# The AI economy: Interconnected forces, feedback loops, and speeds of change # AI 经济学：相互关联的力量、反馈循环与变革速度 > **McKinsey Global Institute (MGI) Discussion Paper** > **麦肯锡全球研究院（MGI）研讨报告** > **Date / 发布日期**: September 10, 2026 / 2026年9月10日 > **Authors / 作者**: Shubham Singhal, Arvind Govindarajan, Tanguy Catlin, María Jesús Ramírez, Xavier Rigby > **Original URL / 原文链接**: [McKinsey Global Institute](https://www.mckinsey.com/mgi/our-research/The-AI-economy-Interconnected-forces-feedback-loops-and-speeds-of-change) --- ### Executive Summary / 核心导读 **AI is shaped by forces far beyond the technology itself, and its effects reach just as widely. Mapping how the pieces interact helps us see what could influence its course and where hurdles and opportunities may emerge.** 人工智能（AI）的发展受到技术本身之外诸多力量的深刻塑造，其产生的影响同样极其深远。系统性地绘制出各个要素之间如何相互作用的图谱，有助于我们洞察究竟是什么在决定其发展轨迹，并提前预判潜在瓶颈与机遇所在。 --- ## At a glance / 内容速览 * **AI’s advance will leave few of us unaffected.** It has been less than four years since the launch of ChatGPT (in late 2022), yet generative AI is already changing how people think and work—and where it goes from here is deeply uncertain. Will it improve our lives or take our jobs? Lead to unprecedented growth or upheaval? Who will benefit, and who might lose? **AI 的演进将影响到几乎每一个人。** 距离 2022 年底 ChatGPT 的问世尚不足四年，生成式人工智能（Generative AI）已在深刻重塑人们的思考与工作方式——而它未来的走向仍充满巨大的不确定性：它究竟会改善我们的生活还是取代我们的工作？会带来前所未有的经济增长还是引发社会动荡？谁将从中获益，谁又可能成为受损者？ * **Answering big questions about AI means seeing the whole picture, not just focusing on the technology.** To help do that, we mapped the connections between AI and physical resources, institutions, businesses, markets, geopolitics, and people—all of which shape AI and change in response to it. **解答关于 AI 的重大命题，需要具备全局视野，而非仅仅局限于技术本身。** 为此，我们系统梳理了 AI 与物理资源、制度框架、企业生态、资本市场、地缘政治以及个体人群之间的动态关联——这些要素既在共同塑造 AI 的发展路径，也在因 AI 的影响而发生深刻变革。 * **Risks and bottlenecks arise because different parts of the AI system are moving at vastly different paces.** AI capabilities are advancing exponentially by some measures. Progress in building data centers, power grids, chip factories, and mines is more linear. Redesigning how a business and its people work spans years. As one constraint eases, others can be exposed. **系统各组成部分发展速度的严重不同步，是导致风险与瓶颈产生的根源。** 从某些衡量标准来看，AI 算法与模型能力正呈指数级飞跃；而数据中心、电网基础设施、半导体晶圆厂以及关键矿产开采等物理建设，却遵循着相对线性的扩张周期；企业组织架构与工作流程的彻底重构更需要历时数年。随着某一处短板的缓解，新的瓶颈往往会随之暴露。 * **In the AI era, leaders must therefore learn to manage at multiple speeds.** Progress may slow when infrastructure or organizations cannot keep up, while risks can grow when adoption outruns safeguards, governance, or public acceptance. Sometimes the answer is to accelerate what is lagging; other times it may be to slow down. A systems view can help us navigate these choices. **在 AI 时代，决策者必须掌握应对“多速发展”的管理艺术。** 当底层基础设施或组织适配能力滞后时，整体演进速度会被拖慢；而当技术普及应用的速度超越安全防御、治理机制或公众接受度时，系统性风险便会急剧累积。应对之道有时在于为滞后环节提速，有时则需要适度控制推进节奏。系统性视角正是指引我们做出明智抉择的核心工具。 --- ## Introduction: Why a Systems View is Needed / 引言：为何需要系统性视角 **AI may be the most consequential technology of our time.** But how far will it spread, how quickly, and with what effects? The answers matter to almost everyone. Workers want to know whether predictions of mass job losses are credible or a scare story; one widely reported, admittedly extreme, forecast says AI could eliminate as many as half of entry-level white-collar jobs within five years [1]. Investors want to know whether the enormous sums being spent on infrastructure will pay off—or leave behind the excess capacity and losses seen after earlier railway and dot-com booms [2]. Policymakers ask how tightly to regulate AI’s risks without putting their economies at a disadvantage; China, the European Union, and the United States are following different paths [3]. For CEOs and CTOs, the question is more immediate: We have rolled out the technology, so why aren’t we seeing it in our results? **AI 可能是我们这个时代最具决定性影响的技术。** 但它将普及到何种程度、推进速度有多快、又会产生怎样的连锁反应？这些答案关乎千家万户。职场人急切想知道大规模失业的预言究竟是现实威胁还是夸大其词——曾有被广泛报道的极端预测指出，AI 可能在五年内消除高达一半的初级白领岗位 [1]；投资者迫切希望了解巨额的基础设施资本开支（CapEx）能否产生合理回报，还是会重演历史上铁路泡沫或互联网泡沫时期产能过剩与巨亏的覆辙 [2]；政策制定者则在权衡如何在严控 AI 风险的同时不削弱本国产业竞争力，目前中美欧三地已走上截然不同的监管路径 [3]；而对企业 CEO 与 CTO 而言，挑战更为现实：企业已经全面上线了 AI 工具，为何财务报表与业务成果上却迟迟不见成效？ None of these questions can be answered in isolation. They require analysis that cuts across disciplines and domains, just as the technology does. Whether jobs disappear depends on whether AI can do the tasks that make up a job as well and as cheaply as a person can, and whether new tasks emerge fast enough to replace the ones it takes over. Investment will pay off only if demand for AI services grows fast enough to fill the computing capacity being built for it—the chips, power, and data centers. Regulators face a trade-off: Move too slowly and risks may go unmanaged; move too quickly and lighter-touch rivals may pull ahead. And whether a company’s rollout produces results depends not only on the technology but also on its economics and, importantly, on whether the organization adapts to capture its full value. 上述任何一个问题都无法孤立解答，它们需要跨学科、跨领域的系统性剖析。岗位是否消失，取决于 AI 执行特定任务的综合成本与效果能否超越人类，以及伴随技术进步能否足够快地孕育出全新工作岗位；基础设施投资能否收回成本，取决于终端口径对 AI 算力服务的实际需求能否填满正在大举兴建的芯片算力、电力负荷与数据中心；监管机构面临着速度与安全的权衡：行动过慢则风险失控，干预过严则可能被采取轻度监管策略的竞争对手反超；而企业部署 AI 能否转化为真实收益，不仅取决于算法优劣，更取决于整体经济效益考量以及组织流程是否完成了深度适配与价值捕获。 To understand how these forces shape one another, we need to see the whole system and the connections among its parts. Drawing on our prior research and work with clients, we propose a general framework that connects factors often considered separately and reveals relationships that are otherwise easy to miss. 要透彻理解这些力量如何相互塑造，我们必须俯瞰整个系统及其内在关联。基于麦肯锡全球研究院先前的研究成果与企业咨询实践，我们提出了一个全景分析框架，将以往常被割裂对待的各个因素有机串联，揭示那些容易被忽视的反馈机制。 --- ## Exhibit 1: The AI Transversal Framework / 图表 1：AI 跨领域系统全景图谱 ![Exhibit 1: The AI System Framework](https://www.mckinsey.com/~/media/mckinsey/mckinsey%20global%20institute/our%20research/the%20ai%20economy%20interconnected%20forces%20feedback%20loops%20and%20speeds%20of%20change/mgi-the%20ai%20economy-1206972656-hero-1536x864.jpg?cq=50&mw=1536&car=16:9&cpy=Center) > **图表架构解析（System Architecture Overview）**： > 整个 AI 经济系统由三大逻辑层次构成： > 1. **底层基石（Foundations / Enablers & Constraints）**：位于架构核心，包括 **AI Capabilities（AI 能力）**、**R&D（研发）**、**Physical Resources（物理资源）** 以及 **Institutional Stance（制度立场）** 四大要素； > 2. **传导中枢（Transmission Layer）**：即 **Adoption and Adaptation（应用普及与组织适配）**，将底层潜能转化为实际生产力； > 3. **终端成果（Outcomes Layer）**：分为 **Economic Outcomes（经济成果）**（含生产率增长、创新、新需求、就业、解放的时间、资本转移）与 **Broader Outcomes（更广泛影响）**（含商业格局重塑、地缘政治博弈、公众福祉与社会心理）； > 4. **贯穿要素（Cross-cutting Elements）**：**Risks across the system（贯穿全系统的潜在风险）** 与 **Feedback loops（双向反馈循环机制）**。 --- ### Core Components & Key Metrics / 核心要素详析与关键指标数据 | 模块类别 (Category) | 核心要素 (Element) | 核心内涵与动态特征 (Description & Dynamics) | 关键数据 / 观察指标 (Key Metrics & Signals to Watch) | | :--- | :--- | :--- | :--- | | **Foundations
底层基石** | **AI capabilities
AI 能力** | **演进轨迹：指数级扩张。** 从 LLM 大语言模型向具身智能（Embodied AI/机器人）与智能体（AI Agents）延伸，长程任务执行与多模态复杂推理能力持续攀升。 | • **16+ 小时**：前沿模型在人类需耗时 16 小时以上才能完成的任务中达到 50% 成功率；
• GPQA Diamond 科学基准得分在 2026 年突破 95%（2023年仅为39%）。 | | **Foundations
底层基石** | **R&D
研发投入** | **演进轨迹：渐进累积与阶段性突破并存。** 涵盖模型架构、算法优化、高效计算、新型制冷及能源材料等跨学科基础研究。 | • **2900 亿美元**：2025 年全球民营部门 AI 投资规模，接近 2019 年水平的 5 倍；
• 关注顶级训练集群算力规模、专利授权量及开源社区活跃度。 | | **Foundations
底层基石** | **Physical resources
物理资源** | **演进轨迹：线性扩张。** 受制于物理世界建造与开采周期（算力芯片、电网并网、工业冷却水、用地、铜矿等关键矿产）。 | • **15 GW vs 18 GW**：2025 年全球数据中心新增投产容量 15 吉瓦，落后于约 18 吉瓦的新增需求；
• 关注电网互连排队时长（中位数达5年）与半导体交付周期。 | | **Foundations
底层基石** | **Institutional stance
制度立场** | **演进轨迹：缓慢且非线性演变。** 涉及法律法规、责任认定标准、行业规范、企业数据合规策略及社会公众信任度。 | • **47%**：不足半数受访公众信任使用 AI 的企业能够妥善保护其个人隐私数据；
• 关注各地人工智能立法进展与监管沙盒机制。 | | **Transmission
传导中枢** | **Adoption & Adaptation
应用普及与组织适配** | **个人应用极速渗透，企业级深度整合相对滞后。** 组织需要打破业务壁垒、重塑业务流程并克服内部惯性与变革阻力。 | • **89% vs 46%**：2026 年有 89% 的企业在至少一个业务部门使用 AI，但其中实现规模化落地（跨越概念验证阶段）的仅占 46%；
• 81% 的美国执业医师在 2026 年已将 AI 融入日常诊疗。 | | **Economic Outcomes
经济成果** | **Productivity growth
生产率增长** | **任务级效率提升向宏观生产率转化的“时滞效应”。** 单点任务提效显著，但宏观层面的飞跃取决于端到端业务流的重构。 | • **2.5%**：2022 年底至 2026 年上半年美国非农商业部门劳动生产率年均增速（略高于 1947 年以来的长期均值 2.1%）。 | | **Economic Outcomes
经济成果** | **Innovation
科技与商业创新** | **通用目的技术（GPT）催化跨学科突破。** 大幅压缩药物研发分子筛选、新材料合成周期，催生前所未见的新业态与新产品。 | • **缩短高达 80%**：AI 将新药早期发现周期缩短多达 80%，药企在相同研发预算下可将产出药物种类提升至 1.5 倍。 | | **Economic Outcomes
经济成果** | **New demand
催生新需求** | **价格效应与杰文斯悖论（Jevons Paradox）。** 生产成本降低与人机交互升级激发潜在消费需求，开拓个人助理与服务外包新蓝海。 | • **30 万亿美元**：互联网这一前代通用技术在 2024 年撬动的全球电子商务年交易规模，AI 有望创造类似量级的新市场空间。 | | **Economic Outcomes
经济成果** | **Jobs
就业生态重塑** | **岗位任务构成解构重组，技能溢价迁移。** 岗位完全被替代属于少数，更多是工作内容的混合重构；人际沟通、批判性思维等技能价值凸显。 | • **72%**：到 2035 年，预计 72% 的美国劳动者将有超过 15% 的工作时长被 AI 深度重塑；
• 技能中低端岗位可借助 AI 实现经验弥补。 | | **Economic Outcomes
经济成果** | **Time freed
解放劳动时间** | **时间红利的再分配。** 节约的工时可能流向高价值业务、持续技能培训、休闲生活，亦可能导致非自愿工时减少。 | • **30%**：近三分之一使用 AI 的职场人在受访前一周表示 AI 累计为其节省了 3 小时以上工作时间。 | | **Economic Outcomes
经济成果** | **Capital shifts
