the chinese AI models are dominating, along with the obvious split of the AI model market into 2 very different games

the first is revenue

the second is volume

OpenAI, Anthropic and Google still dominate revenue because enterprises buy trust, support, security, procurement, and integration. that part of the market moves slowly and is heavily shaped by cloud GTM, solution architects, and existing enterprise relationships

but OpenRouter is showing the leading edge of volume

by may 2026, chinese open-weight models were roughly 61% of tokens consumed on OpenRouter. four of the top five models were Chinese. Llama, once the default open-weight leader, has effectively disappeared from the top rankings

the cost-performance center of gravity for open-weight inference has shifted east

DeepSeek-V4-Pro pricing at roughly 12x below GPT-5.5 makes the point clear. the mass market does not always pay for the absolute frontier. it pays for the cheapest model that is good enough for the task

openRouter usage shifted heavily toward code, with programming rising from roughly 11% of usage at the start of 2025 to more than 50% by mid-2026. coding is high-volume, repeatable, and price-sensitive. that is exactly where cheap, capable open-weight models compound

premium reasoning is a high-margin niche. cheap, open, good-enough inference is becoming the volume layer

if the best open weights are increasingly Chinese, large enterprises will hesitate. export controls, procurement risk, data sensitivity, and political optics matter far more to a Fortune 500 buyer than to a startup trying to cut inference cost

so the market likely bifurcates, where startups chase performance per dollar, while enterprises stay with approved Western vendors longer:

- that delay creates a temporary distortion, not a permanent moat

- the real investment implication is that margin is moving away from the model layer

- it accrues above the model through distribution, workflow ownership, and application lock-in

- it accrues below the model through cloud, inference routing, optimization, and compute infrastructure

the model itself still matters, but in the volume tier it is becoming increasingly replaceable


https://x.com/arndxt_xo/status/2069994796048158939