“장보러 가면서 페라리를 모는 격”…사용량 1~5위가 전부 중국 모델“Like driving a Ferrari to the grocery store”—the top five by usage are all Chinese

The top five AI models by usage are now all Chinese.

Axios reported it on 18 July, based on weekly token usage on the model routing platform OpenRouter. All five are open-weight models you can download and modify.

The best-performing model was not Chinese. What the market picked was not the smartest option.

It ranked fourth on performance

Kimi K3, released by Moonshot AI around the same time, scored 57 on Artificial Analysis — fourth, behind Claude Fable 5 at 60, GPT-5.6 Sol Max at 59 and Sol XHigh at 58.

So why did everyone use it? Price. Kimi K3's API cost 100 yuan per million output tokens, about 219 won.

Augusto Marietti, chief executive of the API company Kong, said use of open-weight models had surged over three months because frontier models were prohibitively expensive. Mozilla's chief technology officer Rafi Krikorian likened using a frontier model for everyday work to driving a Ferrari to the grocery store, and said routine costs could fall by up to fiftyfold.

One investor quoted by Axios expects open models to handle 95% of enterprise AI requests, leaving 5% to frontier models. Morgan Stanley said the results show Chinese large language models catching up with American leaders on scale, performance and price.

The founders are in their thirties and forties

So who built them? On 23 July the Hong Kong paper Sing Tao Daily profiled four Chinese large language model founders.

DeepSeek's Liang Wenfeng, 41, holds 84.29% of the company with voting rights close to 100%. Moonshot AI's Yang Zhilin, 34, holds 51.83%, with personal assets estimated around $16.3bn. Zhipu's Tang Jie, 49, is the only academic of the four, and MiniMax's Yan Junjie, 37, holds 21.6%.

These are estimates rather than confirmed figures. The ownership stakes still stand out: a different structure from American AI firms backed by outside capital and big tech.

All four started between 2015 and 2023 — within a decade of where they now stand.

Twice the target arrived

The money followed. Bloomberg reported on 29 July that Moonshot AI had raised $3.5bn, roughly 5 trillion won, at a post-money valuation of $35bn.

The original target was $1bn to $2bn. It closed at nearly double that. In February 2024 the company was valued at $2.5bn.

Revenue moved too. Annual recurring revenue went from $100m in March to $200m in April and $300m in June — tripling in three months.

The next round targets $50bn, with a Hong Kong listing planned within the year. China's national AI industry investment fund came in as a major backer of this round, so it is not purely private money.

Set against the backdrop it reads oddly. The US restricts exports of advanced AI chips to China. The money went to a model built under exactly that constraint.

The line was drawn inside America

The US government began weighing a block on Chinese open models on security grounds. Objections followed — from American companies.

On 24 July more than twenty US firms signed an open letter: Nvidia, Microsoft, Meta, Palantir, IBM, Dell, and the venture firm Andreessen Horowitz among them.

The argument ran three ways. Open models let startups, universities and public bodies build without heavy cost; published weights let researchers inspect behaviour and find vulnerabilities; and illegal distillation should be distinguished from legitimate development.

Jensen Huang, Nvidia's chief executive, shared the letter and said the world needs both frontier closed models and frontier open models.

The absences are notable. OpenAI, Anthropic and Google did not sign. They are the closed camp, and they have been warning Washington about the risks of Chinese open models.

The government held firm. Michael Kratsios, director of the White House Office of Science and Technology Policy, accused Moonshot AI of replicating American models through distillation, and signalled possible sanctions alongside Treasury Secretary Scott Bessent.

Distillation trains a smaller model on a larger one's outputs. Where legitimate learning ends and unauthorised copying begins sits at the centre of this dispute.

What is at stake is the cheap option

The line was not drawn between America and China but inside America. Those selling chips and infrastructure argue for openness; those selling models argue for restriction. Each speaks from where it stands.

So will the restrictions land? Nothing is settled. The government is reviewing and the industry is objecting.

If they do land, the cheap options shrink. A growing number of Korean services already run Chinese open-weight models underneath, so barring US companies would reach that flow too.

The top five slots by usage, founders in their thirties and forties holding their own equity, and an investment round at twice its target. Three scenes, ten days apart.