资本重分配** | **资本密集度攀升与全产业链投资转移。** 涵盖自研芯片、超大型算力中心、独立发电装置等庞大资产负债表扩张。 | • **7560 亿美元**：2025 年全球七大科技巨头在 AI 领域的资本开支与研发总投入，较 2022 年翻倍；
• 警惕循环融资（Circular Financing）与过度建设风险。 | | **Broader Outcomes
更广泛影响** | **Business landscape
商业竞争格局** | **护城河转移。** 通用大模型迅速商品化（Commoditization），核心竞争力向私有专有数据、客户粘性与行业控制节点转移。 | • **1/3**：约三分之一的企业已开始利用智能体编程工具（Agentic Coding）自主构建内部软件以替代向外采购。 | | **Broader Outcomes
更广泛影响** | **Geopolitics
地缘政治博弈** | **科技制高点与供应链武器化。** 芯片禁运、稀土矿物管制、自主主权 AI（Sovereign AI）成为大国战略博弈焦点。 | • **55% vs 23%**：2022 年以来全球代表性 AI 模型中，55% 来自美国，23% 来自中国；
• 算力基础设施与数字主权日益挂钩。 | | **Broader Outcomes
更广泛影响** | **People & Society
社会心理与福祉** | **数字鸿沟、认知负荷与公众预期分化。** 提供普惠专业支持的同时，可能加剧信息过载与职业倦怠；新兴市场乐观度显著高于欧美成熟经济体。 | • **73%**：个人生活类咨询在消费级 ChatGPT 交互中的占比从 2024 年年中的 50% 攀升至 2025 年的 73%；
• 欧美年轻人对 AI 潜在负面效应担忧较重。 | --- ## Mismatched timelines: Bottlenecks, risks, and opportunities / 时间线错配：瓶颈、风险与机遇 General-purpose technologies can transform economies and improve lives, but realizing their full potential takes more than the invention itself. It requires complementary innovation, new infrastructure, new rules, capital, and people and organizations willing to change how they work. The pieces rarely arrive all at once, and the gap between a new capability and its full payoff has historically taken decades to close. 通用目的技术（General-Purpose Technologies, GPTs，如蒸汽机、电力、计算机与互联网）具备彻底变革经济形态与提升人类生活福祉的潜能，但要将其潜在价值完全变现，绝非仅仅依靠技术发明本身。它需要配套的互补性创新、新型基础设施的铺设、制度规则的设立、海量资本的倾注，以及全社会组织与个人工作模式的根本转变。这些拼图极少能在同一时间完全就位；历史上，一项突破性技术从诞生到迎来其经济效益的真正巅峰，往往需要历经数十年漫长的磨合周期。 --- ### Exhibit 2: Historical Diffusion of General-Purpose Technologies / 图表 2：通用目的技术历史扩散周期对比 ![Exhibit 2: GPT Historical Timeline](https://www.mckinsey.com/~/media/mckinsey/mckinsey%20global%20institute/our%20research/the%20ai%20economy%20interconnected%20forces%20feedback%20loops%20and%20speeds%20of%20change/mgi-ai-transversal-impact-ex2-v2.svgz?cq=50&cpy=Center) ``` [历史演进脉络对比 / Historical Diffusion Comparison] 1770s: Steam Power (蒸汽机) ├── 核心突破: 机械动力输出 (Mechanical Power) ├── 主要瓶颈: 动力传输损耗大、需集中建厂 └── 峰值贡献: 耗时约 100 年达到人均 GDP 增长贡献率 30% 1880s: Electricity (电力革命) ├── 核心突破: 分布式动力传输与连续供电 (Distributed Power) ├── 主要瓶颈: 发电机与电网铺设滞后、工厂产线布局重构周期长 └── 峰值贡献: 耗时约 40 年达到人均 GDP 增长贡献率 10% 1970s-1990s: Computers & Internet (计算机与互联网) ├── 核心突破: 信息数字化处理与全球实时互联 (Digital Information) ├── 主要瓶颈: 光纤通信基建投入巨大、商业模式与组织数字化转型滞后 └── 峰值贡献: 两者协同耗时约 25 年达到人均 GDP 增长贡献率 50% 2020s: Artificial Intelligence (人工智能) ├── 核心突破: 认知推理、模式识别与内容自动生成 (Cognitive & Autonomous Systems) ├── 当前制约: 算力芯片制造、电力与电网并网、业务流程重构、数据合规与社会信任 └── 预期周期: 经济影响正以远快于历史前例的速度压缩显现 ``` In the early decades of electrification, it was difficult simply to generate enough power to run factories and industrial machinery. Later, as that hurdle eased, the limits of the power grid became apparent. The big productivity gains arrived only once factories were physically redesigned around the electric motor, a step that came later still [4]. The internet’s path was much the same: First fiber-optic networks had to be built, then new business models were needed, and finally companies had to reorganize around the technology [5]. 在电气化发展的最初几十年里，仅仅是生产出足以驱动工矿机械的充足电能就已困难重重；随着发电技术的攻克，输电网承载力的极限又成为了新的掣肘；而真正的生产率飞跃，直到工厂彻底摒弃原有的蒸汽集中传动轴设计、全面围绕独立电动机完成车间流水线空间重构之后，才最终得以实现 [4]。互联网的发展轨迹亦如出一辙：行业必须先期铺设昂贵的光纤宽带骨干网络，随后探索全新的商业变现逻辑，最终倒逼各类企业组织围绕数字化架构展开业务流程的彻底重构 [5]。 --- ### Mismatches / 发展速度的失衡与错配 AI faces a similar challenge, with many necessary parts coming together on vastly different timelines and with different shapes of change. Model capabilities are improving exponentially, doubling every few months by some recent measures [6]. Physical infrastructure builds out more linearly, often taking years. Change at companies may occur more unevenly and can take longer, as businesses develop new products and ways of working. 当前 AI 发展正遭遇相似的多维度挑战，各关键支撑要素正以截然不同的时间尺度和增长形态向前演进：模型与算法能力展现出指数级跃迁，部分前沿基准下的任务处理能力每隔数月即可翻番 [6]；然而物理基础设施的建设完全遵循物理规律与工程建造的线性周期，动辄历时数年；企业内部的组织变革则更加非线性且漫长，涉及新商业模式的探索与工作习惯的重塑。 --- ### Exhibit 3: Foundations Moving at Different Speeds / 图表 3：AI 基础层要素发展速度的显著分化 ![Exhibit 3: AI Foundations Speed Divergence](https://www.mckinsey.com/~/media/mckinsey/mckinsey%20global%20institute/our%20research/the%20ai%20economy%20interconnected%20forces%20feedback%20loops%20and%20speeds%20of%20change/mgi-ai-transversal-impact-ex3.svgz?cq=50&cpy=Center) > **四项基础要素的速率分化特征：** > 1. **AI Capabilities（AI 能力）**：呈 **指数级（Exponential）** 上升。