亿卖身、已售设备服务器断连彻底变砖——发布会参数与真实留存之间的落差,就是硬件的照妖镜(B)。正面反例同样清晰:一 AI 眼镜系 2025 全年出货逾 700 万台、同比翻三倍。This map believes: thirty-day retention is hardware’s only truth. One AI-native device kept ~5,000 concurrent daily users out of ~100k buyers five months in (≈95% abandonment); another returned more units than it sold, burned $230M, sold for $116M and bricked every shipped unit when the servers cut — the gap between keynote specs and real retention is hardware’s truth-mirror (B). The counter-example is equally crisp: Meta’s Ray-Ban glasses shipped 7M+ units in 2025, tripling year on year.
本图不相信:「AI 超级换机周期」。2024 年出货回暖主要来自库存补货与宏观修复而非 AI;仅 7% 美国受访者称 AI 会让其提前换机;美国换机周期 3.46→3.84 年拉长、全球均值 3.5–3.6 年微动、中国反而 36→33 个月缩短——「3 年变 4 年」是区域性结论,不是全球规律(B)。AI 撬得动交互原语,暂时撬不动换机欲望。It does not believe in the «AI supercycle». The 2024 shipment recovery came from restocking and macro repair, not AI; just 7% of surveyed Americans say AI would make them upgrade sooner; the US cycle stretched 3.46→3.84 years while the global mean idled at 3.5–3.6 and China shortened 36→33 months — «three years to four» is a regional finding, not a global law (B). AI moves the interaction primitive; it does not yet move the urge to upgrade.
主脊:一款硬件的生命周期七环节(立项定义→方案与供应链→制造→渠道→激活留存→生态订阅→换机与入口战),②⑤标红;三条结构带:AI 原生硬件尸检、跑通的品类(眼镜/耳机/手表)、入口战争第三幕(PC 浏览器→超级 App→Agent 入口);判断层:场景刚需与替代成本 ＞ 技术炫技;生态与分发 ＞ 硬件本身。姊妹:算力→energy,制造→factory,汽车座舱→car,陪伴玩具→companion。
The spine: one device’s lifecycle in seven links (definition → solutions and supply chain → manufacturing → channel → activation and retention → ecosystem and subscription → upgrade cycles and the entrance war), ② and ⑤ flagged; three bands: the AI-native autopsy, the categories that work (glasses, earbuds, watches), act three of the entrance war (browser → super-app → agent); the judgment: scene necessity and switching cost ＞ spec theatre; ecosystem and distribution ＞ the hardware itself. Siblings: compute → energy, manufacturing → factory, the car cabin → car, companion toys → companion.
700 万台
一 AI 眼镜系 2025 全年出货(同比翻三倍;自 2023-10 上市累计 200 万副后引爆)——引爆点不是硬件是 AI 功能(持续「看世界」+实时口译):全景图里唯一跑通的新品类(A/B)Meta’s Ray-Ban line: 2025 shipments (tripled year on year, after 2M cumulative since Oct 2023) — ignited not by hardware but by AI (always-on «seeing the world» plus live interpretation): the map’s only new category that works (A/B)
≈5%
一 AI 原生设备发布约五个月后的同时在线日活比例(约 5000/10 万);另一台退货>销量、装机 8000→7000——「替代手机」路线的墓志铭;共同死因:头部模型商的语音功能一上线,独立硬件卖点被手机 App 免费吸收(B)One AI-native device’s concurrent daily-active share ~5 months after launch (~5,000 of 100k); the other returned more than it sold, installed base 8,000→7,000 — the epitaph of the replace-the-phone route; the shared cause of death: the moment OpenAI’s GPT-4o voice shipped, the phone app absorbed the standalone’s selling point for free (B)
130 元
华强北不带摄像头 AI 眼镜的出货价(带摄像头约 380 元;品牌同类 2000 元+,价差 10–20 倍);「硬件壳+APP 芯」上午设计下午打样;一开源方案已赋能硬件超 200 万台(B)——深圳把 AI 硬件边际成本打到地板The Huaqiangbei price of camera-less AI glasses (with camera ~¥380; brands at ¥2,000+ — a 10–20× spread); «shell hardware, app brain» designs in the morning and prototypes by afternoon; one open-source stack already powers 2M+ devices (B) — Shenzhen floors AI hardware’s marginal cost
250 亿美元
一语音助手设备部门 2017–2021 累计亏损(单年一度预亏 100 亿,内部称「巨型失败」):累计售出 5 亿+ 台、周交互 10 亿次,却多是问天气放音乐——装机量大≠变现:AI 只解决了唤醒,不解决刚需(B)One voice-assistant device division’s 2017–21 cumulative loss (once projected at $10B in a single year, «a colossal failure» internally): 500M+ units sold, a billion weekly interactions — mostly weather and music — install base is not monetisation: AI solved the wake word, not the need (B)
⚠️ 口径裁判(先读):① 换机周期结论有区域性冲突(美国 3.46→3.84 年 vs 全球均值 3.5–3.6 年微动 vs 中国 36→33 个月缩短),勿一概说「3 年变 4 年」;② AI 眼镜中国出货两套机构口径(246 万 vs 275 万台);③ 内存占手机 BOM 比例存 10–20% vs 30–40% 口径冲突;④ 一 AI 原生设备出货有 10 万 vs 13 万两说、且无严格「30 天留存」官方数(公开可证的是「约 5 个月后 ≈5% 日活」);⑤ AI 玩具/陪伴市场规模各机构软硬混算、从 25.6 亿到 5525 亿美元不等,本页只作数量级;⑥ 模型公司硬件的量产时点(2026 下半年 vs 2027 初)与产能(3000 万–1 亿台)均为演进中传闻;⑦ 「马斯克 AI 玩具」系误传已校正(其公司做的是儿童版应用,与那款毛绒玩具无关);⑧ 厂商发布会数字按 D 降权;四份深度研究交叉整理(~70 次双语检索+多源核验);渗透%为编辑估值。
⚠️ Basis rulings (read first): ① upgrade-cycle findings conflict by region (US 3.46→3.84 years vs a global mean idling at 3.5–3.6 vs China shortening 36→33 months) — never flatten to «three to four years»; ② China AI-glasses shipments carry two house bases (2.46M vs 2.75M); ③ memory’s share of phone BOM splits 10–20% vs 30–40%; ④ one AI-native device’s shipments run 100k vs 130k with no official 30-day retention (the provable figure: ~5% daily active at ~5 months); ⑤ AI-toy market sizes mix hardware and software across houses ($2.56B to $552.5B) — magnitudes only; ⑥ the OpenAI-io device timing (H2-2026 vs early 2027) and volumes (30M–100M) are evolving rumours; ⑦ the «Musk AI toy» is a corrected misattribution (his firm ships a kids’ app, unrelated to that plush toy); ⑧ keynote numbers grade D; cross-compiled from four deep-research reports (~70 bilingual searches, multi-source verification); penetration %s are editorial.