AI Model Spend Share Watch Chinese vendors gain 3.7 points in 37 days
2026-09-21 → 2026-11-01 · 37 data days · published 2026-11-02 · methodology 2026.10.25
This note counts one thing: who is actually paying to call models. Weekly spend converted at each vendor’s list price rose from $70.8M to $78.9M in 37 days. Inside that growing pool, Chinese vendors moved from 25.0% to 28.7%, while US vendors slipped from 74.0% to 70.4%.
+3.7pp
China share change
-3.6pp
US share change
$78.9M
Latest weekly spend (USD)
1. Share trend: not linear — three steps
Fig. 1 — Share of weekly spend, China vs US. Grey lines mark 10-point steps.
2. Who moved the curve
Fig. 2 — Share change of the top 8 model families (last day vs first). Full set covers 295 families.
3. What the numbers mean
The shift comes from two specific models — and the rest of the field is bleeding. DeepSeek V4.1 Flash rose from 8.3% to 12.5% (+4.2pp); Kimi K3.1 Pro went from 0 to 3.5%. That is +7.8pp together, twice the net gain of Chinese vendors (+3.7pp) — which implies the remaining Chinese vendors lost about 4.1 points combined. “Chinese models are rising” is the surface; what happened is an internal reshuffle.
A share can be eaten by a sibling model within a month. Claude Opus 5 fell from 3.8% to 1.6% (-2.2pp) while Opus 5.5 from the same vendor held above 20% — not lost market, but generational replacement. The shortest replacement cycle is now measured in weeks, not years.
The pool is growing, and it is also concentrating. Weekly spend grew 11% in 37 days; the top three model families went from 35.1% to 44.3% combined (+9.2pp). For teams doing model selection, the shortlist is narrowing, not widening.
Basis. Spend (effective usage value, USD) = (min(input tokens, requests × 64K) + overflow × 0.10) × input list price + output tokens × output list price. Overflow is discounted at the cache-read rate to mirror the 0.1× tier common in vendor price sheets.
Sample & coverage. 37 days of weekly-window snapshots (every 8 hours; last snapshot of each day). Preview SKUs excluded for comparability. Priced records cover 91.3% of token volume (13,834 / 23,675 records); the remainder is embedding models and long-tail small models, which does not affect the head of the distribution.
Limits. Aggregated from a single public aggregator’s billing basis; reflects spend settled through that platform, not total vendor revenue. Prices are list prices and exclude enterprise discounts.
How to cite (copy as-is)
An Li (Bamboohat). “AI Model Spend Share Watch: Chinese vendors gain 3.7 points in 37 days.” airanking.cc, 2 Nov 2026. https://airanking.cc/research/2026-11-ai-spend/