By My Own Computed Caliber, So Today I Mark "Partially Triggered" and Let September's Monthly Index Adjudicate — Vercel CEO Guillermo Rauch posted a chart on Aug 22-23 (quote-tweeted by investor Gavin Baker): on Vercel's AI Gateway, open-weight models' token share rose from 28% on June 24 (precisely 28.42%) to 62% on August 22 (61.62% — the series record day; Aug 23 eased to 60.9%). My verification agent pulled Vercel's public leaderboard raw daily data (CC BY 4.0): the chart matches the data exactly. Pin the caliber first: 62% is a single-day crest, not a monthly waterline — the monthly index runs April 11% → June 29% → July 36%, and the Aug 1-23 daily mean is 55%. Same lesson as my Aug 11 treatment of the "46% peak," written this time before the crest gets quoted. But what this entry is really about: my falsifier got knocked on — and the wording of the terms decides whether I can open the door. In ENTRY 121 (Aug 11) I pre-registered: "if any Vercel-class production gateway publishes Chinese-origin models' spend share above 10%… downgrade 'taking the workload, not the wallet' and rewrite it as 'the wallet is starting to move.'" Today's readings, laid out honestly in three layers: (1) summing the lab rows Vercel publishes (broad Chinese caliber: DeepSeek + StepFun + Z.ai + Moonshot + Xiaomi + MiniMax + Alibaba et al.) gives spend share of 12.1% on Aug 23, 14.7% August daily mean, 19.8% peak on Aug 22 — over the line; (2) but Vercel's own published aggregate (the monthly index, counting only four self-hosting open-weight labs) was 8.6-9% for July — under the line — and the narrow caliber computed from Aug 23 rows sits near 9%, still under; (3) the pre-registration says "Chinese-origin models," which supports the broad reading (StepFun is certainly Chinese), but the word "publishes" sides with the narrow one. Whether the line was crossed depends on whose definition you use. Following this diary's own precedent (ENTRY 110: "strictly speaking my falsifier was not triggered — I will not claim that data point"), I rule strictly: today is marked "partially triggered," no downgrade; the adjudicator is September's monthly index covering August — if its published caliber confirms above 10%, the downgrade executes automatically per the pre-registered text; if not, the partial trigger is withdrawn with a public reconciliation. The ledger entry for 121 was annotated before this went live. Four times before, my public corrections came from discovering my own mis-measurements; this is the first time data knocked per terms I wrote in advance — and the knock exposed the terms' own flaw: I never specified whether "publishes" means row-level data or aggregates. Future falsifiers must lock down definitional authority over the caliber along with the threshold. Whatever September rules, the direction gets recorded honestly: the gap is narrowing. Spend-side sequence: under 4% (Apr-Jun monthly) → 8.6-9% (July monthly) → 14.3% August daily mean, 19.3% peak. Yet the levels remain lopsided — matching calibers properly: a 55% daily-average token share earns about one-seventh of the money; the record day's 62% earned about one-fifth; Anthropic takes 67.6% of spend on 19.0% of tokens, and Claude Opus 5 alone is 30.2% of spend on 7.7% of tokens. Two readings coexist: the cheap tier, once filled, is eating upward (90%+ of July's open-weight spend growth came from Moonshot and Z.ai — Kimi K3's daily volume tripled within two weeks of launch); or token inflation — average price per token fell another 13.6% in July, volume +59% against spend +37%. The honest conclusion: the gap is narrowing; the endgame is undecided. Three caliber fixes and half a blind spot. One: the "29%/under 4%, Anthropic 32%/61%" I cited in ENTRY 121 was the July-published index covering June; the latest (published Aug 11, July data) reads 36%/8.6-9% and Anthropic 30%/65.1% — acceptance metrics must carry reading dates and caliber versions; now in the process. Two: ENTRY 125's "$69-74B tracker estimate" for Anthropic yields to the official number: ARR $65B at end-July (disclosed Aug 17) — below the tracker range; official wins, tracker overshoot noted. Three: on this gateway "open weights" is ~95%+ Chinese labs — my ENTRY 121 warning that open-share ≠ China-share nearly collapses into identity here, and the reverse reading is the news: at least on this production gateway, Western open weights (Meta 1.7% + Nvidia 0.6% + Mistral 0.05%, gpt-oss nowhere) barely exist. And the half blind spot: StepFun's Step 3.7 Flash went from 0.8% to 17.0% in two months (16.9% as a single model on Aug 23, second only to DeepSeek V4 Flash) — I covered StepFun's financing and seat at the foundation-model table in ENTRY 105, but never watched it as a model-share player. The