如 METR 基准测试显示，前沿模型（包括 Anthropic Claude Mythos Preview 与 OpenAI GPT-5.4）能够以 50% 可靠性自主完成的任务时长已突破 15-16 小时； > 2. **R&D（研发投资）**：呈 **稳步爬升（Steady Growth）** 态势，全球民营资本对 AI 研发的投入维持高位； > 3. **Physical Resources（物理基建）**：呈 **线性缓行（Linear & Gradual）**，数据中心总电力装机容量难以跟上激增的算力部署需求； > 4. **Institutional Stance（制度与社会信任）**：呈 **平台期波动（Flat / Inertial）**，公众对科技公司保护个人隐私的信任度长年停留在 50% 以下。 --- ### Bottlenecks: Hardware, Power, and Infrastructure / 瓶颈演变：硬件交付、电力并网与数据中心建设 By 2023, only a year after OpenAI launched ChatGPT, demand for AI chips had outgrown the capacity to package them with the memory they require. Nvidia, which designs most of the processors that frontier AI runs on, had built exactly the chip for the job, the H100. The problem was fusing it with memory in a single package, so that the hardware could work fast enough. Only one company could do this at the scale AI now demanded: TSMC, the Taiwanese firm that manufactures chips for most of the world’s major technology companies, including Nvidia and Apple. But in 2023, demand for that packaging outpaced even TSMC’s ability to keep up [8]. Buyers reportedly faced lead times of up to 11 months for a finished H100-based server [9]. The fix was more capacity, and TSMC built faster than it ever had [10]. 早在 2023 年（即 ChatGPT 发布仅一年后），AI 计算芯片的井喷式需求就迅速击穿了高端封装产能的上限。设计了绝大部分前沿 AI 算力芯片的英伟达（Nvidia）推出了划时代的 H100 GPU，然而核心瓶颈在于如何利用先进封装工艺（CoWoS）将逻辑芯片与高带宽内存（HBM）严密整合，以实现极高的数据吞吐速率。全球仅有台积电（TSMC）具备满足该量级需求的先进制程与封装能力，而在 2023 年，暴涨的订单让台积电的产能也捉襟见肘 [8]，导致市场上搭载 H100 的服务器交付周期一度长达 11 个月之久 [9]。唯一的解决路径便是大规模扩产，台积电随后启动了历史上最快的封装产线建设扩容 [10]。 Other hurdles emerged in a completely different part of the system: electricity. Building and connecting enough power capacity—a challenge for utilities and grid operators, not chipmakers—is another constraint, one familiar from the early era of electrification. A single data center can consume as much power as a small city, and when local grids cannot handle that much new demand, the utility has to build new power lines and substations before the facility can turn on—construction that requires its own permits and land, behind a long queue of other projects waiting for the same thing. 与此同时，新的瓶颈迅速在电力领域显现。为算力中心提供充足稳定的电力供应，已成为电力公用事业公司与电网运营商面临的严峻考验。一座现代超大规模 AI 数据中心的能耗堪比一座中小型城市，当区域电网无法承受瞬间激增的负荷时，必须先行审批并新建高压输电线路和变电站——而这些基建项目面临着漫长的土地审批与环保许可排期。 Several hyperscalers, the handful of companies building AI at the largest scale, have run into this wall. On an October 2025 earnings call, Microsoft’s CFO told investors that the company’s cloud computing platform, Azure, which is used to handle AI workloads, had been short on operating capacity for several quarters. Electricity was the obstacle. Microsoft had AI chips it could not deploy because it lacked powered, ready-to-occupy data center space [11]. The median new US power plant now takes about five years from interconnection request to commercial operation, up from less than two years in 2008 [12]. 全球为数不多的超大规模云服务商（Hyperscalers）纷纷遭遇这堵“电力之墙”。在 2025 年 10 月的财报电话会议上，微软（Microsoft）首席财务官坦陈，其承载 AI 工作负载的 Azure 云计算平台已连续数个季度受制于电力供应瓶颈。微软仓库中积压着大量采购的顶级算力芯片，却因缺乏通电且具备入驻条件的机房空间而无法上线运行 [11]。在美国，一座新建发电厂从提交电网互连申请到最终获批商业运营，中位数等待时长已从 2008 年的不足 2 年飙升至现今的约 5 年 [12]。 Data center construction itself is now shaping up as a bottleneck, especially in the United States, which accounts for nearly half of global installed data center capacity [13]. AI’s compute, power, and water costs per query have fallen as models get more efficient, but total demand is rising faster than efficiency gains or new supply can offset [14]. At least 75 data center projects worth some $130 billion were blocked or delayed in the United States by community opposition in the first quarter of 2026 alone [16]. In July, New York’s governor signed the first-ever statewide permitting freeze on large data centers [17]. Supplies of copper and other critical minerals needed for chips and wiring are also showing signs of stress [18]. 数据中心土建本身正在成为新的核心瓶颈，在占有全球近半数数据中心装机容量的美国尤为突出 [13]。尽管随着算法优化，单次模型推理的能耗与用水成本有所摊薄，但全社会调用总量的爆发式增长彻底抵消了单位效率的提升（即杰文斯悖论）[14]。仅在 2026 年第一季度，全美就有至少 75 个总价值约 1300 亿美元的数据中心项目因地方社区抵制电价上涨和土地占用而被迫推迟或叫停 [16]。2026 年 7 月，纽约州州长签署了全美首个针对超大型数据中心的州级建设行政暂停令 [17]。此外，电网与芯片所必需的精炼铜等关键矿产资源也面临中长期供给短缺压力 [18]。 --- ### Risks: Governance, Security, and Public Acceptance / 风险维度：治理滞后、网络安全与社会阻力 Governance is also lagging. A separate McKinsey survey of about 500 organizations on AI trust and governance, conducted in December 2025 and January 2026, found that only about 30 percent had reached level three or higher on a four-level maturity scale [23]. Strategy remains underdeveloped as well. Only 22 percent of operations leaders report a fully developed and implemented enterprise AI strategy, even though more than half name strategy as the single biggest driver of AI returns [24]. 