China’s artificial-intelligence industry is mounting its most direct challenge yet to Silicon Valley, not by building higher walls around its technology but by giving much of it away.
A new generation of Chinese open-weight models, led by Moonshot AI’s Kimi K3, is pushing into the trillion-parameter range once associated almost exclusively with the most expensive proprietary systems. Kimi K3 contains 2.8 trillion parameters, according to its developer, while Alibaba has previewed a Qwen model with roughly 2.4 trillion. The releases suggest that China’s AI companies are betting that scale, low prices and downloadable models can weaken the commercial advantages enjoyed by OpenAI, Anthropic and Google.
The strategy is also a response to American technology restrictions. Washington has sought to slow China’s progress by limiting its access to Nvidia’s most advanced processors and the equipment needed to manufacture cutting-edge chips. Chinese developers, however, have increasingly focused on efficient model architectures, domestic computing infrastructure and open distribution. Export controls can restrict hardware shipments, but they are less effective against software weights that can be downloaded, copied and modified around the world.
The result is an increasingly uncomfortable prospect for the US: Chinese companies may not need to dominate the most advanced semiconductor supply chain to shape how much of the world uses artificial intelligence.
China Turns Openness Into a Competitive Weapon
Moonshot AI, a Beijing-based startup backed by investors including Alibaba, released Kimi K3 in July as an open-weight model capable of processing text, images and unusually long sequences of information. Its one-million-token context window is designed to handle large software repositories, research archives and complex agent-based tasks without repeatedly breaking material into smaller pieces.
Kimi’s total parameter count is striking, though it doesn’t mean that every parameter is used for every request. Like many of China’s largest systems, Kimi relies on a mixture-of-experts architecture that activates only part of the network at a time. That design can provide the capacity of an enormous model while reducing the computing cost of each response.
The model’s significance lies as much in its distribution as in its size. Open-weight systems allow companies and researchers to download the parameters that encode the model’s learned behaviour. Developers can then fine-tune the system, run it on private infrastructure or modify it for local languages and specialised industries. That contrasts with closed models, which are generally accessed through interfaces controlled by their American developers.
Chinese models have already gained considerable traction on platforms such as Hugging Face. DeepSeek and Alibaba’s Qwen family helped Chinese developers overtake US groups in some measures of open-model downloads, while state-backed media now portray open AI as a public good for developing economies. The appeal is practical: governments and businesses in Asia, Africa and Latin America can adapt a Chinese model without sending sensitive data to a US cloud provider or paying premium fees for every query.
Parameter counts alone don’t guarantee better performance. A smaller model trained on better data can outperform a larger one, and independent researchers have warned that Kimi K3’s enormous size could make it expensive to host. Still, its release demonstrates that Chinese laboratories can train and distribute systems approaching the scale of the world’s most heavily financed AI projects.
Silicon Valley Confronts a New Price War
China’s open-model push threatens to compress the value of the foundational model itself. If developers can obtain strong reasoning and coding capabilities at little or no licensing cost, the profits may migrate toward cloud infrastructure, proprietary data and specialised applications rather than the companies that trained the original systems.
That possibility recalls the market shock caused by DeepSeek in early 2025, when claims of lower-cost Chinese training briefly erased hundreds of billions of dollars from the value of US technology companies. The selloff reflected fears that demand for costly computing clusters—and the pricing power of closed-model providers—had been overstated. Subsequent Chinese releases have reinforced the argument that frontier-level capabilities could become cheaper and more widely available than investors expected.
The pressure is producing unlikely alliances in Washington. Nvidia, Microsoft, Meta, IBM and other technology companies have urged the US government to avoid premature restrictions on open-weight models. Their argument is that open systems support research, startups and American technological influence. OpenAI, despite building much of its business around closed models, has also supported the broad case against sweeping restrictions.
The industry’s concern is that limiting American open models could surrender the global developer ecosystem to China. A programmer barred from using an advanced US model won’t necessarily abandon open AI; the programmer may simply adopt Qwen, DeepSeek or Kimi instead. Once applications and technical expertise become organised around those systems, American companies could struggle to win developers back.
Anthropic has taken a more cautious position, warning that the most capable downloadable models can be modified to remove safeguards. Open weights can aid scientific research and local innovation, but they can also be repurposed for cyberattacks, propaganda or biological research. Unlike a cloud-based service, a downloaded model can’t easily be recalled or remotely disabled.
An AI Contest That Hardware Controls Can’t Settle
China still faces serious constraints. Its companies have limited access to the latest Nvidia accelerators, and running trillion-parameter systems requires vast amounts of memory, electricity and networking equipment. Many users will continue to access Chinese models through cloud services rather than installing them locally, giving the largest technology groups an advantage over smaller competitors.
Yet China’s strategy doesn’t require every business to operate a 2.8-trillion-parameter model. Large systems can train smaller versions, generate synthetic data and anchor families of specialised models. Their weights can also spread through universities, startups and foreign cloud providers, extending China’s influence far beyond the companies that originally developed them.
That changes the nature of the US-China AI contest. Washington has treated advanced chips as the central bottleneck and assumed that control over computing hardware would preserve America’s lead. The rapid improvement of Chinese open-weight models suggests that algorithms, engineering efficiency and distribution networks may be nearly as important.
China isn’t abandoning proprietary AI. ByteDance, Baidu, Tencent and Alibaba are still building commercial cloud platforms and consumer services. Instead, the country’s leading developers are using openness selectively—releasing enough technology to attract global users while earning revenue from hosting, applications and enterprise support.
For Silicon Valley, the danger isn’t simply that one Chinese model will surpass ChatGPT or Claude. It is that capable models will become interchangeable commodities and that the world’s developers will build on whichever system is cheapest, easiest to modify and least constrained. China’s trillion-parameter releases are an audacious attempt to ensure that, when that choice is made, an increasingly large share of the world chooses Chinese.
