「AI 吞噬世界」全景图系列 · 建筑工程篇 · BUILDTHE «AI IS EATING THE WORLD» ATLAS · CONSTRUCTION

机器人征服了工厂,
却没征服工地Robots Conquered the Factory —
Not the Site

图纸与工地 · Blueprints & Mud · 不可移动的制造业Blueprints & Mud · The Immovable Manufacturing Industry

因为工厂造的是产品,工地造的是原型——每栋楼都是一次性的,自动化的编程与调试成本摊不到百万件上。AI 正以两种速度吞噬这个 31 万亿的行业:图纸上的一切以月为单位沦陷(造价、算量、标书、审图),泥土里的一切以十年为单位僵持(砌筑、绑扎、装修);而真正决定生死的垫资、签字与回款,至今不认识任何算法。Because the factory makes products and the site makes prototypes — every building is one of a kind, and automation’s programming and commissioning never amortise across a million units. AI is eating this ¥31-trillion industry at two speeds: everything on paper falls by the month (estimating, take-offs, bids, plan review), everything in the mud holds out by the decade (bricklaying, rebar-tying, finishing); while the advances, signatures and receivables that actually decide survival still recognise no algorithm.

本图相信THIS MAP BELIEVES

冲击是真的:造价、算量、标书、审图正在结构性塌缩(深圳 AI 审图查错 ≈7 倍、效率 ≈9 倍,B);盾构智能掘进、智慧工地、无人机巡检是真落地;劳动力断层(农民工均龄 43.3 岁、50 岁以上占 32%,A)让自动化从选择题变生存题。The shock is real: estimating, take-offs, bid documents and plan review are structurally collapsing (Shenzhen’s AI review finds ~7× the issues at ~9× the speed, B); intelligent tunnelling, smart sites and drone patrols genuinely landed; and the labour rupture (migrant workers average 43.3 years, 32% over 50 — A) turns automation from a choice into survival.

本图不相信THIS MAP DOES NOT BELIEVE

不信「三年无人工地」——建筑机器人市场约 9 亿元,不足总产值的万分之一(C);不信厂商倍数(16h→4h、+75%、2000 亿参数,一律 D);不信 AI 能改写垫资、分包、签字、验收这套操作系统——它只是把账照亮,没有换掉赌桌。Not «unmanned sites in three years» — the construction-robot market runs ~¥900M, under one ten-thousandth of output (C); not vendor multiples (16h→4h, +75%, 200B parameters — all D); not that AI rewrites the operating system of advances, subcontracting, signatures and acceptance — it lights up the ledger; it never changes the table.

蓝 = 图纸侧(AI 渗透)Blue = the paper side (AI advance) 橙 = 工地侧与硬骨头(方块编号)Orange = the mud side & hard bones (square numbers) A–D = 证据分级(A 官方 → D 厂商自述)A–D = evidence grade (A official → D vendor claim) 渗透% = 评估值,非统计值Penetration % = editorial estimate, not statistics

31.2 万亿建筑业年总产值(2022,国家统计口径 · A)——被 AI 撞上的体量Annual construction output (2022, national statistics · A) — the mass AI ran into

43.3 岁 · 32%农民工平均年龄 / 50 岁以上占比(2025 官方监测 · A)——自动化的总时钟Migrant workers’ average age / share over 50 (2025 official survey · A) — automation’s master clock

7× · 9×深圳 AI 审图:查出问题数 / 审查效率对人工的倍数(40 院 53 项目内测 · B)——图纸侧的狂飙Shenzhen AI plan review: issues found / speed vs human reviewers (53-project pilot · B) — the paper side’s blitz

9 亿 vs 31 万亿2024 建筑机器人市场 vs 总产值——占比约万分之 0.3(研究机构口径 · C)——工地侧的僵持The 2024 construction-robot market vs total output — ~0.003% (research-house basis · C) — the mud side’s stalemate

⚠️ 口径裁判(先读):① 43.3 岁为全体农民工口径(2025 官方监测),官方无分行业均龄,「建筑工人逼近 50 岁」系媒体口径本页未采;② 「9 亿」为研究机构口径(C),与早年「4.5 亿」预测不同源,不构成精确序列;③ 7×/9× 为深圳试点内测数据;④ 「信息化投入占产值 0.12–0.13%」为投入占比(单源 C),不是渗透率——与「64% 企业已尝试 AI」(C)并存不矛盾:钱少、面广、点状;⑤ 厂商效率数字(16h→4h、60 分钟标书、审查周期 −70%、2000 亿参数、验收 52→35 天)一律 D 级,理想工况;⑥ 各节点渗透%为编者综合四份定性评级的评估值,非统计数;⑦ 四份来源之一中途截断且引文列表丢失(其独有数字降 C),另一份自认离线撰写(数字全 C、框架照用)。⚠️ The basis rulings (read first): ① 43.3 years covers all migrant workers (2025 official survey); no official per-industry age exists — the media’s «site workers near 50» is not adopted; ② the ¥900M robot market is a research-house basis (C), not the same series as the older ¥450M forecast; ③ 7×/9× are Shenzhen pilot internal-test figures; ④ «IT spend at 0.12–0.13% of output» is a spend ratio (single source, C), not penetration — it coexists with «64% of firms have tried AI» (C): little money, wide surface, scattered dots; ⑤ vendor efficiency figures (16h→4h, 60-minute bids, −70% review cycles, 200B parameters, 52→35-day acceptance) are D-grade throughout, ideal-condition numbers; ⑥ node penetration %s are editorial estimates converted from four sources’ qualitative tiers, not statistics; ⑦ one source truncates mid-text losing its reference list (its unique figures capped at C); another wrote offline (figures C, frameworks kept).

