Indian startups are rebuilding their products on Chinese foundations. Qwen, DeepSeek and Kimi allowed significant cost savings, run almost as well as the American frontier, and lag it by about six months. Nikkei Asia reported in July 2026 that Indian companies are increasingly switching to Chinese large language models (LLMs) to contain Artificial Intelligence (AI) costs, with one venture investor noting that startups are cutting costs by an order of magnitude.

This open-weight generosity is neither charity nor a workaround for chip controls. It is a strategy supported by different independent logics that make it durable. But durable is not permanent. The assessment by these writers is that China will begin graduating access to its frontier open-weight models around late 2028.

China’s strategy rests on five reinforcing logics. The first is cost. DeepSeek trained their R1 model for $294,000, a fraction of the costs OpenAI or Anthropic incur. Distillation from American models and breakthroughs in architectural efficiency have compressed research and development spending.

The second is prestige. DeepSeek’s January 2025 release wiped roughly a trillion dollars off U.S. tech stocks. Chinese President Xi Jinping’s appearance at the Shanghai AI conference, the 29-country World Artificial Intelligence Cooperation Organization (WAICO) bloc, the 5,000 training slots for developing countries are examples of open-weighting converted into diplomacy.

The third is commoditisation. American labs monetise proprietary weights. Free models good enough for most commercial work hurt their pricing power. Chinese AI companies do not have to beat their competitor’s product; only destroy their ability to charge for it.

The fourth is capital. Financial repression traps Chinese household savings in state banks, which lend cheaply to strategic sectors. The result is the same state subsidisation and overcapacity that flattened global solar and electric vehicle markets. In AI, this dynamic has led to 820 LLMs registered with China’s cyberspace authority by early 2026.

The fifth is infrastructure. Free models drive AI adoption which drives demand for complementary products that China dominates — energy, cloud and physical infrastructure. Alibaba’s cloud revenue grew 34% year-on-year while it gave Qwen away.

What would make Beijing close the gates

The Financial Times reported in July that Chinese regulators led by the Ministry of Commerce have been consulting leading AI companies including Alibaba, Bytedance and Zhipu on two issues: limiting the transfer of training data abroad and whether foreign users should continue to be able to freely download the model weights of Chinese AI systems. But it is clear that three conditions have to be met before restriction becomes rational for the Communist Party.

The first is consolidation. Beijing can coordinate five firms; it cannot coordinate 800. Xinhua has already announced the shift from the “Hundred Model War” to the “Top Five Basic Models.” The irony is that U.S. export controls, by raising costs for Chinese labs through chip export controls, are accelerating the very consolidation that makes Chinese restriction feasible.

The second is lock-in. If China restricts access before global developers are deeply embedded in its cloud stack, they simply migrate elsewhere and the flywheel breaks. This threshold is currently far from being reached.

The third is saturation. Once pricing power of frontier American AI labs is sufficiently commoditised and non-Chinese open-weight releases (from Meta, Mistral, Nvidia and others) are sustaining the pressure independently, further Chinese releases buy nothing. Again, the gap is narrowing here but still exists.

The likely outcome is therefore not a switch flipped but different pathways with graduated restrictions: Frontier models served via API first and weights released after a six-month embargo; commercial licensing above a capability threshold, with smaller distilled models left free as the on-ramp; model weights released openly, but tool-use and agentic scaffolding withheld; preferential access for WAICO members. The assessment by these writers is that graduated restriction begins appearing around late 2028.

What India should do about it

India should leverage the open ecosystem while pricing in the switching costs. Government departments and regulated sectors should be built on model-agnostic architectures such as abstraction layers and harnesses that work across stacks. Instead of offering compute subsidies on GPU slices, the Ministry of Electronics and Information Technology (MeitY) should consider using an OpenRouter equivalent for the public sector.

Focus on areas where India can actually win such as applications, industrial and language data, edge inference silicon design, and domain-specific fine-tuning. This is atmashakti (self-strength through increasing capabilities in a few selective segments), rather than full self-sufficiency, which India cannot afford and does not need.

Lastly, India should use the window diplomatically. It should be shaping open-weight norms in multilateral fora, while the commons is still open and Beijing still needs legitimacy for it.

For the most part, India is consuming and integrating AI into applications that address its needs, rather than building the frontier. Rapid diffusion to enhance productivity across sectors matters far more than AI sovereignty theatre, and today’s open models from both the U.S. and Chinese ecosystems are more than adequate for that. But the reason they are adequate is that a strategic competition is currently being fought by giving them away. That competition will not stay in this phase forever. India should leverage the open ecosystems while making sure that it can survive disruptions that come along the way.

Bharath Reddy and Pranay Kotasthane are researchers with the high-tech geopolitics programme at the Takshashila Institution, an independent centre for research and education in public policy. The views expressed are personal.