Systems Research系统研究Nov 4, 20262026年11月4日Note短评

Owning the model lets the labs take the layer above it. The thin reseller stuck in the middle gets squeezed out.握住模型，实验室就能吃掉它上面的应用层。夹在中间、只负责转卖的那一薄层被挤了出去。

A bright, sunlit editorial illustration. On the right, a huge glowing golden spigot pours a ribbon of luminous light into a small counter bolted onto its base, where an anonymous hand passes glowing cups straight to a short queue of faceless silhouettes. In the middle, a flimsy paper-thin vendor kiosk is folding flat and tipping over, bypassed, with empty tin cups rolling away. The left area is open bright sky with the headline 'Own the Input, Skip the Middleman' and a Chinese line beneath it reading 握住输入，越过中间商.一幅明亮阳光的编辑插画。右侧，一个巨大的发光金色水龙头把一道光流注入螺栓固定在其底座上的小柜台，一只匿名的手把发光的杯子直接递给一小队没有面孔的剪影顾客。中间，一个薄如纸片的小摊正被压扁、翻倒、被绕开，空锡杯滚落一地。左侧是开阔明亮的天空，写着英文标题 Own the Input, Skip the Middleman，下面一行中文：握住输入，越过中间商。Illustration: WAKENEX插画：WAKENEX

Menlo Ventures' State of Enterprise AI report, out November 3, puts the AI coding market at $34 billion this year, up 7.7x. Claude Code and Codex, the labs' own tools, took $24 billion of that, about 70 percent, up from roughly $1 billion a year ago.

Read as a scoreboard, the labs crushed everyone. That is the wrong read. They did not win on better autocomplete. They collapsed two layers into one: they already own the scarce input, the model, and now they ship the agent that runs directly on it. Nobody sells model access back to them. The margin that used to live in the middle, a wrapper reselling someone else's model, is gone.

The tell is that the independents still grew. Cursor, Cognition, Factory, Lovable and the rest grew 2.9x, to $10.1 billion. The whole market is growing. What got squeezed is one position: the thin layer whose only job was access to a model it did not control.

If you build with these tools, stop arguing about which CLI to pick. The real question is where your leverage sits. Anything a lab can clone at its own marginal cost is not a moat, just a feature they have not shipped yet. The durable ground sits above the harness, in distribution, domain data, and judgment about what to build, or below it, in specialization the labs will not bother with. The coding agent itself is heading toward costing what the tokens cost.

We'd revise this if next year's Menlo numbers show the independents taking share back from the labs' combined total. Then 70 percent was a subsidized land grab, not structural capture.

Menlo Ventures 11 月 3 日发布的《企业 AI 现状》报告，把今年的 AI 编程市场规模定在 340 亿美元，同比增长 7.7 倍。其中，Claude Code 和 Codex 这两个出自大模型实验室自己的工具，拿走了 240 亿美元，约占 70%，而一年前这块还只有约 10 亿美元。

把它当成一张排行榜来看，会觉得实验室把所有人都碾压了。这是误读。它们赢，不是靠更好的自动补全，而是把两层合并成了一层：它们本来就握着那个稀缺的输入——模型，如今又直接在模型之上把智能体做成产品出货。没有人需要再向它们转售模型的访问权。原先活在中间那一层的利润——一个转卖别人模型的壳——消失了。

真正的破绽在于，独立厂商其实还在长。Cursor、Cognition、Factory、Lovable 这些加起来增长了 2.9 倍，到 101 亿美元。整个市场还在变大。被挤掉的只是一个特定位置：那层唯一的价值，就是能接入一个自己并不掌控的模型的薄壳。

所以如果你用这些工具做东西，别再纠结该选哪个命令行了。真正该问的是，你的杠杆放在哪里。任何实验室能以自己的边际成本复制出来的东西，都不是护城河，只是它们还没顺手做出来的功能而已。真正稳的位置，要么在执行框架之上——分发、领域数据、以及对"该做什么"的判断，要么在它之下——实验室懒得去碰的专精。编程智能体本身，正朝着"只值它消耗的那点 token 钱"滑过去。

如果明年 Menlo 的数据显示独立厂商从实验室的合计份额里把份额抢了回来，我们就会修正这个判断。那就说明这 70% 是补贴烧出来的圈地，而不是结构性的占领。

Sources来源

Primary source: menlovc.com/perspective/2026-the-state-of-enterprise-ai原始来源：menlovc.com/perspective/2026-the-state-of-enterprise-aimenlovc.com ↗

WAKENEX / NOTESWAKENEX / 短评

Add WAKENEX as a preferred source on Google在 Google 把 WAKENEX 设为首选来源

Sources来源

READ NEXT接着读

Legal stock perps change the venue, not the leverage risk you carry.个股永续合约合法了，换的是场地，不是你承担的杠杆风险。

Reserved blockspace with a zero penalty is a promise, not a guarantee.预订的区块空间，只要违约不用付代价，就只是承诺，不是保证。而这正是 ETHGas 眼下在以太坊上向华尔街兜售的东西。

On one cheap AI provider, you only rent its uptime.把工作流押在一个便宜的单一模型上，你买到的是算力，租到的却是它的在线率——它崩，你就崩。