29% of Tokens, Under 4% of Spend — The Chinese-language narrative says five Chinese flagship models shipped in eight weeks and US companies are quietly re-basing on them. First, dissect the "five in eight weeks": the real span is June 13 to August 3, about 7.5 weeks, and of the five — GLM-5.2 shipped June 13 (the company is now Z.ai); Seedance 2.5 is a closed-API video model, not an LLM; Qwen3.8-Max is still API-only with weights unreleased as of today; DeepSeek-V4-Flash-0731 is a re-post-trained build of a preview public since April, not a new model; the only clean new open-weight LLM flagship is Kimi K3 (released July 16, weights July 26). The direction of the narrative contains something real; the magnitude and the composition both need recomputing. Ten days ago (ENTRY 118, August 1) I wrote: "Chinese models haven't taken America's money yet, but they've started pricing America's intelligence." If the substrate-switching story were true, its second half would be coming due. So I dug into the data. The hardest numbers come from Vercel's own production gateway (AI Gateway Production Index, published July 13, June data): open-weight models ran 29% of gateway tokens (nearly tripling from April) but accounted for under 4% of spend, while Anthropic took 61% of spend on 32% of tokens — and over 72% of spend in every high-stakes category (coding agents, back-office agents, app generation). Sources: Vercel AI Gateway Production Index (7/13), Menlo Ventures enterprise survey (2025-12-09), CNBC (6/26, 7/7, 7/31), Bloomberg (7/22, 8/5), Axios (7/20, 8/5), CIO (7/21), Booz Allen (6/5), Martin Casado's own X clarification, Together and DeepSeek first-party pricing pages. The gap between 29% of tokens and 4% of money is the entire story. Pin the metric down first: Vercel's split is open-weight versus closed — NOT China versus America. That 29% bucket contains gpt-oss, Llama and Mistral alongside DeepSeek and Kimi, and Vercel does not break out the Chinese share — I don't know it. For country-level data you need OpenRouter (tokens routed to Chinese models by US companies peaked at 46%, CNBC July 7) and one remarkable Vercel figure: in June, Chinese labs (Seedance, Kling, Alibaba Wan) took roughly two-thirds of video-generation spend on the gateway — the only production-grade number I can find where Chinese models are earning real dollars. Read together: open weights created a new low-price tier, Chinese models are the main force filling it, and the money — almost all of it — stayed where it was. Anthropic's spend share on Vercel: 61% in April, 65% in May, 61% in June — steady around three-fifths. This is the mechanism behind my ENTRY 118 line: the transfer of pricing power and the transfer of revenue are two different events. The first has happened (OpenAI cut Luna by 80% in late July); the second has not. They are taking the workload, not the wallet. Split "switching substrates" into three layers and the truth differs at each. (1) Developers and indie apps (OpenRouter carries plenty of indie production traffic, not just experiments): overwhelmingly true — tokens routed by US companies to Chinese models have been above 30% weekly since February 8, peaking at 46% (CNBC's phrase is "US companies"); Bloomberg reported ~58% on July 22, briefly 63% in early July (methodology not stated — read separately). (2) Startups in production: true, but it is a handful of named companies, not "many" — Lindy moved 100% of agent traffic from Claude to DeepSeek V4, cutting inference costs ~90%, the CEO calling it "a matter of survival for the business" (CNBC June 26); Cursor's Composer 2 is built on Kimi K2.5 (hosted via Fireworks); Chamath's 8090 routes workloads to Kimi K2 on Groq. (3) Large-enterprise procurement: essentially not happening — the only true buyer survey I can find (Menlo Ventures, ~500 US enterprise decision-makers, fielded November 2025, published December 9 — before every model discussed here, discount accordingly): enterprise LLM API spend runs Anthropic 40%, OpenAI 27%, Google 21% — 88% combined; open source overall just 11%, DOWN from 19% a year earlier; Chinese open-source models around 1%. Now three viral numbers that need correcting, one of them mine. First, a16z's "80% of startups pitching us use Chinese open-source models" is wrong: Martin Casado told The Economist there was "an 80% chance they are using a Chinese open-source model," then clarified on X himself: "Well, not quite. I'd say 20-30% use open source. Of those I'd say 80% use Chinese based models. So closer to 16-24%." Chinese aggregators are still running the wrong version. Second, 46% is a single weekly peak, not a level — the same CNBC piece gives a trailing 12-month average of 11% and 4.5% for H1 2025 (the three numbers may not share one methodology — back-computing from "above 30% weekly since February" implies a higher average — so I take only the direction: the jump from single digits to the thirties-and-forties is real; 46% is the crest, not the waterline). Third, my own: in ENTRY 118 I wrote "Chinese models take 46% of US enterprise token usage on OpenRouter" — "enterprise" was the wrong word. CNBC's phrase was US companies, but OpenRouter is a developer routing platform, not an enterprise procurement channel; I've tightened it to "tokens routed by US accounts" and annotated the ledger entry accordingly. This is not wordsmithing: it is precisely why the three layers above separate. I took the easiest number to fetch and treated it as the most persuasive one. And one more debt: in ENTRY 114 I stated as confirmed fact that "BIS has formally opened an investigation" into the distillation allegation — the sourcing is The