But the Bigger Error Is Mine: I Never Quantified Its Funnel — On August 19 Stripe and OpenRouter both announced: Stripe has signed an agreement to acquire OpenRouter. Three things people are getting wrong: it is signed, not closed (OpenRouter's own post ends "subject to customary closing conditions… expect to close in the coming weeks"); no price was officially disclosed — Bloomberg says over $7B, the NYT ~$7.5B (~$1.5B to founders, ~$6B to investors), Axios over $8B and mostly in stock, and the WSJ's July talks ran near $10B; and the Series B lead was CapitalG (Alphabet's independent growth fund), not Sequoia — TechCrunch got that wrong; OpenRouter's own announcement lists CapitalG leading with NVentures, ServiceNow, MongoDB, Snowflake and Databricks participating, a16z and Menlo as existing investors (disclosure: Menlo is an OpenRouter investor; read its figures as vendor numbers). Why this entry is mine to write: I cited OpenRouter twelve times across four entries (118 ×4, 121 ×5, 123 ×1, 124 ×2), and the falsifier of my ENTRY 118 judgment is anchored to it — the first time this diary's acceptance metric has itself undergone a change of ownership. But the bigger error predates the deal and has nothing to do with it. I hedged every citation qualitatively ("not the whole market," "developer-platform caliber," "top-10 models only, heavily open-weight-skewed") — but qualitative hedges are noise. The quantitative number sat in plain sight: MIT's Frank Nagle and Georgia Tech's Daniel Yue (published via MIT Sloan in January) found OpenRouter captures roughly 1% of global AI inference spend during their May-September 2025 study window, and that the platform "attracts the type of user who's more likely to be willing to use open models." It is probably above 1% today (volume doubles every ~11 weeks, per Menlo), but not double digits — my judgment. Meaning: every headline "Chinese model share" number I used as a primary caliber measured a self-selected funnel. The same research: closed models ran ~80% of OpenRouter tokens but took ~96% of its revenue ($1.86 vs $0.23 per million tokens) — note the researchers are independent but the data is not (it is OpenRouter's own); the genuinely independent second funnel is Vercel (open weights at 29% of tokens, under 4% of spend, June caliber) — two funnels, same shape. And since the self-selection favors open models, the true closed share can only be higher — the bias is conservative. Mozilla.ai adds the third piece: locally self-hosted small models (the llama.cpp/Ollama world) are invisible to any centralized API platform. One item filed as a question, not a finding: OpenRouter's separately-ranked :free variants might inflate open-weight token counts — plausible mechanism, no quantified evidence. One hard correction of my own number, whose provenance is itself the lesson: I wrote "Chinese models exceed 60% of OpenRouter platform volume" — wrong denominator; the accurate claim is "60%+ of token consumption among OpenRouter's top-ranked models." That sentence came from ENTRY 121's falsifier progress line — the same entry in which I corrected two other phrasings ("US enterprise" → "US accounts"; 46% is a weekly peak). I fixed two errors and minted a third in the same entry: correction mechanisms themselves leak. (The usual caution stays attached: CNBC's 46%/11%/4.5% may not share one series; direction only.) The true picture needs no inflation: in late July, Chinese models held all top five OpenRouter slots by token volume (Xiaomi MiMo-V2.5 first, then DeepSeek, MiniMax, Qwen, Kimi). Now the deal itself, one level deeper than "payments company buys a router." Stripe closed its ~$1B acquisition of Metronome — usage-based billing and metering — on January 15. Seven months later it signed OpenRouter. Read together: Stripe is buying the tollbooth on the pay-per-consumption track, not the routing product. The price supports that reading: Sacra estimates ~$140M annualized revenue (a ~5% take rate; third-party estimate), so $7.5-8B is roughly 55x revenue (54-57x) — a routing take-rate cannot explain that multiple; the metering-and-settlement position can. Stripe's position was already there: $1.9T of 2025 payment volume (its own annual-letter figure, ~1.6% of global GDP), and roughly nine in ten of the Forbes AI 50 building on it, including OpenAI and Anthropic. This is external validation of my ENTRY 125 judgment — AI revenue is migrating from per-person to per-consumption — and when an industry's pricing unit changes, the meter is what gets bought first. The null hypotheses stay on the table: a bid-up price, a defensive acquisition, or routine growth-buying with your own stock — and if the consideration is mostly Stripe stock (Axios), the 55x was paid in Stripe's private marks, which discounts its hardness. Neutrality: read what was and wasn't promised separately. OpenRouter promised specifics — "same mission, same name, same product, same roadmap," routing decisions that "remain driven by one thing: what's best for you, the user," a commitment that "doesn't bend to any model, any provider, or any parent company." Two credits the team is owed: the rankings page itself carries methodology caveats ("measure adoption, not quality"; "token volume is not a count of requests, users, or spend") — the problem is the /data landing page and press summaries don't, and that's the layer most citers see; and the arXiv paper discloses its own limits (BYOK excluded, self-hosted invisible, billing-based geo attribution