中心悖论 · 原型性经济学THE CENTRAL PARADOX · PROTOTYPE ECONOMICS

不是技术不够,是账算不过来Not that the technology falls short — the arithmetic does

制造业机器人成功的全部前提——结构化环境、重复动作、百万件摊销——在工地一条都不成立:露天泥水粉尘、多工种交叉、「前一小时的平地,下一小时堆满钢筋」;每栋楼是一次性工程,机器人的编程、调试、标定成本摊不到任何批量上,砌筑机器人的现场调试成本甚至高过熟练砖瓦工。于是 AI 选择了绕行:先吃不需要进工地的部分(图纸),再借道工厂(装配式),最后才与泥土正面交锋——四份独立研究对这条进攻路线完全一致。Every premise of the factory robot’s success — structured surroundings, repeated motions, million-unit amortisation — fails on site, all of them: open-air mud and dust, trades crossing paths, «this hour’s level ground buried in rebar the next»; each building is a one-off, so the robot’s programming, commissioning and calibration amortise across nothing — a bricklaying robot’s on-site set-up can cost more than the mason it replaces. So AI took the detour: eat what never enters the site first (the paper), borrow the factory next (prefabrication), and face the mud last — four independent studies agree on this line of attack without exception.

图纸侧 · 以月为单位The paper side · by the month

算量、标书、审图、方案——读图+按规则计算的白领工序第一波被吃:深圳 AI 审图 7×/9×(B),AI 标书小时级(D),造价初级岗结构性收缩(四源共识)。Take-offs, bids, plan review, schemes — white-collar work that reads drawings and applies rules is eaten first: Shenzhen review at 7×/9× (B), hour-scale AI bids (D), junior estimating structurally shrinking (all four concur).

工地侧 · 以十年为单位The mud side · by the decade

砌筑、绑扎、异形装修攻不动;机器人市场 ≈9 亿 vs 产值 31 万亿(C);已商用的是点状工序:地坪、喷涂、测量、巡检——AI 让设计以秒为单位,混凝土仍要养护 28 天。Bricklaying, rebar-tying and irregular finishing hold; the robot market runs ~¥900M against ¥31T of output (C); what’s commercial is scattered trades — floors, spraying, survey, patrol. AI designs by the second; concrete still cures for 28 days.

组织与钱 · 基本不动Organisation & money · barely moves

垫资—低价中标—签证索赔—回款的赌局,分包金字塔的「人管人」,验收与定标的关系——行业真正的操作系统,AI 至今只能在外围留痕。The wager of advances, lowball wins, claims and receivables; the pyramid’s man-managing-men; the relationships behind acceptance and awards — the industry’s true operating system, where AI still only keeps records at the edge.

四层渗透模型 · 越往下,AI 越进不去(十年尺度)THE FOUR-LAYER MODEL · THE DEEPER, THE LESS AI ENTERS (DECADE SCALE)

软件 / 知识层Software / knowledge

造价算量、标书、审图、方案生成、BIM——深度渗透进行中:读图与按规则计算的部分正在整段交给机器。Estimating, bids, plan review, scheme generation, BIM — deep penetration under way: whatever reads drawings and applies rules is being handed over in whole sections.

感知 / 管控层Sensing / control

AI 摄像头、实名制考勤、无人机进度、实测实量、塔机监控——正在标配化:监管强推,替代的是「人盯人、人拍照、人填表」。AI cameras, real-name attendance, drone progress, laser QA, crane telemetry — becoming standard kit: pushed by regulators, replacing the man who watches, photographs and fills forms.

物理执行层Physical execution

砌筑、绑扎、抹灰、机电安装——人为主、机器人点状:盾构与地坪是例外,普遍工况仍属于手感与随机应变。Bricklaying, tying, plastering, MEP installation — human-led, robots in dots: tunnelling and floor works are the exceptions; the general case still belongs to touch and improvisation.

组织 / 资金 / 责任层Organisation / capital / liability

分包金字塔、垫资赌局、注册签字、验收关系——基本不动:这一层的问题从来不是信息问题,是权力与责任结构。The subcontracting pyramid, the advance-money wager, the registered signature, acceptance politics — barely moves: never an information problem, but a structure of power and liability.

读图法一句话:分层渗透、内核免疫——蓝色越深的层塌得越快,最下面一层十年尺度「基本不动」;主脊十二节点按此着色。The reading in one line: layered penetration, immune core — the bluer the layer, the faster it falls, and the bottom layer «barely moves» on a decade scale; the twelve-node spine is coloured accordingly.

Reading the MapReading the Map

从这张图带走的五条规律Five patterns to take away

立场声明:本页是批判性、祛魅的行业结构分析——四份深度研究交叉整理(一份中途截断且引文列表丢失、一份自认离线撰写,其数字全部降级;框架共识部分源于委托预设论纲,已如实声明),厂商效率数字一律 D 级,渗透%为评估值。本页提供行业结构信息,不构成工程、采购、就业或投资建议;对一线工人不猎奇不俯视——被吃掉的是工序,不是人的价值。核心判断一句话:机器人征服了工厂,却没征服工地——AI 把账照亮了,但垫资、签字与回款的赌桌,还是原来那张。 Stance: a critical, demystifying structural analysis — cross-compiled from four deep-research reports (one truncated with its reference list lost, one self-declared offline, all their figures downgraded; where the four converge partly echoes the commissioning brief, declared honestly). Vendor efficiency figures are D-grade throughout; penetration %s are estimates. Industry-structure information only — not engineering, procurement, career or investment advice; no voyeurism toward site workers — what gets eaten is the trade, never the person’s worth. The core judgment in one line: robots conquered the factory, not the site — AI lit up the ledger, but the table where advances, signatures and receivables are wagered is still the same table.