США сдерживают китайский ИИ, а Кремниевая долина использует Kimi K3 и другие модели

Rest of World reported in 2026 on a paradox: while US authorities are trying to contain the development of Chinese artificial intelligence, American companies — including Apple and the startup Thinking Machines — along with developers around the world, are increasingly using open Chinese models like Kimi K3 from Moonshot AI.

Who is using Chinese models

Chinese open-weight models are being adopted not by marginal players, but by industry leaders. According to Rest of World, users include Apple, the AI startup Thinking Machines, and independent developers in various countries.

• Kimi K3 — an open-weight language model from the Chinese company Moonshot AI
• Apple — among the American companies working with Chinese open models
• Thinking Machines — an AI startup using models developed in China
• The trend was documented by Rest of World in 2026

The composition itself is telling: Chinese models are being adopted by the very American players whose country is officially trying to restrict this kind of proliferation. This is specifically about open-weight models: their weights are published openly, so any engineer can download the model and run it on their own infrastructure — without the developer's permission and without being tied to its paid API.

What is Kimi K3

Kimi K3 is the newest generation of language models from the Chinese startup Moonshot AI, continuing the Kimi line (the previous version was Kimi K2). Its key feature is open weights: the model is distributed in a way that lets anyone download and deploy it independently.

It is precisely this openness that made Kimi noticeable outside China. The model enters the Western stack not through marketing or partnerships, but through the developer community, which adopts it because it is available and it works.

What makes Chinese models attractive

Open Chinese models offer something closed American models do not: full control and low cost. A downloaded model can be fine-tuned, run locally, and embedded into a product without paying for every request to someone else's API.

The difference is especially noticeable in price: for a closed flagship model, you pay by the number of tokens, whereas an open one only needs to be deployed once on your own servers. For teams around the world, this solves several problems at once — data privacy, predictable costs, and independence from a single vendor. That is precisely why, as Rest of World notes, Kimi K3 and similar models are spreading among developers faster than American regulators can react.

Why the US cannot stop this

US export controls target hardware, not software. The restrictions concern the supply of advanced chips to China, but the weights of an already-trained model are files that are freely distributed over the internet and through repositories like Hugging Face.

Here, the containment strategy runs into a contradiction: the state can stop the export of GPUs, but it can practically not stop the downloading of an open-weight model once it has been released to the public. Barriers built around hardware do not work against the free distribution of software.

"The US wants to contain Chinese AI.

Silicon Valley keeps using it anyway" — that's how the headline of the Rest of World piece reads.

What this means

Leadership in artificial intelligence is increasingly determined not by access to chips, but by whose models actually run in products. While Washington builds walls around hardware, open Chinese models like Kimi K3 are seeping into the Western tech stack — including companies that advocated for those very walls. The US-China tech rivalry is shifting from the question of "who has better chips" to "whose models the whole world uses."