企业治理框架与安全管控的滞后尤为令人担忧。麦肯锡在 2025 年底至 2026 年初对全球约 500 家企业开展的 AI 信任与治理调研显示，在四级成熟度评估体系中，仅有约 30% 的企业达到了三级或以上的成熟标准 [23]。只有 22% 的运营负责人表示本机构拥有成熟且落地的企业级 AI 战略规划，尽管超过半数的受访者一致认为战略清晰度是决定 AI 投资回报率（ROI）的最关键变量 [24]。 As adoption increases, so does the risk of cyberattack. In April 2026, the UK AI Security Institute (AISI) found that Anthropic’s Claude Mythos Preview could autonomously execute multi-step attacks on test networks that would take human professionals days of work. Just two years earlier, the best available models in AISI’s evaluations could barely complete beginner-level cyber tasks [28]. 随着 AI 技术的普及，网络攻击的威胁面成倍放大。2026 年 4 月，英国人工智能安全研究所（AISI）的测评显示，前沿模型（如 Anthropic 的 Claude Mythos Preview）已能够自主执行高度复杂、多步骤的网络渗透攻击，而在两年前，前沿模型甚至难以完成初级安全任务 [28]。这一能力跃升意味着一旦发生重大安全事故，可能迅速引发监管收紧与公众信任崩塌。 --- ## AI as a system: How the pieces fit together / AI 作为一个系统：各要素如何协同运转 ``` ┌────────────────────────────────────────────────────────────────────────┐ │ THE AI SYSTEM / AI 系统全景 │ ├────────────────────────────────────────────────────────────────────────┤ │ │ │ [ FOUNDATIONS / 底层基石 ] │ │ ┌───────────────┐ ┌───────────────┐ ┌───────────────┐ ┌────────────┐ │ │ │ AI CAPABILITY │ │ R&D │ │ PHYSICAL │ │INSTITUTION │ │ │ │ AI 算法能力 │ │ 研发与创新 │ │ 算力/电力/矿产 │ │ 法律/政策 │ │ │ └───────┬───────┘ └───────┬───────┘ └───────┬───────┘ └─────┬──────┘ │ │ └─────────────────┼─────────────────┘ │ │ │ ▼ │ │ │ [ TRANSMISSION / 传导中枢 ] │ │ │ ┌────────────────────────────────┐ │ │ │ │ ADOPTION & ADAPTATION │ │ │ │ │ 企业与个人应用普及 / 组织流程重构 │◄───────────────┘ │ │ └───────────────┬────────────────┘ │ │ │ │ │ ▼ │ │ [ OUTCOMES / 综合产出成果 ] │ │ ┌───────────────────────────────────┐ ┌────────────────────────────┐ │ │ │ ECONOMIC OUTCOMES (经济成果) │ │ BROADER OUTCOMES (更广泛影响)│ │ │ │ • Productivity Growth (生产率) │ │ • Business Landscape │ │ │ │ • Innovation (科学/商业创新) │ │ (商业格局与护城河变迁) │ │ │ │ • New Demand (激发新需求) │ │ • Geopolitics │ │ │ │ • Jobs (就业与技能重构) │ │ (大国竞争与主权AI) │ │ │ │ • Time Freed (工时释放与分配) │ │ • People & Well-Being │ │ │ │ • Capital Shifts (资本转移) │ │ (社会心理与认知健康) │ │ │ └─────────────────┬─────────────────┘ └──────────────┬─────────────┘ │ │ │ │ │ │ └─────────────────┬────────────────┘ │ │ ▼ │ │ [ FEEDBACK LOOPS / 动态反馈循环 ] │ │ • 经济收益反哺硬件与模型研发 ◄─────────┴─────────► • 社会舆论与合规反向约束部署 │ │ │ └────────────────────────────────────────────────────────────────────────┘ ``` --- ### Sidebar: Why a Systems View Matters / 专栏：为什么系统性视角至关重要 **Systems thinking is probably as old as philosophy itself, but it was applied directly to business and technology in the 20th century by people studying how organizations actually behave.** **系统性思维（Systems Thinking）的历史或许与哲学一样悠久，但直到 20 世纪，它才被系统应用于商业管理与技术组织行为学研究中。** * **Ludwig von Bertalanffy (路德维希·冯·贝塔朗菲)**: Biologist Ludwig von Bertalanffy spent decades arguing that breaking a phenomenon into isolated parts misses essential properties that surface only when the parts are observed together, an idea he finally consolidated in his 1968 book *General System Theory* [1]. 理论生物学家贝塔朗菲耗费数十年阐明一个核心观点：将复杂现象机械地拆解为孤立片段，必然会遗漏那些只有在各要素协同交互时才会显现的“整体性涌现特征”，他在 1968 年的经典著作《一般系统论》中奠定了这一理论基石 [1]。 * **Jay Forrester (贾伊·福雷斯特)**: The engineer Jay Forrester, in the 1950s, studied why factories lurched between hiring booms and mass layoffs even when customer demand barely moved [2]. The cause was not demand; it was information delay... The swings were caused by the system itself, not the macroeconomy. 麻省理工学院系统动力学奠基人福雷斯特在 20 世纪 50 年代研究发现，即使终端消费需求保持平稳，制造工厂依然会陷入“招工狂潮与大规模裁员”的剧烈震荡中 [2]。这一现象的根本原因并非宏观经济波动，而是产业链信息传递的严重时滞（牛鞭效应）。这种震荡是由系统内在结构缺陷所驱动的。 * **Russell Ackoff (罗素·艾考夫)**: The management theorist Russell Ackoff illustrated the idea with a story about cars [3]. Take the world’s best automotive engineers, he said, and ask them to identify the single best engine, the best transmission, the best brakes, and so on, from every car in production. Now put those “best” parts together. The result is not the best car in the world—it may not run at all, because the parts were never designed to work together. The same logic applies to AI. Optimizing one piece in isolation—faster chips, more capital, a new rule—will not improve outcomes if the real constraint lies elsewhere. 著名管理哲学家艾考夫曾用汽车的精妙比喻阐释系统论 [3]：如果你召集全球最顶尖的汽车工程师，让他们从市面所有车型中分别挑选出最好的发动机、最好的变速箱、最好的刹车系统并组装在一起，其结果绝非世界上性能最好的汽车——它甚至根本无法发动，因为这些孤立的“最优零部件”彼此之间无法匹配协同。AI 的商业落地完全符合这一规律：如果系统的真正短板在组织流程或电网接口，那么孤立地追求算力更快的芯片、堆砌更多的资本或出台一项新规，都无法带来预期的生产力跃升。 --- ## Exhibit 4: Detailed Breakdown of the AI System / 图表 4：系统分层演进与影响链路详解 ![Exhibit 4: Deep Dive into Foundations, Transmission, and Outcomes](https://www.mckinsey.com/~/media/mckinsey/mckinsey%20global%20institute/our%20research/the%20ai%20economy%20interconnected%20forces%20feedback%20loops%20and%20speeds%20of%20change/mgi-ai-transversal-impact-ex4a-diagram-v2.svgz?cq=50&cpy=Center) ### 1. Foundations Layer / 底层基石层 * **AI Capabilities (AI 能力)**: Frontier models are transitioning into autonomous reasoning agents and embodied robotic systems. 前沿模型正加速向具备长程规划能力的自主智能体与具身机器人延伸，但从封闭实验环境向复杂的现实开放环境落地仍面临鲁棒性考验。 * **R&D (研发投入)**: High concentration of cutting-edge frontier research within the US and China. 前沿基础模型研发高度集中于中美头部科技机构，研发成果不仅直接提升智能水平，亦在不断提高推理能效比以缓解物理资源压力。 * **Physical Resources (物理资源)**: Land, compute, grid interconnections, power generation, industrial cooling, and rare earth minerals. 从半导体设备制造、高密度变压器定制，到大型机房用地、工业制冷循环水及稀土等关键原材料，物理供应链约束深刻决定了技术扩散的上限。 * **Institutional Stance (制度立场)**: Regulations, data governance, safety liability, and intellectual property. 监管规则、责任溯源机制及数据合规界限直接影响企业的采纳意愿与部署成本，是平衡创新与安全的制度闸门。 ### 2. Transmission Layer / 传导中枢层 * **Adoption and Adaptation (应用普及与组织适配)**: 个人端采纳呈现极高的自发性与爆发力；而企业级深度整合则极为缓慢。具备竞争优势的企业正在跨越“单点功能测试”，走向“端到端业务流重构”；规模较小的企业拥有敏捷性优势，但中等体量企业最易受到资源与转型的双重挤压。 ### 3. Outcomes Layer / 综合产出层 * **Productivity Growth (生产率增长)**: 克服“索洛悖论（Solow Paradox）”需要全产业链完成配套组织资产的积累。 * **Innovation & New Demand (创新与新需求)**: 在生命科学、材料计算领域压缩试错周期，重构人机交互界面，创造全新的增量市场。 * **Jobs & Time Freed (就业重塑与时间解放)**: 岗位被拆解为细分任务集，人机协同使技能偏好发生结构性位移，工作效率提升释放出大量非核心工时。 * **Capital Shifts (资本重分配)**: 全球数据中心投资在 2025–2030 年间预计累计达 7 万亿美元，资本结构正由单纯的风投向全行业债务与私募信贷扩散。 * **Business Landscape & Geopolitics (商业格局与地缘政治)**: 模型同质化促使竞争护城河向私有业务数据与客户信任转移；主权 AI（Sovereign AI）成为大国数字竞争新焦点。 * **People, Society & Well-Being (社会心理与福祉)**: 普惠专业知识获取的同时，认知过载与潜在的技术焦虑需要更健康的制度与文化引导。 --- ## The Value of a Systems View / 系统视角的实践价值 A systems view can help identify where action could have the greatest effect. For businesses, it can reveal what is holding back results; for investors, how opportunities and risks are shifting; for policymakers, where rules or public investment could help; and for individuals, which skills and opportunities may become more valuable. 系统性视角的核心价值在于，帮助不同主体精准识别在哪个环节采取行动能够产生最大的杠杆效应。 --- ### 1. For Business Leaders / 致企业决策者 * **Look beyond efficiency for lasting advantage / 超越单纯降本，构筑不可复制的竞争护城河** AI can reduce costs and make existing work more productive, but competitors with access to the same technology may achieve similar gains, driving down prices and eroding any initial advantage [29]. Long-term outperformance requires companies to combine AI with advantages that rivals cannot easily replicate: proprietary data, scale, strong customer relationships, and access to scarce physical assets [30]. 仅依靠 AI 实现单点效率提升与成本削减，无法构成长期竞争优势，因为竞争对手能够以相同成本采购同等水平的模型，最终导致全行业同质化竞争与价格侵蚀 [29]。企业必须将 AI 深度融入难以被复制的护城河资产中——包括独家私有数据资产、深厚的客户信任粘性、独特的线下服务网络以及难以替代的物理资源 [30]。 * **Build new things / 着眼于开辟增量业务** The first uses of major new technologies are rarely the most consequential. Using the technology to automate existing work is a first step, but leaders should use the resources freed up to build new sources of growth. A small number of standout companies have driven a disproportionate share of economy-wide productivity growth through shifts in strategy and moves into new businesses [31]. 重大技术的初期应用往往局限于替代旧流程，但其最大价值在于催生全新的商业形态。将现有流程自动化只是第一步，管理者应把被释放出来的资本与人才资源，投入到创造新产品与开拓新市场中 [31]。 * **Learn to learn faster / 打造敏捷进化的组织学习机制** AI capabilities improve exponentially, much faster than many companies can change their technology or workflows. Established companies must leverage their operational scale to test new approaches, gather feedback, and iterate workflows at a pace that matches evolving AI capabilities. 技术模型能力的迭代速度远快于企业软件与流程的更新周期。成熟企业必须善用自身庞大的业务场景与客户触点，建立起快速测试、收集反馈并持续调整作业流程的自适应机制。 * **Give people a reason to support change / 建立让员工可感知的收益分享机制** If people experience better products, services, or ways of working, they may be more likely to support change; if they feel that the harms outweigh the benefits—that their jobs and livelihoods are at risk—resistance may mount. 任何组织变革若要成功，必须让基层员工和终端用户切实感受到 AI 带来的赋能价值（如减少机械重复性劳动），而非仅仅产生岗位被替代的危机感与抵触情绪。 --- ### 2. For Investors / 致资本投资者 * **Follow the bottlenecks / 沿着系统瓶颈寻找超额回报机遇** Constraints create investment opportunities because they generate demand for solutions. When computing power was scarce, opportunities emerged around chipmakers and frontier models. As constraints shift toward electricity, grid gear, and organizational enablement, new investment opportunities emerge around solving those problems. 