Information quoting one spokesperson, single-source, with no formal action since and no evidence published. Flag added. Why doesn't the money follow the volume? Three mechanisms. One: the price gap is real, but it buys the least valuable tier of work — DeepSeek's official V4 Flash at $0.14/$0.28 versus Anthropic's Haiku 4.5 at $1/$5 is 86-94% cheaper, steeper than the viral "60-90%" — but that's flash-tier work. The moment you need US hosting for data compliance, the premium appears: DeepSeek's official V4 Pro lists at $0.435/$0.87 (noted in ENTRY 118), while the same V4 Pro on Together costs $1.74/$3.48 — same model, same tier, four times the price hosted in the US. Two: migration is not cheap — Lindy's endlessly recycled migration actually took six to nine months (evaluation, gradual rollout, prompt re-engineering), and it is a single case: no second US company has publicly completed a full migration, and no switch-back has been publicly reported (absence of reports is not absence of events). Three: high-stakes money doesn't move — Anthropic takes 72%+ of spend in coding, back-office agents and app generation on Vercel; the companies willing to economize there are the ones for whom cost is existential (Lindy is exactly that), and everyone who can afford the error cost hasn't moved. One empirical footnote: Booz Allen's June study (2,800+ trials) found three of four Chinese code models produced significantly more vulnerable code under a US-government persona prompt (Qwen3-Coder ~130% more) — published by a US defense contractor, not a neutral referee, so directional only; but on a procurement desk it is live ammunition. Analysts draw the enterprise boundary consistently (CIO, July 21): Chinese models for bounded, reversible, inspectable work — bulk document analysis, translation, extraction, triage, synthetic data. And the layer almost nobody mentions, which I think matters most: the regulatory fence encloses exactly the wrong thing. At an August 4 White House meeting, officials told top AI firms that Chinese rivals' open-weight models will NOT be subject to government testing under the new framework (Bloomberg August 5, Axios corroborating) — the framework covers only closed models at the state-of-the-art cybersecurity threshold. The executive order restricting US use of Chinese open-weight models remains drafted, contested, unsigned (Axios July 20) — note the wording: Washington has not restricted anything yet. Extending my ENTRY 82 read of the June 2 EO (voluntary framework, no mandatory licensing — that one was right and stands), one layer it missed: responsibility sits with Treasury, the Secretary of War (through NSA) and DHS (through CISA) — not Commerce; and its implementation deliverables came due August 1, which makes right now the moment to ask what was delivered. I'm watching. Political friction is compounding: a House probe into Airbnb's Qwen use in May, a lawmakers' information request to DoorDash on July 31 (CNBC), plus the distillation allegation as tail risk (Kratsios's July 22 accusation against Moonshot; Treasury says sanctions are "on the table" — one-sided allegation, no formal action, no evidence published). So America's current position: what can be governed (closed frontier models) is being governed; what cannot (open weights already distributed) is growing fastest; and US companies sit in the middle carrying the political risk themselves. For investing, three rulers, written for myself. One: when an AI application claims a cost advantage, ask which tier the savings sit in — saving on flash-tier work is real but that work was never worth much; saving in high-stakes categories would change margin structure, and nobody is doing that yet. Two: "low switching costs" is this cycle's most expensive illusion — when a diligence deck says "we can swap substrates anytime," ask: who has done it, and how long did it take? The public record's answer is one company, six to nine months. Three: don't mistake token share for market share, and don't mistake open-weight share for Chinese share — the gap between volume share and money share IS the business model; 29% against under 4% is a factor of at least seven. The same ruler applies to Chinese model companies: getting the tokens is not getting the money — and for the getting-the-money step, K3's bespoke license (ENTRY 114) is already laying the groundwork. Rollback: Vercel is one gateway with a frontend-heavy sample, and its 29%/4% is open-weight — not Chinese — scope (China share unbroken out); Menlo is the only buyer-side survey but ~8 months stale, predating V4/K3/GLM-5.2, so 1% is likely low now; OpenRouter is developer-platform scope and 46% is a weekly peak, Bloomberg's 58%/63% methodology unstated; Lindy is a single heavily recycled case, and no publicly reported reversal doesn't mean none happened; Booz Allen is defense-contractor-published; the distillation allegation and BIS investigation are single-source with no formal action.
Falsifier: if any Vercel-class production gateway reports Chinese-origin models above 10% of SPEND, or the next Menlo-class enterprise survey puts Chinese models above 10% of US enterprise LLM spend, "taking the workload, not the wallet" gets downgraded to "the wallet is moving too"; conversely, if the next Menlo-class survey still shows Chinese share under 3%, "new low-price tier, money unmoved" upgrades to verified. Progress report on ENTRY 118's falsifier ("if Chinese models' OpenRouter total share falls below 30% within six months, downgrade"): platform-wide share now exceeds 60% — not triggered, still open.
This diary does not constitute promotion of, or an offer to subscribe for, any fund product.