CNBC: Chinese models 57–67% of OpenRouter tokens

Opening

Chinese AI models accounted for 57%–67% of tokens on OpenRouter in the week of Sep 14 (up from 6%–13% in February) and 55% of Vercel usage in August (from 11% in January), per usage data shared with CNBC — while U.S. frontier models still attract more overall spending. Kai Nicol-Schwarz reports the shift on developer gateways that route multi-provider access; OpenRouter’s cut covers companies in the U.S., Europe, and what it calls the “Global South” (82 countries). On OpenRouter, more than two-thirds (67%) of tokens used by Global South companies were on Chinese models; about half of OpenRouter tokens overall are used by U.S. companies. OpenRouter head of insights Peter Walker told CNBC Chinese open-source models released this year “can credibly perform in advanced agentic use cases, especially in regards to coding,” and are “incredibly cost-effective compared to most models from American labs.” Vercel’s Harpreet Arora tied adoption to price once quality clears the task bar, while saying firms still want U.S. frontier models for more complicated work. CNBC notes two U.S. House committees are investigating rising Chinese-model use; Remler (CNAS) framed sphere-of-influence risk. OpenAI and Anthropic both announced cheaper models earlier in the week (CNBC).

The numbers

OpenRouter token share — Chinese models 57%–67% week of Sep 14, vs 6%–13% in February — CNBC citing OpenRouter usage data.

Vercel share — Chinese models 55% in August, vs 11% in January — CNBC citing Vercel.

Global South on OpenRouter — 67% of those companies’ tokens on Chinese models; ~half of all OpenRouter tokens from U.S. companies — CNBC / OpenRouter.

Why it matters

Puts a dated, platform-level token-mix print under the open-weight price/performance story already in the archive — share of tokens is not share of spend, which CNBC still attributes to U.S. frontier models overall.

Positioning

Inference margin inversion — Platform token share migrating to cheaper Chinese open models is consistent with cost-per-token pressure on list; a break would be disclosed lab gross-margin compression across consecutive price cuts, not router mix alone.

AI capex durability — Rising open-model token share does not by itself cut hyperscaler build; it bears on whether serving demand keeps requiring new frontier capacity at prior intensity.

Connects to

Latent Space OpenRouter / Stripe — Same OpenRouter distribution layer now showing Chinese models as majority token mix on the gateway (longform).

DeepSeek ~$1B ARR — Prior China open-weight commercial scale color (news).

OpenAI/Anthropic cheaper models — Same week’s U.S. lab price response noted in the CNBC piece (news).