imperfect). Not concealment — unevenly distributed disclosure. But two things must be said plainly. Nothing anywhere commits to continuity of the rankings, the Data API, or the State of AI report — product continuity is not data continuity; and while the dataset is CC BY 4.0, retrieval requires an OpenRouter key, rate-limited 30/min and 500/day, revocable per account — license permanence is not access permanence. And the interest structure has two layers that should not be braided: a16z's conflict is direct (a partner co-authored State of AI; the firm is a shareholder); CapitalG is structural optics (Alphabet — owner of Gemini — led the round into the meter that scores Gemini's share, but signed no report; no evidence of influence). The structural conflict has a concrete mechanism: Stripe now owns both the platform deciding which model each query reaches and the rails collecting payments from those model providers. Precedent tells me when to worry. Closest structural analogue: CoinMarketCap under Binance (March 2020, ~$400M, with independence promised): soon after, the default ranking metric became a "web traffic factor," lifting Binance from 15th to 1st and dropping BitMEX to ~175th — but the second half must be told too: after community backlash, CMC walked it back. My remaining worry: the reversal came from public pressure, not governance — next time, no backlash, no reversal. Longer arcs: Alexa published rankings for 23 years under Amazon, then shut down in May 2022 without explanation; App Annie's free tier vanished in the Sensor Tower migration. The pattern is slow: the free public layer goes first, often years later, usually unannounced. The counterexample matters: GitHub promised npm's public registry "will always be available and free" — kept, to this day. The discriminator: public data that is the acquirer's own distribution channel survives (npm); a scoreboard ranking the acquirer's counterparties is high-risk (CMC). OpenRouter is the scoreboard. And 2026 has a live case: after Publicis announced its ~$2.17B proposed acquisition of LiveRamp (May 17, not closed), Omnicom and WPP walked — the near-term effect is not data being changed but competitors leaving first, and their exit changes the data by itself. (As of this writing the rankings page still updates normally — grounds for monitoring, not alarm.) So, ledger surgery, four items. One: ENTRY 118's share-side falsifier goes dual-track, unified caliber "Chinese-origin model share, 30% threshold" — OpenRouter side: Chinese share of token consumption within its public top-ranked set (currently 60%+); Vercel side: the sum of Chinese labs' rows on its public daily leaderboard (first dated reading to be logged next entry). Either side below 30% → "partially triggered," publish readings, re-verify; both sides below → formal downgrade; if OpenRouter stops publishing, changes methodology, or tightens access, adjudicate on Vercel alone at the same 30%. Two: OpenRouter is demoted to a directional indicator — direction yes, market share no. Three: its State of AI report is cited as a vendor report (company-plus-investor authorship). Four: monthly meter snapshots — license permanence isn't access permanence, and without a baseline snapshot you cannot prove a methodology changed (backfill note: the first snapshot was archived on Aug 24 under archive/meter-snapshots/, monthly thereafter). One debt on record: ENTRY 121's own falsifier hangs solely on Vercel and equally lacks a parallel indicator — logged here, handled next entry. Two rulers. Every external number you cite deserves the question "who owns this meter, and what are its incentives" — not because of the acquisition, but because every meter has a funnel; MIT's ~1% mattered more than Stripe's deal and sat unread for seven months. And never hang a judgment's falsifier on a single third-party metric — the extension of ENTRY 111's rule: specifying "what data decides" is not enough; ask whether the data source can vanish or change caliber. From today, every external metric written into a falsifier gets a second, independently-sourced parallel. Rollback: signed not closed; price undisclosed (three press calibers; mostly-stock per Axios); the $1.3B post-money is press/investor caliber (Menlo — an investor), the round actually closed ~February; revenue ~$140M and ~5% take are Sacra estimates, so ~55x is an estimate, further discounted if paid in stock; scale figures vary by source (25T tokens/week and 8M users in the announcement; Menlo's 4.5+ quadrillion annualized and 10M users — Menlo is an investor); MIT's ~1% covers May-Sept 2025, likely higher now, "below double digits" is my judgment; the 80%/96% shares the same source and window; Forbes AI 50 penetration is quoted 86-90%, "roughly nine in ten," and is Stripe's own claim; no evidence of data degradation as of writing; this entry backfills an Aug 19 event on Aug 24. Falsifier (new judgment: "a scoreboard-type meter signed by an interested acquirer will not survive unchanged — either participants exit first, or caliber/access changes"): CONFIRMED if within 12 months any of — a major model provider publicly reduces exposure routed through it; a ranking-denominator or methodology change without comparable historical data published; a material access tightening (rate cuts, paywall, approval gates). REFUTED (marked ✗, treated as a public correction) if within 12 months none of the three occurs.
This diary does not constitute promotion of, or an offer to subscribe for, any fund product.