There is a pattern in how transformative technologies create value, and it almost never looks like a straight line. Brynjolfsson's J-curve framework, recently cited in the 2026 Stanford AI Index, offers the clearest mental model I've found for thinking about where we are in the AI investment cycle — and where the money moves next.

The core idea is simple. When a general-purpose technology arrives, productivity doesn't immediately surge. It dips. Enterprises pour capital into new infrastructure, retrain workers, and redesign workflows. The spending is real and measurable; the returns are not — yet. This is the bottom of the J. Only after organizations fundamentally restructure around the new technology does productivity accelerate, often explosively.

We've seen this before. Factories adopted electric motors in the 1890s but kept the old steam-engine floor layouts for decades. Productivity gains didn't materialize until the 1920s, when manufacturers redesigned plants around distributed electric power — enabling the assembly line. The IT revolution followed the same arc: PCs proliferated in the 1980s, Solow quipped in 1987 that "computers are everywhere except in the productivity statistics," and the payoff only arrived in the late 1990s when businesses went through genuine digital transformation.

AI's J-curve has one critical difference: adoption speed. The Stanford report notes that 53% of organizations deployed AI within three years — a pace that took PCs and the internet over a decade. This compresses the bottom of the J. The investment implication is that the transition from infrastructure beneficiaries to application-layer winners may arrive sooner than historical analogies suggest — perhaps 2027-2028, not 2030+.

This creates a natural two-phase investment strategy.

Phase 1 (now through ~2027): Ride the infrastructure capex cycle. As long as enterprises are in the "investment period" of the J-curve, capital flows to picks and shovels — semiconductors, advanced packaging materials, power infrastructure, cooling systems. The key insight is to look beyond the obvious names (NVIDIA, TSMC) toward the under-covered supply chain: Japanese and Taiwanese materials companies with near-monopoly positions that haven't been repriced by the market. A company with 90% market share in a critical bottleneck material, still valued as a sleepy industrial, represents mispriced optionality on AI capex continuation.

Phase 2 (2027-2028 onward): Rotate toward AI-native applications. When the J-curve inflects — when enterprises move from "bolting AI onto old workflows" to "redesigning workflows around AI" — the value creation shifts from infrastructure providers to application-layer platforms. The historical analogy: electricity's long-term winners weren't the power generation companies, but Ford and General Electric, who reimagined manufacturing around the new energy source.

The operational question is: what signals mark the inflection point? I'm watching four leading indicators. First, enterprise AI spending shifting from capex (buying GPUs, building data centers) to opex (paying for AI SaaS subscriptions) — this signals applications are delivering measurable value. Second, CFOs beginning to quantify AI's revenue contribution in earnings calls, rather than merely gesturing at "AI investments." Third, a wave of AI-native company IPOs, analogous to the 1999-2000 internet IPO cycle. Fourth, a step-change in aggregate productivity data — the U.S. hit 2.7% in 2025 against a decade average of 1.4%; a quarter printing 3.5%+ would suggest the J-curve is bending upward.

There's an additional dynamic worth noting: Jevons' Paradox applied to AI inference. DeepSeek V4, released this week, reduced inference costs by roughly 100x compared to frontier closed-source models. The naive interpretation is that this is deflationary for infrastructure demand. The correct interpretation, I believe, is the opposite — radically cheaper inference stimulates radically greater usage, which increases total compute demand. This is why materials companies at the top of the supply chain (where capacity takes 2-3 years to expand) remain structurally short even as per-unit costs fall.

The strategy, then, is not to choose between infrastructure and applications, but to sequence them. Earn returns from infrastructure today while building conviction on which application-layer companies will matter tomorrow. The J-curve tells you approximately when to shift weight. The specific signals above tell you that the shift is actually happening, not merely anticipated.

One final thought. The electricity analogy is instructive in another way: Ford existed in 1910, well before the productivity boom of the 1920s. The application-layer winners of AI likely already exist today — they're just not yet in their exponential phase. The observation list should be built now, even if the portfolio allocation comes later.

