Agentic Demand · Model-demand pulse
Token Split Pulse
The AI model race has a scoreboard, and it is denominated in tokens. TSP measures — daily — how the world's open-market model demand splits between Chinese and US frontier models, counted from public usage data on OpenRouter, the largest neutral model router. Price moves tell you what investors believe; token flows tell you what builders actually run.
| Window | Chinese models | US models | European models | Rest of world | Total | CN share | EU share | CN ÷ US |
|---|---|---|---|---|---|---|---|---|
| Last day | 6.04T | 2.64T | 0.06T | 0.05T | 8.78T | 68.8% | 0.7% | 2.29× |
| Trailing week | 38.9T | 18.8T | 0.39T | 0.28T | 58.4T | 66.7% | 0.7% | 2.07× |
| Trailing month | 149.9T | 81.2T | 1.61T | 2.15T | 234.9T | 63.8% | 0.7% | 1.85× |
| # | Model | Lab | Origin | Tokens |
|---|---|---|---|---|
| 1 | mimo-v2.5 | xiaomi | China | 10.8T |
| 2 | deepseek-v4-flash | deepseek | China | 6.73T |
| 3 | hy3 | tencent | China | 4.76T |
| 4 | deepseek-v4-pro | deepseek | China | 3.34T |
| 5 | glm-5.2 | z-ai | China | 3.31T |
| 6 | nemotron-3-ultra-550b-a55b | nvidia | United States | 2.52T |
| 7 | minimax-m3 | minimax | China | 2.09T |
| 8 | step-3.7-flash | stepfun | China | 2.02T |
| 9 | claude-4.8-opus | anthropic | United States | 1.33T |
| 10 | kimi-k3 | moonshotai | China | 1.25T |
Why the token split matters
A month-over-week-over-day comparison of the ratio is a built-in momentum read: when the last day's split is wider than the trailing week's, and the week's wider than the month's, Chinese-model share is still accelerating. The split became a mainstream talking point in July 2026 when BlackRock's Investment Institute charted it; TSP keeps the measurement running continuously, on the same public source, joined to the rest of our boards. The mechanism behind it is the one our Kimi-K3 series traces: open-weight Chinese models at competitive quality and a fraction of the price pull the open market's inference volume — and inference volume is where the economics of AI are decided.
Read it WITH the family, not alone: token flows here are the demand side; who captures the resulting compute, memory and power spending is the supply side our Rubin Build-Out tracks, and whether leadership is migrating from building to operating AI is what the AI Handoff Board scores. Method: every model's prompt + completion tokens from OpenRouter's public rankings data, summed into four buckets by the model lab's home region — United States, China, Europe, rest of world (map covers >99.9% of routed tokens; unmapped orgs are surfaced, never silently bucketed). Day, week and month windows are the source's own; the daily series accrues one point per run. Scope: OpenRouter is one router — a live sample of the open, contested API market. It does not see first-party surfaces (ChatGPT, Gemini, native lab APIs) or China's domestic routing, which is precisely why it is the right venue for the question "whose models win when builders can choose freely?" Research diary, not investment advice.
The European line deserves its own sentence. The continent with a sovereign-AI industrial strategy — the one our Euro-AI Sovereign 50 tracks on the infrastructure side — carries around one percent of the open market's model demand: Mistral almost alone, with Germany's entries (Black Forest Labs, TNG) at the margin of measurement. Europe is building the factory while barely shipping the product; whether that gap closes is exactly what a standing daily number makes visible.