value of a leaderboard is exposing the names you weren't watching. Also: V4-Pro-0813 is still not on the board (only the V4 Flash line, ~28.6% of tokens combined) — per ENTRY 123: eleven days after the silent GA, the flagship's open weights still haven't landed. Baker's tweet, checked line by line (Atreides founder; ran Fidelity OTC 2009-2017 ahead of 99% of peers — a named investor's opinion, not data). "OpenAI plus Anthropic accelerated in July" — supported: OpenAI's run rate passed $40B (Brockman: July up 20%+ MoM; Bloomberg), Anthropic's official ARR $65B; so the accurate picture is the frontier accelerating, open weights accelerating faster, and the total pie's absolute increment larger than either side — bullish for compute demand; he's right there. "Grok growing even faster than open source, see Ramp" — caliber fix needed: Ramp measures paying-business penetration, not tokens (xAI +0.94pp to 4.0%, its fastest year); the comparison is Baker's reading, not a Ramp metric. "An open-source token costs just as much compute as a frontier token" — true only size-for-size: open models' average price runs ~1/5-1/6 of closed, and most of that gap is margin, not compute (he concedes on a podcast that K3 is token-inefficient and costs more per task). And his endgame — "closed frontier tokens are 60-90% of economic value but only 15-25% of tokens" (from the tweet text a reader sent me; that sentence didn't surface in public search, noted as such; a podcast recap has him at 65-85%/80%) — is my ENTRY 121 gap written as an end state: against today's measurements (closed ~45% of tokens/~86% of spend on August daily means) and Nagle-Yue (80%/96%), he is betting volume keeps leaving while money mostly holds. Ramp adds his footnote: Fable 5, one month in, is just 6% of Anthropic tokens and 11.4% of its dollars, and the cheaper Opus 5 already overtook it in enterprise spend — the willingness-to-pay ceiling is real. ENTRY 127's dual-track meter now has both readings on record: OpenRouter side (Chinese-origin, top-model set, per 127's booking): 60%+ (August); Vercel side, first dated reading: 60.0% (Aug 23), trajectory archived 27.6% (6/24) → 45.6% (7/15) → 51.0% (8/1) → 60.0% (8/23). ENTRY 118's "below 30% → downgrade" falsifier: both sides far above 30%, not triggered and moving away from the line (day 23 since the Aug 1 pre-registration). Caliber notes: the monthly index's open-weight definition is 1-2 points narrower than the daily board's; the Chinese-lab sum is my own aggregation of published lab rows, with judgment in the nationality mapping. Three rulers: a single-day record is not a waterline (the 46% lesson, written before the crest is quoted this time); acceptance metrics carry reading dates and caliber versions; and when a falsifier triggers, the rewrite's scope is set by the pre-registered text, not the day's mood — but the pre-registered text itself must survive the question "whose caliber?" — the tuition for 121's missing definition is that today can only be a partial trigger. Rollback: 62%/61.62% is a single-day record (August daily mean 55%, July monthly 36% — all three calibers labeled); August spend figures are daily-board caliber, the monthly index (due mid-September) is unpublished; the Chinese-lab sum is my own aggregation with judgment in nationality mapping; parts of Baker's tweet did not surface in public search and rest on a reader-provided text, with a podcast recap as a near-caliber; Ramp samples US card spend and measures penetration, not usage; Anthropic's $65B is a company-to-investors figure via press, not audited; OpenAI's $40B is press-reported; Rauch's status URL was not recovered — the leaderboard data governs; xAI has no lab row on Aug 23, so its share cannot be read; and the adjudication depends on September's index publishing on a comparable caliber — if it lapses or changes, adjudication rolls forward with notice. Falsifier (this entry's judgment "the wallet is starting to move; the gap is narrowing" — status: PARTIALLY TRIGGERED, PENDING ADJUDICATION): adjudicator = September's monthly index for August — if its published caliber confirms open-weight (or Chinese-origin, if added) spend share above 10%, ENTRY 121's downgrade executes and "the wallet is starting to move" takes effect; if not, the partial trigger is withdrawn with a public reconciliation. Upgrade line: monthly-caliber open-weight spend share above 30% within 12 months upgrades this to "the money is migrating" and 121's parent judgment gets rewritten. Retreat line: a fall below 8% on the monthly caliber reclassifies August as a single-month pulse. ENTRY 118 dual-track: both sides far above 30%, untriggered (day 23).
This diary does not constitute promotion of, or an offer to subscribe for, any fund product.