China’s open AI advantage may not last forever

GS III – Science & Technology (Emerging Technologies) | Artificial Intelligence (AI)

• Chinese LLMs: Qwen, DeepSeek and Kimi have allowed considerable cost reductions, achieved comparable results against the best American frontier models, and are moving with a six-month lag.
• Chinese LLMs & Indian Startups: According to Nikkei Asia, July 2026, Indian startups are increasingly opting for Chinese LLMs to control AI costs; one venture investor said startups are reducing their costs by at least an order of magnitude.
• Open-Weight Strategy: China is not a charity case, and its open-weight strategy is not just a loophole to get around chip controls; it operates on distinct, independent logics that lend it long-term durability.
• Future Ban: However, it is ultimately not permanent. The prognosis is that China might start to phase out access to its frontier open-weight models by the end of 2028.

Five Logics Behind China’s Strategy

• Cost: DeepSeek trained its R1 model for $294,000, which is a small fraction of the costs incurred by OpenAI or Anthropic; distillation from American models plus architectural efficiency has compressed R&D spending.
• Bragging & Prestige: DeepSeek’s January 2025 unveiling wiped out circa $1 trillion from U.S. tech stocks; Xi Jinping’s presence at the Shanghai AI conference, the 29-country WAICO bloc, and the announcement of training slots for around 5,000 people from developing countries show how open-weighting is being transferred into diplomacy.
• Commoditisation: Free models that are adequate for most commercial work reduce the pricing power of proprietary models; a Chinese AI company does not necessarily have to beat competitors’ products but can instead reduce their ability to charge for them.
• Capital: The Chinese model of development sees household savings trapped in state banks, which lend cheaply to strategic sectors, contributing to state subsidisation and massive overcapacity, as seen in global solar and electric vehicle markets; this has contributed to 820 LLMs registered with China’s cyberspace authority by early 2026.
• Infrastructure: Free models accelerate AI adoption, which fuels demand for China-dominated goods needed to run it — energy, cloud and physical infrastructure; Alibaba grew its cloud revenue 34% YoY while giving away Qwen.

What Would It Take for Beijing to Shut the Gates?

• Regulatory Consultation: Chinese regulators led by the Ministry of Commerce have sought advice from leading AI companies such as Alibaba, ByteDance and Zhipu on curbing exports of training data abroad and whether foreign users should continue to freely download weights of Chinese AI systems.
• Consolidation: Beijing can coordinate five firms, not 800, moving from the “Hundred Model War” to “Top Five Basic Models”, as reported by Xinhua.
• U.S. Export Controls: U.S. chip export controls are driving up lab costs in China and accelerating consolidation that makes Chinese restriction possible.
• Lock-in: There is no lock-in if global developers can move elsewhere when access is restricted before they are heavily invested in the Chinese cloud stack; the flywheel breaks, and this threshold is currently a long way off.
• Saturation: After frontier American AI labs have sufficiently commoditised their pricing power, non-Chinese open-weight releases from Meta, Mistral, Nvidia and others can maintain this pressure independently; additional Chinese releases may then buy nothing, although a narrowing gap continues to exist.

Likely Graduated Restrictions

• API Policy: Frontier model weights may not remain freely available indefinitely; API access to frontier models could come first, with weights released later.
• Commercial Licensing: Commercial licensing could apply above a particular capability level, while smaller distilled models may continue to remain free as an on-ramp.
• Weights vs Agentic Capabilities: Model weights may be made public, while tool-use structures and agentic scaffolding remain private.
• Preferential Access: Priority access could be provided to WAICO members.
• Timeline: Graduated restrictions could first be seen approximately around late 2028.

What India Should Do

• Make Use of Open Ecosystem: India must leverage the open ecosystem but price in the switching costs.
• Model-Agnostic Architecture: Model-agnostic approaches are important strategies for government departments and regulated sectors; abstraction layers and harnesses that span stacks should be used.
• AI for Public Sector: Instead of giving compute subsidies on GPU slices in the public sector, MeitY should perhaps supply an equivalent of OpenRouter for the general public.
• Priority Areas: India should focus attention on applications, training data, edge inference silicon design and domain-specific fine-tuning.
• Atmashakti: This means self-strength through increasing capabilities in a few selective segments, rather than full self-sufficiency, which India cannot afford and does not need.
• Diplomatic Window: India must influence open-weight norms in multilateral fora while the commons remain accessible and Beijing still requires legitimacy for it.
• AI Adoption: AI is not an end in itself — in India, it is about consuming from the frontier and building applications that serve its needs.
• Productivity: The fast diffusion of AI to enhance productivity across sectors is more important than AI sovereignty theatre.
• Open Model Coverage: Open models today from the major U.S. and Chinese ecosystems are more than sufficient for this purpose.
• Strategic Competition: They are sufficient because strategic competition is currently being fought through giveaways.
• Future Dislocations: This competition will not remain at this stage forever; India should maximise its use of open ecosystems and build the capacity to survive future dislocations.

Conclusion

• Strategic Balance: Harness the open AI ecosystem while preparing for future restrictions.
• Selective Capability: The emphasis should be on AI diffusion, applications and selective capability-building rather than achieving complete self-reliance.
• Resilience: The present open ecosystem should be utilised while ensuring that India has the foundation to withstand future disruptions — RESILIENCE.

UPSC Mains Practice Question

Q. “There is an opportunity and a strategic challenge for India as it grows more reliant on Chinese open-weight AI models.” Investigate the problem and explain what steps India can take to make its AI ecosystem more robust.