系统的核心短板往往蕴藏着最大的投资红利。随着算力瓶颈从早期的芯片封装向电力基础设施、高压输电装备、工业制冷以及企业级定制解决方案转移，投资重心亦应随之动态调整。 * **Expect different growth paths / 识别不同环节的非对称增长曲线** Exponential growth in AI model performance will not translate into exponential growth across every business in the supply chain. Physical infrastructure follows linear construction cycles, and corporate ROI realization takes years. Investors must avoid applying exponential valuation multiples to linearly expanding assets. 算法能力的指数级跃升绝不等于整条产业链各环节都能实现同等斜率的增长。投资者必须清醒区分物理基建的线性回报特征与软件生态的规模效应，避免资产错配。 * **Look for correlated risks / 警惕高度捆绑的系统性关联风险** Investments in chips, data centers, power utilities, and AI software may appear diversified across different sectors, but they are all tightly coupled to the same end-user AI demand trajectory. A sudden deceleration in enterprise spending could simultaneously impact multiple seemingly unrelated asset classes. 虽然芯片制造、数据中心房地产、电力公用事业与 AI 软件看似分布在不同行业，但在底层逻辑上高度绑定于终端真实算力需求的健康度。投资者需要警惕“假性多元化”，审慎评估产业链上下游的联动风险暴露。 --- ### 3. For Policymakers / 致政策制定者 * **Address system constraints / 针对性疏通本地关键要素堵点** Policymakers should identify the constraints that matter most in their specific jurisdictions—whether it is power permitting, grid modernization, digital infrastructure, or technical workforce training—and focus policy interventions on relieving these specific bottlenecks. 各国各地区的要素禀赋与短板各异。监管与产业政策的着力点应当精准聚焦于本地最具约束性的环节——例如加快新型电网互连审批、升级骨干输配电网络、建设主权算力设施或培育跨学科工程技术人才。 * **Make policy responsive / 构建敏捷响应的动态监管框架** AI capabilities evolve faster than traditional legislative cycles. Rules need to be adjusted based on empirical evidence rather than attempting to settle every hypothetical risk in advance. Encouraging industry self-regulatory mechanisms while closely monitoring system-wide impacts enables more balanced governance [32]. 鉴于前沿模型演进的高动态性，传统静态立法模式难以适应。政策制定者应推行包容审慎的沙盒监管与敏捷治理机制，在防范系统性安全风险的同时，避免扼杀初创企业的创新动能 [32]。 * **Foster strategic partnerships / 务实推进主权能力与国际合作** Governments should strategically determine which critical capabilities must be secured domestically and where reliable international partnerships are sufficient, avoiding unnecessary resource misallocation (such as the UK Sovereign AI Fund and EU InvestAI initiative [33]). 明晰哪些核心技术与数据资产必须依托本国自主可控体系（主权 AI），哪些可以通过多元化国际贸易与盟友合作获取，兼顾国家安全与资金运作效率 [33]。 --- ### 4. For Individuals / 致个体从业者 * **Cultivate AI fluency / 培养复合型“AI 驾驭力（AI Fluency）”** AI fluency means understanding how to use the tools effectively and knowing when to rely on human judgment. Skills such as critical thinking, creative problem-solving, empathy, and adaptability become increasingly valuable as AI takes over routine analytical tasks. 掌握 AI 工具的交互与协同已成为职场基本素养，但更核心的是培养“何时采纳 AI、何时依靠人类判断”的元认知能力。批判性思维、同理心沟通、复杂情境决策与跨界整合能力，将构成人类劳动者的长久竞争优势。 * **Expand your personal ambitions / 借助智能杠杆放大个人创造力边界** AI provides individuals with access to capabilities that once required specialized technical teams or significant financial resources. This fundamentally expands what a single person can realistically achieve—whether learning a new discipline, launching an entrepreneurial venture, or creating novel intellectual property. AI 正在大幅降低个体创新的门槛。以往需要数十人工程团队或巨额预算才能完成的原型开发、市场分析与多语言运营，如今单个超级个体借助 AI 智能体即可高效推进。 --- ## Conclusion / 结语 **History suggests that the economic impact of new technologies may be arriving faster with each successive wave.** Steam power took about a century from commercialization to its peak impact on economic growth, electricity about 40 years, and computers and the internet about 25. It is too early to know whether AI will compress that timeline further. But the possibility adds urgency because businesses, workers, and governments may have less time to adjust. **历史规律表明，每一次全新通用目的技术的经济渗透周期都在显著缩短。** 蒸汽机耗费了约一个世纪才迎来其经济增量峰值，电力用了约 40 年，而计算机与互联网将这一过程缩短至 25 年左右。尽管目前尚无法断言 AI 是否会进一步压缩这一历史时钟，但这种可能性无疑为各界敲响了警钟——留给企业重塑战略、劳动者重构技能以及政府完善治理体系的时间窗口正在大幅收窄。 AI’s trajectory will not be determined by the technology alone. Models will improve, costs will fall, and new uses will emerge. But what follows will depend on how capabilities interact with physical resources, institutions, organizations, markets, and people—all moving at different speeds. A constraint in one part of the system can slow progress elsewhere; removing it can reveal another. And as AI spreads, the changes it produces will feed back into the system itself. AI 未来的行进路线绝不会由单一的技术参数所决定。算法会持续演进，推理成本会不断降低，新的应用场景也会层出不穷。