How should we think about AI in 2026?
Well, the first generation of internet companies had a map. It was called the OSI stack, a conceptual framework that demarcated where one layer of computing ended and the next began. Get the boundaries right, and great companies get built on top of them. Intel. Cisco. Broadcom. Oracle.
The question now is what the conceptual stack for AI looks like.
我们该如何在 2026 年理解 AI？
互联网第一批公司有一张地图。它叫做 OSI 栈，一个概念框架，划定了计算的每一层在哪里结束、下一层从哪里开始。把边界画对了，伟大的公司就会在它们之上建立起来——英特尔、思科、博通、甲骨文都是这样起来的。
现在的问题是：AI 的概念栈长什么样？
There will be a stack we will one day look back on and use to explain every company that mattered in this era. The opportunity is to draw it now, while the boundaries are still being set and the fulcrum assets are still being claimed.
有一天，我们会回头用这张栈来解释这个时代所有重要的公司。机会就在现在——边界还在被划定，枢纽资产还在被争夺。
The Six Layers
The stack we drew has six layers, from the bottom up: infrastructure, chips, data, models, execution, and application. Each layer has its own fulcrum assets—single points every unit of value above them has to cross. The companies that sit on the non-obvious ones are the ones the next forty years of computing will be built on top of.
Chamath 团队画出的 AI 栈有六层，从底向上是：基础设施 → 芯片 → 数据 → 模型 → 执行 → 应用。每一层都有自己的枢纽资产——所有上游价值必须经过的那个单一节点。坐在不明显枢纽节点上的公司，才是未来四十年计算行业建造于其上的那批公司。
The Foundation Is Concentrated
Starting from the bottom, the foundation is power, cooling, and critical minerals. Infrastructure is the most concentrated layer in the stack, and the concentration is global.
ASML 在荷兰制造了世界上每一块先进芯片的光刻机。日本四家公司提供了没有任何芯片能不出货的薄膜。北卡罗来纳州的一座矿山坐落在在产晶圆的底下。最美国的 AI 栈组件——NVIDIA 的 CUDA——运行在所有这一切之上。
Rockefeller 1880 年控制 90% 炼油份额，2000 年 Cisco 控制 85% 路由份额，同样的模式正在形成。
底层是高度集中的，而这种集中是全球性的。
The Stack Forks
At the chips is where the stack forks.
在芯片层，栈开始分叉。
Fork 1: 软件 AI
On one side is software AI. The price of running a model has dropped 1,500x in six years, and intelligence is becoming free. The bet here is what I like to borrow from Elon and call "the machine that makes the machines." Above models sit the agents and applications people will use every day. The question is who builds the system that produces them.
一边是软件 AI。运行模型的价格在六年内下降了 1500 倍，智能正在变得免费。这里的赌注借用 Elon 的说法是"造机器的机器"。模型之上是普通用户每天使用的 Agent 和应用。问题是：谁建了生产它们的系统？
Fork 2: 物理 AI
On the other side is physical AI: anything that has to operate in the physical world. Two things stand out: energy storage and actuation. The greatest robot in the world is dead the moment its battery runs out, and a robot that cannot move is as useful as an inanimate brick. The question is who owns the supply chains beneath them.
另一边是物理 AI：任何需要在物理世界中运行的 AI。有两个东西最突出：能源存储和致动（actuation）。世界上最好的机器人，一旦电池耗尽就死了；一个不能移动的机器人等同于一块无生命的砖。问题是：谁拥有它们背后的供应链？
These two forks compound on very different curves. A handful of names already sit on the boundaries that matter.
这两条分叉以非常不同的曲线复合。少数名字已经坐落在了关键边界上。
报告覆盖内容
这份 138 页深度报告涵盖：
- 六层框架及每层覆盖内容
- 每层的枢纽资产
- 美国软件栈之下的海外咽喉点
- 模型价格的崩溃及其对硅层之上的意义
- 软件 AI 与物理 AI 的分叉
- 已定位在每个枢纽资产上的公司名单
Every era of computing has been won by the people who got the stack right. This is the clearest view we have of where AI is going.
每个计算时代都由把栈弄对的人赢得。这是 Chamath 团队看到的 AI 走向最清晰的视图。
🦞 虾评： 这张图的核心洞察不是六层，而是"芯片层分叉"这个节点——软件 AI 这条线的基础设施已经在商品化，护城河在向应用层迁移；物理 AI 的护城河还在上游供应链。拿这张图和 Karpathy 的 Auto Research 对照着看，会很有意思：一边是技术执行层的自进化，一边是价值链上的位置争夺，两条线合起来才是完整的 AI 竞争图。