变革性技术创造价值的方式有一个规律：几乎从来不是一条直线。Brynjolfsson 的 J 曲线框架（2026 年斯坦福 AI Index 报告引用）是我目前找到的最清晰的思维模型——用来理解我们在 AI 投资周期的什么位置，以及钱接下来会往哪里流。

核心逻辑很简单。当一项通用技术出现时，生产率不会立刻飙升，反而会先下降。企业大量投入资本建设新基础设施、重新培训员工、重新设计工作流程。钱花出去了，看得到；回报还没来。这是 J 曲线的底部。只有当组织围绕新技术完成根本性重构之后，生产率才会加速上升——往往是爆发式的。

历史上反复出现过这个模式。1890 年代工厂就开始用电动机了，但厂房布局还是照搬蒸汽机时代的设计，足足用了几十年。直到 1920 年代制造商围绕分布式电力重新设计工厂（催生了流水线），生产率才真正爆发。IT 革命走了同样的弧线：PC 在 1980 年代普及，Solow 在 1987 年吐槽"电脑无处不在，就是不在生产率统计里"，回报要到 1990 年代末企业真正完成数字化转型才兑现。

AI 的 J 曲线有一个关键区别：采用速度。斯坦福报告指出，53% 的组织在三年内部署了 AI——这个渗透速度，PC 和互联网花了十年以上。这意味着 J 曲线的底部会被压缩。投资上的含义是：从基础设施受益者到应用层赢家的切换，可能比历史类比暗示的更早到来——也许是 2027-2028，而不是 2030+。

这自然形成一个两阶段投资策略。

第一阶段（现在到 ~2027）：乘 capex 周期的东风。只要企业还在 J 曲线的"投入期"，资金就流向 picks and shovels——半导体、先进封装材料、电力基础设施、散热系统。关键洞察是看向那些被市场忽视的供应链：在关键瓶颈材料上拥有近垄断地位的日本和台湾公司，估值还没有被市场重新定价。一家在关键材料上占 90% 市场份额、但仍然按无聊工业股估值的公司，本质上是 AI capex 持续性的错误定价期权。

第二阶段（2027-2028 起）：向 AI 原生应用轮动。当 J 曲线拐头——企业从"在旧流程上加 AI"转向"围绕 AI 重新设计流程"——价值创造就从基础设施提供商转向应用层平台。历史类比：电力时代的长期赢家不是发电公司，而是福特和通用电气——他们围绕新能源源重新定义了制造业。

操作层面的问题是：什么信号标志着拐点？我在观察四个领先指标。第一，企业 AI 支出从 capex（买 GPU、建数据中心）转向 opex（付 AI SaaS 订阅费）——这说明应用层在创造可衡量的价值。第二，CFO 开始在 earnings call 里量化 AI 对营收的贡献，而不只是空泛地提"AI 投入"。第三，AI 原生公司出现 IPO 潮，类比 1999-2000 年的互联网 IPO 周期。第四，总体生产率数据出现跳升——美国 2025 年是 2.7%（过去十年均值 1.4%），如果某个季度打到 3.5%+，说明 J 曲线正在向上弯曲。

还有一个值得注意的动态：杰文斯悖论（Jevons' Paradox）在 AI 推理上的应用。本周发布的 DeepSeek V4 把推理成本降低了约 100 倍。天真的解读是这对基础设施需求构成通缩压力。但我认为正确的解读恰恰相反——成本暴降会刺激用量爆发，总计算需求反而增加。这就是为什么供应链最上游的材料公司（产能扩张需要 2-3 年）在单位成本下降的同时仍然结构性供不应求。

所以策略不是在基础设施和应用之间二选一，而是按时间排序。今天在基础设施上赚确定性的钱，同时建立对应用层公司的认知和观察。J 曲线告诉你大概什么时候该调仓。上述四个信号告诉你调仓的时机是否真的到了，而不只是被预期。

最后一个想法。电力的类比还有另一层启示：福特在 1910 年就存在了，远早于 1920 年代的生产率爆发期。AI 应用层的赢家很可能今天就已经存在——只是还没有进入指数增长阶段。观察名单现在就应该建，即使组合配置要等到以后。