然而最终的经济与社会图景，取决于底层能力如何与物理资源、制度环境、企业组织、金融市场以及全人类社会进行错综复杂的互动。**唯有建立起真正的全系统视野，洞悉各要素在不同速率下的错配与共振，我们方能在剧变之中把握主动，做出最具前瞻性的战略抉择。** --- ## Acknowledgments & Research Team / 课题致谢与研究团队 * **Project Leadership / 项目领导团队**: • **Shubham Singhal** (MGI Chairman, Senior Partner, Detroit) • **Arvind Govindarajan** (MGI Partner, Boston) • **Tanguy Catlin** (MGI Director, Senior Partner, Boston) • **María Jesús Ramírez** (MGI Senior Fellow, Bay Area) • **Xavier Rigby** (Consultant, Austin) * **Working Team Leads / 核心执行团队**: Camillo Lamanna, Xavier Rigby * **Editorial & Visual Production / 采编与可视化设计**: Jason Clenfield (Senior Editor), Chuck Burke (Data Visualizations) * **Institutional Note / 研究性质声明**: This discussion paper contributes to McKinsey’s ongoing independent research on AI. It was not commissioned or sponsored by any business, government, or outside institution. --- ## References & Footnotes / 引用文献与文末注释 1. Jim VandeHei and Mike Allen, “Behind the Curtain: A white-collar bloodbath,” *Axios*, May 28, 2025. (Discussion of Anthropic CEO Dario Amodei's forecast on entry-level jobs). 2. Carlota Perez, *Technological Revolutions and Financial Capital: The Dynamics of Bubbles and Golden Ages*, Edward Elgar, 2002. 3. Airlie Hilliard et al., “Artificial intelligence policy worldwide: A comparative analysis,” *Royal Society Open Science*, 2026, Volume 13, Number 2. 4. Paul A. David, “The dynamo and the computer: An historical perspective on the modern productivity paradox,” *American Economic Review*, 1990, Volume 80, Number 2. 5. Erik Brynjolfsson, Daniel Rock, and Chad Syverson, “The productivity J-curve: How intangibles complement general purpose technologies,” *American Economic Journal: Macroeconomics*, January 2021, Volume 13, Number 1. 6. Thomas Kwa et al., “Measuring AI ability to complete long software tasks,” METR, March 19, 2025; and “Time horizon 1.1,” METR, January 29, 2026. 7. *Energy and AI*, International Energy Agency (IEA), April 10, 2025. 8. C. C. Wei, TSMC Q2 2023 Earnings Conference Call, July 20, 2023. 9. Anton Shilov, “Wait times for Nvidia’s AI GPUs ease to three to four months,” *Tom’s Hardware*, February 16, 2024. 10. TSMC Advanced Backend Fab 6 Opening Release (June 2023) and Q4 2025 Earnings Call (January 2026). 11. Microsoft FY2026 Q1 Earnings Call (October 29, 2025); Data Center Dynamics (November 2025). 12. Joseph Rand et al., *Queued up: 2026 edition—Characteristics of power plants seeking transmission interconnection*, Lawrence Berkeley National Laboratory (LBNL), May 2026. 13. McKinsey Global Institute, “Colocation data centers: The infrastructure race behind AI,” June 30, 2026. 14. Ben Cottier et al., “LLM inference prices have fallen rapidly,” Epoch AI, March 2025. 15. *McKinsey Quarterly*, “The cost of compute: A $7 trillion race to scale data centers,” April 2025. 16. *Data Center Watch Report Q1 2026*, June 12, 2026. 17. Office of the Governor of New York, Press Release on Data Center Moratorium, July 14, 2026. 18. *Global Critical Minerals Outlook 2026*, International Energy Agency, July 2026. 19. Drew Robb, “Are there enough engineers for the AI boom?,” *IEEE Spectrum*, January 2026. 20. “Task-completion time horizons of frontier AI models,” METR, May 2026. 21. McKinsey Global Institute, *Agents, robots, and us* Research Series (US, Europe, Latin America editions, 2025–2026). 22. Dan Tinkoff et al., “The state of AI in 2026: On the road to ROI,” McKinsey, August 2026. 23. Gabriel Morgan Asaftei et al., “State of AI trust in 2026: Shifting to the agentic era,” McKinsey, March 2026. 24. *2026 AI Impact Survey Report*, Grant Thornton, 2026. 25. *Global AI Confessions Report: CEO Edition 2026*, Dataiku & The Harris Poll, May 2026. 26. *Global Cybersecurity Outlook 2026*, World Economic Forum & Accenture, January 2026. 27. David Kertai, *Improving State and Local Government Cybersecurity*, ITIF, April 2026. 28. UK AI Security Institute (AISI), “Evaluation of Claude Mythos Preview’s cyber capabilities,” April 13, 2026. 29. Antoine Montard et al., “Where AI will create value—and where it won’t,” McKinsey, April 2026. 30. Dago Diedrich et al., “From AI table stakes to AI advantage: Building competitive moats,” McKinsey, May 2026. 31. McKinsey Global Institute, “The power of one: How standout firms grow national productivity,” May 2025. 32. Mark Thomas, “Designing a FINRA for frontier AI,” *Lawfare*, July 30, 2026. 33. UK DSIT Sovereign AI Fund (April 2026); European Commission InvestAI Initiative (February 2025).

评论