A first stress test of the “lab as… — resolves 2027-07-27

A first stress test of the “lab as integrated device manufacturer” thesis moves all four of its watchpoints. Anthropic has officially confirmed roughly $2.5B in Claude Code revenue, one-fifth to one-third of a total revenue estimate that itself spans a wide $14-30B band. Second-adjacent verticals have converted within two months across multiple sources — Excel, legal, healthcare, shopping — in a pattern that looks like platform absorption, with the shutdown of Atlas standing as the first case of a lab's own first-party product losing that fight. A neutral distribution layer now has real pricing: OpenRouter is carrying 25 trillion tokens a week on a $113M Series B round, and Bedrock's usage in Q1 exceeded any prior year. Meanwhile third parties are not dying, and some are integrating backward into their own models: Cursor is at a $2B ARR run rate with its Composer 2 model, and SWE-1.5. Read together, this upgrades the call from “labs are one-way integrated device manufacturers” to “everyone is becoming an integrator,” with the neutral layer sliding down into distribution and governance, exemplified by the OpenClaw Foundation's sponsorship structure.

Stratechery(Ben Thompson)
「TSMC was unique precisely because they didn't desi…」

Stratechery(Ben Thompson)
「Claude Code alone has achieved an annualized recur…」

SemiAnalysis
「Anthropic's ARR has exploded from $9B to over $44B…」

SemiAnalysis
「AWS margins inflected this past quarter driven pri…」

AI Snake Oil / Normal Technology (Arvind Narayanan & Akash Kapur)
「If we assume a 5% net margin — much higher than ai…」

arXiv (Satoshi Matsuoka)
「a depreciation conveyor delivers newly amortized f…」

Ethan Mollick(Wharton 教授,One Useful Thing)
「Because of this change, you have to consider three…」

Reviewed over the next 12 months, through 2027-07-27: (1) if a shared meta-harness layer becomes the de facto standard across products, that favors vertical integration outright; (2) Menlo Ventures' mid-year report, expected around 2026-09, showing a sharp rise in switching rates would loosen the integration-quality premium; (3) a first-party lab's revenue share turning down, or another Atlas-style retreat from an adjacent vertical, downgrades the integrator reading; (4) twelve months with no measurable evidence of workloads leaving pure inference clouds retires the “token foundry” framing, while evidence to the contrary upgrades the neutral-layer-is-moving thesis.

Jul 27, 2027

2026-09-30(Menlo 年中報告)

Not due yet — no public outcome to report.

Not resolved yet — here is the condition we pre-committed to that would overturn this call (check it yourself, you don't have to wait):

The call is overturned if any of the following holds: a shared meta-harness layer fails to become a de facto standard within 12 months; Menlo's mid-year report shows no sharp rise in switching rates; or a first-party lab's revenue share turns down, or another adjacent-vertical retreat like Atlas's recurs.