From the Jul 27, 2026 daily brief
Core judgment: An AI lab cannot be the TSMC of AI. TSMC won an entire industry's trust with "we will never make our own chips, never compete with our customers"; the other face of that bargain is its monopoly on leading-edge process, which leaves customers no alternative and locks them in. Selling raw tokens, by contrast, is textbook perfect competition — undifferentiated product, near-zero switching — and staying in that seat is a money-losing trap; the labs hold no equivalent of a process monopoly, so every one of them is forced up-stack into first-party applications. One correction to our earlier frame: the integration runs both ways — the application-layer companies whose business is being eaten are training their own models in reverse, and the whole chain is converging on "everyone becomes an integrated player." The genuinely valuable neutral position hasn't disappeared; it has moved to the distribution and governance layer. Governance means who gets to approve a model for launch, manage access, and do the gatekeeping — and the market has already begun pricing that neutral seat.
Why dig now: Two days ago this brief staked out the judgment that the labs are integrated players selling both the substrate and their own finished goods; this piece treats that judgment as a dashboard and runs its first checkup. All four monitorable indicators moved: Anthropic's coding tool, Claude Code, at $2.5 billion annualized revenue went from market estimate to official, on-the-record confirmation; the second batch of vertical applications — spreadsheets, legal, health, shopping — landed in a cluster around July 25, when this brief first posted the judgment (two days ago); the neutral distribution layer produced its first respectable number (OpenRouter's 25 trillion tokens per week); and third-party applications, far from being crushed, are compounding (Cursor, the AI code editor, went from $100 million to $2 billion in revenue in about 13 months). We also narrowed one old comparison: Samsung's foundry troubles are mainly a yield story and can't all be pinned on "not neutral" — a correction that doesn't overturn the main judgment, just rests its foundation on harder stone.
Open ?
What would prove this wrong: ① a general agent layer bound to no model vendor becomes the de facto standard, adopted by two or more large enterprises as their primary execution layer (our self-set watch window); ② Menlo Ventures' mid-year enterprise survey shows the annual model-switching rate sharply up (due around September 2026); ③ the labs' first-party application revenue share turns downward (our self-set watch window); ④ the purportedly neutral pure-foundry clouds go a year without producing measurable evidence of customer defection (our self-set watch window).
Verdict date: The main judgment settles July 27, 2027. Two short windows: Menlo's switching-rate report (around September 2026) and the OpenRouter deal closing or collapsing (event-driven). Today's top item 1 is the first market reaction this judgment received, hours after it was posted.
Just an email address, unsubscribe anytime. This is the only thing we ask of you.
The call: Databricks CEO Ali Ghodsi's prophecy — LLMs are a commodity; the labs' endgame i…
Core judgment: In the second half of 2025, four heavyweights — investor Brad Gerstner, ent…
Core judgment: "Do chip controls work" is an obsolete question in 2026. Keep three ledgers…
The core judgment. That Anthropic's token business has flipped from losing money to making…